南京大学学报(自然科学版) ›› 2023, Vol. 59 ›› Issue (6): 1013–1022.doi: 10.13232/j.cnki.jnju.2023.06.011

• • 上一篇    

基于BoBGSAL⁃Net的文档级实体关系抽取方法

冯超文1,2, 吴瑞刚1,2, 温绍杰1,2, 刘英莉1,2()   

  1. 1.昆明理工大学信息工程与自动化学院，昆明，650500
    2.云南省计算机技术应用重点实验室，昆明理工大学，昆明，650500
  • 收稿日期:2023-08-20 出版日期:2023-11-30 发布日期:2023-12-06
  • 通讯作者: 刘英莉 E-mail:lyl@kust.edu.cn
  • 基金资助:
    国家自然科学基金(52061020);云南计算机技术应用重点实验室开放基金(2020103);云南省重大科技专项计划项目(202302AG050009)

Document⁃level entity relation extraction method based on BoBGSAL⁃NET

Chaowen Feng1,2, Ruigang Wu1,2, Shaojie Wen1,2, Yingli Liu1,2()   

  1. 1.Faculty of Information Engineering and Automation，Kunming University of Science and Technology，Kunming，650500，China
    2.Yunnan Key Laboratory of Computer Technology Application，Kunming University of Science and Technology, Kunming，650500，China
  • Received:2023-08-20 Online:2023-11-30 Published:2023-12-06
  • Contact: Yingli Liu E-mail:lyl@kust.edu.cn

RichHTML

0

PDF (PC)

149

摘要：

文档级实体关系抽取的主要任务是提取文档中实体之间的关系.相较于句内实体关系提取，文档级实体关系抽取需要对文档中多个句子进行推理.为了解决文档中不同实体之间的复杂信息交互问题，提出一个混合提及级图MMLG （Mixed Mention?Level Graph）策略，用于拟合文档中不同实体之间的复杂信息交互，提高模型对于文档级实体关系的感知能力.此外，为了应对实体关系中存在的关系重叠问题，构建了实体关系图ERG （Entity Relation Graph）模块，该模块融合了路径推理机制，主要针对实体间的多个关系路径进行推理学习，更准确地识别提及级节点实体及关系.通过将MMLG策略与ERG模块聚合到实体关系抽取模型中，构建BoBGSAL?Net （Based on Bipartite Graph Structure Aggregate Logic Network）模型，并在公开数据集DocRED和作者实验室构建的数据集AlSiaRED上开展实验，结果证明BoBGSAL?Net在文档级实体关系抽取任务中性能得到提升，其中BoBGSAL?Net+BERT模型在AlSiaRED数据集上的关系抽取任务中F1指标达到66.04%，和其他模型相比，整体性能提升了4.4%，泛化能力突出，综合效果最优.

关键词: 文档级实体关系抽取, 混合提及级图, 实体关系图, BoBGSAL?Net模型

Abstract:

The primary task of document?level entity relation extraction is to extract relationships among entities in a document. Compared to intra?sentence entity relation extraction，document?level entity relation extraction requires reasoning across multiple sentences in the document. To address the challenge of complex information interaction among different entities in the document，this paper proposes a Mixed Mention?Level Graph (MMLG) strategy for modeling intricate information interaction among different entities in the document，thereby enhancing the model's perception of document?level entity relations. Additionally，to handle the issue of relationship overlap within entity relations in documents，an Entity Relation Graph (ERG) module is constructed，incorporating a path reasoning mechanism that focuses on inferring and learning from multiple relationship paths among entities. This module enhances the accurate identification of entity and relation nodes at the mention level.By integrating the MMLG strategy and ERG module into the entity relation extraction model，this paper develops the BoBGSAL?Net (Based on Bipartite Graph Structure Aggregate Logic Network) model. Experimental evaluations are conducted on the publicly available DocRED dataset and the AlSiaRED dataset created by the authors' laboratory. The experimental results demonstrate the performance improvement of BoBGSAL?Net in document?level entity relation extraction tasks. Notably，the BoBGSAL?Net+BERT model achieves an F1 score of 66.04% in relation extraction tasks on the AlSiaRED dataset，showcasing a 4.4% overall performance improvement compared to other models. The model exhibits exceptional generalization capability，culminating in an optimal comprehensive performance.

