Large-scale ontology processing method based on anchor point adaptive dynamic division

Abstract

The invention belongs to the technical field of ontology engineering, and particularly relates to a large-scale ontology processing method based on anchor point self-adaptive dynamic division, which comprises the following steps: block parameter integration based on tree coding: integrating core parameters of blocks through a tree structure, constructing five types of block construction parameters related to layer nodes and decision probability parameters of the five types of block construction parameters, logic layer nodes adopt two types of logic operators; based on block dynamic adjustment and optimization of an evolutionary algorithm, automatically iterating an advanced combination tree form of block features by adopting crossover and variation of the evolutionary algorithm for the tree-shaped structure individuals; based on the decision probability learning of elite knowledge, adjusting the decision probability of constructing layer nodes in the tree by utilizing knowledge contained in an elite solution; and based on the step length self-adjustment of the iteration process, recording current iteration information, and adaptively setting the adjustment step length of the decision probability. According to the method, a large-scale ontology is divided into a plurality of small-scale ontology clusters with good structural information and uniform fragment concepts in a full-dynamic manner, so that the quality of subsequent matching is effectively ensured.

Description

Technical Field

[0001] The invention belongs to the technical field of ontology engineering, and in particular relates to a large-scale ontology processing method based on anchor point adaptive dynamic partitioning. Background Art

[0002] Ontology is a clear and sharable conceptual model used to describe entities and their relationships in a specific domain. It provides a unified framework to express domain knowledge by defining concepts, attributes and structures. Ontology matching refers to the process of identifying semantically equivalent entity pairs between heterogeneous ontologies, aiming to solve the semantic heterogeneity problem caused by different naming, structure or expression methods between different ontologies. With the widespread application of ontology technology in fields such as artificial intelligence, semantic web and knowledge graph, the scale of ontologies has increased dramatically, and matching between large-scale ontologies has become an important research direction in the field of ontology matching.

[0003] An effective method for handling large-scale ontology matching is the "divide and conquer" strategy. Specifically, through ontology partitioning, the large-scale source ontology and target ontology are split into several sub-ontology modules, and then the corresponding sub-modules are matched in combination with similarity features, so that the overall task is converted into a matching problem between multiple small-scale sub-modules. High-quality ontology segmentation is a prerequisite for ensuring high-quality matching results. If the segmentation quality is poor, the distribution of the real matching entities will not match the actual segmentation, and even the most advanced similarity features will find it difficult to improve the overall quality of large-scale ontology matching results. Therefore, in large-scale ontology matching, segmentation quality is usually more important than the construction of similarity features. However, before the completion of the application of this patent, the existing large-scale ontology matching methods all adopt a segmentation strategy based on simple rules, directly determining the segmentation only through one-time or multiple segmentations, and focusing the work on optimizing similarity features. These methods are static segmentation strategies based on simple rules, which weaken the integrity and consistency of the ontology structure, resulting in serious loss of semantic information before and after segmentation, making it difficult to achieve the best matching effect.

[0004] In the research field of ontology partitioning, the existing technology literature Stuckenschmidt H, Schlicht A. Structure-based partitioning of large ontologies [M] / / Modular ontologies: Concepts, theories and techniques for knowledge modularization. Berlin, Heidelberg: Springer Berlin Heidelberg, 2009: 187-210. proposed a structure-based ontology partitioning method, but this method is difficult to apply between source and target ontologies with large structural differences, and it is easy to ignore the semantic association between concepts and lacks sufficient generality; the literature Hamdi F, Safar B, Reynaud C, et al. Alignment-based partitioning of large-scale ontologies[J]. Advances in knowledgediscovery and management, 2010: 251-269. proposed an ontology partitioning method based on matching results, but this method is too dependent on the quality of the initial matching results. For areas with less association or no matching in the preliminary matching, it may lead to incorrect partitioning or omission problems; the literature Xue X, Tsai PW, Chen J. Large-scale complexontology matching through anchor-based semantic partitioning technique andconfidence matrix based evolutionary algorithm[J]. Applied Soft Computing,2022, 128: 109516. An anchor-based ontology partitioning method is proposed, but the anchor selection in this method adopts a fixed strategy, which is a static selection method and may not cover the diverse structural features in the ontology; Literature Xue X, Sun D, ​​Shankar A, et al. Efficient large-scale biomedical ontology matching with anchor-based biomedical ontology partitioning and compact geometric semantic genetic programming[J].Journal of Industrial Information Integration, 2024:100637. A variety of block strategies were considered, but the strategy still relies on fixed center points and diffusion methods, and cannot avoid problems such as missing block structure or unbalanced scale; Reference Lv Qing, Zhou Xin, Li Fenglian. Large-scale anatomical ontology matching under dynamic block adjustment mechanism[J]. Journal of Computer Applications, 2023, 40(01): 136-140,146. A dynamic block adjustment mechanism was proposed, but the mechanism is essentially still a one-time completion of the ontology division, and only a small range of adjustments are made by redistributing the matched entities to the target block, which is far from achieving the effect of full dynamic optimization of the block. .

