基于变形图匹配的知识图谱多跳问答

李香粤; 方全; 胡骏; 钱胜胜; 徐常胜

doi:10.13700/j.bh.1001-5965.2022.0375

基于变形图匹配的知识图谱多跳问答

doi: 10.13700/j.bh.1001-5965.2022.0375

1.
郑州大学河南先进技术研究院，郑州 450000
2.
中国科学院自动化研究所模式识别国家重点实验室，北京 100190

基金项目: 国家自然科学基金(62072456,62036012,62106262)；之江实验室开放课题(2021KE0AB05)

详细信息

通讯作者:
E-mail：qfang@nlpr.ia.ac.cn

中图分类号: V221.3；TB553
计量
- 文章访问数: 556
- HTML全文浏览量: 25
- PDF下载量: 11
- 被引次数: 0
出版历程
- 收稿日期: 2022-05-18
- 录用日期: 2022-06-23
- 网络出版日期: 2022-08-15
- 整期出版日期: 2024-02-27

Multi-hop knowledge graph question answering based on deformed graph matching

1.
Zhengzhou University，Zhengzhou 450000，China
2.
National Laboratory of Pattern Recognition，Institute of Automation，Chinese Academy of Sciences，Beijing 100190，China

Funds: National Natural Science Foundation of China (62072456,62036012,62106262); Open Research Projects of Zhejiang Lab (2021KE0AB05)

More Information

Corresponding author: E-mail：qfang@nlpr.ia.ac.cn

摘要

摘要:
知识图谱问答(KGQA)是给定自然语言问题，对问题进行语义理解和解析，进而利用知识图谱进行查询、推理得出答案的过程。但知识图谱通常是不完整的，链接缺失给多跳问答带来许多挑战。许多方法在利用知识图谱嵌入时忽略了重要的路径信息来评估路径和多关系问题之间的相关性；且使用文本语料库也会限制文本增强模型的可扩展性。针对这些现有方法的缺陷，提出了基于变形图匹配的知识图谱问答(DGM-KGQA)模型，该模型同时利用问题和主题实体构建语义子图，与知识图谱的局部结构匹配并找到正确答案。在基准数据集MetaQA上的实验结果验证了DGM-KGQA的有效性，该模型在完整知识图谱上检索到的答案准确率分别比PullNet、EmbedKGQA增加了4.2%、0.8%；在完整度仅有一半的知识图谱上检索到的答案准确率分别比PullNet、EmbedKGQA增加了11.1%、0.5%。实验证明提出的变形图匹配模型能够有效地增强知识图谱的关联性及多跳问答的答案准确率。
- 自然语言问题 /
- 链接缺失 /
- 文本语料库 /
- 多跳问答 /
- 变形图匹配
Abstract:
Knowledge Graph Question Answering (KGQA) is a process in which a given natural language question is semantically understood and parsed, and then the knowledge graph is used to query and reason to get the answer. But knowledge graphs which lack links, bring many challenges to multi-hop question answering. Many methods ignore important path information to evaluate the correlation between paths and multi-relationship problems when using knowledge graph embeddings, and text corpora also limit the scalability of text-enhanced models. Due to the drawbacks of these existing approaches, the Multi-hop Knowledge Graph Question Answering Based on Deformed Graph Matching (DGM-KGQA) method is proposed. This method builds semantic subgraphs using both question and topic entities, which then match the local structure of the knowledge graph to determine the correct solution. The experimental results on the benchmark dataset MetaQA verify the effectiveness of DGM-KGQA. In comparison to PullNet and EmbedKGQA, the accuracy of the answers retrieved on the completed knowledge graph is 4.2% higher than that of PullNet and 0.8% higher than that of EmbedKGQA. The accuracy of the answers retrieved on half of the knowledge graphs is 11.1% higher than that of PullNet and 0.5% higher than that of EmbedKGQA. Experiments show that the proposed deformed graph matching model can effectively enhance the relevance of knowledge graphs and the answer accuracy of multi-hop question answering.
- natural language problems /
- lack of links /
- text corpora /
- multi-hop question answering /
- deformed graph matching

HTML全文

图 1 基于变形图匹配的知识图谱问答模型

Figure 1. A knowledge graph Q&A model based on deformed graph matching

下载: 全尺寸图片幻灯片

表 1 MetaQA数据集中不同跳数的问答对数量

Table 1. The number of question answering pairs with different hops in MetaQA dataset

MetaQA	训练集	验证集	测试集
1-hop	96 106	9 992	9 947
2-hop	118 948	14 872	14 872
3-hop	114 196	14 274	14 274

下载: 导出CSV

表 2 MetaQA数据集不同完整度的实验结果

Table 2. Experimental results for MetaQA datasets with different degrees of completeness

模型	总体性能
模型	KG-Full	KG-Half
GraftNet	89.1	57.9
PullNet	96.1	58.2
KV-Mem	74.0	45.8
EmbedKGQA	97.0	81.9
DGM-KGQA	98.2	90.0

下载: 导出CSV

表 3 完整和不完整知识图谱中不同跳数的总体性能

Table 3. Overall performance of different hops in complete and incomplete knowledge graphs