Key words: document?level entity relation extraction, mixed mention?level graph, entity relation graph, BoBGSAL?Net model

中图分类号: 

  • TP183

图1

BoBGSAL?Net结构图"

表1

服务器的详细配置"

操作系统Ubuntu 20.04 LST
CPU型号Inter Xeon Gold 5120 (56) CPU @2.2GHZ
CPU存储256 G
GPU型号NVIDIA Tesla V100
GPU存储16 G

表2

核心依赖工具包"

安装包版本
CUDA10.2
Python3.7.5
Matplotlib3.3.5
NumPy1.19.4
Torch1.6.0
Transformers3.1.0
Scikit⁃learn0.23.2

表3

BoBGSAL?Net模型和其他模型在DocRED数据集上的命名实体识别实验结果的对比"

模型验证集测试
Ign F1Ign AUCF1AUCIgn F1F1
BoBGSAL⁃Net54.33%53.75%55.84%54.97%54.14%55.08%
CNN[12]41.58%36.85%43.45%39.39%40.33%42.26%
LSTM[5]48.44%46.62%50.68%49.48%47.71%50.07%
BiLSTM[6]48.87%47.61%50.94%50.26%48.78%51.06%
Context⁃Aware[21]48.94%47.22%51.09%50.17%48.40%50.70%
HIN⁃GloVe[7]51.06%—52.95%—51.15%53.30%
CFER⁃GloVe[27]54.29%—55.31%—53.70%54.06%
SSAN⁃BERT⁃base[28]54.03%—54.95%—53.44%53.16%
GAIN+SIEF[29]53.82%—54.24%—53.87%54.79%

表4

BoBGSAL?Net模型和其他模型在AlSiaRED数据集上的命名实体识别实验结果的对比"

模型验证集测试
Ign F1Ign AUCF1AUCIgn F1F1
BoBGSAL⁃Net53.66%53.19%55.39%55.23%52.55%54.83%
CNN[12]39.53%31.47%40.15%32.44%38.73%39.20%
LSTM[5]41.34%40.43%43.03%41.09%41.26%42.97%
BiLSTM[6]44.08%43.65%46.57%45.13%43.24%45.16%
Context⁃Aware[21]46.09%45.36%48.85%47.33%46.13%48.17%
HIN⁃GloVe[7]48.38%—50.35%—48.24%50.18%
CFER⁃GloVe[27]53.34%—54.27%—52.45%53.60%
SSAN⁃BERT⁃base[28]53.45%—53.25%—52.34%53.27%
GAIN+SIEF[29]53.82%—54.24%—53.87%53.29%

表5

BoBGSAL?Net模型和其他模型在DocRED数据集上的关系抽取实验结果的对比"

模型验证集测试
Ign F1Ign AUCF1AUCIgn F1F1
GAT[23]45.17%—51.44%—47.36%49.15%
GCNN[8]46.22%—51.52%—49.59%51.62%
EOG[24]45.94%—52.15%—49.48%51.82%
AGGCN[25]46.29%—52.47%—48.89%51.45%
LSR⁃GloVe[22]48.82%—55.17%—52.15%54.18%
GAIN⁃GloVe[26]53.05%52.57%55.29%55.44%52.66%55.08%
HIN⁃BERT⁃base[7]54.29%—55.43%—53.70%55.60%
LSR+BERT⁃base[30]58.93%—60.89%—57.71%59.94%
CGM2IR⁃RoBERTa[31]62.03%—63.95%—61.96%62.89%
BoBGSAL⁃Net54.32%53.47%55.20%54.43%53.62%54.57%
BoBGSAL⁃Net+GloVe56.15%54.39%57.33%57.63%54.35%56.97%
BoBGSAL⁃Net+BiLSTM60.62%58.27%61.45%59.72%58.47%60.54%
BoBGSAL⁃Net+BERT65.20%64.47%64.38%64.58%62.43%65.32%

