Cross-language entity alignment method and system based on relational semantic enhancement

Abstract

The invention relates to the field of natural language processing, and discloses a cross-language entity alignment method and system based on relational semantic enhancement, and the method comprises the following steps: constructing entity initial embedding and relational embedding representation of a cross-language knowledge graph; a key neighbor screening stage; based on the entity similarity matrix and the relationship similarity matrix, screening key neighbor entities with structure matching and semantic matching in combination with a preset semantic relevancy threshold; embedding and updating the neighbor enhanced entity; a graph attention network is adopted to fuse semantic features of a target entity and key neighbors of the target entity, and context-aware entity embedding is generated; semantic consistency alignment propagation: jointly optimizing first-order entity feature alignment and second-order relation structure alignment through a shared permutation matrix; and dynamic seed expansion iteration: dynamically identifying high-confidence misaligned entity pairs according to Manhattan distance embedded by the entities, adding the entity pairs of which the distance is lower than a threshold value into a seed set, and iteratively executing the previous steps until convergence.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of natural language processing, in particular to a cross-language entity alignment method and system based on relationship semantic enhancement. BACKGROUND

[0002] Traditional entity alignment methods, especially those based on vector space mapping, often ignore the relationship information between entities, resulting in inaccurate alignment results. To this end, this paper proposes an entity alignment method based on entity and relationship interaction, which effectively models the semantic association between entities and relationships through knowledge graph embedding technology, thereby improving the accuracy and consistency of entity alignment between different data sources.

[0003] In current cross-language Knowledge Graph (KG) alignment research, the accuracy of entity embedding directly affects the alignment performance.

[0004] However, the same entity across languages has structural heterogeneity, and its neighbor structure often differs significantly due to language or cultural differences, thereby introducing noise into embedding and reducing alignment performance. Existing methods mainly rely on nested similarity between entities to select "important neighbors", but ignore the role of relationship semantics between entities and their neighbors, which cannot fully capture the real semantic relationship between entities, resulting in the selected neighbors being unrepresentative and ultimately affecting alignment performance. SUMMARY

[0005] The purpose of the present application is to provide a cross-language entity alignment method and system based on relationship semantic enhancement, which solves the above technical problems.

[0006] The purpose of the present application can be achieved by the following technical solutions: A cross-language entity alignment method based on relationship semantic enhancement, comprising the following steps: S1, constructing entity initial embedding and relationship embedding representation of cross-language knowledge graph; S2, key neighbor selection stage; Based on the entity similarity matrix and the relationship similarity matrix, combined with the preset semantic correlation threshold, the key neighbor entities with structural matching and semantic matching are selected; S3, neighbor-enhanced entity embedding update; The semantic features of the target entity and its key neighbors are fused using a graph attention network to generate context-aware entity embedding; S4, semantic consistency alignment propagation: First-order entity feature alignment and second-order relationship structure alignment are jointly optimized through a shared permutation matrix; S5, dynamic seed expansion iteration: The high-confidence unaligned entity pairs are dynamically identified according to the Manhattan distance of the entity embedding, the entity pairs with a distance less than a threshold are added to a seed set, and steps S2 to S4 are iteratively performed until convergence.

[0007] As a further technical solution, the key neighbor screening in step S2 specifically includes: The entity similarity matrix is calculated by the L1 distance of the entity embedding vector; The relationship similarity matrix is calculated by the semantic correlation of the relationship embedding vector; Only the neighbor entities with both entity similarity and relationship similarity exceeding a preset threshold are retained.

[0008] As a further technical solution, the second-order relationship structure alignment in step S4 fuses the spatial structure information and semantic association information of the connection relationship, and the specific implementation manner is: By minimizing the relationship structure matrix difference Optimizing the shared permutation matrix P; wherein , The correlation matrix fusing the entity spatial structure and the relationship semantics; The shared permutation matrix P is constrained to satisfy row normalization, column normalization, and element binary value {0, 1}.

[0009] As a further technical solution, the construction manner of the relationship structure matrix includes: If there is a direct relationship connection between two entities, the relationship embedding is fused into the entity spatial structure matrix in a dot product manner; If there is no direct relationship, the correlation matrix is calculated based on only the entity spatial structure information.

