AI knowledge graph construction and maintenance method based on ontology engineering

Abstract

The invention relates to the technical field of knowledge graph construction and maintenance, and discloses an AI knowledge graph construction and maintenance method based on ontology engineering, and the method comprises the steps: obtaining multi-source heterogeneous data to construct an initial graph ontology model; loading the constraint rules and the entity alignment parameters to carry out cross-modal knowledge fusion, and generating a dynamic knowledge evaluation model; executing ontology topology reasoning based on the model, and constructing time sequence knowledge evolution simulation data; training the semantic conflict resolution model to generate a knowledge fusion optimization model, and outputting map maintenance parameters; knowledge increment simulation information is generated, a multi-dimensional optimization model is constructed to iteratively adjust the ontology structure, and optimal knowledge fusion path parameters are generated; and dynamic ontology matching is realized by combining real-time reasoning and historical track updating maintenance strategies. According to the method, the construction efficiency, the fusion quality and the self-adaptive maintenance capability of the knowledge graph are improved, the method is suitable for an intelligent knowledge management scene of multi-source heterogeneous data, and the defects of a traditional method in the aspects of semantic conflict resolution, ontology dynamic evolution and the like are overcome.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph construction and maintenance technology, specifically to an AI knowledge graph construction and maintenance method based on ontology engineering. Background Technology

[0002] Knowledge graphs, as a key technology in the field of artificial intelligence, aim to represent entities and their relationships in the real world in a structured way, providing underlying knowledge support for applications such as intelligent question answering, semantic search, and decision analysis. However, the construction and maintenance of knowledge graphs currently face many challenges, which restricts their application effectiveness in complex scenarios.

[0003] From a data perspective, integrating multi-source heterogeneous data is the primary challenge. Data from various channels, including the internet, sensors, and databases, exhibits characteristics such as heterogeneous formats (e.g., the coexistence of structured, semi-structured, and unstructured data), semantic ambiguity (different expressions of the same concept, and the same term for different concepts), and diverse modalities (text, images, audio, etc.). Traditional methods struggle to effectively extract ontological concepts and semantic relationships from cross-modal data, leading to issues such as missing concepts and incorrect relationships during the initial knowledge graph construction, thus affecting the fundamental quality of the knowledge graph.

[0004] In the knowledge fusion stage, the consistency processing of cross-modal knowledge faces technical bottlenecks. Descriptions of the same entity often differ across modalities; for example, text descriptions emphasize attribute features, while image data focuses on visual features. Achieving cross-modal entity alignment and knowledge fusion, eliminating semantic conflicts, and maintaining knowledge consistency are weak points in existing technologies. Existing fusion frameworks lack dynamic evaluation mechanisms, making it difficult to adapt to changes in knowledge structure brought about by real-time data updates, resulting in redundancy or inconsistency in the fused knowledge graph.

[0005] Regarding ontology evolution and maintenance, existing methods lack effective modeling of the temporal evolution of knowledge. As domain knowledge is constantly updated, the ontology structure needs to be dynamically adjusted to reflect the latest knowledge system. However, traditional methods struggle to capture dynamic features such as knowledge update frequency and semantic conflict evolution, causing ontology models to fail to adapt to domain changes in a timely manner. This results in problems such as confused concept levels and outdated relation edges, affecting the timeliness and usability of knowledge graphs.

[0006] The inadequacy of semantic conflict resolution mechanisms also hinders the improvement of knowledge graph quality. Existing conflict resolution methods are mostly based on static rules and lack the ability to perform deep learning and dynamic optimization on conflict data. They are unable to accurately identify the root causes of conflicts in complex semantic networks and provide optimal resolution strategies, resulting in low efficiency and poor accuracy in conflict resolution during knowledge fusion, which affects the reliability of knowledge graphs.

[0007] Furthermore, the lack of multi-dimensional optimization and dynamic ontology matching capabilities makes it difficult for knowledge graphs to achieve a balance between multiple objectives such as knowledge coverage efficiency, semantic conflict control, and ontology evolution cost. Existing methods mostly employ single optimization strategies, failing to comprehensively consider the influence of multiple variables such as entity node density, relation confidence threshold, and knowledge update frequency. This results in optimized ontology structures that are difficult to meet the complex needs of practical applications, limiting the expansion of knowledge graph application scenarios. Summary of the Invention

[0008] The purpose of this invention is to provide a method for constructing and maintaining AI knowledge graphs based on ontology engineering, so as to solve the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a method for constructing and maintaining an AI knowledge graph based on ontology engineering, the method comprising:

[0010] Acquire multi-source heterogeneous data and extract ontological concepts and semantic relationship sets to construct an initial graph ontological model;

[0011] Domain knowledge constraint rules are loaded onto the initial graph ontology model to generate a standardized knowledge fusion framework. Cross-modal knowledge fusion is then performed in conjunction with entity alignment parameters to obtain a dynamic knowledge evaluation model.

[0012] Ontology topology reasoning is performed based on a dynamic knowledge evaluation model to generate entity relationship extension paths, and time-series knowledge evolution simulation data is constructed by combining knowledge update frequency parameters.

[0013] A semantic conflict resolution model is constructed and trained using temporal knowledge evolution simulation data to generate a knowledge fusion optimization model, which then outputs graph maintenance parameters. The graph maintenance parameters include at least entity disambiguation weights, relation confidence thresholds, and ontology evolution priority sequences.

[0014] Based on graph maintenance parameters, dynamic knowledge evaluation model and entity alignment parameters, knowledge increment simulation information is generated;

[0015] By combining knowledge increment simulation information with ontology evolution constraints, a multi-dimensional optimization model is constructed and the ontology structure is iteratively adjusted to generate the optimal knowledge fusion path parameters.

[0016] Based on the real-time ontology topology network, reason about the current knowledge fusion path parameters, calculate the entity semantic consistency index, and combine the optimal knowledge fusion path parameters with the historical evolution trajectory to update the knowledge graph maintenance strategy to achieve the goal of dynamic ontology matching.

[0017] Preferably, the temporal knowledge evolution simulation data includes at least a semantic conflict set, attribute constraint parameters, ontology evolution boundary set, and dynamic resolution priority sequence;

[0018] The knowledge increment simulation information includes at least entity mapping deviation, knowledge coverage completeness rate, and semantic consistency evaluation index.

[0019] Preferably, the step of acquiring multi-source heterogeneous data and extracting ontology concepts and semantic relationship sets to construct an initial graph ontology model includes the following steps:

[0020] Structured parsing and cleaning of multi-source heterogeneous data to generate a standardized set of knowledge units;

[0021] Ontology concept extraction is performed on a set of knowledge units, wherein the extraction includes one or more of semantic clustering, attribute association mapping, concept hierarchy partitioning, and synonym disambiguation;

[0022] Based on the conceptual hierarchy partitioning results, an initial graph ontology model is constructed, wherein the model includes a set of entity nodes, a set of relation edges, and attribute constraint rules;

[0023] Semantic relation strength analysis is performed on the initial graph ontology model to generate a dynamic relation weight matrix, and domain knowledge constraint labels are assigned to each entity node.

[0024] Preferably, the step of loading domain knowledge constraint rules onto the initial graph ontology model, generating a standardized knowledge fusion framework, and performing cross-modal knowledge fusion in conjunction with entity alignment parameters to obtain a dynamic knowledge evaluation model includes the following steps:

[0025] The initial knowledge fusion framework is generated by loading domain knowledge constraint rules onto the initial graph ontology model to constrain the range of entity attribute values.

[0026] Load cross-modal alignment parameters onto the first knowledge fusion framework to generate a standardized knowledge fusion framework;

[0027] Cross-modal knowledge fusion based on a standardized knowledge fusion framework includes the following specific processes:

[0028] A multimodal entity alignment equation is constructed, which includes at least a semantic similarity equation, an attribute matching degree equation, and a context consistency equation. The multimodal entity alignment equation is iteratively solved using a dynamic programming algorithm to obtain a set of semantic conflicts, attribute constraint parameters, and an ontology evolution boundary set.