表6

BoBGSAL?Net模型和其他模型在AlSiaRED数据集上的关系抽取实验结果的对比"

模型验证集测试
Ign F1Ign AUCF1AUCIgn F1F1
BoBGSAL⁃Net+BERT66.14%65.59%65.40%65.32%64.73%66.04%
GAT[23]46.33%—48.20%—45.54%47.39%
GCNN[8]48.46%—50.36%—47.85%49.83%
EOG[24]45.57%—46.91%—45.31%46.32%
AGGCN[25]49.19%—50.95%—48.89%49.63%
LSR⁃GloVe[22]51.35%—53.44%—51.27%53.29%
GAIN⁃GloVe[26]57.88%56.47%59.29%57.89%57.57%59.14%
HIN⁃BERT⁃base[7]53.62%—54.44%—52.56%54.72%
LSR+BERT⁃base[30]59.23%—61.47%—59.62%60.20%
CGM2IR⁃RoBERTa[31]63.53%—62.74%—63.38%63.26%
BoBGSAL⁃Net55.43%54.64%56.51%55.78%54.84%55.73%
BoBGSAL⁃Net+GloVe60.45%56.47%59.29%57.89%57.57%59.14%
BoBGSAL⁃Net+BiLSTM61.58%59.73%62.50%60.48%59.76%61.48%

表7

BoBGSAL?Net模型和其他模型在DocRED数据集上的实体抽取实验结果的对比"