[0010] As a further technical solution, the relationship embedding generation manner in step S1 includes: The relationship vector is constructed by splicing the average embedding of the head entity and the average embedding of the tail entity ; wherein represents a connection operation, , respectively represent the average embedding of all different head entities and tail entities of the relationship ; Introducing a TransE-based triple regularization loss Optimizing the relationship embedding.

[0011] As a further technical solution, the dynamic seed expansion of step S5 specifically includes: In each iteration, the unaligned entities of the source knowledge graph are traversed, and the candidate entity with the smallest Manhattan distance in the target graph is found; If the distance is less than a threshold If the entity pair is not in the seed alignment set, then the entity pair is added to the seed alignment set and the embedding representation is updated.

[0012] As a further technical solution, the generation manner of the entity embedding in the step S1 is: The entity representation is updated by using the multi-layer propagation rule of the graph convolutional neural network: ; Wherein, is the adjacency matrix of the graph plus the unit matrix , is the entity feature vector of the first layer, is the weight matrix of the first layer, represents an activation function, is the degree matrix of ; The embedding is optimized based on a boundary loss function: ; Wherein, represents a loss function of the entity embedding, represents a boundary parameter greater than 0, represents an initial seed alignment set of the entity, represents a set of negative alignment sets generated by nearest neighbor sampling, represents the Manhattan distance of the entity embedding, represents the Manhattan distance of the negative sample entity pair.

[0013] As a further technical solution, a semi-supervised iterative strategy is used to train the model: The entity alignment loss and the relationship regularization loss are combined, and the influence of the two is balanced by a weight coefficient; The entity distance matrix and the relationship distance matrix are updated synchronously in the iteration process.

[0014] A cross-language entity alignment system based on relationship semantic enhancement, which is used to execute the cross-language entity alignment method based on relationship semantic enhancement.

[0015] The beneficial effects of the present application are: The present application captures the neighbor information of the entity by the first-order entity feature alignment, and mines the deep semantic relationship between the entities by combining the second-order relationship structure alignment, so that the equivalent entities in different knowledge graphs can be accurately matched. In addition, the present application introduces a shared permutation matrix, learns the neighbor feature similarity of the entity and the corresponding relationship of the relationship space structure by optimizing the matrix, so that the entity and the relationship can realize positive interaction, and the stability and accuracy of the alignment are improved. BRIEF DESCRIPTION OF DRAWINGS

[0016] The application will be further described below with reference to the drawings.

[0017] Figure 1 Diagram for neighbor heterogeneity problem in cross-language entity alignment; Figure 2 Diagram for the framework of the application; Figure 3 Flow for relationship structure alignment. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the application.

[0019] Referring to Figures 1-3 The application is a cross-language entity alignment method based on relationship semantic enhancement, and the method comprises the following steps: S1, constructing entity initial embedding and relationship embedding representation of a cross-language knowledge graph; S2, key neighbor screening stage; Based on the entity similarity matrix and the relationship similarity matrix, the key neighbor entities of structural matching and semantic matching are screened in combination with a preset semantic correlation threshold value; S3, neighbor-enhanced entity embedding update; The semantic features of the target entity and its key neighbors are fused by using a graph attention network to generate context-aware entity embedding; S4, semantic consistency alignment propagation: First-order entity feature alignment and second-order relationship structure alignment are jointly optimized through a shared permutation matrix; S5, dynamic seed expansion iteration: High-confidence unaligned entity pairs are dynamically identified according to the Manhattan distance of entity embedding, entity pairs with a distance lower than a threshold value are added to a seed set, and steps S2 to S4 are iteratively executed until convergence.

