Large-scale knowledge fusion system and method based on ontology mapping and conflict resolution

Abstract

The invention discloses a large-scale knowledge fusion system based on ontology mapping and conflict resolution. The large-scale knowledge fusion system comprises an ontology standardization module, an ontology analysis module and an ontology analysis module, the multi-strategy ontology mapping module extracts the multi-dimensional features of the standardized ontology, calculates the similarity, and generates a comprehensive similarity matrix in combination with the dynamic weight to obtain an ontology mapping rule; the intelligent entity alignment module extracts multi-dimensional features of entities, fuses the multi-dimensional features to generate a joint entity embedding vector, screens candidate entity pairs and then obtains an entity alignment result through a classifier; the deep conflict detection and resolution module detects conflicts in combination with entity attribute values and optimizes and resolves the conflicts through semantic reasoning, structure assistance and reinforcement learning to obtain an optimal conflict resolution scheme; the knowledge graph fusion module obtains a fusion knowledge graph through distributed optimization, and carries out incremental updating to obtain an updated fusion knowledge graph. According to the invention, the precision, efficiency, intelligence and expandability of large-scale knowledge fusion are improved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and knowledge engineering technology, specifically to a large-scale knowledge fusion system and method based on ontology mapping and conflict resolution. Background Technology

[0002] With the rapid development of artificial intelligence and knowledge engineering technologies, knowledge graphs, as the core carrier of structured knowledge, have been widely used in numerous fields. The fusion of multi-source heterogeneous knowledge graphs is a crucial step in breaking down data barriers and building a unified knowledge foundation. Its core objective is to integrate knowledge graphs from different sources, with different structures and semantic expressions, into a consistent and complete knowledge representation, providing reliable support for downstream analytical applications. The core tasks of knowledge fusion include ontology mapping (concept and attribute alignment), entity alignment (instance-level fusion), and conflict resolution (handling contradictory attribute values), which are prerequisites for realizing knowledge reuse and value mining.

[0003] Typical knowledge fusion techniques include rule-based methods, dictionary-based matching methods, embedding vector-based methods, and deep learning-based end-to-end alignment methods. Compared to other methods, embedding vector-based methods and deep learning methods offer some improvement in automation, but still have significant limitations: traditional ontology mapping methods (such as the Levenshtein algorithm based on string edit distance and the structure-based Graph Match algorithm) rely on shallow features and struggle to capture deep semantic relationships; conflict resolution often relies on manual rule bases, resulting in poor flexibility and high maintenance costs; the computational complexity of existing methods increases exponentially with the size of the knowledge graph, making it difficult to meet real-time fusion requirements in large-scale scenarios with hundreds of millions of nodes; automation is low, failing to deeply integrate machine learning, deep learning, and large language model capabilities, leading to high costs for manual intervention; and the lack of incremental update mechanisms means that new knowledge sources require full recalculation, limiting practicality and efficiency. Summary of the Invention

[0004] The purpose of this invention is to provide a large-scale knowledge fusion system and method based on ontology mapping and conflict resolution. This invention improves the accuracy, efficiency, intelligence and scalability of large-scale knowledge fusion.

[0005] To achieve this objective, the present invention designs a large-scale knowledge fusion system based on ontology mapping and conflict resolution, which includes: The ontology normalization module is used to normalize the ontology of the multi-source heterogeneous knowledge graph to be fused, so as to obtain a normalized ontology. The multi-strategy ontology mapping module is used to extract multi-dimensional features of the normalized ontology, calculate the similarity of the multi-dimensional features respectively, generate a comprehensive similarity matrix by combining dynamic weights, and perform ontology mapping on the comprehensive similarity matrix to obtain ontology mapping rules. The intelligent entity alignment module is used to extract multi-dimensional features of each entity based on the normalized ontology and the ontology mapping rules, and to fuse the multi-dimensional features of each entity to generate a corresponding joint entity embedding vector. By calculating the similarity between any two joint entity embedding vectors, the top K similarity-ranked entity pairs are selected as candidate entity pairs. The candidate entity pairs are then classified by a machine learning classifier to obtain the entity alignment result. The deep conflict detection and resolution module is used to detect conflicts based on the ontology mapping rules and the entity alignment results, combined with entity attribute values, and then resolve conflicts through semantic reasoning, structural assistance and reinforcement learning optimization to obtain the optimal conflict resolution scheme and the normalized entity attribute values ​​after conflict resolution. The knowledge graph fusion module is used to obtain a fused knowledge graph through distributed optimization based on the normalized ontology, the ontology mapping rules, the entity alignment results, the optimal conflict resolution scheme, and the normalized entity attribute values, and to perform incremental updates on the fused knowledge graph to obtain an updated fused knowledge graph.

[0006] The preferred method for standardizing the ontology of the multi-source heterogeneous knowledge graph to be fused is as follows: The ontology of the multi-source heterogeneous knowledge graph to be integrated is standardized, including format unification, namespace standardization and terminology clarification; Among them, format unification is to convert the ontology of the multi-source heterogeneous knowledge graph to be integrated into a standard ontology language format; Namespace standardization is to unify the namespace of the ontology of all multi-source heterogeneous knowledge graphs to be integrated into a global namespace; Terminology clarification identifies ambiguous and unclear terms in the ontology after standardization of format. It calls a large language model to construct a contextual description that includes term definitions, relationships, and application scenarios. The input model generates standard semantic labels, and the output is an ontology with standardized terms.

