Industrial ontology matching method and system based on OWL2Vec* and graph isomorphic network
By using graph isomorphic network methods and combining semantic embedding with structural embedding, a comprehensive embedding vector is generated, which solves the matching error of nodes with similar semantics but different structures in industrial ontology, and achieves efficient ontology matching and data fusion.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV OF POSTS & TELECOMM
- Filing Date
- 2025-12-05
- Publication Date
- 2026-04-17
AI Technical Summary
Existing technologies struggle to effectively distinguish semantically similar but structurally different nodes in industrial ontology matching, leading to mismatches and ambiguities. Traditional methods fail to adequately model the hierarchical relationships and adjacency structures of the ontology.
A graph isomorphic network-based approach is adopted to convert the ontology into a graph structure. The semantic embedding vector and the RDF graph structure are input into the graph isomorphic network. The multi-head attention mechanism is used to capture the information interaction between nodes, perform message passing and feature aggregation, generate a comprehensive embedding vector, and achieve adaptive weighted fusion of semantics and structure through a gating fusion module. Finally, the similarity score of concept pairs is calculated.
It effectively solves the semantic inconsistency problem of industrial ontology, achieves accurate matching of heterogeneous ontology, and improves the quality of data fusion and system interoperability.
Smart Images
Figure CN121880951A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of semantic networks and natural language processing technology, and relates to... A method and system for matching industrial ontology in graph isomorphic networks. Background Technology
[0002] With the rapid development of the Industrial Internet, intelligent manufacturing, and digital twin technologies, the multi-source heterogeneity of industrial data is becoming increasingly prominent. Different enterprises, factories, or business systems often construct industrial ontologies based on their own standards to describe concepts such as equipment, products, processes, orders, materials, and customers, as well as their relationships. However, due to differences in modeling languages, domain specifications, data granularity, and semantic expression methods, even when referring to the same industrial entities, their semantic definitions and structural representations still differ significantly across systems, resulting in a large number of heterogeneous ontologies.
[0003] This heterogeneity leads to semantic inconsistencies between data and difficulties in interoperability between systems, severely restricting the implementation of core applications such as the integration of industrial knowledge graphs, intelligent scheduling, supply chain collaboration, and cross-enterprise data sharing.
[0004] Ontology matching is a key technology for solving the above problems. Its core objective is to automatically identify semantically identical or similar concepts and relationships in different ontologies, thereby achieving semantic alignment and fusion. In the industrial field, ontology matching is not only related to semantic interoperability, but also directly affects the transferability of production knowledge and the quality of data fusion.
[0005] Traditional ontology matching methods are generally divided into three categories: String-based matching methods match concepts by comparing the similarity of their labels or names. Methods based on lexical or semantic resources utilize knowledge bases such as WordNet to measure semantic similarity; Structure-based matching methods calculate similarity based on the hierarchical structure or adjacency relationship of concepts. However, these methods generally have the following limitations: (i) Semantic embedding models primarily focus on semantic similarity at the text level, but fail to adequately model the hierarchical relationships and adjacency structures of entities within the ontology. Consequently, they struggle to distinguish between nodes that are semantically similar but differ significantly in their structural positions. This problem is particularly pronounced in industrial scenarios. For instance, "finished product" and "semi-finished product" may be semantically similar, but their hierarchical positions within the production process ontology are entirely different. Relying solely on semantic embedding vectors can easily lead to mismatches.
[0006] (II) While structural encoding methods based on traditional graph convolutional networks (GCN, GAT) can capture local topological features, their aggregation mechanisms are mostly based on averaging or weighted summing of neighboring node features, lacking the ability to express the uniqueness of node structure. When structurally isomorphic or similar subgraphs exist in the graph, nodes with different semantics may obtain almost identical embedding representations, thus causing matching ambiguity. This problem is particularly serious in industrial ontology; for example, "assembly line nodes" in different workshops may have similar structural neighborhoods but semantically represent different entity roles. Summary of the Invention
[0007] In view of this, the object of the present invention is to provide a method based on... A method and system for industrial ontology matching based on graph isomorphic networks is proposed. This method extracts complex structural embedding vectors between entities and simultaneously fuses semantic and structural embedding vectors, effectively achieving ontology matching. The invention first converts the ontology into a graph structure, transforming entities into nodes and relationships into edges, using node information as... The model takes semantic embedding vectors as input to obtain the ontology. Then, these semantic embedding vectors and the RDF graph structure are input into a graph isomorphic network to model higher-order structural dependencies and generate structural embedding vectors. Further, a gating fusion mechanism is used to adaptively weight and fuse the semantic and structural embeddings to obtain a comprehensive embedding vector. Finally, a mapping selector is used to calculate the correct matching result.