[0020] In this embodiment, firstly, entities and relations in two knowledge graphs are embedded using GCNs with TransE regularizers to obtain entity and relation embedding vectors, respectively. Secondly, the entity embedding vectors and relation embedding vectors are associated to construct a graph embedding representation matrix. Then, a first-order entity feature alignment module is constructed based on the entity's neighbor information to capture the entity's contextual information. Simultaneously, a second-order relation structure alignment module is further constructed using the spatial structure and semantic association information of the connection relations, with the aim of mining deep semantic relationships between entities. Next, a shared permutation matrix P is learned by jointly using the first-order entity feature alignment and the second-order relation structure alignment to effectively promote positive interactions between entities and relations, thereby improving alignment performance. Finally, the positive interaction between the entity alignment task and the relation alignment task is jointly learned in a semi-supervised iterative manner. In each iteration, the embedding representations of entities and relations are updated using the shared permutation matrix, and the new entity and relation alignment is determined by judging the similarity of the embedding vectors.

[0021] Entity embedding: Entity embedding takes into account the structural information of the two knowledge graphs by aligning them with a seed set of entities given two knowledge graphs and using GCNs to embed all entities in the two knowledge graphs into the same latent space.

[0022] Specifically, firstly, entity representations are initialized using pre-trained word embeddings to provide necessary semantic information for the entities. This initialization method not only ensures the similarity of all entities on a common basis but also lays the foundation for subsequent fine-grained processing in the GCN layer. Secondly, GCNs capture the contextual information of two KGs, effectively embedding all entities into the shared latent space. The core of GCNs lies in its multi-layer propagation rule, which continuously updates the representation of each entity by aggregating the features of neighboring entities. In the above process, the representation of each entity is adjusted and optimized based on the information of its neighbors, thereby achieving a more accurate and comprehensive entity representation. In GCN, the update rule of a single layer can be expressed as: Update entity representations using multi-layer propagation rules of graph convolutional neural networks: ; in, It is the adjacency matrix of the graph Add the identity matrix , For the first The entity feature vector of the layer, For the first The weight matrix of the layer, This represents the activation function. yes The degree matrix; During the initialization phase, is the entity embedding vector obtained from the pre-trained word embedding; the activation function is the ReLU function, i.e. The neighbor aggregation strategy in GCNs can effectively capture the topological structure of the graph and the entity feature information. The entity final representation is defined as where denotes the dimension of the entity embedding, denotes the number of entities; In order to embed the entities in two into the same latent space, seed alignment is used as training samples. At the same time, a boundary-based loss function is designed to encourage clear separation between different classes by maximizing the interval between classes. Optimize the embedding based on the boundary loss function: ; where denotes the loss function of the entity embedding, denotes the boundary parameter greater than 0, denotes the initial set of seed alignments of entities, denotes a set of negative alignments generated by nearest neighbor sampling, denotes the Manhattan distance of the entity embedding, denotes the Manhattan distance of the negative sample entity pair.

[0023] is the L1 norm of the vector, The lower the value of , the stronger the alignment between the entities ; the principle of this loss function is that the distance between aligned entity pairs should be close to zero; on the contrary, the distance between negative samples should be maximized as much as possible.

[0024] Relation embedding: In , facts are encoded as triples, i.e. , , ), where the meaning of the relationship is associated with the two entities connected to it; in order to utilize the semantic association information of the connected entities, the embeddings of the head entity and the tail entity learned from GCNs are used to represent the relationship in the knowledge graph: ; , denotes the relationship , denotes the connection operation, , denote the average embedding of all different head entities and tail entities of the relationship , respectively. Introducing TransE-based triple regularization loss Optimizing relation embeddings. The goal is to promote sparsity of the model through L1 regularization, which helps the model to focus on the most important features and thus reduces the risk of overfitting.

[0025] ; where, and represent the set of triples of two given knowledge graphs and respectively; denotes the transformation matrix from the latent space of relations to the latent space of entities, which is the model parameter to be learned.

[0026] To jointly learn the embeddings of entities and relations, the objective function is designed as: ; is the weight coefficient, which is used to balance the entity alignment loss and the regularization loss while considering the influence of relation embeddings; the main goal is to minimize the objective function after the pre-training of entity embeddings is completed.

