Cross-domain entity identity matching and information fusion methods and systems

Abstract

This invention relates to a method and system for cross-domain entity identity matching and information fusion, and pertains to the field of entity matching technology. It includes: performing entity matching on a standardized entity dataset using at least two of the following methods: rule-based entity matching, similarity-based entity matching, machine learning-based entity matching, and graph-based entity matching, and outputting entity matching results; filtering the entity matching results using at least two of the following methods: timestamp-based method, confidence-weighted method, voting method, and model-assisted judgment method, and outputting unique attribute values; completing the entity matching results using at least two of the following methods: rule-based completion method, similarity-based completion method, and large model-based completion method, and outputting completed entity information; then fusing and outputting initial unified entity information; verifying the initial unified entity information, and outputting final entity information. This invention can improve the accuracy of cross-domain entity recognition.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of entity matching and information fusion, and particularly relates to a cross-domain entity identity matching and information fusion method and system. BACKGROUND

[0002] Under the background of continuous digital transformation, multi-source heterogeneous data is experiencing explosive growth, and the phenomenon of "data silos" between different fields and different systems is becoming more and more prominent, and related problems of entity matching and information fusion have also arisen. The same real-world entity is often recorded in multiple fields or systems, but these records often show significant differences. Taking personnel information as an example, it may be stored in the superior and subordinate units or multiple business systems, and due to the essential differences in the design specifications, data collection standards and methods of each system, the records of the same entity often have multiple heterogeneous problems: first, the field names are inconsistent, such as "mobile phone" and "contact phone" pointing to the same attribute but expressing different meanings; second, the number of fields is different, such as A system records the user's email information completely, while B system does not set this field; third, the field value formats are not uniform, for example, the date exists in both "YYYY-MM-DD" and "MM / DD / YYYY" formats; fourth, the field values are in conflict, such as the permanent address of the same user recorded completely differently in two systems. The root cause of these problems lies in the differences in the construction purpose and implementation method of each data source, resulting in heterogeneous attribute description, semantic ambiguity, and non-uniform structure of the same entity in different systems, while the data itself also has problems such as quality good and bad, redundant information accumulation, and missing key information.

[0003] If the heterogeneous entity information is not effectively processed and directly applied, it may cause a series of problems: first, there is a risk of data application failure. When the same person is labeled as "25 years old" and "30 years old" in different systems, if there is no effective matching and fusion mechanism, it will cause information splitting, making the application based on these data lose practical significance. Second, it is easy to cause decision bias and resource waste. For example, the duty, post and training experience of a certain person have changed, but the related information has not been updated synchronously between systems, which may lead to misjudgment in decision-making such as human resource allocation, work task allocation, etc., causing unnecessary consumption of human and time resources. Third, it will form obstacles to system cooperation and data governance. In the cross-department data sharing scenario, slight differences in entity attribute description may lead to data docking failure, hinder cross-field business cooperation, and increase the difficulty and cost of data governance.

[0004] To solve the above problems, the core lies in carrying out entity matching and information fusion to ultimately form unified and complete entity information. Entity matching (also known as entity alignment) plays a crucial role in eliminating entity conflicts and unclear referencing in heterogeneous data, clarifying the correspondence between different records and the same entity. Information fusion, through the integration, disambiguation, reasoning, and verification of heterogeneous data, transforms scattered and fragmented knowledge into a high-quality knowledge base within a unified framework. Together, these two processes constitute a key path to breaking down "data silos" and improving data usability, providing a solid foundation for various advanced data application scenarios such as intelligent decision support, knowledge graph construction, and cross-domain business collaboration.

[0005] Currently, mainstream technologies for entity matching and information fusion can be divided into two aspects. Entity matching primarily employs rule-based, similarity-based, and model-assisted decision-making methods; information fusion typically involves rule-based and large-model-based methods. However, existing technologies still face many challenges that urgently need to be overcome: in large-scale data scenarios, entity heterogeneity is diverse, and the limitations and application scenarios of each method make it difficult to achieve an effective balance between accuracy and efficiency. Summary of the Invention

[0006] Therefore, it is necessary to provide a method and system for cross-domain entity identity matching and information fusion that can solve the problems of cross-domain entity recognition accuracy and the need to improve the effect of information fusion conflict handling, in order to address the above-mentioned technical issues.

[0007] A method for cross-domain entity identity matching and information fusion, the method comprising:

[0008] Step 1: Obtain cross-domain entity data, preprocess the cross-domain entity data, and output a standardized entity dataset;

[0009] Step 2: Employ a multi-method complementary matching strategy, using at least two of the following methods—rule-based entity matching, similarity-based entity matching, machine learning-based entity matching, and graph-based entity matching—to perform entity matching on the standardized entity dataset, and output entity matching results. The entity matching results include multi-source attribute records of the same entity.

[0010] Step 3: Adopt a conflict resolution strategy, and use at least two of the following methods: timestamp-based method, confidence weight-based method, voting method, and model-assisted judgment method, to filter the entity matching results based on the multi-source attribute records of the same entity, and output a unique attribute value.

[0011] Step 4, using an information completion strategy, performing information completion on the entity matching result by at least two of a rule-based completion method, a similarity-based completion method, and a large model-based completion method, and outputting completed entity information;

[0012] Step 5, fusing the unique attribute value and the completed entity information, and outputting initial unified entity information;

[0013] Step 6, verifying the initial unified entity information, if the verification is passed, outputting final entity information, if the verification is not passed, adjusting parameters of the corresponding completion method in step 4 and processing again, outputting final entity information, if the parameter correction cannot be passed, triggering manual verification.

[0014] On the other hand, a cross-domain entity identity matching and information fusion system is also provided, comprising:

[0015] An entity data acquisition module is configured to acquire cross-domain entity data, pre-process the cross-domain entity data, and output a standardized entity data set;

[0016] An entity matching result output module is configured to use a multi-method complementary matching strategy to perform entity matching on the standardized entity data set by at least two of a rule-based entity matching, a similarity-based entity matching, a machine learning-based entity matching, and a graph-based entity matching, and output an entity matching result, wherein the entity matching result includes multiple-source attribute records of the same entity.

[0017] A unique attribute value output module is configured to use a conflict resolution strategy to filter the entity matching result based on the multiple-source attribute records of the same entity by at least two of a timestamp-based method, a confidence weight-based method, a voting-based method, and a model-assisted determination method, and output a unique attribute value.

[0018] An entity information completion module is configured to use an information completion strategy to perform information completion on the entity matching result by at least two of a rule-based completion method, a similarity-based completion method, and a large model-based completion method, and output completed entity information.

[0019] A fusion module is configured to fuse the unique attribute value and the completed entity information, and output initial unified entity information.

[0020] A verification module is configured to verify the initial unified entity information, if the verification is passed, output final entity information, if the verification is not passed, adjust parameters of the corresponding completion method in the entity information completion module and process again, output final entity information, if the parameter correction cannot be passed, trigger manual verification.

[0021] Compared with the prior art, the cross-domain entity identity matching and information fusion method and system provided by the application has the following beneficial effects:

[0022] 1. By cross-domain entity data preprocessing, heterogeneous data is uniformly converted into a standardized entity data set, ensuring data format compatibility and attribute description uniformity, laying a standardized foundation for subsequent entity processing and avoiding processing logic breaks caused by data heterogeneity.

[0023] 2. By constructing a multi-method complementary matching strategy using at least two types of matching methods, the problem of insufficient scene adaptability of single matching method in the prior art is effectively solved. The method can be flexibly combined according to the heterogeneity of cross-domain data, realizing the collaborative optimization of cross-domain entity matching efficiency and accuracy.

[0024] 3. For the conflict problem of multiple source attribute records of the same entity, at least two types of conflict resolution methods are used to construct a multi-dimensional screening logic, effectively avoiding the inherent bias of a single conflict resolution method. Compared with the single conflict resolution strategy in the prior art, the multi-dimensional conflict resolution strategy proposed in the application can cross-verify from multiple dimensions such as timeliness, quality, consensus, etc., ensuring that the output unique attribute value has timeliness and accuracy, and laying a reliable foundation for subsequent entity information fusion.

[0025] 4. By constructing an information completion strategy using at least two types of information completion methods, the completeness of cross-domain entity data can be significantly improved, reducing the impact of missing key information on subsequent applications.

[0026] 5. Through the closed-loop mechanism of "verification, parameter adjustment, and manual backup", errors in the initial unified entity information are checked from multiple dimensions, ensuring the accuracy and compliance of the final output entity information.

[0027] 6. The method proposed in the application can integrate heterogeneous entity data scattered in different domains into unified, complete and accurate entity information, effectively breaking down data silos and providing high-quality data support for intelligent decision support, cross-domain business collaboration, knowledge graph construction and other advanced data applications, with strong practical application value. BRIEF DESCRIPTION OF DRAWINGS

[0028] In order to more clearly illustrate the technical solutions of the embodiments of the application, the following will briefly introduce the drawings needed in the embodiments. It should be understood that the following drawings only show some embodiments of the application, and those skilled in the art can obtain other related drawings from these drawings without creative labor.

[0029] Figure 1 The flowchart of the cross-domain entity identity matching and information fusion method in embodiment 1 is shown in the figure.

