CN112417159A - Cross-language entity alignment method of context alignment enhanced graph attention network - Google Patents

Cross-language entity alignment method of context alignment enhanced graph attention network Download PDF

Info

Publication number
CN112417159A
CN112417159A CN202011201832.1A CN202011201832A CN112417159A CN 112417159 A CN112417159 A CN 112417159A CN 202011201832 A CN202011201832 A CN 202011201832A CN 112417159 A CN112417159 A CN 112417159A
Authority
CN
China
Prior art keywords
entity
graph
knowledge
entities
aligned
Prior art date
Legal status (The legal status 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 status listed.)
Granted
Application number
CN202011201832.1A
Other languages
Chinese (zh)
Other versions
CN112417159B (en
Inventor
刘进
赵焜松
谢志文
崔晓晖
黄勃
周光有
匡秋明
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Wuhan University WHU
Original Assignee
Wuhan University WHU
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 Wuhan University WHU filed Critical Wuhan University WHU
Priority to CN202011201832.1A priority Critical patent/CN112417159B/en
Publication of CN112417159A publication Critical patent/CN112417159A/en
Application granted granted Critical
Publication of CN112417159B publication Critical patent/CN112417159B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/36—Creation of semantic tools, e.g. ontology or thesauri
    • G06F16/367—Ontology
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/20—Natural language analysis
    • G06F40/205—Parsing
    • G06F40/211—Syntactic parsing, e.g. based on context-free grammar [CFG] or unification grammars
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/20—Natural language analysis
    • G06F40/279—Recognition of textual entities
    • G06F40/289—Phrasal analysis, e.g. finite state techniques or chunking
    • G06F40/295—Named entity recognition
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Computational Linguistics (AREA)
  • General Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • Biomedical Technology (AREA)
  • Computing Systems (AREA)
  • Molecular Biology (AREA)
  • Evolutionary Computation (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Biophysics (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Animal Behavior & Ethology (AREA)
  • Databases & Information Systems (AREA)
  • Machine Translation (AREA)

Abstract

The invention provides a cross-language entity alignment method of a context alignment enhancement graph attention network. Introducing a first knowledge graph and a second knowledge graph, screening and aligning a seed entity set, and translating each entity name into English; constructing a training set and a testing set from the translated aligned seed set, and converting the entity name into a word vector by using a word2vec algorithm; initial features of the two maps are constructed separately by summing the word vectors for each entity name. Dividing a training set into a context alignment seed set and a target alignment seed set, and taking the context alignment seed set and the target alignment seed set together with the initial characteristics as input data; the features of each entity containing map fusion information and multi-hop neighbor information are obtained by a cross-knowledge-map aggregation layer and an attention-based neural network. The method fully utilizes the context to align the seed set, and transmits the information between the maps through the cross-knowledge map aggregation layer; communicating entity neighbor information and entity alignment information across atlases is collected by an attention-based graph neural network.

