WO2022011681A1 - 一种基于迭代补全的知识图谱融合方法 - Google Patents

一种基于迭代补全的知识图谱融合方法 Download PDF

Info

Publication number
WO2022011681A1
WO2022011681A1 PCT/CN2020/102683 CN2020102683W WO2022011681A1 WO 2022011681 A1 WO2022011681 A1 WO 2022011681A1 CN 2020102683 W CN2020102683 W CN 2020102683W WO 2022011681 A1 WO2022011681 A1 WO 2022011681A1
Authority
WO
WIPO (PCT)
Prior art keywords
entity
vector
knowledge graph
similarity
entities
Prior art date
Application number
PCT/CN2020/102683
Other languages
English (en)
French (fr)
Inventor
赵翔
曾维新
唐九阳
李欣奕
谭真
郭得科
罗来龙
Original Assignee
国防科技大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 国防科技大学 filed Critical 国防科技大学
Priority to PCT/CN2020/102683 priority Critical patent/WO2022011681A1/zh
Publication of WO2022011681A1 publication Critical patent/WO2022011681A1/zh
Priority to US18/097,292 priority patent/US20230206127A1/en

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00Machine learning
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/28Databases characterised by their database models, e.g. relational or object models
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/36Creation of semantic tools, e.g. ontology or thesauri
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/042Knowledge-based neural networks; Logical representations of neural networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/044Recurrent networks, e.g. Hopfield networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/045Combinations of networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • G06N3/084Backpropagation, e.g. using gradient descent
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00Computing arrangements using knowledge-based models
    • G06N5/02Knowledge representation; Symbolic representation
    • G06N5/022Knowledge engineering; Knowledge acquisition