[0029] The generation of the ontology evolution boundary set includes the following steps:

[0030] Based on the semantic conflict set, calculate the conflict density of each entity node;

[0031] Identify regions in the standardized knowledge fusion framework where the conflict density is no greater than a preset threshold, and generate an ontology evolution boundary set.

[0032] Preferably, the step of constructing a semantic conflict resolution model and training it with temporal knowledge evolution simulation data to generate a knowledge fusion optimization model, and then outputting graph maintenance parameters, includes the following steps:

[0033] A semantic conflict resolution model is constructed based on the initial graph ontology model;

[0034] The semantic conflict resolution model is trained and validated using time-series knowledge evolution simulation data to generate a knowledge fusion optimization model.

[0035] Input the real-time entity alignment parameters into the knowledge fusion optimization model to predict the set of ontology evolution boundaries;

[0036] Based on the predicted set of ontology evolution boundaries, graph maintenance parameters are output, wherein the graph maintenance parameters include at least entity disambiguation weights, relation confidence thresholds, and ontology evolution priority sequences.

[0037] Preferably, the method further includes feature reconstruction of the time-series knowledge evolution simulation data, specifically:

[0038] An initial multimodal knowledge input tensor is constructed based on the semantic conflict set, attribute constraint parameters, and ontology evolution boundary set.

[0039] The initial multimodal knowledge input tensor is normalized and its features are enhanced to generate the final multimodal knowledge input tensor.

[0040] Based on the dynamic resolution priority sequence, a knowledge fusion optimization label tensor is constructed;

[0041] The training sample set is formed by combining the final multimodal knowledge input tensor with the knowledge fusion optimization label tensor.

[0042] Preferably, the output map maintenance parameters based on the predicted ontological evolution boundary set include the following steps:

[0043] Extract the entity node with the lowest semantic conflict density from the predicted ontology evolution boundary set and generate entity disambiguation weights;

[0044] Based on the topological connectivity of the predicted ontology evolution boundary set, the distribution structure of the relation confidence threshold is fitted to generate an ontology evolution priority sequence.

[0045] The semantic similarity gradient direction of the predicted ontology evolution boundary set is calculated and normalized into a knowledge fusion baseline vector, which is the direction of relation confidence adjustment.

[0046] Preferably, the step of generating knowledge increment simulation information based on graph maintenance parameters, dynamic knowledge evaluation model, and entity alignment parameters includes the following steps:

[0047] The entity disambiguation weights are mapped to the dynamic knowledge evaluation model, the relationship confidence threshold and ontology evolution priority sequence are matched, the entity node density in the conflict area is adjusted, the dynamic knowledge evaluation model is updated, the knowledge increment triggering conditions, ontology reconstruction rules and parameter update increments are defined, and the knowledge increment simulation model is generated.

[0048] Based on the knowledge increment simulation model, the multimodal entity alignment equation is iteratively solved using a dynamic programming algorithm to generate knowledge increment simulation information, specifically including:

[0049] When the knowledge increment triggering condition is met, update the entity mapping deviation, knowledge coverage completeness rate and semantic consistency evaluation index, and resolve the multimodal entity alignment equation until the simulation termination condition is reached.

[0050] The knowledge increment triggering condition includes triggering parameter updates when the current knowledge coverage completeness rate is not greater than a preset completeness rate threshold; the ontology reconstruction rule includes adjusting the relationship confidence threshold based on the semantic similarity gradient direction; the parameter update increment has a piecewise linear relationship with the current semantic consistency evaluation index.

[0051] Preferably, the step of constructing a multi-dimensional optimization model and iteratively adjusting the ontology structure to generate optimal knowledge fusion path parameters includes the following steps:

[0052] A multi-dimensional optimization model is constructed, in which the optimization variables include entity node density, relation confidence threshold and knowledge update frequency, the optimization objectives include maximizing knowledge coverage efficiency and minimizing semantic conflict density, and the constraints include ontology evolution boundary constraints and attribute matching accuracy threshold.

[0053] The multi-dimensional optimization model is initially solved using a genetic algorithm to generate an initial set of optimization paths.

[0054] Based on the initial set of optimized paths, the simulated annealing algorithm is used for global optimization to generate the optimal knowledge fusion path parameters.

[0055] Preferably, the step of using a genetic algorithm to initially solve the multi-dimensional optimization model and generate an initial set of optimization paths includes the following steps:

[0056] A chromosome coding structure is constructed based on entity node density and relationship confidence threshold to generate an initial population;

[0057] The fitness function is calculated based on knowledge coverage efficiency and semantic conflict density. Crossover and mutation operations are performed on the initial population to generate a set of intermediate optimized paths.

[0058] A local search and elite retention filter are performed on the intermediate optimized path set to generate an initial optimized path set.

[0059] Compared with the prior art, the beneficial effects of the present invention are:

[0060] In processing multi-source heterogeneous data, a standardized set of knowledge units is generated through structured parsing and cleaning. Various techniques, such as semantic clustering and attribute association mapping, are comprehensively used to extract ontology concepts, ensuring that the initial graph ontology model can accurately capture the core concepts and hierarchical structure of domain knowledge. Simultaneously, the introduction of a dynamic relation weight matrix and domain knowledge constraint labels enhances the model's ability to represent the strength of semantic relationships and entity attribute constraints, laying a solid foundation for subsequent knowledge fusion.

[0061] In the cross-modal knowledge fusion stage, a standardized knowledge fusion framework is constructed by loading domain knowledge constraint rules and cross-modal alignment parameters. Iterative solutions based on multimodal entity alignment equations and dynamic programming algorithms effectively resolve semantic conflicts and attribute matching issues in cross-modal data. In particular, by calculating conflict density and generating ontology evolution boundary sets, precise segmentation of the knowledge fusion region is achieved, improving fusion efficiency and accuracy and ensuring seamless integration of different modal knowledge into a unified ontology model.

[0062] In terms of ontology topological reasoning and temporal knowledge evolution modeling, entity relationship expansion paths are generated based on a dynamic knowledge evaluation model, and temporal knowledge evolution simulation data is constructed by combining knowledge update frequency parameters. This effectively captures the evolutionary patterns of the knowledge system over time. By modeling dynamic features such as semantic conflict sets and attribute constraint parameters, a scientific basis is provided for the temporal maintenance of the knowledge graph, enabling the ontology structure to dynamically adapt to the updates of domain knowledge and maintain the timeliness and integrity of the graph.

[0063] The semantic conflict resolution mechanism constructs a semantic conflict resolution model and trains it using temporal knowledge evolution simulation data to generate a knowledge fusion optimization model, achieving intelligent identification and resolution of semantic conflicts. Through feature reconstruction and tensor modeling, the model's ability to represent multimodal knowledge is enhanced. Combined with a dynamic resolution priority sequence, it can provide optimal resolution strategies for different conflict types, significantly improving the efficiency and accuracy of conflict resolution and ensuring the semantic consistency of the knowledge graph.

[0064] The knowledge increment simulation and multi-dimensional optimization stages achieve dynamic adjustment and multi-objective optimization of the ontology structure by generating knowledge increment simulation information and constructing a multi-dimensional optimization model. Based on a combined optimization strategy of genetic algorithm and simulated annealing algorithm, the optimal knowledge fusion path parameters can be searched in a multi-variable space, including entity node density and relation confidence threshold, balancing multiple objectives such as knowledge coverage efficiency and semantic conflict control. This ensures that the ontology structure is always in an optimal state, improving the overall performance of the knowledge graph.

[0065] In terms of dynamic ontology matching, real-time ontology topology network reasoning and historical evolution trajectory analysis enable dynamic updates to the knowledge graph maintenance strategy, achieving adaptive adjustments to the ontology structure. By combining graph maintenance parameters with optimal knowledge fusion path parameters, the knowledge graph can respond to data changes and domain requirements in real time, maintaining accurate matching with actual application scenarios and significantly enhancing the graph's practicality and generalization ability. Attached Figure Description

[0066] Figure 1 This is a schematic diagram illustrating the working principle of the AI ​​knowledge graph construction and maintenance method based on ontology engineering described in this invention.

