Adaptive industrial ontology matching method and system based on multi-view structure decoupling

Abstract

The invention relates to a self-adaptive industrial ontology matching method and system based on multi-view structure decoupling, and belongs to the technical field of semantic network and natural language processing. Firstly, an unsupervised training data set of double-layer filtering is constructed based on a bidirectional reciprocal constraint mechanism of rule matching and vector space, then a coarse recall model of transverse and longitudinal double-view decoupling is constructed, longitudinal lineage reasoning features and transverse neighborhood semantic field features are extracted by adopting a parameter freezing strategy, then the longitudinal lineage reasoning features and the transverse neighborhood semantic field features are fused and adversarial disturbance is introduced, and a final final final result is obtained. The method comprises the following steps: generating global entity representation, calculating a similarity score, carrying out adaptive routing shunting by using a dual threshold and a competitive difference to obtain a definite sample and a fuzzy sample, inputting the fuzzy sample and a structured context into a full interactive cross encoder, and carrying out deep semantic disambiguation and preferential selection to complete matching. According to the method, dependence on manual annotation is reduced, the video memory bottleneck of a traditional graph algorithm is effectively overcome, and the matching efficiency and accuracy of the large-scale industrial ontology are improved.

Description

Technical Field

[0001] This invention belongs to the field of semantic networks and natural language processing technology, and relates to an adaptive industrial ontology matching method and system based on multi-view structure decoupling. Background Technology

[0002] With the rapid development of the Industrial Internet, intelligent manufacturing, and digital twin technologies, data governance technologies based on semantic web and knowledge graphs are being applied more and more widely in industrial scenarios. Ontologies, as formal representation models of knowledge, are used to standardize the description of core concepts such as equipment, products, processes, and BOM structures, as well as their complex relationships. However, due to significant differences in the modeling standards, business focuses, and organizational structures followed by different enterprises, factories, or business systems when constructing ontologies, a large number of multi-source heterogeneous ontologies exist in the industrial data space.

[0003] Conflicts in naming conventions and inconsistent hierarchical structures between different ontologies have led to the formation of "data silos" between systems, which seriously restricts the implementation of core applications such as industrial knowledge graph fusion, cross-enterprise supply chain collaboration, full lifecycle management, and intelligent production scheduling.

[0004] Ontology matching is a key technology for addressing the semantic heterogeneity challenge of multi-source data. Essentially, it uses algorithms to automatically mine and identify sets of entities representing the same industrial concept across different namespaces. In smart manufacturing scenarios, the robustness of the matching results directly constrains the efficiency of the flow of underlying data to upper-level applications, and is crucial for ensuring the quality of knowledge reuse.

[0005] Traditional ontology matching methods are generally classified into the following categories:

[0006] String-based matching methods match concepts by comparing the string similarity of their labels or names. Syntax-based matching methods: These methods calculate edit distance or N-gram similarity, relying solely on surface features of the strings for matching. The matching method based on external resources utilizes general dictionaries such as WordNet to assist in disambiguation and calculate semantic similarity; Structure-based matching methods utilize graph structures to propagate similarity.

[0007] However, these methods generally have the following limitations: (I) Limitations of Semantic Embedding Models: These methods (such as BERT-based BertMap and S-Match) typically employ a dual-tower architecture, where pre-trained language models independently encode source and target entities, followed by matching based on vector cosine similarity. Due to this "independent encoding" mechanism, source and target entities lack deep semantic interaction during feature extraction, leading to the loss of fine-grained semantic differences during vector compression. In industrial scenarios, there are numerous adversarial entities that are literally highly similar but semantically distinct (e.g., "inlet valve" and "outlet valve," "primary pump" and "secondary pump" in a BOM structure). Dual-tower models often map these entities to extremely similar vector representations, causing semantic ambiguity and producing high-confidence false matches, failing to meet the stringent precision requirements of industrial applications.

[0008] (II) Limitations of Structure-Based Embedding Models: These methods (such as GCN-Align, GAT, and other graph neural network methods) attempt to enhance matching performance by leveraging the topological information of the ontology and aggregating neighbor features through a recursive message passing mechanism. However, they face a dual challenge in industrial applications: First, memory bottlenecks and computational scalability issues. Industrial ontologies typically contain hundreds of thousands of nodes. As the number of GNN layers increases, the receptive field of a node expands exponentially (i.e., the "neighbor explosion" phenomenon), leading to a rapid expansion of the computational graph during training and inference. This makes it difficult to deploy such algorithms on resource-constrained industrial edge devices or consumer-grade graphics cards, greatly limiting their practical application capabilities. Second, matching ambiguity caused by structural isomorphism. Industrial scenarios often contain subgraphs with highly similar or even isomorphic structures (for example, production lines in different workshops may have almost identical equipment connection structures, but their physical meanings are different). In this case, nodes with different semantics may obtain almost identical embedding representations, making it difficult for the model to distinguish them, thus leading to matching errors.

[0009] (III) The Universality Challenges of Existing Supervised Learning Paradigms: Whether semantic or structural models, high performance often relies on extensive manually labeled matching data for supervised training. However, the industrial sector suffers from a severe "cold start" problem: different companies lack publicly available matching annotations for their private ontologies, and manual annotation by domain experts is extremely costly. Traditional supervised learning methods experience a sharp performance drop when labeled data is lacking, while simple unsupervised methods (such as those based on string rules) easily introduce a large amount of noise, making it difficult to construct high-quality training samples. Summary of the Invention

[0010] In view of this, existing industrial ontology matching technologies suffer from problems such as high memory consumption, difficulty in fine-grained semantic disambiguation, and lack of high-quality labeled data. This invention provides an adaptive industrial ontology matching method and system based on multi-view structural decoupling. By using horizontal and vertical dual-view decoupling and parameter freezing strategies, it aims to overcome the memory bottleneck of large-scale graph computation. It utilizes context-aware cross-coding and adaptive routing mechanisms to solve the semantic ambiguity problem in industrial scenarios. Furthermore, it combines reciprocal nearest neighbor mining technology to achieve cold-start matching under unsupervised conditions.

