Knowledge graph embedding method based on approximate implication co-occurrence graph convolutional network

Abstract

The invention provides a knowledge graph embedding method based on an approximate implication co-occurrence graph convolutional network, and the method comprises the steps: extracting co-occurrence features of entities and relationships in a knowledge graph through employing an approximate implication co-occurrence graph convolutional network model CoAE-GCN, and taking the co-occurrence features as input vectors; learning local features of the knowledge graph by using a graph convolutional network to obtain embedding of entities and relationships; mining an implicit relationship rule in the knowledge graph by using approximate implications; in combination with an attention mechanism, local structure features of nodes are extracted by using a graph convolutional network, and intermediate embedded representation is generated; the graph convolution network and the full connection layer are used for combining entity embedding relation embedding, and a binary cross entropy loss function is minimized through excitation function calculation; and optimizing a loss function, learning optimal parameter configuration, and realizing completion of the knowledge graph.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a knowledge graph embedding method based on an approximate implication co-occurrence graph convolutional network. Background Art

[0002] Knowledge graph is a form of structured data storage. The inherent incompleteness of real-world knowledge graphs has inspired the study of knowledge graph completion (KGC). Incomplete knowledge graphs limit the effectiveness of search and retrieval algorithms in information retrieval. Knowledge graph completion in the form of knowledge graph embedding (KGE) has become a proactive approach to address the challenge of predicting missing facts. KGE achieves accurate predictions in knowledge graphs by obtaining distributed low-dimensional representations of entities and relations.

[0003] Knowledge graphs (KGs) have been widely used in many artificial intelligence applications. The complex common sense in nature leads to the need for knowledge graphs to be constantly constructed, supplemented, and updated. In order to cope with the changes in incomplete knowledge triples, the knowledge graph completion task (KGC) was introduced, which aims to complete the missing triples by successfully predicting the missing relations. KGC is committed to solving the problem of incomplete knowledge triples and enhancing the comprehensiveness of knowledge graphs. However, current methods face three major challenges: (1) how to make full use of the co-occurrence features of entities and relations to enhance the expressiveness of the model; (2) how to extract the local structural features of KG; and (3) how to model relations with complex semantics.

[0004] In knowledge graph embedding, the first paradigm adopted is translation embedding, which actively represents entities and relations using a low-dimensional vector space. Translation embedding assumes that each relation transforms between a head and a tail entity within the embedding vector space. The TransE model is a notable model among translation embedding methods, which is both simple and effective. When considering a full triple (h, r, t), the embedding of the tail entity t is determined by adding the embedding of the head entity h to the relation r, h+r≈t. Although TransE is effective in a variety of scenarios, it encounters some limitations when dealing with complex relation types and patterns.

[0005] Researchers have proposed many translation embedding models to address the limitations of TransE and improve its performance from different perspectives. These models include TransH, TransD, and TransR. TransH is able to generate customized entity representations based on different relations. TransD uses dynamic matrices to project entities and relations onto a low-dimensional embedding space. TransR learns representations of entities and relations in different entity and relation spaces. Although the translation embedding model has efficient inference speed, it must effectively fuse the structural information in the entire knowledge graph.

[0006] Neural network embedding methods play a vital role in improving the generalization ability of models. The two main methods of neural network embedding are traditional neural networks and graph convolutional networks. Traditional neural network methods use complex architectures such as ConvE, ConvKB, and RSN to learn to express features. However, these models are limited to processing Euclidean data. Knowledge graphs store triples in the form of graph structures, and traditional neural networks cannot effectively process knowledge graphs.

[0007] Graph Convolutional Networks (GCNs) are able to effectively process non-Euclidean data, and these models have achieved success in KGC tasks. Among the popular GCN models in KGC are SACN, A2N, and KBGAT, which have provided inspiration for many other GCN-based methods. SACN aggregates neighbor information using a weighted graph convolutional network encoder and a backpropagation decoder. A2N learns entity embeddings through a dedicated attention mechanism applied to the query. Each neighbor computes a probability based on the query, and these probabilities serve as key features to represent the source entity. The KBGAT model introduces auxiliary edges to establish connections between neighbors and source entities, enabling information collection through an attention mechanism.

