Multi-data standard definition conflict resolution method based on large model and knowledge graph

Abstract

The invention relates to a multi-data standard definition conflict resolution method based on a large model and a knowledge graph, and belongs to the field of public data governance and data standardization. The method comprises the following steps: extracting and preprocessing a multi-data standard definition; defining semantic representation and large model analysis: performing semantic analysis on the definition of each data item by using a large model, and generating a semantic embedding vector and a key feature; knowledge graph construction and entity alignment: constructing a knowledge graph containing all data items, and connecting data item nodes with similar semantics in different standards by adopting an entity alignment algorithm to form a candidate alignment relationship; conflict detection and type identification: aiming at the aligned data item definition, comparing attributes and values, identifying definition conflict points, and classifying and marking conflict types; and carrying out conflict resolution and unified definition generation by using a globally optimized conflict resolution algorithm. According to the method, the problems of data islands and semantic conflicts caused by inconsistent data standard definitions in the prior art are solved.

Description

Technical Field

[0001] The present invention relates to a multi-data standard definition conflict resolution method based on a large model and a knowledge graph, and belongs to the technical field of public data governance and data standardization. Background Art

[0002] In public data governance, various departments and information systems often develop their own data standards and lack unified data definitions and rules, resulting in horizontal data fragmentation and vertical standard non-uniformity. Different departments have inconsistent descriptions of the same data and different data formats. These inconsistencies in data standards have caused data semantic conflicts across departments and systems, hindering the sharing, exchange, and integrated utilization of data. In order to solve this problem, the industry generally hopes to avoid semantic conflicts by establishing a complete, unified, and standard data standard system with unified data definitions. However, in practice, coordinating multiple data standards and eliminating conflicts among them usually requires a lot of manual participation. Traditional methods mainly rely on manually formulating mapping rules or writing conversion scripts to align data definitions of different systems, which is not only labor-intensive, but also prone to missing conflicts and difficult to maintain in a timely manner. There have been some relevant technical solutions and research around the problem of inconsistency in the definition of multiple data standards, but there are still many shortcomings:

[0003] Manual mapping and static rule-based approaches: Traditional approaches often rely on domain experts to manually develop standard mapping tables or write fixed conversion rules, manually mapping corresponding items across different standards. This approach was used in early standardization practices. For example, research has proposed technical strategies for resolving conflicts in data encoding standards by manually pre-defining the corresponding relationships between standard codes. Similarly, in the field of electronic data interchange (EDI), systematic solutions have emerged for data conversion by centrally maintaining mappings between standards. The advantages of this approach are its simplicity, the ability to develop targeted rules for known conflicts, and low initial investment. However, its disadvantages are also significant: First, manual mapping is time-consuming, error-prone, and difficult to exhaustively account for all conflict scenarios. Second, once established, rules lack flexibility, requiring frequent updates to accommodate new standards or changing definitions. Third, semantic differences (such as subtle differences in the connotation and extension of concepts) are often difficult to detect and define in a timely manner. Therefore, static mapping rule-based approaches may work adequately when the number of standards is small and changes are minimal. However, they are insufficient in the context of the continuous evolution of multi-source, heterogeneous data standards, and a more intelligent solution is urgently needed.

[0004] Knowledge graph-based conflict detection methods: Recently, some technologies have begun to utilize knowledge graphs to address conflicting standard definitions. For example, patent CN113449118A from South China University of Technology proposes first structuring the unstructured content of standard documents into a standard knowledge graph. Then, through techniques such as text similarity matching and multi-entity list matching, conflicting entities are identified across multiple standard knowledge graphs, thereby enabling conflict detection between standard documents. The advantage of this approach lies in the use of a knowledge graph to unify the knowledge representation of different standards, enabling a certain degree of automatic comparison of content across different standards and identifying obvious inconsistencies. Compared to purely manual methods, knowledge graph-based methods improve the coverage and efficiency of conflict detection. Furthermore, in areas such as smart homes, researchers have also studied implicit conflict detection methods based on knowledge graphs to identify conflicts in device linkage rules, further demonstrating the potential of knowledge graphs for complex conflict identification. However, these methods currently focus primarily on conflict detection and rely on methods such as similarity matching, which can miss cases where the definitions differ in wording but are inherently conflicting. Furthermore, and more importantly, there remains a lack of a clear solution for automatically resolving detected conflicts. Existing patent CN113449118A focuses on conflict detection, but detected conflicts often require manual analysis and resolution, without providing a systematic automated resolution mechanism. Therefore, while existing knowledge graph-based methods improve conflict detection capabilities, they are not sufficient to completely resolve conflicts and require integration with stronger semantic understanding and decision-making mechanisms.

[0005] Standard alignment methods combining knowledge graphs with pre-trained models: In specific fields, attempts have been made to combine knowledge graphs and pre-trained models for standardization. For example, Chinese patent CN113191156A proposes a medical examination item standardization system and method based on a medical knowledge graph and a pre-trained model. This method leverages a medical knowledge graph and a pre-trained language model to map examination item names used by different hospitals or institutions to standardized names, thereby improving the accuracy of term alignment. This method's advantage lies in the use of pre-trained models (such as BERT) to understand the semantics of medical terminology. Compared to pure rule-based or dictionary-based approaches, it offers better intelligent matching results and addresses the issue of standardizing different terms for the same medical examination item. However, such solutions primarily focus on standardizing terminology, falling within the domain of name mapping and failing to address complex definitional conflicts. Furthermore, they primarily use pre-trained models as auxiliary matching tools and lack the ability to proactively resolve conflicts. In the broader context of cross-domain data standards conflicts, terminology matching alone is insufficient: different standards may define different calculation formulas, units of measurement, or compliance thresholds for the same concept. These conflicts go beyond simple name inconsistencies and are not addressed by existing approaches like medical standardization. This invention addresses these complex conflicts by providing a comprehensive solution: focusing not only on semantic matching but also on resolving logical conflicts.

[0006] In summary, existing technologies either rely on manual / hard-coded approaches, resulting in low intelligence and high maintenance costs; or they can only detect conflicts or address local name standardization, failing to automatically and comprehensively resolve definition conflicts arising from the coexistence of multiple data standards. These shortcomings hinder the integration of diverse data standards, create risks in data sharing and exchange, and severely impact system interaction and data quality. Therefore, it is necessary to propose new inventive solutions that combine the latest AI big model technology and knowledge graph reasoning capabilities to specifically address these shortcomings and achieve the automated resolution and unified integration of definition conflicts across multiple data sources. Summary of the Invention

[0007] The technical problem solved by the present invention is: the present invention provides a multi-data standard definition conflict resolution method based on a big model and knowledge graph, which is used to solve the data island and semantic conflict problems caused by inconsistent data standard definitions in the existing technology; the present invention introduces big model semantic recognition technology and knowledge graph construction means to comprehensively understand the data definitions in different standards and accurately detect the conflicts therein; combines conflict network modeling and intelligent optimization algorithms to realize automatic conflict resolution; and cooperates with standard updates and continuous feedback supervision mechanisms to continuously improve the data standard system, thereby improving the consistency, accuracy and efficiency of data fusion and sharing in public data governance.

[0008] The technical solution of the present invention is: a multi-data standard definition conflict resolution method based on a large model and a knowledge graph, the method comprising:

[0009] S1. Extraction and preprocessing of multiple data standard definitions: Extract data item definitions from various data standard documents, clean them, and normalize their representations;

[0010] S2. Define semantic representation and large model analysis: Use the large model to perform semantic analysis on the definition of each data item and generate semantic embedding vectors and key features;

[0011] S3. Knowledge graph construction and entity alignment: Build a knowledge graph containing all data items and use an entity alignment algorithm to connect semantically similar data item nodes in different standards to form candidate alignment relationships.

[0012] S4, conflict detection and type identification: Define and compare attributes and values ​​of aligned data items, identify definition conflict points and classify and mark conflict types;

[0013] S5. Globally optimized conflict resolution algorithm for conflict resolution and unified definition generation: Design a globally optimized conflict resolution algorithm to resolve conflicts; dynamically adjust matching relationships or attribute values ​​to generate a unified solution, and output the fused data standard definition.

[0014] Furthermore, the S1 includes:

[0015] Process multiple data standard documents and perform unified parsing, which includes several data items and their definitions; Process multiple data standard documents and perform unified parsing, including:

[0016] Use text parsing and format conversion technologies to convert standard documents from various sources into structured data representations; write customized parsing scripts to extract "data item names," "definition text," and related attribute fields; if document formats are inconsistent or definitions are buried in natural language descriptions, call the big model API to assist in parsing complex text, including having the big model identify key fields in definition sentences;

[0017] S1.1. First, input each data standard document and load it into the parsing tool;

[0018] For structured formats, read fields directly;

[0019] For unstructured text, use regular rules or the NLP capabilities of large models to locate the "data item name: definition..." pattern and process and extract the name and definition content;

[0020] If the definition contains attribute information, the attribute name and attribute value are also parsed;

[0021] S1.2. Then, data cleaning and normalization are performed: unify the terms used in different sources and fill in the missing information; in this process, the large model is called to simply rewrite or fill in the definition text; and a standardized data item definition list or initial knowledge base is output: containing structured records for each data item.