Description

Cross-language entity alignment method of context alignment enhanced graph attention network
Technical Field
The invention relates to a cross-language entity alignment method, in particular to a method for collecting and propagating heterogeneous knowledge graph information by using a cross-knowledge graph neural network model to solve the problem of cross-language knowledge graph entity alignment, which comprises a cross-knowledge graph aggregation layer and an attention-based information propagation layer. The pre-aligned seed entity pair is regarded as a medium for information transfer between two heterogeneous knowledge graphs of different languages, the information transfer is carried out between the two heterogeneous knowledge graphs, so that more neighbor alignment characteristics are provided for equivalent entities in different knowledge graphs, and then the equivalent entities in different knowledge graphs are predicted through the learned characteristic representation.
Background
In recent years, knowledge-graphs have shown great potential in many natural language processing tasks, such as language modeling and question and answer. With the rapid growth of multi-language knowledge maps (e.g., DBpedia, YAGO), cross-language entity alignment has attracted much researchers' attention due to the lack of connections between cross-language entities. The cross-language entity alignment task aims to automatically search equivalent entities from different single-language knowledge maps so as to close gaps of different languages.
Recently, many Graph Neural Network (GNNs) based methods have been proposed for entity alignment tasks. GNN-based methods achieve good performance because GNN can learn representations of graph structure data by aggregating neighborhood information. However, existing GNN-based approach models separately model two knowledge-graphs across languages, ignoring the useful pre-aligned connections (seed entities) between the two knowledge-graphs. These GNN-based methods only optimize the objective function using the seed entities during the training process, but do not make full use of the seed entities that provide context alignment information, resulting in non-optimal results.
Disclosure of Invention
The invention provides a cross-language entity alignment method based on a context alignment enhancement graph attention network, aiming at the defects of the prior art in cross-language entity alignment, and the method comprises a cross-knowledge graph aggregation layer and an attention-based cross-knowledge graph propagation layer. The information across the knowledge graphs is propagated through the pre-aligned pairs of seed entities, so that feature representations of different knowledge graphs are obtained, and then equivalent entities in different knowledge graphs are predicted through learned feature representations (embedding).
In order to achieve the above object, the present invention is conceived as follows: firstly, regarding a seed entity pair as a medium for information transfer between two heterogeneous knowledge graphs, and collecting cross-knowledge graph information by using a cross-knowledge graph aggregation layer; the neighborhood information for the entity is then collected using an attention-based graph neural network. And learning multi-hop neighbor information by stacking a plurality of the two layers. Finally, model parameters are optimized using an edge-based loss function.
According to the conception, the invention adopts a technical scheme that: a cross-language entity alignment method of a context alignment enhanced graph attention network is provided, which comprises the following steps:
step 1: introducing a first knowledge graph and a second knowledge graph, screening an aligned seed entity set according to the first knowledge graph and the second knowledge graph, translating the name of each entity in each aligned entity pair in the aligned seed entity set into English, defining the translated name of each entity in each aligned entity pair in the aligned seed entity set, and constructing a training set and a test set from the translated aligned seed entity set;
step 2: converting the translated name of each entity in the aligned entity pair in the aligned seed entity set into a word vector of the entity name by using a word2vec algorithm, and 2, summing the word vectors of each entity name to serve as the initialization feature of the entity, and respectively constructing the initialization feature of the first knowledge graph and the initialization feature of the second knowledge graph;
and step 3: randomly dividing a training set into a context alignment seed set and a target alignment seed set, and constructing input data of a neural network through the context alignment seed set, the target alignment seed set, the initialization feature of a first knowledge graph and the initialization feature of a second knowledge graph;
and 4, step 4: information transmission of different knowledge graphs is carried out through a cross-knowledge graph aggregation layer;
and 5: collecting neighbor information of each entity in the first knowledge graph and the second knowledge graph through the attention-based neural network;
step 6: aligning the context with the seed set and inputting the initialized characteristics of the entity into the model, and obtaining the characteristics of each entity containing map fusion information and multi-hop neighbor information through the steps 4 and 5;
preferably, the first knowledge-graph in step 1 is:
G1＝(E1，R1，T1)
step 1 the second knowledge-graph is:
G2＝(E2，R2，T2)
wherein E is1Set of entities, R, representing a first knowledge-graph1Set of relationships, T, representing a first knowledge-graph1Three tuple sets representing a first knowledge graph, E2Set of entities, R, representing a second knowledge-graph2Set of relationships, T, representing a second knowledge-graph2A set of triplets representing a second knowledge-graph,
step 1, aligning the seed entity sets:
A＝{ak＝(ek,1,i，ek,2,j)|ek,1,i∈E1，ek,2,j∈E2}
i∈[1,M]，j∈[1,N]
k∈[1,K]
wherein A represents an aligned set of seed entities, akRepresents the K-th aligned entity pair, K represents the number of aligned entity pairs in the aligned set of seed entities, ek,1,iRepresenting the ith entity from the entity set of the first knowledge-graph in the k-th aligned pair of entities, ek,2,jRepresenting the jth entity of the k aligned pair of entities from the set of entities of the second knowledge-graph, ek,1,iAnd ek,2,jThe two Chinese meanings are the same, M represents the number of the entities from the first knowledge graph in A, and N represents the number of the entities from the second knowledge graph in A;
step 1, translating the name of each aligned entity pair in the aligned seed entity set into english as follows:
align the names of the entities in each aligned entity pair in A, i.e. ek,1,i,ek,2,jAll translated into English, and the set of aligned seed entities after translation is marked as A*Specifically defined as:
A*＝{a* k＝(e* k,1,i，e* k,2,j)}
i∈[1,M]，j∈[1,N]
k∈[1,K]
wherein A is*Representing a set of post-translationally aligned seed entities, a* kRepresenting the K-th aligned entity pair after translation, K representing the number of aligned entity pairs in the set of seed entities aligned after translation, e* k,1,iRepresenting the translated ith entity from the entity set of the first knowledge-graph in the k aligned pair of entities, e* k,2,jRepresenting the i-th translated entity from the entity set of the first knowledge-graph in the k-th aligned entity pair, M represents A*The number of entities from the first knowledge-graph in (A), N represents*The number of entities from the second knowledge-graph;
step 1, defining the name of each translated entity as:
the translated name of each entity is contained in a plurality of English words, and is specifically represented as follows:
Figure BDA0002755608780000031
Figure BDA0002755608780000032
wherein, wordk，1，i，tRepresenting the tth word, of the translated ith entity from the entity set of the first knowledge-graph in the kth aligned pair of entitiesk，2，j，tRepresenting the tth word of the translated jth entity from the entity set of the second knowledge-graph in the kth aligned entity pair, n being the total number of words of the translated ith entity from the entity set of the first knowledge-graph in the kth aligned entity pair, m being the total number of words of the translated jth entity from the entity set of the second knowledge-graph in the kth aligned entity pair;
step 1, the construction of a training set and a test set from the set of translated aligned seed entities is as follows:
from a set of post-translationally aligned seed entities, namely A*Randomly selecting P entities from the K aligned entity pairs as a training set, and using AtrainIs represented by A*The remaining K-P aligned entity pairs in the test set, using AtestRepresents;
preferably, the word vector of each entity name in step 2 is:
Figure BDA0002755608780000041
i∈[1,M]，j∈[1,N]
k∈[1,K]
wherein the content of the first and second substances,
Figure BDA0002755608780000042
representing a translated set of entities from the first knowledge-graph in the k-th aligned pair of entitiesThe word vector for the t word for the ith entity,
Figure BDA0002755608780000043
a word vector representing the t word of the translated j entity from the entity set of the second knowledge-graph in the K aligned entity pair, K represents the number of aligned entity pairs in the set of translated aligned seed entities, M represents A*The number of entities from the first knowledge-graph in (A), N represents*The number of entities from the second knowledge-graph;
step 2, summing the word vectors of each entity name as the initialization feature of the entity, specifically as follows:
Figure BDA0002755608780000044
Figure BDA0002755608780000045
where n is the total number of words of the translated ith entity from the entity set of the first knowledge-graph in the kth aligned pair of entities, m is the total number of words of the translated ith entity from the entity set of the first knowledge-graph in the kth aligned pair of entities,
Figure BDA0002755608780000046
representing initialization features of the translated ith entity from the entity set of the first knowledge-graph in the kth aligned pair of entities,