[0005] The existing ontology partitioning methods have the following disadvantages: (1) Faced with the complex structure of the ontology, it is necessary to manually select the block strategy and formulate rules, which easily leads to the loss of semantic information and has a negative impact on the subsequent matching quality; (2) The developed block strategy is difficult to reuse in different scenarios, which increases the actual operation block cost; (3) Once the partitioning parameters are determined, they cannot be flexibly adjusted. There is a lack of dynamic optimization mechanism, making it difficult to discover the best entity partitioning solution. Summary of the invention

[0006] The purpose of the present invention is to provide a large-scale ontology processing method based on anchor point adaptive dynamic partitioning to solve the deficiencies in the above-mentioned background technology.

[0007] In order to achieve the above object, the present invention adopts the following technical scheme: A large-scale ontology processing method based on anchor point adaptive dynamic partitioning, the specific steps are as follows: S1: Block parameter integration based on tree coding, which integrates the core parameters of the block through a tree structure. The construction layer nodes involve five types of block construction parameters and decision probability parameters that determine the values ​​of these parameters. The two directly control the diffusion strategy of anchor points in the block construction process. The block construction parameters include branch constraints, anchor point constraints, structure constraints, scale constraints and depth constraints; the logic layer nodes use two types of logical operators, namely AND operations and OR operations, which are used to merge the block results of each branch. The tree coding can flexibly represent various block rules; S2: Dynamic optimization of blocks based on evolutionary algorithms, directly operates on tree-structured individuals, uses evolutionary algorithm crossover and mutation to automatically iterate out advanced combination tree forms of block features, and fully dynamically optimizes blocks; S3: Decision probability learning based on elite knowledge, using the knowledge contained in the elite solution to adjust the decision probability of the building layer nodes in the tree, guide the generation of block building parameters, and improve the decision quality; S4: Based on the self-adjustment of the step size of the iterative process, the current iteration information is recorded and the adjustment step size of the decision probability is adaptively set.

[0008] Preferably, the specific contents of the block parameter integration based on tree coding in step S1 include: The core parameters of building blocks are integrated into the syntax tree. Each syntax tree corresponds to a given block strategy. It adopts a multi-level tree encoding structure and a bottom-up tree decoding method. The tree structure contains three types of nodes: leaf nodes, construction layer nodes, and logic layer nodes: (1) Leaf nodes are used to input anchor point candidate sets.

[0009] (2) The construction layer nodes are used to store the operators required in the block process and the corresponding decision probabilities in the node generation process. The specific contents are as follows: The operator branch constraint indicates the maximum number of branches that the syntax tree is allowed to generate to integrate the block parameters; The operator anchor constraint represents the anchor selected from the anchor candidate set for expanding the block during the evolution process; The operator structure constraint means that when each anchor point spreads to form its own block, it looks for a structural form of highly cohesive and low-coupling entities belonging to the block. There are three common structures: is-a, part-of, and equivalent. The operator size constraint indicates the maximum block size allowed, that is, the maximum number of entities that the current block can accommodate; The operator depth constraint represents the shortest path for each diffusion anchor point to reach a certain entity c along the current structure, expressed as: ; in, anc represents the diffusion anchor point, c represents the entity to be diffused, fa Indicates the common ancestor of the diffusion anchor point and the entity to be diffused; Indicates the structural depth of the closest common ancestor between the diffusion anchor point and the entity to be diffused; Indicates the common ancestor of the diffusion anchor point and the entity to be diffused. anc The shortest path; Indicates the common ancestor of the diffusion anchor point and the entity to be diffused. c The shortest path; In addition to the above five types of operators, each construction layer node also contains an additional parameter - decision probability, which represents the probability of each value being selected for multiple value choices of this type of operator in a specific type of construction layer node. Multiple decision probability fragments are integrated through a probability vector, in which each element corresponds to the probability value of a specific value selection, and the number of decision probability fragments is equal to the number of values ​​supported by this type of operator.