[0022] Furthermore, the S2 includes:

[0023] Use large-scale LLM or pre-trained models to perform natural language processing on each data item definition;

[0024] On the one hand, the data item definition is converted into a high-dimensional semantic vector representation through the embedding interface of the large model;

[0025] On the other hand, the parsing capabilities of the large model are used to extract key semantic features in the data item definition, including:

[0026] Use the pre-trained sentence vector model to generate the embedding vector of the defined text or directly call the large model API to request the output of the defined structured information, which includes semantic summary and synonym list;

[0027] The specific process includes:

[0028] First, the "name" and "definition description" are spliced ​​into a complete description sentence as the model processing text;

[0029] The processing then proceeds in two parts:

[0030] (1) Semantic embedding vector generation: Using the vectorization function provided by the large model, each data item definition is fed into the large model to obtain a semantic embedding representation;

[0031] (2) Key feature extraction: Leveraging the question-answering or parsing capabilities of large models to extract key attributes and semantic labels from data item definitions;

[0032] The specific approach is to construct prompts to query the big model; the big model outputs a structured result based on the trained knowledge; the big model is used to supplement the attributes implicit in the plain text definition; and the semantic labels and normalized attribute values ​​defined for each data item are obtained.

[0033] Furthermore, the S3 includes:

[0034] Construct a knowledge graph to represent multi-source data standards, and use entity alignment algorithms to find identical or equivalent data items in the graph; the knowledge graph uses nodes to represent data items and edges to represent potential equivalent relationships between data items of different standards;

[0035] During the alignment process, multi-dimensional similarities are comprehensively considered. Specifically, a fusion similarity calculation method is used, a custom similarity formula is set, and a threshold is set to determine candidate matches. At the same time, a combination of heuristics and algorithms is adopted: preliminary corresponding candidates are first established by name or keyword, and then the matching score is accurately calculated through cosine similarity. For complex matching problems, graph algorithms are introduced to find the global optimal matching solution; and candidate alignment relationships are generated.

[0036] Furthermore, the specific steps of S3 include:

[0037] S3.1. The input is the node representation of each data item; first, build the knowledge graph:

[0038] Create a subgraph for each data standard and put all data items into the graph as nodes; each node has attributes and also carries semantic vectors. ;

[0039] S3.2. Next, calculate the node similarity across standards; for any two data item nodes from different standards and , define comprehensive semantic similarity , integrating multiple factors: name term similarity, definition text semantic similarity, attribute value similarity; comprehensive semantic similarity The calculation formula is as follows:

[0040]

[0041] in, is the name similarity, It defines the semantic similarity of texts. is the attribute similarity, is the weighting coefficient, satisfying , used to balance the importance of each part, these weights are adjusted based on experience or training sets;

[0042] S3.3. Calculate all cross-standard node pairs of After that, determine the candidate alignment relationship: The value is higher than the set threshold , mark the two data items as a “candidate matching pair”;

[0043] In the knowledge graph, a temporary "equivalence" edge is added between each pair of candidate matching data item nodes, and the edge weight is recorded. value;

[0044] For multiple criteria, the process of building a knowledge graph is reduced to finding similar groups in the multi-partite graph;

[0045] Output a preliminary fused knowledge graph: nodes are data items of each standard, and highly similar candidates across standards are connected by edges; each node is connected to multiple nodes from other standards.

[0046] Furthermore, in S3.2, the specific calculation details of the comprehensive semantic similarity are as follows:

[0047] (1) Name similarity : Compare the similarity of data item names or terms; if the two names are exactly the same, set If the names are different but have an intersection after word segmentation or synonym normalization, the Jaccard coefficient is used to calculate the overlap of the word sets. The calculation formula is as follows:

[0048]

[0049] , Refers to the data item node and The corresponding data item name, Tokens is the "word set" after the text is segmented. is to take the absolute value;

[0050] At the same time, using the synonym list extracted by S2, synonymous terms are considered as matches;

[0051] (2) Text semantic similarity :Use the semantic embedding vector generated by the large model to calculate the semantic similarity of the data item definition description; use the cosine similarity formula to embed the semantic embedding vectors of two data items Perform the calculation:

[0052]

[0053] in," " represents the vector dot product, is the Euclidean norm of the semantic embedding vector, The value range is [0,1];

[0054] (3) Attribute similarity :Compare the structured attribute values ​​of two data items, including data type, length limit, unit, value range, and calculate the matching degree of these attributes; For all considered attribute sets, the attribute similarity is defined as the weighted average of the similarities of each attribute:

[0055]

[0056] in, Represents a data item In the properties The value on is an attribute The importance weight of ,|P| This is the number of attributes in the attribute set P, assuming that each attribute has equal weight or highlights the main key attribute; is the similarity function of attribute values;

[0057] By comparing attributes one by one and taking weighted sum, it quantifies how close the two data items are to each other in terms of structural constraints.

[0058] Furthermore, the S4 includes:

[0059] Perform consistency checks and conflict pattern identification on candidate matching relationships in the knowledge graph:

[0060] A rule-based algorithm combined with a large model-assisted analysis method is used to determine whether there is a true semantic conflict or a difference in expression;

[0061] The method of using rule-based algorithms combined with large models to assist in analysis includes: using predefined conflict determination rules to quickly screen conflicts, while calling large models to perform comparison analysis on matching definition texts;

[0062] By combining the two, the location and type of conflicts can be accurately identified; common conflict types include: hit ambiguity conflicts, attribute inconsistency conflicts, and definition semantic conflicts.

[0063] Furthermore, the specific step S4 includes:

[0064] S4.1. Input is the knowledge graph from step S3. First, perform one-to-many matching conflict detection: scan candidate equivalent edges for each node to check whether a node is connected to multiple other nodes with similar similarity scores.

[0065] S4.2. Next, for each pair of candidate matches ,Perform attribute consistency check: compare the attribute values ​​of the two nodes and find out the actual conflicts;

[0066] Define a conflict indicator function: If there is an attribute Make And the difference cannot be ignored, then the attribute is determined Conflict; formally expressed as:

[0067]

[0068] For each detected attribute conflict, the conflict degree is also evaluated;

[0069] Then, define semantic conflict checks: use the large model to deeply compare definition descriptions;

[0070] The S4 can organize the conflict detection result list; each conflict record includes: the data item node involved, the conflict type, the specific performance and the conflict severity index.

[0071] Furthermore, the S5 includes:

[0072] Design a globally optimized conflict resolution algorithm, combined with intelligent optimization methods, to search for the optimal alignment and fusion solution. The globally optimized conflict resolution algorithm uses the candidate alignment relationship obtained by S3 as the initial solution, and through random perturbation and step-by-step optimization, it searches for the best solution in the solution space to maximize the overall similarity and minimize the conflict.

[0073] At the same time, a dynamic weight adjustment mechanism is introduced to adjust the similarity weight coefficient in real time according to the conflict situation during the optimization iteration process. , gradually strengthen the penalty for conflicting terms and guide the algorithm to get rid of the local optimal solution;

[0074] In addition, specific strategies are combined in the conflict resolution process: value fusion is adopted for attribute conflicts, selection or splitting is made in matching for one-to-many conflicts, a large model is called to generate a unified definition text, multi-standard differences are balanced, and high-quality fusion results are output.

[0075] Furthermore, the specific step S5 includes:

[0076] The input is the conflict list and the current matching graph state; the globally optimized conflict resolution algorithm is first initialized:

[0077] Establish the objective function of the optimization problem; set the objective function Maximize the "overall matching quality" minus the "conflict cost" ; Objective function Formally:

[0078]

[0079] in, Represents a set of selected matching pairs, the first item is the total similarity score of all selected matches, and the second Is the attribute conflict penalty; the third is the number of violations of the one-to-one principle in the matching set;

[0080] Then it enters the iterative optimization phase, using the simulated annealing strategy to traverse the solution space. The specific operations include:

[0081] (1) Define neighborhood operation: randomly select a pair of matching sets The modification is considered as a new solution candidate;

[0082] (2) Design neighborhood operations for attribute conflicts, which generate different new solutions ;

[0083] (2) Evaluation and acceptance criteria: Calculation of new solutions The objective function ;Compare ;

[0084] if , then unconditionally accept As the current solution;

[0085] if , then according to the Metropolis criterion of simulated annealing, the inferior solution is accepted with a certain probability to escape the local optimum;

[0086] The specific acceptance probability depends on the current "temperature" parameter ;

[0087] (3) Dynamic weight adjustment: During the iteration process, the similarity weight coefficient or the objective function penalty coefficient is regularly adjusted according to the remaining conflicts;

[0088] The dynamic adjustment function is formulated based on the current conflict rate: Represents the attributes in the current matching set The conflict ratio, then update the attribute Importance weight , the updated weight is:

[0089]

[0090] in is the adjustment coefficient; It refers to the importance weight of attribute 𝑝 in the current iteration, which represents the current evaluation of the importance of the attribute when it is first assigned to this iteration; It means that in the next iteration, the importance weight of attribute p will be updated to this value;

[0091] (4) Iterative convergence and optimal solution generation: As the temperature decreases and the weights are adjusted, the algorithm will gradually converge to a stable matching solution. ; Output the final matching relationship set ;

[0092] Based on this final alignment result, a unified knowledge graph is constructed: matching nodes are merged into a single concept node or a mapping table is established to indicate their equivalence. For each set of equivalent data items, the final standard definition value is determined. If there are differences in attributes, they are merged according to a specific strategy. Common strategies include credibility priority, retaining the union, and majority rule.

[0093] If a unified definition description text needs to be generated during the merging process, the large model is called again, the multi-source data item definitions of the same group are input, and the model is asked to write a comprehensive version;

[0094] If the definition generated by the large model is not ideal, the prompts are constrained according to the attribute fusion results; ultimately, a standardized definition of each unified concept is obtained.