[0011] To achieve the above objectives, the present invention provides the following technical solution: On the one hand, an adaptive industrial ontology matching method based on multi-view structure decoupling is provided, which includes the following steps: S1. Based on rule matching, an initial seed set is generated from the standardized text of the source ontology and the target ontology. A bidirectional reciprocal constraint mechanism based on vector space is established, and a pre-trained language model is used to mine the implicit matching samples that are the nearest neighbors. The initial seed set and the implicit matching samples are used together to construct a two-layer filtered unsupervised training dataset. S2. Construct a coarse recall model with horizontal and vertical dual-view decoupling, which includes a text semantic view encoding branch, a vertical lineage view encoding branch and a horizontal neighborhood view encoding branch. Extract global semantic features, vertical reasoning features and horizontal structural features respectively, and fuse multi-view features and introduce adversarial perturbation to generate a global representation of the entity. S3. Establish an adaptive route based on prediction confidence, and dynamically divide the samples into a set of sure-match samples, sure-mismatch samples, and fuzzy samples according to the double reset confidence threshold. S4. For the diverged samples that are sure to match and sure to not match, directly output the final matching result; for the diverged fuzzy sample set, explicitly extract the structured context information of the entities and input it into the fully interactive cross encoder for deep semantic disambiguation and optimization, and output the final matching result.

[0012] Furthermore, in step S1, the initial seed set generation process includes: firstly, parsing the input source ontology and target ontology into structured data; for each entity node, extracting its name label and synonyms to form an initial concept feature set; performing fine-grained standardization cleaning on the initial concept feature set, and using a pre-trained language model to map the cleaned text into a high-dimensional dense vector; then generating pseudo-labels by combining rule matching and reciprocal nearest neighbor mining strategies, wherein an ontology parsing tool is used to traverse the input ontology file; for each entity node in the ontology... Index its main label Label and all aliases Synonym to construct an entity text set. ; For entity text collections The standardized text cleaning process includes phrase segmentation, noise filtering, and normalization, ultimately forming a standardized text feature set. Then, based on this standardized text feature set, a string matching strategy is executed, traversing entity pairs between the source and target ontologies and calculating the string overlap of their standardized names. If the names of two entities or any synonyms achieve a complete string match, they are classified as high-confidence positive samples and directly included in the initial seed set. This yields the initial seed set. The process of mining implicit matching samples and constructing a two-layer filtered unsupervised training dataset is as follows: using a pre-trained language model to encode the standardized entity text into high-dimensional semantic vectors. Construct the source entity vector matrix With the target entity vector matrix And calculate the cosine similarity matrix between the two. Then execute the two-way reciprocity verification logic: only if the target entity It is the source entity The term with the highest similarity in the target space, and also the source entity Also the target entity The pair with the highest similarity in the source space, and whose similarity is greater than the high confidence threshold, satisfies the two-way reciprocity constraint. Then, entity pairs satisfying the following two-way reciprocity constraint are selected. Add to the training set, i.e.:

[0013] In the formula, Represents the source entity With the target entity Similarity between them Represents the source entity With each entity in the target space Similarity; Represents the target entity With each entity in the source space similarity, High confidence threshold.

[0014] Furthermore, in the text semantic view encoding branch of step S2, the global semantic features of the central entity are extracted using the parameter-updatable first semantic encoding branch. This involves... The standardized name sequence is input into the BERT-Base model, and the vector at the last layer [CLS] position is extracted as the global semantic feature. During this process, the parameters of the first semantic encoding branch remain unfrozen and are updated with the training gradient to capture the core lexical meaning of entities.

[0015] Furthermore, in the longitudinal phylogenetic view encoding branch of step S2, a second semantic encoding branch with parameter freezing is used to capture the logical reasoning features from the root node to the leaf node, including phylogenetic path projection, feature vectorization, multi-scale convolutional encoding, and pooling output processes. Lineage path projection: for any entity in the ontology By backtracking to the root node using the shortest path algorithm, a unique lineage path sequence is constructed. :

[0016] In the formula, For the current entity, For the first Skip to ancestor node, This refers to the hierarchical relationship type between connecting nodes; Feature vectorization: The semantic vectors of each node in the path are extracted using the second semantic encoding branch with parameter freezing, and then reduced to dimensionality through a linear layer. Simultaneously, the relation type is mapped to an embedding vector, and the node vector and relation vector are concatenated in sequence to form a path feature matrix. ; Multi-scale convolutional coding: Introducing a one-dimensional convolutional neural network, utilizing a set of convolutional kernels. For the characteristic matrix Perform sliding convolution operation to extract local inference features. The convolution calculation method is as follows:

[0017] in, The size of the convolution window corresponds to the single-hop inheritance relationship and the double-hop inference logic, respectively. Represents a local segment of a path sequence; For bias terms; Pooling output: Global max pooling is performed on the feature map generated by convolution to obtain the vertical structure feature vector. .

[0018] Furthermore, in the lateral neighborhood view encoding branch of step S2, the semantic field features of the local neighborhood are aggregated using the parameter-frozen second semantic encoding branch, including isomorphic neighborhood sampling, attention aggregation, and feature weighting processes. Homogenization neighborhood sampling selects entities The direct parent node and its sibling nodes constitute the neighborhood set. ; Attention aggregation utilizes parameter-frozen secondary towers to extract neighbor nodes. Text features and target entity Text features This involves introducing a graph attention mechanism, which calculates attention weights based on the semantic relevance between the target entity and its neighboring nodes. :

[0019] in, It is a learnable linear transformation matrix. For learnable attention vectors, This represents a vector concatenation operation. It is a non-linear activation function; Finally, the lateral structural feature vector is obtained by weighted summation. .