[0007] Preferably, the process of extracting multi-dimensional features from the normalized ontology, calculating the similarity of each multi-dimensional feature, generating a comprehensive similarity matrix by combining dynamic weights, and performing ontology mapping on the comprehensive similarity matrix to obtain the ontology mapping rules is as follows: Feature extraction is performed on the normalized ontology to extract semantic features, structural features, and instance features. The semantic similarity of the semantic features is calculated using the cosine similarity algorithm. : in, Representing a normalized ontology The attributes, Representing a normalized ontology The attributes, Representing a normalized ontology The concept, Representing a normalized ontology The concept, Representing a normalized ontology semantic embedding vector, Representing a normalized ontology semantic embedding vector; The norm symbol is used to ensure that the similarity value is in the range of [-1, 1]. The structural similarity of the structural features is calculated using the cosine similarity algorithm. : in, Representing a normalized ontology The structural embedding vector, Representing a normalized ontology The structural embedding vector; The instance similarity of the instance features is calculated using the cosine similarity algorithm. : in, Representing a normalized ontology Concept The corresponding instance feature matrix, Representing a normalized ontology Concept The corresponding instance feature matrix; Based on semantic similarity Structural similarity Similarity to instances The overall similarity is calculated through dynamic weight fusion. : in, Weights for semantic similarity The weights for structural similarity, The weights for instance similarity; A comprehensive similarity matrix is ​​generated based on the comprehensive similarity between different normalized ontologies; Set a comprehensive similarity threshold, select mapping pairs in the comprehensive similarity matrix whose comprehensive similarity is not lower than the comprehensive similarity threshold as candidate mapping pairs, and generate ontology mapping rules after logically verifying the candidate mapping pairs.

[0008] Preferably, the specific process for obtaining the entity alignment result is as follows: Based on the standardized ontology and the ontology mapping rules, multi-dimensional features of entities are extracted, including attribute features, text description features and graph structure features. A graph neural network is used to iteratively aggregate the features of entity neighbor nodes to generate structural embeddings. A pre-trained language model is used to encode the text description features to generate text embeddings. An attention fusion strategy is used to fuse the structural embeddings and text embeddings to obtain a unified joint entity embedding vector. Calculate the cosine similarity of the joint entity embedding vectors of different entities in the multi-source heterogeneous knowledge graph to be fused. : in, Representing entities The joint entity embedding vector, Representing entities The joint entity embedding vector; The top K entity pairs with the highest cosine similarity are selected as candidate entity pairs. A machine learning classifier is used to perform binary classification on the candidate entity pairs to obtain the entity alignment result.

[0009] Preferably, the specific process of conflict detection based on the ontology mapping rules and the entity alignment results, combined with entity attribute values, is as follows: Based on the ontology mapping rules and the entity alignment results, the association relationship between the aligned entities in the multi-source heterogeneous knowledge graph to be fused is established. The entity attribute values ​​of the associated entities are traversed, and the differences in the values ​​of the same attribute of the same entity in different knowledge graphs are compared to identify the conflict detection results. The conflict detection results include type conflict, numerical conflict and logical conflict.

[0010] Preferably, the specific process of resolving conflicts through semantic reasoning, structural assistance, and reinforcement learning optimization to obtain the optimal conflict resolution scheme and the normalized entity attribute values ​​after conflict resolution is as follows: Conflict resolution is achieved through semantic reasoning, structural assistance, and reinforcement learning optimization. Semantic reasoning refers to converting conflict scenarios into structured natural language descriptions that include conflict entities, conflict attributes, KG attribute values, and knowledge source background. The large language model uses these structured natural language descriptions to generate conflict resolution decision logic. Structural assistance refers to analyzing the topological structure of conflict attributes in the ontology layer of the multi-source heterogeneous knowledge graph to be fused through graph neural networks, and using topological consistency to evaluate the rationality of the conflict resolution decision logic; Reinforcement learning optimization refers to constructing a dynamic weight decision model based on reinforcement learning, using the DPO algorithm, and combining conflict resolution decision logic, the topological structure, and the authority of the knowledge source to dynamically adjust the weights of each evaluation dimension, thereby obtaining the optimal conflict resolution scheme and the normalized entity attribute values ​​after conflict resolution.

[0011] Preferably, the specific process for obtaining the normalized entity attribute values ​​after conflict resolution is as follows: The entity attribute values ​​after conflict resolution are formatted and standardized by using a deep learning model to eliminate format differences and obtain entity attribute values ​​in a standard format. The entity attribute values ​​in the standard format are numerically converted. Based on the domain knowledge base, the unit conversion and precision unification of the standard format entity attribute values ​​are performed to obtain the numerically converted entity attribute values. The formula is as follows: in, For transformation functions in the domain knowledge base, For entity attribute values ​​in a standard format, To unify the target units, To standardize target accuracy, These are the entity attribute values ​​after numerical conversion; Text standardization is performed on the entity attribute values ​​after numerical conversion. Regular expressions and large language models are used to correct text errors and inconsistencies in the entity attribute values ​​after numerical conversion, and finally, the standardized entity attribute values ​​after conflict resolution are obtained.