[0030] Figure 2Flowchart for the layered progressive strategy in embodiment 1;

[0031] Figure 3 Flowchart for the parallel invocation strategy in embodiment 1;

[0032] Figure 4 Flowchart for the layered feature extraction strategy in embodiment 1;

[0033] Figure 5 Flowchart for the serial fusion strategy in embodiment 1;

[0034] Figure 6 Flowchart for the parallel fusion strategy in embodiment 1;

[0035] Figure 7 Flowchart for the layered completion strategy in embodiment 1;

[0036] Figure 8 Structure block diagram of the cross-domain entity identity matching and information fusion system in embodiment 2;

[0037] Figure 9 Internal structure diagram of the computer device in embodiment 3.

[0038] The purposes, functional features and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0039] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0040] It should be noted that the description such as "first", "second" and the like in the present application is only for the purpose of description, and cannot be understood as indicating or implying the relative importance of the technical features indicated or the number of the technical features indicated. Therefore, the features limited by "first", "second" can explicitly or implicitly include at least one of the features. In the description of the present application, the meaning of "multiple" is at least two, for example, two, three, etc., unless otherwise specifically limited.

[0041] It can be understood that the technical solutions of various embodiments of the present application can be combined with each other, but it must be based on the realization of those of ordinary skill in the art, and when the combination of technical solutions appears to be contradictory or unachievable, it should be considered that the combination of technical solutions does not exist, nor is it within the protection scope required by the present application.

[0042] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0043] Example 1

[0044] This embodiment discloses a cross-domain entity identity matching and information fusion method, the entire process including two stages: entity matching and information fusion. The core objective of entity matching is to accurately determine whether two or more records from different domains or data sources point to the same entity in the real world; the accuracy of entity matching directly determines the quality of subsequent information fusion.

[0045] After identifying the same entity through entity matching, the next step is information fusion. This stage merges scattered and heterogeneous records into a unified, complete, and accurate standardized record, primarily focusing on conflict resolution and information completion. Conflict resolution requires eliminating inconsistencies in attribute values. When different data sources contain inconsistent values ​​for the same attribute of the same entity, conflict resolution is essential to ensure the accuracy of the merged record. Conflict resolution uses pre-defined rules or algorithms to identify and filter conflicting attribute values, ultimately determining a unique and accurate value to avoid logical contradictions in the merged data.

[0046] Information completion requires ensuring both data integrity and uniqueness. The purpose of information completion is to integrate valid information from multiple sources, improve the completeness of physical archives, and remove redundant information to guarantee data uniqueness. In terms of integrity improvement, the complementarity of records from different data sources is used to fill information gaps. Regarding uniqueness assurance, duplicate information in multi-source data needs to be deduplicated to avoid data redundancy consuming storage resources and affecting the efficiency of subsequent data applications.

[0047] This invention solves the core problems in the prior art, namely, the need to improve the accuracy of cross-domain entity recognition and the effectiveness of information fusion conflict handling.

[0048] like Figure 1 As shown, the cross-domain entity identity matching and information fusion method provided in this embodiment includes the following steps:

[0049] Step 1: Obtain cross-domain entity data, preprocess the cross-domain entity data, and output a standardized entity dataset.

[0050] Cross-domain entity data refers to entity records from different domains or systems, primarily obtained through API calls, direct database connections (such as JDBC / ODBC), or standardized file imports (such as CSV and JSON). For example, for a given cross-domain entity data, the original data of the entity to be compared in domain A and domain B are obtained synchronously using the above methods to ensure the integrity of the data fields.

[0051] The preprocessing is a key operation for eliminating the heterogeneity of cross-domain data, mainly including three links: 1. Structuring of unstructured data, including: extracting key attributes by regular matching for semi-structured text; identifying text information by OCR technology for picture data; and converting various data into key-value pair format of "attribute name-attribute value" by natural language processing methods such as named entity recognition.

[0052] 2. Cleaning of structured data, deleting redundant fields, eliminating fields such as "notes" and "temporary markers" without value; processing noise and outliers, removing special characters in text fields by regular matching, eliminating abnormal data by numerical range checking (such as mobile phone number should be 11 digits, age should be between 0-120 years old); processing missing values, data with missing core identifier fields (such as ID number) are directly marked as "to be manually verified", and non-core fields are filled with the mean value of the same entity cluster or default value.

[0053] 3. Data format unification, mapping of synonymous fields, establishing a global field mapping table to map "mobile phone number, contact number, mobile phone number" to the standard field "telephone"; format standardization, date unified to "YYYY-MM-DD", amount unified to "yuan" (2 decimal places), weight unified to "kg"; unit normalization, such as converting "grams" and "tons" to "kg", and "ten thousand yuan" to "yuan".

[0054] The standardized entity data set refers to an entity data set with unified field name, format and type after the above preprocessing; and the cross-domain entity data is converted, cleaned and unified to lay a foundation for subsequent matching.

[0055] Step 2, a multi-method complementary matching strategy is adopted, at least two types of methods of rule-based entity matching, similarity-based entity matching, machine learning-based entity matching and graph-based entity matching are used for entity matching of the standardized entity data set, and an entity matching result is output, the entity matching result including multiple source attribute records of the same entity.

[0056] It can be understood that the present application combines the advantages of at least two types of entity matching methods through a multi-method complementary matching strategy, makes up for the limitations of single method in scene adaptability, precision or efficiency, and realizes collaborative optimization.

[0057] Among them, the rule-based entity matching refers to defining deterministic matching rules and deterministic exclusion rules based on domain knowledge, and directly screening entity pairs through the deterministic matching rules and the deterministic exclusion rules.

[0058] The similarity-based entity matching is to calculate attribute similarity and weighted sum, and determine the matching state according to a preset threshold. For suspected entity pairs not covered by the rules, the similarity-based entity matching is used to further narrow the range and improve the matching efficiency. The advantage is intuitive, fast, and interpretable, and it is suitable for the scenes of unified field structure, clear core identifier, small and medium scale, high real-time, strong interpretability demand.

[0059] For entity pairs with similar similarity, the entity matching based on machine learning converts the matching problem into a classification or regression task, learns the mapping relationship between the feature vector of the entity pair and the matching probability through the model, improves the matching accuracy, and solves the matching problem in the complex semantic scene. The advantage is strong generalization, controllable precision, and support for dynamic optimization, and it is suitable for cross-industry / system, semantic field, dynamic data, large-scale complex integration, high-precision and long-term optimization scenes.

[0060] The graph-based entity matching is to abstract entities as nodes and the association between entities as edges, construct a local association subgraph, and determine the matching through graph features. The advantage is to break the field dependence, support incremental matching, and capture implicit associations. It is suitable for scenes with explicit entity association, high graph connectivity, dense relationships, sparse fields, and frequent incremental updates.

[0061] The output entity matching result includes matching pairs and non-matching pairs, wherein the matching pair refers to the entity record pair determined as the same entity, and the non-matching pair refers to the entity record pair determined as different entities. The multi-source attribute record of the same entity refers to the attribute record set from different data sources that is confirmed by matching to point to the same real entity.

[0062] Step 3, adopt a conflict resolution strategy, filter the entity matching result based on the same entity multi-source attribute record by at least two types of methods based on timestamp method, confidence weight method, voting method, and model-assisted determination method, and output unique attribute values.

[0063] It can be understood that the conflict resolution strategy eliminates the attribute value contradiction in the same entity multi-source attribute record by at least two types of methods, and ensures the accuracy of the attribute value after filtering.

[0064] Specifically, the timestamp-based method is to preferentially select the attribute value with the latest timestamp, which is suitable for time-sensitive attributes and invalid for static attributes. The advantage is the fastest speed, and it is suitable for time-sensitive scenes.

[0065] The confidence weight-based method is to assign a confidence weight to each data source, and preferentially use the attribute value of the data source with high confidence when there is a conflict. The advantage is that it can distinguish data quality and has high precision.

[0066] The voting-based method is to group multi-source records according to attribute values, count the number of votes in each group, and adopt the majority result with the highest number of votes, which is suitable for scenarios where the data source quality is similar. The advantages are fast speed and low cost.

[0067] The model-based auxiliary determination method is to output the optimal probability of each candidate attribute value through a pre-trained model, and determine the optimal value based on the probability.

[0068] The output unique attribute value refers to the single attribute value determined after screening and without contradiction, providing a consistent attribute basis for subsequent information fusion.

[0069] Step 4, adopt information completion strategy, through at least two kinds of methods of rule-based completion method, similarity-based completion method, and large model-based completion method to complete the entity matching result, and output the completed entity information.

[0070] It can be understood that the information completion strategy is to fill in the null fields in the entity record and integrate multi-source conflicting information through at least two kinds of methods to improve the completeness and consistency of the entity information.

[0071] Among them, the rule-based completion method is to fill in the null values according to the pre-defined rules based on domain knowledge or data association, including deduction type rules, enumeration type rules, and fixed format rules. The advantages are high accuracy and strong interpretability.

[0072] The similarity-based completion method is to calculate the feature similarity between the entity to be completed and other entities, and select the common information with high similarity, and complete the missing attributes of another entity with the known attributes of the high similarity entity. The advantage is that it does not need to manually develop complex rules or large-scale annotated data.

[0073] The large model-based completion method is to use the semantic understanding ability of large language models to integrate multi-source conflicting fields and verify the consistency of the completion results with other information of the entity. The completed entity information refers to the entity record with null fields filled in, conflicting fields integrated, and more complete information. The advantage is that it can capture non-linear associations of attributes and has strong generalization; the disadvantage is poor interpretability.