[0067] Figure 2 Design diagrams generated for a standardized knowledge fusion framework;

[0068] Figure 3 Design diagram for training a semantic conflict resolution model;

[0069] Figure 4 Design drawings for iterative adjustments to the main body structure;

[0070] Figure 5 A design diagram for a dynamic maintenance strategy for knowledge graphs. Detailed Implementation

[0071] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0072] Please see Figures 1-5 This invention relates to a method for constructing and maintaining an AI knowledge graph based on ontology engineering, the specific implementation steps of which are as follows:

[0073] To construct an initial graph ontology model, multi-source heterogeneous data is acquired and ontology concepts and semantic relationship sets are extracted. First, the multi-source heterogeneous data undergoes structured parsing and cleaning to remove redundant information and noise, generating a standardized set of knowledge units. Then, ontology concept extraction is performed on the knowledge unit set. Through semantic clustering, attribute association mapping, concept hierarchy partitioning, and synonym disambiguation, the logical hierarchy and attribute relationships between concepts are clarified. Based on the concept hierarchy partitioning results, an initial graph ontology model is constructed, including a set of entity nodes, a set of relation edges, and attribute constraint rules. Semantic relationship strength analysis is performed on this model to generate a dynamic relation weight matrix. Simultaneously, domain knowledge constraint labels are assigned to each entity node to represent the specific rules and restrictions of its domain.

[0074] Domain knowledge constraint rules and entity alignment parameters are loaded to achieve cross-modal knowledge fusion and generate a dynamic knowledge evaluation model: Domain knowledge constraint rules are loaded onto the initial graph ontology model to constrain the range of entity attribute values, generating a first knowledge fusion framework. Cross-modal alignment parameters are further loaded to form a standardized knowledge fusion framework. When performing cross-modal knowledge fusion based on this framework, a multimodal entity alignment equation is constructed, including semantic similarity equations, attribute matching degree equations, and context consistency equations. This equation is iteratively solved using a dynamic programming algorithm to obtain a semantic conflict set, attribute constraint parameters, and an ontology evolution boundary set. The generation process of the ontology evolution boundary set is as follows: Based on the semantic conflict set, the conflict density of each entity node is calculated, and regions with conflict densities not exceeding a preset threshold are identified as ontology evolution boundaries.

[0075] Perform ontology topological reasoning and construct temporal knowledge evolution simulation data: Based on a dynamic knowledge evaluation model, perform ontology topological reasoning, analyze the association paths between entities, and generate entity relationship extension paths. Combined with knowledge update frequency parameters, construct temporal knowledge evolution simulation data that includes a semantic conflict set, attribute constraint parameters, an ontology evolution boundary set, and a dynamic resolution priority sequence to simulate the evolution of knowledge over time.

[0076] Construct and train a semantic conflict resolution model to generate a knowledge fusion optimization model and graph maintenance parameters: Based on the initial graph ontology model, construct a semantic conflict resolution model, train and validate it using time-series knowledge evolution simulation data, and generate a knowledge fusion optimization model. Input real-time entity alignment parameters into this model to predict the ontology evolution boundary set, and output graph maintenance parameters based on the prediction results, including at least entity disambiguation weights, relation confidence thresholds, and ontology evolution priority sequences.

[0077] Generate knowledge increment simulation information: Map entity disambiguation weights to a dynamic knowledge evaluation model, match relationship confidence thresholds and ontology evolution priority sequences, and adjust entity node density in conflict regions to update the model. Define knowledge increment triggering conditions (e.g., triggering parameter updates when the current knowledge coverage completeness rate is not greater than a preset threshold), ontology reconstruction rules (e.g., adjusting relationship confidence thresholds based on semantic similarity gradient direction), and parameter update increments (showing a piecewise linear relationship with the semantic consistency evaluation index) to generate a knowledge increment simulation model. Based on this model, iteratively solve the multimodal entity alignment equation using a dynamic programming algorithm to generate knowledge increment simulation information containing entity mapping deviation, knowledge coverage completeness rate, and semantic consistency evaluation index.

[0078] A multi-dimensional optimization model is constructed and the ontology structure is iteratively adjusted: An optimization model is built with entity node density, relation confidence threshold, and knowledge update frequency as optimization variables, maximizing knowledge coverage efficiency and minimizing semantic conflict density as optimization objectives, and ontology evolution boundary constraints and attribute matching accuracy thresholds as constraints. An initial set of optimization paths is generated through preliminary solution using a genetic algorithm, followed by global optimization using simulated annealing to obtain the optimal knowledge fusion path parameters.

[0079] Update the knowledge graph maintenance strategy: Based on the real-time ontology topology network, reason about the current knowledge fusion path parameters, calculate the entity semantic consistency index, and combine the optimal knowledge fusion path parameters with the historical evolution trajectory to dynamically adjust the knowledge graph maintenance strategy and achieve the goal of dynamic ontology matching.

[0080] Example 1:

[0081] In the process of constructing the initial graph ontology model, the processing of multi-source heterogeneous data requires key steps such as structured parsing, cleaning, and ontology concept extraction. The specific implementation methods are as follows:

[0082] Firstly, for the structured analysis of multi-source heterogeneous data, targeted processing methods need to be adopted according to the differences in data types. For text data, the Natural Language Processing (NLP) technology chain is used to transform raw text into structured knowledge units: continuous text is segmented into independent lexical units using word segmentation tools (such as jieba), and each word is assigned a part-of-speech tag (noun, verb, adjective, etc.) using part-of-speech tagging models (such as sequence tagging models based on BiLSTM-CRF). Then, named entity recognition (NER) models (such as BERT+CRF) are used to identify entity references in the text, such as names of people, organizations, and products, and classify them into preset entity types (such as "person", "organization", "item"). For structured data (such as relational database tables and CSV files), data schema mapping rules are defined to match fields in the database tables with entity attributes in the knowledge graph. For example, the "User ID" field in the "User Table" is mapped to the "Unique Identifier" attribute of the "User" entity, and the "Name" field is mapped to the "Name" attribute. For unstructured image data, computer vision (CV) techniques are used to extract feature vectors. Convolutional neural networks (CNNs, such as ResNet) are used to extract features from images, generating high-dimensional feature vectors containing information such as color, texture, and shape. Then, image annotation models (such as Faster R-CNN) are used to identify target objects in the images and generate descriptive labels (such as "red car" and "hospital building"), establishing a mapping relationship between image features and textual concepts. Through the collaboration of the above multi-technology stack, structured parsing of multi-source data such as text, structured data, and images is achieved, generating a standardized set of knowledge units in a unified format. Each knowledge unit contains information such as entity identifier, type, attribute key-value pairs, and source annotation.

[0083] The data cleaning process aims to remove noise, correct errors, and improve data quality. For missing value handling, statistical analysis is first used to identify missing attribute fields. For numerical attributes (such as "age" and "price"), if the missing percentage is below a preset threshold (e.g., 10%), domain knowledge is used for filling (e.g., if the "age" field is missing, and the entity type is "adult," the default value "30" is used), or the K-Nearest Neighbors (KNN) algorithm is used to interpolate based on the attribute values ​​of neighboring entities. For non-numerical attributes (such as "address" and "category"), association rule mining (e.g., the Apriori algorithm) is used to analyze the relationship between other attributes of the entity and the missing attribute, and matching values ​​are extracted from similar entities for filling. Duplicate data detection uses a combination of hash encoding and cosine similarity: first, a hash value is generated for the key attributes of each knowledge unit (e.g., entity name, unique identifier), quickly locating candidate duplicate units with the same hash value. Then, the cosine similarity between candidate units (based on attribute value vectors) is calculated. When the similarity exceeds a set threshold (e.g., 0.9), it is determined to be duplicate data; the unit with the most complete attribute is retained, and the remaining duplicates are deleted. Outlier handling involves identifying data points that deviate from the normal range using statistical methods. For numerical attributes, the Z-score algorithm is used to calculate the standard deviation multiple of each value, and values ​​exceeding ±3σ are considered outliers, corrected using regression models (such as linear regression) or binning methods (such as equal-frequency binning). For non-numerical attributes, outliers that do not conform to the rules are filtered out by establishing a domain rule base (such as "drug names must contain keywords of active ingredients"), and then manually verified, corrected, or deleted. This cleaning process generates a high-quality set of standardized knowledge units, laying the foundation for subsequent ontology concept extraction.