[0012] Preferably, based on the normalized ontology, the ontology mapping rules, the entity alignment results, the optimal conflict resolution scheme, and the normalized entity attribute values, a fused knowledge graph is obtained through distributed optimization. The specific process of incrementally updating the fused knowledge graph to obtain the updated fused knowledge graph is as follows: Based on the normalized ontology, the ontology mapping rules, the entity alignment results, the optimal conflict resolution scheme, and the normalized entity attribute values, distributed optimization is performed. The tasks of ontology mapping rules, entity alignment results, optimal conflict resolution scheme, and normalized entity attribute values ​​are split into corresponding independent partitions for parallel processing. Container orchestration technology is used, and resources are allocated to each independent partition according to task type through a large language model inference cluster, a graph neural network computing cluster, and a general computing cluster. Elastic distributed datasets and vector databases are used to store the intermediate results and embedding vectors of each independent partition, ultimately resulting in a fused knowledge graph. Incremental updates are performed on the fused knowledge graph. By comparing the differences in structure, concepts, entities, and attributes between the newly added knowledge sources and the fused knowledge graph using a version hash algorithm, the added, modified, or deleted modules are located. Ontology mapping, entity alignment, and conflict resolution are re-executed only on the differing modules to obtain local update results. The local update results are then integrated into the fused knowledge graph through consistency verification to obtain the updated fused knowledge graph.

[0013] A large-scale knowledge fusion method based on ontology mapping and conflict resolution includes the following steps: The ontology of the multi-source heterogeneous knowledge graph to be integrated is normalized to obtain a normalized ontology; Extract the multi-dimensional features of the normalized ontology, calculate the similarity of the multi-dimensional features respectively, generate a comprehensive similarity matrix by combining dynamic weights, and perform ontology mapping on the comprehensive similarity matrix to obtain the ontology mapping rules. Based on the normalized ontology and the ontology mapping rules, multi-dimensional features of each entity are extracted, and the multi-dimensional features of each entity are fused to generate corresponding joint entity embedding vectors. By calculating the similarity between any two joint entity embedding vectors, the top K similarity pairs are selected as candidate entity pairs. The candidate entity pairs are then classified by a machine learning classifier to obtain the entity alignment result. Based on the ontology mapping rules and the entity alignment results, conflict detection is performed by combining entity attribute values. Then, conflict resolution is performed through semantic reasoning, structural assistance, and reinforcement learning optimization to obtain the optimal conflict resolution scheme and the normalized entity attribute values ​​after conflict resolution. Based on the normalized ontology, the ontology mapping rules, the entity alignment results, the optimal conflict resolution scheme, and the normalized entity attribute values, a fused knowledge graph is obtained through distributed optimization. The fused knowledge graph is then incrementally updated to obtain an updated fused knowledge graph.

[0014] A computer program product includes a computer program that, when executed by a processor, implements the steps of the above-described method.

[0015] The beneficial effects of this invention are: This invention improves the accuracy, efficiency, intelligence, and scalability of large-scale knowledge fusion through multi-strategy feature fusion, intelligent model collaboration, and distributed architecture optimization. By combining multi-strategy fusion with logical verification of a large language model, leveraging its hybrid expert architecture and long contextual capabilities, it effectively captures deep semantic relationships and solves the problem of mismatched alignment of cross-domain concepts and attributes. Combined with semantic reasoning and reinforcement learning optimization of the large language model, it automates the handling of complex conflicts, ensuring the transparency and traceability of decisions. Based on a distributed computing architecture and parallel task processing mechanism, this invention significantly reduces computational complexity, meeting the real-time requirements of large-scale knowledge fusion applications. Furthermore, through an incremental update mechanism, this invention achieves dynamic access and partial updates of new knowledge sources, avoiding full recalculation, improving the adaptability to dynamic knowledge scenarios, and reducing long-term maintenance costs. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the structure of the present invention; Figure 2 This is a flowchart of the present invention. Detailed Implementation

[0017] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: Example 1 A large-scale knowledge fusion system based on ontology mapping and conflict resolution, such as Figure 1 As shown, it includes: The ontology normalization module is used to normalize the ontology (ontology refers to the semantic framework of a knowledge graph, which defines the concepts, attributes and relationships between concepts in the knowledge graph) of the multi-source heterogeneous knowledge graph to be fused (multi-source heterogeneous knowledge graph refers to a collection of multiple knowledge graphs with different sources, structures and semantic expressions) to obtain a normalized ontology. This design solves the data heterogeneity problem from the source by transforming the ontology of multi-source heterogeneous knowledge graph into a unified standard normalized ontology, avoiding errors or inefficient processing caused by inconsistent basic data, and improving the overall fusion efficiency and accuracy. The multi-strategy ontology mapping module is used to extract multi-dimensional features of the normalized ontology, calculate the similarity of the multi-dimensional features respectively, generate a comprehensive similarity matrix by combining dynamic weights, and perform ontology mapping on the comprehensive similarity matrix to obtain (high confidence) ontology mapping rules. This design can improve the reliability and accuracy of the generated ontology mapping rules by finding the association between different normalized ontology. The intelligent entity alignment module is used to extract multi-dimensional features of each entity (an entity refers to a specific object, thing, or instance described in a knowledge graph; an entity is a concretization of a concept defined in an ontology) based on the normalized ontology and the ontology mapping rules. It then fuses the multi-dimensional features of each entity to generate a corresponding joint entity embedding vector. By calculating the similarity between any two joint entity embedding vectors, the top K similarity pairs are selected as candidate entity pairs. A machine learning classifier performs binary classification on the candidate entity pairs to obtain the entity alignment result. This design avoids the limitations of single features by extracting multi-dimensional features of each entity; the binary classification using a machine learning classifier achieves high recall while filtering out erroneous candidate entity pairs, thus improving precision. The deep conflict detection and resolution module is used to detect conflicts based on the ontology mapping rules and the entity alignment results, combined with entity attribute values ​​(entity attribute values ​​refer to the specific values ​​of a certain attribute of an entity in the knowledge graph). Then, it performs conflict resolution through semantic reasoning, structural assistance, and reinforcement learning optimization to obtain the optimal conflict resolution scheme and the normalized entity attribute values ​​after conflict resolution. This design achieves comprehensive coverage of conflicts through conflict detection, fully identifies various conflicts of entity attributes in multi-source knowledge graphs, generates the optimal conflict resolution scheme, and obtains normalized entity attribute values. It solves the data inconsistency problem and ensures the accuracy and reliability of the fused knowledge graph. The knowledge graph fusion module is used to obtain a fused knowledge graph through distributed optimization based on the normalized ontology, the ontology mapping rules, the entity alignment results, the optimal conflict resolution scheme, and the normalized entity attribute values. The fused knowledge graph is then incrementally updated to obtain an updated fused knowledge graph. This design efficiently integrates multi-source heterogeneous knowledge graphs, generates a complete and consistent fused knowledge graph through distributed optimization, and supports incremental updates, thereby improving the efficiency and timeliness of knowledge fusion and meeting the needs of large-scale dynamic knowledge fusion.