[0010] (3) The logic layer nodes are used to store the operators used in the secondary optimization block process: The AND operator represents the intersection of the diffusion blocks of each branch, that is, extracting the same anchor concept and block structure between each branch; The or operator represents the union of the diffusion blocks of each branch, that is, merging the different anchor point concepts and block structures between branches.

[0011] Preferably, the process of obtaining the anchor point candidate set of leaf nodes in (1) the block parameter integration based on tree coding in step S1 is as follows: Source Get the context concept of a certain structural relationship (is-a, part-of or equivalent) for the source ontology All entities in the dataset are sorted in descending order according to their importance, and the top 50% of entities are selected. These entities are limited in size and rich in context information. Entities selected from different structures are aggregated to generate an anchor source entity training set. The same method is used for the target ontology. Generate the anchor target entity training set, which together with the anchor source entity training set constitutes the anchor training set: Similarity analysis: Utilize string-based similarity measurement technology to analyze the similarity between the source entity and the target entity in the anchor training set and generate a similarity matrix.

[0012] Similarity screening: For the obtained similarity matrix, extract the similarity values ​​that are both the maximum values ​​of the rows and columns in the matrix, and filter them in combination with the preset similarity threshold. If the similarity of a pair of source entity-target entity exceeds the threshold, it is considered to be a correct matching pair and the matching pair is added to the anchor candidate set; the above extraction and threshold filtering operations are repeated to generate an anchor candidate set, which contains one-to-one corresponding anchor source entities and anchor target entities.

[0013] Preferably, the values ​​of different types of operators in the construction layer in the block parameter integration based on tree coding in step S1 (2) are based on: The branch constraint sets the value range according to the total number of constraint conditions and the maximum number of permutations and combinations allowed by the grammatical structure; The anchor point constraint sets the value range according to the size of the anchor point candidate set; The structural constraint is one of is-a, part-of, or equivalent; The depth constraint sets the value range based on the maximum in-and-out degree of all structures; The scale constraint sets the value range based on twice the average block size corresponding to the current number of anchor points.

[0014] Preferably, the basis for taking values ​​of different types of operators in the construction layer in step S1 also includes that when the value supported by the operator is a continuous variable, the continuous variable needs to be divided into multiple sub-intervals, and a probability vector is constructed based on the range of each sub-interval, and then the target interval is determined according to the decision probability, and finally the specific value of the operator is randomly generated within the selected interval.

[0015] Preferably, the specific process of the block dynamic tuning based on the evolutionary algorithm in step S2 is as follows: (1) Configure the control parameters of the evolutionary algorithm, such as population size, number of iterations, crossover probability, and mutation probability; (2) Randomly initialize individuals in the population - syntax trees. Individuals divide large-scale source and target ontologies according to syntax rules, generate corresponding block results, and initialize the decision probabilities of construction layer nodes with equal probability; (3) Execute evolution, evaluate the fitness value of individual block results, generate the next generation of individuals through syntax-based single-point crossover, subtree crossover, single-point mutation, and subtree mutation, and automatically build a more advanced syntax tree structure; (4) According to the elite solution, the decision probability of the nodes in the construction layer is updated, and half of the individuals are selected from the population after crossover and mutation using the roulette method, which is dominated by the evolutionary strategy; the other half of the individuals are generated based on the updated decision probability, which is dominated by the decision probability. The two together constitute the next generation population; (5) Return to step (3) until the termination condition is met, output the elite solution, and obtain the optimal partitioning result of the source ontology and the target ontology.

[0016] Preferably, the specific content of the decision probability learning based on elite knowledge in step S3 includes: At the end of each generation, the operator value of the construction layer node in the elite tree structure is obtained and compared with the stored probability vector. If the operator value is the same, the corresponding decision probability is increased by one learning step. , the update formula is: ; For decision probabilities with different operator values, the mean of the learning step length is reduced in turn, and the update formula is: , num It represents the total number of categories of the current operator value. All probability vectors are updated, guiding the generation direction of the next generation of individuals and strengthening elite knowledge.

[0017] Preferably, the specific contents of the self-adjustment of the step length based on the iterative process in step S4 include: Calculate learning step length based on iteration information , adaptively adjust the change range of decision probability, the specific formula is: ; Indicates the current elite solution f Metrics, Indicates the last elite solution f Metrics; is a constant used to limit the initial value of the step length; T Indicates the current iteration number of the population; is a scaling factor that limits the rate at which the step size changes.