[0074] Step 5, fuse the unique attribute value and the completed entity information to output the initial unified entity information.

[0075] It can be understood that the fusion is to integrate the unique attribute value obtained after conflict resolution and the completed entity information obtained after information completion according to a unified framework, that is, to retain the non-conflicting fields after completion and replace the fields still in conflict after completion with the unique attribute value.

[0076] The initial unified entity information is the preliminary unified record containing complete attributes of the entity after integration, which needs to be further verified to ensure accuracy.

[0077] Step 6, the initial unified entity information is checked, if the check is passed, the final entity information is output; if the check is not passed, the parameters of the corresponding completion method in step 4 are adjusted and then processed again, and the final entity information is output; if it cannot be corrected by parameters, manual checking is triggered.

[0078] It can be understood that the check is a multi-dimensional quality check on the initial unified entity information, including format check, logic check, and integrity check; manual check is when the parameter adjustment still cannot solve the check error, the domain expert determines the final information according to the data source authority and business logic. The final entity information refers to the unified entity record that passes the check, has no format error, is logically consistent, and has complete core fields, which can be directly used for intelligent decision-making, knowledge graph construction and other downstream applications.

[0079] The cross-domain entity identity matching and information fusion method provided by the application eliminates cross-domain data heterogeneity through data preprocessing, improves entity recognition accuracy through multi-method complementary matching, ensures attribute value consistency through multi-dimensional conflict resolution, improves data integrity through hierarchical information completion, and finally integrates scattered cross-domain entity data into unified, accurate and complete entity information, effectively breaking down data silos and providing high-quality data support for cross-domain business collaboration, intelligent decision-making support and other scenarios.

[0080] In one embodiment, as shown in Figure 2 In step 2, the multi-method complementary matching strategy is a hierarchical progressive strategy, including:

[0081] Step 201, through rule-based entity matching, according to the preset deterministic matching rules and deterministic exclusion rules, the entity pairs in the standardized entity data set are filtered, and the first matching result and the first to be determined result are output.

[0082] It can be understood that the deterministic matching rule refers to a preset rule based on the unique identifier attribute or key attribute combination of the entity, which can unambiguously determine that the multi-source records point to the same real entity, such as complete identity card number, complete name and mobile phone number, etc.; if the deterministic matching rule is met, it belongs to the entity pair that meets the matching requirements, and it is directly included in the first matching result without further verification.

[0083] The deterministic exclusion rule refers to a preset rule based on the key attribute conflict or attribute feature of the entity, which can unambiguously determine that the multi-source records point to different real entities, such as inconsistent gender without reasonable explanation, missing core identifier field without other matching attributes, etc.; if the deterministic exclusion rule is met, it belongs to the entity pair that excludes the matching possibility, and is directly regarded as a non-matching result.

[0084] If the entity pair does not satisfy the deterministic matching rule and the deterministic exclusion rule, it is regarded as a state of uncertain matching, is included in the first to-be-determined result, and jumps to step 202.

[0085] In step 202, the entity pairs in the first to-be-determined result are screened according to the calculation attribute similarity, and the second matching result and the second to-be-determined result are outputted by the similarity-based entity matching.

[0086] For the suspected entity pairs not covered by the rules, the efficiency of the similarity method is further reduced, and the standardized field in step 1 is reused for the first to-be-determined result, without additional processing of the attribute name.

[0087] Specifically, first assume that the attribute name set of the entity pair is , where is the number of attributes, the corresponding attribute value set is , and the weight set of the attribute is .

[0088] When calculating, for the text attribute in the entity pair, the character similarity is calculated by using the edit distance (Levenshtein), and the semantic similarity is calculated by using Word2Vec, and the average of the two is taken as the text attribute similarity.

[0089] For the numerical, identification attribute (such as mobile phone number, ID number) in the entity pair, the exact match (consistent as 1, inconsistent as 0) is adopted.

[0090] For the date attribute in the entity pair, the normalized value of the time difference is calculated (such as difference ≤ 30 days as 1, difference > 1 year as 0).

[0091] Through the above method, the attribute name similarity set of the entity pair is calculated as , and the attribute value similarity set is . The similarity of the attribute name and the attribute value is respectively weighted and integrated, and then the average of the sum of the evaluation values of the attribute name and the attribute value is calculated according to the same weight, to calculate the total similarity score after weighted summation. The threshold interval is set as ; if the total similarity score of the attribute corresponding to the entity pair is located in the range of , it is included in the second to-be-determined result, and jumps to step 203; if the total similarity score of the attribute corresponding to the entity pair is greater than , it is included in the second matching result; if the total similarity score of the attribute corresponding to the entity pair is less than , the entity pair is directly regarded as a non-matching result, and is excluded.

[0092] In step 203, through entity matching based on machine learning, the model learns the mapping of features and results, and is used to screen the entity pairs in the second to-be-determined result, and output the third matching result and the third to-be-determined result.

[0093] For entity pairs with similar degrees of suspicion, a machine learning model is called to improve accuracy and solve the matching problem in complex semantic scenarios.

[0094] Specifically, first, a multi-dimensional feature set of the entity pair is constructed, including: basic features (attribute similarity scores in step 202); semantic features (vectorization of long text such as names and addresses through BERT, and then calculation of cosine similarity); statistical features (attribute missing rate, field consistency proportion), etc.

[0095] The model uses a pre-trained Siamese-BERT binary classification model, and the input is a multi-dimensional feature set, and the output is a matching confidence score.

[0096] If the matching confidence score is in the range of , it is included in the third to-be-determined result, and step 204 is skipped; if the matching confidence score is greater than , it is included in the third matching result; if the matching confidence score is less than , the entity pair is directly considered as a non-matching result and excluded.

[0097] In step 204, through graph-based entity matching, the entity and the associated relationship are abstracted as nodes and edges to construct a local association subgraph, and the entity pairs in the third to-be-determined result are screened to output the fourth matching result.

[0098] For high-suspected entity pairs that cannot be determined by machine learning, entity association relationships are used for supplementary verification to break through the limitations of single attribute dependence.

[0099] Specifically, first, the entity in the third to-be-determined result is taken as the core node, its first-level associated entities are taken as the expansion nodes, and the relationship between entities is taken as the edge to construct a local subgraph, avoiding the complexity of full graph calculation.

[0100] Then, the Jaccard coefficient is used to calculate the proportion of common associated nodes; the relationship path similarity is calculated according to the path extraction, path feature quantization, and similarity aggregation method, such as two entities being directly connected to the same unit node, then the path similarity is 1; the graph feature score is obtained according to the Adamic-Adar index neighborhood structure similarity.

[0101] The graph feature score and the matching confidence score in step 203 are weighted and fused according to the weight to obtain the final confidence score. If the final confidence score is greater than If the final confidence score is less than , the entity pair is directly regarded as a non-matching result and excluded.

[0102] Step 205, integrating the first matching result, the second matching result, the third matching result and the fourth matching result, and outputting the final entity matching result.

[0103] It can be understood that the integration is to combine the first matching result to the fourth matching result, remove the repeated matching pairs, and combine the determination basis of each result to form the final entity matching result, to ensure that the result is traceable.

[0104] Further, the error matching and missing matching samples confirmed by manual review are fed back to the corresponding steps of steps 201 to 204 to adjust and optimize the parameters; the optimized parameters and the model are automatically synchronized to the matching process for subsequent cross-domain entity comparison tasks. Among them, in the optimization of rule-based entity matching, new coverage rules are added for missing matching samples, and constraint rules are supplemented for error matching samples. In the optimization of similarity-based entity matching, the attribute weight is recalculated based on the feedback samples, and the threshold is fine-tuned. In the optimization of machine learning-based entity matching, the entity pairs of error matching / missing matching are labeled as training samples and supplemented to the model training set, and the Siamese-BERT model is fine-tuned again to optimize the feature weight. In the optimization of graph-based entity matching, for the relationship mis-checking samples, the incorrect association edges in the graph are corrected, the missing association edges are supplemented, and the accuracy of the graph structure is improved.

[0105] It can be understood that the hierarchical progressive strategy uses the logic of first rough screening and then accurate judgment to quickly filter the clear matching and non-matching entity pairs by rule matching first, and then gradually processes the suspected pairs with higher complexity by similarity matching, machine learning matching and graph matching, which not only guarantees the accuracy of entity matching, but also controls the consumption of computing resources through layer-by-layer screening, realizing the collaborative optimization of efficiency and accuracy.

[0106] In one embodiment, as shown in Figure 3 , in step 2, the multi-method complementary matching strategy is a parallel calling strategy, including:

[0107] Step 211, through rule-based entity matching, the entity pairs in the standardized entity data set are screened according to the preset deterministic matching rules and deterministic exclusion rules, and the first matching result and the first to-be-determined result are output.

[0108] Step 211 is completely consistent with step 201, that is, the first matching result and the first to-be-determined result are screened through the deterministic matching rules and the deterministic exclusion rules, which will not be repeated here.

[0109] The difference is that if an entity pair does not meet either the deterministic matching rule or the deterministic exclusion rule, it is considered an undeterminable matching state, included in the first undetermined result, and the process jumps to step 212.