[0018] In the above technical solution, the specific method for standardizing the ontology of the multi-source heterogeneous knowledge graph to be fused is as follows: The ontology of the multi-source heterogeneous knowledge graph to be integrated is standardized, including format unification, namespace standardization and terminology clarification; Among them, format unification is to convert the ontology of the multi-source heterogeneous knowledge graph to be integrated into a standard ontology language format; Namespace standardization is to unify the namespace of the ontology of all multi-source heterogeneous knowledge graphs to be integrated into a global namespace; Terminology clarification identifies ambiguous and unclear terms in the ontology after format unification. It calls a large language model to construct a contextual description that includes term definitions, relationships, and application scenarios. The input model generates standard semantic tags, and the output is an ontology with standardized terms. The above design converts the ontology of multi-source heterogeneous knowledge graphs into a standard ontology language format through format unification, ensuring the consistency of data format. The namespace is standardized and unified into a global namespace to avoid naming conflicts. Terminology clarification resolves term ambiguity by generating standard semantic tags through a large language model.

[0019] In the above technical solution, the process of extracting multi-dimensional features of the normalized ontology, calculating the similarity of each multi-dimensional feature, generating a comprehensive similarity matrix by combining dynamic weights, and performing ontology mapping on the comprehensive similarity matrix to obtain the ontology mapping rules is as follows: Feature extraction is performed on the normalized ontology to extract semantic features, structural features, and instance features. The semantic similarity of the semantic features is calculated using the cosine similarity algorithm. : in, Representing a normalized ontology The attributes, Representing a normalized ontology The attributes, Representing a normalized ontology The concept, Representing a normalized ontology The concept, Representing a normalized ontology semantic embedding vector, Representing a normalized ontology semantic embedding vector; The structural similarity of the structural features is calculated using the cosine similarity algorithm. : in, Representing a normalized ontology The structural embedding vector, Representing a normalized ontology The structural embedding vector; The instance similarity of the instance features is calculated using the cosine similarity algorithm. : in, Representing a normalized ontology Concept The corresponding instance feature matrix, Representing a normalized ontology Concept The corresponding instance feature matrix; Based on semantic similarity Structural similarity Similarity to instances The overall similarity is calculated through dynamic weight fusion. : in, Weights for semantic similarity The weights for structural similarity, The weights for instance similarity; A comprehensive similarity matrix is ​​generated based on the comprehensive similarity between different normalized ontologies; A comprehensive similarity threshold is set (ranging from 0.5 to 0.8). Mapping pairs with a comprehensive similarity score not lower than the threshold are selected as high-confidence candidate mapping pairs. Logical verification is performed on the candidate mapping pairs to generate verification results. The verification results include three categories: "reasonable", "unreasonable", and "awaiting manual review". Candidate mapping pairs with reasonable verification results are retained to generate ontology mapping rules. The above design extracts semantic features, structural features, and instance features, covering the core attributes (semantics), organizational form (structure), and actual application scenarios (instances) of the ontology. Then, the cosine similarity algorithm is used to calculate the similarity of each dimension, and a comprehensive similarity matrix is ​​generated by combining dynamic weights, which is more comprehensive than single-dimensional mapping. By setting a comprehensive similarity threshold and logical verification to select candidate mapping pairs, invalid or erroneous mappings are effectively filtered out, improving the reliability of the generated ontology mapping rules.