[0095] This invention provides a method for resolving conflicts in the definition of multiple data standards based on a large model (a pre-trained artificial intelligence model) and a knowledge graph. This method aims to address the technical issues of inconsistent definitions and difficulty in automatically identifying and resolving conflicts when multiple data standards coexist. Specific technical points protected include:

[0096] For standard documents or data dictionaries from different sources, this paper first uses natural language processing technology to parse unstructured descriptions such as term definitions, extract key elements and relationships, and construct a corresponding knowledge graph representation. This step converts the definitions of each data standard into a structured semantic network that can be understood by machines, laying the foundation for subsequent conflict detection.

[0097] Large-scale pre-trained models (such as domain-based pre-trained language models) are introduced to deeply understand and vectorize the semantics of each standard definition. Leveraging the powerful language understanding capabilities of these large models, semantic alignment is performed across different representations of the same concept, determining whether they refer to the same or similar entities. Compared to traditional keyword matching, these large models can identify implicit synonymy relationships and semantic conflicts, providing a basis for knowledge graph node alignment and conflict resolution.

[0098] The method of the present invention combines the results of knowledge graph and large model analysis to automatically detect conflicting entities and conflicting relationships between multiple data standard knowledge graphs. By comparing the definition differences of attributes, constraints, value ranges, etc. of the same entity in different standards, incompatible or contradictory parts are identified. For example, inconsistent units, different classification granularity, overlapping or contradictory definition ranges, etc. will all be detected. Unlike the existing technology that is only based on simple text similarity or preset rule detection, the present invention combines the structured association of knowledge graphs with the semantic reasoning of large models to discover more complex and hidden conflicts.

[0099] After detecting a definition conflict, the present invention proposes an automated resolution strategy. With the help of the reasoning ability of the large model, a solution to the conflict is generated, such as suggesting the unified adoption of a certain standard definition, or introducing a higher-level concept in the knowledge graph to associate concepts of different standards to form a compatible new definition. The device of the present invention includes a conflict resolution module, which can fuse conflicting multi-source definitions based on the suggestions generated by the pre-trained model and the preset resolution rules; for example, the units of numerical caliber differences can be automatically converted and the thresholds can be adjusted, and the fine-grained concepts of concept level differences can be merged into upper-level concepts through the knowledge graph relationship. After the resolution is completed, the knowledge graph is updated accordingly to save the consistent standard definition after fusion.

[0100] The beneficial effects of the present invention are:

[0101] 1. This invention significantly improves the accuracy and recall of cross-standard semantic conflict identification through a large model's deep understanding of semantics, and can discover hidden conflicts that are difficult to detect with manual rules;

[0102] 2. This invention uses knowledge graphs to uniformly model data standards and conduct conflict network analysis, making the conflict location and resolution process more intuitive and comprehensive. Compared with traditional decentralized processing, it is easier to find globally consistent solutions.

[0103] 3. This invention introduces a simulated annealing algorithm for global optimization search, avoiding the suboptimal alignment that may result from greedy strategies, minimizing residual inconsistencies between multi-standard definitions, and ensuring more accurate and reliable fused data. The simulated annealing algorithm achieves a high degree of automation in handling data standard conflicts, significantly reducing the workload of manual comparison and modification, and shortening the cycle of cross-departmental data integration. When public data sources follow consistent standards, data shareability and fusion are significantly improved, and data quality issues (such as inconsistent definitions and duplicate storage) are significantly reduced.

[0104] 4. Through a continuous feedback learning mechanism, the present invention enables the system to adapt to ever-changing development needs, helping to maintain the robustness and advancement of the data standards system over the long term. This invention will help significantly improve the efficiency and quality of data integration and sharing in public data governance.

[0105] 5. The conflict resolution method provided by the present invention can serve as an important component of the application of standard knowledge graphs, assisting in the automatic integration and unification of cross-standard knowledge and improving the quality and efficiency of standard digitization. BRIEF DESCRIPTION OF THE DRAWINGS

[0106] Figure 1 is a flow chart of the present invention;

[0107] Figure 2 Schematic diagram of the technical structure of the multi-data standard definition conflict resolution method based on a large model and knowledge graph in the present invention. DETAILED DESCRIPTION

[0108] Example 1: Figure 1-Figure 2 As shown, a multi-data standard definition conflict resolution method based on a large model and a knowledge graph includes:

[0109] S1. Extraction and preprocessing of multiple data standard definitions: Extract data item definitions from various data standard documents, clean them, and normalize their representation to lay a good foundation for subsequent processing;

[0110] Furthermore, the S1 includes:

[0111] Process multiple data standard documents (e.g., national, provincial, industry, and internal enterprise standards) and perform unified parsing. These documents contain multiple data items and their definitions. These documents may have different formats (e.g., Excel data dictionary, Word manual, database metadata export, etc.) and require unified parsing. Each data item typically has attributes such as name, definition description, data type, value range, and unit.

[0112] Process multiple data standard documents and perform unified parsing, including:

[0113] Use text parsing and format conversion technologies to convert standard documents from various sources into structured data representations. Write customized parsing scripts (for example, using Python's Pandas to read Excel and regular expressions to extract Word paragraphs) to extract "data item names," "definition text," and related attribute fields. If document formats are inconsistent or definitions are buried in natural language descriptions, call the Big Model API to assist in parsing complex text, including having the Big Model identify key fields in definition sentences.

[0114] S1.1. First, input each data standard document and load it into the parsing tool;

[0115] For structured formats, read fields directly;

[0116] For unstructured text, use regular rules or the NLP capabilities of large models to locate the "data item name: definition..." pattern and process and extract the name and definition content;

[0117] If the definition contains attribute information (such as "type: integer, length: 10"), the attribute name and attribute value are also parsed;

[0118] S1.2. Next, data cleaning and normalization are performed: terminology used across different sources is standardized (for example, "string" and "text" are synonymous), and missing information is supplemented (this can be accomplished through inference or comparison with other standards). During this process, the large model is used to simply rewrite or complete the definition text to make it more standardized (for example, the expression of units of measurement is standardized). A standardized data item definition list or initial knowledge base is output: a structured record containing each data item, such as {standard source, name, definition description, data type, unit, value range, etc.}. This data will serve as input for subsequent steps. Each data item record includes a source identifier to distinguish between identical names from different standards. This preprocessing ensures that data from different standards can be compared and analyzed on the same platform.

[0119] The standardized data item definitions output from this step lay the foundation for subsequent semantic analysis and knowledge graph construction. Only through uniformly cleansed definition data can the large model accurately understand its meaning and maintain a consistent attribute structure for entities in the knowledge graph.

[0120] S2. Define semantic representation and large model analysis: Use the large model to perform semantic analysis on the definition of each data item, generate semantic embedding vectors and key features, and provide support for entity alignment;

[0121] Furthermore, the S2 includes:

[0122] Process each data item definition record from step 1, paying special attention to the "Name" and "Definition Description" fields, as well as related attributes (type, unit, etc.). These will be used for semantic analysis to determine whether data items in different standards refer to the same concept;

[0123] Use large-scale LLM or pre-trained models to perform natural language processing on each data item definition;

[0124] On the one hand, the data item definition is converted into a high-dimensional semantic vector representation through the embedding interface of the large model;

[0125] On the other hand, the parsing capabilities of the large model are used to extract key semantic features in the data item definition, including:

[0126] Use a pre-trained sentence vector model (such as BERT / DeepSeek) to generate an embedding vector for the definition text, or directly call the large model API to request output of the structured information of the definition, including semantic summaries and synonym lists. The large model can understand the meaning of professional definitions and extract implicit semantic relationships that are difficult to discover manually.

[0127] The specific process includes:

[0128] First, the "name" and "definition description" are spliced ​​into a complete description sentence as the model processing text; for example: "Data item name: X. Definition:...";

[0129] The processing then proceeds in two parts:

[0130] (1) Semantic embedding vector generation: Using the vectorization function provided by the big model, each data item definition is fed into the big model to obtain the semantic embedding representation. ; For example, call the embedding interface of models such as DeepSeek or Sentence-BERT, input the definition text, and output a dimensional vector ( is the embedding dimension, such as 768 or 1024). Vector It captures the semantic features defined by the data item. The formula is:

[0131]

[0132] in The text embedding function representing the large model maps the concatenated text of the data item name and definition into a semantic vector space. In this way, all definitions are projected into the same vector space, and semantically similar definitions have closer vector distances.

[0133] (2) Key feature extraction: Leveraging the question-answering or parsing capabilities of large models to extract key attributes and semantic labels from data item definitions;

[0134] The specific approach is to construct a prompt to query the big model, for example: "Please extract the data type, unit, value range, and definition keywords from the following data item definition: definition text definition text"; the big model outputs a structured result based on the trained knowledge, such as: {data type: "Int", unit: "meter", value range: "0-100", keywords: ["length", "distance"]}; for attributes implicit in the plain text definition (for example, the unit is not explicitly given but can be inferred from the context), the big model is used to supplement them; the semantic label (synonyms, superordinate concepts, etc.) and normalized attribute value of each data item definition are obtained.

[0135] The above two sub-processes can be used in combination: for example, first use the keywords and synonyms extracted by the large model to assist in enriching the vector representation (append the keywords to the text and then generate the vector to incorporate additional semantics), or use the nearest neighbors of the embedded vector to roughly determine which definitions are related, and then call the large model for detailed analysis of possible conflicting definitions. The output of this step is an enhanced definition representation list: each data item contains its semantic embedding As well as the extracted semantic features (normalized attributes and keywords, etc.). These results will be provided to the next step of knowledge graph construction to calculate the similarity between definitions and perform entity alignment.