Figure BDA0002755608780000047
representing initialized features of a translated jth entity of the kth aligned pair of entities from the entity set of the second knowledge-graph;
step 2, respectively constructing the initialization features from the first knowledge graph and the initialization features from the second knowledge graph as follows:
initializing a feature of the first knowledge-graph
Figure BDA0002755608780000048
The specific definition is as follows:
Figure BDA0002755608780000051
initialization features of the second knowledge-graph are noted
Figure BDA0002755608780000052
The specific definition is as follows:
Figure BDA0002755608780000053
wherein E is1Set of entities representing a first knowledge-graph, E2Set of entities representing a second knowledge-graph, ek,1,iRepresenting the ith entity from the entity set of the first knowledge-graph in the k-th aligned pair of entities, ek,2,jRepresenting the jth entity, A, of the k-th aligned pair of entities from the set of entities of the second knowledge-graph*Representing a set of post-translationally aligned seed entities, e* k,1,iRepresenting the translated ith entity from the entity set of the first knowledge-graph in the k aligned pair of entities, e* k,2,jRepresenting a translated jth entity from the entity set of the second knowledge-graph in the kth aligned pair of entities;
preferably, the training set in step 3 is Atrain；
Step 3, the context alignment seed set is ActxThe input data is used as the input data of the model and is used for transmitting information between the first knowledge graph and the second knowledge graph;
step 3, the target alignment seed set is AobjFor calculating a loss function;
step 3, the initialization characteristic of the first knowledge graph is
Figure BDA0002755608780000054
As input to the modelInputting data;
step 3 the initialization characteristic of the second knowledge-graph is
Figure BDA0002755608780000055
As input data for the model;
preferably, the information propagation of different knowledge graphs through the cross-knowledge-graph aggregation layer in the step 4 is as follows:
step 4.1: calculating the weight of the information fusion of the first knowledge graph and the second knowledge graph by using a door mechanism according to the initialization characteristics of the translated entity from the entity set of the first knowledge graph in the seed set and the aligned entity pair in the following text alignment and the initialization characteristics of the translated entity from the entity set of the second knowledge graph in the aligned entity pair, and specifically calculating the weight as follows:
Figure BDA0002755608780000056
Figure BDA0002755608780000057
i∈[1,M]，j∈[1,N]
k∈[1,K]
wherein, | | represents the concatenation operation of the vector, W is the weight matrix learned in the training process, b is the deviation, and σ is the sigmoid activation function.
Figure BDA0002755608780000061
Is A*From knowledge-graph G in the k-th entity pair1The embedding of the ith entity of (1),
Figure BDA0002755608780000062
is A*From knowledge-graph G in the k-th entity pair2Embedding of the j-th entity in (a). l refers to the l-th layer of the network. K represents the number of aligned entity pairs in the set of post-translation aligned seed entities, M represents A*From the first knowledge-graph G1The number of entities, N represents A*The number of entities from the second knowledge-graph G2;
Figure BDA0002755608780000063
is represented by A*From knowledge-graph G in the k-th entity pair1The fusion weight of the ith entity of (1),
Figure BDA0002755608780000064
is represented by A*From knowledge-graph G in the k-th entity pair2The fusion weight of the jth entity of (1);
step 4.2: and fusing the information of the first knowledge graph and the second knowledge graph by using a door mechanism to obtain a fused representation of the first knowledge graph and the second knowledge graph, wherein the specific calculation is as follows:
Figure BDA0002755608780000065
Figure BDA0002755608780000066
wherein the content of the first and second substances,
Figure BDA0002755608780000067
is represented by A*From knowledge-graph G in the k-th entity pair1The fusion weight of the ith entity of (1),
Figure BDA0002755608780000068
is represented by A*From knowledge-graph G in the k-th entity pair2The fusion weight of the jth entity of (1);
Figure BDA0002755608780000069
is represented by A*From knowledge-graph G in the k-th entity pair1The embedding of the ith entity of (1),
Figure BDA00027556087800000610
A*the k th best inBody centered from knowledge graph G1(ii) embedding of the ith entity of (1);
Figure BDA00027556087800000611
is represented by A*From knowledge-graph G in the k-th entity pair1The ith entity of (2) is subjected to embedding after a door mechanism,
Figure BDA00027556087800000612
is represented by A*From knowledge-graph G in the k-th entity pair2The jth entity of (1) is subjected to embedding after a door mechanism,
step 4.3: calculating characteristic matrixes after cross-knowledge-map aggregation layers for all entities in the first knowledge map and the second knowledge map;
the characteristic matrix after the cross-knowledge-map aggregation layer is as follows:
Figure BDA0002755608780000071
Figure BDA0002755608780000072
wherein, the calculation formula of crossAggr is as follows:
Figure BDA0002755608780000073
wherein the content of the first and second substances,
Figure BDA0002755608780000074
representing a feature representation of an ith entity from the first knowledge-graph in a kth entity pair of the set of translated aligned seed entities,
Figure BDA0002755608780000075
representing post-translation aligned set of seed entities A*(ii) a feature representation of an ith entity from the second knowledge-graph in the kth pair of entities;
preferably, the step 5 collects the neighbor information of each entity in the first knowledge graph and the second knowledge graph by the attention-based neural network as follows:
step 5.1: the weights, i.e. attention, are calculated according to the neighbor entities of the entity, and the specific calculation formula is as follows:
Figure BDA0002755608780000076
wherein, the first and second guide rollers are arranged in a row,
Figure BDA0002755608780000077
representing neutralizing entity e in a first knowledge-graphk，1，iSet of linked neighbour entities, exp is an exponential function;
Figure BDA0002755608780000078
representing the characteristic representation of the ith entity from the first knowledge graph in the kth entity pair in the translated aligned seed entity set after a door mechanism;
Figure BDA0002755608780000079
and
Figure BDA00027556087800000710
respectively represent A*Of the kth pair of entities from the ith entity of the first knowledge-graph, a1，k，iRepresenting the weight of the ith entity from the first knowledge-graph G1 in the kth pair of entities in a;
Figure BDA0002755608780000081
the calculation formula is as follows:
Figure BDA0002755608780000082
wherein LeakyRelu is the activation function and V is the mathematical formulaThe network parameter matrix to be learned is,
Figure BDA0002755608780000083
show that
Figure BDA0002755608780000084
And
Figure BDA0002755608780000085
both features represent the execution of the splicing operation,
Figure BDA0002755608780000086
as entity ek，1，iIs determined to be a certain one of the neighbour entities,
Figure BDA0002755608780000087
representing feature representations of the ith entity from the first knowledge-graph in the kth entity pair of the post-translation aligned seed entity set after a door mechanism
Step 5.2: and (3) fusing neighbor information according to the weight calculated in the step (5.1) to collect the neighbor information of the entity, wherein the calculation formula is as follows:
Figure BDA0002755608780000088
wherein Relu is the activation function, αk，1，i，pIs the first knowledge-graph G1The weight of the p-th neighbor of the ith entity in (1). crossAtt is the attention-based graph neural network layer, crosssaggr is the cross-knowledgegraph aggregation layer,
Figure BDA0002755608780000089
is represented by A*From knowledge-graph G in the k-th entity pair1Is characterized by the i-th entity of
Figure BDA00027556087800000810
Is represented by A*In the kth entity pair, the ith entity from the first knowledge graph is represented by the characteristics of the gate mechanism;
Figure BDA00027556087800000811
is represented by A*(ii) a characteristic representation of a p-th neighbor of the k-th entity pair from the i-th entity of the first knowledge-graph; alpha is alpha1，k，iRepresenting the weight of the ith entity from the first knowledge-graph G1 in the kth pair of entities in a;
step 5.3: calculating weights of all entities of the first knowledge graph and the second knowledge graph according to neighbor entities of the entities and fusing neighbor information according to the calculated weights so as to collect the neighbor information of the entities, namely step 5.1 and step 5.2, and then obtaining a feature matrix after the neighbor information is collected by a graph neural network layer based on attention, namely:
Figure BDA0002755608780000091
Figure BDA0002755608780000092
wherein, crossAtt is the neural network layer of the attention-based map,
Figure BDA0002755608780000093
is the output feature representation of the first knowledge-graph through the cross-knowledge-graph aggregation layer;
Figure BDA0002755608780000094
is the output feature representation of the second knowledge-graph by the cross-knowledge-graph aggregation layer;
preferably, the characteristics of each entity containing the graph fusion information and the multi-hop neighbor information in step 6 are as follows:
Figure BDA0002755608780000095
Figure BDA0002755608780000096
wherein the content of the first and second substances,
Figure BDA0002755608780000097
is the output of the first knowledge-graph through the l-th CGAT layer; wherein
Figure BDA0002755608780000098
Is the output of the second knowledge-graph through the l-th CGAT layer;
Figure BDA0002755608780000099
features initialized for the first knowledge-graph
Figure BDA00027556087800000910
Features initialized for the second knowledge-graph
Figure BDA00027556087800000911