[0020] In the above technical solution, the specific process of obtaining the entity alignment result is as follows: Based on the standardized ontology and the ontology mapping rules, multi-dimensional features of entities are extracted, including attribute features (the types of attribute features include text, numeric, date, and boolean), text description features, and graph structure features. A graph neural network (Graph Sampling and Aggregate) is used to iteratively aggregate the features of the entity's neighbor nodes to generate structural embeddings. A pre-trained language model is used to encode the text description features to generate text embeddings. An attention fusion strategy is used to fuse the structural embeddings and text embeddings to obtain a unified joint entity embedding vector. Calculate the cosine similarity of the joint entity embedding vectors of different entities in the multi-source heterogeneous knowledge graph to be fused. : in, Representing entities The joint entity embedding vector, Representing entities The joint entity embedding vector; The top K entity pairs with the highest cosine similarity are selected as candidate entity pairs. A machine learning classifier performs binary classification on these candidate entity pairs, outputting a probability value that indicates the confidence level of the match. By setting a threshold (e.g., 0.5), the probability value is converted into a binary classification result: 1 (match) or 0 (no match). All candidate entity pairs judged as 1 by the machine learning classifier are considered aligned entities, thus obtaining the final entity alignment result. The above design clarifies the specific types of multi-dimensional features of entities (attributes, text descriptions, graph structures). By filtering and classifying candidate entity pairs, erroneous candidate entity pairs can be filtered out, improving the accuracy of the entity alignment result.

[0021] In the above technical solution, the specific process of conflict detection based on the ontology mapping rules and the entity alignment results, combined with entity attribute values, is as follows: Based on the ontology mapping rules and the entity alignment results, the association relationships between aligned entities in the multi-source heterogeneous knowledge graph to be fused are established. The entity attribute values ​​of the associated entities are traversed, and the differences in the values ​​of the same attribute of the same entity in different knowledge graphs are compared to identify conflict detection results. The conflict detection results include type conflicts (inconsistent attribute value format or data type), numerical conflicts (attribute values ​​are numerical but the units of value are different), and logical conflicts (logical contradictions exist between attribute values). The above design, by establishing the association relationships between associated entities and by traversing attribute values ​​and comparing the differences in the values ​​of the same attribute, accurately identifies the three types of conflicts: type, numerical, and logical. It clarifies the specific types of conflicts, avoids the omission or misjudgment of conflicts, and can improve the efficiency and effectiveness of subsequent conflict resolution.

[0022] In the above technical solution, the specific process of resolving conflicts through semantic reasoning, structural assistance, and reinforcement learning optimization to obtain the optimal conflict resolution scheme and the normalized entity attribute values ​​after conflict resolution is as follows: Conflict resolution is achieved through semantic reasoning, structural assistance, and reinforcement learning optimization. Semantic reasoning refers to converting conflict scenarios into structured natural language descriptions that include conflicting entities (conflicting entities are entities that are confirmed to have at least one conflicting attribute after conflict detection), conflicting attributes (conflicting attributes are detected attributes that have different values ​​for the same conflicting entity in different knowledge graphs), KG attribute values ​​(KG attribute values ​​are the set of specific values ​​corresponding to the same conflicting attribute of the same entity in different knowledge graphs in the conflict resolution scenario), and knowledge source background. The large language model (using GLM-4.5) uses the structured natural language descriptions to generate conflict resolution decision logic. Structural assistance refers to analyzing the topological structure of conflict attributes in the ontology layer of the multi-source heterogeneous knowledge graph to be fused through graph neural networks, and using topological consistency to evaluate the rationality of the conflict resolution decision logic; Reinforcement learning optimization involves first identifying three core evaluation dimensions: logical semantic compatibility of conflict resolution decisions, topological consistency, and knowledge source authority. Then, a dynamic weighted decision-making model based on reinforcement learning is constructed. The quantified values ​​of these three core evaluation dimensions serve as the model's input, and adjusting the weight ratios of each dimension is the dynamic weighted decision-making model's action. The optimization objective is to improve the adaptability of the conflict resolution results, employing the DPO algorithm (Direct Preference Optimization). The optimization process completes the training of the dynamic weight decision model. Specifically, it involves collecting high-quality and low-quality decision samples from historical conflict resolution cases, labeling the weight allocation preferences for each evaluation dimension in the samples, using the constructed reinforcement learning model as the base model, maximizing the log probability difference between high-quality and low-quality decisions through the DPO algorithm, optimizing the parameters of the dynamic weight decision model, enabling the dynamic weight decision model to learn weight adjustment logic that conforms to the domain scenario, and finally, combining the conflict resolution decision logic, the aforementioned topological structure, and the authority of the knowledge source, and dynamically adjusting the weight allocation ratio of each evaluation dimension based on the current conflict scenario characteristics, and finally obtaining the optimal conflict resolution scheme and the normalized entity attribute values ​​after conflict resolution through weighted scoring. This design clarifies the collaborative resolution logic of semantic reasoning, structural assistance, and reinforcement learning optimization, avoiding the limitations of a single technology, ensuring the rationality and optimality of the resolution scheme, effectively solving the inconsistency problem of multi-source data, ensuring the data quality of the fused knowledge graph, and improving the data consistency of the fused knowledge graph.

[0023] In the above technical solution, the specific process of obtaining the normalized entity attribute values ​​after conflict resolution is as follows: The entity attribute values ​​after conflict resolution are formatted and standardized by using a deep learning model (a Seq2Seq deep learning model based on the Transformer architecture) to eliminate format differences and obtain entity attribute values ​​in a standard format. The entity attribute values ​​in the standard format are numerically converted. Based on the domain knowledge base, the unit conversion and precision unification of the standard format entity attribute values ​​are performed to obtain the numerically converted entity attribute values. The formula is as follows: in, For transformation functions in the domain knowledge base, For entity attribute values ​​in a standard format, To unify the target unit (a unified unit determined according to industry standards or user-preset rules in the application field). To standardize target precision (using a unified unit determined according to industry standards or user-preset rules in the application field). These are the entity attribute values ​​after numerical conversion; Text standardization is performed on the entity attribute values ​​after numerical conversion. Regular expressions and large language models are used to correct text errors and inconsistencies in the entity attribute values ​​after numerical conversion, and finally, the standardized entity attribute values ​​after conflict resolution are obtained. The above design, through format unification, numerical conversion and text standardization, standardizes the entity attribute values ​​after conflict resolution, which can improve the data standardization and consistency of the integrated knowledge graph.