模型	KG-Full			KG-Half
模型	1-hop	2-hop	3-hop	1-hop	2-hop	3-hop
GraftNet	97.0	94.8	77.7	64.0	52.6	59.2
PullNet	97.0	99.9	91.4	65.1	52.1	59.7
KV-Mem	96.2	82.7	48.9	63.6	41.8	37.6
EmbedKGQA	97.5	98.8	94.8	83.9	91.8	70.3
DGM-KGQA	97.2	98.4	95.6	78.3	92.5	70.8

下载: 导出CSV

表 4 本文模型的消融实验结果

Table 4. The ablation results of the model in this paper

Import	总体性能
	KG-Full			KG-Half
	1-hop	2-hop	3-hop	1-hop	2-hop	3-hop
-local matching	86.4	90.3	85.2	59.8	53.5	54.3
-Overall puzzle	93.2	95.7	91.5	72.4	85.8	64.5
DGM-KGQA(final)	97.2	98.4	95.6	78.3	92.5	70.8
注：-local matching, -Overall puzzle中的-为消融实验去掉该模块后的结果。

下载: 导出CSV

参考文献(19)

[1]	WANG R Z, YAN J C, YANG X K. Learning combinatorial embedding networks for deep graph matching [C]//2019 IEEE/CVF International Conference on Computer Vision. Piscatawa: IEEE Press, 2019: 3056-3065.
[2]	KOCH O, KRIEGE N M, HUMBECK L. Chemical similarity and substructure searches [M]//Encyclopedia of Bioinformatics and Computational Biology. Amsterdam: Elsevier, 2019: 640-649.
[3]	SHARAN R, IDEKER T. Modeling cellular machinery through biological network comparison[J]. Nature Biotechnology, 2006, 24: 427-433. doi: 10.1038/nbt1196
[4]	SINGH R, XU J B, BERGER B. Global alignment of multiple protein interaction networks with application to functional orthology detection[J]. Proceedings of the National Academy of Sciences of the United States of America, 2008, 105(35): 12763-12768.
[5]	ZHANG J W, YU P S. Multiple anonymized social networks alignment [C]//2015 IEEE International Conference on Data Mining. Piscataway: IEEE Press, 2015: 599-608.
[6]	VENTO M, FOGGIA P. Graph matching techniques for computer vision[M]//BAI X, AHENG J, HANCOCK. Graph-based methods in computer vision. Hershey: IGI Global, 2012: 1-41.
[7]	STAUFFER M, TSCHACHTLI T, FISCHER A, et al. A survey on applications of bipartite graph edit distance [C]// International Workshop on Graph-Based Representations in Pattern Recognition. Cham: Springer, 2017: 242-252.
[8]	BUNKE H, SHEARER K. A graph distance metric based on the maximal common subgraph[J]. Pattern Recognition Letters, 1998, 19(3-4): 255-259.
[9]	YAN J C, YIN X C, LIN W Y, et al. A short survey of recent advances in graph matching [C]//Proceedings of the 2016 ACM on International Conference on Multimedia Retrieval. New York: ACM, 2016: 167-174.
[10]	XU H T, LUO D X, ZHA H Y, et al. Gromov-wasserstein learning for graph matching and node embedding [EB/OL]. (2016-06-06)[2022-05-18]. http://arxiv.org/abs/1901.06003.
[11]	ROTA BULÒ S, PELILLO M, BOMZE I M. Graph-based quadratic optimization: a fast evolutionary approach[J]. Computer Vision and Image Understanding, 2011, 115(7): 984-995. doi: 10.1016/j.cviu.2010.12.004
[12]	HEIMANN M, SHEN H M, SAFAVI T, et al. REGAL: representation learning-based graph alignment [C] //Proceedings of the 27th ACM International Conference on Information and Knowledge Management. New York: ACM, 2018: 117-126.
[13]	DERR T, KARIMI H, LIU X R, et al. Deep adversarial network alignment [C]//Proceedings of the 30th ACM International Conference on Information & Knowledge Management. New York: ACM, 2021: 352-361.
[14]	FEY M, LENSSEN J E, MORRIS C, et al. Deep graph matching consensus [EB/OL]. (2020-01-27)[2022-05-18]. https://arxiv.org/abs/2001.09621.
[15]	MILLER A, FISCH A, DODGE J, et al. Key-value memory networks for directly reading documents [EB/OL]. (2016-10-10)[2022-05-18].https://arxiv.org/abs/1606.03126.
[16]	SUN H T, DHINGRA B, ZAHEER M, et al. Open domain question answering using early fusion of knowledge bases and text [C]//Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. Stroudsburg: Association for Computational Linguistics, 2018: 4231-4242.
[17]	SUN H T, BEDRAX-WEISS T, COHEN W. PullNet: Open domain question answering with iterative retrieval on knowledge bases and text [C]//Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing. Stroudsburg: Association for Computational Linguistics, 2019: 2380-2390.
[18]	SAXENA A, TRIPATHI A, TALUKDAR P. Improving multi-hop question answering over knowledge graphs using knowledge base embeddings [C]//Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. Stroudsburg: Association for Computational Linguistics, 2020: 4498-4507.
[19]	ZHANG Y Y, DAI H J, KOZAREVA Z, et al. Variational reasoning for question answering with knowledge graph [EB/OL]. (2017-11-27)[2022-05-18].https://arxiv.org/abs/1709.04071.