Definitions

  • the invention belongs to the technical field of natural language processing, relates to knowledge graph generation and fusion, and in particular relates to a knowledge graph fusion method based on iterative completion.
  • a feasible method is to introduce relevant knowledge from other knowledge graphs, because knowledge graphs constructed in different ways have knowledge redundancy and complementarity.
  • a general knowledge graph constructed from web pages may only contain the names of scientists, while more information can be found in an academic knowledge graph constructed based on academic data.
  • the most important step is to align the different knowledge graphs.
  • the entity alignment (EA) task was proposed and received extensive attention. This task aims to find entity pairs expressing the same meaning in different knowledge graphs. These entity pairs serve as a hub to link different knowledge graphs and serve subsequent tasks.
  • mainstream entity alignment methods mainly rely on the structural features of knowledge graphs to determine whether two entities point to the same thing. Such methods assume that entities expressing the same meaning in different knowledge graphs have similar adjacency information.
  • the entity generation structure vector proposed by Lingbing Guo et al., and then the recognition of entity pairs has achieved certain results (Reference: Lingbing Guo, Zequn Sun, and Wei Hu. 2019. Learning to Exploit Long-term Relational Dependencies in Knowledge Graphs. In Proceedings of the 36th International Conference on Machine Learning,ICML 2019,9-15 June 2019,Long Beach,California,USA.2505–2514), on artificially constructed datasets, this kind of method achieved the best experimental results .
  • the knowledge graphs in these artificially constructed datasets are denser than the real-world knowledge graphs, and the entity alignment method based on structural features is much less effective on the knowledge graphs with normal distribution.
  • the purpose of the present invention is to propose a knowledge graph fusion method based on iterative completion, which overcomes the deficiencies of the prior art and is used for identifying and aligning the same or similar entities from multiple knowledge graphs, and then Realize the knowledge fusion of multiple knowledge graphs, and improve the coverage and accuracy of knowledge graphs.
  • a knowledge graph fusion method based on iterative completion includes the following steps:
  • Step 1 obtain multiple knowledge graph data, and identify all entities in the knowledge graph
  • Step 2 perform structure vector representation learning on all entities to obtain the structure vector of each entity; perform entity name vector representation learning on all entities to obtain the entity name vector of each entity;
  • Step 3 calculates the structural similarity between entities according to the described structure vector, calculates the entity name similarity between the entities according to the described entity name vector;
  • Step 4 establish a mutual attention network based on degree perception, and calculate the entity similarity between entities after fusion;
  • Step 5 Select high-confidence entity pairs according to the entity similarity, and use iterative training to complete the knowledge map to obtain a fused knowledge map.
  • the calculation process of the mutual attention network includes the following steps:
  • Step 401 construct feature matrix: construct a feature matrix for each entity, which is represented by the entity name vector of the entity. structure vector and the entity degree vector composition, the entity degree vector is in, is the one-hot vector of the entity degree, is the fully connected parameter matrix, d g is the dimension of the degree vector, and for the entity e 1 , its feature matrix is further expressed as:
  • d m max ⁇ d n ,d s ,d g ⁇ , d n represents the dimension of the entity name vector, and d s represents the dimension of the structure vector;
  • Step 402 Calculate the mutual attention similarity matrix: a feature matrix that dynamically characterizes the entity e 1 and the feature matrix of entity e 2 The correlation between, constructs a mutual attention similarity matrix Wherein the degree of similarity between an i-th feature e e 1 and j-th 2 wherein:
  • Step 403 assign weights, calculate entity similarity: use mutual attention similarity matrix Generate attention vector and first It is sent to the softmax layer, and then sent to the averaging layer to generate the attention vector, where Wherein the degree of characterization related to e 1 e 2 features, and Represents the degree of correlation between the e 2 feature and the e 1 feature. Finally, by multiplying the similarity values of different features with their weights, the fused entity similarity value is obtained:
  • the first and second values of respectively represent the weights corresponding to the structural similarity Sim s (e 1 , e 2 ) and the entity name similarity Sim t (e 1 , e 2 ).
  • the entity name vector is a power-averaged word vector.
  • the word vectors of all words that constitute the entity name are expressed in matrix form as where l represents the number of words, d represents the dimension of the embedding, and Perform the power average operation to generate the power average word vector,
  • the power average operation formula is:
  • the entity name vector adopts the splicing K-th power average word vector.
  • the word vectors of all the words that constitute the entity name are expressed in matrix form as: Among them, l represents the number of words, and d represents the dimension of the embedding.
  • the K-th power average word vector is calculated for the word vector of the entity name, and then the K-th power average word vector is spliced to generate the entity name vector. which is:
  • the specific values of the K different power averages are respectively 1, negative infinity and positive infinity.
  • the structural similarity Sim s (e 1 , e 2 ) is the structural vector of the two entities and The cosine similarity of
  • the entity name similarity Sim t (e 1 , e 2 ) is the entity name vector of the two entities and The cosine similarity of .
  • the step of selecting a high-confidence entity pair is: for each entity e 1 in the original knowledge graph, it is assumed that the most similar entity in the external knowledge graph is e 2 , and the second similar entity is e' 2 , the similarity difference is For e 2 in the external knowledge graph, the most similar entity in the original knowledge graph is exactly e 1 , the second similar entity is e′ 1 , and the similarity difference is If the similarity difference values ⁇ 1 , ⁇ 2 are both higher than a predetermined value, then (e 1 , e 2 ) is considered to be a high-confidence entity pair;
  • the iterative training process for the completion of the knowledge graph is multi-round. For each triple in the external knowledge graph, if the head entity and the tail entity are both in the original knowledge graph, the entity in the external knowledge graph will be replaced. Then use the added knowledge graph to re-learn the structure vector, calculate the entity similarity, and generate a new high-confidence entity pair, continue The knowledge graph is added and completed until the stop condition is met, and the iterative training is stopped.
  • the present invention has the following advantages and beneficial effects:
  • a degree-aware mutual attention network is proposed to fuse entity name information and structure information, so that the alignment effect is better;
  • Fig. 1 is the overall flow schematic diagram of the embodiment of the present invention.
  • FIG. 2 is a structural diagram of a mutual attention network according to an embodiment of the present invention.
  • FIG. 3 is an overall flow frame diagram of an embodiment of the present invention.
  • a knowledge graph fusion method based on iterative completion includes the following steps:
  • Step 1 obtain multiple knowledge graph data, and identify all entities in the knowledge graph
  • Step 2 perform structure vector representation learning on all entities to obtain the structure vector of each entity; perform entity name vector representation learning on all entities to obtain the entity name vector of each entity;
  • Step 3 calculates the structural similarity between entities according to the described structure vector, calculates the entity name similarity between the entities according to the described entity name vector;
  • Step 4 establish a mutual attention network based on degree perception, and calculate the entity similarity between entities after fusion;
  • Step 5 Select high-confidence entity pairs according to the entity similarity, and use iterative training to complete the knowledge map to obtain a fused knowledge map.
  • the learning of the structure vector can use the existing method in the background technology to generate the structure vector, and the structure matrix is expressed as: where n represents the number of entities and d s represents the dimension of the structure vector.
  • the entity name vector can be a power-averaged word vector.
  • the word vector of all the words constituting the entity name is expressed in matrix form as where l represents the number of words, d represents the dimension of the embedding, and Perform the power average operation to generate the power average word vector,
  • the power average operation formula is:
  • the entity name vector can be spliced with the K-th power average word vector. After the average word vector is spliced, the entity name vector is generated which is:
  • K represents the specific value of K different power averages.
  • the specific numerical values of the K different power averages are respectively 1, negative infinity and positive infinity.
  • the concatenated power average word vector can capture more information of the entity name and reduce the uncertainty of the vector representation.
  • the input of the mutual attention network is the structural similarity Sim s (e 1 , e 2 ) between the two entities, the entity name similarity Sim t (e 1 , e 2 ) and the degree of the entity.
  • the calculation process includes the following step:
  • Step 401 construct feature matrix: construct a feature matrix for each entity, which is represented by the entity name vector of the entity. structure vector and the entity degree vector composition, the entity degree vector is in, is the one-hot vector of the entity degree, is the fully connected parameter matrix, d g is the dimension of the degree vector, and for the entity e 1 , its feature matrix is further expressed as:
  • Step 402 Calculate the mutual attention similarity matrix: a feature matrix that dynamically characterizes the entity e 1 and the feature matrix of entity e 2 The correlation between, constructs a mutual attention similarity matrix Wherein the degree of similarity between an i-th feature e e 1 and j-th 2 wherein:
  • Step 403 assign weights, calculate entity similarity: use mutual attention similarity matrix Generate attention vector and first It is sent to the softmax layer, and then sent to the averaging layer to generate the attention vector, where Wherein the degree of characterization related to e 1 e 2 features, and Represents the degree of correlation between the e 2 feature and the e 1 feature. Finally, by multiplying the similarity values of different features with their weights, the fused entity similarity value is obtained:
  • the first and second values of respectively represent the weights corresponding to the structural similarity Sim s (e 1 , e 2 ) and the entity name similarity Sim t (e 1 , e 2 ).
  • Long-tail entities may have little structural information in the original knowledge graph, but richer structural information in the external knowledge graph. If the structural information in the external knowledge graph can be introduced to supplement the structural information of the long-tail entities in the original knowledge graph, the long-tail problem can be alleviated to a certain extent and the coverage of the knowledge graph can be improved.
  • the augmented knowledge graph can generate more accurate structure vectors and improve the effect of entity alignment.
  • the step of selecting a high-confidence entity pair is: for each entity e 1 in the original knowledge graph, it is assumed that the most similar entity in the external knowledge graph is e 2 , the second similar entity is e' 2 , and the similarity
  • the degree difference is For e 2 in the external knowledge graph, the most similar entity in the original knowledge graph is exactly e 1 , the second similar entity is e′ 1 , and the similarity difference is If the similarity difference values ⁇ 1 , ⁇ 2 are both higher than a predetermined value, then (e 1 , e 2 ) is considered to be a high-confidence entity pair;
  • the iterative training process for the completion of the knowledge graph is multi-round. For each triple in the external knowledge graph, if the head entity and the tail entity are both in the original knowledge graph, the entity in the external knowledge graph will be replaced. Then use the added knowledge graph to re-learn the structure vector, calculate the entity similarity, and generate a new high-confidence entity pair, continue The knowledge graph is added and completed until the stop condition is met, and the iterative training is stopped.
  • the method of the present invention proposes a new entity alignment framework, as shown in Figure 3, so as to better realize the fusion of knowledge graphs.
  • the main technical effects of the present invention are as follows:
  • the method of the present invention takes the entity name as a new alignment information.
  • the method of the present invention uses the entity name as a separate feature, and represents the entity name by splicing the power average word vector, which can capture more of the entity name. information, and reduce the uncertainty of vector representation; in the alignment stage, it is observed that for entities with different degrees, the importance of structural information and entity name information is also different, a mutual attention network is designed to determine under the guidance of degrees weights of different features, and effectively integrate multi-source information; in the post-alignment processing stage, an iterative training algorithm based on knowledge graph completion is proposed. While supplementing the knowledge graph structure information, it iteratively improves the entity alignment effect, thereby making Long tailed entities are easier to align.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Software Systems (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Databases & Information Systems (AREA)
  • Medical Informatics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