[0008] To achieve the above objectives, the present invention provides the following technical solution: On the one hand, it provides a basis An industrial ontology matching method for graph isomorphic networks, the method comprising: S1. First, input the ontology file. The model generates semantic embedding vectors and RDF graph structures for each concept; S2. Input the semantic embedding vector and the RDF graph structure into the graph isomorphic network to model structural information, and optimize the model through an objective function to finally obtain the structural embedding vector. This includes the following sub-steps: S21. In graph isomorphic networks, semantic embedding vectors are used as the initial features of the corresponding nodes in the RDF graph structure. S22, No. Layer node feature vectors A multi-head attention mechanism is used to achieve information exchange and dependency capture between nodes, and to calculate the aggregate attention weight. ; S23, Based on Aggregated Attention Weights Execute a message passing mechanism to aggregate the characteristic information of neighboring nodes. ; S24. The aggregated features are fused with the historical states of the nodes, and the node representations are updated through nonlinear transformation to generate structural embedding vectors. ; S25. Employ the Manhattan distance metric to embed the similarity between vector pairs, aiming to minimize the distance between positive sample pairs and maximize the distance between negative sample pairs, and train a graph isomorphic network. S3. The semantic embedding vector and the structural embedding vector are weighted and fused through the gating fusion module to generate a comprehensive embedding vector; S4. Calculate the Euclidean distance between the composite embedding vectors using the FAISS vector library, and define the similarity score accordingly; calculate the similarity score between concept pairs based on the composite embedding vectors, and determine the optimal matching relationship based on the score, thereby completing the ontology matching task.
[0009] Furthermore, in step S22, the first... The process of aggregating attention weights at the nodes of a layer includes: First, the query matrices are obtained through linear transformations. Key matrix :
[0010]
[0011] In the formula, , The number of head nodes. For trainable parameters, The total dimension of the input embedding vector, index Indicates the corresponding node ; Next, regarding the first The head calculates the attention coefficient between nodes to measure the nodes' importance. For neighboring nodes Level of attention:
[0012] In the formula, This is a scaling factor used to prevent the inner product value from becoming too large, which could lead to gradient instability. The operation guarantees that the sum of the attention weights of each node to its neighboring nodes is 1; Finally, the attention weights of each head calculated by the multi-head attention algorithm are aggregated to obtain the aggregated attention weights:
[0013] In the formula, This refers to The attention weights of each attention head are concatenated into a vector. It is a trainable linear transformation matrix. This means mapping the concatenated vector to the attention weight space, achieving a dimensionality transformation so that the attention weights can match the feature dimensions of subsequent nodes.
[0014] Furthermore, in step S23, the message passing process is represented as follows:
[0015]
[0016] In the formula, Represents a node semantic embedding vector, It is the first The first layer The slave node of the head calculation To the node The message being delivered It is in the Layer from node To the node The message being delivered; It is in the Layer A learnable matrix for each node. The semantic embedding vector is projected onto the message passing space; For the target node , its first The neighbor aggregation of a layer is represented as:
[0017] in, Represents a node The set of neighbors, the aggregation process By adjusting attention weights, semantically relevant neighbors contribute more to node features, thereby achieving graph-based contextual semantic fusion. node In the Neighbor information after layer aggregation.
[0018] Furthermore, in step S24, the specific process of updating node features is as follows: First, compute the intermediate representation of the node:
[0019] in, For learnable residual weights, This represents a multilayer perceptron structure used for nonlinear feature extraction. Then, a preactivation function is introduced. Residual enhancement is performed using two-layer linear mapping, and batch normalization is employed to improve feature stability.
[0020]
[0021] In the formula, This indicates a batch normalization operation. Represents a node In the The feature vector of the layer, Indicates the first Layer nodes Features after residual enhancement and normalization , Indicates the first Layer-trainable linear mappings This represents a pre-activated nonlinear function. Indicates the first Layer nodes The structure embedding vector.