[0024] In the above technical solution, based on the normalized ontology, the ontology mapping rules, the entity alignment results, the optimal conflict resolution scheme, and the normalized entity attribute values, a fused knowledge graph is obtained through distributed optimization. The specific process of incrementally updating the fused knowledge graph to obtain the updated fused knowledge graph is as follows: Based on the normalized ontology, the ontology mapping rules, the entity alignment results, the optimal conflict resolution scheme, and the normalized entity attribute values, distributed optimization is performed. The tasks of ontology mapping rules, entity alignment results, optimal conflict resolution scheme, and normalized entity attribute values ​​are split into corresponding independent partitions for parallel processing. Container orchestration technology is adopted, and each independent partition is allocated to a corresponding dedicated computing cluster and configured with appropriate resources according to the task type through a large language model inference cluster, a graph neural network computing cluster, and a general computing cluster. An elastic distributed dataset and a vector database are used to store the intermediate results of each independent partition and cache the embedding vectors of concepts and entities, ultimately resulting in a fused knowledge graph. Incremental updates are performed on the fused knowledge graph. A version hash algorithm is used to compare the structural, conceptual, entity, and attribute differences between newly added knowledge sources and the fused knowledge graph. The specific process of comparison using the version hash algorithm is as follows: Hash values ​​are calculated for the core information of the structural, conceptual, entity, and attribute layers of both types of knowledge graphs. Then, the hash values ​​at each corresponding level are compared one by one. Hash values ​​existing only in newly added knowledge sources are identified as new elements; elements with the same unique identifier but different hash values ​​are identified as modified elements; and hash values ​​existing only in the existing fused knowledge graph are identified as deleted elements. Finally, these elements are aggregated into difference modules according to their relationships, completing the difference localization, which is crucial for subsequent partial updates and final updates. The resulting updated fused knowledge graph provides accurate object location for newly added, modified, or deleted modules; it re-executes ontology mapping, entity alignment, and conflict resolution only on the differing modules to obtain partial update results; and it integrates the partial update results into the fused knowledge graph through consistency verification to obtain the updated fused knowledge graph. The above-mentioned distributed optimization strategy improves the processing speed of large-scale knowledge fusion through task splitting, parallel processing, dedicated cluster allocation, and efficient storage; incremental updates locate differing modules through version hashing algorithms, processing only the differing parts, avoiding the resource waste of full updates, ensuring that the fused knowledge graph can absorb new knowledge in a timely manner, and guaranteeing update consistency.

[0025] Example 2 A large-scale knowledge fusion method based on ontology mapping and conflict resolution, such as Figure 2 As shown, the ontology of the multi-source heterogeneous knowledge graph is normalized; multi-dimensional features of the normalized ontology are extracted, similarity is calculated, and a comprehensive similarity matrix is ​​generated by combining dynamic weights to obtain the ontology mapping rules; multi-dimensional features of entities are extracted and fused to generate joint entity embedding vectors, and after screening candidate entity pairs, entity alignment results are obtained through a classifier; conflicts are detected by deep integration of entity attribute values, and conflicts are resolved through semantic reasoning, structural assistance, and reinforcement learning to obtain the optimal conflict resolution scheme; a fused knowledge graph is obtained through distributed optimization, and incremental updates are performed to obtain the updated fused knowledge graph.

[0026] The specific methods for large-scale knowledge fusion include the following steps: The ontology of the multi-source heterogeneous knowledge graph to be integrated is normalized to obtain a normalized ontology; Extract the multi-dimensional features of the normalized ontology, calculate the similarity of the multi-dimensional features respectively, generate a comprehensive similarity matrix by combining dynamic weights, and perform ontology mapping on the comprehensive similarity matrix to obtain the ontology mapping rules. Based on the normalized ontology and the ontology mapping rules, multi-dimensional features of each entity are extracted, and the multi-dimensional features of each entity are fused to generate corresponding joint entity embedding vectors. By calculating the similarity between any two joint entity embedding vectors, the top K similarity pairs are selected as candidate entity pairs. The candidate entity pairs are then classified by a machine learning classifier to obtain the entity alignment result. Based on the ontology mapping rules and the entity alignment results, conflict detection is performed by combining entity attribute values. Then, conflict resolution is performed through semantic reasoning, structural assistance, and reinforcement learning optimization to obtain the optimal conflict resolution scheme and the normalized entity attribute values ​​after conflict resolution. Based on the normalized ontology, the ontology mapping rules, the entity alignment results, the optimal conflict resolution scheme, and the normalized entity attribute values, a fused knowledge graph is obtained through distributed optimization. The fused knowledge graph is then incrementally updated to obtain an updated fused knowledge graph.

[0027] Example 3 A computer program product includes a computer program that, when executed by a processor, implements the steps of the method described in Embodiment 2.

[0028] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0029] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A system that specifies functions in one or more boxes.

[0030] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including an instruction set implemented in a process. Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0031] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0032] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit its scope of protection. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading the present invention, they can still make various changes, modifications or equivalent substitutions to the specific implementation of the invention, but these changes, modifications or equivalent substitutions are all within the scope of protection of the pending claims of the invention.