一种基于迭代补全的知识图谱融合方法,包括以下步骤:获取多个知识图谱数据,识别知识图谱中的所有实体;获得每一个实体的结构向量和实体名向量;计算实体间的结构相似度和实体名相似度;建立基于度感知的互注意力网络,计算融合后实体间的实体相似度;根据所述的实体相似度选择高置信度实体对,并采用迭代训练进行知识图谱补全,获得融合后的知识图谱。本发明方法提出了度感知的互注意力网络以融合实体名向量和结构向量,使得对齐效果更好;提出了使用拼接幂平均词向量表征实体名,能够捕捉实体名的更多信息;提出了一种基于知识图谱补全的迭代训练算法,迭代式地提升实体对齐效果,使得长尾实体更容易对齐。

Description

一种基于迭代补全的知识图谱融合方法 技术领域
本发明属于自然语言处理技术领域,涉及知识图谱生成与融合,具体涉及一种基于迭代补全的知识图谱融合方法。
背景技术
近年来,涌现出一大批知识图谱(knowledge graph,KG),诸如YAGO,DBpedia,Knowledge Vault等。这些大规模知识图谱在问答系统、个性化推荐等智能服务中起到重要作用。此外,为满足特定领域相关需求,衍生出越来越多的领域知识图谱,如学术知识图谱。但任何一个知识图谱,都无法达到完备或者完全正确。
为提升知识图谱的覆盖率及正确率,一种可行方法是从其它知识图谱中引入相关知识,因为以不同方式构建得到的知识图谱间存在知识的冗余以及互补。例如,从网页上抽取构建的通用知识图谱中可能仅包含科学家的名字,而更多的信息可在基于学术数据构建的学术知识图谱中找到。为将外部知识图谱中的知识整合到目标知识图谱中,最重要的一步是对齐不同的知识图谱。为此,实体对齐(entity alignment,EA)任务被提出并受到广泛关注。该任务旨在找到不同知识图谱中表达同一含义的实体对。而这些实体对则作为链接不同知识图谱的枢纽,服务于后续任务。
目前,主流实体对齐方法主要借助知识图谱结构特征判断两实体是否指向同一事物。这类方法假设不同知识图谱中表达同一含义的实体具有类似的邻接信息。Lingbing Guo等人提出的实体生成结构向量,进而实现实体对的识别取 得了一定的效果(参考文献:Lingbing Guo,Zequn Sun,and Wei Hu.2019.Learning to Exploit Long-term Relational Dependencies in Knowledge Graphs.In Proceedings of the 36th International Conference on Machine Learning,ICML 2019,9-15 June 2019,Long Beach,California,USA.2505–2514),在人工构建的数据集上,这类方法取得了最好的实验结果。但最近一项工作指出,这些人工构建的数据集中的知识图谱比真实世界的知识图谱更加稠密,而基于结构特征的实体对齐方法在具有正常分布的知识图谱上效果大打折扣。
事实上,通过分析真实世界知识图谱中的实体分布可知,超过半数的实体只与一两个其它实体相连。这些实体被称为长尾实体(long-tail entities),占据了知识图谱实体的大部分,使得图谱整体呈现较高的稀疏性。这也符合对真实世界知识图谱的认知:只有很少一部分实体被经常使用并具有丰富的邻接信息;绝大部分实体很少被提及,包含微少的结构信息。因此,当前基于结构信息的实体对齐方法和知识图谱融合方法在真实世界数据集上的表现不尽人意。
发明内容
有鉴于此,本发明的目的在于提出一种基于迭代补全的知识图谱融合方法,所述方法克服现有技术的不足,用于从多个知识图谱中进行相同或者相似实体识别和对齐,进而实现多个知识图谱的知识融合,提高知识图谱的覆盖率及正确率。
基于上述目的,一种基于迭代补全的知识图谱融合方法,包括以下步骤:
步骤1,获取多个知识图谱数据,识别知识图谱中的所有实体;
步骤2,对所有实体进行结构向量表示学习,获得每一个实体的结构向量;对所有实体进行实体名向量表示学习,获得每一个实体的实体名向量;
步骤3,根据所述的结构向量计算实体间的结构相似度,根据所述的实体名向量计算实体间的实体名相似度;
步骤4,建立基于度感知的互注意力网络,计算融合后实体间的实体相似度;
步骤5,根据所述的实体相似度选择高置信度实体对,并采用迭代训练进行知识图谱补全,获得融合后的知识图谱。
所述的互注意力网络的计算过程包括以下步骤:
步骤401,构建特征矩阵:为每个实体构建一个特征矩阵,由所述实体的实体名向量
Figure PCTCN2020102683-appb-000001
结构向量
Figure PCTCN2020102683-appb-000002
以及实体度向量
Figure PCTCN2020102683-appb-000003
组成,实体度向量为
Figure PCTCN2020102683-appb-000004
其中,
Figure PCTCN2020102683-appb-000005
是所述实体度数的one-hot向量,
Figure PCTCN2020102683-appb-000006
是全连接参数矩阵,d g是度向量的维度,对于实体e 1,其特征矩阵进一步表示为:
Figure PCTCN2020102683-appb-000007
其中;代表沿着列的拼接,d m=max{d n,d s,d g},d n表示实体名向量的维度,d s表示结构向量的维度;
步骤402,计算互注意力相似度矩阵:为动态刻画实体e 1的特征矩阵
Figure PCTCN2020102683-appb-000008
和实体e 2的特征矩阵
Figure PCTCN2020102683-appb-000009
之间的关联,构建一个互注意力相似度矩阵
Figure PCTCN2020102683-appb-000010
其中e 1的第i个特征和e 2的第j个特征之间的相似度为:
Figure PCTCN2020102683-appb-000011
其中,
Figure PCTCN2020102683-appb-000012
是特征矩阵
Figure PCTCN2020102683-appb-000013
第i个行向量,
Figure PCTCN2020102683-appb-000014
是特征矩阵
Figure PCTCN2020102683-appb-000015
第j个列向量,i=1,2,3;j=1,2,3,
Figure PCTCN2020102683-appb-000016
是一个用于生成相似度的可训练标量函数,
Figure PCTCN2020102683-appb-000017
是参数向量,
Figure PCTCN2020102683-appb-000018
代表沿着行的拼接操作,ο代表点乘;
步骤403,分配权重,计算实体相似度:利用互注意力相似度矩阵
Figure PCTCN2020102683-appb-000019
生成注意力向量
Figure PCTCN2020102683-appb-000020
Figure PCTCN2020102683-appb-000021
先将
Figure PCTCN2020102683-appb-000022
送入softmax层,再送入平均化层,进而生成注 意力向量,其中
Figure PCTCN2020102683-appb-000023
表征e 1特征与e 2特征的相关程度,而