[0020] Furthermore, the process of fusing global semantic features, vertical reasoning features, and horizontal structural features in step S2 to generate a global representation is as follows: Feature fusion: combining global semantic features Vertical structural features and lateral structural features Dimensional concatenation is performed, and features are fused through a learnable linear mapping layer to obtain the final multi-view fused feature vector of the entity. The calculation formula is as follows:

[0021] in, This represents a vector concatenation operation; and These are the weight matrix and bias term of the fusion layer, respectively; Adversarial perturbation introduction: Calculating the total loss function Regarding the input layer embedding matrix gradient Generate bounded adversarial perturbations based on gradient direction :

[0022] in This is the preset perturbation amplitude hyperparameter; Adversarial optimization: perturbation Superimposed onto the original embedding matrix, adversarial examples are constructed, and the optimization objective is to minimize the loss in the worst case:

[0023] Iterative training is performed on a coarse recall model with decoupled horizontal and vertical views. The model parameters are updated using the gradient descent algorithm until the total loss function converges or the preset maximum number of iterations is reached. After training, the final multi-view fusion feature vector is output. As a unified representation of the source entity and the target entity in subsequent steps, they are respectively denoted as... and .

[0024] Furthermore, in step S3, an adaptive routing mechanism is used to split the coarse recall results, wherein the unified representation of the source entity and the target entity is used as the basis for the splitting. and Calculate cosine similarity score :

[0025] in, Represents the vector dot product. The L2 norm of a vector; Establish a dual-reset confidence threshold and triage the data, including a high-confidence threshold. and low confidence threshold and competitive difference threshold The diversion process is as follows: like and If the sample is determined to be a sure match, the matching result is output directly. like If the sample is determined to be a mismatch, it will be filtered out directly. like If the sample is a numerically fuzzy sample, then the sample is determined to be a numerically fuzzy sample; if > But the difference If it is, then it is judged as a competing fuzzy sample. The highest similarity score is given. It received the second-highest similarity score.

[0026] Furthermore, in step S4, the input and output processes of the fully interactive cross-encoder model are as follows: For each pair of entities to be decided in the set Explicitly extract the parent class name and sibling class name of the entity; The entity names and their structured contexts are concatenated into a long sequence using delimiters specific to the pre-trained model. The formula is constructed as follows:

[0027] in, This represents a sequence concatenation operation; The name text representing the entity; The structural context text representing the entity, ; , Used as a separator; The constructed long sequence The input fully interactive cross-encoder model is processed as follows: First, extract the hidden state vector at the [CLS] position of the last layer of the model. As a semantic aggregation representation of the entire sequence:

[0028] Then, the vector is fed into a fully connected layer, and the output is mapped to a sigmoid activation function. Interval matching probability :

[0029] Finally, the best candidate is selected, and the source entity is calculated. The probability set of all corresponding candidate target entities Execute the ArgMax optimization strategy: Select the target entity with the highest probability value As the optimal matching candidate:

[0030] If the target entity probability value If the result is greater than the preset value, the final matching result will be output. Otherwise, determine the source entity. No valid match found.

[0031] Furthermore, in step S4, during the training of the fully interactive cross-encoder model, a dedicated training set is constructed using a hard example mining strategy based on coarse recall results, and training is performed, including: Static hard example generation: Using a coarse recall model that has been trained and converged and whose parameters are frozen, a full database search is performed on each source entity in the training set to recall the Top-K candidate targets; Hard-to-bearing examples screening: Among the recalled Top-K candidates, entities with probability values ​​higher than the preset value but negative true labels are screened out and marked as hard-to-bearing examples; Structured augmentation training: Mix hard negative examples with real positive samples, and construct parent and sibling class contexts according to the input sequence construction method to build triples or classification datasets for fine-tuning training of the fully interactive cross-encoder model.

[0032] On the other hand, a system for performing the aforementioned adaptive industrial ontology matching method based on multi-view structure decoupling is also provided, the system comprising: The unsupervised data mining module is configured to solve the cold start problem and automatically build training sets, including a rule seed generation unit and a mutual nearest neighbor mining unit. The coarse recall module with horizontal and vertical dual-view decoupling is configured to extract entity features and calculate global similarity, and includes a text semantic encoding unit, a vertical phylogenetic convolutional encoding unit, a horizontal neighborhood attention encoding unit, and an adversarial fusion training unit. The adaptive routing control module is configured to divide entity pairs into a set of sure-match, sure-mismatch, and fuzzy samples based on dual thresholds, and distribute the fuzzy samples to the next module. The context-aware fine-grained rearrangement module is configured for deep disambiguation of fuzzy samples and includes a context construction unit and a cross-coding inference unit.

[0033] The beneficial effects of this invention are as follows: This invention overcomes the memory bottleneck in large-scale graph computation, achieving superior resource efficiency. Unlike traditional Graph Neural Networks (GNNs), which rely on recursive aggregation leading to exponential expansion of the computational graph with each layer (i.e., the "neighbor explosion" phenomenon), this invention employs a "horizontal and vertical dual-view decoupling" architecture combined with a "parameter freezing" strategy to cut off the gradient backpropagation paths of massive neighbor nodes; and transforms complex tree-structured inference into lightweight 1D-CNN sequence operations. This design reduces memory consumption from being strongly correlated with graph size to being linearly correlated only with batch size, enabling the model to be trained on large-scale industrial ontologies (hundreds of thousands of nodes) with significantly reduced memory usage, thus allowing deployment on resource-constrained industrial edge devices.

[0034] This invention solves the problems of semantic ambiguity and high-score competition in industrial scenarios, significantly improving matching accuracy. Addressing the weakness of traditional dual-tower models in distinguishing between entities with highly similar literal meanings but opposite semantics (such as "inlet valve" and "outlet valve"), this invention designs a cascade mechanism of "adaptive routing + context-aware fine-grained reordering." The system automatically filters out "high-score competition" difficult examples that not only have high absolute scores but also have minimal differences from the suboptimal solution by introducing a "competitive difference threshold." It also utilizes a fully interactive cross-encoder containing structured context for deep disambiguation. In particular, by introducing a hard negative mining training strategy, the model is forced to focus on fine-grained structural differences, thereby significantly improving the precision of Top-1 matching while maintaining recall.