[0022] Furthermore, in step S25, the Manhattan distance is expressed as:
[0023] In the formula, Nodes representing the source ontology Structural features, Nodes representing the target entity Structural features, Represents the source ontology nodes and the target entity nodes The Manhattan distance between them.
[0024] In step S35, the distance between positive sample pairs is minimized and the distance between negative sample pairs is increased:
[0025] In the formula, Indicates positive samples of ontology matching. Indicates negative samples of ontology matching. These represent positive samples respectively. Zhongyuan ontology nodes and the target entity nodes , They represent negative samples respectively. Zhongyuan ontology nodes and the target entity nodes , This represents the marginal margin, ensuring that the distance between positive and negative samples is at least less than [value missing]. , express ,only Loss is calculated only if the value is greater than 0; otherwise, it is 0.
[0026] Furthermore, in step S3, the process of gating and fusing the semantic embedding vector and the structural embedding vector through the gating aggregation module includes: First, for concept nodes Its ontology semantic embedding vector is Its ontology's structure embedding vector is Then its gating weight is:
[0027] In the formula, , These represent the learnable parameters in the gated aggregator. For activation function, node The gating weight vector represents the importance of semantic information in the final representation vector.
[0028] The final integrated embedding vector representation is calculated based on the gating weights:
[0029] It dynamically balances the semantic embedding vector and the structural embedding vector through gating weights.
[0030] Furthermore, in step S4, matching is achieved by calculating the similarity between cross-ontology concept embeddings. The Euclidean distance is calculated using the FAISS vector retrieval library, and the similarity is defined based on its inverse distance.
[0031] In the formula, The composite embedding vector representing the source ontology. The comprehensive embedding vector representing the target ontology; The pair of embedded vectors with the highest similarity score is selected as the final mapping relationship to obtain the ontology matching result.
[0032] On the other hand, a method for performing the aforementioned based on A method and system for matching industrial ontology in graph isomorphic networks, the system including The modules include an embedding module, a graph isomorphic network module, a gated aggregation module, and a mapping selection module; among them, Embedded module, input ontology file The model generates semantic embedding vectors for each concept; The graph isomorphic network module obtains the structural information of the ontology through the graph isomorphic model and outputs the structural embedding vector of the entity. The gated aggregation module combines semantic embedding vectors and structural embedding vectors to obtain a comprehensive embedding vector.
[0033] The mapping selection module calculates similarity scores between cross-ontology concepts based on comprehensive embedding vectors, and selects the optimal matching entity pair by similarity ranking to achieve ontology matching between heterogeneous ontologies.
[0034] The beneficial effects of this invention are as follows: The present invention first inputs the ontology file. The model generates a semantic embedding vector for each concept. Next, these semantic embedding vectors, along with the RDF graph structure, are input into a graph isomorphic network model. This model uses a multi-head attention mechanism to model inter-node information interaction and dependencies, and sequentially performs message passing, feature aggregation, and node representation update operations to obtain the ontology's structural embedding vector. The model is then trained using an objective function. Next, the semantic and structural embedding vectors are input into a gating aggregation module to generate a comprehensive embedding vector. Finally, the optimal matching relationship is determined by calculating the similarity score between entity pairs, thus completing the ontology matching task.
[0035] The present invention The semantic embedding vectors acquired by the module aim to compensate for terminological differences and reduce mismatches caused by semantic inconsistencies in the ontology. Meanwhile, the graph isomorphic network module adaptively learns the influence of different neighboring nodes on the target node through attention weights, effectively mitigating structural heterogeneity. Compared to traditional graph neural networks such as GCN and GAT, graph isomorphic networks, while maintaining node isomorphism invariance, effectively capture complex graph topological patterns and high-order structural information by modeling dependencies between nodes through attention mechanisms. This invention utilizes graph isomorphic networks to achieve dynamic modeling of the structural context between nodes in an RDF graph structure, enabling effective alignment of structural patterns between different ontologies.
[0036] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description
[0037] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 Based on the embodiments of the present invention A schematic diagram of the overall process of industrial ontology matching method with graph isomorphic networks; Figure 2 Embodiments of the present invention A schematic diagram of the model's processing flow; Figure 3 This is a schematic diagram of the execution flow of the graph isomorphic network module in an embodiment of the present invention. Detailed Implementation
[0038] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.