[0136] This step leverages the cleaned data from step 1 to ensure the large model processes uniformly formatted and complete definition text, thereby improving parsing accuracy. The resulting semantic vectors and features are directly used in step 3 and form the core basis for calculating similarity links in the knowledge graph.

[0137] S3. Knowledge graph construction and entity alignment: Build a knowledge graph containing all data items and use an entity alignment algorithm to connect semantically similar data item nodes in different standards to form candidate alignment relationships.

[0138] The output from step S2 is processed to represent all data item definitions, including node information for each data item (name, semantic vector, attribute features, etc.). These data items come from multiple different data standards, and there may be duplicate entities referring to the same concept. These need to be represented in the knowledge graph, and "aligned" entity pairs across standards need to be identified.

[0139] Furthermore, the S3 includes:

[0140] Construct a knowledge graph to represent multi-source data standards, and use entity alignment algorithms to find identical or equivalent data items in the graph; the knowledge graph uses nodes to represent data items and edges to represent potential equivalent relationships between data items of different standards;

[0141] During the alignment process, multi-dimensional similarity is comprehensively considered. Specifically, a fusion similarity calculation method is used, a custom similarity formula is set, and a threshold is set to determine candidate matches. At the same time, a combination of heuristics and algorithms is adopted: preliminary corresponding candidates are first established by name or keyword, and then the matching score is accurately calculated through cosine similarity. For complex matching problems, graph algorithms are introduced (such as simulated annealing, which will be discussed in detail in step S5) to find the globally optimal matching solution; and candidate alignment relationships are generated.

[0142] Furthermore, the specific steps of S3 include:

[0143] S3.1. The input is the node representation of each data item; first, build the knowledge graph:

[0144] Create a subgraph for each data standard and put all data items into the graph as nodes; each node has attributes such as "data type = Int, unit = meter" (extracted from step 1 / 2), and the node also carries a semantic vector ;

[0145] S3.2. Next, calculate the node similarity across standards; for any two data item nodes from different standards and , define comprehensive semantic similarity , integrating multiple factors: name term similarity, definition text semantic similarity, attribute value similarity; comprehensive semantic similarity The calculation formula is as follows:

[0146]

[0147] in, is the name similarity, It defines the semantic similarity of texts. is the attribute similarity, is the weighting coefficient, satisfying , used to balance the importance of each part, these weights are adjusted based on experience or training sets;

[0148] S3.3. Calculate all cross-standard node pairs of After that, determine the candidate alignment relationship: The value is higher than the set threshold (Based on experience, such as 0.8) or when they rank first among their respective closest matches, the two data items are marked as a "candidate matching pair";

[0149] In the knowledge graph, a temporary "equivalence" edge is added between each pair of candidate matching data item nodes, and the edge weight is recorded. value;

[0150] For multiple criteria, the process of building a knowledge graph is reduced to finding similar groups in a multi-partite graph. For example, for three criteria A, B, and C, we can first calculate the similarity edges A–B, A–C, and B–C in pairs, and add all high-similarity links to the graph.

[0151] The output is a preliminary converged knowledge graph: nodes represent data items from each standard, with highly similar candidates across standards connected by edges. Each node is connected to multiple nodes from other standards (representing multiple candidate equivalence relationships). This graph provides the basis for conflict detection in the next step. It is important to emphasize that the alignment relationships at this stage are "candidate" and may contain one-to-many or many-to-many conflicts, which need to be resolved through subsequent optimization.

[0152] This step utilizes the semantic representation and attribute extraction results from the large model obtained in Step 2 to calculate cross-criteria similarity and construct a knowledge graph. The innovative combination of similarity calculation formulas ensures comprehensive consideration of both text and structure, resulting in more accurate and comprehensive alignment candidates. The candidate matching results from this step are passed to Step 4 for conflict analysis and further optimization and confirmation in Step 5.

[0153] Furthermore, in S3.2, the specific calculation details of the comprehensive semantic similarity are as follows:

[0154] (1) Name similarity : Compare the similarity of data item names or terms; if the two names are exactly the same, set If the names are different but have an intersection after word segmentation or synonym normalization, the Jaccard coefficient is used to calculate the overlap of the word sets. The calculation formula is as follows:

[0155]

[0156] , Refers to the data item node and The corresponding data item name, Tokens is the "word set" after the text is segmented. is to take the absolute value;

[0157] At the same time, the synonym list extracted by S2 is used to treat synonymous terms as matches; for example, if "age" and "age" or "year of birth" are recognized as synonyms by the large model, the name similarity score can be improved.

[0158] (2) Text semantic similarity :Use the semantic embedding vector generated by the large model to calculate the semantic similarity of the data item definition description; use the cosine similarity formula to embed the semantic embedding vectors of two data items Perform the calculation:

[0159]

[0160] in," " represents the vector dot product, is the Euclidean norm of the semantic embedding vector, The value range is [0, 1]; larger values ​​indicate closer semantics between two definitions. Leveraging the deep semantic understanding of the large model, this similarity not only captures explicit keywords but also reflects the degree of conceptual similarity within the definitions.

[0161] (3) Attribute similarity :Compare the structured attribute values ​​of two data items, including data type, length limit, unit, value range, and calculate the matching degree of these attributes; For all considered attribute sets (e.g. {data type, unit, length, allowed value range, ...}), attribute similarity is defined as the weighted average of the similarities of each attribute (Note: if some attributes are missing, only the common attributes are calculated):

[0162]

[0163] in, Represents a data item In the properties The value on is an attribute The importance weight of ,|P| This is the number of attributes in the attribute set P, assuming that each attribute has equal weight or highlights the main key attribute; is a similarity function of attribute values; for example, for the "data type" attribute, When the types are the same (both are integers, etc.), different types (Or set partial similarity, such as integers and long integers are considered similar to 0.8); for the "unit" attribute, if there is a convertible relationship, you can define (For example, although meters and centimeters have different characters, they can be converted and are considered equivalent units), otherwise 0; for the "value range" attribute, the similarity can be calculated by the interval overlap rate (for example, if one standard allows 0-100 and the other allows 0-120, the ratio of the interval overlap to the total range is used as the similarity). ); For example, the "code value set" can use Jaccard similarity to compare set differences.

[0164] By comparing attributes one by one and taking weighted sum, it quantifies how close the two data items are to each other in terms of structural constraints.

[0165] S4. Conflict detection and type identification: Define and compare attributes and values ​​of aligned data items, identify definition conflicts, and classify and mark conflict types to provide a basis for resolution decisions;

[0166] The preliminary aligned knowledge graph generated in step S3 includes all candidate matching pairs of data items, their similarities, and attribute information. Conflicts may arise at this point, for example, where a data item from one standard has a high degree of similarity with multiple data items from another standard (a one-to-many conflict), or where two data items match but have conflicting attributes (a definition conflict). These conflicts need to be detected and categorized.

[0167] Technical means adopted: S4 includes:

[0168] Perform consistency checks and conflict pattern identification on candidate matching relationships in the knowledge graph:

[0169] A rule-based algorithm combined with a large model-assisted analysis method is used to determine whether there is a true semantic conflict or a difference in expression;

[0170] The method of using rule-based algorithms combined with large models to assist in analysis includes: using predefined conflict determination rules to quickly screen conflicts (such as one-to-many matching and inconsistent attributes), while calling the large model to perform comparative analysis on the matching definition text;

[0171] By combining the two, the location and type of conflicts can be accurately identified; common conflict types include: hit ambiguity conflicts (one-to-many, many-to-many matching), attribute inconsistency conflicts (inconsistent hard constraints such as data types or units), definition semantic conflicts (contradictory descriptions or incompatible scopes), etc.

[0172] Furthermore, the specific step S4 includes:

[0173] S4.1. Input is the knowledge graph from step S3 (nodes, edges, and similarities). First, perform one-to-many matching conflict detection: scan each node's candidate equivalent edges to check whether a node connects to multiple other nodes with similar similarity scores. For example, if node X under criterion A matches both node Y and node Z under criterion B (both edges have high Sim values ​​and are similar), this is a one-to-many conflict. Record this conflict and label the relevant node set {X, Y, Z}. Similarly, record many-to-many conflicts (multiple nodes with high similarity loops). These conflicts indicate uncertainty in the matching relationship and require subsequent resolution to select the optimal pairing.

[0174] S4.2. Next, for each pair of candidate matches ,Perform attribute consistency check: compare the attribute values ​​of the two nodes and find out the actual conflicts;

[0175] Define a conflict indicator function: If there is an attribute Make And the difference cannot be ignored, then the attribute is determined Conflict; formally expressed as:

[0176]

[0177] For example: if one data type is "integer" and the other is "string", then (Not compatible); the units are "meter" and "centimeter", which are different but can be converted. In this case, it can be set to be compatible. (This can be resolved by unit standardization); if one specifies a range of 1-100 and the other specifies 1-80, there will be a range conflict (one allows values ​​in the interval [81,100] but the other does not allow it). Summarize all attributes and define attribute conflict identifiers:

[0178]

[0179] like , records that the matching pair has conflicts in attributes and lists the specific conflicting attribute categories. For each detected attribute conflict, the conflict severity is also evaluated; for example, for numerical ranges, the range overlap rate or difference value size can be calculated to quantify the severity of the conflict.