模型验证集测试
Ign F1Ign AUCF1AUCIgn F1F1
BoBGSAL⁃Net+BERT66.14%65.59%65.40%65.32%64.73%66.04%
DocRED⁃CNN[32]40.27%32.75%43.35%34.17%36.44%42.33%
MRN+BERT[33]59.47%—60.20%—59.52%61.74%
DRN⁃GloVe[34]54.61%—56.49%—54.35%56.33%
BoBGSAL⁃Net55.43%54.64%56.51%55.78%54.84%55.73%
BoBGSAL⁃Net+GloVe60.45%56.47%59.29%57.89%57.57%59.14%
BoBGSAL⁃Net+BiLSTM61.58%59.73%62.50%60.48%59.76%61.48%
1 Yuan C S， Huang H Y， Feng C，et al. Document?level relation extraction with entity?selection attention. Information Sciences，2021（568）：163-174.
2 Zhang Q Q， Chen M D， Liu L Z. A review on entity relation extraction ∥ Proceedings of the 2nd Inter?national Conference on Mechanical，Control and Computer Engineering. Harbin，China： IEEE，2017：178-183.
3 Li Z H， Yang Z H， Xiang Y，et al. Exploiting sequence labeling framework to extract document?level relations from biomedical texts. BMC Bioinformatics，2020，21(1)：125.
4 Han X Y， Wang L. A novel document?level relation extraction method based on BERT and entity information. IEEE Access，2020(8)：96912-96919.
5 Geng Z Q， Chen G F， Han Y M，et al. Semantic relation extraction using sequential and tree?structured LSTM with attention. Information Sciences，2020（509）：183-192.
6 Luo L， Yang Z H， Yang P，et al. An attention?based BiLSTM?CRF approach to document?level chemical named entity recognition. Bioinformatics，2018，34(8)：1381-1388.
7 Tang H Z， Cao Y N， Zhang Z Y，et al. HIN：Hierarchical inference network for document?level relation extraction∥Proceedings of the 24th Pacific?Asia Conference on Knowledge Discovery and Data Mining. Springer Berlin Heidelberg，2020：197-209.
8 Najibi M， Rastegari M， Davis L S. G?CNN：An iterative grid based object detector ∥Proceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas，NV，USA：IEEE，2016：2369-2377.
9 Gu J X， Wang Z H， Kuen J，et al. Recent advances in convolutional neural networks. Pattern Recognition，2018（77）：354-377.
10 Li Z W， Liu F， Yang W J，et al. A survey of convolutional neural networks：Analysis，applications，and prospects. IEEE Transactions on Neural Networks and Learning Systems，2022，33(12)：6999-7019.
11 O'Shea K, Nash R. An introduction to convolutional neural networks. 2015，arXiv：1511.08458.
12 Lavin A， Gray S. Fast algorithms for convolutional neural networks∥Proceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas，NV，USA：IEEE，2016：4013-4021.
13 Huang J W， Abadi D J. Leopard：Lightweight edge?oriented partitioning and replication for dynamic graphs. Proceedings of the VLDB Endowment，2016，9(7)：540-551.
14 刘英莉,吴瑞刚,么长慧，等. 铝硅合金实体关系抽取数据集的构建方法. 浙江大学学报(工学版)，2022，56(2)：245-253.
Liu Y L， Wu R G， Yao C H，et al. Construction method of extraction dataset of Al?Si alloy entity relationship. Journal of Zhejiang University (Engineering Science)，2022，56(2)：245-253.
15 Sheng D M， Wang D， Shen Y，et al. Summarize before aggregate：A global?to?local heterogeneous graph inference network for conversational emotion recognition∥Proceedings of the 28th International Conference on Computational Linguistics. Barcelona，Spain： International Committee on Computational Linguistics，2020：4153-4163.
16 Auten A， Tomei M， Kumar R. Hardware acceleration of graph neural networks∥ Proceedings of 2020 57th ACM/IEEE Design Automation Conference (DAC). San Francisco，CA，USA：IEEE，2020：1-6.
17 Abadal S， Jain A， Guirado R，et al. Computing graph neural networks：A survey from algorithms to accelerators. ACM Computing Surveys，2022，54(9)：191.
18 Pennington J， Socher R， Manning C. GloVe：Global vectors for word representation ∥ Proceedings of 2014 Conference on Empirical Methods in Natural Language Processing. Doha，Qatar：ACL，2014：1532-1543.
19 Tanvir R， Shawon T R， Mehedi H K，et al. A GAN?BERT based approach for bengali text classification with a few labeled examples∥ Proceedings of the 19th International Symposium on Distributed Computing and Artificial Intelligence. Springer Berlin Heidelberg，2022：20-30.