[0084] The ontology concept extraction stage requires the comprehensive application of techniques such as semantic clustering, attribute association mapping, concept hierarchy segmentation, and synonym disambiguation. Semantic clustering employs hierarchical clustering algorithms (such as agglomerative hierarchical clustering) to group knowledge units: first, each knowledge unit is treated as an independent cluster, and the cosine similarity between clusters (based on attribute value vectors) is calculated. The two clusters with the highest similarity are merged, and this process is repeated until the similarity of all clusters is below a set threshold (e.g., 0.6), forming concept clusters of different granularities. Attribute association mapping analyzes the relationships between attributes within concept clusters using association rule mining algorithms (such as the Apriori algorithm). For example, in the "medicine" concept cluster, the attributes "ingredients" and "indications" are found to co-occur frequently. An association rule of "ingredients → indications" is established, and an attribute-concept mapping table is generated, recording the concept to which each attribute belongs and the strength of the association. Conceptual hierarchy segmentation combines domain expert knowledge with semantic networks (such as WordNet) to construct a tree-like hierarchical structure: First, the top-level concept (such as "entity") is determined, and then it is subdivided step by step according to the semantic category of the concept. For example, "entity" is divided into "physical entity" and "abstract entity". "Physical entity" is further divided into "organism", "item", "location", etc. Each sub-concept inherits the attributes and relationships of the parent concept and adds unique attributes (such as "medicine" as a subclass of "item", adding unique attributes such as "dosage form" and "specification"). Synonym disambiguation uses a thesaurus (such as the "Extended Thesaurus") and cosine similarity calculation to identify synonymous concepts. For example, "computer" and "computer" are regarded as the same concept. By merging synonym concept clusters, conceptual redundancy is eliminated, and a unique canonical concept name is assigned to each synonym set to establish a synonym-canonical word mapping table.

[0085] After extracting the ontology concepts, an initial graph ontology model is constructed based on the concept hierarchy partitioning results. This model is represented by a directed graph structure, containing a set of entity nodes, a set of relation edges, and attribute constraint rules: entity nodes correspond to normalized concepts, node labels are concept names (e.g., "medicine" or "disease"), and node attributes contain all attribute key-value pairs of that concept (e.g., the "name," "ingredients," and "indications" attributes of the "medicine" node); relation edges represent semantic relationships between entities (e.g., "medicine-treatment-disease" or "company-production-product"), edge labels are relation types, and edge weights are generated through semantic relation strength analysis. Semantic relation strength analysis employs the edge betweenness centrality algorithm in graph theory, calculating the frequency of each relation edge in the shortest paths of all node pairs. Higher frequencies result in greater edge betweenness centrality values, corresponding to stronger relations and higher weight values. A dynamic relation weight matrix records the weight values ​​of all relation edges, used for assessing relation importance in subsequent knowledge fusion and reasoning processes.

[0086] Simultaneously, domain knowledge constraint labels are assigned to each entity node. The label content is based on the specific rules and standards of the entity's domain. For example, in the medical field, the "Drugs" entity node is assigned the "Drug Administration Law Constraint" label, restricting its "Indications" attribute to be within the list of diseases approved by the National Medical Products Administration; the "Diseases" entity node is assigned the "ICD-10 Coding Constraint" label, requiring its "Coding" attribute to conform to the International Classification of Diseases (ICD). Domain knowledge constraint labels are automatically generated through a predefined domain rule base, which includes entity attribute constraints and relationship existence constraints for each domain (e.g., the "Medical Devices" entity must be associated with the "Manufacturer" entity). A rule engine (such as Drools) performs real-time validation of entity nodes to ensure the model conforms to domain specifications.

[0087] Through the complete process described above, the initial ontology model of the knowledge graph is constructed from multi-source heterogeneous data, providing a structured knowledge foundation for subsequent knowledge fusion, evolution, and maintenance. This process not only solves the heterogeneity problem of multi-source data and improves data quality through standardized parsing and cleaning, but also ensures the semantic accuracy and structural rationality of the knowledge graph through systematic ontology concept extraction and model construction, laying a crucial foundation for the entire lifecycle management of the knowledge graph.

[0088] Example 2:

[0089] In the cross-modal knowledge fusion stage, the construction of the standardized knowledge fusion framework and the multimodal entity alignment process need to be implemented through domain knowledge constraint rule loading, cross-modal parameter configuration, and conflict resolution algorithms. The specific implementation methods are as follows:

[0090] Domain knowledge constraint rules are loaded onto the initial graph ontology model to generate the first knowledge fusion framework. These domain knowledge constraint rules are derived from domain standards, industry norms, and expert experience, and are defined in the form of logical expressions or regular expressions. For example, in the medical field, for the "indication" attribute of the "drug" entity, the constraint rule is defined as "indication ∈ {hypertension, diabetes, coronary heart disease}" (a preset set of disease categories). If the "indication" attribute value of a drug entity in the input data is "cold," a conflict detection mechanism is triggered, marking the entity as needing correction. For the "transaction" entity in the financial field, a numerical constraint rule "transaction amount > 0" is defined. If a negative amount occurs, a data cleaning process is automatically triggered. The constraint rules are applied to the entity attributes of the initial graph ontology model using a rule engine (such as Apache CommonsValidator). Attribute values ​​that do not conform to the rules are intercepted, marked, or automatically corrected, generating the first knowledge fusion framework that conforms to the domain norms. This framework ensures that entity attribute values ​​are within a reasonable range and initially eliminates obvious semantic conflicts.

[0091] Load cross-modal alignment parameters into the first knowledge fusion framework to generate a standardized knowledge fusion framework. The cross-modal alignment parameters are designed based on the feature differences of different modal data, including visual semantic embedding parameters for text-image alignment, semantic mapping matrices for structured data-text alignment, etc. Taking the text and image modalities as an example, a contrastive learning method is used to train a vision-language model (such as CLIP), mapping text descriptions (such as "red sedan") and image feature vectors (extracted by ResNet) to a shared semantic space, and generating an embedding parameter matrix containing text-image alignment relationships. For the alignment of structured data (such as database tables) and text data, by constructing a domain ontology mapping table, the corresponding relationship between structured fields and text concepts is defined manually or automatically. For example, the "age" field in the "user table" is mapped to the text concept "age", and a semantic mapping matrix containing data type conversion rules (such as integer to string) is generated. After integrating multi-modal alignment parameters, the standardized knowledge fusion framework forms a general framework that supports unified processing of multi-modal data such as text, images, and structured data, providing an infrastructure for cross-modal knowledge fusion.

[0092] When performing cross-modal knowledge fusion under the standardized knowledge fusion framework, first construct a multi-modal entity alignment equation, which includes a semantic similarity equation, an attribute matching degree equation, and a context consistency equation. The semantic similarity equation uses a pre-trained language model (such as BERT) to calculate the cosine similarity of semantic vectors between text entities. The specific process is as follows: input the entity name and description text into the BERT model to generate a 768-dimensional semantic vector, and calculate the similarity value between the vectors through the cosine similarity formula. The value range is [0,1], and the larger the value, the more similar the semantics. The attribute matching degree equation calculates the similarity of attribute value strings using the edit distance algorithm for the attribute value differences of structured entities. For example, for the "address" attribute, compare the edit distance between "XX Road, Chaoyang District, Beijing" and "XX Road, Haidian District, Beijing". The smaller the distance, the closer the attribute values. The context consistency equation uses a graph neural network (GNN, such as GraphSAGE) to analyze the context topological structure similarity of entities in the knowledge graph. By aggregating the feature information of entity neighbor nodes, a node embedding vector containing context relationships is generated, and the Euclidean distance is used to calculate the difference between the vectors. The smaller the difference, the higher the context consistency.