[0110] Step 212: The first result to be determined is processed in parallel by entity matching based on similarity, entity matching based on machine learning, and entity matching based on graph, and the total similarity score, matching confidence score, and graph feature score are output respectively.

[0111] It is understandable that parallel processing refers to simultaneously applying three methods—similarity-based entity matching, machine learning-based entity matching, and graph-based entity matching—to each suspected entity pair in the first result to be determined, rather than processing them sequentially.

[0112] The similarity-based entity matching is calculated in the same way as step 202, with the first result to be determined as the input and the total similarity score as the output.

[0113] The calculation method for entity matching based on machine learning is the same as in step 203, except that the input is the first result to be determined, and the output is the matching confidence score.

[0114] The calculation method for graph-based entity matching is the same as in step 204, except that the input is the first result to be determined, and the output is the graph feature score.

[0115] Step 213: The total similarity score, matching confidence score, and graph feature score are corrected by cross-validation to obtain the corrected total similarity score, matching confidence score, and graph feature score.

[0116] Specifically, a result difference threshold is set. If the difference between the scores of any two matching methods exceeds this threshold, feature backtracking verification is triggered. By extracting the core features relied upon by the two methods, the cause of the difference is determined, outliers are corrected, and the corresponding scores are adjusted. For example, the core feature relied upon for machine learning-based entity matching is address semantic matching; the core feature relied upon for similarity-based entity matching is phone number matching. If the phone numbers are inconsistent but the addresses and associations are consistent, the low overall similarity score caused by the incorrect phone number is removed, and the similarity score is recalculated. It is worth noting that after correction, it is necessary to ensure that the results of the three matching methods are logically consistent to avoid outliers from a single matching method affecting the judgment.

[0117] Step 214: Calculate the weighted composite score of the corrected total similarity score, the matching confidence score, and the graph feature score; if the weighted composite score is greater than or equal to the preset high-precision threshold, it is determined as a new matching result; if the weighted composite score is less than the preset high-precision threshold, it is determined as a non-matching result.

[0118] Specifically, a high-precision threshold is pre-set. Then, based on the method's accuracy and domain adaptability, weights are assigned to similarity-based entity matching, machine learning-based entity matching, and graph-based entity matching. The weight for similarity-based entity matching is... The weights for entity matching based on machine learning are: The weights for graph-based entity matching are: And the weights satisfy .

[0119] Then, the weighted composite score for each entity pair is calculated using the following expression:

[0120] Weighted composite score = total similarity score × +Match confidence score × +Final confidence score × ;

[0121] If the weighted composite score is greater than or equal to the preset high-precision threshold, it is determined as a new matching result; if the weighted composite score is less than the preset high-precision threshold, it is determined as a non-matching result.

[0122] Step 215: Integrate the first matching result with the newly added matching result, and output the final entity matching result.

[0123] It is understandable that integration involves merging the first matching result with the newly added matching result, while recording the score and correction basis for each type of matching method, to form the final entity matching result.

[0124] It is understandable that the parallel invocation strategy significantly shortens the processing time of suspected entity pairs by simultaneously calculating multiple methods. At the same time, it corrects the deviation of a single method through cross-validation and ensures high accuracy by combining weight skew. It is suitable for scenarios with high timeliness requirements for pair processing and can improve the reliability of matching results through cross-validation of multi-dimensional results.

[0125] In one embodiment, such as Figure 4 As shown, in step 2, the multi-method complementary matching strategy is replaced with a hierarchical feature extraction strategy. This constructs entity feature vectors from three dimensions: basic, semantic, and relational, comprehensively characterizing entity differences and associations. The hierarchical feature extraction strategy includes:

[0126] Step 221: Extract basic attribute features, semantic association features, and relational network features from the standardized entity dataset, and output basic attribute feature vectors, semantic association feature vectors, and relational network feature vectors.

[0127] It is understandable that in the basic attribute feature extraction, quantitative indicators are used to describe the degree of difference in the surface matching relationship and data normalization of entity fields. A total of 5 types of features are extracted, and all 5 types of features are normalized to the range [0,1]. Among them, the first type is the core identifier matching degree, which is 1 for unique identifier fields such as ID card numbers, and 0 otherwise; the second type is the text field similarity, which is calculated for text fields such as names and addresses, and then normalized, with larger values ​​indicating greater similarity; the third type is the numerical field difference, which is calculated for numerical fields such as age and amount. The variance is calculated as follows: for example, if the age range is 0-120 years, and values ​​A=25 and B=30, then the variance = 1 - 5 / 120 ≈ 0.96. The fourth category is data type consistency: 1 is awarded for data types (string / numeric / date) that are completely consistent across the same standard field; otherwise, it is awarded as 0 (e.g., if all "gender" fields are strings, then it is 1). The fifth category is field completeness: the percentage of non-missing values ​​in common standard fields is calculated. For example, if one of six common fields is missing, then the completeness = 5 / 6 ≈ 0.83. The final output is a 5-dimensional basic attribute feature vector.

[0128] Semantic association feature extraction breaks through the limitations of surface-level fields by capturing the semantic relationships between field values ​​through pre-trained language models, correcting misjudgments that are literally different but semantically consistent. For example:

[0129] For text-type field values ​​only, such as addresses “XX City XX Road XX No.” and “XX City XX District XX Road XX No.”, there is no need to calculate semantic similarity for field names that have been mapped to a unified standard.

[0130] Then, the BERT model, fine-tuned based on entity matching corpus, is used to convert text field values ​​into 768-dimensional semantic vectors.

[0131] Then, semantic similarity is calculated, which is divided into the following types: Text-type value: directly calculate the cosine similarity of the corresponding field value vectors of the two entities; Enumerated value: such as gender "male / female" and education "undergraduate / master", use exact matching, if they match, it is 1, otherwise it is 0; Date-type value: calculated according to the formula 1-|time difference (days)| / 365, normalized according to a 1-year cycle, the smaller the time difference, the larger the value.

[0132] Then, semantic feature aggregation is performed. The average semantic similarity of all text, enumeration, and date fields is taken as the overall semantic association feature of the entity and normalized to [0,1] to obtain a 1-dimensional semantic association feature vector.

[0133] Relationship network feature extraction is suitable for scenarios where there is entity-related data. It supplements the matching criteria for cases with low field similarity but close association by using relationship network features. The specific process is as follows:

[0134] First, collect associated entities and extract a list of first-level associated entities of entity A and entity B from the graph, such as a person's direct superior, and clarify the type of association relationship.

[0135] Then, the core relationship features are calculated and normalized to [0,1]. Specifically, the calculation is performed from the following perspectives: the proportion of jointly associated entities: (0 for no common entity, 1 for complete consistency); Relationship path similarity: For hierarchical relationships, such as personnel → department → unit, calculate the proportion of common superior nodes. For example, if A and B both belong to "Technology Department", the path similarity is 1; Association frequency consistency: Count the number of interactions between entities and related entities in the past 3 months, and calculate 1 - |frequency A - frequency B| / (frequency A + frequency B + 1) (avoid denominator of 0).

[0136] Then, relational feature aggregation is performed: corresponding weights are assigned to the common association ratio, relational path similarity, and association frequency consistency, and weighted summation is performed to generate a 1-dimensional relational network feature vector.

[0137] Step 222: Concatenate the basic attribute feature vector, semantic association feature vector, and relational network feature vector to output the entity feature vector.

[0138] It can be understood that the concatenation is to combine the 5-dimensional basic attribute feature vector, the 1-dimensional semantic association feature vector, and the 1-dimensional relational network feature vector in sequence to form a 7-dimensional initial entity feature vector. Then, the initial vector is processed by Min-Max normalization to ensure that the dimensions of all feature dimensions are uniform and fall within the [0,1] interval, so as to avoid a single feature dominating the model training and output a standardized entity feature vector.

[0139] Step 223: Construct a matching prediction model. Input the entity feature vector into the matching prediction model for processing and output the final entity matching result.

[0140] The Siamese-Transformer dual-tower model was selected for the matching prediction model. During model preparation, domain fine-tuning was performed based on labeled samples (10,000+ "matching pairs" and "non-matching pairs," covering different cross-domain scenarios). The input was a pair of 7-dimensional entity feature vectors for entity A and entity B, and the output was the matching confidence score in the range [0,1]. A higher output value indicated a greater likelihood of the entities being the same. A threshold of t was set. During prediction, the entity feature vector pairs were input into the fine-tuned matching prediction model, and the matching confidence score was output. The final entity matching result was determined as follows: if the matching confidence score ≥ t, the pair was marked as a matching pair of the same entity; if the score < t, the pair was marked as different entities. Additionally, core matching criteria were recorded simultaneously, such as basic features (0.92), semantic features (0.88), relational features (0.90), and a comprehensive score (0.91 ≥ 0.85). It is worth noting that the threshold t was determined based on the optimal F1-score on the validation set.

[0141] It is understandable that the hierarchical feature extraction strategy comprehensively portrays the differences and associations of entities through three dimensions of features: basic, semantic, and relational, avoiding the limitations of a single field or a single method. Combined with the adapted Siamese-Transformer model, it can effectively learn the complex mapping relationship of multi-dimensional features, greatly improving the accuracy of cross-domain entity matching. At the same time, after matching, a full lifecycle profile of the entity containing feature basis can be constructed simultaneously to support data traceability and subsequent reuse.