[0035] This invention eliminates the reliance on manually labeled data, enabling low-cost cold start. Addressing the lack of high-quality matching annotations in the industrial sector, this invention innovatively constructs a two-layer filtering mechanism of 'rule-guided + two-way reciprocity'. Compared to traditional supervised learning methods that heavily rely on expert annotations and simple rule-based methods that are prone to introducing noise, this invention first generates high-confidence anchor points based on rules, and then establishes a two-way reciprocity constraint mechanism based on vector space. This mechanism can automatically mine implicit matching samples that are literally dissimilar but semantically highly consistent from scratch. This dual-constraint design, while ensuring sample quality, significantly reduces the labor costs of industrial implementation, allowing the method to be quickly transferred to different industrial verticals.

[0036] This invention resolves matching ambiguities caused by structural isomorphism and enhances the robustness of the model. Traditional methods often overemphasize local neighborhood structures, easily leading to ambiguity in structurally isomorphic scenarios (such as different workshops having the same equipment connection structure). This invention, by introducing a vertical phylogenetic view, encodes the absolute hierarchical position of entities into features using shortest path projection, achieving an organic unity between global structure and local semantics. Simultaneously, adversarial training is introduced during the training process, simulating data perturbations, significantly improving the model's generalization ability and robustness in the face of industrial data noise (such as non-standard naming and missing synonyms).

[0037] 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

[0038] 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 This is a schematic diagram of the overall process of the adaptive industrial ontology matching method based on multi-view structure decoupling in an embodiment of the present invention; Figure 2 This is a schematic diagram of the coarse recall module with horizontal and vertical dual-view decoupling according to an embodiment of the present invention. Figure 3 This is a logical diagram illustrating adaptive routing and fine-grained reordering in an embodiment of the present invention; Figure 4 This is a functional block diagram of an adaptive industrial ontology matching system based on multi-view structure decoupling, provided in an embodiment of the present invention. Detailed Implementation

[0039] 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.

[0040] 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.

[0041] 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.

[0042] Please see Figures 1-4 This presents an adaptive industrial ontology matching method and system based on multi-view structure decoupling.

[0043] Example 1 This embodiment first provides a detailed flowchart of the adaptive industrial ontology matching method based on multi-view structural decoupling, such as... Figure 1 The diagram shown illustrates the overall flow of the method, which includes the following steps: S1. Constructing a two-layer filtering unsupervised training dataset: The source and target ontologies to be matched are text-standardized, and an initial seed set for explicit matching is generated based on rule matching. At the same time, a bidirectional reciprocal constraint mechanism based on vector space is established, and a pre-trained language model is used to mine implicit matching samples that are literally dissimilar but whose semantic vectors are the nearest neighbors. The initial seed set and the implicit matching samples are used together to construct a two-layer filtering unsupervised training dataset. S2. Constructing a coarse recall model with horizontal and vertical dual-view decoupling: Entity features are extracted using a horizontal and vertical dual-view decoupling architecture and a parameter freezing strategy. Specifically, textual semantic features are extracted using a parameter-updatable first semantic encoding branch. A vertical lineage view and a horizontal neighborhood view are constructed using a parameter-freezing second semantic encoding branch. Vertical inference features are extracted using a one-dimensional convolutional neural network, and horizontal neighborhood features are aggregated using a graph attention mechanism. Finally, the multi-view features are fused and adversarial perturbations are introduced to generate a global representation of the entity. S3. Adaptive routing based on prediction confidence: Calculate the cosine similarity score between the source entity and the target entity, and dynamically divide the samples into a set of sure-match samples, sure-mismatch samples and fuzzy samples according to the preset double reset confidence threshold. S4. Context-aware fine-grained rearrangement: For the fuzzy sample set diverted in step S3, the structured context information of the entities is explicitly extracted, a fully interactive cross-encoder model is constructed with long sequence input fine-tuned by a hard example mining strategy, deep semantic disambiguation and optimization are performed, and the final matching result is output.

[0044] In step S1 of this embodiment, the main task is to construct the unsupervised training dataset with two-layer filtering: The input source and target ontology are parsed into structured data. For each entity node, its name label (rdfs:label) and synonyms (skos:altLabel) are extracted to form an initial concept feature set. Fine-grained standardization and cleaning are performed on the text feature set, and a pre-trained language model is used to map the cleaned text into a high-dimensional dense vector. Based on this, pseudo-labels are generated using rule matching and reciprocal nearest neighbor mining strategies. Specifically, the process includes the following steps: Ontology parsing and fine-grained text normalization: Ontology parsing tools (such as OWL API or RDFLib) are used to traverse the input OWL or RDF format ontology file; for each entity node in the ontology... It indexes its main label (Label) and all aliases (Synonyms) to construct an entity text set. .

[0045] Subsequently, Perform a standardized cleaning process on the text sequence: Phrase segmentation: Identify non-natural language delimiters in text (such as "_", "-", or camelCase) and convert them back to natural language space-separated forms (e.g., convert Water_Valve or waterValve to water valve). Noise filtering: Use a stop word list to remove common words such as "the", "of", and "system" that do not contribute to the meaning. Normalization process: All characters are converted to lowercase and lemmatizer is used to convert plural nouns or verbs back to their root prototypes, ultimately forming a standardized set of text features.

[0046] The rule-based initial seed generation executes a strict string matching strategy based on a standardized text feature set. It iterates through entity pairs in both the source and target ontologies, calculating the string overlap of their standardized names. If the names or any synonyms of two entities achieve an exact match at the string level, they are classified as high-confidence positive samples and directly included in the initial seed set. , serving as "anchor points" for model training; Reciprocal Nearest Neighbor (RNN) mining based on vector space aims to discover implicit matches that are literally dissimilar but semantically equivalent. It utilizes pre-trained language models (such as BERT) to encode standardized entity text into high-dimensional semantic vectors. Construct the source entity vector matrix. With the target entity vector matrix And calculate the cosine similarity matrix between the two. Based on this, execute the two-way reciprocal verification logic: only if the target entity... It is the source entity The item with the highest similarity in the target space (Top-1), and also the source entity Also the target entity When the term with the highest similarity in the source space is found, that is, when the condition is met... and , Represents the source entity With each entity in the target space Similarity; Represents the target entity With each entity in the source space The similarity.