[0039] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.
[0040] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.
[0041] Please see Figures 1-3 , for the basis A method and system for matching industrial ontology in graph isomorphic networks.
[0042] Example 1 This embodiment first gives a basis A detailed flowchart of the industrial ontology matching method based on graph isomorphic networks, such as... Figure 1 The diagram shown illustrates the overall flow of the method, which includes the following steps: S1. First, input the ontology file. The model generates semantic embedding vectors and RDF graph structures for each concept; S2. Input the semantic embedding vector and the RDF graph structure into the graph isomorphic network to model the structural information, and optimize the model through the objective function to finally obtain the structural embedding vector.
[0043] S3. The semantic embedding vector and the structural embedding vector are weighted and fused through the gating fusion module to generate a comprehensive embedding vector; S4. Calculate the Euclidean distance between the composite embedding vectors using the FAISS vector library, and define the similarity score accordingly; calculate the similarity score between concept pairs based on the composite embedding vectors, and determine the optimal matching relationship based on the score, thereby completing the ontology matching task.
[0044] In step S1 of this embodiment, the main task is to form a corpus by combining label and definition information: the input ontology is parsed into an RDF graph structure; for each concept node, its label, synonym, and definition are extracted to form a concept feature set; the RDFLib tool is used to complete the structural transformation from OWL to RDF graph; a random walk is performed on the RDF graph structure to generate a path sequence reflecting the structural context; the path is linearized into a text sequence; and the corpus is formed by combining label and definition information. Specifically, the process includes the following steps: S11. Parse the input ontology into an RDF graph structure. ,in For a set of concept nodes, For relational edge sets; S12, For each concept node Extract its label, synonym, and definition to form a concept feature set. Use the RDFLib tool to perform the structural conversion of OWL to RDF graph; S13. Perform random walks on the RDF graph structure to generate path sequences that reflect the structural context. The path is linearized into a text sequence, and a corpus is formed by combining labels and definition information. .
[0045] like Figure 2 As shown, using The specific process by which the model generates semantic embedding vectors is as follows: The Word2vec model is trained using a corpus, with the objective function being to maximize the context prediction probability.
[0046] In the formula, This represents the central word in the corpus. The context window that represents the headword, Model parameters, Given the central word Predict context words The probability of; For a concept node The semantic embedding vector at the node level is obtained by averaging the vector encoded by Word2vec.
[0047] in, The semantic vector of each text element after Word2vec encoding; Representing concept nodes The corpus is used. In step S2 of this embodiment, the semantic embedding vector and RDF graph structure are input into the graph isomorphic network model for structural information extraction. The graph isomorphic network model calculates the correlation between nodes through a multi-head attention mechanism to obtain the importance weight of neighbors. The values of neighbors are weighted according to the attention weight to form messages. All neighbor messages are aggregated at the node, and the historical representation of the node is merged with the aggregated message. The node embedding vector is updated through MLP transformation, using activation, feedforward layer, batch normalization, residuals, etc. The objective function is used to minimize the distance between positive sample pairs and increase the distance between negative sample pairs to train the model, so that the model can learn how to distinguish between similar and dissimilar concepts in different ontologies. Finally, the structural embedding vector of the ontology is obtained. The process is as follows. Figure 3 As shown, the specific execution process is as follows: (1) Input RDF graph structure and semantic embedding vector encoding. In this embodiment, the RDF graph structure constructed in the preprocessing stage will be used as input, and the semantic embedding vector encoding will be used as the node feature of the RDF graph structure node.
[0048] (2) Based on the input, the graph isomorphism model first achieves information interaction and dependency capture between nodes through a multi-head self-attention mechanism. For the node feature matrix... (No. (Node representation of the layer) First, the query matrix is obtained through linear transformation. Key matrix :
[0049]
[0050]
[0051] In the formula, , The number of head nodes. For trainable parameters, The total dimension of the input embedding vector, index Indicates the corresponding node ;
[0052] in This is a scaling factor used to prevent the gradient from becoming unstable due to excessively large inner product values. The operation guarantees that the sum of the attention weights of each node to its neighbors is 1. (This is achieved through...) This mechanism allows the model to dynamically adjust the information propagation weights between nodes on the graph structure, capturing semantically important neighbor relationships. The aggregated attention weights are:
[0053] In the formula, This refers to The attention weights of each attention head are concatenated into a vector. It is a trainable linear transformation matrix. This means mapping the concatenated vector to the attention weight space, achieving a dimensionality transformation so that the attention weights can match the feature dimensions of subsequent nodes.
