融合实体和上下文信息的篇章关系抽取研究

黄河燕; 袁长森; 冯冲

doi:10.16383/j.aas.c220966

融合实体和上下文信息的篇章关系抽取研究

doi: 10.16383/j.aas.c220966

黄河燕^{1, 2,},
袁长森^{1, 2,},
冯冲^{1, 2,}

1.
北京理工大学计算机学院北京 100081
2.
北京理工大学自然语言处理实验室北京 100081

详细信息

作者简介:
黄河燕：北京理工大学计算机学院教授. 主要研究方向为语言信息智能化处理, 社交网络, 数据分析和云计算. E-mail: hhy63@bit.edu.cn

袁长森：北京理工大学计算机学院博士后. 主要研究方向为知识图谱, 信息抽取. 本文通信作者. E-mail: yuanchangsen@bit.edu.cn

冯冲：北京理工大学计算机学院教授. 主要研究方向为机器翻译, 信息抽取和信息检索. E-mail: fengchong@bit.edu.cn

计量
- 文章访问数: 570
- HTML全文浏览量: 199
- PDF下载量: 76
- 被引次数: 0
出版历程
- 收稿日期: 2022-12-12
- 录用日期: 2023-03-29
- 网络出版日期: 2023-08-28
- 刊出日期: 2024-10-21

Document-level Relation Extraction With Entity and Context Information

HUANG He-Yan^{1, 2
,},
YUAN Chang-Sen^{1, 2
,},
FENG Chong^{1, 2
,}

1.
School of Computer Science and Technology, Beijing Institute of Technology, Beijing 100081
2.
Natural Language Processing Laboratory, Beijing Institute of Technology, Beijing 100081

More Information

Author Bio:
HUANG He-Yan　Professor at the School of Computer Science and Technology, Beijing Institute of Technology. Her research interest covers intelligent processing of language information, social network, data analysis, and cloud computing

YUAN Chang-Sen　Postdoctor at the School of Computer Science and Technology, Beijing Institute of Technology. His research interest covers knowledge graph and information extraction. Corresponding author of this paper

FENG Chong　Professor at the School of Computer Science and Technology, Beijing Institute of Technology. His research interest covers machine translation, information extraction, and information retrieval

摘要

摘要: 篇章关系抽取旨在识别篇章中实体对之间的关系. 相较于传统的句子级别关系抽取, 篇章级别关系抽取任务更加贴近实际应用, 但是它对实体对的跨句子推理和上下文信息感知等问题提出了新的挑战. 本文提出融合实体和上下文信息(Fuse entity and context information, FECI)的篇章关系抽取方法, 它包含两个模块, 分别是实体信息抽取模块和上下文信息抽取模块. 实体信息抽取模块从两个实体中自动地抽取出能够表示实体对关系的特征. 上下文信息抽取模块根据实体对的提及位置信息, 从篇章中抽取不同的上下文关系特征. 本文在三个篇章级别的关系抽取数据集上进行实验, 效果得到显著提升.
- 篇章关系抽取 /
- 实体信息 /
- 上下文信息 /
- 提及位置信息 /
- 跨句子推理
Abstract: Document-level relation extraction aims to identify the relations among entities from the document. Compared with traditional sentence-level relation extraction, document-level relation extraction is more realistic and poses new challenges of cross-sentence inference and context information understanding. In this paper, we propose a novel method for document-level relation extraction by fusing entity and context information (FECI), which contains two modules: Entity information extraction module and context information extraction module. Entity information extraction module automatically extracts crucial relation features about entity pair. Context information extraction module extracts different context relation features from the document according to mentions＇ position information of entity pair. We have conducted experiments on three document-level relation extraction datasets, and the effect has been significantly improved.
- Document-level relation extraction /
- entity information /
- context information /
- mentions＇ position information /
- cross-sentence inference

HTML全文

图 1 篇章级别关系抽取数据集DocRED中的一个实例

Fig. 1 An example of document-level relation extraction dataset DocRED

下载: 全尺寸图片幻灯片

图 2 模型框架图主要有两个部分, 分别是实体信息抽取模块和上下文信息抽取模块

Fig. 2 Architecture of the proposed model, which contains two parts: Entity information extraction module and context information extraction module