[0008] The above methods are effective in knowledge graph embedding. However, there are some problems with current embedding models. First, most models do not consider the co-occurrence characteristics of entities and relations. When a specific entity or relationship often appears with a central entity throughout the knowledge graph, these co-occurring elements are crucial to describe the central entity. Second, graph neural networks exhibit weak representation of local structural features. GCN needs to distinguish the local structural features of a single entity while effectively capturing the structural features of KG. It is worth noting that when vectorizing a single entity, the impact of different connected nodes on the resulting vector is different. Therefore, a method is needed to model the local structural features in the knowledge graph. Finally, a satisfactory solution to the complex semantic representation of relations has not yet been found. Relations in knowledge graphs usually have multiple semantics. For example, the "owner" and "owner_company" relations have similar semantic features, but are not exactly the same. Therefore, a model that can adjust and accommodate different semantic representations of relations is needed. Summary of the invention

[0009] According to the technical problems raised above, a knowledge graph embedding method based on approximate implication co-occurrence graph convolutional network is provided. The present invention mainly uses graph convolutional network CoAE-GCN to complete the knowledge graph, which can be applied to the large language model question-answering system based on the knowledge graph, and assists the large language model to capture the answer entity more accurately and quickly.

[0010] The technical means adopted by the present invention are as follows:

[0011] A knowledge graph embedding method based on approximate implication co-occurrence graph convolutional network, comprising:

[0012] S1. Use the approximate implication co-occurrence graph convolutional network model CoAE-GCN to extract the co-occurrence features of entities and relationships in the knowledge graph, and use the co-occurrence features as input vectors;

[0013] S2. Use graph convolutional networks to learn local features of knowledge graphs and obtain embeddings of entities and relationships.

[0014] S3, using approximate implication to mine implicit relational rules in knowledge graphs;

[0015] S4. Combined with the attention mechanism, the graph convolutional network is used to extract the local structural features of the nodes and generate intermediate embedding representations;

[0016] S5. Use graph convolutional networks and fully connected layers to merge entity embeddings and relation embeddings, and minimize the binary cross entropy loss function through activation function calculation.

[0017] S6. Repeat S4 and S5 to optimize the loss function, learn the optimal parameter configuration, and complete the knowledge graph.

[0018] Furthermore, the step S1 specifically includes:

[0019] The vector representations of entities e and relations r are randomly initialized, and the set of entities is represented as e i The vector representing the i-th entity; the CoAE-GCN model is used to extract the co-occurrence features of entities and relations e co , and the co-occurrence feature e is aggregated through the aggregation function co Fuse with the initial vector to form a co-occurrence embedding;

[0020] The set of entities that co-occur with entity e is denoted as coe(e), and the set of entities that co-occur with relation r is denoted as Cor(e). The co-occurrence features of entities and relations are aggregated using a fully connected network and weighted summation:

[0021] e co =Agg(concat[e 0 ,M(coe(e)),M(cor(e))])

[0022] Where M(·) is the function that calculates the mean of each row in the input vector, Agg(·) is the aggregation function, and concat[.] represents the concatenation function;

[0023] The co-occurrence characteristics of aggregated entities and relationships are:

[0024] r co=Agg(concat[r 0 ,M(coe(r))])

[0025] Among them, r co is the vector of relation r, r 0 Represents the initial vector of the relationship.

[0026] Furthermore, the step S2 specifically includes:

[0027] The graph convolutional network GCN is combined with the attention mechanism to calculate the incoming and outgoing neighbor features of the node respectively; the embedding of the triple is represented by c, and the head entity h co , relationship co , tail entity t co The calculation formula is as follows:

[0028]

[0029] Among them, c ijk Represents a triple, including the head entity h co , relationship co , tail entity t co The eigenvector of c (·) represents a fully connected layer, i represents the i-th entity in the entity set, as the head entity, j represents the j-th relationship in the relationship set, and k represents the k-th entity in the entity set, as the tail entity;

[0030] Through the multi-head attention mechanism, the importance weights of different neighbor nodes are dynamically adjusted, the features of the input neighbors and the output neighbors are combined, and the vector representation of the central node is generated; the attention of the input neighbors and the output neighbors is calculated:

[0031]

[0032] in, and is the attention weight, and is a fully connected layer, and softmax represents the activation function;

[0033] Fusion of input neighbor and output neighbor features to generate a comprehensive embedding representation of the node, calculate the input neighbor feature x in And output neighbor features x out :

[0034]

[0035] Among them, σ(·) represents the activation function, represents the input neighbor feature weights of the triplet, represents the weight of the output neighbor feature, represents the entity set pointing to entity i, Represents the entity set pointed to by entity i;