[0027] Entity feature alignment: By comparing the similarity of the first-order neighbor feature information of entities, the correspondence between entities in different knowledge graphs is maximized; the first-order feature information of each entity is obtained from the entity embedding module, which contains neighbor and context information related to the entity. In the context of aligning the first-order entity features, if and , which is the th entity in , and , which is the th entity in , show greater similarity than other entity pairs, then the possibility of a strong correspondence between entities and is increased; to achieve alignment, an optimization objective is used to minimize the difference between entity features. The expression of this objective is shown in the formula: , , ; where, denotes the embedding representation of all entities in , denotes the embedding representation of all entities in , and denote and the number of entities in is denoted as Frobenius norm, which aims to measure the overall size of a matrix, mainly by calculating the square root of the sum of the squares of the absolute values of each element in the matrix; in the above formula, denotes a shared permutation matrix, which aims to incorporate permutation-based operations into the alignment task; is a model training parameter, which facilitates the exploration of multiple alignment possibilities and captures various alignment patterns, which is essential for establishing different correspondence between entities in denotes that when the matrix is multiplied by a vector consisting entirely of 1s, the resulting vector also consists entirely of 1s, indicating that the matrix the sum of the elements in each row is equal to 1, which ensures that performs row normalization, i.e., the sum of each row is 1, thus maintaining the integrity of the row distribution; Similarly, equation confirms that the sum of the elements in each column of the matrix is equal to 1, ensuring column normalization; the above formula ensures that the matrix is simultaneously normalized in rows and columns; at the same time, a constraint is imposed on the shared permutation matrix , which can only contain 0 and 1, thus ensuring a one-to-one correspondence between entities in the two knowledge graphs, thereby improving alignment performance.

[0028] Relationship structure alignment: On the basis of first-order entity feature alignment, second-order relationship structure alignment is introduced; its goal is to ensure that the connection between entities and their neighbors remains comparable, mainly considering the spatial structure of the connection and its semantic association information to improve alignment performance in different knowledge graphs.

[0029] First, find the one-to-one similarity in the spatial structure of the relationship by optimizing the shared permutation matrix to help align entities.

[0030] For example, as Figure 3 , there are two different knowledge graphs; if and are equivalent entities, and are also equivalent entities, then the spatial structure of the relationship connecting the entities ( to , to ) should also be similar.

[0031] Therefore, when knowing and When they are equivalent, the connection relationship is found by optimizing the shared permutation matrix. arrive , arrive Spatial structural similarity information can be used to determine arrive They are similar.

[0032] Secondly, this approach combines semantic association information of connection relationships to enhance alignment; The inherent semantic associations between the relationships in two knowledge graphs can be better facilitated by one-to-one matching to promote entity alignment.

[0033] By combining the semantic association information of the relation embedding vector with the spatial structure information, a matrix containing the spatial structure and semantic association information of the connection relationship is formed.

[0034] For example, such as Figure 3 As shown, Existence Relationship connect and ,as well as Existence Relationship connect and ,and Equivalent to When found through the shared permutation matrix and When similar, it can be inferred and They are similar, thus improving entity alignment performance. Therefore, by minimizing the function shown in the formula, alignment of second-order relation structures can be achieved: ; , , ; in, , , and They represent and A matrix containing spatial structure and semantic association information of connection relationships. and They are and The embedded representations of all relations obtained from the relation embedding module contain semantic association information of the relations. and They described respectively and The internal spatial structure of two entity sets; The specific design method is as follows: If there is a connection relationship between and , the relationship embedding vector is embedded into the spatial structure information between entities in the form of dot product, so as to better preserve the spatial structure and semantic association information of the relationship. If and there is no connection relationship in the form of triple, the spatial structure information between entities is considered for alignment, such as Figure 3 between and , and the same operation is performed in to improve the alignment performance; by minimizing the formula, the relationship structure alignment converges to high consistency; the shared permutation matrix plays a crucial role in capturing the interaction between entity features and relationship structures, thereby enhancing the comprehensiveness of the alignment method.

[0035] Joint interaction learning optimization goal: The optimization goal of joint interaction learning is to guide the finding of one-to-one corresponding similar entity neighbors and relationship spatial structures in different knowledge graph subgraphs by training the shared permutation matrix P through the combination of entity feature alignment and relationship structure alignment modules, so as to minimize the optimization goal of joint interaction learning; as shown in the formula: ; , , ; Among them, is a balance hyperparameter, and the constraints of the optimization problem are and , that is, the sum of the elements of each row and each column of the permutation matrix is equal to 1; in addition, the value constraint of P is , that is, The elements of can only take 0 or 1; by optimizing the shared permutation matrix , the interaction between entity features and relationship structures can be effectively promoted; the above comprehensive alignment method considers the complexity between entities and relationships in the knowledge graph.

