您的位置:山东大学 -> 科技期刊社 -> 《山东大学学报(工学版)》

山东大学学报 (工学版) ›› 2022, Vol. 52 ›› Issue (6): 105-114.doi: 10.6040/j.issn.1672-3961.0.2021.304

• 机器学习与数据挖掘 • 上一篇    下一篇

基于空洞卷积块架构的命名实体识别模型

袁钺1(),王艳丽2,刘勘2,*()   

  1. 1. 北京大学信息管理系,北京 100871
    2. 中南财经政法大学信息与安全工程学院,湖北 武汉 430073
  • 收稿日期:2021-06-09 出版日期:2022-12-20 发布日期:2022-12-23
  • 通讯作者: 刘勘 E-mail:yuangyue@qq.com;liukan@zuel.edu.cn
  • 作者简介:袁钺(1996—),男,湖北武汉人,博士研究生,主要研究方向为计算型情报分析、自然语言处理、深度学习。E-mail: yuangyue@qq.com
  • 基金资助:
    中央高校基本科研业务费交叉学科创新研究项目(2722021EK016)

Named entity recognition model based on dilated convolutional block architecture

Yue YUAN1(),Yanli WANG2,Kan LIU2,*()   

  1. 1. Department of Information Management, Peking University, Beijing 100871, China
    2. School of Information and Safety Engineering, Zhongnan University of Economics and Law, Wuhan 430073, Hubei, China
  • Received:2021-06-09 Online:2022-12-20 Published:2022-12-23
  • Contact: Kan LIU E-mail:yuangyue@qq.com;liukan@zuel.edu.cn

摘要:

受到空洞卷积的启发提出面向二维文本嵌入的列式空洞卷积,设计空洞卷积块架构,基于此架构提出命名实体识别模型并开展进一步试验。在命名实体识别试验中,提出的模型的精密度、召回率和F1超越了其他基线模型,分别达到了0.918 7、0.879 4和0.898 6,表明空洞卷积块架构能够获取包含更多上下文信息的文本特征,从而支持模型对上下文长距离依赖特征的捕获和处理。感受野试验表明需要适当调整空洞率以减轻空洞卷积给模型带来的“网格效应”。提出的基于空洞卷积块架构能有效执行命名实体识别任务。

关键词: 命名实体识别, 空洞卷积块架构, 感受野, 神经网络, 深度学习

Abstract:

Inspired by the dilated convolution, a column-wise dilated convolution towards two dimensional text embedding was proposed and a dilated convolutional block architecture was designed. A named entity recognition model based on the architecture was built for further experiments. In the named entity recognition experiment, the model surpassed other baseline models in the metrics of precision, recall, and F1 value, respectively reaching 0.918 7, 0.879 4, and 0.898 6, indicating that the dilated convolutional block architecture obtained features from context information, thereby supporting the extraction of the long-term dependency. The receptive field experiment showed that it was necessary to jointly adjust the dilation rate and the convolution kernel size to reduce the "gridding effect". The dilated convolutional block architecture proposed could effectively perform the task of named entity recognition.

Key words: named entity recognition, dilated convolutional block architecture, receptive field, neural network, deep learning

中图分类号: 

  • TP391

图1

增加卷积核尺寸或空洞率感受野变化示例"

图2

堆叠卷积层时标准卷积和空洞卷积感受野变化"

图3

空洞卷积块结构"

图4

二维文本嵌入与处理图像的空洞卷积"

图5

列式空洞卷积示例"

图6

空洞卷积块架构"

图7

基于空洞卷积块架构的命名实体识别模型"

图8

基于空洞卷积层的命名实体识别模型"

表1

数据集统计"

新闻句子数量 平均句子长度(汉字字符) 最长句子长度(汉字字符) 最短句子长度(汉字字符) 人物标签数量 地点标签数量 组织标签数量
50 729 47 1476 5 19 588 39 394 21 902

表2

命名实体识别试验结果(地点)"

是否有CRF层 模型 P R F1
没有CRF层 Bi-LSTM-Softmax 0.750 2 0.715 0 0.732 2
IDCNN-Softmax 0.769 9 0.770 9 0.770 4
DCL-Bi-LSTM-Softmax 0.792 9 0.758 4 0.775 3
DCBA-Bi-LSTM-Softmax 0.797 6 0.824 8 0.811 0
有CRF层 Bi-LSTM-CRF 0.869 9 0.808 5 0.838 0
IDCNN-CRF 0.900 9 0.824 8 0.861 2
DCL-Bi-LSTM-CRF 0.906 1 0.861 7 0.883 3
DCBA-Bi-LSTM-CRF 0.918 7 0.879 4 0.898 6