[0036] Use the multi-head attention mechanism to represent incoming neighbor features and outgoing neighbor features:

[0037]

[0038]

[0039] Among them, M is the number of heads of multi-head attention, represents the tail entity set of the triple containing the head entity i and the relation j; the input neighbor feature and the output neighbor feature vector are fused through the connection operation, and the fused feature x of entity e is:

[0040]

[0041] in, is the activation function, representing a linear operation, and bias is the process bias vector;

[0042] The graph convolutional network is used to calculate the fusion features x of all entities to obtain the embeddings of all entities and relations, and the head entity embedding h is retrieved based on the triples. G , tail entity embedding t G and relation embedding rG.

[0043] Furthermore, the step S3 specifically includes:

[0044] The approximate implication is introduced to automatically mine implicit relationship rules in the knowledge graph; the rule system includes a standard confidence level and a PCA confidence level, the standard confidence level and the PCA confidence level are used as scoring criteria for the rules, and the principal component analysis is used as a measure expressed in units of variables;

[0045] The approximate implication is expressed as (r p , r q ,λ),r p and r q Represents the involved relationship, λ represents the confidence level; the rule system is used to automatically obtain approximate implications from the knowledge graph, and the approximate implications set is expressed as:

[0046]

[0047] The rule system is used to obtain high-confidence approximate implications and encode them into the relational representation. p The characteristics of the relationship are integrated into r q Among the characteristics of a relationship:

[0048]

[0049] Among them, N represents the inclusion relationship r q The number of sets of approximate implication relations, represents the i-th approximate implication relation, λ i is the extreme part, indicating the confidence level associated with the ith approximate implication.

[0050] Furthermore, the step S5 specifically includes:

[0051] Entities are embedded into h using graph convolutional networks and fully connected layers. G ,t G and relation embedding q AE Merge, for a given triple (h, r, t), the score function is expressed as:

[0052]

[0053] Where f(·) represents a nonlinear function, vec(·) function is used to flatten a vector into a one-dimensional form, re(·) function is used to reshape a two-dimensional structure vector, * is a convolution operator, and W is a linear transformation parameter matrix;

[0054] Use the activation function σ(·) to calculate the score function, score Minimize the binary cross entropy loss function, expressed as:

[0055]

[0056] in, is the loss function and t is the tail entity.

[0057] Compared with the prior art, the present invention has the following advantages:

[0058] The knowledge graph embedding method based on the approximate implication co-occurrence graph convolutional network provided by the present invention shows significant advantages in the knowledge graph completion task. By introducing the co-occurrence features of entities and relationships, the model can more comprehensively capture the semantic associations between entities and relationships, thereby improving the expressiveness and prediction accuracy of the model. Through the approximate implication technology, the model can better simulate the complex semantics of the relationship, further improve the quality of the relationship representation, and make the model perform better when processing complex relationship patterns such as many-to-one and one-to-many. Experimental results show that the present invention has achieved excellent performance on multiple benchmark datasets, especially the MRR indicators on the FB15k-237 and WN18RR datasets reached 0.379 and 0.482 respectively, which is significantly better than the existing methods. By fusing multiple features, the model effectively reduces the impact of noise, improves the stability and robustness of the model, and is suitable for the completion task of large-scale knowledge graphs.

[0059] Based on the above reasons, the present invention can be widely promoted in fields such as artificial intelligence. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0061] Figure 1 This is a flow chart of the knowledge graph embedding method based on the approximate implication co-occurrence graph convolutional network in the present invention.

[0062] Figure 2 Schematic diagram of the model architecture in an embodiment of the present invention.

[0063] Figure 3 This is the process of learning co-occurrence features for the model in the embodiment of the present invention.

[0064] Figure 4 The figure compares the knowledge graph embedding effects of the CoAE-GCN model in an embodiment of the present invention and the existing model for the datasets FB15k-237 and WN18RR.