下载: 全尺寸图片幻灯片

图 3 篇章级别关系抽取开发集中的一个实例分析

Fig. 3 An example analysis on the document-level relation extraction development set

下载: 全尺寸图片幻灯片

表 1 数据集的统计

Table 1 Statistics of the datasets

统计	DocRED	CDR	GDA
训练集	3053	500	23353
开发集	1000	500	5839
测试集	1000	500	1000
关系种类	97	2	2
每篇的关系数量	19.5	7.6	5.4

下载: 导出CSV

表 2 模型的超参数

Table 2 Hyper-parameters of model

参数名称	DocRED	CDR	GDA
批次大小	4	4	4
迭代次数	30	30	10
学习率 (编码)	$5\times 10^{-5}$	$5\times 10^{-5}$	$5\times 10^{-5}$
学习率 (分类)	$1\times 10^{-4}$	$1\times 10^{-4}$	$1\times 10^{-4}$
分组大小	64	64	64
Dropout	0.1	0.1	0.1
梯度裁剪	1.0	1.0	1.0

下载: 导出CSV

表 3 在DocRED开发集和测试集上的实验结果(%)

Table 3 Experiment results on the development and test sets of DocRED (%)

模型	开发集		测试集
模型	Ign F1	F1	Ign F1	F1
CNN	41.58	43.45	40.33	42.26
LSTM	48.44	50.68	47.71	50.07
Bi-LSTM	48.87	50.94	48.78	51.06
Context-Aware	48.94	51.09	48.40	50.70
HIN-GloVe	51.06	52.95	51.15	53.30
GAT-GloVe	45.17	51.44	47.36	49.51
GCNN-GloVe	46.22	51.52	49.59	51.62
EoG-GloVe	45.94	52.15	49.48	51.82
AGGCN-GloVe	46.29	52.47	48.89	51.45
LSR-GloVe	48.82	55.17	52.15	54.18
BERT-RE_BASE	—	54.16	—	53.20
RoBERTa_BASE	53.85	56.05	53.52	55.77
BERT-Two-Step_BASE	—	54.42	—	53.92
HIN-BERT_BASE	54.29	56.31	53.70	55.60
CorefBERT_BASE	55.32	57.51	54.54	56.96
LSR-BERT_BASE	52.43	59.00	56.97	59.05
BERT-E_BASE	56.51	58.52	—	—
GAIN_BASE	59.14	61.22	59.00	61.24
FECI_BASE	59.74	61.38	59.81	61.22

下载: 导出CSV

表 4 在CDR和GDA数据集上测试集F1值(%)

Table 4 F1 values of test set on CDR and GDA datasets (%)

模型	CDR	GDA
BRAN	62.1	—
CNN	62.3	—
EoG	63.6	81.5
LSR-BERT	64.8	82.2
SciBERT_BASE	65.1	82.5
SciBERT-E_BASE	65.9	83.3
FECI_BASE	69.2	83.7

下载: 导出CSV

表 5 FECI_BASE在开发集上的消融研究结果

Table 5 Ablation study results of FECI_BASE on the development set

模型	开发集
模型	Ign F1 (%)	F1 (%)	P (M)	T (s)
FECI_BASE	59.74	61.38	133.4	2962.4
w/o Entity	58.16	60.07	132.2	2831.7
w/o Context	58.67	60.89	130.5	482.3

下载: 导出CSV

表 6 FECI_BASE在开发集上噪声实体和噪声上下文的实验结果(%)

Table 6 The experiment results of noisy entity and noisy context of FECI_BASE on the development set (%)

模型	开发集
模型	Ign F1	F1
FECI_BASE	59.74	61.38
Head entity	58.42	60.14
Tail entity	57.97	60.08
Entity pair	58.91	60.85
Tradition	57.42	59.72
Co-occurrence	58.27	61.01
Non co-occurrence	56.72	58.86

下载: 导出CSV

表 7 FECI_BASE在开发集上不同上下文信息的实验结果(%)

Table 7 The experiment results of different context information of FECI_BASE on the development set (%)