[0093] The multimodal entity alignment equation is solved iteratively using a dynamic programming algorithm, gradually adjusting the entity alignment parameters to minimize overall conflict. The state transition equation of the dynamic programming algorithm is defined as follows: In the t-th iteration, the optimal solution of the current entity alignment state is derived from the optimal solution of the previous t-1 iterations. By comparing the conflict values ​​of different alignment schemes (a weighted sum of semantic similarity, attribute matching degree, and contextual consistency), the scheme with the smallest conflict value is selected as the current optimal solution. During the iteration process, if the semantic similarity of an entity pair is lower than a preset threshold (e.g., 0.5) and the attribute matching degree is lower than a preset threshold (e.g., 0.6), it is determined to be a conflicting entity pair and added to the semantic conflict set. At the same time, the actual value range of each entity attribute is recorded, attribute constraint parameters are generated (e.g., the value range of the "age" attribute is [18, 80]), and an ontology evolution boundary set is generated based on the concept boundaries involved in the entity alignment process.

[0094] The generation process of the ontology evolution boundary set includes two steps: conflict density calculation and boundary identification. The conflict density calculation uses the sliding window method, with the entity node as the center and a window size of k (e.g., k=5, indicating that it includes the center node and its neighboring nodes within k hops). The ratio of the number of conflicting edges within the window to the total number of edges is calculated using the following formula:

[0095]

[0096] Here, conflict relationship edges refer to the relationship edges connecting conflicting entity pairs. By traversing all entity nodes in the knowledge graph, the conflict density value of each node is calculated to form a conflict density distribution matrix.

[0097] In the boundary identification phase, a conflict density threshold θ is set (e.g., θ=0.3). The conflict density distribution matrix is ​​traversed, and all regions with conflict densities ≤ θ are identified as ontology evolution boundaries. Specifically, a region growing algorithm is used: starting from the node with the lowest conflict density, nodes with conflict densities ≤ θ among its neighbors are merged into the same region. This process is repeated until no new nodes can be merged, forming one or more connected ontology evolution boundary regions. Each boundary region contains a set of entity nodes and corresponding relation edges, serving as the scope of subsequent ontology evolution operations. This restricts evolution to within the boundary, avoiding knowledge graph structure instability caused by global adjustments.

[0098] During the iterative process of cross-modal knowledge fusion, the size of the semantic conflict set is gradually reduced and the accuracy of the attribute constraint parameters is improved by dynamically adjusting cross-modal alignment parameters (such as visual semantic embedding parameters and semantic mapping matrix) and entity alignment strategies (such as prioritizing alignment of high-frequency entity pairs). When the change in the size of the semantic conflict set is less than a preset value (such as 5%) in n consecutive iterations (such as n=5), the cross-modal knowledge fusion process is considered to have converged. The final semantic conflict set, attribute constraint parameters, and ontology evolution boundary set are output for subsequent construction of dynamic knowledge evaluation models and training of semantic conflict resolution models.

[0099] This implementation method achieves effective semantic fusion of multi-source heterogeneous data by pre-validating domain knowledge constraint rules, configuring cross-modal alignment parameters in a structured manner, and iteratively solving multi-modal entity alignment equations, thus solving the semantic gap problem of cross-modal data. Simultaneously, through conflict density analysis and ontology evolution boundary identification, it provides clear range guidance for the local evolution of the knowledge graph, balancing the stability and scalability of the knowledge graph. This ensures accurate location of conflict regions when introducing new data, avoiding excessive disturbance to the global structure, and laying a crucial foundation for subsequent temporal knowledge evolution simulation and dynamic ontology matching.

[0100] Example 3:

[0101] The construction and training process of the semantic conflict resolution model needs to combine the graph structure features of the initial graph ontology model with the multimodal features of the temporal knowledge evolution simulation data. The specific implementation method is as follows:

[0102] First, a semantic conflict resolution model is constructed based on the initial knowledge graph ontology model. This model employs a Graph Convolutional Network (GCN) architecture, taking the directed graph structure of the knowledge graph as input. The feature vector of each entity node contains node attribute values ​​(e.g., "ingredients" and "indications" for the "medicine" entity) and one-hot encodings of domain knowledge constraint labels. The feature vector of each relation edge contains the relation type (e.g., "treatment" and "production") and weight values ​​from a dynamic relation weight matrix. The first convolutional layer of the GCN model updates the node representation by aggregating the feature information of neighboring nodes. The calculation formula is as follows:

[0103]

[0104] in, To add a self-loop adjacency matrix, This is the original adjacency matrix. It is the identity matrix; It is a degree matrix; For the first The node feature matrix of the layer; For the first The trainable weight matrix of the layer; The activation function is ReLU. Through the stacking of multiple convolutional layers, the model gradually learns the high-order semantic features of entity nodes and their neighborhoods, and finally outputs the conflict resolution strategy of each node (such as the probability distribution of "entity merging", "relationship correction" and "attribute update") through a fully connected layer.

[0105] Feature reconstruction of temporal knowledge evolution simulation data is a key preprocessing step for model training. First, an initial multimodal knowledge input tensor is constructed based on the semantic conflict set, attribute constraint parameters, and ontology evolution boundary set. The semantic conflict set is represented by a binary matrix, and the matrix elements... Representing entities With entity There is a semantic conflict, otherwise it is Attribute constraint parameters are represented as numerical vectors, with each element corresponding to the boundary of the value range of an entity attribute (e.g., the minimum and maximum values ​​of the "age" attribute). The ontology evolution boundary set is represented as a list of node indices, converted into a sparse vector through one-hot encoding. These three types of features are concatenated dimensionally to form the initial multimodal knowledge input tensor. ,in The number of entity nodes. For semantic conflict feature dimension, As an attribute constraint feature dimension, This represents the dimension of the ontological evolution boundary features.

[0106] The initial tensor is normalized and its features are enhanced. Normalization uses the Z-score standardization method for each feature dimension. Calculate the mean and standard deviation The standardized value is This ensures that all feature dimensions have the same numerical range. Feature enhancement is achieved through Principal Component Analysis (PCA), which projects high-dimensional features into a low-dimensional space, retaining principal components with a cumulative variance contribution rate of up to 90%, and generating the final multimodal knowledge input tensor. ,in Principal component dimension.

[0107] Dynamically resolving priority sequences is used to construct a knowledge fusion optimization label tensor. The priority sequence is determined based on conflict severity and domain importance: conflict severity is determined by the type of semantic conflict (e.g., entity homonyms, contradictory relationships) and the number of entities involved; domain importance is determined by the business priority of the entity's domain (e.g., in the medical field, "emergency medicines" have a higher priority than ordinary medicines). The priority sequence is mapped to continuous values ​​between 0 and 1, and the label tensor is generated through linear interpolation. A larger value indicates a higher resolution priority. This is combined with the final multimodal knowledge input tensor. Optimize label tensors by integrating knowledge To form a training sample set It is used to supervise the parameter learning of the semantic conflict resolution model.

[0108] The training and validation process of the knowledge fusion optimization model employs a cross-validation strategy. The sample set is divided into a training set (80%) and a validation set (20%). Mean squared error (MSE) is used as the loss function, and the weight parameters of the GCN model are updated using a backpropagation algorithm (such as the Adam optimizer). During training, the loss value on the validation set is monitored. When the loss value no longer decreases after 10 consecutive iterations, the model is considered to have converged, and the final knowledge fusion optimization model is generated.

[0109] Real-time entity alignment parameters are input into the knowledge fusion optimization model for inference. These parameters include cross-modal alignment parameter updates within the current iteration cycle (such as the increment of the text-image embedding matrix) and entity mapping deviation (a vector reflecting alignment error). The model calculates through forward propagation and outputs a predicted set of ontology evolution boundaries, represented as a list of node indices corresponding to regions in the knowledge graph currently suitable for evolutionary operations.