crossAtt is the attention-based graph neural network layer; crossknowledgeable aggregation layer. By superposing L CGAT layers, the characteristics of the first knowledge graph and the second knowledge graph can be updated for L times, and finally the updated characteristics are output
Figure BDA00027556087800000912
And
Figure BDA00027556087800000913
calculating a target loss function on the target alignment seed set, wherein the formula is as follows:
Figure BDA0002755608780000101
wherein the content of the first and second substances,
Figure BDA0002755608780000102
represents the final embedding of the entity,
Figure BDA0002755608780000103
represents the L1 distance between two entities, φ is a model parameter that can be trained, e-Refers to a negative entity corresponding to a certain entity, AobjIs the data used to optimize the model parameters.
And optimizing and updating model parameters phi by using an Adam algorithm, wherein the model parameters phi comprise parameters which are trainable across the knowledge graph aggregation layer and the attention-based graph neural network layer, and constructing a context-based alignment enhancement graph attention network model according to the optimized parameters phi.
The method has the advantages that the context alignment seed set is fully utilized, and the information among the maps is transmitted through the cross-knowledge map aggregation layer; communicating entity neighbor information and entity alignment information across atlases is collected by an attention-based graph neural network.
Drawings
FIG. 1 is a schematic diagram of a context enhancement graph attention network in the present invention.
Detailed Description
The present invention will be described in further detail with reference to examples for the purpose of facilitating understanding and practice of the invention by those of ordinary skill in the art, and it is to be understood that the embodiments described herein are merely illustrative and explanatory of the invention and are not restrictive thereof.
The following describes an embodiment of the present invention with reference to fig. 1:
step 1, introducing a first knowledge graph and a second knowledge graph, screening an aligned seed entity set according to the first knowledge graph and the second knowledge graph, translating the name of each entity in each aligned entity pair in the aligned seed entity set into English, defining the translated name of each entity in each aligned entity pair in the aligned seed entity set, and constructing a training set and a test set from the translated aligned seed entity set;
step 1 the first knowledge-graph is:
G1＝(E1，R1，T1)
step 1 the second knowledge-graph is:
G2＝(E2，R2，T2)
wherein E is1Set of entities, R, representing a first knowledge-graph1Set of relationships, T, representing a first knowledge-graph1Three tuple sets representing a first knowledge graph, E2Set of entities, R, representing a second knowledge-graph2Set of relationships, T, representing a second knowledge-graph2A set of triplets representing a second knowledge-graph,
step 1, aligning the seed entity sets:
A＝{ak＝(ek,1,i，ek,2,j)|ek,1,i∈E1，ek,2,j∈E2}
i∈[1,M]，j∈[1,N]
k∈[1,K]
wherein A represents an aligned set of seed entities, akDenotes the K-th aligned entity pair, K15000 denotes the number of aligned entity pairs in the set of aligned seed entities, ek,1,iRepresenting the ith entity from the entity set of the first knowledge-graph in the k-th aligned pair of entities, ek,2,jRepresenting the jth entity of the k aligned pair of entities from the set of entities of the second knowledge-graph, ek,1,iAnd ek,2,jBoth belong to the kth aligned entity pair and have the same Chinese meaning, M66469 represents the number of entities from the first knowledge-graph in a, and N98125 represents the number of entities from the second knowledge-graph in a;
step 1, translating the name of each aligned entity pair in the aligned seed entity set into english as follows:
align the names of the entities in each aligned entity pair in A, i.e. ek,1,i,ek,2,jAll translated into English, and the set of aligned seed entities after translation is marked as A*Specifically defined as:
A*＝{a* k＝(e* k,1,i，e* k,2,j)}
i∈[1,M]，j∈[1,N]
k∈[1,K]
wherein A is*Representing a set of post-translationally aligned seed entities, a* kDenotes the K-th aligned entity pair after translation, K15000 denotes the number of aligned entity pairs in the set of aligned seed entities after translation, e* k,1,iRepresenting the translated ith entity from the entity set of the first knowledge-graph in the k aligned pair of entities, e* k,2,jRepresenting the i-th translated entity from the entity set of the first knowledge-graph in the k-th aligned entity pair, M66469 represents a*Wherein N is 98125 represents a*The number of entities from the second knowledge-graph;
step 1, defining the name of each translated entity as:
the translated name of each entity is contained in a plurality of English words, and is specifically represented as follows:
Figure BDA0002755608780000121
Figure BDA0002755608780000122
wherein, wordk，1，i，tRepresenting the tth word, of the translated ith entity from the entity set of the first knowledge-graph in the kth aligned pair of entitiesk，2，j，tRepresenting the tth word of the translated jth entity from the entity set of the second knowledge-graph in the kth aligned entity pair, n being the total number of words of the translated ith entity from the entity set of the first knowledge-graph in the kth aligned entity pair, m being the total number of words of the translated jth entity from the entity set of the second knowledge-graph in the kth aligned entity pair;
step 1, the construction of a training set and a test set from the set of translated aligned seed entities is as follows:
from a set of post-translationally aligned seed entities, namely A*In 15000 aligned entity pairs, P4500 are randomly selected as training set, and A is usedtrainIs represented by A*The remaining 10500 aligned entity pairs of K-P in (A) are used as test settestRepresents;
step 2, converting the translated name of each entity in the aligned entity pair in the aligned seed entity set into a word vector of the entity name by using a word2vec algorithm, and summing the word vectors of each entity name as the initialization feature of the entity in the step 2 to respectively construct the initialization feature of the first knowledge graph and the initialization feature of the second knowledge graph;
step 2, the word vector of each entity name is:
Figure BDA0002755608780000123
i∈[1,M]，j∈[1,N]
k∈[1,K]
wherein the content of the first and second substances,
Figure BDA0002755608780000124
a word vector representing the t word of the translated ith entity from the entity set of the first knowledge-graph in the k aligned pair of entities,
Figure BDA0002755608780000125
a word vector representing the t word of the translated j entity from the entity set of the second knowledge-graph in the K aligned entity pair, K4500 representing the number of aligned entity pairs in the set of translated aligned seed entities, M4500 representing a*Wherein N is 4500, which represents a*The number of entities from the second knowledge-graph;
step 2, summing the word vectors of each entity name as the initialization feature of the entity, specifically as follows:
Figure BDA0002755608780000131
Figure BDA0002755608780000132
where n is the total number of words of the translated ith entity from the entity set of the first knowledge-graph in the kth aligned pair of entities, m is the total number of words of the translated ith entity from the entity set of the first knowledge-graph in the kth aligned pair of entities,
Figure BDA0002755608780000133
representing initialization features of the translated ith entity from the entity set of the first knowledge-graph in the kth aligned pair of entities,
Figure BDA0002755608780000134
representing initialized features of a translated jth entity of the kth aligned pair of entities from the entity set of the second knowledge-graph;
step 2, respectively constructing the initialization features from the first knowledge graph and the initialization features from the second knowledge graph as follows:
initializing a feature of the first knowledge-graph
Figure BDA0002755608780000135
The specific definition is as follows:
Figure BDA0002755608780000136
initialization features of the second knowledge-graph are noted
Figure BDA0002755608780000137
The specific definition is as follows:
Figure BDA0002755608780000138
wherein E is1Set of entities representing a first knowledge-graph, E2Set of entities representing a second knowledge-graph, ek,1,iTo representThe ith entity from the entity set of the first knowledge-graph in the k aligned pair of entities, ek,2,jRepresenting the jth entity, A, of the k-th aligned pair of entities from the set of entities of the second knowledge-graph*Representing a set of post-translationally aligned seed entities, e* k,1,iRepresenting the translated ith entity from the entity set of the first knowledge-graph in the k aligned pair of entities, e* k,2,jRepresenting a translated jth entity from the entity set of the second knowledge-graph in the kth aligned pair of entities;
and step 3, the context alignment enhancement graph attention network (CGAT) mainly comprises a cross-knowledge graph aggregation layer and an attention-based graph neural network layer. The cross-knowledge-graph aggregation layer is used to pass cross-knowledge-graph information between two knowledge graphs, while the attention-based graph neural network layer is used to collect neighbor information for each entity in the knowledge graphs. By superposing a plurality of CGAT layers, multi-hop cross-knowledge map information and neighbor information are spread in a knowledge map. The following describes the construction and training process of the model in detail.
And 4, step 4: randomly dividing a training set into a context alignment seed set and a target alignment seed set, and constructing input data of a neural network through the context alignment seed set, the target alignment seed set, the initialization feature of a first knowledge graph and the initialization feature of a second knowledge graph;
step 4 the training set is Atrain；
Step 4, the context alignment seed set is ActxThe input data is used as the input data of the model and is used for transmitting information between the first knowledge graph and the second knowledge graph;