模型	开发集
模型	Ign F1	F1
FECI_BASE	59.74	61.38
Random	58.47	60.61
Mean	59.56	60.94
Tradition	58.19	60.06

下载: 导出CSV

表 8 不同方法在开发集上的效率

Table 8 Efficiency of different methods on the development set

模型	开发集
模型	P (M)	Train T (s)	Decoder T (s)
LSR-BERT_BASE	112.1	282.9	38.8
GAIN_BASE	217.0	2271.6	817.2
FECI_BASE	133.4	2962.4	829.0

下载: 导出CSV

参考文献(39)

[1]	Yu M, Yin W P, Hasan K S, Santos C D, Xiang B, Zhou B W. Improved neural relation detection for knowledge base question answering. arXiv preprint arXiv: 1704.06194, 2017.
[2]	Chen Z Y, Chang C H, Chen Y P, Nayak J, Ku L W. UHop: An unrestricted-hop relation extraction framework for knowledge-based question answering. arXiv preprint arXiv: 1904.01246, 2019.
[3]	Yu H Z, Li H S, Mao D H, Cai Q. A relationship extraction method for domain knowledge graph construction. World Wide Web, 2020, 23(2): 735−753 doi: 10.1007/s11280-019-00765-y
[4]	Ristoski P, Gentile A L, Alba A, Gruhl D, Welch S. Large-scale relation extraction from web documents and knowledge graphs with human-in-the-loop. Journal of Web Semantics, 2020, 60: Article No. 100546 doi: 10.1016/j.websem.2019.100546
[5]	Macdonald E, Barbosa D. Neural relation extraction on Wikipedia tables for augmenting knowledge graphs. In: Proceedings of the 29th ACM International Conference on Information & Knowledge Management. New York, USA: ACM, 2020. 2133–2136
[6]	Mintz M, Bills S, Snow R, Jurafsky D. Distant supervision for relation extraction without labeled data. In: Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP. Suntec, Singapore: ACL, 2009. 1003–1011
[7]	Lin Y K, Shen S Q, Liu Z Y, Luan H B, Sun M S. Neural relation extraction with selective attention over instances. In: Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics. Berlin, Germany: ACL, 2016. 2124–2133
[8]	Miwa M, Bansal M. End-to-end relation extraction using LSTMs on sequences and tree structures. arXiv preprint arXiv: 1601.00770, 2016.
[9]	Zhang Y H, Qi P, Manning C D. Graph convolution over pruned dependency trees improves relation extraction. arXiv preprint arXiv: 1809.10185, 2018.
[10]	Guo Z J, Zhang Y, Lu W. Attention guided graph convolutional networks for relation extraction. arXiv preprint arXiv: 1906.07510, 2019.
[11]	Yao Y, Ye D M, Li P, Han X, Lin Y K, Liu Z H, et al. DocRED: A large-scale document-level relation extraction dataset. arXiv preprint arXiv: 1906.06127, 2019.
[12]	Zhou W X, Huang K, Ma T Y, Huang J. Document-level relation extraction with adaptive thresholding and localized context pooling. arXiv preprint arXiv: 2010.11304, 2020.
[13]	Zeng S, Xu R X, Chang B B, Li L. Double graph based reasoning for document-level relation extraction. arXiv preprint arXiv: 2009.13752, 2020.
[14]	Santos C N D, Xiang B, Zhou B W. Classifying relations by ranking with convolutional neural networks. arXiv preprint arXiv: 1504.06580, 2015.
[15]	Cho K, Merrienboer B V, Gulcehre C, Bahdanau D, Bougares F, Schwenk H, et al. Learning phrase representations using RNN encoder-decoder for statistical machine translation. arXiv preprint arXiv: 1406.1078, 2014.
[16]	Liu Y, Wei F R, Li S J, Ji H, Zhou M, Wang H F. A dependency-based neural network for relation classification. In: Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Short Papers). Beijing, China: ACL, 2015. 285–290
[17]	Christopoulou F, Miwa M, Ananiadou S. A walk-based model on entity graphs for relation extraction. arXiv preprint arXiv: 1902.07023, 2019.
[18]	Christopoulou F, Miwa M, Ananiadou S. Connecting the dots: Document-level neural relation extraction with edge-oriented graphs. arXiv preprint arXiv: 1909.00228, 2019.