[0018] Compared with the prior art, the present invention has the following beneficial effects: (1) The syntax tree structure is used to integrate the block parameters. Driven by the ontology matching quality, the blocks with reasonable structure, optimal scale and complete semantic information are constructed in a fully dynamic manner, which improves the block quality of large-scale ontologies. This is a pioneering and new idea. (2) The dual learning mechanism is adopted, with adaptive learning of decision probability and adaptive change of tuning step size, which avoids the limitations of subjective parameter setting and enumeration parameters and realizes flexible parameter adjustment. The block strategy can be reused in different scenarios, reducing the actual operation block cost. This mechanism balances the diversity and convergence of the population and promotes the model to extract rich high-level features. (3) This method does not use all anchor points indiscriminately for diffusion segmentation, but avoids concept duplication and structural confusion through reasonable selection. The final segmentation result may not completely cover all entities in the source (or target) ontology, but the number of isolated entities is significantly reduced, and the entities that are automatically divided out will not have a substantial impact on the semantics of the ontology and the final matching results. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flow chart of the method of the present invention; Figure 2 A schematic diagram of block-by-block individual encoding of an embodiment of the method of the present invention; Figure 3 A schematic diagram of anchor point selection in an embodiment of the method of the present invention; Figure 4 A flow chart of population iterative optimization in an embodiment of the method of the present invention; Figure 5 A schematic diagram of decision probability update in an embodiment of the method of the present invention; Figure 6The figure is a comparison diagram of the block size and standard deviation of the embodiment of the method of the present invention and other clustering block methods in 30 independent running experiments. DETAILED DESCRIPTION

[0020] The technical solution of the present invention is described clearly and completely below in conjunction with the accompanying drawings and embodiments.

[0021] The ontology matching datasets used in this embodiment are six large-scale public test datasets officially provided by OAEI. The relevant descriptions are shown in Table 1.

[0022] Table 1 Description of OAEI anatomical and large biomedical ontologies

[0023] like Figure 1 As shown, a large-scale ontology processing method based on anchor point adaptive dynamic partitioning includes the following steps: S1: Block parameter integration based on tree coding, which integrates the core parameters of the block through a tree structure. The construction layer nodes involve five types of block construction parameters and decision probability parameters that determine the values ​​of these parameters. The two directly control the diffusion strategy of anchor points in the block construction process. The block construction parameters include branch constraints, anchor constraints, structure constraints, scale constraints and depth constraints; the logic layer nodes use two types of logical operators, AND operations and OR operations, to merge the block results of each branch. The tree coding can flexibly represent various block rules, including: Integrate the core parameters of the building block into the syntax tree, such as Figure 2 As shown in the figure, each syntax tree corresponds to an established block strategy. It adopts a multi-level tree encoding structure and a bottom-up tree decoding method. The tree structure contains three types of nodes: leaf nodes, construction layer nodes, and logic layer nodes. The generation of the syntax tree should strictly follow the arrow direction and logical hierarchy. The dotted box represents the operator that can be generated optionally, rather than the part that must be generated.

[0024] (1) Leaf nodes are anchors input from the anchor candidate set. They are contextual concepts based on a certain structural relationship (is-a, part-of, or equivalent) and are used to represent the source ontology. All entities in the dataset are sorted in descending order according to their importance, and the top 50% of entities are selected. These entities are limited in size and rich in context information. Entities selected from different structures are aggregated to generate an anchor source entity training set. The same method is used for the target ontology. Generate an anchor target entity training set, which together with the anchor source entity training set constitutes the anchor training set.

[0025] The anchor point selection process is as follows Figure 3As shown, the specific contents include: using the string-based similarity measurement technology to calculate the similarity between the source entity and the target entity in the anchor training set, and generate a similarity matrix. The string-based similarity measurement method is: ;in, 、 Respectively represent the ontology in the anchor training set and ontology The corresponding entity; Representing Entities and The common feature score between them is used to measure the common similarity between two entities; Representing Entities and The difference score between them reflects the degree to which information cannot be shared between the two entities; It is an improvement on the traditional SMOA measurement method, which indicates that the entity and The common prefix length of the strings between them.

[0026] Extract the similarity values ​​that are both the maximum values ​​of the rows and columns in the matrix, and filter them with the preset similarity threshold of 0.99. If the similarity of a source entity-target entity pair exceeds the threshold, it is considered to constitute a correct matching pair, and the matching pair is added to the anchor candidate set. The above extraction and threshold filtering operations are repeated, and all source entity-target entity matching pairs that meet the requirements are added to the anchor candidate set, as follows: ;in, represents the determined anchor candidate set, Indicates the anchor point under the is-a structure. Indicates the anchor point under the part-of structure. Indicates the anchor point under the equivalent structure.