[0142] In one embodiment, such as Figure 5 As shown, in step 3, the conflict resolution strategy is a serial fusion strategy, including:

[0143] Step 301: Obtain the multi-source attribute records of the same entity in the entity matching results, including conflict version ID, attribute value, source domain name, timestamp, confidence weight of the source domain, historical accuracy of each source domain, and attribute type.

[0144] It is understandable that when the system detects an entity attribute version conflict, it collects basic data to ensure that subsequent steps have data to rely on. Specifically, multi-source attribute records for the same entity include version information, which includes the conflicting version ID, attribute value, source domain name, timestamp (update time), and source domain confidence weight. Auxiliary information includes the historical accuracy of each source domain and attribute type.

[0145] Among them, the conflict version ID is a unique identifier that distinguishes attribute versions from different data sources; the attribute value is the entity attribute content recorded in each data source; the source domain name is the data source to which the attribute record belongs; the timestamp is the update time of the attribute record; the confidence weight of the source domain is a value set based on the historical accuracy of the data source and the data collection standards; the historical accuracy of each source domain is the matching ratio between the past attribute values ​​and the true values ​​of that data source; the attribute type is divided into time-sensitive (such as address, job title, salary, which need to reflect the latest status) and static (such as ID number, gender, date of birth, which do not change in the long term).

[0146] Step 302: Using a timestamp-based method, based on the attribute type, the attribute value with the latest timestamp is filtered by the time-sensitive attribute, and the source domain corresponding to the attribute value with the latest timestamp is taken as the first candidate source; if it is a static attribute, all source domains corresponding to the attribute value are extracted as the second candidate source; the first candidate source and the second candidate source are integrated to output the candidate source set, and then the process jumps to step 303.

[0147] Specifically, the filtering logic for time-sensitive attributes is as follows: compare the update times of all conflicting versions, select the attribute value with the latest timestamp, and its corresponding source domain is the first candidate source; static attributes do not need to be updated, so all source domains that provide the attribute value are used as the second candidate source, and then proceed to step 303.

[0148] Step 303: Using a confidence weight-based method, obtain the confidence weight corresponding to each candidate source in the candidate source set; determine whether the confidence weight is greater than the high confidence threshold; if so, output the candidate source attribute value corresponding to the confidence weight as the unique attribute value; if not, record the candidate source with the highest confidence weight as the core candidate source, and jump to step 304.

[0149] Specifically, a high confidence threshold is set; if the confidence weight of a candidate source in the candidate source set is greater than the high confidence threshold, its attribute value is directly used as the unique attribute value; if the confidence weight of all candidate sources is less than or equal to the high confidence threshold, the candidate source with the highest confidence weight is recorded as the core candidate source, and the process jumps to step 304.

[0150] Step 304: Using a voting-based method, count the number of votes for all possible values ​​of the current attribute in the multi-source attribute records of the same entity, select the value with the highest number of votes as the optimal voting value, and record the voting association domain corresponding to the optimal voting value; determine whether the core candidate source is in the voting association domain; if so, directly output the attribute value corresponding to the core candidate source as the unique attribute value; if not, select the candidate source with the highest confidence weight from the voting association domain as the optimal voting domain, and jump to step 305.

[0151] It is understandable that a voting-based approach is used to achieve consensus across multiple versions. By statistically analyzing the consensus values ​​of all conflicting versions, the rationality of the previous screening results is further evaluated, reflecting the logic of majority rule.

[0152] Specifically, the voting statistics logic is as follows: First, all conflicting versions are grouped according to "attribute value", and the number of votes in each group is counted, i.e., the number of data sources supporting that value; then, the value with the highest number of votes is selected as the optimal voting value, and all source domains corresponding to the optimal voting value are recorded as voting association domains, such as voting association domains = A and C; then, based on the core candidate source in step 303 and the voting association domain in step 304, it is determined whether the core candidate source is in the voting association domain; if so, it means that the previous screening is consistent with the voting consensus, and the attribute value corresponding to the core candidate source is directly output as the unique attribute value; if not, the candidate source with the highest confidence weight is selected from the voting association domain as the optimal voting domain, and the process jumps to step 305.

[0153] Step 305: Using a model-assisted judgment method, construct an auxiliary judgment model. Input the timestamp freshness, confidence weight, and voting support corresponding to the optimal voting domain of each candidate source in the candidate source set into the auxiliary judgment model for processing, and output the optimal probability of each candidate source; determine whether the optimal probability is greater than the auxiliary judgment threshold; if so, output the attribute value corresponding to the candidate source with the highest optimal probability as the unique attribute value; if not, jump to step 306.

[0154] It is understandable to use a model-based decision-making method to resolve complex disagreements. When the results of preceding steps (timeliness, reliability, consensus) are inconsistent, a simple model can be used to make an accurate judgment without the need for complex operations.

[0155] The timestamp freshness calculation formula is "1 - (current time - update time) / 365", and normalized to [0,1]. The closer the update time, the higher the score. The voting support calculation formula is the number of votes for the corresponding value of the domain / the total number of versions. For example, A corresponds to the number of votes for "28 years old" 2 → 2 / 3≈0.67. The confidence weight directly reuses the confidence weight of the source domain in step 301.

[0156] The auxiliary decision-making model is first trained based on historical conflict data to obtain a trained auxiliary decision-making model. During training, the input consists of three-dimensional features: timestamp freshness, confidence weight, and vote support, and the output is the optimal probability of each candidate source.

[0157] Then, set an auxiliary judgment threshold; determine whether the optimal probability is greater than the auxiliary judgment threshold; if so, output the attribute value corresponding to the candidate source with the highest optimal probability as the unique attribute value; if not, jump to step 306.

[0158] Step 306: Manually count the number of times each candidate source appears in the candidate source set in Step 302, the core candidate source in Step 303, the optimal voting domain in Step 304, and the candidate source with the highest optimal probability in Step 305. Output the attribute value corresponding to the candidate source with the most occurrences as the unique attribute value. If the number of occurrences of candidate sources is the same, output the attribute value corresponding to the candidate source with the highest confidence weight as the unique attribute value.

[0159] Specifically, all candidate sources from steps 302 to 305 are listed, and the frequency of each candidate source is counted. If the frequencies are the same, the candidate source with the higher confidence weight is selected, and its attribute value is used as the unique attribute value. If the frequency of candidate sources is the same, the attribute value corresponding to the candidate source with the highest confidence weight is output as the unique attribute value. If there are candidate sources with the same frequency of occurrence and similar confidence weights, they are pushed to domain experts for review. The experts directly select the source based on the authority of the data source and the business logic, and record the review criteria.

[0160] It is understandable that this embodiment adopts a serial fusion strategy for version conflict issues in entity information fusion, starting with simple and then moving to complex, and starting with rules and then providing a fallback. The candidate range is gradually narrowed down by using timestamps, confidence levels, and voting methods. Then, a model is used to assist in the determination and resolution of complex disagreements. Finally, manual statistics are used as a fallback. This approach not only ensures the efficiency of conflict resolution but also improves accuracy through cross-validation based on multiple dimensions, thus balancing efficiency and precision.

[0161] In one embodiment, such as Figure 6 As shown, in step 3, the conflict resolution strategy is a parallel fusion strategy, including:

[0162] Step 311: Obtain the multi-source attribute records of the same entity in the entity matching results, including conflict version ID, attribute value, source domain name, timestamp, confidence weight of the source domain, historical accuracy of each source domain, and attribute type.

[0163] It is understandable that step 311 is completely consistent with step 301 of the serial fusion strategy, that is, to obtain key information such as conflict version ID and attribute value of multi-source attribute records of the same entity, which will not be elaborated here.

[0164] Step 312: The multi-source attribute records of the same entity are processed in parallel using the timestamp-based method, the confidence weight-based method, the voting method, and the model-assisted judgment method, respectively, and the candidate attribute values ​​and corresponding confidence scores are output.

[0165] Parallel processing can be understood as processing multiple source attribute records of the same entity simultaneously using four types of conflict resolution methods to output the values ​​of each candidate attribute and their corresponding confidence scores.

[0166] For the timestamp-based method, the candidate attribute value is the attribute value with the latest timestamp, and the corresponding confidence score is calculated as: 1 - (current time - timestamp) / 365. For the confidence weight-based method, the candidate attribute value is the attribute value from a high-confidence data source, and the corresponding confidence score directly reuses the confidence weight of the source domain from step 301. For the voting method, the candidate attribute value is the optimal voting value, and the corresponding confidence score is calculated as: the number of votes for this attribute value / the total number of conflicting versions. For the model-assisted decision method, the candidate attribute value is the attribute value with the highest optimal probability, and the confidence score is the optimal probability corresponding to this attribute value.

[0167] Step 313: Assign preset weights to the timestamp-based method, confidence weight-based method, voting-based method, and model-assisted judgment method; according to the assigned weights, sum the confidence scores of the same attribute value to obtain the comprehensive score of each candidate attribute value.

[0168] Specifically, the preset weights are set based on the accuracy of the method and the adaptability to the scenario. Generally, model-assisted judgment methods have the highest accuracy and the highest weight allocation; confidence-based methods can distinguish the quality of data sources, so the weight allocation is next; voting-based methods can reflect multi-source consensus, so the weight allocation is next; and timestamp-based methods are suitable for time-sensitive scenarios, so the weight allocation is the lowest. The weight allocation can also be adjusted according to other logic and is not limited to the weight allocation approach provided in this embodiment. When allocating weights, ensure that the sum of the weights is 1.