[0180] Then, define the semantic conflict check: Even if the attributes are consistent, there may be subtle differences in the meaning of the definition texts of the two standards. To this end, use the big model to deeply compare the definition descriptions; for example, input the two matching definition texts into the big model together, and ask "Do these two definitions describe exactly the same concept? What are the differences?" Let the big model give an analysis. If the big model answers that some semantic differences are included (such as "Definition A includes specific business rules, while B does not mention it"), it is considered that there is a semantic conflict. The big model can also help classify the conflict type: for example, if the answer is "Definition A is stricter than Definition B", it will be marked as an "inconsistent scope" conflict. In this way, description differences that are difficult to detect with manual rules are captured.

[0181] S4 compiles a list of conflict detection results; each conflict record includes the data item nodes involved (and their sources), the conflict type (one-to-many, inconsistent attributes, semantic differences, etc.), the specific manifestation (such as "inconsistent data type: integer vs. string" or "inconsistent semantic range determined by the large model"), and a conflict severity indicator. This list provides a checklist for conflict resolution. Regarding dependencies, only by first identifying the specific conflicts can we select targeted resolution strategies. For example, one-to-many conflicts require trade-offs in matching, while attribute conflicts can be resolved by adjusting values.

[0182] This step uses the alignment results generated in Step 3 as input and verifies them with the attributes and semantic information extracted in Step 2. The accuracy of conflict detection directly impacts the resolution plan: only when detection is sufficient can resolution accurately address each conflict point. The results of this step will guide the selection of appropriate conflict resolution measures in Step 5.

[0183] S5. Globally optimized conflict resolution algorithm for conflict resolution and unified definition generation: Design a globally optimized conflict resolution algorithm to resolve conflicts; dynamically adjust matching relationships or attribute values ​​to generate a unified solution, and output the fused data standard definition.

[0184] The conflicts identified in step 4 are listed, along with the complete aligned knowledge graph data. The goal is to resolve these conflicts globally and simultaneously, outputting a consistent, multi-source integrated set of data standard definitions. This involves optimizing and adjusting the candidate matching relationships, and merging or discarding any diverging definitions. Ultimately, a unified knowledge graph (or equivalent set of data standards) is obtained, free of conflicts.

[0185] Furthermore, the S5 includes:

[0186] Design a globally optimized conflict resolution algorithm, combined with intelligent optimization methods such as simulated annealing, to search for the optimal alignment and fusion solution. The globally optimized conflict resolution algorithm uses the candidate alignment relationship obtained by S3 as the initial solution. Through random perturbations and step-by-step optimization, it searches for the best solution in the solution space to maximize the overall similarity and minimize conflicts.

[0187] At the same time, a dynamic weight adjustment mechanism is introduced to adjust the similarity weight coefficient in real time according to the conflict situation during the optimization iteration process. , gradually strengthen the penalty for conflicting terms and guide the algorithm to get rid of the local optimal solution;

[0188] Furthermore, the conflict resolution process incorporates specific strategies: value fusion (such as unit conversion and range expansion) for attribute conflicts, selection or splitting during matching for one-to-many conflicts, and the use of a large model to generate unified definition text, balancing differences across multiple criteria and delivering high-quality fusion results. This entire process demonstrates superior performance compared to traditional static rules or greedy matching, automatically balancing differences across multiple criteria and delivering high-quality fusion results.

[0189] Furthermore, the specific step S5 includes:

[0190] The input is the conflict list and the current matching graph state (including all candidate matches and similarity scores); the globally optimized conflict resolution algorithm is first initialized:

[0191] Establish the objective function of the optimization problem; set the objective function Maximize the "overall matching quality" minus the "conflict cost" ; Objective function Formally:

[0192]

[0193] in, represents a set of selected matching pairs (i.e., the final set of equivalence relations selected from the candidate edges), the first item is the total similarity score of all selected matches, and the second Is the attribute conflict penalty (for each selected match, if there is an attribute conflict, a penalty will be deducted , is the weight coefficient reflecting our penalty for attribute inconsistency); the third item It is the number of violations of the one-to-one principle in the matching set, for example, when a node matches multiple nodes, multiplied by a large coefficient Severe penalties are imposed (to ensure that the final result tends to a one-to-one matching relationship). With such a goal design, ideally, The maximum value is obtained under the optimal matching scheme, when there is almost no conflict (the conflict cost term tends to zero), and the matching The sum is as large as possible (keeping as many high similarity matching pairs as possible). The algorithm uses the current candidate matching set as the initial solution. (This may contain conflicts);

[0194] Then it enters the iterative optimization phase, using the simulated annealing strategy to traverse the solution space. The specific operations include:

[0195] (1) Define neighborhood operation: randomly select a pair of matching sets Modifications are used as new solution candidates; for example: for one-to-many conflicts, one of the matching edges is randomly cancelled (a certain matching is cancelled); for unmatched node pairs, they are tried to be added as new matchings; or random swap adjustments are made between two different matching pairs (for example, node X of standard A is originally matched with node Y of standard B, and node Z of standard A is matched with node W of standard B. The neighborhood operation may be swapped to X–W, Z–Y matching) [This operation is particularly effective when there are cross-matching errors, simulating the process of human swapping matching;

[0196] (2) Design neighborhood operations for attribute conflicts. For example, if two matches conflict in units, try to "consider them compatible after converting units" as a state (equivalent to adjusting the attributes to make the conflict disappear). These operations generate different new solutions. ;

[0197] (2) Evaluation and acceptance criteria: Calculation of new solutions The objective function ;Compare ;

[0198] if (The new solution reduces conflicts and / or improves matching score), then accept it unconditionally As the current solution;

[0199] if (The new solution becomes worse in some aspects), then according to the Metropolis criterion of simulated annealing, the inferior solution is accepted with a certain probability to escape the local optimum;

[0200] The specific acceptance probability depends on the current "temperature" parameter ; For example: when hour, Initial temperature Higher, some poor solutions may be accepted to explore more combinations; gradually reduce the temperature as the iteration proceeds , when the algorithm converges, only the better solution is retained. The temperature update uses the cooling scheduling function, such as Each round of decay, or simulated annealing, involves lowering the temperature after a certain number of steps.

[0201] (3) Dynamic weight adjustment: During the iteration process, the similarity weight coefficient is adjusted regularly according to the remaining conflicts ( etc.) or objective function penalty coefficient ( ). For example, in the initial stage, in order to match semantically similar items as much as possible, (Text semantic weight) can be set larger to make the algorithm tend to connect nodes with high semantic similarity; once the main match is established, the attribute weight is gradually increased in the refinement stage. , strengthen the requirements for consistency of data types, etc., and avoid overly loose matching;

[0202] The dynamic adjustment function is formulated based on the current conflict rate: Represents the attributes in the current matching set The conflict ratio, then update the attribute Importance weight , the updated weight is:

[0203]

[0204] in is the adjustment coefficient; It refers to the importance weight of attribute 𝑝 in the current iteration, which represents the current evaluation of the importance of the attribute when it is first assigned to this iteration; It means that in the next iteration, the importance weight of attribute p will be updated to this value;

[0205] This formula shows that if a certain attribute conflicts frequently ( If the conflict is large, its weight is increased, prompting subsequent iterations to place greater emphasis on consistency of this attribute. Conversely, if the conflict has been reduced, the weight will not change much or may even decrease. Through multiple rounds of adjustments, we gradually transition from "primarily semantic matching" to "strict consistency." The final output matches are both semantically sound and fully compatible in terms of attributes.

[0206] (4) Iterative convergence and optimal solution generation: As the temperature decreases and the weights are adjusted, the algorithm will gradually converge to a stable matching solution. ;at this time When the maximum value is reached or no longer significantly improved, and all conflicting items are either eliminated (resolved by changing the match or adjusting the attributes) or reduced to an acceptable range (for example, some small differences that cannot be completely eliminated are recorded but do not affect the overall consistency), the final set of matching relationships is output. ;

[0207] Based on this final alignment, we construct a unified knowledge graph: matching nodes are merged into a single concept node (or a mapping table is established to indicate their equivalence). For each set of equivalent data items, the final standard definition value is determined. If there are attribute differences, they are merged according to a specific strategy. Common strategies include prioritizing credibility (e.g., referencing the data source credibility mentioned in Step 4: selecting the value of the authoritative standard), preserving the union (e.g., taking the union of the ranges of each standard to encompass all), and majority rule (following the majority consensus value if a minority of sources differ). For example, unit conflicts are converted to a single standard unit; for value range conflicts, a wider range may be chosen to cover all requirements (or labeled as multiple sub-constraints of the concept). To generate a unified definition description during the merging process, the master model can be invoked again: input the same set of multi-source definitions and have the model compile a comprehensive version. For example, if prompted with the following: "Please merge the following definitions into a unified definition" and provided with a list of the original definitions, the master model will generate a complete new definition that strives to cover all key points of the standard and eliminate any conflicts. If the definition generated by the model is not ideal, the prompt can be constrained based on the attribute fusion results (for example, by specifying the value range of the new definition). Ultimately, a standardized definition of each unified concept is obtained.

[0208] Through the above optimization process, the algorithm weighs all matches and conflicts globally, which is more efficient and comprehensive than manually resolving them one by one. The final output of this step is: a set of integrated data standard definitions (or a integrated knowledge graph), in which each data item concept is a consistent definition with all sources aligned and conflicts resolved. The specific output form can be an updated knowledge graph (a graph after equivalent nodes are merged, with unified attributes), and a corresponding data standard document (listing unified data item names, definition descriptions, data types, etc.). In addition, the system can generate a conflict resolution report to record which data items have been merged from which sources, and how the original differences have been handled, providing users with change traceability.