Figure PCTCN2020102683-appb-000024
代表e 2特征与e 1特征的相关程度,最后,通过将不同特征的相似度值与其权重相乘,得到融合后的实体相似度值:
Figure PCTCN2020102683-appb-000025
其中,
Figure PCTCN2020102683-appb-000026
Figure PCTCN2020102683-appb-000027
为注意力向量
Figure PCTCN2020102683-appb-000028
的第1个和第2个值,分别代表结构相似度Sim s(e 1,e 2)和实体名相似度Sim t(e 1,e 2)所对应的权重。
具体地,所述的实体名向量为幂平均词向量,对于某一实体的实体名s,构成该实体名的所有词的词向量用矩阵形式表示为
Figure PCTCN2020102683-appb-000029
其中l代表词的数目,d代表嵌入的维度,对
Figure PCTCN2020102683-appb-000030
进行幂平均操作,即可生成的幂平均词向量,
Figure PCTCN2020102683-appb-000031
幂平均操作公式为:
Figure PCTCN2020102683-appb-000032
Figure PCTCN2020102683-appb-000033
表示将
Figure PCTCN2020102683-appb-000034
处理后生成的幂平均词向量。
更进一步地,所述的实体名向量采用拼接K次幂平均词向量,对于某一实体的实体名s,构成该实体名的所有词的词向量用矩阵形式表示为
Figure PCTCN2020102683-appb-000035
其中l代表词的数目,d代表嵌入的维度,先对实体名的词向量计算K次幂平均词向量,然后将这K次幂平均词向量拼接后,生成实体名向量
Figure PCTCN2020102683-appb-000036
即:
Figure PCTCN2020102683-appb-000037
其中
Figure PCTCN2020102683-appb-000038
代表沿着行的拼接操作,而p 1,...,p K代表K个不同的幂平均的具体数值。
更进一步地,K个不同的幂平均的具体数值分别取1,负无穷和正无穷三个数值。
具体地,所述的结构相似度Sim s(e 1,e 2)为两个实体的结构向量
Figure PCTCN2020102683-appb-000039
Figure PCTCN2020102683-appb-000040
的 余弦相似度,所述的实体名相似度Sim t(e 1,e 2)为两个实体的实体名向量
Figure PCTCN2020102683-appb-000041
Figure PCTCN2020102683-appb-000042
的余弦相似度。
具体地,所述选择高置信度实体对的步骤为:对于原有知识图谱中每一个实体e 1,假定其在外部知识图谱中最相似的实体是e 2,第二相似的实体是e′ 2,相似度差值为
Figure PCTCN2020102683-appb-000043
若对于外部知识图谱中的e 2,其在原有知识图谱中最相似的实体正好是e 1,第二相似的实体是e′ 1,并且相似度差值为
Figure PCTCN2020102683-appb-000044
若相似度差值Δ 12均高于某一预设值的话,则认为(e 1,e 2)是一个高置信度实体对;
所述知识图谱补全的迭代训练过程是多轮的,对于外部知识图谱中的每一个三元组,如果其头实体和尾实体均在原有知识图谱中,则将外部知识图谱中的实体换成原有知识图谱中对应的实体,并将其添入到原有知识图谱中;接着利用添入后的知识图谱重新学习结构向量、计算实体相似度,生成新的高置信度实体对,继续进行知识图谱添入补全,直到满足停止条件,停止迭代训练。
与现有技术相比,本发明以下优点和有益效果:
(1)提出了度感知的互注意力网络以融合实体名信息和结构信息,使得对齐效果更好;
(2)提出了使用拼接幂平均词向量表征实体名,与平均词向量相比,拼接幂平均词向量更够捕捉实体名的更多信息,并减少向量表示的不确定性;
(3)提出了一种基于知识图谱补全的迭代训练算法,在补充知识图谱结构信息的同时,迭代式地提升实体对齐效果,使得长尾实体更容易对齐。
附图说明
图1为本发明实施例的整体流程示意图;
图2为本发明实施例的互注意力网络结构图;
图3为本发明实施例的整体流程框架图。
具体实施方式
下面结合附图对本发明作进一步的说明,但不以任何方式对本发明加以限制,基于本发明教导所作的任何变换或替换,均属于本发明的保护范围。
如图1所示,一种基于迭代补全的知识图谱融合方法,包括以下步骤:
步骤1,获取多个知识图谱数据,识别知识图谱中的所有实体;
步骤2,对所有实体进行结构向量表示学习,获得每一个实体的结构向量;对所有实体进行实体名向量表示学习,获得每一个实体的实体名向量;
步骤3,根据所述的结构向量计算实体间的结构相似度,根据所述的实体名向量计算实体间的实体名相似度;
步骤4,建立基于度感知的互注意力网络,计算融合后实体间的实体相似度;
步骤5,根据所述的实体相似度选择高置信度实体对,并采用迭代训练进行知识图谱补全,获得融合后的知识图谱。
所述的结构向量的学习可以采用背景技术中已有的方法生成结构向量,结构矩阵表示为
Figure PCTCN2020102683-appb-000045
其中n代表实体个数,d s代表结构向量维度。
给定两个实体e 1和e 2,它们的结构相似度Sim s(e 1,e 2)为
Figure PCTCN2020102683-appb-000046
Figure PCTCN2020102683-appb-000047
的余弦相似度,其中,
Figure PCTCN2020102683-appb-000048
代表e 1的结构向量,
Figure PCTCN2020102683-appb-000049
代表e 2的结构向量。
所述的实体名向量可以采用幂平均词向量,对于某一实体的实体名s,构成该实体名的所有词的词向量用矩阵形式表示为
Figure PCTCN2020102683-appb-000050
其中l代表词的数目,d代表嵌入的维度,对
Figure PCTCN2020102683-appb-000051
进行幂平均操作,即可生成的幂平均 词向量,
Figure PCTCN2020102683-appb-000052
幂平均操作公式为:
Figure PCTCN2020102683-appb-000053
Figure PCTCN2020102683-appb-000054
表示将
Figure PCTCN2020102683-appb-000055
处理后生成的幂平均词向量。
更进一步地,为了捕捉更多的实体名特征,所述的实体名向量可以采用拼接K次幂平均词向量,先对实体名的词向量计算K次幂平均词向量,然后将这K次幂平均词向量拼接后,生成实体名向量
Figure PCTCN2020102683-appb-000056
即:
Figure PCTCN2020102683-appb-000057
其中
Figure PCTCN2020102683-appb-000058
代表沿着行的拼接操作,而p 1,...,p K代表K个不同的幂平均的具体数值。本实施例中,K个不同的幂平均的具体数值分别取1,负无穷和正无穷三个数值。