[0065] Figure 5 The figure compares the knowledge graph embedding effects of the CoAE-GCN model in the embodiment of the present invention and the existing model for the datasets FB15k and WN18. DETAILED DESCRIPTION

[0066] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0067] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0068] like Figure 1 As shown, the present invention provides a knowledge graph embedding method based on an approximate implication co-occurrence graph convolutional network, comprising:

[0069] S1. Use the approximate implication co-occurrence graph convolutional network model CoAE-GCN to extract the co-occurrence features of entities and relationships in the knowledge graph, and use the co-occurrence features as input vectors;

[0070] In specific implementation, as a preferred embodiment of the present invention, the step S1 specifically includes:

[0071] The vector representations of entities e and relations r are randomly initialized, and the set of entities is represented as e i represents the vector of the i-th entity; Figure 2 As shown in Figure 2, the CoAE-GCN model is used to extract the co-occurrence features of entities and relations. co , and the co-occurrence feature e is aggregated through the aggregation function co Fuse with the initial vector to form a co-occurrence embedding;

[0072] The set of entities that co-occur with entity e is denoted as coe(e), and the set of entities that co-occur with relation r is denoted as cor(e). The co-occurrence features of entities and relations are aggregated using a fully connected network and weighted summation:

[0073] e co =Agg(concat[e 0 ,M(coe(e)),M(cor(e))])

[0074] Where M(·) is the function that calculates the mean of each row in the input vector, Agg(·) is the aggregation function, and concat[.] represents the concatenation function;

[0075] Co-occurrence embedding provides semantic information and enhances the expressiveness of the model by capturing the co-occurrence frequency of entities and relations. The co-occurrence features of aggregated entities and relations are:

[0076] r co =Agg(concat[r 0 ,M(coe(r))])

[0077] Among them, r co is the vector of relation r, r 0 Represents the initial vector of the relationship.

[0078] Step S1 is performed on all entities and relations, and the co-occurrence features of entities and relations are fused using a fully connected network. The initial dimension of entities and relations is 100 to generate an input vector.

[0079] S2, such as Figure 3 As shown in the figure, the graph convolutional network is used to learn the local features of the knowledge graph and obtain the embedding of entities and relationships;

[0080] In specific implementation, as a preferred embodiment of the present invention, step S2 specifically includes:

[0081] The graph convolutional network GCN is combined with the attention mechanism to calculate the incoming and outgoing neighbor features of the node respectively; the embedding of the triple is represented by c, and the head entity h co , relationship co , tail entity t co The calculation formula is as follows:

[0082]

[0083] Among them, c ijk Represents a triple, including the head entity h co , relationship co , tail entity t co The eigenvector of c (·) represents a fully connected layer, i represents the i-th entity in the entity set, as the head entity, j represents the j-th relationship in the relationship set, and k represents the k-th entity in the entity set, as the tail entity;

[0084] Through the multi-head attention mechanism, the importance weights of different neighbor nodes are dynamically adjusted to effectively extract the local structural features of the node, combine the features of the input neighbors and the output neighbors, and generate a vector representation of the central node; calculate the attention of the input neighbors and the output neighbors:

[0085]

[0086] in, and is the attention weight, and is a fully connected layer, and softmax represents the activation function;

[0087] Fusion of input neighbor and output neighbor features generates a comprehensive embedding representation of the node, further improving the structural modeling capability of the model and calculating the input neighbor feature x in And output neighbor features x out :

[0088]

[0089] Among them, σ(·) represents the activation function, represents the input neighbor feature weights of the triplet, represents the weight of the output neighbor feature, represents the entity set pointing to entity i, Represents the entity set pointed to by entity i;

[0090] Use the multi-head attention mechanism to represent incoming neighbor features and outgoing neighbor features:

[0091]

[0092] Among them, M is the number of heads of multi-head attention, Represents the tail entity set of triples containing head entity i and relation j;

[0093] The input neighbor feature and the output neighbor feature vector are fused through the connection operation, and the fused feature x of entity e is:

[0094]

[0095] in, is the activation function, representing a linear operation, and bias is the process bias vector;

[0096] The graph convolutional network is used to calculate the fusion features x of all entities to obtain the embeddings of all entities and relations, and the head entity embedding h is retrieved based on the triples. G , tail entity embedding t G and relation embedding G .

[0097] S3, using approximate implication to mine implicit relational rules in knowledge graphs;

[0098] When implemented, the representation power of model relations is enhanced by using approximate entailment. Approximate entailment refers to a relationship between two aspects where one aspect can hint at the possibility of the other aspect to some extent. Approximate entailment captures associations between different pieces of information and is less strict than exact logical entailment. In knowledge graphs, approximate entailment involves identifying patterns that suggest certain relations based on other relations.