[0027] (2) The construction layer nodes are used to store the operators required in the block process and the corresponding decision probabilities in the node generation process: The operator branch constraint indicates the maximum number of branches that the syntax tree is allowed to generate to integrate the block parameters; The operator anchor constraint represents the anchor selected from the anchor candidate set for expanding the block during the evolution process; The operator structure constraint means that when each anchor point spreads to form its own block, it looks for a structural form of highly cohesive and low-coupling entities belonging to the block. There are three common structures: is-a, part-of, and equivalent. The operator size constraint indicates the maximum block size allowed, that is, the maximum number of entities that the current block can accommodate; The operator depth constraint represents the shortest path for each diffusion anchor point to reach a certain entity c along the current structure, expressed as: ;in, anc represents the diffusion anchor point, c represents the entity to be diffused, fa Indicates the common ancestor of the diffusion anchor point and the entity to be diffused; Indicates the structural depth of the closest common ancestor between the diffusion anchor point and the entity to be diffused; Indicates the common ancestor of the diffusion anchor point and the entity to be diffused. anc The shortest path; Indicates the common ancestor of the diffusion anchor point and the entity to be diffused. c The shortest path; In addition to the above five types of operators, each construction layer node also contains an additional parameter - decision probability, which represents the probability of each value being selected for multiple value choices of this type of operator in a specific type of construction layer node. Multiple decision probability fragments are integrated through a probability vector, in which each element corresponds to the probability value of a specific value selection, and the number of decision probability fragments is equal to the number of values ​​supported by this type of operator.

[0028] The values ​​of different types of operators in the construction layer, the value range of branch constraints is set according to the total number of constraints and the maximum number of permutations and combinations allowed by the grammatical structure; the value range of anchor constraints is set according to the size of the anchor candidate set; the structure constraint is one of is-a, part-of, and equivalent; the value range of depth constraints is set according to the maximum in-and-out degree of all structures; the value range of scale constraints is set according to twice the average block size corresponding to the current number of anchors. The value ranges of different operators in this embodiment are initialized as follows: Branch constraint: Comprehensively consider all constraints that satisfy the grammatical rules, set to 1 to 15, which is a discrete variable; Anchor constraint: Take the first 10% to 90% of the anchor candidate set after sorting in descending order according to the number of all context concepts, which is a continuous variable with a division interval of 10%; Structural constraints: is-a / part-of / equivalent, which is a discrete variable; Size limit: set to 10 to , is a continuous variable, divided into intervals of 10, Represents the source entity and target ontology The maximum number of entities included; Depth constraint: set to 1 to 5 based on the maximum in-and-out degree, a discrete variable; It should be noted that when the value supported by the operator is a continuous variable, the continuous variable needs to be divided into multiple sub-intervals, and a probability vector is constructed based on the range of each sub-interval. The target interval is then determined based on the decision probability, and finally the specific value of the operator is randomly generated within the selected interval.

[0029] (3) The logic layer nodes are used to store the operators used in the secondary optimization block process: The AND operator represents the intersection of the diffusion blocks of each branch, that is, extracting the same anchor concept and block structure between each branch; The or operator represents the union of the diffusion blocks of each branch, that is, merging the different anchor point concepts and block structures between branches.

[0030] S2: Dynamic optimization of blocks based on evolutionary algorithms, direct operation of tree-structured individuals, automatic iteration of advanced combination tree forms of block features using evolutionary algorithm crossover and mutation, full dynamic optimization of blocks, population iteration optimization process such as Figure 4 The specific contents include: (1) Configure the control parameters of the evolutionary algorithm. In this embodiment, the population size is set to 100, the number of iterations is set to 2000, the crossover probability is set to 0.8, and the mutation probability is set to 0.1. (2) Randomly initialize individuals in the population - syntax trees. Individuals divide large-scale source and target ontologies according to syntax rules, generate corresponding block results, and initialize the decision probabilities of construction layer nodes with equal probability; (3) Execute evolution, evaluate the fitness value of individual block results, generate the next generation of individuals through syntax-based single-point crossover, subtree crossover, single-point mutation, and subtree mutation, and automatically build a more advanced syntax tree structure; The string-based, grammar-based, and structure-based similarity feature methods with equal weights are used as the ontology matching model to evaluate the performance of the block, and the matching results are generated at the block level. The formula for evaluating the performance of the block is: ; ; ;in, , Respectively represent the source entity and target ontology The number of entities involved in the match; Representing Entities and Calculated similarity score; Coverage Represents the recall index, which is used to evaluate the completeness of matching information; ConfidenceRepresents the precision index, which is used to evaluate the accuracy of matching information; Indicates the evaluation results that comprehensively consider completeness and accuracy; (4) According to the elite solution, the decision probability of the nodes in the construction layer is updated, and half of the individuals are selected from the population after crossover and mutation using the roulette method, which is dominated by the evolutionary strategy; the other half of the individuals are generated based on the updated decision probability, which is dominated by the decision probability. The two together constitute the next generation population; (5) Return to step (3) until the termination condition is met, output the elite solution, and obtain the optimal partitioning result of the source ontology and the target ontology.