将所有实体的名字向量表示为矩阵
Figure PCTCN2020102683-appb-000059
其中d n=d×K表示实体名向量的维度。给定两个实体e 1和e 2,它们的实体名相似度Sim t(e 1,e 2)为实体名向量
Figure PCTCN2020102683-appb-000060
和实体名向量
Figure PCTCN2020102683-appb-000061
的余弦相似度。
与平均词向量相比,拼接幂平均词向量更够捕捉实体名的更多信息,并减少向量表示的不确定性。
不同的信息从不同的方面刻画实体。因此,需要通过特征融合有效结合多方面信息。对于不同度数的实体,各种信息的重要度是不同的。对于只具有少许结构信息的长尾实体来说,实体名信息更加重要;相反的,对于常出现的实体,结构信息更为重要。为刻画这种动态变化,设计一个度感知的互注意力网络,如图2所示。
所述的互注意力网络的输入为两个实体间的结构相似度Sim s(e 1,e 2)、实体名相似度Sim t(e 1,e 2)和实体的度数,计算过程包括以下步骤:
步骤401,构建特征矩阵:为每个实体构建一个特征矩阵,由所述实体的 实体名向量
Figure PCTCN2020102683-appb-000062
结构向量
Figure PCTCN2020102683-appb-000063
以及实体度向量
Figure PCTCN2020102683-appb-000064
组成,实体度向量为
Figure PCTCN2020102683-appb-000065
其中,
Figure PCTCN2020102683-appb-000066
是所述实体度数的one-hot向量,
Figure PCTCN2020102683-appb-000067
是全连接参数矩阵,d g是度向量的维度,对于实体e 1,其特征矩阵进一步表示为:
Figure PCTCN2020102683-appb-000068
其中;代表沿着列的拼接,d m=max{d n,d s,d g};
步骤402,计算互注意力相似度矩阵:为动态刻画实体e 1的特征矩阵
Figure PCTCN2020102683-appb-000069
和实体e 2的特征矩阵
Figure PCTCN2020102683-appb-000070
之间的关联,构建一个互注意力相似度矩阵
Figure PCTCN2020102683-appb-000071
其中e 1的第i个特征和e 2的第j个特征之间的相似度为:
Figure PCTCN2020102683-appb-000072
其中,
Figure PCTCN2020102683-appb-000073
是特征矩阵
Figure PCTCN2020102683-appb-000074
第i个行向量,
Figure PCTCN2020102683-appb-000075
是特征矩阵
Figure PCTCN2020102683-appb-000076
第j个列向量,i=1,2,3;j=1,2,3,
Figure PCTCN2020102683-appb-000077
是一个用于生成相似度的可训练标量函数,
Figure PCTCN2020102683-appb-000078
是参数向量,
Figure PCTCN2020102683-appb-000079
代表沿着行的拼接操作,°代表点乘。
步骤403,分配权重,计算实体相似度:利用互注意力相似度矩阵
Figure PCTCN2020102683-appb-000080
生成注意力向量
Figure PCTCN2020102683-appb-000081
Figure PCTCN2020102683-appb-000082
先将
Figure PCTCN2020102683-appb-000083
送入softmax层,再送入平均化层,进而生成注意力向量,其中
Figure PCTCN2020102683-appb-000084
表征e 1特征与e 2特征的相关程度,而
Figure PCTCN2020102683-appb-000085
代表e 2特征与e 1特征的相关程度,最后,通过将不同特征的相似度值与其权重相乘,得到融合后的实体相似度值:
Figure PCTCN2020102683-appb-000086
其中,
Figure PCTCN2020102683-appb-000087
Figure PCTCN2020102683-appb-000088
为注意力向量
Figure PCTCN2020102683-appb-000089
的第1个和第2个值,分别代表结构相似度Sim s(e 1,e 2)和实体名相似度Sim t(e 1,e 2)所对应的权重。
要注意到,Sim(e 1,e 2)≠Sim(e 2,e 1)。该网络通过最大化正负例相似度的差值进行训练。得到实体相似度值后,对于每一个待对齐实体,可选取与其相似度最大的外部知识图谱中的实体作为其的对应实体,实现实体对齐。
长尾实体可能在原有知识图谱中具有很少的结构信息,但在外部知识图谱中具有较为丰富的结构信息。若能将外部知识图谱中的结构信息引入,补充原有知识图谱中长尾实体的结构信息,能够在一定程度上缓解长尾问题,提升知识图谱的覆盖程度。而扩增后的知识图谱则能生成更精准的结构向量,提升实体对齐的效果。
所述选择高置信度实体对的步骤为:对于原有知识图谱中每一个实体e 1,假定其在外部知识图谱中最相似的实体是e 2,第二相似的实体是e′ 2,相似度差值为
Figure PCTCN2020102683-appb-000090
若对于外部知识图谱中的e 2,其在原有知识图谱中最相似的实体正好是e 1,第二相似的实体是e′ 1,并且相似度差值为
Figure PCTCN2020102683-appb-000091
若相似度差值Δ 12均高于某一预设值的话,则认为(e 1,e 2)是一个高置信度实体对;
所述知识图谱补全的迭代训练过程是多轮的,对于外部知识图谱中的每一个三元组,如果其头实体和尾实体均在原有知识图谱中,则将外部知识图谱中的实体换成原有知识图谱中对应的实体,并将其添入到原有知识图谱中;接着利用添入后的知识图谱重新学习结构向量、计算实体相似度,生成新的高置信度实体对,继续进行知识图谱添入补全,直到满足停止条件,停止迭代训练。
从发明内容和实施例内容可知,为了解决结构信息不足情况下的实体对齐问题,本发明方法提出新的实体对齐框架,如图3所示,从而较好地实现了知识图谱的融合。本发明主要的技术效果如下:
在预对齐阶段,本发明方法将实体名作为一种新的对齐信息。区别于现有将实体名向量作为初始特征用于学习结构表示的实体对齐工作,本发明方法将实体名作为单独的特征,通过拼接幂平均词向量来表征实体名,能够捕捉实体名的更多信息,并减少向量表示的不确定性;在对齐阶段,观察到对于不同度 数的实体,结构信息和实体名信息的重要度也是不同的,设计了一个互注意力网络,在度数的引导下确定不同特征的权重,并有效融合多源信息;在对齐后处理阶段,提出一种基于知识图谱补全的迭代训练算法,在补充知识图谱结构信息的同时,迭代式地提升实体对齐效果,进而使得长尾实体更容易对齐。
上述实施例为本发明方法用于知识图谱融合的一种实施方式,但本发明的实施方式并不受所述实施例的限制,其他的任何背离本发明的精神实质与原理下所做的改变、修饰、代替、组合、简化,均应为等效的置换方式,都包含在本发明的保护范围之内。