[0047] Finally, further high confidence thresholds were used. (e.g., 0.85) Perform secondary filtering on reciprocal pairs, and add the entity pairs that pass the verification as incremental positive samples to the training set, thereby achieving cold start without manual annotation.

[0048] Next, proceed to step S2. For example... Figure 2The diagram shown illustrates the architecture of the coarse recall model with horizontal and vertical dual-view decoupling in this embodiment of the invention. This step aims to utilize three views—horizontal (neighborhood), vertical (lineage), and semantic (text)—and employ a "dual-stream decoupling" and "parameter freezing" strategy to extract multi-dimensional features while overcoming the memory bottleneck of large-scale graph computation, and then perform full coarse recall. The specific implementation process includes the following sub-steps: Construct a text semantic view (V1) and extract global semantic features of the central entity using the parameter-updatable first semantic encoding branch (main tower). Specifically, the central entity... The standardized name sequence is input into the BERT-Base model, and the vector at the last layer [CLS] position is extracted as the global semantic feature. During this process, the parameters of the first semantic encoding branch remain unfrozen and are updated with the training gradient to capture the core lexical meaning of entities.

[0049] Construct a vertical lineage view (V2) and use the second semantic encoding branch (sub-tower) with parameter freezing to capture the logical reasoning features from the root node to the leaf node.

[0050] Lineage path projection: for any entity in the ontology By backtracking to the root node using the shortest path algorithm, a unique lineage path sequence is constructed. :

[0051] in, For the current entity, For the first Skip to ancestor node, This refers to the hierarchical relationship type between connecting nodes.

[0052] Feature vectorization: Semantic vectors of each node in the path are extracted using the parameter-frozen secondary tower, and then reduced to dimensionality through a linear layer. Simultaneously, relation types are mapped to embedding vectors. Node vectors and relation vectors are concatenated sequentially to form a path feature matrix. .

[0053] Multi-scale convolutional coding: Introducing a one-dimensional convolutional neural network (1D-CNN), utilizing a set of convolutional kernels. For the characteristic matrix Perform a sliding convolution operation to extract local inference features. The convolution calculation formula is:

[0054] in, The size of the convolution window corresponds to the single-hop inheritance relationship and the double-hop inference logic, respectively. Represents a local segment of a path sequence; This is a bias term.

[0055] Pooling output: Global max-pooling is performed on the feature map generated by convolution to obtain the vertical structure feature vector. .

[0056] The horizontal neighborhood view (V3) is constructed by using the parameter-frozen sub-tower to aggregate the semantic field features of the local neighborhood.

[0057] Homogenized neighborhood sampling, selecting entities The direct parent node and its sibling nodes constitute the neighborhood set. To reduce computational complexity, this step ignores the connection relationships between nodes and only retains the text sequences of neighboring nodes.

[0058] Attention aggregation extracts neighbor nodes using the parameter-frozen sub-tower. Text features and target entity Text features Although edge attributes are ignored, this step introduces a graph attention mechanism, calculating attention weights based on the semantic relevance between the target entity and its neighboring nodes. The formula is:

[0059] in, It is a learnable linear transformation matrix. For learnable attention vectors, This represents a vector concatenation operation. It is a non-linear activation function.

[0060] Feature weighting: The final lateral structure feature vector is obtained by weighted summation.

[0061] Multi-view fusion and adversarial training: Finally, the features of the three views mentioned above are fused, and adversarial perturbations are introduced in each iteration of model training to enhance robustness.

[0062] Feature fusion: combining global semantic features Vertical structural features and lateral structural features Dimensional concatenation is performed, and feature fusion is carried out through a learnable linear mapping layer to obtain the final multi-view fused feature vector of the entity. The calculation formula is as follows:

[0063] in, This represents a vector concatenation operation; and These are the weight matrix and bias term of the fusion layer, respectively. This vector... That is, to serve as a unified representation of the source entity and the target entity in subsequent steps (denoted as respectively). and ).

[0064] Adversarial perturbation introduction: Calculating the total loss function Regarding the input layer embedding matrix gradient Generate bounded adversarial perturbations based on gradient direction. The formula is as follows:

[0065] in The preset perturbation amplitude hyperparameter (e.g., 1.0).

[0066] Adversarial optimization: perturbation Superimposed onto the original embedding matrix, adversarial examples are constructed, and the optimization objective is to minimize the loss in the worst case:

[0067] Next, steps S3 and S4 are executed. For example... Figure 3 The diagram shown illustrates the logic of adaptive routing and fine-grained reordering in this embodiment of the invention. The system utilizes an adaptive routing mechanism to distribute the coarse recall results and performs context-aware deep disambiguation only on fuzzy samples. The specific implementation process is as follows: Step S3 is adaptive routing (traffic splitting strategy) based on predicted confidence. This step utilizes the cosine similarity score output from step S2. The samples are dynamically split based on a preset dual threshold and competitive difference.

[0068] Calculate the cosine similarity score to obtain the source entity. The final fusion features With the target entity The final fusion features The score is calculated using the cosine similarity formula. :

[0069] in, Represents the vector dot product. This represents the L2 norm of a vector. (Score) The range of values ​​is .

[0070] Set up a dual-reset confidence threshold and split the stream based on a preset high confidence threshold. and low confidence threshold (For example, in this embodiment) =0.85, =0.45), and introduces a "competitive difference threshold". (In this embodiment, 0.05 is preferred), the calculated score Perform the following routing decision: (1) Confident Match: Satisfies and The sample.

[0071] (2) Confident Non-Match: Satisfies The sample.