[0209] The conflict resolution scheme proposed in this step overcomes the problem of greedy matching easily falling into local optima through simulated annealing global optimization, enabling the identification of a more optimal overall solution for complex multi-source matching. Furthermore, by combining large-scale semantic understanding with knowledge graph attribute constraints, this approach achieves conflict resolution that combines "soft semantics" with "hard rules," making it more intelligent than traditional methods based solely on fixed rules (such as simple priority or majority voting). A mechanism for dynamically adjusting similarity weights allows the algorithm to focus on different priorities at different stages, ensuring that potential matches are not missed initially while ensuring rigorous consistency in later results. This adaptive process is a major innovation of this solution. Conflict resolution decisions not only consider the selected value but also incorporate attribute conversion and definition fusion (such as automatic unit conversion and text synthesis), further improving the quality and applicability of the fusion results. The resulting unified data standard definition is semantically and format-free, providing a standardized data foundation for business applications. The entire technical process is logically rigorous and sequential, and any conflicts are effectively identified and resolved through the aforementioned method, ensuring the successful integration of multiple data standards.

[0210] The improvement and innovation of the present invention lies in the organic combination of the semantic understanding of large models and the structured reasoning of knowledge graphs, which realizes the automatic processing of the entire process from detection to resolution of conflicts in multi-data standard definitions. Compared with the existing technology, this solution can: 1) automatically understand the meaning of standard definitions and identify conflicts that are difficult for humans to detect; 2) provide intelligent conflict resolution methods, rather than just prompting conflicts; 3) greatly reduce manual participation and improve the consistency of cross-standard data fusion. The current existing technology is still unable to effectively handle the complex semantic conflicts between cross-domain and cross-format standard definitions. For example, existing solutions mostly stay at the detection level or rely on manual formulation of mapping rules, which makes it difficult to cope with dynamically added standard differences. The present invention proposes a new solution to these shortcomings.

[0211] Simulation Experiment: To further verify the effectiveness of the present method, a simulation experiment was designed. In this experiment, three virtual data standards were constructed, each containing approximately 50 data elements. These standards were artificially set to have a certain degree of duplication and conflict between them.

[0212] Specifically, this invention randomly generates a batch of common data items (such as name, gender, age, address, phone number, company, and code) and their different definition versions, and assigns them to three standards. Some of these data items have different names across standards but refer to the same concept (e.g., "residence" vs. "address"), while others have the same name but different meanings (e.g., "employee number" refers to the internal employee ID in Standard 1 but to the external partner ID in Standard 2).

[0213] A total of 20 pairs of conflicting data items were injected, covering various types such as inconsistent formats, inconsistent value ranges, and confusing naming.

[0214] This paper used the present method and two comparative methods to detect and resolve conflicts on the simulated data: Comparative method 1 performed simple mapping based on string matching rules (without utilizing a semantic model); comparative method 2 used semantic similarity to detect conflicts without global optimization (resolving conflicts pairwise without considering overall constraints). Evaluation metrics included the accuracy and recall of conflict identification, as well as the effectiveness of the final conflict resolution (the proportion of remaining conflicts).

[0215] Experimental results show that the method of the present invention achieved the highest accuracy and recall rate in the conflict identification stage: among the 20 pairs of conflicting data items implanted in the simulation, the method successfully identified 18 pairs, and missed 2 pairs with extremely obscure definitions; the string matching method only identified 10 pairs, missing most conflicts with different semantic expressions; the semantic detection method without optimization identified 15 pairs, but failed to detect some conflicts that required global trade-offs. Entering the conflict resolution stage, the method correctly merged semantically equivalent definitions through global optimization and split the concepts that needed to be distinguished. Finally, the remaining conflict indicators in the knowledge graph With only one conflict (a few requiring manual decision-making remained unresolved), the conflict resolution rate reached 95%. In contrast, string matching methods, due to their lack of semantic analysis, mistakenly treated some items that should not be merged as identical, resulting in approximately 40% of conflicts remaining unresolved or introducing new ones after resolution. Semantic methods without global optimization, while avoiding significant mismatches, ultimately left approximately 20% of conflicts unresolved due to a lack of consistency in their individual decisions. Overall, the method of the present invention achieved the highest conflict resolution completeness and accuracy in a simulated data environment. Furthermore, the efficiency of the method of the present invention was evaluated. With the data set of the aforementioned scale (approximately 150 data items and 20 conflict pairs), the system's semantic analysis and knowledge graph construction were completed within seconds, and the simulated annealing optimization process converged and obtained a stable solution in just tens of seconds. This time consumption is quite acceptable, considering that data standard adjustment is typically performed in an offline batch process. If the data scale is further increased, the method of the present invention can also be accelerated through parallel computing and a suitable reduction in the number of simulated annealing iterations. Therefore, the method of the present invention has a certain scalability while ensuring the effect, and can be applied to actual large-scale public data standard conflict management tasks.

[0216] Experiments have verified the effectiveness of the present invention: it can efficiently discover complex conflicts and provide reasonable resolution solutions, significantly improving the consistency and fusion quality of multi-source data. The performance improvements achieved in the simulation experiments are consistent with other research results cited, such as a more than 50% increase in the recall rate of knowledge conflict resolution, demonstrating the advanced nature and reliability of the present invention's solution. It should be noted that the above examples and simulation results are intended to illustrate the present invention's method, and any matters not covered are subject to flexible adjustment within the scope of the present invention. For example, the choice of large model can be adjusted based on the application domain and resource conditions: general domains can use general large models such as the DeepSeek series, while specialized domains can prefer domain pre-trained models or knowledge-enhanced models. When a model is temporarily unavailable, traditional ontology dictionary methods can be used to assist in semantic recognition, although the effectiveness will be relatively reduced. The present invention uses a simulated annealing algorithm for conflict optimization, but this is not the only option. Other heuristic algorithms, such as genetic algorithms and tabu search, can also be used, or the conflict resolution model can be converted to an integer linear program and accurately solved by an optimization solver. When the data scale is small, an accurate solution can obtain the global optimal solution; when the data volume is large, heuristic algorithms are more effective in balancing efficiency and effectiveness. In the implementation of knowledge graphs, semantic web standards such as RDF / OWL can be used to represent standard data knowledge, enabling integration with existing standardization platforms and ontology libraries. Semantic similarity calculations can also incorporate prior knowledge provided by domain ontologies or lexicons (e.g., WordNet, Synonym Forest) to improve accuracy. For example, similarity thresholds in conflict detection and weight parameters in optimization models can be adjusted and optimized based on actual needs.

[0217] Example 2: Multi-data standard definition conflict resolution method based on large model and knowledge graph. Assume that there are two data standards A and B formulated by different departments, which contain the following corresponding data items:

[0218] Standard A:

[0219] Address: Defined as a complete residential address string (including all elements such as province, city, district, and street).

[0220] Population: defined as the number of permanent residents within the jurisdiction (including local registered residents and non-registered residents holding long-term residence permits).

[0221] Standard B:

[0222] Address: It is defined as the residential address information of the residents, which is divided into three fields: "Province / City / Detailed Address" for storage.

[0223] Total population: defined as the number of registered residents within the jurisdiction (including only residents with local household registration).

[0224] In the above standards A and B, the meanings of “address” and “residence” are actually the same, both indicating the place of residence of residents, but standard A uses the address as a single text field, while standard B subdivides the address into multiple field records, and there are differences in format specifications. In addition, the “population” in standard A and the “total population” in standard B both refer to the population size of a certain area, but the former’s statistical scope includes non-household registered long-term residents, while the latter only counts registered residents, and there is a conflict between the two in terms of statistical scope. These two examples reflect the common problem of inconsistent standards in actual public data governance: synonymous with different forms (different names but the same semantics, such as address / residence) and homonyms (same or similar names but different meanings, such as population / total population). For the above data standards A and B, the method of the present invention performs conflict detection and resolution in steps:

[0225] S1. Semantic Recognition. A pre-trained large model was used to analyze the definition text for Standards A and B. The model output indicated that the semantic similarity score for "address" and "residence" was high (close to 1), indicating that they represent the same concept. The semantic similarity for "population" and "total population" was also high, suggesting that the model tends to believe that the population concepts they refer to are highly related. However, through analysis of the definition details, the model noted that the definition scope of "total population" is "only registered residents," while Standard A's "population" includes non-registered permanent residents, suggesting a semantic difference between the two. After the semantic recognition stage, we obtained a semantic vector representation for each data item and preliminarily identified the associations: {address ≈ residence}, {population ≈ total population}.

[0226] S2. Knowledge Graph Construction. Based on the semantic recognition results, a cross-standard knowledge graph is constructed. The system creates nodes representing "address" in Standard A and "residence" in Standard B, and establishes a sameAs equivalence relationship between them. An abstract node, "address (concept)," is introduced to connect the "address" and "residence" nodes, indicating that they belong to the same address concept. The system also establishes a sameAs relationship for "population" and "total population," connecting them with the abstract node, "population (concept)," indicating that they are considered the same "population" concept. This creates two conflict nodes in the knowledge graph: an "address" conflict node (connecting the address in Standard A with the residence in Standard B) and a "population" conflict node (connecting the population in Standard A with the total population in Standard B). Furthermore, the graph records a containment relationship: the "total population" data in Standard B is actually contained in the population data in Standard A (registered population is a subset of the permanent population). Therefore, a "total population ⊂ population" relationship is added to the knowledge graph for reference.