[0099] In specific implementation, as a preferred embodiment of the present invention, step S3 specifically includes:

[0100] The approximate implication is introduced to automatically mine implicit relationship rules in the knowledge graph; the rule system includes a standard confidence level and a PCA confidence level, the standard confidence level and the PCA confidence level are used as scoring criteria for the rules, and the principal component analysis is used as a measure expressed in units of variables;

[0101] The approximate implication is expressed as (r p , r q ,λ),r p and r q Denotes the involved relations, λ denotes the confidence level; in implementation, a threshold of 0.8 or higher is established for the PCA confidence level associated with approximate implications. The approximate implications are automatically obtained from the knowledge graph using a rule system, and the approximate implications set is expressed as:

[0102]

[0103] The rule system is used to obtain high-confidence approximate implications and encode them into the relational representation to enhance the semantic representation of the relation. p The characteristics of the relationship are integrated into r q Among the characteristics of a relationship:

[0104]

[0105] Among them, N represents the inclusion relationship r q The number of sets of approximate implication relations, represents the i-th approximate implication relation, λ i is the extreme part, indicating the confidence level associated with the ith approximate implication.

[0106] S4, repeat steps S2 and S3, with the number of hidden units being 200, the number of GCN layers being 1, and the dropout rate being 0.1. Combined with the attention mechanism, the graph convolutional network is used to extract the local structural features of the nodes and generate an intermediate embedding representation;

[0107] S5. Use graph convolutional networks and fully connected layers to merge entity embeddings and relation embeddings, and minimize the binary cross entropy loss function through activation function calculation. In implementation, use graph convolutional networks and approximate implication techniques to optimize relation representations and generate the final embedding representation to obtain the head entity embedding h G , tail entity embedding t G and relation embedding G The ConvE model is used with a dropout rate of 0.3, a convolution kernel size of 7, and a filter number of 200.

[0108] In specific implementation, as a preferred embodiment of the present invention, step S5 specifically includes:

[0109] Entities are embedded into h using graph convolutional networks and fully connected layers. G ,t G and relation embedding q AE Merge, for a given triple (h, r, t), the score function is expressed as:

[0110]

[0111] Among them, f(·) represents a nonlinear function, vec(·) function is used to flatten the vector into a one-dimensional form, re(·) function is used to reshape the two-dimensional structure vector, * is the convolution operator, W is the linear transformation parameter matrix,

[0112] Use the activation function σ(·) to calculate the score function, score Minimize the binary cross entropy loss function, expressed as:

[0113]

[0114] in, is the loss function and t is the tail entity.

[0115] S6. Repeat S4 and S5 to optimize the loss function, learn the optimal parameter configuration, and complete the knowledge graph.

[0116] Example

[0117] like Figure 1 As shown in FIG, the present invention provides a knowledge graph embedding method based on an approximate implication co-occurrence graph convolutional network. The experimental results of this embodiment on four real data sets (FB15k, FB15-237, WN18, WN18RR) are shown in FIG. Figure 4 and Figure 5 As shown in the figure, the CoAE-GCN model performs well overall in all indicators, especially when dealing with complex relationship patterns, which verifies the effectiveness and superiority of the CoAE-GCN model.

[0118] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0119] In the above embodiments of the present invention, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0120] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of the units can be a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0121] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0122] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0123] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk and other media that can store program codes.

[0124] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A knowledge graph embedding method based on approximate implication co-occurrence graph convolutional network, characterized in that: include: S1. Use the approximate implication co-occurrence graph convolutional network model CoAE-GCN to extract the co-occurrence features of entities and relationships in the knowledge graph, and use the co-occurrence features as input vectors; S2. Use graph convolutional networks to learn local features of knowledge graphs and obtain embeddings of entities and relationships. S3, using approximate implication to mine implicit relational rules in knowledge graphs; S4. Combined with the attention mechanism, the graph convolutional network is used to extract the local structural features of the nodes and generate intermediate embedding representations; S5. Use graph convolutional networks and fully connected layers to merge entity embeddings and relation embeddings, and minimize the binary cross entropy loss function through activation function calculation. S6. Repeat S4 and S5 to optimize the loss function, learn the optimal parameter configuration, and complete the knowledge graph.