[0033] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

Claims

1. A large scale knowledge fusion system based on ontology mapping and conflict resolution, characterized in that, include: The ontology normalization module is used to normalize the ontology of the multi-source heterogeneous knowledge graph to be fused, so as to obtain a normalized ontology. The multi-strategy ontology mapping module is used to extract multi-dimensional features of the normalized ontology, calculate the similarity of the multi-dimensional features respectively, generate a comprehensive similarity matrix by combining dynamic weights, and perform ontology mapping on the comprehensive similarity matrix to obtain ontology mapping rules. The intelligent entity alignment module is used to extract multi-dimensional features of each entity based on the normalized ontology and the ontology mapping rules, and to fuse the multi-dimensional features of each entity to generate a corresponding joint entity embedding vector. By calculating the similarity between any two joint entity embedding vectors, the top K similarity-ranked entity pairs are selected as candidate entity pairs. The candidate entity pairs are then classified by a machine learning classifier to obtain the entity alignment result. The deep conflict detection and resolution module is used to detect conflicts based on the ontology mapping rules and the entity alignment results, combined with entity attribute values, and then resolve conflicts through semantic reasoning, structural assistance and reinforcement learning optimization to obtain the optimal conflict resolution scheme and the normalized entity attribute values ​​after conflict resolution. The knowledge graph fusion module is used to obtain a fused knowledge graph through distributed optimization based on the normalized ontology, the ontology mapping rules, the entity alignment results, the optimal conflict resolution scheme, and the normalized entity attribute values, and to perform incremental updates on the fused knowledge graph to obtain an updated fused knowledge graph.

2. The system for large scale knowledge fusion based on ontology mapping and conflict resolution according to claim 1, characterized in that: The specific method for standardizing the ontology of the multi-source heterogeneous knowledge graph to be integrated is as follows: The ontology of the multi-source heterogeneous knowledge graph to be integrated is standardized, including format unification, namespace standardization and terminology clarification; Among them, format unification is to convert the ontology of the multi-source heterogeneous knowledge graph to be integrated into a standard ontology language format; Namespace standardization is to unify the namespace of the ontology of all multi-source heterogeneous knowledge graphs to be integrated into a global namespace; Terminology clarification identifies ambiguous and unclear terms in the ontology after standardization of format. It calls a large language model to construct a contextual description that includes term definitions, relationships, and application scenarios. The input model generates standard semantic labels, and the output is an ontology with standardized terms.

3. The large-scale knowledge fusion system based on ontology mapping and conflict resolution according to claim 1, characterized in that: The process of extracting multi-dimensional features from the normalized ontology, calculating the similarity of each feature, generating a comprehensive similarity matrix using dynamic weights, and then mapping the comprehensive similarity matrix to the ontology to obtain the ontology mapping rules is as follows: Feature extraction is performed on the normalized ontology to extract semantic features, structural features, and instance features. The semantic similarity of the semantic features is calculated using the cosine similarity algorithm. : in, Representing a normalized ontology The attributes, Representing a normalized ontology The attributes, Representing a normalized ontology The concept, Representing a normalized ontology The concept, Representing a normalized ontology semantic embedding vector, Representing a normalized ontology semantic embedding vector; The structural similarity of the structural features is calculated using the cosine similarity algorithm. : in, Representing a normalized ontology The structural embedding vector, Representing a normalized ontology The structural embedding vector; The instance similarity of the instance features is calculated using the cosine similarity algorithm. : in, Representing a normalized ontology Concept The corresponding instance feature matrix, Representing a normalized ontology Concept The corresponding instance feature matrix; Based on semantic similarity Structural similarity Similarity to instances The overall similarity is calculated through dynamic weight fusion. : in, Weights for semantic similarity The weights for structural similarity, The weights for instance similarity; A comprehensive similarity matrix is ​​generated based on the comprehensive similarity between different normalized ontologies; Set a comprehensive similarity threshold, select mapping pairs in the comprehensive similarity matrix whose comprehensive similarity is not lower than the comprehensive similarity threshold as candidate mapping pairs, and generate ontology mapping rules after logically verifying the candidate mapping pairs.

4. The large-scale knowledge fusion system based on ontology mapping and conflict resolution according to claim 1, characterized in that: The specific process for obtaining the entity alignment result is as follows: Based on the standardized ontology and the ontology mapping rules, multi-dimensional features of entities are extracted, including attribute features, text description features and graph structure features. A graph neural network is used to iteratively aggregate the features of entity neighbor nodes to generate structural embeddings. A pre-trained language model is used to encode the text description features to generate text embeddings. An attention fusion strategy is used to fuse the structural embeddings and text embeddings to obtain a unified joint entity embedding vector. Calculate the cosine similarity of the joint entity embedding vectors of different entities in the multi-source heterogeneous knowledge graph to be fused. : in, Representing entities The joint entity embedding vector, Representing entities The joint entity embedding vector; The top K entity pairs with the highest cosine similarity are selected as candidate entity pairs. A machine learning classifier is used to perform binary classification on the candidate entity pairs to obtain the entity alignment result.