20 Niu Z Y， Zhong G Q， Yu H. A review on the attention mechanism of deep learning. Neuro?computing，2021（452）：48-62.
21 Harter A， Hopper A， Steggles P，et al. The anatomy of a context?aware application. Wireless Networks，2002，8(2-3)：187-197.
22 Mrityunjay K, Ravindra G. Learning to fingerprint the latent structure in question articulation∥2018 17th IEEE International Conference on Machine Learning and Applications (ICMLA).. Orlando,FL,USA：IEEE,2018：73-80.
23 Veli?kovi?， Cucurull G， Casanova A，et al. Graph attention networks. 2017，arXiv：1710.10903.
24 Chen L， Tian F L. Skew?rank of an oriented graph with edge?disjoint cycles. Linear and Multilinear Algebra，2016，64(6)：1197-1206.
25 Li Z X， Sun Y R， Zhu J W，et al. Improve relation extraction with dual attention?guided graph convolutional networks. Neural Computing and Applications，2021，33(6)：1773-1784.
26 Zeng S, Xu R, Chang B，et al. Double graph based reasoning for document?level relation extraction. 2020，arXiv：2009.13752.
27 Dai D M， Ren J， Zeng S，et al. Coarse?to?fine entity representations for document?level relation extraction. 2020，arXiv：2012.02507.
28 Xu B F， Wang Q， Lyu Y J，et al. Entity structure within and throughout：Modeling mention dependencies for document?level relation extraction ∥ Proceedings of the 35th AAAI Conference on Artificial Intelligence. Online：AAAI Press，2021：14149-14157.
29 Xu W， Chen K H， Mou L L，et al. Document?level relation extraction with sentences importance estimation and focusing∥Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics：Human Language Technologies. Seattle，WA，United States：ACL，2022：2920-2929.
30 Nan G S， Guo Z J， Sekuli I，et al. Reasoning with latent structure refinement for document?level relation extraction∥Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. Online：ACL，2020：1546-1557，DOI：10.18653/v1/2020.acl-main.141 .
31 Zhao C， Zeng D J， Xu L，et al. Document?level relation extraction with context guided mention integration and inter?pair reasoning.2022， arXiv：2201.04826.
32 Yao Y， Ye D M， Li P，et al. DocRED：A large?scale document?level relation extraction dataset∥Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. Florence，Italy：ACL，2019：764-777，DOI：10.18653/v1/P19-1074.
33 Li J Y， Xu K， Li F，et al. MRN：A locally and globally mention?based reasoning network for document?level relation extraction∥Proceedings of the Findings of the Association for Computational Linguistics. Online：ACL， 2021：1359-1370.
34 Xu W， Chen K H， Zhao T J. Discriminative reasoning for document?level relation extraction ∥Proceedings of the Findings of the Association for Computational Linguistics. ACL， 2021：1653-1663，DOI：10.18653/v1/2021.findings-acl.144 .
[1] 姚启皓, 王伟昊, 尤鸣宇. 基于稳定性优化的三维装配补全方法[J]. 南京大学学报(自然科学版), 2023, 59(5): 742-751.
[2] 严镕宇, 李伟, 陈玉明, 黄宏, 王文杰, 宋宇萍. 一种基于孪生网络的图片匹配算法[J]. 南京大学学报(自然科学版), 2023, 59(5): 770-776.
[3] 王昱翔, 葛洪伟. 基于U2⁃Net的金属表面缺陷检测算法[J]. 南京大学学报(自然科学版), 2023, 59(3): 413-424.
[4] 蔡香香, 王磊, 王冲, 刘斌. 一种结合勾股模糊相似度的多属性决策方法[J]. 南京大学学报(自然科学版), 2023, 59(1): 35-42.
[5] 房笑宇, 曹陈涵, 夏彬. 基于注意力机制的大规模系统日志异常检测方法[J]. 南京大学学报(自然科学版), 2021, 57(5): 785-792.
[6] 普志方, 陈秀宏. 基于卷积神经网络的细胞核图像分割算法[J]. 南京大学学报(自然科学版), 2021, 57(4): 566-574.
[7] 陶砚蕴, 岳国旗, 王凯欣, 张宇祯, 蒋彬, 黄杏梅. 心电图信号双任务学习的时空级联神经网络及心律失常分类模型[J]. 南京大学学报(自然科学版), 2021, 57(2): 318-326.
[8] 张萌, 韩冰, 王哲, 尤富生, 李浩然. 基于深度主动学习的甲状腺癌病理图像分类方法[J]. 南京大学学报(自然科学版), 2021, 57(1): 21-28.
[9] 顾健伟, 曾　诚, 邹恩岑, 陈　扬, 沈　艺, 陆　悠, 奚雪峰. 基于双向注意力流和自注意力结合的机器阅读理解[J]. 南京大学学报(自然科学版), 2019, 55(1): 125-132.
[10] 安　晶, 艾　萍, 徐　森, 刘　聪, 夏建生, 刘大琨. 一种基于一维卷积神经网络的旋转机械智能故障诊断方法[J]. 南京大学学报(自然科学版), 2019, 55(1): 133-142.
Viewed
Full text


Abstract

Cited
  Shared   
  Discussed   
No Suggested Reading articles found!

版权所有 © 2021 《南京大学学报(自然科学)》

地址：江苏省南京市鼓楼区汉口路22号，南京大学学报编辑部，210093

电话：025-83592704 , 传真：025-83592704        技术支持：北京玛格泰克科技发展有限公司