Claims (7)

  1. 一种基于迭代补全的知识图谱融合方法,其特征在于,包括以下步骤:
    步骤1,获取多个知识图谱数据,识别知识图谱中的所有实体;
    步骤2,对所有实体进行结构向量表示学习,获得每一个实体的结构向量;对所有实体进行实体名向量表示学习,获得每一个实体的实体名向量;
    步骤3,根据所述的结构向量计算实体间的结构相似度,根据所述的实体名向量计算实体间的实体名相似度;
    步骤4,建立基于度感知的互注意力网络,计算融合后实体间的实体相似度;
    步骤5,根据所述的实体相似度选择高置信度实体对,并采用迭代训练进行知识图谱补全,获得融合后的知识图谱。
  2. 根据权利要求1所述的知识图谱融合方法,其特征在于,所述的互注意力网络的计算过程包括以下步骤:
    步骤401,构建特征矩阵:为每个实体构建一个特征矩阵,由所述实体的实体名向量
    Figure PCTCN2020102683-appb-100001
    结构向量
    Figure PCTCN2020102683-appb-100002
    以及实体度向量
    Figure PCTCN2020102683-appb-100003
    组成,实体度向量为
    Figure PCTCN2020102683-appb-100004
    其中,
    Figure PCTCN2020102683-appb-100005
    是所述实体度数的one-hot向量,
    Figure PCTCN2020102683-appb-100006
    是全连接参数矩阵,d g是度向量的维度,对于实体e 1,其特征矩阵进一步表示为:
    Figure PCTCN2020102683-appb-100007
    其中;代表沿着列的拼接,d m=max{d n,d s,d g},d n表示实体名向量的维度,d s表示结构向量的维度;
    步骤402,计算互注意力相似度矩阵:为动态刻画实体e 1的特征矩阵
    Figure PCTCN2020102683-appb-100008
    和实体e 2的特征矩阵
    Figure PCTCN2020102683-appb-100009
    之间的关联,构建一个互注意力相似度矩阵
    Figure PCTCN2020102683-appb-100010
    其中实体e 1的第i个特征和实体e 2的第j个特征之间的相似度为:
    Figure PCTCN2020102683-appb-100011
    其中,
    Figure PCTCN2020102683-appb-100012
    是特征矩阵
    Figure PCTCN2020102683-appb-100013
    第i个行向量,
    Figure PCTCN2020102683-appb-100014
    是特征矩阵
    Figure PCTCN2020102683-appb-100015
    第j个列向量, i=1,2,3;j=1,2,3,
    Figure PCTCN2020102683-appb-100016
    是一个用于生成相似度的可训练标量函数,
    Figure PCTCN2020102683-appb-100017
    是参数向量,
    Figure PCTCN2020102683-appb-100018
    代表沿着行的拼接操作,ο代表点乘;
    步骤403,分配权重,计算实体相似度:利用互注意力相似度矩阵
    Figure PCTCN2020102683-appb-100019
    生成注意力向量
    Figure PCTCN2020102683-appb-100020
    Figure PCTCN2020102683-appb-100021
    先将
    Figure PCTCN2020102683-appb-100022
    送入softmax层,再送入平均化层,进而生成注意力向量,其中
    Figure PCTCN2020102683-appb-100023
    表征e 1特征与e 2特征的相关程度,而
    Figure PCTCN2020102683-appb-100024
    代表e 2特征与e 1特征的相关程度,最后,通过将不同特征的相似度值与其权重相乘,得到融合后的实体相似度值:
    Figure PCTCN2020102683-appb-100025
    其中,
    Figure PCTCN2020102683-appb-100026
    Figure PCTCN2020102683-appb-100027
    为注意力向量
    Figure PCTCN2020102683-appb-100028
    的第1个和第2个值,分别代表结构相似度Sim s(e 1,e 2)和实体名相似度Sim t(e 1,e 2)所对应的权重。
  3. 根据权利要求1或2所述的知识图谱融合方法,其特征在于,所述的实体名向量为幂平均词向量,对于某一实体的实体名s,构成该实体名的所有词的词向量用矩阵形式表示为
    Figure PCTCN2020102683-appb-100029
    其中l代表词的数目,d代表嵌入的维度,对
    Figure PCTCN2020102683-appb-100030
    进行幂平均操作,即可生成的幂平均词向量,
    Figure PCTCN2020102683-appb-100031
    幂平均操作公式为:
    Figure PCTCN2020102683-appb-100032
    Figure PCTCN2020102683-appb-100033
    表示将
    Figure PCTCN2020102683-appb-100034
    处理后生成的幂平均词向量。
  4. 根据权利要求1或2所述的知识图谱融合方法,其特征在于,所述的实体名向量采用拼接K次幂平均词向量,对于某一实体的实体名s,构成该实体名的所有词的词向量用矩阵形式表示为
    Figure PCTCN2020102683-appb-100035
    其中l代表词的数目,d代表嵌入的维度,先对实体名的词向量计算K次幂平均词向量,然后将这K次幂平均词向量拼接后,生成实体名向量
    Figure PCTCN2020102683-appb-100036
    即:
    Figure PCTCN2020102683-appb-100037
    其中,
    Figure PCTCN2020102683-appb-100038
    代表沿着行的拼接操作,而p 1,...,p K代表K个不同的幂平均的具体数值。
  5. 根据权利要求4所述的知识图谱融合方法,其特征在于,K个不同的幂平均的具体数值分别取1,负无穷和正无穷三个数值。
  6. 根据权利要求1或5所述的知识图谱融合方法,其特征在于,所述的结构相似度Sim s(e 1,e 2)为两个实体的结构向量
    Figure PCTCN2020102683-appb-100039
    Figure PCTCN2020102683-appb-100040
    的余弦相似度,所述的实体名相似度Sim t(e 1,e 2)为两个实体的实体名向量
    Figure PCTCN2020102683-appb-100041
    Figure PCTCN2020102683-appb-100042
    的余弦相似度。
  7. 根据权利要求1所述的知识图谱融合方法,其特征在于,所述选择高置信度实体对的步骤为:对于原有知识图谱中每一个实体e 1,假定其在外部知识图谱中最相似的实体是e 2,第二相似的实体是e′ 2,相似度差值为
    Figure PCTCN2020102683-appb-100043
    若对于外部知识图谱中的e 2,其在原有知识图谱中最相似的实体正好是e 1,第二相似的实体是e′ 1,并且相似度差值为
    Figure PCTCN2020102683-appb-100044
    若相似度差值Δ 12均高于某一预设值的话,则认为(e 1,e 2)是一个高置信度实体对;
    所述知识图谱补全的迭代训练过程是多轮的,对于外部知识图谱中的每一个三元组,如果其头实体和尾实体均在原有知识图谱中,则将外部知识图谱中的实体换成原有知识图谱中对应的实体,并将其添入到原有知识图谱中;接着利用添入后的知识图谱重新学习结构向量、计算实体相似度,生成新的高置信度实体对,继续进行知识图谱添入补全,直到满足停止条件,停止迭代训练。
PCT/CN2020/102683 2020-07-17 2020-07-17 一种基于迭代补全的知识图谱融合方法 WO2022011681A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
PCT/CN2020/102683 WO2022011681A1 (zh) 2020-07-17 2020-07-17 一种基于迭代补全的知识图谱融合方法
US18/097,292 US20230206127A1 (en) 2020-07-17 2023-01-16 Knowledge graph fusion method based on iterative completion