[0054] (3) After the multi-head self-attention mechanism, the model undergoes message passing and aggregation. This aggregation process adjusts the attention weights to make semantically related neighbors contribute more to the node features, thereby achieving contextual semantic fusion based on the graph structure. The graph isomorphic network model executes the message passing mechanism based on these weights to aggregate the feature information of neighboring nodes.
[0055]
[0056]
[0057] In the formula, Represents a node semantic embedding vector, It is the first The first layer The slave node of the head calculation To the node The message being delivered It is in the Layer from node To the node The message being delivered; It is in the Layer A learnable matrix for each node. The semantic embedding vector is projected onto the message passing space; For the target node , its first The neighbor aggregation of a layer is represented as:
[0058] in, Represents a node The set of neighbors, the aggregation process By adjusting attention weights, semantically relevant neighbors contribute more to node features, thereby achieving graph-based contextual semantic fusion. node In the Neighbor information after layer aggregation.
[0059] (4) After undergoing contextual semantic fusion of the graph structure, the next step is node representation update. Specifically, after message aggregation, the node's features need to be fused with its historical state and nonlinearly transformed to form a new node representation. First, the model calculates the intermediate representation of the node:
[0060] in, For learnable residual weights, MLP represents a multilayer perceptron architecture used for nonlinear feature extraction. Subsequently, a pre-activation function is introduced. Residual enhancement is performed using two-layer linear mapping, and batch normalization is employed to improve feature stability.
[0061]
[0062] In the formula, This indicates a batch normalization operation. Represents a node In the The feature vector of the layer, Indicates the first Layer nodes Features after residual enhancement and normalization , Indicates the first Layer-trainable linear mappings This represents a pre-activated nonlinear function. Indicates the first Layer nodes The structure embedding vector.
[0063] (5) After obtaining the updated nodes, train the graph isomorphic network model and optimize the structural embedding vectors. The Manhattan distance is used to measure the similarity between concept pairs in the training set. The model is trained by minimizing the distance between positive sample pairs and increasing the distance between negative sample pairs, thus enabling the model to learn how to distinguish between similar and dissimilar concepts in different ontologies. The Manhattan distance formula is as follows:
[0064] In the formula, Nodes representing the source ontology Structural features, Nodes representing the target entity Structural features, Represents the source ontology nodes and the target entity nodes The Manhattan distance between them.
[0065] Minimize the distance between positive sample pairs and increase the distance between negative sample pairs:
[0066] in, Indicates positive samples of ontology matching. Indicates negative samples of ontology matching. These represent positive samples respectively. Zhongyuan ontology nodes and the target entity nodes , They represent negative samples respectively. Zhongyuan ontology nodes and the target entity nodes , This represents the marginal margin, ensuring that the distance between positive and negative samples is at least less than [value missing]. , express ,only Loss is calculated only if the value is greater than 0; otherwise, it is 0.
[0067] In step S3 of this embodiment, the purpose of the gating aggregation module is to aggregate semantic embedding vectors and structural embedding vectors. Specifically, firstly, for nodes... Its ontology semantic embedding vector is Its ontology's structure embedding vector is Then its gating weight is:
[0068] In the formula, , These represent the learnable parameters in the gated aggregator. For activation function, node The gating weight vector represents the importance of semantic information in the final representation vector.
[0069] The final integrated embedding vector representation is calculated based on the gating weights:
[0070] It dynamically balances the semantic embedding vector and the structural embedding vector through gating weights.
[0071] In step S4 of this embodiment, matching is achieved by calculating the similarity between cross-ontology concept embeddings. The Euclidean distance is calculated using the FAISS vector retrieval library, and the similarity is defined based on its inverse distance.
[0072] In the formula, The composite embedding vector representing the source ontology. The comprehensive embedding vector representing the target ontology; The pair of embedded vectors with the highest similarity score is selected as the final mapping relationship to obtain the ontology matching result.