[19]	Yang B S, Mitchell T. Joint extraction of events and entities within a document context. arXiv preprint arXiv: 1609.03632, 2016.
[20]	Swampillai K, Stevenson M. Extracting relations within and across sentences. In: Proceedings of the Recent Advances in Natural Language Processing. Hissar, Bulgaria: DBLP, 2011. 25–32
[21]	Jia R, Wong C, Poon H. Document-level n-ary relation extraction with multiscale representation learning. arXiv preprint arXiv: 1904.02347, 2019.
[22]	Verga P, Strubell E, McCallum A. Simultaneously self-attending to all mentions for full-abstract biological relation extraction. arXiv preprint arXiv: 1802.10569, 2018.
[23]	Nan G S, Guo Z J, Sekulic I, Lu W. Reasoning with latent structure refinement for document-level relation extraction. arXiv preprint arXiv: 2005.06312, 2020.
[24]	Wang D F, Hu W, Cao E, Sun W J. Global-to-local neural networks for document-level relation extraction. arXiv preprint arXiv: 2009.10359, 2020.
[25]	Devlin J, Chang M W, Lee K, Toutanova K. BERT: Pre-training of deep bidirectional Transformers for language understanding. arXiv preprint arXiv: 1810.04805, 2019.
[26]	Vaswani A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez A N, et al. Attention is all you need. In: Proceedings of the 31st International Conference on Neural Information Processing Systems. Long Beach, USA: Curran Associates Inc., 2017. 6000–6010
[27]	Sennrich R, Haddow B, Birch A. Neural machine translation of rare words with subword units. arXiv preprint arXiv: 1508.07909, 2016.
[28]	Li J, Sun Y P, Johnson R J, Sciaky D, Wei C, Leaman R, et al. BioCreative V CDR task corpus: A resource for chemical disease relation extraction. The Journal of Biological Databases and Curation, 2016: Article No. baw068
[29]	Wu Y, Luo R B, Leung H C M, Ting H, Lam T. RENET: A deep learning approach for extracting gene-disease associations from literature. In: Proceedings of the International Conference on Research in Computational Molecular Biology. Washington, USA: Springer, 2019. 272–284
[30]	Liu Y H, Ott M, Goyal N, Du J F, Joshi M, Chen D Q, et al. RoBERTa: A robustly optimized BERT pretraining approach. arXiv preprint arXiv: 1907.11692, 2019.
[31]	Beltagy I, Lo K, Cohan A. SciBERT: A pretrained language model for scientific text. arXiv preprint arXiv: 1903.10676, 2019.
[32]	Micikevicius P, Narang S, Alben J, Diamos G, Elsen E, Garca D, et al. Mixed precision training. arXiv preprint arXiv: 1710.03740, 2018.
[33]	Loshchilov I, Hutter F. Decoupled weight decay regularization. arXiv preprint arXiv: 1711.05101, 2019.
[34]	Velickovic P, Cucurull G, Casanova A, Romero A, Liò P, Bengio Y. Graph attention networks. arXiv preprint arXiv: 1710.10903, 2018.
[35]	Wang H, Focke C, Sylvester R, Mishra N, Wang W. Fine-tune BERT for DocRED with two-step process. arXiv preprint arXiv: 1909.11898, 2019.
[36]	Tang H Z, Cao Y N, Zhang Z Y, Cao J X, Fang F, Wang S, et al. HIN: Hierarchical inference network for document-level relation extraction. arXiv preprint arXiv: 2003.12754, 2020.
[37]	Pennington J, Socher R, Manning C D. GloVe: Global vectors for word representation. In: Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP). Doha, Qatar: Association for Computational Linguistics, 2014. 1532–1543
[38]	Ye D M, Lin Y K, Du J J, Liu Z H, Sun M S, Liu Z Y. Coreferential reasoning learning for language representation. arXiv preprint arXiv: 2004.06870, 2020.
[39]	Nguyen D Q, Verspoor K. Convolutional neural networks for chemical-disease relation extraction are improved with character-based word embeddings. arXiv preprint arXiv: 1805.10586, 2018.