表3

命名实体识别试验结果(人物)"

是否有CRF层 模型 P R F1
没有CRF层 Bi-LSTM-Softmax 0.764 5 0.770 7 0.767 6
IDCNN-Softmax 0.797 8 0.783 8 0.790 7
DCL-Bi-LSTM-Softmax 0.816 9 0.816 5 0.816 7
DCBA-Bi-LSTM-Softmax 0.823 0 0.796 9 0.809 7
有CRF层 Bi-LSTM-CRF 0.846 0 0.817 0 0.831 3
IDCNN-CRF 0.849 7 0.815 0 0.832 0
DCL-Bi-LSTM-CRF 0.897 4 0.837 7 0.866 5
DCBA-Bi-LSTM-CRF 0.889 1 0.824 1 0.855 3

表4

命名实体识别试验结果(组织)"

是否有CRF层 模型 P R F1
没有CRF层 Bi-LSTM-Softmax 0.543 3 0.622 1 0.580 0
IDCNN-Softmax 0.631 8 0.676 9 0.653 6
DCL-Bi-LSTM-Softmax 0.633 8 0.652 9 0.643 2
DCBA-Bi-LSTM-Softmax 0.681 1 0.733 3 0.706 2
有CRF层 Bi-LSTM-CRF 0.755 6 0.731 8 0.743 5
IDCNN-CRF 0.774 9 0.770 8 0.772 9
DCL-Bi-LSTM-CRF 0.835 4 0.777 6 0.805 4
DCBA-Bi-LSTM-CRF 0.835 3 0.815 2 0.825 1

表5

命名实体识别试验结果(合计)"

是否有CRF层 模型 P R F1
没有CRF层 Bi-LSTM-Softmax 0.704 4 0.712 9 0.708 6
IDCNN-Softmax 0.747 1 0.754 8 0.751 0
DCL-Bi-LSTM-Softmax 0.765 0 0.754 4 0.759 6
DCBA-Bi-LSTM-Softmax 0.779 0 0.796 2 0.787 5
有CRF层 Bi-LSTM-CRF 0.837 0 0.794 7 0.815 3
IDCNN-CRF 0.855 8 0.810 1 0.832 3
DCL-Bi-LSTM-CRF 0.888 3 0.835 9 0.861 3
DCBA-Bi-LSTM-CRF 0.891 0 0.847 9 0.868 9

图9

使用CRF的模型学习曲线"

图10

使用Softmax的模型学习曲线"

表6

感受野试验设置及结果(设置1)"

模型 r S(高×宽×数量) 感受野 P R F1
DCL-Bi-LSTM-Softmax 1 10×50×1 10 0.642 8 0.646 2 0.644 5
DCBA-Bi-LSTM-Softmax 1 10×50×1 10 0.689 0 0.721 7 0.705 0

表7

感受野试验设置及结果(设置2)"

模型 r S(高×宽×数量) 感受野 P R F1
DCL-Bi-LSTM-Softmax 2 10×50×1 19 0.658 0 0.646 5 0.652 2
DCBA-Bi-LSTM-Softmax 2 10×50×1 19 0.723 4 0.744 2 0.733 6

表8

感受野试验设置及结果(设置3)"

模型 r S(高×宽×数量) 感受野 P R F1
DCL-Bi-LSTM-Softmax 4 10×50×1 37 0.673 2 0.657 5 0.665 3
DCBA-Bi-LSTM-Softmax 4 10×50×1 37 0.754 1 0.756 1 0.755 1

表9

感受野试验设置及结果(设置4)"