[0031] S3: Decision probability learning based on elite knowledge, using the knowledge contained in the elite solution to adjust the decision probability of the construction layer nodes in the tree, further guide the generation of block construction parameters, and improve the quality of decision making. Specific contents include: At the end of each generation, the operator value of the construction layer node in the elite tree structure is obtained and compared with the stored probability vector. If the operator value is the same, the corresponding decision probability is increased by one learning step. , the update formula is: ; For decision probabilities with different operator values, the mean of the learning step length is reduced in turn, and the update formula is: , num It represents the total number of categories of the current operator value. All probability vectors are updated, guiding the generation direction of the next generation of individuals and strengthening elite knowledge.

[0032] The decision probability is updated as Figure 5 As shown, taking the depth constraint as an example, the discrete probability vector is initialized with equal probability for different values ​​of the depth constraint = 0.2, and the individual randomly generates a depth constraint = 2 according to probability. If the individual is selected as an elite individual, the decision probability corresponding to the depth constraint value of 2 is updated to 0.228 according to the learning step LR = 0.028, and the decision probabilities of the other depth constraint values ​​are proportionally reduced to 0.193.

[0033] Taking the anchor point constraint as an example of the continuous probability vector, the decision probability of different anchor point constraint value intervals is initialized with equal probability = 0.125, the individual randomly generates the anchor point constraint value interval = 21%~30% according to probability, and then randomly generates the anchor point constraint = 25% in the selected interval. If the individual is selected as an elite individual, the decision probability corresponding to the anchor point constraint value interval of 21%~30% is updated to 0.153 according to the learning step LR=0.028, and the decision probabilities of the remaining anchor point constraint value intervals are proportionally reduced to 0.121; the updated decision probability is used to guide the generation of the next generation of individuals.

[0034] S4: Based on the self-adjustment of the step size of the iterative process, record the current iteration information, and adaptively set the adjustment step size of the decision probability. The specific contents include: Calculate learning step length based on iteration information , adaptively adjust the change range of decision probability, the specific formula is: ; Indicates the current elite solution f Metrics, Indicates the last elite solution f Metrics; is a constant used to define the initial value of the step length. In this embodiment, =1; T Indicates the current iteration number of the population; is a proportional factor used to limit the rate of change of the step size. In this embodiment, =10.

[0035] This example compares the results of the anchor adaptive dynamic partitioning algorithm (DAPA) with the matching systems AML, LogMap, LogMapBio, ATMatch, ATBox, Wiktionary, and Lily participating in OAEI. The experimental results of DAPA are the average of 30 independent runs of the algorithm.

[0036] Table 2 Evaluation results of DAPA and OAEI participants on all datasets

[0037] As can be seen from Table 2, the anchor adaptive dynamic partitioning algorithm adopted by the present invention can effectively carry out experiments on large-scale ontology, and the final matching quality is better than that of OAEI participants, which shows that the present invention has broad application prospects in dealing with large-scale ontology matching problems.

[0038] Table 3 Comparison of the number of structurally isolated entities in the anchor-based adaptive dynamic partitioning algorithm and the existing structure-based, context-based, and matching result-oriented partitioning methods

[0039] As can be seen from Table 3, compared with the structure-based, context-based and matching result-oriented block segmentation methods in the prior art, the present invention has obvious advantages in maintaining the structure, semantics and other related information of the ontology entity. Specifically, the other three block segmentation methods failed to fully retain the important attributes and structural relationships between the ontology entities during the segmentation process, resulting in the loss of a large amount of key information, and once the block is determined, it cannot be adjusted, which seriously affects the accuracy of the subsequent ontology matching results. The present invention significantly reduces the loss of ontology information by adopting a fully dynamic block tuning strategy, and can adaptively adjust the partitioning mode according to different ontologies, and flexibly obtain a block solution that is more suitable for problem solving.