[0110] The process of maintaining parameters in the output graph of the predicted ontological evolution boundary set is as follows:

[0111] Entity disambiguation weight generation: Extract the entity node with the lowest semantic conflict density from the boundary set. Semantic conflict density is defined as the ratio of the number of conflicting relation edges that node participates in to the total number of relation edges. Let node... The conflict density is Then the entity disambiguation weight The calculation formula is:

[0112]

[0113] in, and Let these be the maximum and minimum values ​​of the conflict density of nodes within the boundary set. For very small positive numbers (e.g.) To avoid the denominator being zero, a larger weight value indicates a lower disambiguation priority for that node, prioritizing the retention of nodes with stable structures.

[0114] Relationship confidence threshold distribution and ontology evolution priority sequence generation: The topological connections of the boundary set are analyzed, and the shortest path algorithm is used to calculate the path length between nodes to fit the distribution structure of the relationship confidence threshold. For high-frequency connections (such as "medicine-treatment-disease"), a higher confidence threshold is set to ensure higher semantic similarity during alignment; for low-frequency connections, a lower threshold is set to improve flexibility. The ontology evolution priority sequence is sorted based on a combination of node degree (number of connections) and conflict density, with nodes having high degree and low conflict density evolving first. The formula is:

[0115]

[0116] in, This is the degree weighting coefficient (value range 0-1, set according to domain requirements, such as 0.7 for knowledge-intensive domains).

[0117] Relationship confidence adjustment direction generation: Calculate the semantic similarity gradient direction between entities within the boundary set. Use principal component analysis to extract the eigenvector corresponding to the largest eigenvalue as the principal gradient direction, and obtain the knowledge fusion baseline vector through L2 normalization. This vector indicates the optimal direction for adjusting the confidence level of a relationship. For example, if the gradient direction points to a region of high similarity, it is recommended to increase the confidence threshold of the relationship in that region to enhance semantic consistency.

[0118] Through the above process, the semantic conflict resolution model achieves end-to-end mapping from multimodal knowledge features to graph maintenance parameters. Combined with the dynamic characteristics of time-series knowledge evolution simulation data, it provides a quantitative basis for real-time maintenance of the knowledge graph. This implementation effectively captures the contextual dependencies between entities and relationships in the knowledge graph through the structural feature learning capability of graph neural networks. Simultaneously, through feature reconstruction and priority modeling, it improves the model's generalization ability to complex conflict scenarios, ensuring the rationality and timeliness of graph maintenance parameters, and providing crucial support for subsequent knowledge increment simulation and ontology structure optimization.

[0119] Example 4:

[0120] In the process of generating knowledge increment simulation information, it is necessary to combine the map maintenance parameters with the dynamic knowledge evaluation model, and achieve incremental fusion of multimodal data through conflict region adjustment, trigger condition judgment and iterative solution. The specific implementation method is as follows:

[0121] Taking a medical knowledge graph as an example, assume that the current graph contains entities such as "drugs," "diseases," and "indications," along with their corresponding relationships. First, the entity disambiguation weights are mapped to the dynamic knowledge evaluation model: the disambiguation weight values ​​of the "drugs" entity nodes are obtained from the graph maintenance parameters (e.g., node A has a weight of 0.8, node B has a weight of 0.3). The weight values ​​reflect the stability of the nodes in conflict resolution; higher weights indicate that the node should be prioritized for retention. All nodes of the "drugs" entity are traversed using a graph traversal algorithm (e.g., breadth-first search), and the weight values ​​are updated as node attributes in the dynamic knowledge evaluation model. Simultaneously, the relationship confidence threshold (e.g., the confidence threshold for the "drug-treatment-disease" relationship is 0.7) and the ontology evolution priority sequence (e.g., sorted by the number of people covered by the drug's indication) are matched.

[0122] For conflict regions (such as regions with identically named but different drug nodes), the model automatically adjusts the density of entity nodes: if the density of "drug" nodes in a region exceeds a preset threshold (e.g., 15 conflicting nodes out of every 100 nodes), a node merging operation is triggered, merging identically named nodes with a semantic similarity higher than 0.9 and retaining the node with the most complete attributes; if the node density is lower than the threshold (e.g., less than 3 conflicting nodes out of every 100 nodes), isolated nodes are split, generating new child nodes to refine the classification. For example, in a conflict region, there are "Aspirin" node A (attribute: antipyretic and analgesic, manufacturer X) and node B (attribute: antithrombotic, manufacturer Y). The semantic similarity is calculated to be 0.3 (below the threshold of 0.5), which is determined to be a heteronym. Instead of merging, they are assigned different unique identifiers and the relation edges are adjusted to point to the corresponding disease nodes.

[0123] After the model update is completed, the knowledge increment triggering conditions, ontology reconstruction rules, and parameter update increments are defined. The knowledge increment triggering conditions monitor the knowledge coverage completeness rate in real time; this metric is defined as the ratio of the number of aligned entities to the number of known entities in the domain. For example, when the coverage completeness rate of the "disease" entity in the graph is below 85% (a preset threshold), the parameter update process is triggered, automatically importing new disease data and initiating the alignment procedure. The ontology reconstruction rules adjust the relationship confidence threshold based on the semantic similarity gradient direction: if the semantic similarity gradient between drugs and diseases in a certain region points to the "cardiovascular and cerebrovascular diseases" category, the confidence threshold for the "drug-treatment-cardiovascular and cerebrovascular diseases" relationship is increased to 0.8, requiring the cosine similarity of their semantic vectors to be higher than this value during alignment; conversely, the relationship threshold for low-similarity regions is reduced to 0.6. The parameter update increment has a piecewise linear relationship with the current semantic consistency evaluation metric. For example, when the metric is above 0.8, each update increment is 0.1 times the current value; when it is below 0.5, the increment is reduced to 0.01 times to avoid update oscillations due to excessive conflict.

[0124] Based on the above settings, a knowledge increment simulation model is generated, and the multimodal entity alignment equation is iteratively solved using a dynamic programming algorithm. Taking the newly added text description of "new antihypertensive drug" and the corresponding drug instruction image data as an example, the text data is used to generate semantic vectors through BERT, and the image data is used to extract feature vectors through ResNet. These are input into the multimodal entity alignment equation to calculate semantic similarity, attribute matching degree (such as dosage form and dosage), and contextual consistency (such as whether it is associated with the known "hypertension" disease node). In the first iteration, if the "indication" attribute value in the text description is "refractory hypertension," while the label in the image is "hypertension," the attribute matching degree calculated by edit distance is 0.9 (preset threshold 0.6), and the semantic similarity is 0.85 (higher than the threshold 0.5), which is considered a valid alignment. An entity mapping deviation (e.g., 0.12, reflecting the difference between text and image features) is generated, and the knowledge coverage completeness is updated (assuming an increase from 84% to 86%).

[0125] If the knowledge coverage completeness rate is still lower than the preset threshold (e.g., 85%), a second parameter update is triggered: the confidence threshold of the "drug-indication" relationship is adjusted from 0.7 to 0.75, and key information (such as manufacturer and contraindications) is re-extracted from the image and added to the entity attributes. When solving the alignment equation again, the contextual consistency score of the newly added "contraindications" attribute with the existing "drug-contraindications-population" relationship is improved by 0.15, and the semantic consistency evaluation index is increased from 0.78 to 0.89, satisfying the iteration termination condition (two consecutive index changes less than 0.05). The knowledge increment simulation information is output, including the entity mapping deviation (0.12→0.08), the knowledge coverage completeness rate (86%), and the semantic consistency evaluation index (0.89).

[0126] During the iteration process, the ontology reconstruction rules are executed synchronously: based on the semantic similarity gradient direction (pointing to the "cardiovascular system disease" category), the confidence threshold of the "drug-treatment-cardiovascular disease" relationship is adjusted from 0.7 to 0.78, while the threshold of the "drug-manufacturer-manufacturer" relationship remains unchanged due to the smaller gradient influence. The parameter update increment is generated with an increment value of 0.08 according to the current semantic consistency evaluation index (0.89) being in the high segment, and is used to adjust the visual semantic embedding matrix in the cross-modal alignment parameters.