[0169] Then, based on the assigned weights, the confidence scores for the same attribute value are weighted and summed to obtain the comprehensive score for each candidate attribute value. The expression for calculating the comprehensive score is as follows:

[0170] ;

[0171] In the formula, Indicates the first The overall score of each candidate attribute value; This represents the weights assigned based on the model-assisted decision-making method; This indicates the weights assigned based on the confidence-weighting method; This indicates the weights assigned based on the voting method; This indicates the weights assigned based on the timestamp method; Indicates the first The credibility score corresponding to each candidate attribute value of the model; Indicates the first The confidence score corresponding to each candidate attribute value; Indicates the first The credibility score corresponding to the attribute value of each voting candidate; Indicates the first The credibility score corresponding to each timestamp candidate attribute value.

[0172] Step 314: Determine whether the overall score of each candidate attribute value is greater than the overall judgment threshold; if so, output the attribute value corresponding to the candidate source with the highest overall score as the unique attribute value; otherwise, proceed directly to step 315.

[0173] All domains are ranked by overall score Sort the data from highest to lowest; then set a comprehensive judgment threshold. If the highest comprehensive score is greater than or equal to the comprehensive judgment threshold, the data source attribute value corresponding to the candidate attribute value will be used as the unique attribute value; if the highest comprehensive score is less than the comprehensive judgment threshold, proceed to the manual review stage in step 315.

[0174] Step 315: Push all candidate values ​​and judgment criteria to manual review, and use the manual confirmation result as the unique attribute value output.

[0175] Specifically, manual review requires submitting all candidate attribute values ​​and the criteria for judging the four types of methods, such as the scores, weights, and calculation processes of each method. Experts then confirm the unique attribute values ​​based on the authority of the data source and the business logic.

[0176] It is understood that the parallel fusion strategy in this embodiment synchronously calls four methods and independently calculates the credibility score of each domain. It then performs cross-method aggregation by multiplying the method weight by the domain score, and finally selects the optimal domain based on the total score, thus balancing the efficiency of parallel computing with the accuracy of multi-dimensional fusion.

[0177] In one embodiment, such as Figure 7 As shown, in step 4, the information completion strategy is a hierarchical completion strategy, including:

[0178] Step 401: Obtain multi-source attribute records of the same entity, and filter out null value fields as objects to be completed; based on the synonym field mapping table, unify the heterogeneous fields in the objects to be completed into standard fields.

[0179] It is understandable that null fields refer to fields that are not filled in in the multi-source attribute records of the same entity; the object to be completed is the collection of all null fields; the synonym field mapping table is a preset mapping rule between heterogeneous fields and standard fields. After being unified into standard fields, it ensures that subsequent completion methods process the same fields.

[0180] Then, the unified standard fields are traversed and divided into four core scenarios according to the field status, laying the foundation for subsequent method matching. These scenarios are: Scenario A: Null fields (need to be filled in, such as "Training Experience" being empty); Scenario B: Consistent fields from multiple sources (no need to fill in / merge, just keep them, such as "Name: Zhang San" from multiple sources); Scenario C: Conflicting fields from multiple sources (need to be merged, such as "Age 28" in domain A and "Age 29" in domain B); Scenario D: Unique fields (need to be kept and their source marked, such as the "Emergency Contact" field unique to domain A).

[0181] Step 402: For null value fields with clear logical deduction in the standard fields, perform classification and completion according to the rule-based completion method, output the first completed entity record, and retain the uncompleted text-type null value fields, then jump to step 403.

[0182] It is understandable that a null field with a clear logical derivation refers to a null field that can be determined by calculation or enumeration of other non-null fields.

[0183] Rule-based completion methods leverage the advantages of high accuracy and interpretability of rule-based approaches. They prioritize completion and classification scenarios that can be derived through explicit logic, enabling accurate completion and adapting to all structured fields, such as scenario A, scenario B, and scenario D.

[0184] Specifically, for fields with consistent multi-source information (Scenario B), the value is directly retained and labeled "consistent multi-source information"; for unique fields (Scenario D), the complete value is retained and the source domain is labeled, such as "Emergency contact person: Li Si (source domain A)".

[0185] For null fields (Scenario A), there are completion rules such as deductive rules, enumeration rules, and fixed format rules. Deductive rules calculate completion based on associated non-null fields; for example, if the birth date is null but the age is not, the derivation is: birth date = current year - age (format: YYYY-MM-DD). Enumeration rules complete based on domain enumeration values; for example, if gender is null but there is participation in the men's sports meet, the result is "gender = male". Fixed format rules complete default values ​​based on field standards (non-core fields); for example, if education level is null, it is marked as "to be supplemented" (not applicable to core fields).

[0186] After completing the rule-coverable completion and classification, output the first completed entity record. The first completed entity record contains the completed fields and the uncompleted text-type null value fields. Keep the null value fields (scenario A1) and multi-source conflict fields (scenario C) that cannot be completed by the rule-based completion method, and jump to step 403.

[0187] Step 403: For the text-type null value fields that were not completed in step 402, complete them according to the semantic completion method based on similarity, and output the second completed entity record.

[0188] It is understandable that for text-based null fields (Scenario A1) that cannot be completed by the rules, semantic similarity is used to filter reference entities and expand the scope of completion.

[0189] Specifically, text-based null fields refer to null fields that are recorded in text form and have no explicit derivation rules (such as training experience, job description, etc.).

[0190] The specific process of the similarity-based semantic completion method is as follows:

[0191] The first step is to screen similar entities, which involves three steps: feature extraction, screening scope, and threshold screening. Feature extraction uses Word2Vec / FastText to convert the core non-empty fields of the entity to be completed (such as name, department, and job title) into vectors. The screening scope is limited to entities within the same domain and job title (e.g., if the entity to be completed is a technical engineer, only entities within the same department and job title are screened). Threshold screening calculates cosine similarity and selects the top three entities with similarity scores higher than the entity screening threshold as reference entities.

[0192] Then, the completion content is generated. First, common information is extracted from the null fields of the target reference entity. Then, a lightweight language model (such as BERT-base) is used to deduplicate and sort the common information to generate structured completion content, outputting the second completed entity record. It can be understood that the second completed entity record is the entity record after semantic completion, where text-based null fields are filled.

[0193] Step 404: Construct an entity completion model. Input the first completed entity record, the second completed entity record, and the original non-empty fields of the entity into the entity completion model for processing. Verify the logical consistency between the completed entity record and the non-empty fields, and output the completed entity information.

[0194] Entity completion models employ large-scale language models (such as LLMs) to leverage their semantic understanding and generation capabilities, thereby addressing the issues of multi-source conflict fusion and post-completion information integration and improving the reasonableness of the results.

[0195] Specifically, for multi-source conflict fields (Scenario C), the input to the large language model includes the first completed entity record (rule completion result), the second completed entity record (semantic completion result), and the original non-empty fields of the entity. The model processing includes two parts: conflict integration and logical consistency verification. The logical consistency between the completed entity record and the non-empty fields is verified. If the verification passes, the expression is optimized (such as unifying the format and correcting redundancy) and the completed entity information is output. If the verification fails, the process is backtracked to step 403 to adjust the entity filtering threshold and complete the data again.

[0196] It is understandable that the layered completion strategy follows the logic of classification scenarios, matching methods, and verification loops. Through the layered logic of rule-based completion, similarity-assisted completion, and large-model enhancement, it accurately adapts to different types of null value scenarios. Rule-based completion handles deterministic null values, ensuring accuracy; similarity-based completion expands the coverage of text-based null values, improving completeness; and large-model completion integrates conflicts and verification consistency, ensuring accuracy. The three-layered approach complements each other, avoiding the limitations of a single method and ensuring the reliability of the completion results through logical verification, significantly improving the completeness and usability of cross-domain entity information.

[0197] In one embodiment, step 6 involves verifying the initial unified entity information, including:

[0198] Step 601: Perform format verification to check whether the values ​​of each field in the initial unified entity information conform to the preset standard format.

[0199] It is understandable that the default standard format is a uniform format set according to the field type, such as 11-digit mobile phone number, YYYY-MM-DD date, and gender only "male / female".

[0200] Step 602: Perform logical verification to confirm the consistency of associated fields in the initial unified entity information.

[0201] It is understandable that related fields refer to a combination of fields that have a logical relationship; verifying the consistency of related fields, such as end time > start time, age = current year - birth year ± 1, salary ≥ minimum wage standard for the position.

[0202] Step 603: Perform an integrity check to confirm that no core fields are missing in the initial unified entity information.

[0203] As can be understood, core fields refer to the key fields that identify the entity's identity and support core business; integrity verification is completed by checking whether there are null values ​​in the core fields, and marking the core fields with null values.

[0204] Step 604: If steps 601 to 603 all pass, the final entity information is output; if any verification fails, the parameters of the corresponding completion method in step 4 are adjusted and processed again to output the final entity information; if parameter correction fails, manual verification is triggered, and the final entity information is output based on manual judgment.