[0073] Example 2 This embodiment provides a matching method based on the above-described matching system, which includes... The modules include an embedding module, a graph isomorphic network module, a gated aggregation module, and a mapping selection module; among them, Embedded module: Input ontology file The model generates semantic embedding vectors and RDF graph structures for each concept; The graph isomorphic network module takes the RDF graph structure and semantic embedding vectors as input. It models the dependencies between nodes through a multi-head attention mechanism, enabling cross-node information interaction and feature propagation. This module extracts structural features through message passing and feature aggregation, then fuses and nonlinearly transforms the node features with their historical states to update node features, outputting a structural embedding vector containing structural information.
[0074] Gated aggregation module: It takes semantic embedding vectors and structural embedding vectors as inputs, performs weighted fusion through a gating mechanism, and generates a comprehensive embedding representation that fully integrates semantic and structural feature information.
[0075] Mapping selection module: Based on the comprehensive embedding vector, it calculates the similarity score between cross-ontology concepts, selects the optimal matching entity pair by similarity ranking, and realizes ontology matching between heterogeneous ontologies.
[0076] In summary, through detailed implementation instructions, those skilled in the art can clearly understand the specific implementation steps of the present invention and can implement it using conventional software and hardware platforms.
[0077] This invention proposes a method based on An industrial ontology matching system and method for graph isomorphic networks. The system first inputs the ontology file... The model generates semantic embedding vectors and RDF graph structures for each concept. Further, the semantic embedding vectors and RDF graph structures are input into a graph isomorphic network model. This model models the information interaction and structural dependencies between nodes through a multi-head self-attention mechanism, and extracts high-order structural features through message passing and feature aggregation. Subsequently, the aggregated features are fused with the historical states of the nodes, and the node representations are updated through nonlinear transformations to finally generate structural embedding vectors. The model is then trained using an objective function optimization. Finally, the semantic embedding vectors and structural embedding vectors are input into a gated aggregation module, which adaptively fuses them to generate a comprehensive embedding vector. Based on this comprehensive embedding, the similarity score between concept pairs is calculated to determine the optimal matching relationship, thereby completing the ontology matching task.
[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. Based on The industrial ontology matching method based on graph isomorphic networks is characterized by: The method includes: S1. First, input the ontology file. The model generates semantic embedding vectors and RDF graph structures for each concept; S2. Input the semantic embedding vector and the RDF graph structure into the graph isomorphic network to model structural information, and optimize the model through an objective function to finally obtain the structural embedding vector; specifically including the following sub-steps: S21. In graph isomorphic networks, semantic embedding vectors are used as the initial features of the corresponding nodes in the RDF graph structure. S22, No. Layer node feature vectors A multi-head attention mechanism is used to achieve information exchange and dependency capture between nodes, and to calculate the aggregate attention weight. ; S23, Based on Aggregated Attention Weights Execute a message passing mechanism to aggregate the characteristic information of neighboring nodes. ; S24. The aggregated features are fused with the historical states of the nodes, and the node representations are updated through nonlinear transformation to generate structural embedding vectors. ; S25. Employ the Manhattan distance metric to embed the similarity between vector pairs, aiming to minimize the distance between positive sample pairs and maximize the distance between negative sample pairs, and train a graph isomorphic network. S3. The semantic embedding vector and the structural embedding vector are weighted and fused through the gating fusion module to generate a comprehensive embedding vector; S4. Calculate the Euclidean distance between the composite embedding vectors using the FAISS vector library, and define the similarity score accordingly; calculate the similarity score between concept pairs based on the composite embedding vectors, and determine the optimal matching relationship based on the score, thereby completing the ontology matching task.
2. The method based on claim 1 The industrial ontology matching method based on graph isomorphic networks is characterized by: In step S22, the first... The process of aggregating attention weights at the nodes of a layer includes: First, the query matrices are obtained through linear transformations. Key matrix : In the formula, , The number of head nodes. For trainable parameters, The total dimension of the input embedding vector, index Indicates the corresponding node ; Next, regarding the first The head calculates the attention coefficient between nodes to measure the nodes' importance. For neighboring nodes Level of attention: In the formula, This is a scaling factor used to prevent the inner product value from becoming too large, which could lead to gradient instability. The operation guarantees that the sum of the attention weights of each node to its neighboring nodes is 1; Finally, the attention weights calculated by each head in the multi-head attention process are aggregated to obtain the aggregated attention weights: In the formula, This refers to The attention weights of each attention head are concatenated into a vector. It is a trainable linear transformation matrix. This means mapping the concatenated vector to the attention weight space, achieving a dimensionality transformation so that the attention weights can match the feature dimensions of subsequent nodes.