[0127] Throughout the process, the knowledge increment simulation model continuously monitors the alignment status of each entity pair. For conflicting pairs that still cannot meet the threshold after multiple iterations (such as a name conflict between a traditional Chinese medicine decoction piece and a chemical drug), they are automatically marked as "manual review" and the automatic processing flow is paused. When the preset maximum number of iterations (such as 10 times) is reached or the index converges, the simulation is terminated, and the final knowledge increment simulation information is generated to provide data support for subsequent ontology structure optimization.

[0128] This implementation achieves precise adjustment of conflict regions through real-time interaction between entity disambiguation weights and a dynamic model; a triggering mechanism based on knowledge coverage completeness ensures proactive graph updates; and a piecewise linear parameter update strategy balances update efficiency and accuracy. Taking the medical field as an example, this process can effectively handle the incremental fusion of multimodal data such as drug instruction text, structured data from quality inspection reports, and drug images. While avoiding global structural instability, it gradually improves the coverage and semantic consistency of the knowledge graph, providing a reliable knowledge foundation for applications such as clinical decision support.

[0129] Example 5:

[0130] In the process of multi-dimensional optimization model construction and ontology structure iterative adjustment, taking the knowledge graph in the financial field as an example, it is necessary to comprehensively consider variables such as entity node density, relationship confidence threshold, and knowledge update frequency to achieve a balance between knowledge coverage and conflict resolution. The specific implementation method is as follows:

[0131] Assuming a financial knowledge graph contains entities such as "enterprises," "transactions," and "risk indicators," the goal is to optimize the graph structure to improve risk assessment efficiency. First, a multi-dimensional optimization model is constructed: optimization variables include the node density of "enterprise" entities (number of nodes per unit graph area), the confidence threshold of the "transaction-related enterprise" relationship, and the knowledge update frequency of risk indicator data (e.g., daily / weekly updates). The optimization objectives are to maximize knowledge coverage efficiency (number of covered enterprises / total number of enterprises in the industry) and minimize semantic conflict density (number of conflicting transaction relationships / total number of transaction relationships), with their weights set at 7:3 according to regulatory requirements. Constraints include ontology evolution boundaries (e.g., only enterprise nodes with transaction records within the past year are allowed to be adjusted) and attribute matching accuracy thresholds (e.g., the matching error of the "risk level" attribute must be less than 5%).

[0132] The initial population generation of the genetic algorithm is based on a feasible range of entity node density and relationship confidence threshold. For example, the node density ranges from [50, 200] nodes / m², the confidence threshold is [0.6, 0.9], each individual is represented by a two-dimensional vector (density, threshold), and 50 individuals are randomly generated to form the initial population. The fitness function is calculated as follows:

[0133]

[0134] Where 1000 represents the total number of enterprises in the industry, and 5000 represents the total number of transaction relationships. A crossover operation (e.g., single-point crossover, crossover probability 0.8) is performed on the initial population, randomly selecting two individuals to exchange some gene loci to generate offspring individuals; a mutation operation (mutation probability 0.01) randomly perturbs the gene loci of individuals, such as node density ±10 and threshold ±0.05. An intermediate optimized path set is generated through crossover and mutation. Then, through local search (e.g., fine-tuning the gene loci of each individual by ±5%) and an elite retention strategy (retaining the top 20% of individuals in fitness), 20 optimal individuals are selected to form the initial optimized path set, such as optimized combinations including (120, 0.75) and (150, 0.8).

[0135] Simulated annealing starts with an initial set of optimized paths and performs global optimization. Taking an individual (120, 0.75) as an example, a perturbation vector (+8, -0.03) is randomly generated, resulting in a new individual (128, 0.72). The fitness value of the new individual is calculated. If the new fitness is higher than the original fitness, it is directly accepted; if it is lower, the acceptance probability is calculated using the Metropolis criterion.

[0136]

[0137] The initial temperature T is set to 100, and is reduced by a factor of 0.95 after each iteration. For example, if the original fitness is 0.82 and the new fitness is 0.80, the acceptance probability at T=100 is exp(0.02 / 100)≈0.9998, indicating a high probability of accepting a poor solution to escape local optima. After 200 iterations, the temperature drops to near 0, and the optimal solution (135, 0.78) converges. This means that when the node density is 135 nodes / m² and the confidence threshold is 0.78, the knowledge coverage efficiency and conflict density are balanced.

[0138] The optimal knowledge fusion path parameters are applied to ontology structure adjustments. For example, based on the optimal node density of 135 nodes / ㎡, "zombie enterprise" nodes in low-density areas (e.g., 80 nodes / ㎡) are merged, and their transaction relationships are migrated to active enterprise nodes. In high-density areas (e.g., 180 nodes / ㎡), industry sub-sector nodes are split, such as splitting "technology enterprises" into subcategories like "artificial intelligence enterprises" and "big data enterprises." The confidence threshold for "transaction-related enterprise" relationships is adjusted to 0.78, requiring that transaction records between related enterprises simultaneously meet the following conditions: transaction amount ≥ 1 million RMB (attribute constraint) and transaction frequency ≥ once per quarter (topology constraint); otherwise, the relationship is considered invalid. The knowledge update frequency is increased from once per week to once per day to capture risk changes of high-frequency trading enterprises in real time.

[0139] Based on real-time ontology topology network reasoning, the semantic consistency index of entities is calculated. Taking "high-risk enterprises" as an example, the model analyzes the risk level distribution of related enterprises by traversing their transaction relationship chains: if 6 out of a certain enterprise's 10 related enterprises have a risk level of "high", the semantic consistency index is 0.6 (the preset threshold is 0.5), triggering a risk warning. Combining the optimal knowledge fusion path parameters and historical evolution trajectory (such as the node density being adjusted three times in the past six months), a dynamic maintenance strategy is generated: the real-time monitoring frequency of enterprise nodes with frequent fluctuations in risk level is increased, and a manual verification process is initiated for "guarantee relationship" edges with confidence levels below the threshold.

[0140] In e-commerce analogies, multi-dimensional optimization models can be constructed for entities such as "products," "users," and "purchase records." The initial population of the genetic algorithm can contain individuals with densities of 200 product nodes / m² and a confidence threshold of 0.7, or 250 nodes / m² and a threshold of 0.65. The fitness function focuses on knowledge coverage efficiency (e.g., the number of product categories covered). After optimization using simulated annealing, a combination of 220 nodes / m² and a confidence threshold of 0.72 can be obtained, guiding the merging of nodes from less popular product categories and improving the confidence requirement for the "bestselling product - user preference" relationship. The knowledge update frequency is adjusted from hourly to minutely to respond in real-time to changes in user behavior during promotional activities.

[0141] This implementation effectively solves multivariate optimization problems by combining genetic algorithms and simulated annealing algorithms, avoiding getting trapped in local optima. Taking the financial sector as an example, the optimized knowledge graph can more accurately identify the risk transmission paths of related enterprises. Simultaneously, density adjustment and threshold settings ensure the simplicity of the graph structure and query efficiency. The dynamic maintenance strategy combines real-time topology analysis with historical data, enabling the adaptive evolution of the knowledge graph. This meets regulatory requirements for risk coverage while avoiding decreased reasoning efficiency due to overly complex structures, providing a highly interpretable and dynamically adaptable knowledge infrastructure for domain applications.