2. The knowledge graph embedding method based on approximate implication co-occurrence graph convolutional network according to claim 1 is characterized in that: The step S1 specifically includes: The vector representations of entities e and relations r are randomly initialized, and the set of entities is represented as e i The vector representing the i-th entity; the CoAE-GCN model is used to extract the co-occurrence features of entities and relations e co , and the co-occurrence feature e is aggregated through the aggregation function co Fuse with the initial vector to form a co-occurrence embedding; The set of entities that co-occur with entity e is denoted as coe(e), and the set of entities that co-occur with relation r is denoted as cor(e). The co-occurrence features of entities and relations are aggregated using a fully connected network and weighted summation: e co =Agg(concat[e 0 ,M(coe(e)),M(cor(e))]) Where M(·) is the function that calculates the mean of each row in the input vector, Agg(·) is the aggregation function, and concat[.] represents the concatenation function; The co-occurrence characteristics of aggregated entities and relationships are: r co =Agg(concat[r 0 ,M(coe(r))]) Among them, r co is the vector of relation r, r 0 Represents the initial vector of the relationship.

3. The knowledge graph embedding method based on approximate implication co-occurrence graph convolutional network according to claim 1 is characterized in that: The step S2 specifically includes: The graph convolutional network GCN is combined with the attention mechanism to calculate the incoming and outgoing neighbor features of the node respectively; the embedding of the triple is represented by c, and the head entity h co , relationship co , tail entity t co The calculation formula is as follows: Among them, c ijk Represents a triple, including the head entity h co , relationship co , tail entity t co The eigenvector of c (·) represents a fully connected layer, i represents the i-th entity in the entity set, as the head entity, j represents the j-th relationship in the relationship set, and k represents the k-th entity in the entity set, as the tail entity; Through the multi-head attention mechanism, the importance weights of different neighbor nodes are dynamically adjusted, the features of the input neighbors and the output neighbors are combined, and the vector representation of the central node is generated; the attention of the input neighbors and the output neighbors is calculated: in, and is the attention weight, and is a fully connected layer, and softmax represents the activation function; Fusion of input neighbor and output neighbor features to generate a comprehensive embedding representation of the node, calculate the input neighbor feature x in And output neighbor features x out : Among them, σ(·) represents the activation function, represents the input neighbor feature weights of the triplet, represents the weight of the output neighbor feature, represents the entity set pointing to entity i, Represents the entity set pointed to by entity i; Use the multi-head attention mechanism to represent incoming neighbor features and outgoing neighbor features: Among them, M is the number of heads of multi-head attention, Represents the set of tail entities of triples containing head entity i and relation j; The input neighbor feature and the output neighbor feature vector are fused through the connection operation, and the fused feature x of entity e is: in, is the activation function, representing a linear operation, and bias is the process bias vector; The graph convolutional network is used to calculate the fusion features x of all entities to obtain the embeddings of all entities and relations, and the head entity embedding h is retrieved based on the triples. G , tail entity embedding t G and relation embedding rG.

4. The knowledge graph embedding method based on approximate implication co-occurrence graph convolutional network according to claim 1 is characterized in that: The step S3 specifically includes: The approximate implication is introduced to automatically mine implicit relationship rules in the knowledge graph; the rule system includes a standard confidence level and a PCA confidence level, the standard confidence level and the PCA confidence level are used as scoring criteria for the rules, and the principal component analysis is used as a measure expressed in units of variables; The approximate implication is expressed as (r p , r q ,λ),r p and r q Represents the involved relationship, λ represents the confidence level; the rule system is used to automatically obtain approximate implications from the knowledge graph, and the approximate implication set AE is expressed as: The rule system is used to obtain high-confidence approximate implications and encode them into the relational representation. p The characteristics of the relationship are integrated into r q Among the characteristics of a relationship: Among them, N represents the inclusion relationship r q The number of sets of approximate implication relations, represents the i-th approximate implication relation, λ i is the extreme part, indicating the confidence level associated with the ith approximate implication.

5. The knowledge graph embedding method based on approximate implication co-occurrence graph convolutional network according to claim 1 is characterized in that: The step S5 specifically includes: Entities are embedded into h using graph convolutional networks and fully connected layers. G ,t G and relation embedding q AE Merge, for a given triple (h, r, t), the score function is expressed as: Where f(·) represents a nonlinear function, vec(·) function is used to flatten a vector into a one-dimensional form, re(·) function is used to reshape a two-dimensional structure vector, * is a convolution operator, and W is a linear transformation parameter matrix; Use the activation function σ(·) to calculate the score function, score Minimize the binary cross entropy loss function, expressed as: in, is the loss function and t is the tail entity.

Images