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2020/102683 WO2022011681A1 (zh) 2020-07-17 2020-07-17 一种基于迭代补全的知识图谱融合方法

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US18/097,292 Continuation US20230206127A1 (en) 2020-07-17 2023-01-16 Knowledge graph fusion method based on iterative completion

Publications (1)

Publication Number Publication Date
WO2022011681A1 true WO2022011681A1 (zh) 2022-01-20

Family

ID=79554468

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/102683 WO2022011681A1 (zh) 2020-07-17 2020-07-17 一种基于迭代补全的知识图谱融合方法

Country Status (2)

Country Link
US (1) US20230206127A1 (zh)
WO (1) WO2022011681A1 (zh)

Cited By (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114417845A (zh) * 2022-03-30 2022-04-29 支付宝(杭州)信息技术有限公司 一种基于知识图谱的相同实体识别方法和系统
CN114626530A (zh) * 2022-03-14 2022-06-14 电子科技大学 一种基于双边路径质量评估的强化学习知识图谱推理方法
CN114708929A (zh) * 2022-04-08 2022-07-05 南京邮电大学 药物与药物相互作用预测模型及其训练方法
CN114942998A (zh) * 2022-04-25 2022-08-26 西北工业大学 融合多源数据的知识图谱邻域结构稀疏的实体对齐方法
CN115203436A (zh) * 2022-07-15 2022-10-18 国网江苏省电力有限公司信息通信分公司 一种基于有向图数据融合的电力知识图谱构建方法和装置
CN115292504A (zh) * 2022-09-29 2022-11-04 北京如炬科技有限公司 实体关系分类方法、装置、设备及存储介质
CN115795060A (zh) * 2023-02-06 2023-03-14 吉奥时空信息技术股份有限公司 一种基于知识增强的实体对齐方法
CN116187446A (zh) * 2023-05-04 2023-05-30 中国人民解放军国防科技大学 基于自适应注意力机制的知识图谱补全方法、装置和设备
CN116702899A (zh) * 2023-08-07 2023-09-05 上海银行股份有限公司 一种适用于公私联动场景的实体融合方法
CN117235285A (zh) * 2023-11-09 2023-12-15 支付宝(杭州)信息技术有限公司 融合知识图谱数据的方法及装置
CN117251583A (zh) * 2023-11-20 2023-12-19 湖北大学 基于局部图结构的文本增强知识图谱表示学习方法及系统
CN117370583A (zh) * 2023-12-08 2024-01-09 湘江实验室 一种基于生成对抗网络的知识图谱实体对齐方法及系统
CN117725555A (zh) * 2024-02-08 2024-03-19 暗物智能科技(广州)有限公司 多源知识树的关联融合方法、装置、电子设备及存储介质

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110245299A (zh) * 2019-06-19 2019-09-17 中国人民解放军国防科技大学 一种基于动态交互注意力机制的序列推荐方法及其系统
CN110489554A (zh) * 2019-08-15 2019-11-22 昆明理工大学 基于位置感知互注意力网络模型的属性级情感分类方法
CN110851613A (zh) * 2019-09-09 2020-02-28 中国电子科技集团公司电子科学研究院 基于实体概念的知识图谱补全、推演、存储方法及装置
US20200104311A1 (en) * 2018-09-27 2020-04-02 Babylon Partners Limited Method and system for extracting information from graphs
CN111291139A (zh) * 2020-03-17 2020-06-16 中国科学院自动化研究所 基于注意力机制的知识图谱长尾关系补全方法
CN111339313A (zh) * 2020-02-18 2020-06-26 北京航空航天大学 一种基于多模态融合的知识库构建方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200104311A1 (en) * 2018-09-27 2020-04-02 Babylon Partners Limited Method and system for extracting information from graphs
CN110245299A (zh) * 2019-06-19 2019-09-17 中国人民解放军国防科技大学 一种基于动态交互注意力机制的序列推荐方法及其系统
CN110489554A (zh) * 2019-08-15 2019-11-22 昆明理工大学 基于位置感知互注意力网络模型的属性级情感分类方法
CN110851613A (zh) * 2019-09-09 2020-02-28 中国电子科技集团公司电子科学研究院 基于实体概念的知识图谱补全、推演、存储方法及装置
CN111339313A (zh) * 2020-02-18 2020-06-26 北京航空航天大学 一种基于多模态融合的知识库构建方法
CN111291139A (zh) * 2020-03-17 2020-06-16 中国科学院自动化研究所 基于注意力机制的知识图谱长尾关系补全方法