模型 r S(高×宽×数量) 感受野 P R F1
DCL-Bi-LSTM-Softmax 8 10×50×1 73 0.650 2 0.643 7 0.647 0
DCBA-Bi-LSTM-Softmax 8 10×50×1 73 0.753 2 0.751 6 0.752 4
1 PANCHENDRARAJAN R, AMARESAN A. Bidirectional LSTM-CRF for named entity recognition[C]//Proceedings of the 32nd Pacific Asia Conference on Language, Information and Computation. Hong Kong, China: Association for Computational Linguistics, 2018: 531-540.
2 LI L , XU W , YU H . Character-level neural network model based on Nadam optimization and its application in clinical concept extraction[J]. Neurocomputing, 2020, 414 (16): 182- 190.
3 SHARMA R , MORWAL S , AGARWAL B , et al. A deep neural network-based model for named entity recognition for Hindi language[J]. Neural Computing and Applications, 2020, 32 (20): 16191- 16203.
doi: 10.1007/s00521-020-04881-z
4 WU C , WU F , QI T , et al. Detecting entities of works for chinese chatbot[J]. ACM Transactions on Asian and Low-Resource Language Information Processing (TALLIP), 2020, 19 (6): 1- 13.
5 LI X , ZHANG H , ZHOU X H . Chinese clinical named entity recognition with variant neural structures based on BERT methods[J]. Journal of Biomedical Informatics, 2020, 107 (18): 103422.
6 JIA C, SHI Y, YANG Q, et al. Entity enhanced bert pre-training for chinese NER[C]//Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). Punta Cana, Dominica: Association for Computational Linguistics, 2020: 6384-6396.
7 HAN Y, YAN Y, HAN Y, et al. Chinese grammatical error diagnosis based on RoBERTa-BiLSTM-CRF model[C]//Proceedings of the 6th Workshop on Natural Language Processing Techniques for Educational Applications. Suzhou, China: Association for Computational Lingu-istics, 2020: 97-101.
8 YANG Z, DAI Z, YANG Y, et al. Xlnet: generalized autoregressive pretraining for language understanding[C]// Proceedings of Advances in Neural Information Processing Systems. Vancouver, Canada: MIT Press, 2019: 5753-5763.
9 DEVLIN J, CHANG M W, LEE K, et al. Bert: pre-training of deep bidirectional transformers for language understanding[C]//Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Minneapolis, USA: Association for Computational Linguistics, 2019: 4171-4186.
10 ZHANG Z, HAN X, LIU Z, et al. ERNIE: enhanced language representation with informative entities[C]// Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. Florence, Italy: Association for Computational Linguistics, 2019: 1441-1451.
11 CHEN L C , PAPANDREOU G , KOKKINOS I , et al. Deeplab: semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 40 (4): 834- 848.
12 WANG Z , JI S . Smoothed dilated convolutions for improved dense prediction[J]. Data Mining and Knowledge Discovery, 2021, 35 (4): 1- 27.
13 MEHTA S, RASTEGARI M, CASPI A, et al. Espnet: efficient spatial pyramid of dilated convolutions for semantic segmentation[C]//Proceedings of the European Conference on Computer Vision (ECCV). Munich, Germany: Springer, 2018: 552-568.
14 STRUBELL E, VERGA P, BELANGER D, et al. Fast and accurate entity recognition with iterated dilated convolutions[C]//Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing. Copenhagen, Denmark: Association for Computational Linguistics, 2017: 2670-2680.
15 KALCHBRENNER N, GREFENSTETTE E, BLUNSOM P. A convolutional neural network for modelling sentences[C]// Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics. Baltimore, USA: Association for Computational Linguistics, 2014: 655-665.
16 GLOROT X, BORDES A, BENGIO Y. Deep sparse rectifier neural networks[C]//Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics. Ft. Lauderdale, USA: AISTATS, 2011: 315-323.
17 HE K, ZHANG X, REN S, et al. Deep residual learning for image recognition[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas, USA: IEEE, 2016: 770-778.
18 WANG P, CHEN P, YUAN Y, et al. Understanding convolution for semantic segmentation[C]//Proceedings of the 2018 IEEE Winter Conference on Applications of Computer Vision (WACV). Tahoe City, USA: IEEE, 2018: 1451-1460.
19 LIPPMANN R, CAMPBELL W, CAMPBELL J. An overview of the darpa data driven discovery of models (d3m) program[C]//Proceedings of 29th Conference on Neural Information Processing Systems. Barcelona, Spain: MIT Press, 2016: 1-2.
[1] 李旭涛,杨寒玉,卢业飞,张玮. 基于深度学习的遥感图像道路分割[J]. 山东大学学报 (工学版), 2022, 52(6): 139-145.
[2] 王智伟,徐海超,郭相阳,马炯,褚云龙,陈前昌,卢治. 基于卷积神经网络和层次分析的新能源电源调频能力智能预测方法[J]. 山东大学学报 (工学版), 2022, 52(5): 70-76.
[3] 孟令灿,聂秀山,张雪. 基于遮挡目标去除的公交车拥挤度分类算法[J]. 山东大学学报 (工学版), 2022, 52(4): 83-88.
[4] 黄皓,周丽华,黄亚群,姜懿庭. 基于混合深度模型的虚假信息早期检测[J]. 山东大学学报 (工学版), 2022, 52(4): 89-98.
[5] 孟祥飞,张强,胡宴才,张燕,杨仁明. 欠驱动船舶自适应神经网络有限时间跟踪控制[J]. 山东大学学报 (工学版), 2022, 52(4): 214-226.
[6] 杨霄,袭肖明,李维翠,杨璐. 基于层次化双重注意力网络的乳腺多模态图像分类[J]. 山东大学学报 (工学版), 2022, 52(3): 34-41.
[7] 张学思,张婷,刘兆英,江天鹏. 基于轻量型卷积神经网络的海面红外显著性目标检测方法[J]. 山东大学学报 (工学版), 2022, 52(2): 41-49.
[8] 王心哲,邓棋文,王际潮,范剑超. 深度语义分割MRF模型的海洋筏式养殖信息提取[J]. 山东大学学报 (工学版), 2022, 52(2): 89-98.
[9] 尹旭,刘兆英,张婷,李玉鑑. 基于弱监督和半监督学习的红外舰船分割方法[J]. 山东大学学报 (工学版), 2022, 52(2): 99-106.
[10] 蒋桐雨,陈帆,和红杰. 基于非对称U型金字塔重建的轻量级人脸超分辨率网络[J]. 山东大学学报 (工学版), 2022, 52(1): 1-8, 18.
[11] 吴建清,宋修广. 同步定位与建图技术发展综述[J]. 山东大学学报 (工学版), 2021, 51(5): 16-31.
[12] 丁飞,江铭炎. 基于改进狮群算法和BP神经网络模型的房价预测[J]. 山东大学学报 (工学版), 2021, 51(4): 8-16.
[13] 尹晓敏,孟祥剑,侯昆明,陈亚潇,高峰. 一种计及空间相关性的光伏电站历史出力数据的修正方法[J]. 山东大学学报 (工学版), 2021, 51(4): 118-123.
[14] 杨修远,彭韬,杨亮,林鸿飞. 基于知识蒸馏的自适应多领域情感分析[J]. 山东大学学报 (工学版), 2021, 51(3): 15-21.
[15] 陶亮,刘宝宁,梁玮. 基于CNN-LSTM 混合模型的心律失常自动检测[J]. 山东大学学报 (工学版), 2021, 51(3): 30-36.
Viewed
Full text