5. The large-scale knowledge fusion system based on ontology mapping and conflict resolution according to claim 1, characterized in that: The specific process of conflict detection based on the ontology mapping rules and the entity alignment results, combined with entity attribute values, is as follows: Based on the ontology mapping rules and the entity alignment results, the association relationship between the aligned entities in the multi-source heterogeneous knowledge graph to be fused is established. The entity attribute values ​​of the associated entities are traversed, and the differences in the values ​​of the same attribute of the same entity in different knowledge graphs are compared to identify the conflict detection results. The conflict detection results include type conflict, numerical conflict and logical conflict.

6. The large-scale knowledge fusion system based on ontology mapping and conflict resolution according to claim 1, characterized in that: The specific process of resolving conflicts through semantic reasoning, structural assistance, and reinforcement learning optimization, to obtain the optimal conflict resolution scheme and the normalized entity attribute values ​​after conflict resolution, is as follows: Conflict resolution is achieved through semantic reasoning, structural assistance, and reinforcement learning optimization. Semantic reasoning refers to converting conflict scenarios into structured natural language descriptions that include conflict entities, conflict attributes, KG attribute values, and knowledge source background. The large language model uses these structured natural language descriptions to generate conflict resolution decision logic. Structural assistance refers to analyzing the topological structure of conflict attributes in the ontology layer of the multi-source heterogeneous knowledge graph to be fused through graph neural networks, and using topological consistency to evaluate the rationality of the conflict resolution decision logic; Reinforcement learning optimization refers to constructing a dynamic weight decision model based on reinforcement learning, using the DPO algorithm, and combining conflict resolution decision logic, the topological structure, and the authority of the knowledge source to dynamically adjust the weights of each evaluation dimension, thereby obtaining the optimal conflict resolution scheme and the normalized entity attribute values ​​after conflict resolution.

7. The large-scale knowledge fusion system based on ontology mapping and conflict resolution according to claim 6, characterized in that: The specific process for obtaining the normalized entity attribute values ​​after conflict resolution is as follows: The entity attribute values ​​after conflict resolution are formatted and standardized by using a deep learning model to eliminate format differences and obtain entity attribute values ​​in a standard format. The entity attribute values ​​in the standard format are numerically converted. Based on the domain knowledge base, the unit conversion and precision unification of the standard format entity attribute values ​​are performed to obtain the numerically converted entity attribute values. The formula is as follows: in, For transformation functions in the domain knowledge base, For entity attribute values ​​in a standard format, To unify the target units, To standardize target accuracy, These are the entity attribute values ​​after numerical conversion; Text standardization is performed on the entity attribute values ​​after numerical conversion. Regular expressions and large language models are used to correct text errors and inconsistencies in the entity attribute values ​​after numerical conversion, and finally, the standardized entity attribute values ​​after conflict resolution are obtained.

8. The large-scale knowledge fusion system based on ontology mapping and conflict resolution according to claim 1, characterized in that: Based on the normalized ontology, the ontology mapping rules, the entity alignment results, the optimal conflict resolution scheme, and the normalized entity attribute values, a fused knowledge graph is obtained through distributed optimization. The specific process of incrementally updating the fused knowledge graph to obtain the updated fused knowledge graph is as follows: Based on the normalized ontology, the ontology mapping rules, the entity alignment results, the optimal conflict resolution scheme, and the normalized entity attribute values, distributed optimization is performed. The tasks of ontology mapping rules, entity alignment results, optimal conflict resolution scheme, and normalized entity attribute values ​​are split into corresponding independent partitions for parallel processing. Container orchestration technology is used, and resources are allocated to each independent partition according to task type through a large language model inference cluster, a graph neural network computing cluster, and a general computing cluster. Elastic distributed datasets and vector databases are used to store the intermediate results and embedding vectors of each independent partition, ultimately resulting in a fused knowledge graph. Incremental updates are performed on the fused knowledge graph. By comparing the differences in structure, concepts, entities, and attributes between the newly added knowledge sources and the fused knowledge graph using a version hash algorithm, the added, modified, or deleted modules are located. Ontology mapping, entity alignment, and conflict resolution are re-executed only on the differing modules to obtain local update results. The local update results are then integrated into the fused knowledge graph through consistency verification to obtain the updated fused knowledge graph.

9. A large-scale knowledge fusion method based on ontology mapping and conflict resolution, characterized in that, It includes the following steps: The ontology of the multi-source heterogeneous knowledge graph to be integrated is normalized to obtain a normalized ontology; Extract the multi-dimensional features of the normalized ontology, calculate the similarity of the multi-dimensional features respectively, generate a comprehensive similarity matrix by combining dynamic weights, and perform ontology mapping on the comprehensive similarity matrix to obtain the ontology mapping rules. Based on the normalized ontology and the ontology mapping rules, multi-dimensional features of each entity are extracted, and the multi-dimensional features of each entity are fused to generate corresponding joint entity embedding vectors. By calculating the similarity between any two joint entity embedding vectors, the top K similarity pairs are selected as candidate entity pairs. The candidate entity pairs are then classified by a machine learning classifier to obtain the entity alignment result. Based on the ontology mapping rules and the entity alignment results, conflict detection is performed by combining entity attribute values. Then, conflict resolution is performed through semantic reasoning, structural assistance, and reinforcement learning optimization to obtain the optimal conflict resolution scheme and the normalized entity attribute values ​​after conflict resolution. Based on the normalized ontology, the ontology mapping rules, the entity alignment results, the optimal conflict resolution scheme, and the normalized entity attribute values, a fused knowledge graph is obtained through distributed optimization. The fused knowledge graph is then incrementally updated to obtain an updated fused knowledge graph.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method of claim 9.

Images