[0227] S3. Conflict Detection. The system conducts in-depth analysis of each conflicting node in the knowledge graph, extracting the specific attributes of each relevant data item. For example, for the "address" conflict node, it discovered that Standard A's address is a single string field, while Standard B's address consists of multiple fields. Although both have text data types, their structural formats differ, resulting in a format conflict. For the "population" conflict node, the system compares the population definition attributes of Standards A and B: While the units of measurement are the same (number of people) and the data types are both integers, the statistical scopes defined differ. Standard A's population includes non-hukou residents, while Standard B's total population only includes registered residents. This constitutes a scope conflict. Therefore, the system marks the "population" conflict as a scope inconsistency conflict and assigns it a higher conflict weight (because semantic scope differences directly affect data statistics). Conversely, the "address" conflict, which involves only a format conversion, is assigned a lower weight. Furthermore, the system notes that while "population" and "total population" have different names, their meanings are similar. This is a terminological difference, not a naming ambiguity, so no naming conflict is flagged. After conflict detection, the system identified two conflicts: one in address format and one in demographic caliber.

[0228] S4. Conflict resolution. The system establishes an optimization model for the above conflicts and runs a simulated annealing algorithm to search for the optimal solution. For the "address" conflict, since the semantics of the two fields are completely consistent and only the representation is different, the algorithm tends to merge them by unifying the format. Specifically, the algorithm assumes no conversion in the initial solution, which will leave conflict costs; in subsequent iterations, it attempts to take actions such as "splitting and storing the addresses of standard A" or "splicing the addresses of standard B" to eliminate format conflicts. After the format conversion measures are taken, the "address" conflict is resolved. From 1 to 0, the objective function The algorithm faces two strategies for the "population" conflict: merging the two concepts (treating them as the same concept and attempting to unify their definition) or splitting them (retaining them as distinct concepts). Choosing a merge strategy requires compromises in definition, such as expanding the scope of registered population or labeling the data source; otherwise, the conflict cannot be completely resolved. Splitting, however, means no longer forcing the two indicators to align, but rather recognizing their separate statistical calibers. The simulated annealing algorithm randomly tried both strategies during a high-temperature phase and calculated the conflict costs. The merge strategy consistently retained a certain conflict cost, while the split strategy reduced the conflict cost to zero. As the temperature decreased, the algorithm gradually converged to a split approach for the "population" concept: revoking the equivalence between Standard A (population) and Standard B (total population), treating them as distinct concepts. Ultimately, the simulated annealing algorithm outputted the following conflict resolution solution: standardizing the address data format and distinguishing the statistical calibers of the population data. That is to say, the address fields of Standard A and Standard B will adopt a unified storage format or provide conversion mapping to ensure consistent format; while the two indicators "population" and "total population" will maintain their respective definitions and no longer attempt to force alignment to avoid semantic conflicts.

[0229] S5. Standard Updates and Feedback. Based on the aforementioned mitigation solution, Standards A and B were adjusted and configured accordingly. First, regarding the address field, both parties agreed to adopt Standard B's "three-segment" storage structure as the standard format, as split storage provides richer information and facilitates comparison. Accordingly, Standard A will add structure to the address field, or provide a conversion program to split a single address string into the standard format during data exchange. This allows direct mapping of address data between the two standards, eliminating format conflicts. Second, regarding population indicators, it was decided to rename Standard A's "population" to "permanent population" within the system, while Standard B's "total population" will be clearly defined as "registered population." The difference in scope between the two standards will be clearly noted in the metadata. This update eliminates confusion between the two standards' population indicators: in the knowledge graph, "permanent population" and "registered population" exist as distinct nodes, linked by the relationship "registered population ⊂ permanent population." The previously conflicting "population" node has been removed. Finally, the data governance team will provide feedback on these changes to the relevant standard development departments, which will serve as the basis for subsequent standard revisions. At the same time, the system records the decision-making process of this conflict resolution, especially the decision to split rather than merge the "population" in the simulated annealing algorithm, and stores it in the conflict rule library. If a similar "partial overlap" conflict arises in the future, the system can refer to this experience and make decisions more quickly. Through the above implementation, the method of the present invention successfully eliminated the two major data definition conflicts between standards A and B. Objective function The initial non-zero value drops to 0, proving that the scheme has achieved the goal of conflict resolution.

[0230] Example 3: The present invention also provides a multi-data standard definition conflict resolution system based on a large model and a knowledge graph, the system comprising:

[0231] Data collection and preprocessing module, used to extract data item definitions from multiple data standard documents, and perform cleaning and normalization processing;

[0232] Semantic representation and large model parsing module, which is used to perform semantic analysis on data item definitions using large models to generate embedding vectors and key features;

[0233] The knowledge graph construction and entity alignment module is used to build a knowledge graph containing all data items and use the entity alignment algorithm to connect the data item nodes with similar semantics in different standards.

[0234] The conflict detection and resolution optimization module is used to identify conflicting points in the aligned data item definitions and resolve them through optimization algorithms to generate a unified solution.

[0235] The standard update and feedback supervision module is used to feed back the optimization results into the original standard document, establish a continuous feedback and supervision mechanism, monitor the use of the standard, and promptly discover new conflicts.

[0236] Furthermore, the data acquisition and preprocessing module includes: a data acquisition module, a data cleaning module, and a data normalization module;

[0237] Data collection module, used to support the import of multiple data formats (such as JSON, XML, CSV, etc.);

[0238] Data cleaning module, used to remove redundant information, fill missing values, and unify data formats;

[0239] Data normalization module, used to convert data item definitions from different sources into a unified representation;

[0240] The semantic representation and large model parsing module includes: an embedding generation module, a feature extraction module, and a semantic enhancement module;

[0241] Embedding generation module, which is used to generate semantic embedding vectors of data items using large models (such as BERT and RoBERTa);

[0242] Feature extraction module, used to extract key features from the embedding vector for subsequent entity alignment;

[0243] Semantic enhancement module, used to enhance the semantic representation of data item definitions by combining contextual information;

[0244] The knowledge graph construction and entity alignment layer module includes: a knowledge graph construction module, an entity alignment module, and an alignment relationship generation module;

[0245] The knowledge graph construction module is used to construct a knowledge graph based on data items and their relationships.

[0246] The entity alignment module is used to encode the knowledge graph using graph neural networks (such as GCN and GAT) and identify potential alignment entities;

[0247] The alignment relationship generation module is used to generate alignment relationships and form candidate alignment data item pairs.

[0248] The conflict detection and resolution optimization module includes: a conflict detection module, a catalog function scoring module, an optimization algorithm module, and a unified definition generation module;

[0249] Conflict detection module, used to compare attributes and values, identify definition conflict points and classify and mark conflict types;

[0250] The catalog function scoring module is used to design multi-dimensional catalog functions and score conflicting nodes by comprehensively considering factors such as definition similarity, usage frequency, and context consistency;

[0251] The optimization algorithm module is used to apply the simulated annealing algorithm to search for the optimal conflict resolution solution and dynamically adjust the matching relationship or attribute value;

[0252] Unified definition generation module, used to generate the fused data standard definition and output a unified solution;

[0253] The standard update and feedback supervision module includes: a standard update module, a feedback supervision module, and a version management module;

[0254] The standard update module is used to feed the optimization results back to the original standard document and update the standard definition;

[0255] Feedback and supervision module, which is used to introduce continuous feedback and supervision mechanisms to monitor the use of standards and promptly identify new conflicts;

[0256] The version management module is used to manage different versions of the standard, record change history, and support backtracking and auditing.

[0257] The specific embodiments of the present invention are described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made within the knowledge of ordinary technicians in this field without departing from the scope of the present invention.

Claims

1. A multi-data standard definition conflict resolution method based on a large model and knowledge graph, characterized by: The method comprises: S1. Extraction and preprocessing of multiple data standard definitions: Extract data item definitions from various data standard documents, clean them, and normalize their representations; S2. Define semantic representation and large model analysis: Use the large model to perform semantic analysis on the definition of each data item and generate semantic embedding vectors and key features; S3. Knowledge graph construction and entity alignment: Build a knowledge graph containing all data items and use an entity alignment algorithm to connect semantically similar data item nodes in different standards to form candidate alignment relationships. S4, conflict detection and type identification: Define and compare attributes and values ​​of aligned data items, identify definition conflict points and classify and mark conflict types; S5. Globally optimized conflict resolution algorithm for conflict resolution and unified definition generation: Design a globally optimized conflict resolution algorithm to resolve conflicts; dynamically adjust matching relationships or attribute values ​​to generate a unified solution, and output the fused data standard definition.

2. The method for resolving conflicts in multiple data standard definitions based on a large model and knowledge graph according to claim 1, characterized in that: Said S1 comprises: Process multiple data standard documents and perform unified parsing, which includes several data items and their definitions; Process multiple data standard documents and perform unified parsing, including: Use text parsing and format conversion technologies to convert standard documents from various sources into structured data representations. Write customized parsing scripts to extract "data item names," "definition text," and related attribute fields. If document formats are inconsistent or definitions are buried in natural language descriptions, call the Big Model API to assist in parsing complex text, including having the Big Model identify key fields in definition sentences. S1.

1. First, input each data standard document and load it into the parsing tool; For structured formats, read fields directly; For unstructured text, use regular rules or the NLP capabilities of large models to locate the "data item name: definition..." pattern and process and extract the name and definition content; If the definition contains attribute information, the attribute name and attribute value are also parsed; S1.

2. Then, data cleaning and normalization are performed: unify the terms used in different sources and fill in the missing information; in this process, the large model is called to simply rewrite or fill in the definition text; and a standardized data item definition list or initial knowledge base is output: containing structured records for each data item.