[0040] Depend on Figure 6 It can be seen that compared with the other three typical clustering and segmentation methods, the present invention performs better in scale and can divide large-scale ontology into smaller sub-ontologies, which helps to reduce the time and space complexity of large-scale ontology matching problems; on the other hand, the ontology segmentation result of the present invention has a smaller standard deviation and a more stable segmentation result, which can meet the ontology segmentation needs under different scales and structures.

[0041] The above description is only a preferred embodiment of the present invention. Any equivalent replacement made according to the contents of this specification and the drawings, or any direct or indirect application in the relevant technical field, shall fall within the scope of patent protection of the present invention.

Claims

1. A large-scale ontology processing method based on anchor point adaptive dynamic partitioning, characterized in that: The specific steps are as follows: S1: Block parameter integration based on tree coding, which integrates the core parameters of the block through a tree structure. The construction layer nodes involve five types of block construction parameters and decision probability parameters that determine the values ​​of these parameters. The two directly control the diffusion strategy of anchor points in the block construction process. The block construction parameters include branch constraints, anchor point constraints, structure constraints, scale constraints and depth constraints; the logic layer nodes use two types of logical operators, namely AND operations and OR operations, which are used to merge the block results of each branch. The tree coding can flexibly represent various block rules; S2: Dynamic optimization of blocks based on evolutionary algorithms, directly operates on tree-structured individuals, uses evolutionary algorithm crossover and mutation to automatically iterate out advanced combination tree forms of block features, and fully dynamically optimizes blocks; S3: Decision probability learning based on elite knowledge, using the knowledge contained in the elite solution to adjust the decision probability of the building layer nodes in the tree, guide the generation of block building parameters, and improve the decision quality; S4: Based on the self-adjustment of the step size of the iterative process, the current iteration information is recorded and the adjustment step size of the decision probability is adaptively set.

2. According to claim 1, a large-scale ontology processing method based on anchor point adaptive dynamic partitioning is characterized in that: The specific contents of the block parameter integration based on tree coding in step S1 include: The core parameters of building blocks are integrated into the syntax tree. Each syntax tree corresponds to a given block strategy. It adopts a multi-level tree encoding structure and a bottom-up tree decoding method. The tree structure contains three types of nodes: leaf nodes, construction layer nodes, and logic layer nodes: (1) Leaf nodes are used to input anchor point candidate sets; (2) The construction layer nodes are used to store the operators required in the block process and the corresponding decision probabilities in the node generation process. The specific contents are as follows: The operator branch constraint indicates the maximum number of branches that the syntax tree is allowed to generate to integrate the block parameters; The operator anchor constraint represents the anchor selected from the anchor candidate set for expanding the block during the evolution process; The operator structure constraint means that when each anchor point spreads to form its own block, it looks for a structural form of highly cohesive and low-coupling entities belonging to the block. There are three common structures: is-a, part-of, and equivalent. The operator size constraint indicates the maximum block size allowed, that is, the maximum number of entities that the current block can accommodate; The operator depth constraint represents the shortest path for each diffusion anchor point to reach a certain entity c along the current structure, expressed as: ; in, anc represents the diffusion anchor point, c represents the entity to be diffused, fa Indicates the common ancestor of the diffusion anchor point and the entity to be diffused; Indicates the structural depth of the closest common ancestor between the diffusion anchor point and the entity to be diffused; Indicates the common ancestor of the diffusion anchor point and the entity to be diffused. anc The shortest path; Indicates the common ancestor of the diffusion anchor point and the entity to be diffused. c The shortest path; In addition to the above five types of operators, each construction layer node also contains an additional parameter - decision probability, which represents the probability of each value being selected for multiple value selections of this type of operator in a specific type of construction layer node. Multiple decision probability fragments are integrated through a probability vector, in which each element corresponds to the probability value of a specific value selection, and the number of decision probability fragments is equal to the number of values ​​supported by this type of operator; (3) The logic layer nodes are used to store the operators used in the secondary optimization block process: The AND operator represents the intersection of the diffusion blocks of each branch, that is, extracting the same anchor concept and block structure between each branch; The or operator represents the union of the diffusion blocks of each branch, that is, merging the different anchor point concepts and block structures between branches.