Cited By (21)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114626530A (zh) * 2022-03-14 2022-06-14 电子科技大学 一种基于双边路径质量评估的强化学习知识图谱推理方法
CN114417845A (zh) * 2022-03-30 2022-04-29 支付宝(杭州)信息技术有限公司 一种基于知识图谱的相同实体识别方法和系统
CN114708929A (zh) * 2022-04-08 2022-07-05 南京邮电大学 药物与药物相互作用预测模型及其训练方法
CN114708929B (zh) * 2022-04-08 2024-03-08 南京邮电大学 药物与药物相互作用预测模型及其训练方法
CN114942998A (zh) * 2022-04-25 2022-08-26 西北工业大学 融合多源数据的知识图谱邻域结构稀疏的实体对齐方法
CN114942998B (zh) * 2022-04-25 2024-02-13 西北工业大学 融合多源数据的知识图谱邻域结构稀疏的实体对齐方法
CN115203436B (zh) * 2022-07-15 2023-12-15 国网江苏省电力有限公司信息通信分公司 一种基于有向图数据融合的电力知识图谱构建方法和装置
CN115203436A (zh) * 2022-07-15 2022-10-18 国网江苏省电力有限公司信息通信分公司 一种基于有向图数据融合的电力知识图谱构建方法和装置
CN115292504A (zh) * 2022-09-29 2022-11-04 北京如炬科技有限公司 实体关系分类方法、装置、设备及存储介质
CN115795060A (zh) * 2023-02-06 2023-03-14 吉奥时空信息技术股份有限公司 一种基于知识增强的实体对齐方法
CN116187446A (zh) * 2023-05-04 2023-05-30 中国人民解放军国防科技大学 基于自适应注意力机制的知识图谱补全方法、装置和设备
CN116187446B (zh) * 2023-05-04 2023-07-04 中国人民解放军国防科技大学 基于自适应注意力机制的知识图谱补全方法、装置和设备
CN116702899B (zh) * 2023-08-07 2023-11-28 上海银行股份有限公司 一种适用于公私联动场景的实体融合方法
CN116702899A (zh) * 2023-08-07 2023-09-05 上海银行股份有限公司 一种适用于公私联动场景的实体融合方法
CN117235285B (zh) * 2023-11-09 2024-02-02 支付宝(杭州)信息技术有限公司 融合知识图谱数据的方法及装置
CN117235285A (zh) * 2023-11-09 2023-12-15 支付宝(杭州)信息技术有限公司 融合知识图谱数据的方法及装置
CN117251583A (zh) * 2023-11-20 2023-12-19 湖北大学 基于局部图结构的文本增强知识图谱表示学习方法及系统
CN117251583B (zh) * 2023-11-20 2024-01-26 湖北大学 基于局部图结构的文本增强知识图谱表示学习方法及系统
CN117370583A (zh) * 2023-12-08 2024-01-09 湘江实验室 一种基于生成对抗网络的知识图谱实体对齐方法及系统
CN117370583B (zh) * 2023-12-08 2024-03-19 湘江实验室 一种基于生成对抗网络的知识图谱实体对齐方法及系统
CN117725555A (zh) * 2024-02-08 2024-03-19 暗物智能科技(广州)有限公司 多源知识树的关联融合方法、装置、电子设备及存储介质

Also Published As

Publication number Publication date
US20230206127A1 (en) 2023-06-29

Similar Documents

Publication Publication Date Title
WO2022011681A1 (zh) 一种基于迭代补全的知识图谱融合方法
CN108399428B (zh) 一种基于迹比准则的三元组损失函数设计方法
CN110021051A (zh) 一种基于生成对抗网络通过文本指导的人物图像生成方法
CN110110642A (zh) 一种基于多通道注意力特征的行人重识别方法
CN111797703A (zh) 基于鲁棒深度语义分割网络的多源遥感影像分类方法
CN109255002B (zh) 一种利用关系路径挖掘解决知识图谱对齐任务的方法
CN105550649A (zh) 基于全耦合局部约束表示的极低分辨率人脸识别方法及系统
CN114090783A (zh) 一种异构知识图谱融合方法及系统
CN110347857A (zh) 基于强化学习的遥感影像的语义标注方法
CN115438197B (zh) 一种基于双层异质图的事理知识图谱关系补全方法及系统
CN107885787A (zh) 基于谱嵌入的多视角特征融合的图像检索方法
CN111027700A (zh) 基于wcur算法的知识库补全方法
CN109145763A (zh) 基于自然语言描述的视频监控行人搜索图像文本融合方法
CN112464816A (zh) 基于二次迁移学习的地方手语识别方法、装置
Zhang et al. An open set domain adaptation algorithm via exploring transferability and discriminability for remote sensing image scene classification
CN116258990A (zh) 一种基于跨模态亲和力的小样本参考视频目标分割方法
Wang et al. Graph meta transfer network for heterogeneous few-shot hyperspectral image classification
CN114519107A (zh) 一种联合实体关系表示的知识图谱融合方法
Li et al. Msvit: training multiscale vision transformers for image retrieval
CN113807214A (zh) 基于deit附属网络知识蒸馏的小目标人脸识别方法
CN117350330A (zh) 基于混合教学的半监督实体对齐方法
CN116862080A (zh) 一种基于双视角对比学习的碳排放预测方法及系统
Guo et al. TLRM: Task-level relation module for GNN-based few-shot learning
CN113160291B (zh) 一种基于图像配准的变化检测方法
Liu et al. Data assimilation network for generalizable person re-identification

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20945094

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20945094

Country of ref document: EP

Kind code of ref document: A1

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 060723)

122 Ep: pct application non-entry in european phase

Ref document number: 20945094

Country of ref document: EP

Kind code of ref document: A1