Abstract

Cited

  Shared   
  Discussed   
[1] 何东之, 张吉沣, 赵鹏飞. 不确定性传播算法的MapReduce并行化实现[J]. 山东大学学报(工学版), 0, (): 22 -28 .
[2] 刘云,邱晓国 . 内插TOC系数法测定水体中COD研究[J]. 山东大学学报(工学版), 2007, 37(4): 108 -117 .
[3] 顿月芹 闵越 袁建生. 阵列侧向测井正演响应的特性分析[J]. 山东大学学报(工学版), 2010, 40(1): 121 -125 .
[4] 吴俊飞,王威强,胡德栋,崔玉良, . 平阴尿塔塔体爆炸能量分析与计算[J]. 山东大学学报(工学版), 2008, 38(4): 80 -83 .
[5] 田伟,乔谊正,马志强 . 基于DWT的二次特征提取脱机中文签名鉴定[J]. 山东大学学报(工学版), 2007, 37(3): 55 -59 .
[6] 林彦,魏东 . 铸钢空心球管节点的破坏机理分析与承载力影响因素[J]. 山东大学学报(工学版), 2006, 36(3): 103 -107 .
[7] 郑桂兰,马庆雷,潘秀华 . 旧水泥混凝土路面沥青加铺层结构设计方法研究[J]. 山东大学学报(工学版), 2007, 37(4): 93 -97 .
[8] 杨国辉1,孙晓瑜1,2,椿范立1. 应用沸石胶囊催化剂制备生物汽油(英文)[J]. 山东大学学报(工学版), 2009, 39(2): 92 -97 .
[9] 赵伟,艾洪奇. pH对Aβ42小纤维的结构影响[J]. 山东大学学报(工学版), 2018, 48(2): 134 -138 .
[10] 钟倩倩,岳钦艳*,李倩,李颖,许醒,高宝玉. 改性麦草秸秆对活性艳红的吸附动力学研究[J]. 山东大学学报(工学版), 2011, 41(1): 133 -139 .