[0072] (3) Ambiguous Sample: Numerical ambiguity: The highest score is in The sample.

[0073] Competition is ambiguous: Although > But the difference .

[0074] Step S4, context-aware fine-grained rearrangement (deep disambiguation and optimization), targets the fuzzy sample set derived in step S3 (i.e., for the same source entity). There are one or more candidate target entities. (All meet the fuzzy judgment threshold), this step introduces structural context for deep disambiguation and selection.

[0075] Structured context extraction and input sequence construction, for each entity pair to be decided in the set. To address the challenge of semantic indistinguishable entity names alone, we explicitly extract the parent name and sibling names of the entity.

[0076] The entity names and their structured contexts are concatenated according to a pre-trained model-specific delimiter (such as [SEP]) to form a long sequence. The formula is constructed as follows:

[0077] in, This represents a sequence concatenation operation; The name text representing the entity; The structural context text representing the entity is specifically defined as the concatenation of the parent class name and the sibling class name:

[0078] Cross-coding reasoning will construct long sequences Input a fully interactive Transformer model (Cross-Encoder) to enable deep interaction of source and target entity information at the token level.

[0079] First, extract the hidden state vector at the [CLS] position of the last layer of the model. As a semantic aggregation representation of the entire sequence:

[0080] Then, the vector is input into a fully connected linear layer, and the output is mapped to a sigmoid activation function. Interval matching probability :

[0081] Finally, the best candidate is selected, and the source entity is calculated. The probability set of all corresponding candidate target entities Execute the ArgMax optimization strategy: Select the target entity with the highest probability value As the optimal matching candidate:

[0082] like Then the final matching result will be output. Otherwise, determine the source entity. No valid match found.

[0083] Hard example mining and training strategy for cross-encoders: To enable the fine-grained rearrangement module to handle "fuzzy samples," this embodiment employs a hard example mining strategy based on coarse recall results to construct a dedicated training set for S4. Static hard example generation: Using the S2 coarse recall model which has been trained and converged and whose parameters have been frozen, a full database search is performed on each source entity in the training set to recall Top-K (e.g., Top-10) candidate targets.

[0084] Hard negative example screening: Among the recalled Top-K candidates, entities with "high prediction scores but negative labels" are selected (i.e. However, these entities are not the actual entities that match the target. These entities are often very similar to the source entity literally (such as "inlet valve" and "outlet valve"), which is the main reason why the dual-tower model fails.

[0085] Structured augmentation training: Mix the aforementioned difficult negative examples with real positive samples, and concatenate the parent class and sibling class contexts as described in S4 to construct triples or classification datasets to fine-tune the Cross-Encoder.

[0086] By employing this difficult example extraction and training-specific strategy, the Cross-Encoder is forced to pay attention to structural contextual differences beyond entity names (e.g., similar names but different parent classes), thereby achieving high-precision secondary decision-making within the "fuzzy range" that S2 cannot distinguish.

[0087] Example 2 like Figure 4 The diagram shown is a block diagram of the system functional modules provided in an embodiment of the present invention. This embodiment provides a system for executing the above-described adaptive industrial ontology matching method based on multi-view structure decoupling. The system mainly includes: an unsupervised data mining module 10, a coarse recall module 20 for horizontal and vertical dual-view decoupling, an adaptive routing control module 30, and a context-aware fine-grained reordering module 40. The specific structure and function description of each module are as follows: The unsupervised data mining module 10 is configured to address the cold start problem and automatically construct the training dataset in the absence of manual annotation, specifically corresponding to step S1 in the aforementioned method embodiment. This module internally includes a text normalization unit, a rule seed generation unit, and a mutual nearest neighbor mining unit.

[0088] Its specific functions are as follows: receiving source and target ontology, generating initial seeds through fine-grained cleaning and rule matching; calculating vectors using pre-trained models, and mining implicit matching samples by verifying bidirectional Top-1 reciprocal relationships, and finally outputting an unsupervised training set.

[0089] The coarse recall module 20, which decouples the horizontal and vertical views, is configured to extract multi-dimensional entity features and calculate global similarity with low memory consumption, specifically corresponding to step S2 in the aforementioned method embodiment. This module employs a parameter freezing strategy and internally contains three view processing units: Text Semantic Encoding Unit (V1): Extracts global semantic features using an updatable master tower; Vertical phylogenetic convolutional coding unit (V2): Performs phylogenetic path projection and extracts logical reasoning features using frozen subtowers and 1D-CNN; Lateral Neighborhood Attention Encoding Unit (V3): Performs neighborhood sampling and aggregates local semantic features with frozen subtowers and GAT.

[0090] This module also includes an adversarial fusion training unit for feature fusion and adversarial perturbation generation.

[0091] The adaptive routing control module 30 is configured to dynamically allocate computing resources, specifically corresponding to step S3 in the aforementioned method embodiment. This module receives the cosine similarity score output by the coarse recall module 20 and splits entity pairs according to preset high confidence thresholds, low confidence thresholds, and competitive difference thresholds. The splitting logic is as follows: only when the highest score is higher than the high threshold and the difference between it and the second highest score is greater than the difference threshold, will a sure-match sample be directly output; sure-non-match samples will be directly filtered; otherwise, a set of ambiguous samples in the middle range or with too small a difference will be sent to the next module.

[0092] The context-aware fine-grained rearrangement module 40 is configured to perform deep disambiguation on ambiguous samples, specifically corresponding to step S4 in the aforementioned method embodiment. This module internally includes a context construction unit, a cross-coding inference unit, an optimal selection unit, and a hard example mining and training unit. Its specific functions are as follows: Context construction unit: Explicitly extract the parent and sibling class structural contexts of fuzzy samples to construct long sequences; Cross-Encoder Inference Unit: Inputs the sequence into the Cross-Encoder model and uses a deep interaction mechanism to calculate the fine-grained matching probability; Optimal matching unit: Executes ArgMax strategy and threshold filtering, and outputs the final matching result set; Hard example mining and training unit: Configures and fine-tunes a dedicated training set for building the model. This unit uses a coarse recall module with frozen parameters to retrieve the training set, selects entities with high prediction scores but negative labels as hard negative examples, and mixes them with positive samples to train the cross-encoding inference unit.