step 4, the target alignment seed set is AobjFor calculating a loss function;
step 4 the initialization characteristic of the first knowledge-graph is
Figure BDA0002755608780000141
As input data for the model;
step 4 the initialization characteristic of the second knowledge-graph is
Figure BDA0002755608780000142
As input data for the model;
and 5: information transmission of different knowledge graphs is carried out through a cross-knowledge graph aggregation layer;
and 5, the information propagation of different knowledge graphs through the cross-knowledge-graph aggregation layer comprises the following steps:
step 5.1: calculating the weight of the information fusion of the first knowledge graph and the second knowledge graph by using a door mechanism according to the initialization characteristics of the translated entity from the entity set of the first knowledge graph in the seed set and the aligned entity pair in the following text alignment and the initialization characteristics of the translated entity from the entity set of the second knowledge graph in the aligned entity pair, and specifically calculating the weight as follows:
Figure BDA0002755608780000143
Figure BDA0002755608780000144
i∈[1,M]，j∈[1,N]
k∈[1,K]
wherein, | | represents the concatenation operation of the vector, W is the weight matrix learned in the training process, b is the deviation, and σ is the sigmoid activation function.
Figure BDA0002755608780000145
Is A*From knowledge-graph G in the k-th entity pair1The embedding of the ith entity of (1),
Figure BDA0002755608780000146
is A*From knowledge-graph G in the k-th entity pair2Embedding of the j-th entity in (a). l refers to the l-th layer of the network. K4500 denotes the number of aligned entity pairs in the set of post-translation aligned seed entities, M4500 denotes a*From a first knowledge-graph G1 entityThe number, N4500, represents A*The number of entities from the second knowledge-graph G2;
Figure BDA0002755608780000151
is represented by A*From knowledge-graph G in the k-th entity pair1The fusion weight of the ith entity of (1),
Figure BDA0002755608780000152
is represented by A*From knowledge-graph G in the k-th entity pair2The fusion weight of the jth entity of (1);
step 5.2: and fusing the information of the first knowledge graph and the second knowledge graph by using a door mechanism to obtain a fused representation of the first knowledge graph and the second knowledge graph, wherein the specific calculation is as follows:
Figure BDA0002755608780000153
Figure BDA0002755608780000154
wherein the content of the first and second substances,
Figure BDA0002755608780000155
is represented by A*From knowledge-graph G in the k-th entity pair1The fusion weight of the ith entity of (1),
Figure BDA0002755608780000156
is represented by A*From knowledge-graph G in the k-th entity pair2The fusion weight of the jth entity of (1);
Figure BDA0002755608780000157
is represented by A*From knowledge-graph G in the k-th entity pair1The embedding of the ith entity of (1),
Figure BDA0002755608780000158
is represented by A*To middlek pairs of entities from knowledge graph G1(ii) embedding of the ith entity of (1);
Figure BDA0002755608780000159
is represented by A*From knowledge-graph G in the k-th entity pair1The ith entity of (2) is subjected to embedding after a door mechanism,
Figure BDA00027556087800001510
is represented by A*From knowledge-graph G in the k-th entity pair2The jth entity of (1) is subjected to embedding after a door mechanism,
step 5.3: calculating characteristic matrixes after cross-knowledge-map aggregation layers for all entities in the first knowledge map and the second knowledge map;
the characteristic matrix after the cross-knowledge-map aggregation layer is as follows:
Figure BDA00027556087800001511
Figure BDA00027556087800001512
wherein, the calculation formula of crossAggr is as follows:
Figure BDA0002755608780000161
wherein the content of the first and second substances,
Figure BDA0002755608780000162
representing a feature representation of an ith entity from the first knowledge-graph in a kth entity pair of the set of translated aligned seed entities,
Figure BDA0002755608780000163
representing post-translation aligned set of seed entities A*The feature table of the ith entity from the second knowledge-graph in the kth entity pairShown in the specification;
step 6: collecting neighbor information of each entity in the first knowledge graph and the second knowledge graph through the attention-based neural network;
step 6, collecting neighbor information of each entity in the first knowledge graph and the second knowledge graph based on the attention neural network is as follows:
step 6.1: the weights, i.e. attention, are calculated according to the neighbor entities of the entity, and the specific calculation formula is as follows:
Figure BDA0002755608780000164
wherein, the first and second guide rollers are arranged in a row,
Figure BDA0002755608780000165
representing neutralizing entity e in a first knowledge-graphk，1，iSet of linked neighbour entities, exp is an exponential function;
Figure BDA0002755608780000166
representing the characteristic representation of the ith entity from the first knowledge graph in the kth entity pair in the translated aligned seed entity set after a door mechanism;
Figure BDA0002755608780000167
and
Figure BDA0002755608780000168
respectively represent A*Of the kth pair of entities from the ith entity of the first knowledge-graph, a1，k，iRepresenting the weight of the ith entity from the first knowledge-graph G1 in the kth pair of entities in a;
Figure BDA0002755608780000169
the calculation formula is as follows:
Figure BDA00027556087800001610
wherein LeakyRelu is an activation function, V is a learnable network parameter matrix,
Figure BDA0002755608780000171
show that
Figure BDA0002755608780000172
And
Figure BDA0002755608780000173
both features represent the execution of the splicing operation,
Figure BDA0002755608780000174
as entity ek，1，iIs determined to be a certain one of the neighbour entities,
Figure BDA0002755608780000175
representing feature representations of the ith entity from the first knowledge-graph in the kth entity pair of the post-translation aligned seed entity set after a door mechanism
Step 6.2: and fusing neighbor information according to the weight calculated in the step 6.1 to collect the neighbor information of the entity, wherein the calculation formula is as follows:
Figure BDA0002755608780000176
wherein Relu is the activation function, αk，1，i，pIs the first knowledge-graph G1The weight of the p-th neighbor of the ith entity in (1). crossAtt is the attention-based graph neural network layer, crosssaggr is the cross-knowledgegraph aggregation layer,
Figure BDA0002755608780000177
is represented by A*From knowledge-graph G in the k-th entity pair1Is characterized by the i-th entity of
Figure BDA0002755608780000178
Is represented by A*In the kth entity pair, the ith entity from the first knowledge graph is represented by the characteristics of the gate mechanism;
Figure BDA0002755608780000179
is represented by A*(ii) a characteristic representation of a p-th neighbor of the k-th entity pair from the i-th entity of the first knowledge-graph; alpha is alpha1，k，iRepresenting the weight of the ith entity from the first knowledge-graph G1 in the kth pair of entities in a;
step 6.3: calculating weights of all entities of the first knowledge graph and the second knowledge graph according to neighbor entities of the entities and fusing neighbor information according to the calculated weights so as to collect the neighbor information of the entities, namely step 6.1 and step 6.2, and then obtaining a feature matrix after the neighbor information is collected by a graph neural network layer based on attention, namely:
Figure BDA0002755608780000181
Figure BDA0002755608780000182
wherein, crossAtt is the neural network layer of the attention-based map,
Figure BDA0002755608780000183
is the output feature representation of the first knowledge-graph through the cross-knowledge-graph aggregation layer;
Figure BDA0002755608780000184
is the output feature representation of the second knowledge-graph by the cross-knowledge-graph aggregation layer;
and 7: the context is aligned to the seed set and the initialized characteristic of the entity is input into the model, and the characteristic representation of each entity containing the map fusion information and the multi-hop neighbor information can be obtained through the steps 5 and 6, and the formula is as follows:
Figure BDA0002755608780000185
Figure BDA0002755608780000186
wherein the content of the first and second substances,
Figure BDA0002755608780000187
is the output of the first knowledge-graph through the l-th CGAT layer; wherein
Figure BDA0002755608780000188
Is the output of the second knowledge-graph through the l-th CGAT layer;
Figure BDA0002755608780000189
features initialized for the first knowledge-graph
Figure BDA00027556087800001810
Features initialized for the second knowledge-graph
Figure BDA00027556087800001811
crossAtt is the attention-based graph neural network layer; crossknowledgeable aggregation layer. By superposing L CGAT layers, the characteristics of the first knowledge graph and the second knowledge graph can be updated for L times, and finally the updated characteristics are output
Figure BDA00027556087800001812
And
Figure BDA00027556087800001813
calculating a target loss function on the target alignment seed set, wherein the formula is as follows:
Figure BDA0002755608780000191
wherein the content of the first and second substances,
Figure BDA0002755608780000192
represents the final embedding of the entity,
Figure BDA0002755608780000193
represents the L1 distance between two entities, φ is a model parameter that can be trained, e-Refers to a negative entity corresponding to a certain entity, AobjIs the data used to optimize the model parameters.
It should be understood that parts of the application not described in detail are prior art.
It should be understood that the above description of the preferred embodiments is given for clearness of understanding and no unnecessary limitations should be understood therefrom, and all changes and modifications may be made by those skilled in the art without departing from the scope of the invention as defined by the appended claims.