[0036] Semi-supervised iterative strategy and implementation details: In order to train an effective EA model, the objective function of embedding learning in the formula and the objective function of joint interaction learning are combined, and the final optimization goal is represented as: ; wherein, is the weight coefficient of the joint interactive learning objective function, by changing the value of , the influence of joint interactive learning on the entire optimization process can be controlled; during the training process, different objectives may guide the model to optimize in different directions, and the parameter can ensure that the above two objectives can promote each other; The ultimate goal of is to learn the interaction between entity features and relationship structures, thereby realizing a comprehensive alignment method. To this end, a semi-supervised framework is proposed, which uses the positive interaction between entity and relationship alignment tasks to iteratively enhance alignment, thereby improving the robustness and accuracy of the model.

[0037] The distance matrix of entity pairs from is represented as: The distance matrix of relationship pairs from to is represented as: .

[0038] Using the learned entity and relationship embeddings, and the initialization definition is as shown in the formula: ; ; Within a specified number of iterations , the model refines the alignment process by using the corresponding embeddings, thereby realizing the alignment of entities and relationships between two different knowledge graphs; the above iteration process is guided by the distance thresholds of entities and relationships and , in each iteration, the algorithm traverses each entity and relationship in , and finds the closest corresponding item in ; if the distance between the aligned entity and the aligned relationship is lower than the respective threshold, the algorithm identifies a potential alignment; then, the entity alignment set and the relationship alignment set are updated accordingly; subsequently, the model updates the distance matrix of entities and the distance matrix of relationships by optimizing the first-order entity feature alignment and the second-order relationship structure ; the above iterative alignment and distance matrix refinement process continues until the maximum number of iterations ; finally, the algorithm outputs an updated distance matrix and and an updated entity alignment set and a relation alignment set .

[0039] The present application aims to improve the accuracy and robustness of cross-language entity alignment. The core idea is to identify and utilize key neighbor information during entity alignment, integrate entity and relationship semantic information, and optimize entity embedding. The main steps are as follows: First, data preprocessing and initial entity representation construction; Extracting multi-modal information such as structure, attribute, and text description of entities in cross-language knowledge graph; Using existing embedding methods (such as TransE, GCN) to generate initial representations for entities; According to the existing aligned entities, construct a seed set to provide initial supervision signals for subsequent alignment training; Second, entity-relation similarity matrix calculation and neighbor selection; Calculate entity similarity matrix and relation similarity matrix respectively to measure the semantic similarity between cross-language entities and their relationships; Set a threshold for the similarity matrix to ensure that only one optimal neighbor entity is selected for each row; Eliminate low-quality and semantically misleading neighbors and only keep key neighbors that are highly matched in structure and semantics; Third, neighbor-enhanced embedding update mechanism design; Based on the graph attention network (GAT), design an embedding update module and introduce high-quality neighbor information after screening; Fuse the entity's own representation with its key neighbor representation, recalculate the entity embedding, and strengthen the semantic context; Fourth, semantic alignment propagation mechanism; Optimize the embedding space through the triple embedding criterion ( ) for the initial aligned entity pair; Use the semantic consistency between the aligned entity pairs to guide the update of the unaligned entity embedding; Achieve the linkage optimization of the entire embedding space and improve the global alignment effect; Fifth, dynamic seed expansion mechanism; Use Manhattan distance as the alignment criterion. If the distance between unaligned entity pairs is less than the set threshold, it is determined as a new aligned entity; Add the new aligned entity that meets the conditions to the seed set and optimize the embedding and alignment boundary in a loop; Repeat the training for several rounds until convergence, and finally complete large-scale cross-language entity alignment.

[0040] The above detailed description has shown, by way of example, an embodiment of the application. It is specifically contemplated that the application is not limited to the embodiments described herein, but rather the scope of the application is defined by the claims.