[0093] 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.

[0094] This invention proposes an adaptive industrial ontology matching method and system based on multi-view structural decoupling. The system first constructs an unsupervised training dataset through mutually beneficial nearest neighbor mining and rule constraints to address the cold-start problem. Further, it constructs a coarse recall model with horizontal and vertical dual-view decoupling, employing a parameter freezing strategy. Vertical lineage reasoning features are extracted using a multi-scale one-dimensional convolutional neural network (1D-CNN), and horizontal neighborhood semantic field features are aggregated using a graph attention mechanism (GAT). Subsequently, multi-view features are fused with global semantic features, and adversarial perturbations are introduced to generate robust entity representations. Similarity scores are calculated based on these representations, and adaptive routing is performed using dual thresholds and competitive differences to directly output confident sample results. Finally, the selected fuzzy samples and their structured context are input into a fully interactive cross-encoder optimized by a hard example mining strategy for deep semantic disambiguation and optimization, thereby achieving high-precision industrial ontology matching with low memory consumption.

[0095] 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. An adaptive industrial ontology matching method based on multi-view structural decoupling, characterized in that: The method includes the following steps: S1. Based on rule matching, an initial seed set is generated from the standardized text of the source ontology and the target ontology. A bidirectional reciprocal constraint mechanism based on vector space is established, and a pre-trained language model is used to mine the implicit matching samples that are the nearest neighbors. The initial seed set and the implicit matching samples are used together to construct a two-layer filtered unsupervised training dataset. S2. Construct a coarse recall model with horizontal and vertical dual-view decoupling, which includes a text semantic view encoding branch, a vertical lineage view encoding branch and a horizontal neighborhood view encoding branch. Extract global semantic features, vertical reasoning features and horizontal structural features respectively, and fuse multi-view features and introduce adversarial perturbation to generate a global representation of the entity. S3. Establish an adaptive route based on prediction confidence, and dynamically divide the samples into a set of sure-match samples, sure-mismatch samples, and fuzzy samples according to the double reset confidence threshold. S4. For the diverged samples that are sure to match and sure to not match, directly output the final matching result; for the diverged fuzzy sample set, explicitly extract the structured context information of the entities and input it into the fully interactive cross encoder for deep semantic disambiguation and optimization, and output the final matching result.

2. The adaptive industrial ontology matching method based on multi-view structure decoupling according to claim 1, characterized in that: In step S1, the initial seed set generation process includes: first, parsing the input source ontology and target ontology into structured data; for each entity node, extracting its name label and synonyms to form an initial concept feature set; performing fine-grained standardization cleaning on the initial concept feature set, and using a pre-trained language model to map the cleaned text into a high-dimensional dense vector; then generating pseudo-labels by combining rule matching and reciprocal nearest neighbor mining strategies, wherein an ontology parsing tool is used to traverse the input ontology file; for each entity node in the ontology... Index its main label Label and all aliases Synonym to construct an entity text set. ; For entity text collections The standardized text cleaning process includes phrase segmentation, noise filtering, and normalization, ultimately forming a standardized text feature set. Then, based on this standardized text feature set, a string matching strategy is executed, traversing entity pairs between the source and target ontologies and calculating the string overlap of their standardized names. If the names of two entities or any synonyms achieve a complete string match, they are classified as high-confidence positive samples and directly included in the initial seed set. This yields the initial seed set. The process of mining implicit matching samples and constructing a two-layer filtered unsupervised training dataset is as follows: using a pre-trained language model to encode the standardized entity text into high-dimensional semantic vectors. Construct the source entity vector matrix With the target entity vector matrix And calculate the cosine similarity matrix between the two. Then execute the two-way reciprocity verification logic: only if the target entity It is the source entity The item with the highest similarity in the target space, and also the source entity Also the target entity The pair with the highest similarity in the source space, and whose similarity is greater than the high confidence threshold, satisfies the two-way reciprocity constraint. Then, entity pairs satisfying the following two-way reciprocity constraint are selected. Add to the training set, i.e.: In the formula, Represents the source entity With the target entity Similarity between them Represents the source entity With each entity in the target space Similarity; Represents the target entity With each entity in the source space similarity, High confidence threshold.

3. The adaptive industrial ontology matching method based on multi-view structure decoupling according to claim 2, characterized in that: In the text semantic view encoding branch of step S2, the global semantic features of the central entity are extracted using the parameter-updatable first semantic encoding branch. The standardized name sequence is input into the BERT-Base model, and the vector at the last layer [CLS] position is extracted as the global semantic feature. During this process, the parameters of the first semantic encoding branch remain unfrozen and are updated with the training gradient to capture the core lexical meaning of entities.

4. The adaptive industrial ontology matching method based on multi-view structure decoupling according to claim 3, characterized in that: In the longitudinal phylogenetic view encoding branch of step S2, a second semantic encoding branch with parameter freezing is used to capture the logical reasoning features from the root node to the leaf node, including phylogenetic path projection, feature vectorization, multi-scale convolutional encoding, and pooling output processes. Lineage path projection: for any entity in the ontology By backtracking to the root node using the shortest path algorithm, a unique lineage path sequence is constructed. : In the formula, For the current entity, For the first Skip to ancestor node, This refers to the hierarchical relationship type between connecting nodes; Feature vectorization: The semantic vectors of each node in the path are extracted using the second semantic encoding branch with parameter freezing, and then reduced to dimensionality through a linear layer. Simultaneously, the relation type is mapped to an embedding vector, and the node vector and relation vector are concatenated in sequence to form a path feature matrix. ; Multi-scale convolutional coding: Introducing a one-dimensional convolutional neural network, utilizing a set of convolutional kernels. For the characteristic matrix Perform sliding convolution operation to extract local inference features. The convolution calculation method is as follows: in, The size of the convolution window corresponds to the single-hop inheritance relationship and the double-hop inference logic, respectively. Represents a local segment of a path sequence; For bias terms; Pooling output: Global max pooling is performed on the feature map generated by convolution to obtain the vertical structure feature vector. .