3. The method based on claim 1 The industrial ontology matching method based on graph isomorphic networks is characterized by: In step S23, the message passing process is represented as follows: In the formula, Represents a node semantic embedding vector, It is the first The first layer The slave node of the head calculation To the node The message being delivered It is in the Layer from node To the node The message being delivered; It is in the Layer A learnable matrix for each node. The semantic embedding vector is projected onto the message passing space; For the target node , its first The neighbor aggregation of a layer is represented as: in, Represents a node The set of neighbors, the aggregation process By adjusting attention weights, semantically relevant neighbors contribute more to node features, thereby achieving graph-based contextual semantic fusion. Represents a node In the Neighbor information after layer aggregation.
4. The method based on claim 1 The industrial ontology matching method based on graph isomorphic networks is characterized by: In step S24, the specific process of updating node features is as follows: First, compute the intermediate representation of the node. : in, For learnable residual weights, This represents a multilayer perceptron structure used for nonlinear feature extraction. Then, a pre-activation nonlinear function is introduced with two layers of linear mapping for residual enhancement, and batch normalization is used to improve feature stability. In the formula, This indicates a batch normalization operation. Represents a node In the The feature vector of the layer, Indicates the first Layer nodes Features after residual enhancement and normalization , Indicates the first Layer-trainable linear mappings This represents a pre-activated nonlinear function. Indicates the first Layer nodes The structure embedding vector.
5. The method based on claim 1 The industrial ontology matching method based on graph isomorphic networks is characterized by: In step S25, the Manhattan distance is expressed as: In the formula, Nodes representing the source ontology Structural features, Nodes representing the target entity Structural features, Represents the source ontology nodes and the target entity nodes Manhattan distance between them; In step S25, the distance between positive sample pairs is minimized and the distance between negative sample pairs is increased: In the formula, Indicates positive samples of ontology matching. Indicates negative samples of ontology matching. These represent positive samples respectively. Zhongyuan ontology nodes and the target entity nodes , They represent negative samples respectively. Zhongyuan ontology nodes and the target entity nodes , This represents the marginal margin, ensuring that the distance between positive and negative samples is at least less than [value missing]. , express ,only Loss is calculated only if the value is greater than 0; otherwise, it is 0.
6. The method based on claim 1 The industrial ontology matching method based on graph isomorphic networks is characterized by: In step S4, the gating fusion process of semantic embedding vectors and structural embedding vectors through the gating aggregation module includes: First, for nodes Its ontology semantic embedding vector is Its ontology's structure embedding vector is Then its gating weight is: In the formula, , These represent the learnable parameters in the gated aggregator. For activation function, node The gated weight vector represents the importance of semantic information in the final representation vector; The final integrated embedding vector representation is calculated based on the gating weights: It dynamically balances the semantic embedding vector and the structural embedding vector through gating weights.
7. The method based on claim 1 The industrial ontology matching method based on graph isomorphic networks is characterized by: In step S5, matching is achieved by calculating the similarity of the ontology's integrated embedding vectors. The Euclidean distance is calculated using the FAISS vector retrieval library, and the similarity is defined based on its inverse distance. In the formula, The composite embedding vector representing the source ontology. The comprehensive embedding vector representing the target ontology; The concept pair with the highest similarity score is selected as the final mapping relationship to obtain the ontology matching result.
8. A method for performing the action based on any one of claims 1-7. The system of industrial ontology matching method based on graph isomorphic networks is characterized by: The system includes The modules include an embedding module, a graph isomorphic network module, a gated aggregation module, and a mapping selection module; among them, Embedded module, input ontology file The model generates semantic embedding vectors and RDF graph structures for each concept; The graph isomorphic network module obtains the structural information of the ontology through the graph isomorphic model and outputs the structural embedding vector of the entity. The gated aggregation module combines semantic embedding vectors and structural embedding vectors to obtain a comprehensive embedding vector. The mapping selection module calculates similarity scores between cross-ontology concepts based on comprehensive embedding vectors, and selects the optimal matching entity pair by similarity ranking to achieve ontology matching between heterogeneous ontologies.