Claims

1. A cross-language entity alignment method based on relational semantic enhancement, characterized in that, Includes the following steps: S1. Construct initial entity embeddings and relation embedding representations for cross-language knowledge graphs; S2, Key Neighbor Screening Stage; Based on the entity similarity matrix and the relation similarity matrix, and combined with the preset semantic relevance threshold, key neighbor entities for structural and semantic matching are selected. S3, Neighbor-enhanced entity embedding update; A graph attention network is used to fuse the semantic features of the target entity and its key neighbors to generate context-aware entity embeddings. S4, Semantic Consistency Alignment Propagation: The alignment of first-order entity features and second-order relation structures are jointly optimized by sharing a permutation matrix. S5, Dynamic Seed Expansion Iteration: Based on the Manhattan distance of entity embedding, high-confidence misaligned entity pairs are dynamically identified. Entity pairs with a distance below the threshold are added to the seed set, and steps S2 to S4 are iteratively executed until convergence.

2. The cross-language entity alignment method based on relational semantic enhancement according to claim 1, characterized in that, The key neighbor screening in step S2 specifically includes: The entity similarity matrix is ​​calculated using the L1 distance between entity embedding vectors; The relation similarity matrix is ​​calculated using the semantic relevance of the relation embedding vectors; Only retain neighboring entities whose entity similarity and relationship similarity both exceed the preset threshold.

3. The cross-language entity alignment method based on relational semantic enhancement according to claim 1, characterized in that, The spatial structure information and semantic association information of the second-order relation structure alignment and fusion connection relationship in step S4 are specifically implemented as follows: By minimizing the differences in the relational structure matrix Optimize the shared permutation matrix P; where , An association matrix that integrates entity spatial structure and relational semantics; The constraint-shared permutation matrix P satisfies row normalization, column normalization, and its elements are binary values ​​{0,1}.

4. The cross-language entity alignment method based on relational semantic enhancement according to claim 3, characterized in that, The methods for constructing the relational structure matrix include: If there is a direct relationship between two entities, the relationship is embedded into the entity space structure matrix by dot product. If there is no direct relationship, the association matrix is ​​calculated based solely on the spatial structure information of the entities.

5. The cross-language entity alignment method based on relational semantic enhancement according to claim 4, characterized in that, The relation embedding generation method in step S1 includes: A relation vector is constructed by concatenating the average embeddings of the head and tail entities. ;in Indicates a connection operation. , Representing the relationship The average embedding of all different head and tail entities; Introducing TransE-based triple regularization loss Optimize relationship embedding.

6. The cross-language entity alignment method based on relational semantic enhancement according to claim 1, characterized in that, The dynamic seed expansion in step S5 specifically includes: In each iteration, the unaligned entities in the source knowledge graph are traversed, and the candidate entity with the smallest Manhattan distance is found in the target graph. If the distance is less than the threshold If so, the entity pair is added to the seed alignment set and the embedding representation is updated.

7. The cross-language entity alignment method based on relational semantic enhancement according to claim 1, characterized in that, The entity embedding in step S1 is generated in the following way: Update entity representations using multi-layer propagation rules of graph convolutional neural networks: ； in, It is the adjacency matrix of the graph. Add the identity matrix , For the first The entity feature vector of the layer, For the first The weight matrix of the layer, This represents the activation function. yes The degree matrix, ; Embedding optimization based on boundary loss function: ； in, The loss function represents entity embedding. This represents a boundary parameter greater than 0. Represents the initial seed alignment set of entities. This represents a negatively aligned set generated through nearest neighbor sampling. , representing the Manhattan distance of the entity embedding, This represents the Manhattan distance between negative sample entity pairs.

8. The cross-language entity alignment method based on relational semantic enhancement according to claim 1, characterized in that, The model is trained using a semi-supervised iterative strategy: The joint entity alignment loss and relation regularization loss are balanced by weighting coefficients. The entity distance matrix and relation distance matrix are updated synchronously during the iteration process.

9. A cross-language entity alignment system based on relational semantic enhancement, characterized in that, The system is used to execute the cross-language entity alignment method based on relational semantic enhancement as described in any one of claims 1-8.

Images