[0205] It is understandable that parameter adjustments are mainly aimed at identifying the parameters of the corresponding completion method based on the reasons for the validation failure. For example, if the format validation fails, the fixed format rules of the rule-based completion method are adjusted, and the process is backtracked to step 402 to re-execute the information completion process after parameter adjustment; if the logic validation fails, the entity filtering threshold of the similarity-based semantic completion method is adjusted, and the process is backtracked to step 403 to re-execute the information completion process after parameter adjustment; if the completeness validation fails, the input template of the entity completion model is adjusted, and the process is backtracked to step 404 to re-execute the information completion process after parameter adjustment.

[0206] After repeating the information completion process with adjusted parameters, a second verification is performed. If the verification still fails after parameter adjustment, manual verification is triggered. This involves pushing abnormal fields and relevant evidence, such as the outputs of each completion step and data source records, to domain experts who combine business logic and data source queries to supplement information and output the final entity information; or confirming it as abnormal.

[0207] This embodiment comprehensively checks the quality of the initial unified entity information through three dimensions: format, logic, and completeness. It combines parameter adjustment with manual verification to form a closed-loop quality control system, preventing erroneous information from flowing into downstream applications and ensuring the accuracy, compliance, and usability of the final entity information.

[0208] It should be understood that, although this embodiment Figures 1 to 7 The steps are shown sequentially as indicated by the arrows, but they are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order in which these steps are performed; they can be executed in other orders. Figures 1 to 7 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0209] Example 2

[0210] Based on the cross-domain entity identity matching and information fusion method in Embodiment 1, this embodiment discloses a cross-domain entity identity matching and information fusion system, such as... Figure 8 As shown, the cross-domain entity identity matching and information fusion system includes: an entity data acquisition module 701, an entity matching result output module 702, a unique attribute value output module 703, an entity information completion module 704, a fusion module 705, and a verification module 706, wherein:

[0211] The entity data acquisition module 701 is used to acquire cross-domain entity data, preprocess the cross-domain entity data, and output a standardized entity dataset.

[0212] Specifically, cross-domain entity data acquisition is achieved through API calls, direct database connections, and standardized file imports (such as CSV and JSON format files). This involves collecting the original entity data to be processed across domains, simultaneously acquiring the entity's core attributes (such as name and ID number), associated information (such as affiliated organization), and data source metadata (such as source domain and update timestamp), and outputting the structured collection results to the preprocessing module.

[0213] Preprocessing involves three levels of processing on cross-domain entity data: field cleaning, format unification, and mapping alignment. The final output is a standardized entity dataset with unified field names, formats, and types, providing consistent data input for subsequent modules.

[0214] The entity matching result output module 702 is used to employ a multi-method complementary matching strategy to perform entity matching on the standardized entity dataset using at least two of the following methods: rule-based entity matching, similarity-based entity matching, machine learning-based entity matching, and graph-based entity matching. The entity matching result includes multi-source attribute records of the same entity.

[0215] Specifically, the multi-method complementary matching strategy includes three optional strategies: a hierarchical progressive strategy, a parallel invocation strategy, and a hierarchical feature extraction strategy. The module can dynamically select the appropriate strategy based on the business scenario. When using the hierarchical progressive strategy, matching methods based on rules, similarity, machine learning, and graphs are invoked sequentially to filter entity pairs layer by layer. When using the parallel invocation strategy, similarity, machine learning, and graph matching methods are invoked in parallel after rule-based filtering, and then weighted and fused after cross-validation. When using the hierarchical feature extraction strategy, three dimensions of features—basic, semantic, and relational—are extracted and input into the Siamese-Transformer model to predict the matching results. Entity matching results include matched pairs, non-matched pairs, and records of multiple attributes of the same entity. The module synchronously records the judgment criteria for each matching result to ensure traceability.

[0216] The unique attribute value output module 703 is used to adopt a conflict resolution strategy and filter entity matching results based on multiple source attribute records of the same entity using at least two of the following methods: timestamp-based method, confidence weight-based method, voting-based method, and model-assisted judgment method, and output a unique attribute value.

[0217] Specifically, the conflict resolution strategy includes two optional strategies: serial fusion and parallel fusion. When using the serial fusion strategy, attribute conflicts are processed in the following order: timestamp filtering, confidence filtering, voting filtering, model-assisted judgment, and manual statistical fallback. When using the parallel fusion strategy, the four conflict resolution methods are called in parallel, and the output results of each method are weighted and fused.

[0218] The core output of the module is a unique attribute value, that is, a single attribute value that is consistent with the conflict resolution process. At the same time, the judgment criteria for each type of conflict resolution method are recorded to ensure that the results are traceable.

[0219] The entity information completion module 704 is used to employ an information completion strategy, and complete the entity matching results using at least two of the following methods: rule-based completion method, similarity-based completion method, and large model-based completion method, and output the completed entity information.

[0220] Specifically, the entity information completion module 704 first filters out null fields in the entity matching results, and unifies heterogeneous null fields into standard fields based on the synonym field mapping table; then it performs layered processing according to the logic of rule completion, similarity completion, and large model verification, and finally outputs the entity information after null filling and logical consistency, and records the basis for completion.

[0221] The fusion module 705 is used to fuse the unique attribute value with the completed entity information and output the initial unified entity information.

[0222] Specifically, the fusion module 705 performs fusion according to priority integration logic, retaining fields without conflicts in the completed entity information; for fields that still have attribute conflicts after completion, it replaces them with unique attribute values; at the same time, it integrates all attributes according to the preset entity information framework to form initial unified entity information containing complete entity attributes, ensuring that the field structure is standardized and the attribute values ​​are consistent, providing a complete quality inspection object for the subsequent verification module.

[0223] The verification module 706 is used to verify the initial unified entity information. If the verification passes, the final entity information is output. If the verification fails, the parameters of the corresponding completion method in the entity information completion module are adjusted and processed again to output the final entity information. If the parameter correction fails, manual verification is triggered.

[0224] Specifically, the module performs three types of checks in sequence: format check checks the compliance of field format through regular expression matching, logic check verifies consistency through related field calculations, and integrity check confirms that core fields have no empty values. If all three checks pass, the final entity information is directly output. If any check fails, the root cause of the problem is located, the corresponding parameters of the entity information completion module are adjusted, the completion process is retried, and the check is performed again. If the parameter adjustment still cannot solve the problem, manual check is triggered, and the abnormal fields and the judgment criteria for the entire process are pushed to the domain experts to supplement information or confirm the abnormality.

[0225] In this embodiment, the specific working process and working principle of the entity data acquisition module 701, entity matching result output module 702, unique attribute value output module 703, entity information completion module 704, fusion module 705, and verification module 706 are the same as those in Embodiment 1, and therefore will not be described again in this embodiment. Each unit module can be implemented entirely or partially through software, hardware, or a combination thereof. Each unit module can be embedded in or independent of the processor in the computer device in hardware form, or it can be stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to the above unit modules.

[0226] Example 3

[0227] like Figure 9 The diagram illustrates a terminal device disclosed in this embodiment, comprising a transmitter, a receiver, a memory, and a processor. The transmitter transmits instructions and data, the receiver receives instructions and data, the memory stores computer-executed instructions, and the processor executes the computer-executed instructions stored in the memory to implement the method described in Embodiment 1 above.

[0228] It is important to note that the aforementioned memory can be either standalone or integrated with the processor. When the memory is set up independently, the terminal device also includes a bus for connecting the memory and the processor.

[0229] Example 4

[0230] This embodiment discloses a computer-readable storage medium storing computer-executable instructions. When a processor executes the computer-executable instructions, it implements the method in Embodiment 1 above.

[0231] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0232] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0233] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A method for cross-domain entity identity matching and information fusion, characterized in that, The method comprises: Step 1, obtaining cross-domain entity data, preprocessing the cross-domain entity data, and outputting a standardized entity data set; Step 2, adopting a multi-method complementary matching strategy, performing entity matching on the standardized entity data set by at least two types of methods selected from the group consisting of rule-based entity matching, similarity-based entity matching, machine learning-based entity matching, and graph-based entity matching, and outputting an entity matching result, wherein the entity matching result comprises multi-source attribute records of the same entity; Step 3, adopting a conflict resolution strategy, filtering the entity matching result based on the multi-source attribute records of the same entity by at least two types of methods selected from the group consisting of a timestamp-based method, a confidence weight-based method, a voting-based method, and a model-assisted judgment method, and outputting a unique attribute value; Step 4, adopting an information completion strategy, performing information completion on the entity matching result by at least two types of methods selected from the group consisting of a rule-based completion method, a similarity-based completion method, and a large model-based completion method, and outputting completed entity information; Step 5, fusing the unique attribute value and the completed entity information, and outputting initial unified entity information; Step 6, verifying the initial unified entity information, and if the verification is passed, outputting final entity information; if the verification is not passed, adjusting parameters of the corresponding completion method in step 4 and processing again, and outputting final entity information; if the parameter correction cannot be passed, triggering manual verification.

2. The cross-domain entity identity matching and information fusion method of claim 1, wherein, In step 2, the multi-method complementary matching strategy is a hierarchical progressive strategy, comprising: Step 201, filtering entity pairs in the standardized entity data set according to preset deterministic matching rules and deterministic exclusion rules by the rule-based entity matching, and outputting a first matching result and a first to-be-determined result; Step 202, filtering entity pairs in the first to-be-determined result according to a calculation attribute similarity method by the similarity-based entity matching, and outputting a second matching result and a second to-be-determined result; Step 203, filtering entity pairs in the second to-be-determined result by model learning features and result mapping by the machine learning-based entity matching, and outputting a third matching result and a third to-be-determined result; Step 204, filtering entity pairs in the third to-be-determined result by constructing a local association subgraph in a manner of abstracting entities and association relationships into nodes and edges by the graph-based entity matching, and outputting a fourth matching result; Step 205, integrating the first matching result, the second matching result, the third matching result, and the fourth matching result, and outputting a final entity matching result.