[0142] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0143] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for constructing and maintaining an AI knowledge graph based on ontology engineering, characterized in that, Includes the following steps: Acquire multi-source heterogeneous data and extract ontological concepts and semantic relationship sets to construct an initial graph ontological model; Domain knowledge constraint rules are loaded onto the initial graph ontology model to generate a standardized knowledge fusion framework. Cross-modal knowledge fusion is then performed by combining entity alignment parameters to obtain a dynamic knowledge evaluation model. Ontology topology reasoning is performed based on a dynamic knowledge evaluation model to generate entity relationship extension paths, and time-series knowledge evolution simulation data is constructed by combining knowledge update frequency parameters. A semantic conflict resolution model is constructed and trained using temporal knowledge evolution simulation data to generate a knowledge fusion optimization model, which then outputs graph maintenance parameters. The graph maintenance parameters include at least entity disambiguation weights, relation confidence thresholds, and ontology evolution priority sequences. Based on graph maintenance parameters, dynamic knowledge evaluation model and entity alignment parameters, knowledge increment simulation information is generated; By combining knowledge increment simulation information with ontology evolution constraints, a multi-dimensional optimization model is constructed and the ontology structure is iteratively adjusted to generate the optimal knowledge fusion path parameters. Based on the real-time ontology topology network, reason about the current knowledge fusion path parameters, calculate the entity semantic consistency index, and combine the optimal knowledge fusion path parameters with the historical evolution trajectory to update the knowledge graph maintenance strategy to achieve the goal of dynamic ontology matching.

2. The method for constructing and maintaining an AI knowledge graph based on ontology engineering according to claim 1, characterized in that, The temporal knowledge evolution simulation data includes at least a semantic conflict set, attribute constraint parameters, ontology evolution boundary set, and dynamic resolution priority sequence. The knowledge increment simulation information includes at least entity mapping deviation, knowledge coverage completeness rate, and semantic consistency evaluation index.

3. The method for constructing and maintaining an AI knowledge graph based on ontology engineering according to claim 1, characterized in that, The process of acquiring multi-source heterogeneous data, extracting ontological concepts and semantic relationship sets, and constructing an initial graph ontological model includes the following steps: Structured parsing and cleaning of multi-source heterogeneous data to generate a standardized set of knowledge units; Ontology concept extraction is performed on a set of knowledge units, wherein the extraction includes one or more of semantic clustering, attribute association mapping, concept hierarchy partitioning, and synonym disambiguation; Based on the conceptual hierarchy partitioning results, an initial graph ontology model is constructed, wherein the model includes a set of entity nodes, a set of relation edges, and attribute constraint rules; Semantic relation strength analysis is performed on the initial graph ontology model to generate a dynamic relation weight matrix, and domain knowledge constraint labels are assigned to each entity node.

4. The method for constructing and maintaining an AI knowledge graph based on ontology engineering according to claim 1, characterized in that, The process of loading domain knowledge constraint rules onto the initial graph ontology model, generating a standardized knowledge fusion framework, and combining entity alignment parameters to perform cross-modal knowledge fusion to obtain a dynamic knowledge evaluation model includes the following steps: The initial knowledge fusion framework is generated by loading domain knowledge constraint rules onto the initial graph ontology model to constrain the range of entity attribute values. Load cross-modal alignment parameters onto the first knowledge fusion framework to generate a standardized knowledge fusion framework; Cross-modal knowledge fusion based on a standardized knowledge fusion framework includes the following specific processes: A multimodal entity alignment equation is constructed, which includes at least a semantic similarity equation, an attribute matching degree equation, and a context consistency equation. The multimodal entity alignment equation is iteratively solved using a dynamic programming algorithm to obtain a set of semantic conflicts, attribute constraint parameters, and an ontology evolution boundary set. The generation of the ontology evolution boundary set includes the following steps: Based on the semantic conflict set, calculate the conflict density of each entity node; Identify regions in the standardized knowledge fusion framework where the conflict density is no greater than a preset threshold, and generate an ontology evolution boundary set.

5. The method for constructing and maintaining an AI knowledge graph based on ontology engineering according to claim 1, characterized in that, The process of constructing a semantic conflict resolution model, training it with temporal knowledge evolution simulation data, generating a knowledge fusion optimization model, and then outputting graph maintenance parameters includes the following steps: A semantic conflict resolution model is constructed based on the initial graph ontology model; The semantic conflict resolution model is trained and validated using time-series knowledge evolution simulation data to generate a knowledge fusion optimization model. Input the real-time entity alignment parameters into the knowledge fusion optimization model to predict the set of ontology evolution boundaries; Based on the predicted set of ontology evolution boundaries, graph maintenance parameters are output, wherein the graph maintenance parameters include at least entity disambiguation weights, relation confidence thresholds, and ontology evolution priority sequences.

6. The method for constructing and maintaining an AI knowledge graph based on ontology engineering according to claim 5, characterized in that, This also includes feature reconstruction of time-series knowledge evolution simulation data, specifically: An initial multimodal knowledge input tensor is constructed based on a set of semantic conflicts, attribute constraint parameters, and ontology evolution boundaries. The initial multimodal knowledge input tensor is normalized and its features are enhanced to generate the final multimodal knowledge input tensor. Based on the dynamic resolution priority sequence, a knowledge fusion optimization label tensor is constructed. The training sample set is formed by combining the final multimodal knowledge input tensor with the knowledge fusion optimization label tensor.

7. The method for constructing and maintaining an AI knowledge graph based on ontology engineering according to claim 6, characterized in that, The prediction-based ontological evolution boundary set outputs graph maintenance parameters, including the following steps: Extract the entity node with the lowest semantic conflict density from the predicted ontology evolution boundary set and generate entity disambiguation weights; Based on the topological connectivity of the predicted ontology evolution boundary set, the distribution structure of the relation confidence threshold is fitted to generate an ontology evolution priority sequence. The semantic similarity gradient direction of the predicted ontology evolution boundary set is calculated and normalized into a knowledge fusion baseline vector, which is the direction of relation confidence adjustment.

8. The method for constructing and maintaining an AI knowledge graph based on ontology engineering according to claim 1, characterized in that, The process of generating knowledge increment simulation information based on graph maintenance parameters, dynamic knowledge evaluation model, and entity alignment parameters includes the following steps: The entity disambiguation weights are mapped to the dynamic knowledge evaluation model, the relationship confidence threshold and ontology evolution priority sequence are matched, the entity node density in the conflict area is adjusted, the dynamic knowledge evaluation model is updated, the knowledge increment triggering conditions, ontology reconstruction rules and parameter update increments are defined, and the knowledge increment simulation model is generated. Based on the knowledge increment simulation model, the multimodal entity alignment equation is iteratively solved using a dynamic programming algorithm to generate knowledge increment simulation information, specifically including: When the knowledge increment triggering condition is met, update the entity mapping deviation, knowledge coverage completeness rate and semantic consistency evaluation index, and resolve the multimodal entity alignment equation until the simulation termination condition is reached. The knowledge increment triggering condition includes triggering parameter updates when the current knowledge coverage completeness rate is not greater than a preset completeness rate threshold; the ontology reconstruction rule includes adjusting the relationship confidence threshold based on the semantic similarity gradient direction; the parameter update increment has a piecewise linear relationship with the current semantic consistency evaluation index.

9. The method for constructing and maintaining an AI knowledge graph based on ontology engineering according to claim 1, characterized in that, The process of constructing a multi-dimensional optimization model and iteratively adjusting the ontology structure to generate optimal knowledge fusion path parameters includes the following steps: A multi-dimensional optimization model is constructed, in which the optimization variables include entity node density, relation confidence threshold and knowledge update frequency, the optimization objectives include maximizing knowledge coverage efficiency and minimizing semantic conflict density, and the constraints include ontology evolution boundary constraints and attribute matching accuracy threshold. The multi-dimensional optimization model is initially solved using a genetic algorithm to generate an initial set of optimization paths. Based on the initial set of optimized paths, the simulated annealing algorithm is used for global optimization to generate the optimal knowledge fusion path parameters.

10. The method for constructing and maintaining an AI knowledge graph based on ontology engineering according to claim 9, characterized in that, The process of using a genetic algorithm to initially solve the multi-dimensional optimization model and generate an initial set of optimization paths includes the following steps: A chromosome coding structure is constructed based on entity node density and relationship confidence threshold to generate an initial population; The fitness function is calculated based on knowledge coverage efficiency and semantic conflict density. Crossover and mutation operations are performed on the initial population to generate a set of intermediate optimized paths. A local search and elite retention filter are performed on the intermediate optimized path set to generate an initial optimized path set.

Images