5. The adaptive industrial ontology matching method based on multi-view structure decoupling according to claim 4, characterized in that: In the lateral neighborhood view encoding branch of step S2, the semantic field features of the local neighborhood are aggregated using the parameter-frozen second semantic encoding branch, including isomorphic neighborhood sampling, attention aggregation, and feature weighting processes. Homogenization neighborhood sampling selects entities The direct parent node and its sibling nodes constitute the neighborhood set. ; Attention aggregation extracts neighbor nodes from secondary towers using parameter freezing. Text features and target entity Text features This involves introducing a graph attention mechanism, which calculates attention weights based on the semantic relevance between the target entity and its neighboring nodes. : in, It is a learnable linear transformation matrix. For learnable attention vectors, This represents a vector concatenation operation. It is a non-linear activation function; Finally, the lateral structural feature vector is obtained by weighted summation. .

6. The adaptive industrial ontology matching method based on multi-view structural decoupling according to claim 5, characterized in that: The process of fusing global semantic features, vertical reasoning features, and horizontal structural features in step S2 to generate a global representation is as follows: Feature fusion: combining global semantic features Vertical structural features and lateral structural features Dimensional concatenation is performed, and features are fused through a learnable linear mapping layer to obtain the final multi-view fused feature vector of the entity. The calculation formula is as follows: in, This represents a vector concatenation operation; and These are the weight matrix and bias term of the fusion layer, respectively; Adversarial perturbation introduction: Calculating the total loss function Regarding the input layer embedding matrix gradient Generate bounded adversarial perturbations based on gradient direction : in This is the preset perturbation amplitude hyperparameter; Adversarial optimization: perturbation Superimposed onto the original embedding matrix, adversarial examples are constructed, and the optimization objective is to minimize the loss in the worst case: Iterative training is performed on a coarse recall model with decoupled horizontal and vertical views. The model parameters are updated using the gradient descent algorithm until the total loss function converges or the preset maximum number of iterations is reached. After training, the final multi-view fusion feature vector is output. As a unified representation of the source entity and the target entity in subsequent steps, they are respectively denoted as... and .

7. The adaptive industrial ontology matching method based on multi-view structural decoupling according to claim 6, characterized in that: In step S3, an adaptive routing mechanism is used to split the coarse recall results, wherein the unified representation of the source entity and the target entity is used. and Calculate cosine similarity score : in, Represents the vector dot product. The L2 norm of a vector; Establish a dual-reset confidence threshold and triage the data, including a high-confidence threshold. and low confidence threshold and competitive difference threshold The diversion process is as follows: like and If the sample is determined to be a sure match, the matching result is output directly. like If the sample is determined to be a mismatch, it will be filtered out directly. like If the sample is a numerically fuzzy sample, then the sample is determined to be a numerically fuzzy sample; if > But the difference If it is, then it is judged as a competing fuzzy sample. The highest similarity score is given. It received the second-highest similarity score.

8. The adaptive industrial ontology matching method based on multi-view structural decoupling according to claim 7, characterized in that: In step S4, the input and output processes of the fully interactive cross-encoder model are as follows: For each pair of entities to be decided in the set Explicitly extract the parent class name and sibling class name of the entity; The entity names and their structured contexts are concatenated into a long sequence using delimiters specific to the pre-trained model. The formula is constructed as follows: in, This represents a sequence concatenation operation; The name text representing the entity; The structural context text representing the entity, ; , Used as a separator; The constructed long sequence The input fully interactive cross-encoder model is processed as follows: First, extract the hidden state vector at the [CLS] position of the last layer of the model. As a semantic aggregation representation of the entire sequence: Then, the vector is fed into a fully connected layer, and the output is mapped to a sigmoid activation function. Interval matching probability : Finally, the best candidate is selected, and the source entity is calculated. The probability set of all corresponding candidate target entities Execute the ArgMax optimization strategy: Select the target entity with the highest probability value As the optimal matching candidate: If the target entity probability value If the result is greater than the preset value, the final matching result will be output. Otherwise, determine the source entity. No valid match found.

9. The adaptive industrial ontology matching method based on multi-view structure decoupling according to claim 8, characterized in that: In step S4, during the training of the fully interactive cross-encoder model, a dedicated training set is constructed using a hard example mining strategy based on coarse recall results, and training is performed, including: Static hard example generation: Using a coarse recall model that has been trained and converged and whose parameters are frozen, a full database search is performed on each source entity in the training set to recall the Top-K candidate targets; Hard-to-bearing examples screening: Among the recalled Top-K candidates, entities with probability values ​​higher than the preset value but negative true labels are screened out and marked as hard-to-bearing examples; Structured augmentation training: Mix hard negative examples with real positive samples, and construct parent and sibling class contexts according to the input sequence construction method to build triples or classification datasets for fine-tuning training of the fully interactive cross-encoder model.

10. A system for executing the adaptive industrial ontology matching method based on multi-view structural decoupling as described in any one of claims 1-9, characterized in that: The system includes: The unsupervised data mining module is configured to solve the cold start problem and automatically build training sets, including a rule seed generation unit and a mutual nearest neighbor mining unit. The coarse recall module with horizontal and vertical dual-view decoupling is configured to extract entity features and calculate global similarity, and includes a text semantic encoding unit, a vertical phylogenetic convolutional encoding unit, a horizontal neighborhood attention encoding unit, and an adversarial fusion training unit. The adaptive routing control module is configured to divide entity pairs into a set of sure-match, sure-mismatch, and fuzzy samples based on dual thresholds, and distribute the fuzzy samples to the next module. The context-aware fine-grained rearrangement module is configured for deep disambiguation of fuzzy samples and includes a context construction unit and a cross-coding inference unit.

Images