3. The large-scale ontology processing method based on anchor point adaptive dynamic partitioning according to claim 2 is characterized in that: The process of obtaining the anchor candidate set of leaf nodes in the block parameter integration based on tree coding (1) in step S1 is as follows: Source Get the context concept of a certain structural relationship (is-a, part-of or equivalent) for the source ontology All entities in the dataset are sorted in descending order according to their importance, and the top 50% of entities are selected. These entities are limited in size and rich in context information. Entities selected from different structures are aggregated to generate an anchor source entity training set. The same method is used for the target ontology. Generate the anchor target entity training set, which together with the anchor source entity training set constitutes the anchor training set: Similarity analysis: Analyze the similarity between the source entity and the target entity in the anchor training set using string-based similarity measurement technology to generate a similarity matrix; Similarity screening: For the obtained similarity matrix, extract the similarity values ​​that are both the maximum values ​​of the rows and columns in the matrix, and filter them in combination with the preset similarity threshold. If the similarity of a pair of source entity-target entity exceeds the threshold, it is considered to be a correct matching pair, and the matching pair is added to the anchor candidate set. The above extraction and threshold filtering operations are repeated to generate an anchor candidate set, which contains one-to-one corresponding anchor source entities and anchor target entities.

4. The large-scale ontology processing method based on anchor point adaptive dynamic partitioning according to claim 2 is characterized in that: In the block parameter integration based on tree coding in step S1, the values ​​of different types of operators in the (2) construction layer are based on: The branch constraint sets the value range according to the total number of constraint conditions and the maximum number of permutations and combinations allowed by the grammatical structure; The anchor point constraint sets the value range according to the size of the anchor point candidate set; The structural constraint is one of is-a, part-of, or equivalent; The depth constraint sets the value range based on the maximum in-and-out degree of all structures; The scale constraint sets the value range based on twice the average block size corresponding to the current number of anchor points.

5. The large-scale ontology processing method based on anchor point adaptive dynamic partitioning according to claim 4 is characterized in that: The basis for taking values ​​of different types of operators in the construction layer in step S1 also includes that when the value supported by the operator is a continuous variable, the continuous variable needs to be divided into multiple sub-intervals, and a probability vector is constructed based on the range of each sub-interval, and then the target interval is determined according to the decision probability, and finally the specific value of the operator is randomly generated within the selected interval.

6. A large-scale ontology processing method based on anchor point adaptive dynamic partitioning according to any one of claims 1 or 5, characterized in that: The specific process of the block dynamic optimization based on the evolutionary algorithm in step S2 is as follows: (1) Configure the control parameters of the evolutionary algorithm, such as population size, number of iterations, crossover probability, and mutation probability; (2) Randomly initialize individuals in the population - syntax trees. Individuals divide large-scale source and target ontologies according to syntax rules, generate corresponding block results, and initialize the decision probabilities of construction layer nodes with equal probability; (3) Execute evolution, evaluate the fitness value of individual block results, generate the next generation of individuals through syntax-based single-point crossover, subtree crossover, single-point mutation, and subtree mutation, and automatically build a more advanced syntax tree structure; (4) According to the elite solution, the decision probability of the nodes in the construction layer is updated, and half of the individuals are selected from the population after crossover and mutation using the roulette method, which is dominated by the evolutionary strategy; the other half of the individuals are generated based on the updated decision probability, which is dominated by the decision probability. The two together constitute the next generation population; (5) Return to step (3) until the termination condition is met, output the elite solution, and obtain the optimal partitioning result of the source ontology and the target ontology.

7. The large-scale ontology processing method based on anchor point adaptive dynamic partitioning according to claim 1 is characterized in that: The specific contents of the decision probability learning based on elite knowledge in step S3 include: At the end of each generation, the operator value of the construction layer node in the elite tree structure is obtained and compared with the stored probability vector. If the operator value is the same, the corresponding decision probability is increased by one learning step. , the update formula is: ; For decision probabilities with different operator values, the mean of the learning step length is reduced in turn, and the update formula is: , num It represents the total number of categories of the current operator value. All probability vectors are updated, guiding the generation direction of the next generation of individuals and strengthening elite knowledge.

8. The large-scale ontology processing method based on anchor point adaptive dynamic partitioning according to claim 1 is characterized in that: The specific contents of step S4 based on the self-adjustment of the step length of the iterative process include: Calculate learning step length based on iteration information , adaptively adjust the change range of decision probability, the specific formula is: ； Indicates the current elite solution f Metrics, Indicates the last elite solution f Metrics; is a constant used to limit the initial value of the step length; T Indicates the current iteration number of the population; is a scaling factor that limits the rate at which the step size changes.

Images