Claims (7)

1. A cross-language entity alignment method of a context alignment enhanced graph attention network is characterized by comprising the following steps:
step 1: introducing a first knowledge graph and a second knowledge graph, screening an aligned seed entity set according to the first knowledge graph and the second knowledge graph, translating the name of each entity in each aligned entity pair in the aligned seed entity set into English, defining the translated name of each entity in each aligned entity pair in the aligned seed entity set, and constructing a training set and a test set from the translated aligned seed entity set;
step 2: converting the translated name of each entity in the aligned entity pair in the aligned seed entity set into a word vector of the entity name by using a word2vec algorithm, and 2, summing the word vectors of each entity name to serve as the initialization feature of the entity, and respectively constructing the initialization feature of the first knowledge graph and the initialization feature of the second knowledge graph;
and step 3: randomly dividing a training set into a context alignment seed set and a target alignment seed set, and constructing input data of a neural network through the context alignment seed set, the target alignment seed set, the initialization feature of a first knowledge graph and the initialization feature of a second knowledge graph;
and 4, step 4: information transmission of different knowledge graphs is carried out through a cross-knowledge graph aggregation layer;
and 5: collecting neighbor information of each entity in the first knowledge graph and the second knowledge graph through the attention-based neural network;
step 6: and (5) inputting the context alignment seed set and the initialized characteristic of the entity into the model, and obtaining the characteristic of each entity containing the map fusion information and the multi-hop neighbor information through the steps 4 and 5.
2. The method of cross-language entity alignment for a context alignment enhanced graph attention network of claim 1, wherein:
step 1 the first knowledge-graph is:
G1＝(E1，R1，T1)
step 1 the second knowledge-graph is:
G2＝(E2，R2，T2)
wherein E is1Set of entities, R, representing a first knowledge-graph1Set of relationships, T, representing a first knowledge-graph1Three tuple sets representing a first knowledge graph, E2Set of entities, R, representing a second knowledge-graph2Set of relationships, T, representing a second knowledge-graph2A set of triplets representing a second knowledge-graph,
step 1, aligning the seed entity sets:
A＝{ak＝(ek,1,i，ek,2,j)|ek,1,i∈E1，ek,2,j∈E2}
i∈[1,M]，j∈[1,N]
k∈[1,K]
wherein A represents alignedSet of seed entities, akRepresents the K-th aligned entity pair, K represents the number of aligned entity pairs in the aligned set of seed entities, ek,1,iRepresenting the ith entity from the entity set of the first knowledge-graph in the k-th aligned pair of entities, ek,2,jRepresenting the jth entity of the k aligned pair of entities from the set of entities of the second knowledge-graph, ek,1,iAnd ek,2,jThe two Chinese meanings are the same, M represents the number of the entities from the first knowledge graph in A, and N represents the number of the entities from the second knowledge graph in A;
step 1, translating the name of each aligned entity pair in the aligned seed entity set into english as follows:
align the names of the entities in each aligned entity pair in A, i.e. ek,1,i,ek,2,jAll translated into English, and the set of aligned seed entities after translation is marked as A*Specifically defined as:
A*＝{a* k＝(e* k,1,i，e* k,2,j)}
i∈[1,M]，j∈[1,N]
k∈[1,K]
wherein A is*Representing a set of post-translationally aligned seed entities, a* kRepresenting the K-th aligned entity pair after translation, K representing the number of aligned entity pairs in the set of seed entities aligned after translation, e* k,1,iRepresenting the translated ith entity from the entity set of the first knowledge-graph in the k aligned pair of entities, e* k,2,jRepresenting the i-th translated entity from the entity set of the first knowledge-graph in the k-th aligned entity pair, M represents A*The number of entities from the first knowledge-graph in (A), N represents*The number of entities from the second knowledge-graph;
step 1, defining the name of each translated entity as:
the translated name of each entity is contained in a plurality of English words, and is specifically represented as follows:
Figure FDA0002755608770000021
Figure FDA0002755608770000022
wherein, wordk，1，i，tRepresenting the tth word, of the translated ith entity from the entity set of the first knowledge-graph in the kth aligned pair of entitiesk，2，j，tRepresenting the tth word of the translated jth entity from the entity set of the second knowledge-graph in the kth aligned entity pair, n being the total number of words of the translated ith entity from the entity set of the first knowledge-graph in the kth aligned entity pair, m being the total number of words of the translated jth entity from the entity set of the second knowledge-graph in the kth aligned entity pair;
step 1, the construction of a training set and a test set from the set of translated aligned seed entities is as follows: from a set of post-translationally aligned seed entities, namely A*Randomly selecting P entities from the K aligned entity pairs as a training set, and using AtrainIs represented by A*The remaining K-P aligned entity pairs in the test set, using AtestAnd (4) showing.
3. The method of cross-language entity alignment for a context alignment enhanced graph attention network of claim 1, wherein:
step 2, the word vector of each entity name is:
Figure FDA0002755608770000031
i∈[1,M]，j∈[1,N]
k∈[1,K]
wherein the content of the first and second substances,
Figure FDA0002755608770000032
a word vector representing the t word of the translated ith entity from the entity set of the first knowledge-graph in the k aligned pair of entities,
Figure FDA0002755608770000033
a word vector representing the t word of the translated j entity from the entity set of the second knowledge-graph in the K aligned entity pair, K represents the number of aligned entity pairs in the set of translated aligned seed entities, M represents A*The number of entities from the first knowledge-graph in (A), N represents*The number of entities from the second knowledge-graph;
step 2, summing the word vectors of each entity name as the initialization feature of the entity, specifically as follows:
Figure FDA0002755608770000034
Figure FDA0002755608770000035
where n is the total number of words of the translated ith entity from the entity set of the first knowledge-graph in the kth aligned pair of entities, m is the total number of words of the translated ith entity from the entity set of the first knowledge-graph in the kth aligned pair of entities,
Figure FDA0002755608770000036
representing initialization features of the translated ith entity from the entity set of the first knowledge-graph in the kth aligned pair of entities,
Figure FDA0002755608770000037
representing initialized features of a translated jth entity of the kth aligned pair of entities from the entity set of the second knowledge-graph;
step 2, respectively constructing the initialization features from the first knowledge graph and the initialization features from the second knowledge graph as follows:
initializing a feature of the first knowledge-graph
Figure FDA0002755608770000038
The specific definition is as follows:
Figure FDA0002755608770000041
initialization features of the second knowledge-graph are noted
Figure FDA0002755608770000042
The specific definition is as follows:
Figure FDA0002755608770000043
wherein E is1Set of entities representing a first knowledge-graph, E2Set of entities representing a second knowledge-graph, ek,1,iRepresenting the ith entity from the entity set of the first knowledge-graph in the k-th aligned pair of entities, ek,2,jRepresenting the jth entity, A, of the k-th aligned pair of entities from the set of entities of the second knowledge-graph*Representing a set of post-translationally aligned seed entities, e* k,1,iRepresenting the translated ith entity from the entity set of the first knowledge-graph in the k aligned pair of entities, e* k,2,jRepresenting a translated jth entity from the entity set of the second knowledge-graph in the kth aligned pair of entities.
4. The method of cross-language entity alignment for a context alignment enhanced graph attention network of claim 1, wherein:
step 3 the training set is Atrain；
Step 3, the context alignment seed set is ActxAs input data for the model, for the first knowledge-graph, the second knowledge-graphInformation is transmitted between knowledge maps;
step 3, the target alignment seed set is AobjFor calculating a loss function;
step 3, the initialization characteristic of the first knowledge graph is
Figure FDA0002755608770000044
As input data for the model;
step 3 the initialization characteristic of the second knowledge-graph is
Figure FDA0002755608770000045
As input data for the model.
5. The method of cross-language entity alignment for a context alignment enhanced graph attention network of claim 1, wherein:
and 4, the information propagation of different knowledge graphs through the cross-knowledge-graph aggregation layer comprises the following steps:
step 4.1: calculating the weight of the information fusion of the first knowledge graph and the second knowledge graph by using a door mechanism according to the initialization characteristics of the translated entity from the entity set of the first knowledge graph in the seed set and the aligned entity pair in the following text alignment and the initialization characteristics of the translated entity from the entity set of the second knowledge graph in the aligned entity pair, and specifically calculating the weight as follows:
Figure FDA0002755608770000046
Figure FDA0002755608770000047
i∈[1,M]，j∈[1,N]
k∈[1,K]
wherein, | | represents the splicing operation of the vector, W is a weight matrix learned in the training process, b is a deviation, and σ is a sigmoid activation function;
Figure FDA0002755608770000051
is A*From knowledge-graph G in the k-th entity pair1The characteristic representation of the i-th entity of (1),
Figure FDA0002755608770000052
is A*From knowledge-graph G in the k-th entity pair2A characteristic representation of the jth entity in (a); l refers to the l-th layer of the network; k represents the number of aligned entity pairs in the set of post-translation aligned seed entities, M represents A*The number of entities from the first knowledge-graph G1, N denotes A*The number of entities from the second knowledge-graph G2;
Figure FDA0002755608770000053
is represented by A*From knowledge-graph G in the k-th entity pair1The fusion weight of the ith entity of (1),
Figure FDA0002755608770000054
is represented by A*From knowledge-graph G in the k-th entity pair2The fusion weight of the jth entity of (1);
step 4.2: and fusing the information of the first knowledge graph and the second knowledge graph by using a door mechanism to obtain a fused representation of the first knowledge graph and the second knowledge graph, wherein the specific calculation is as follows:
Figure FDA0002755608770000055
Figure FDA0002755608770000056
wherein the content of the first and second substances,
Figure FDA0002755608770000057
is represented by A*From knowledge-graph G in the k-th entity pair1The fusion weight of the ith entity of (1),
Figure FDA0002755608770000058
is represented by A*From knowledge-graph G in the k-th entity pair2The fusion weight of the jth entity of (1);
Figure FDA0002755608770000059
is represented by A*From knowledge-graph G in the k-th entity pair1The characteristic representation of the i-th entity of (1),
Figure FDA00027556087700000510
from knowledge-graph G in the k-th entity pair1A characterization of the ith entity of (1);
Figure FDA00027556087700000511
is represented by A*From knowledge-graph G in the k-th entity pair1Characterized by the door mechanism of the ith entity,
Figure FDA00027556087700000512
is represented by A*From knowledge-graph G in the k-th entity pair2Characterized by the door mechanism of the jth entity of (a),
step 4.3: calculating characteristic matrixes after cross-knowledge-map aggregation layers for all entities in the first knowledge map and the second knowledge map;
the characteristic matrix after the cross-knowledge-map aggregation layer is as follows:
Figure FDA0002755608770000061
Figure FDA0002755608770000062
wherein, the calculation formula of crossAggr is as follows:
Figure FDA0002755608770000063
wherein the content of the first and second substances,
Figure FDA0002755608770000064
representing a feature representation of an ith entity from the first knowledge-graph in a kth entity pair of the set of translated aligned seed entities,
Figure FDA0002755608770000065
representing post-translation aligned set of seed entities A*The feature representation of the ith entity from the second knowledge-graph in the kth pair of entities.
6. The method of cross-language entity alignment for a context alignment enhanced graph attention network of claim 1, wherein:
step 5, collecting neighbor information of each entity in the first knowledge graph and the second knowledge graph based on the attention neural network is as follows:
step 5.1: the weights, i.e. attention, are calculated according to the neighbor entities of the entity, and the specific calculation formula is as follows:
Figure FDA0002755608770000066
wherein, the first and second guide rollers are arranged in a row,
Figure FDA0002755608770000067
representing neutralizing entity e in a first knowledge-graphk，1，iSet of linked neighbour entities, exp is an exponential function;
Figure FDA0002755608770000068
representing post-translational aligned seedsPerforming door mechanism on the ith entity from the first knowledge graph in the kth entity pair in the entity set to represent the characteristics;
Figure FDA0002755608770000069
and
Figure FDA00027556087700000610
respectively represent A*Of the kth pair of entities from the ith entity of the first knowledge-graph, a1，k，iRepresenting the weight of the ith entity from the first knowledge-graph G1 in the kth pair of entities in a;
Figure FDA0002755608770000071
the calculation formula is as follows:
Figure FDA0002755608770000072
wherein LeakyRelu is an activation function, V is a learnable network parameter matrix,
Figure FDA0002755608770000073
show that
Figure FDA0002755608770000074
And
Figure FDA0002755608770000075
both features represent the execution of the splicing operation,
Figure FDA0002755608770000076
as entity ek，1，iIs determined to be a certain one of the neighbour entities,
Figure FDA0002755608770000077
representing post-translation aligned sets of seed entitiesIn the kth entity pair, the ith entity from the first knowledge-graph is represented by the characteristics of the gate mechanism
Step 5.2: and (3) fusing neighbor information according to the weight calculated in the step (5.1) to collect the neighbor information of the entity, wherein the calculation formula is as follows:
Figure FDA0002755608770000078
wherein Relu is the activation function, αk，1，i，pIs the first knowledge-graph G1The weight of the p-th neighbor of the ith entity in (1); crossAtt is the attention-based graph neural network layer, crosssaggr is the cross-knowledgegraph aggregation layer,
Figure FDA0002755608770000079
is represented by A*From knowledge-graph G in the k-th entity pair1Is characterized by the i-th entity of
Figure FDA00027556087700000710
Is represented by A*In the kth entity pair, the ith entity from the first knowledge graph is represented by the characteristics of the gate mechanism;
Figure FDA00027556087700000711
is represented by A*(ii) a characteristic representation of a p-th neighbor of the k-th entity pair from the i-th entity of the first knowledge-graph; alpha is alpha1，k，iRepresenting the weight of the ith entity from the first knowledge-graph G1 in the kth pair of entities in a;
step 5.3: calculating weights of all entities of the first knowledge graph and the second knowledge graph according to neighbor entities of the entities and fusing neighbor information according to the calculated weights so as to collect the neighbor information of the entities, namely step 5.1 and step 5.2, and then obtaining a feature matrix after the neighbor information is collected by a graph neural network layer based on attention, namely:
Figure FDA0002755608770000081
Figure FDA0002755608770000082
wherein, crossAtt is the neural network layer of the attention-based map,
Figure FDA0002755608770000083
is the output feature representation of the first knowledge-graph through the cross-knowledge-graph aggregation layer;
Figure FDA0002755608770000084
is that the second knowledge-graph is represented by output features across knowledge-graph aggregation layers.
7. The method of cross-language entity alignment for a context alignment enhanced graph attention network of claim 1, wherein:
step 6, the characteristics of each entity containing the map fusion information and the multi-hop neighbor information are as follows:
Figure FDA0002755608770000085
Figure FDA0002755608770000086
wherein the content of the first and second substances,
Figure FDA0002755608770000087
is the output of the first knowledge-graph through the l-th CGAT layer; wherein
Figure FDA0002755608770000088
Is the output of the second knowledge-graph through the l-th CGAT layer;
Figure FDA0002755608770000089
features initialized for the first knowledge-graph
Figure FDA00027556087700000810
Figure FDA00027556087700000811
Features initialized for the second knowledge-graph
Figure FDA00027556087700000812
crossAtt is the attention-based graph neural network layer; cross-knowledge-map aggregation layer; by superposing L CGAT layers, the characteristics of the first knowledge graph and the second knowledge graph can be updated for L times, and finally the updated characteristics are output
Figure FDA0002755608770000091
And
Figure FDA0002755608770000092
calculating a target loss function on the target alignment seed set, wherein the formula is as follows:
Figure FDA0002755608770000093
wherein the content of the first and second substances,
Figure FDA0002755608770000094
the final characteristic representation of the representation entity,
Figure FDA0002755608770000095
represents the L1 distance between two entities, φ is a model parameter that can be trained, e-Refers to a negative entity corresponding to a certain entity, AobjIs data used to optimize model parameters;
and optimizing and updating model parameters phi by using an Adam algorithm, wherein the model parameters phi comprise parameters which are trainable across the knowledge graph aggregation layer and the attention-based graph neural network layer, and constructing a context-based alignment enhancement graph attention network model according to the optimized parameters phi.
CN202011201832.1A 2020-11-02 2020-11-02 Cross-language entity alignment method of context alignment enhanced graph attention network Active CN112417159B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011201832.1A CN112417159B (en) 2020-11-02 2020-11-02 Cross-language entity alignment method of context alignment enhanced graph attention network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011201832.1A CN112417159B (en) 2020-11-02 2020-11-02 Cross-language entity alignment method of context alignment enhanced graph attention network