3. The method for resolving conflicts in multiple data standard definitions based on a large model and knowledge graph according to claim 1, characterized in that: The S2 includes: Use large-scale LLM or pre-trained models to perform natural language processing on each data item definition; On the one hand, the data item definition is converted into a high-dimensional semantic vector representation through the embedding interface of the large model; On the other hand, the parsing capabilities of the large model are used to extract key semantic features in the data item definition, including: Use the pre-trained sentence vector model to generate the embedding vector of the defined text or directly call the large model API to request the output of the defined structured information, which includes semantic summary and synonym list; The specific process includes: First, "name" and "definition description" are concatenated into a complete description sentence as the model processing text; The processing then proceeds in two parts: (1) Semantic embedding vector generation: Using the vectorization function provided by the large model, each data item definition is fed into the large model to obtain a semantic embedding representation; (2) Key feature extraction: Leveraging the question-answering or parsing capabilities of large models to extract key attributes and semantic labels from data item definitions; The specific approach is to construct prompts to query the big model; the big model outputs a structured result based on the trained knowledge; the big model is used to supplement the attributes implicit in the plain text definition; and the semantic labels and normalized attribute values ​​defined for each data item are obtained.

4. The method for resolving conflicts in multiple data standard definitions based on a large model and knowledge graph according to claim 1, characterized in that: The S3 includes: Construct a knowledge graph to represent multi-source data standards, and use entity alignment algorithms to find identical or equivalent data items in the graph; the knowledge graph uses nodes to represent data items and edges to represent potential equivalent relationships between data items of different standards; During the alignment process, multi-dimensional similarities are comprehensively considered. Specifically, a fusion similarity calculation method is used, a custom similarity formula is set, and a threshold is set to determine candidate matches. At the same time, a combination of heuristics and algorithms is adopted: preliminary corresponding candidates are first established by name or keyword, and then the matching score is accurately calculated through cosine similarity. For complex matching problems, graph algorithms are introduced to find the global optimal matching solution; and candidate alignment relationships are generated.

5. The method for resolving conflicts in multiple data standard definitions based on a large model and knowledge graph according to claim 1, characterized in that: The specific steps of S3 include: S3.

1. The input is the node representation of each data item; first, build the knowledge graph: Create a subgraph for each data standard and put all data items into the graph as nodes; each node has attributes and also carries semantic vectors. ; S3.

2. Next, calculate the node similarity across standards; for any two data item nodes from different standards and , define comprehensive semantic similarity , integrating multiple factors: name term similarity, definition text semantic similarity, attribute value similarity; comprehensive semantic similarity The calculation formula is as follows: ； in, is the name similarity, It defines the semantic similarity of texts. is the attribute similarity, is the weighting coefficient, satisfying , used to balance the importance of each part, these weights are adjusted based on experience or training sets; S3.

3. Calculate all cross-standard node pairs of After that, determine the candidate alignment relationship: The value is higher than the set threshold , mark the two data items as a "candidate matching pair"; In the knowledge graph, add a temporary "equivalence" edge between each pair of candidate matching data item nodes, and record the edge weight value; For multiple criteria, the process of building a knowledge graph is reduced to finding similar groups in the multi-partite graph; Output a preliminary fused knowledge graph: nodes are data items of each standard, and highly similar candidates across standards are connected by edges; each node is connected to multiple nodes from other standards.

6. The method for resolving conflicts in multiple data standard definitions based on a large model and knowledge graph according to claim 5, characterized in that: In S3.2, the specific calculation details of the comprehensive semantic similarity are as follows: (1) Name similarity : Compare the similarity of data item names or terms; If the two names are exactly the same, If the names are different but have an intersection after word segmentation or synonym normalization, the Jaccard coefficient is used to calculate the overlap of the word sets. The calculation formula is as follows: ； , Refers to the data item node and The corresponding data item name, Tokens is the "word set" after the text is segmented. is to take the absolute value; At the same time, using the synonym list extracted by S2, synonymous terms are considered as matches; (2) Text semantic similarity :Use the semantic embedding vector generated by the large model to calculate the semantic similarity of the data item definition description; use the cosine similarity formula to embed the semantic embedding vectors of two data items Perform the calculation: ； in," " represents the vector dot product, is the Euclidean norm of the semantic embedding vector, The value range is [0,1]; (3) Attribute similarity :Compare the structured attribute values ​​of two data items, including data type, length limit, unit, value range, and calculate the matching degree of these attributes; For all considered attribute sets, the attribute similarity is defined as the weighted average of the similarities of each attribute: ； in, Represents a data item In the properties The value on is an attribute The importance weight of ,|P| This is the number of attributes in the attribute set P, assuming that each attribute has equal weight or highlights the main key attribute; is the similarity function of attribute values; By comparing attributes one by one and taking weighted sum, it quantifies how close the two data items are to each other in terms of structural constraints.

7. The method for resolving conflicts in multiple data standard definitions based on a large model and knowledge graph according to claim 1, characterized in that: The S4 includes: Perform consistency checks and conflict pattern identification on candidate matching relationships in the knowledge graph: A rule-based algorithm combined with a large model-assisted analysis method is used to determine whether there is a true semantic conflict or a difference in expression; The method of using rule-based algorithms combined with large models to assist in analysis includes: using predefined conflict determination rules to quickly screen conflicts, while calling large models to perform comparison analysis on matching definition texts; By combining the two, the location and type of conflicts can be accurately identified; common conflict types include: hit ambiguity conflicts, attribute inconsistency conflicts, and definition semantic conflicts.

8. The method for resolving conflicts in multiple data standard definitions based on a large model and knowledge graph according to claim 1, characterized in that: The specific steps of S4 include: S4.

1. Input is the knowledge graph from step S3. First, perform one-to-many matching conflict detection: scan candidate equivalent edges for each node to check whether a node is connected to multiple other nodes with similar similarity scores. S4.

2. Next, for each pair of candidate matches ,Perform attribute consistency check: compare the attribute values ​​of the two nodes and find out the actual conflicts; Define a conflict indicator function: If there is an attribute Make And the difference cannot be ignored, then the attribute is determined Conflict; formally expressed as: ； For each detected attribute conflict, the conflict degree is also evaluated; Then, define semantic conflict checks: use the large model to deeply compare definition descriptions; The S4 can organize the conflict detection result list; each conflict record includes: the data item node involved, the conflict type, the specific performance and the conflict severity index.

9. The method for resolving conflicts in multiple data standard definitions based on a large model and knowledge graph according to claim 1, characterized in that: The S5 includes: Design a globally optimized conflict resolution algorithm, combined with intelligent optimization methods, to search for the optimal alignment and fusion solution. The globally optimized conflict resolution algorithm uses the candidate alignment relationship obtained by S3 as the initial solution, and through random perturbation and step-by-step optimization, it searches for the best solution in the solution space to maximize the overall similarity and minimize the conflict. At the same time, a dynamic weight adjustment mechanism is introduced to adjust the similarity weight coefficient in real time according to the conflict situation during the optimization iteration process. , gradually increase the penalty for conflicting terms and guide the algorithm to get rid of the local optimal solution; In addition, specific strategies are combined in the conflict resolution process: value fusion is adopted for attribute conflicts, selection or splitting is made in matching for one-to-many conflicts, a large model is called to generate a unified definition text, multi-standard differences are balanced, and high-quality fusion results are output.

10. The method for resolving conflicts in multiple data standard definitions based on a large model and knowledge graph according to claim 1, characterized in that: The specific steps of S5 include: The input is the conflict list and the current matching graph state; the globally optimized conflict resolution algorithm is first initialized: Establish the objective function of the optimization problem; set the objective function Maximize the "overall matching quality" minus the "conflict cost" ; Objective function Formally: ； in, Represents a set of selected matching pairs, the first item is the total similarity score of all selected matches, and the second Is the attribute conflict penalty; the third is the number of violations of the one-to-one principle in the matching set; Then it enters the iterative optimization phase, using the simulated annealing strategy to traverse the solution space. The specific operations include: (1) Define neighborhood operation: randomly select a pair of matching sets The modification is considered as a new solution candidate; (2) Design neighborhood operations for attribute conflicts, which generate different new solutions ; (2) Evaluation and acceptance criteria: Calculation of new solutions The objective function ;Compare ; if , then unconditionally accept As the current solution; if , then according to the Metropolis criterion of simulated annealing, the inferior solution is accepted with a certain probability to escape the local optimum; The specific acceptance probability depends on the current "temperature" parameter ; (3) Dynamic weight adjustment: During the iteration process, the similarity weight coefficient or the objective function penalty coefficient is regularly adjusted according to the remaining conflicts; The dynamic adjustment function is formulated based on the current conflict rate: Represents the attributes in the current matching set The conflict ratio, then update the attribute Importance weight , the updated weight is: ； in is the adjustment coefficient; It refers to the importance weight of attribute 𝑝 in the current iteration, which represents the current evaluation of the importance of the attribute when it is first assigned to this iteration; It means that in the next iteration, the importance weight of attribute p will be updated to this value; (4) Iterative convergence and optimal solution generation: As the temperature decreases and the weights are adjusted, the algorithm will gradually converge to a stable matching solution. ; Output the final matching relationship set ; Based on this final alignment result, a unified knowledge graph is constructed: matching nodes are merged into a single concept node or a mapping table is established to indicate their equivalence. For each set of equivalent data items, the final standard definition value is determined. If there are differences in attributes, they are merged according to a specific strategy. Common strategies include credibility priority, retaining the union, and majority rule. If a unified definition description text needs to be generated during the merging process, the large model is called again, the multi-source data item definitions of the same group are input, and the model is asked to write a comprehensive version; If the definition generated by the large model is not ideal, the prompts are constrained according to the attribute fusion results; ultimately, a standardized definition of each unified concept is obtained.

Images