3. The cross-domain entity resolution and information fusion method of claim 1, wherein, In step 2, the multi-method complementary matching strategy is a parallel calling strategy, comprising: Step 211, filtering entity pairs in the standardized entity data set according to preset deterministic matching rules and deterministic exclusion rules by the rule-based entity matching, and outputting a first matching result and a first to-be-determined result; Step 212, the first to be determined result is processed in parallel by the similarity-based entity matching, the machine learning-based entity matching, and the graph-based entity matching, and total similarity scores, matching confidence scores, and graph feature scores are respectively outputted; Step 213, the total similarity scores, the matching confidence scores, and the graph feature scores are corrected by cross-validation to obtain corrected total similarity scores, matching confidence scores, and graph feature scores; Step 214, a weighted comprehensive score of the corrected total similarity scores, the matching confidence scores, and the graph feature scores is calculated; if the weighted comprehensive score is greater than or equal to a preset high-precision threshold, it is determined as a new matching result; if the weighted comprehensive score is less than the preset high-precision threshold, it is determined as a non-matching result; Step 215, the first matching result and the new matching result are integrated to output a final entity matching result.

4. The cross-domain entity identity matching and information fusion method according to claim 2 or 3, c h a r a c t e r i z e d by, In step 2, the multi-method complementary matching strategy is replaced by a hierarchical feature extraction strategy, which includes: Step 221, the standardized entity data set is subjected to basic attribute feature extraction, semantic association feature extraction, and relationship network feature extraction respectively to output basic attribute feature vectors, semantic association feature vectors, and relationship network feature vectors; Step 222, the basic attribute feature vectors, the semantic association feature vectors, and the relationship network feature vectors are spliced to output an entity feature vector; Step 223, a matching prediction model is constructed, the entity feature vector is inputted into the matching prediction model for processing, and a final entity matching result is outputted.

5. The cross-domain entity resolution and information fusion method of claim 1, wherein, In step 3, the conflict resolution strategy is a serial fusion strategy, which includes: Step 301, same entity multi-source attribute records of the entity matching result are acquired, including conflict version ID, attribute value, source domain name, timestamp, confidence weight of the source domain, historical accuracy rate of each source domain, and attribute type; Step 302, a time stamp-based method is adopted, based on the attribute type, the latest attribute value is filtered through time-sensitive attributes, and the source domain corresponding to the latest attribute value is taken as a first candidate source; if it is a static attribute, all source domains corresponding to the attribute value are taken as a second candidate source; the first candidate source and the second candidate source are integrated to output a candidate source set, and then step 303 is jumped to; Step 303, a confidence weight-based method is adopted to acquire confidence weights corresponding to each candidate source in the candidate source set; it is judged whether the confidence weight is greater than a high-confidence threshold; if yes, the candidate source attribute value corresponding to the confidence weight is taken as the only attribute value and outputted; if no, the candidate source with the highest confidence weight is recorded as a core candidate source, and step 304 is jumped to; Step 304, a voting-based method is used to count all value votes of the current attribute value in the same entity multi-source attribute record, filter the value with the highest number of votes as the optimal value, and record the voting association domain corresponding to the optimal value; determine whether the core candidate source is in the voting association domain; if yes, directly output the attribute value corresponding to the core candidate source as the unique attribute value; if not, select the candidate source with the highest confidence weight from the voting association domain as the optimal voting domain, and jump to step 305; Step 305, an auxiliary decision model is constructed based on a model-assisted decision method, the timestamp freshness of each candidate source in the candidate source set, the confidence weight, and the voting support degree corresponding to the optimal voting domain are input into the auxiliary decision model for processing, and the optimal probability of each candidate source is output; determine whether the optimal probability is greater than the auxiliary decision threshold; if yes, output the attribute value corresponding to the candidate source with the highest optimal probability as the unique attribute value; if not, jump to step 306; Step 306, manually count the number of times each candidate source appears in the candidate source set in step 302, the core candidate source in step 303, the optimal voting domain in step 304, and the candidate source with the highest optimal probability in step 305, and output the attribute value corresponding to the candidate source with the most appearances as the unique attribute value. If the number of times each candidate source appears is the same, output the attribute value corresponding to the candidate source with the highest confidence weight as the unique attribute value.

6. The cross-domain entity resolution and information fusion method of claim 1, wherein, In step 3, the conflict resolution strategy is a parallel fusion strategy, which includes: Step 311, obtaining the same entity multi-source attribute record of the entity matching result, including conflict version ID, attribute value, source domain name, timestamp, confidence weight of source domain, historical accuracy rate of each source domain, and attribute type; Step 312, the same entity multi-source attribute record is processed in parallel by the timestamp-based method, the confidence weight-based method, the voting-based method, and the model-assisted decision method, respectively, to output each candidate attribute value and the corresponding confidence score; Step 313, assign a preset weight to the timestamp-based method, the confidence weight-based method, the voting-based method, and the model-assisted decision method; according to the assigned weight, weight and sum the confidence scores of the same attribute value to obtain the comprehensive score of each candidate attribute value; Step 314, determine whether the comprehensive score of each candidate attribute value is greater than the comprehensive decision threshold; if yes, output the attribute value corresponding to the candidate source with the highest comprehensive score as the unique attribute value; if not, directly enter step 315; Step 315, push all candidate values and decision basis to manual audit, and output the result confirmed by manual as the unique attribute value.

7. The cross-domain entity disambiguation and information fusion method of claim 6, wherein, In step 313, the comprehensive score of each candidate attribute value is calculated, and the calculation expression is: ； wherein, denotes the combined score of the th candidate attribute value; denotes the weight assigned based on the model-assisted decision method; denotes the weight assigned based on the confidence weight method; denotes the weight assigned based on the voting method; denotes the weight assigned based on the timestamp method; denotes the confidence score corresponding to the th model candidate attribute value; denotes the confidence score corresponding to the th confidence candidate attribute value; denotes the confidence score corresponding to the th voting candidate attribute value; denotes the confidence score corresponding to the th timestamp candidate attribute value.

8. The cross-domain entity resolution and information fusion method of claim 1, wherein, In step 4, the information completion strategy is a hierarchical completion strategy, which includes: Step 401, obtaining the same entity multi-source attribute record, filtering the null field as the object to be completed; based on the synonym field mapping table, unify the heterogeneous fields in the object to be completed into standard fields; In step 402, for the null fields with explicit logical derivation in the standard field, the rule-based completion method is used for classified completion, the first completed entity record is output, and the text type null field which is not completed is reserved, and step 403 is jumped to; In step 403, for the text type null field which is not completed in step 402, the similarity-based semantic completion method is used for completion, and the second completed entity record is output; In step 404, the entity completion model is constructed, the first completed entity record, the second completed entity record and the original non-null field of the entity are input into the entity completion model for processing, the logical consistency of the completed entity record and the non-null field is verified, and the completed entity information is output.

9. The cross-domain entity disambiguation and information fusion method of any one of claims 5 to 8, wherein, In step 6, the initial unified entity information is verified, including: In step 601, format verification is performed, and whether the field values in the initial unified entity information meet the preset standard format is checked; In step 602, logical verification is performed, and the consistency of the associated fields in the initial unified entity information is verified; In step 603, integrity verification is performed, and it is confirmed that there is no missing in the core field of the initial unified entity information; In step 604, if steps 601 to 603 are all passed, the final entity information is output; if any verification fails, the parameters of the corresponding completion method in step 4 are adjusted, and the processing is performed again, and the final entity information is output; if the parameter correction cannot be passed, manual verification is triggered, and the final entity information is output by manual determination.

10. A cross-domain entity identity matching and information fusion system, comprising: The system comprises: An entity data acquisition module is configured to acquire cross-domain entity data, pre-process the cross-domain entity data, and output a standardized entity data set; An entity matching result output module is configured to use a multi-method complementary matching strategy to perform entity matching on the standardized entity data set by at least two types of methods selected from the group consisting of rule-based entity matching, similarity-based entity matching, machine learning-based entity matching, and graph-based entity matching, and output an entity matching result, wherein the entity matching result comprises multiple-source attribute records of the same entity; A unique attribute value output module is configured to use a conflict resolution strategy to filter the entity matching result based on the multiple-source attribute records of the same entity by at least two types of methods selected from the group consisting of a timestamp-based method, a confidence weight-based method, a voting-based method, and a model-assisted determination method, and output a unique attribute value; An entity information completion module is configured to use an information completion strategy to perform information completion on the entity matching result by at least two types of methods selected from the group consisting of a rule-based completion method, a similarity-based completion method, and a large model-based completion method, and output completed entity information; A fusion module is configured to fuse the unique attribute value and the completed entity information, and output initial unified entity information; A verification module is configured to verify the initial unified entity information, output final entity information if the verification is passed, adjust the parameters of the corresponding completion method in the entity information completion module and perform processing again if the verification fails, and output final entity information if the parameter correction cannot be passed, and trigger manual verification.

Images