Publications (2)

Publication Number Publication Date
CN112417159A true CN112417159A (en) 2021-02-26
CN112417159B CN112417159B (en) 2022-04-15

Family

ID=74827829

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011201832.1A Active CN112417159B (en) 2020-11-02 2020-11-02 Cross-language entity alignment method of context alignment enhanced graph attention network

Country Status (1)

Country Link
CN (1) CN112417159B (en)

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112905807A (en) * 2021-03-25 2021-06-04 北京邮电大学 Multi-source space-time knowledge fusion method
CN113157935A (en) * 2021-03-16 2021-07-23 中国科学技术大学 Graph neural network model and method for entity alignment based on relationship context
CN113392216A (en) * 2021-06-23 2021-09-14 武汉大学 Remote supervision relation extraction method and device based on consistency text enhancement
CN113407759A (en) * 2021-08-18 2021-09-17 中国人民解放军国防科技大学 Multi-modal entity alignment method based on adaptive feature fusion
CN113704495A (en) * 2021-08-30 2021-11-26 合肥智能语音创新发展有限公司 Entity alignment method and device, electronic equipment and storage medium
CN113761221A (en) * 2021-06-30 2021-12-07 中国人民解放军32801部队 Knowledge graph entity alignment method based on graph neural network
CN116257643A (en) * 2023-05-09 2023-06-13 鹏城实验室 Cross-language entity alignment method, device, equipment and readable storage medium

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110929041A (en) * 2019-11-20 2020-03-27 北京邮电大学 Entity alignment method and system based on layered attention mechanism
CN111723587A (en) * 2020-06-23 2020-09-29 桂林电子科技大学 Chinese-Thai entity alignment method oriented to cross-language knowledge graph

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110929041A (en) * 2019-11-20 2020-03-27 北京邮电大学 Entity alignment method and system based on layered attention mechanism
CN111723587A (en) * 2020-06-23 2020-09-29 桂林电子科技大学 Chinese-Thai entity alignment method oriented to cross-language knowledge graph

Cited By (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113157935A (en) * 2021-03-16 2021-07-23 中国科学技术大学 Graph neural network model and method for entity alignment based on relationship context
CN113157935B (en) * 2021-03-16 2024-02-27 中国科学技术大学 Entity alignment based on relation context and graph neural network system and method
CN112905807A (en) * 2021-03-25 2021-06-04 北京邮电大学 Multi-source space-time knowledge fusion method
CN113392216A (en) * 2021-06-23 2021-09-14 武汉大学 Remote supervision relation extraction method and device based on consistency text enhancement
CN113392216B (en) * 2021-06-23 2022-06-17 武汉大学 Remote supervision relation extraction method and device based on consistency text enhancement
CN113761221A (en) * 2021-06-30 2021-12-07 中国人民解放军32801部队 Knowledge graph entity alignment method based on graph neural network
CN113407759A (en) * 2021-08-18 2021-09-17 中国人民解放军国防科技大学 Multi-modal entity alignment method based on adaptive feature fusion
CN113407759B (en) * 2021-08-18 2021-11-30 中国人民解放军国防科技大学 Multi-modal entity alignment method based on adaptive feature fusion
CN113704495A (en) * 2021-08-30 2021-11-26 合肥智能语音创新发展有限公司 Entity alignment method and device, electronic equipment and storage medium
CN113704495B (en) * 2021-08-30 2024-05-28 合肥智能语音创新发展有限公司 Entity alignment method, device, electronic equipment and storage medium
CN116257643A (en) * 2023-05-09 2023-06-13 鹏城实验室 Cross-language entity alignment method, device, equipment and readable storage medium
CN116257643B (en) * 2023-05-09 2023-08-25 鹏城实验室 Cross-language entity alignment method, device, equipment and readable storage medium

Also Published As

Publication number Publication date
CN112417159B (en) 2022-04-15

Similar Documents

Publication Publication Date Title
CN112417159B (en) Cross-language entity alignment method of context alignment enhanced graph attention network
Han et al. A survey on metaheuristic optimization for random single-hidden layer feedforward neural network
CN108095716B (en) Electrocardiosignal detection method based on confidence rule base and deep neural network
JPWO2018051841A1 (en) Model learning apparatus, method thereof and program
CN112508334A (en) Personalized paper combining method and system integrating cognitive characteristics and test question text information
CN113140254B (en) Meta-learning drug-target interaction prediction system and prediction method
CN115131627B (en) Construction and training method of lightweight plant disease and pest target detection model
CN110322959B (en) Deep medical problem routing method and system based on knowledge
CN113190688A (en) Complex network link prediction method and system based on logical reasoning and graph convolution
CN114818703B (en) Multi-intention recognition method and system based on BERT language model and TextCNN model
CN114898121A (en) Concrete dam defect image description automatic generation method based on graph attention network
CN114491039B (en) Primitive learning few-sample text classification method based on gradient improvement
CN114398976A (en) Machine reading understanding method based on BERT and gate control type attention enhancement network
CN113344044A (en) Cross-species medical image classification method based on domain self-adaptation
CN114564596A (en) Cross-language knowledge graph link prediction method based on graph attention machine mechanism
US20230401390A1 (en) Automatic concrete dam defect image description generation method based on graph attention network
Tan Information Analysis of Advanced Mathematics Education‐Adaptive Algorithm Based on Big Data
CN107766895A (en) A kind of induction type is non-negative to project semi-supervised data classification method and system
CN114841318A (en) Intelligent contract vulnerability detection method based on cross-modal knowledge distillation
CN115101145A (en) Medicine virtual screening method based on adaptive meta-learning
Zang et al. [Retracted] The Application of 5G and Artificial Intelligence Technology in the Innovation and Reform of College English Education
CN112529188B (en) Knowledge distillation-based industrial process optimization decision model migration optimization method
CN110874392A (en) Text network information fusion embedding method based on deep bidirectional attention mechanism
CN117131933A (en) Multi-mode knowledge graph establishing method and application
CN117058394A (en) Zero sample semantic segmentation method

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant