A knowledge graph-based cross-platform business system integration method

Abstract

This invention discloses a cross-platform business system integration method based on knowledge graphs, comprising the following steps: S1, reading data from the business platform database and generating field text, entity identifiers, and relation records; S2, inputting the field text into an ALBERT model, generating field semantic vectors through a domain word vector gating layer; S3, forming triples based on entity identifiers and relation records and writing them into the knowledge graph; S4, inputting the triples into a TransH model, generating entity structure embedding vectors through dynamic rotation of relation vectors and hyperplane normal vectors; S5, generating entity embedding vectors based on field semantic vectors and entity structure embedding vectors; S6, calculating various distances based on entity embedding vectors to obtain the similarity of cross-platform candidate entity pairs; S7, performing duplicate entity determination on the similarity and updating the knowledge graph. This invention achieves accurate alignment of cross-platform business entities, improves data fusion efficiency and consistency, and is suitable for multi-source business integration scenarios.

Description

Technical Field

[0001] This invention relates to the field of cross-platform business data integration technology, and in particular to a cross-platform business system integration method based on knowledge graphs. Background Technology

[0002] Current cross-platform business data integration methods generally rely on rule configuration, field name comparison, or manual mapping. When there are significant differences in field naming, inconsistent field descriptions, or ambiguous semantic expressions between different business platforms, traditional methods struggle to achieve accurate alignment. Entities on different platforms often have multiple ways of being represented, such as abbreviations, aliases, or non-uniform naming. String-based matching methods cannot effectively identify synonymous entities, leading to a large amount of manual confirmation work, resulting in low integration efficiency and a high risk of errors.

[0003] Existing knowledge graph construction methods are mostly based on static field parsing and inherent relational structures, lacking deep extraction of the semantics of field texts, resulting in insufficient semantic dimension in the generated entity representations. Existing embedding models typically use fixed relational constraints during relation learning, failing to consider the dynamic changes of entities in different business relational environments, making it difficult for entity structure embeddings to accurately reflect business semantics. Cross-platform entity alignment often relies on a single distance formula to evaluate similarity, failing to measure the proximity between entities from multiple perspectives, and easily leading to unstable similarity or misjudgment issues.

[0004] Existing methods for identifying duplicate entities generally lack multi-level screening mechanisms. They cannot make comprehensive judgments based on different distance results, nor do they have clear quantitative standards for duplicate entities, making cross-platform entity alignment lack a reliable basis.

[0005] Therefore, how to provide a cross-platform business system integration method based on knowledge graphs is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a cross-platform business system integration method based on knowledge graphs. This invention combines the ALBERT model and the TransH model to uniformly model the business fields and entity relationships of multiple platforms, realizes automatic matching of cross-platform entities and determination of duplicate entities, and has the advantages of accurate semantic understanding, high entity alignment accuracy and high degree of automation of data integration.

[0007] According to an embodiment of the present invention, a cross-platform business system integration method based on knowledge graphs includes the following steps:

[0008] S1. Retrieve field names, field descriptions, entity identifiers, and relationship records from the business platform database. Perform preprocessing on the retrieved data to generate field text, entity identifiers, and relationship records.

[0009] S2. Construct the ALBERT model, load the domain word vector library, input the field text into the ALBERT language model one by one, and generate field semantic vectors through the domain word vector gating layer;

[0010] S3. Generate entity node data and relation edge data based on entity identifiers and relation records, organize them into triples, and write them into the knowledge graph;

[0011] S4. Construct the TransH model, set the relation vector and the hyperplane normal vector, generate the rotation reference vector based on the head entity vector and the tail entity vector when each triplet is input, perform rotation update on the hyperplane normal vector, and iteratively train to obtain the entity structure embedding vector.

[0012] S5. Based on the correspondence between entity identifiers and field texts, concatenate, linearly transform, and normalize the field semantic vector and the entity structure embedding vector to generate the entity embedding vector.

[0013] S6. Construct cross-platform candidate entity pairs from entity embedding vectors, calculate the distance for each pair of entity embedding vectors according to different distance formulas, and generate multi-distance similarity results.

[0014] S7. Based on the multi-distance similarity results, perform duplicate entity determination on the candidate entity pairs. If there are duplicate entities, write them into the duplicate entity list. If there are no duplicate entities, do not perform cross-platform entity alignment processing.

[0015] Optionally, S2 specifically includes:

[0016] S21. Using the field text as input, construct a domain word vector library. The domain word vector library performs word segmentation, stop word removal, word frequency statistics, and co-occurrence statistics on historical business field text, industry terminology text, and entity name text. Then, it trains a word vector matrix based on the co-occurrence window and stores it in a fixed manner.

[0017] S22. Input the field text into the input embedding layer of the ALBERT model. The input embedding layer performs word vector mapping on each token in the field text to generate basic word vectors.

[0018] S23. Input the basic word vector and the corresponding domain word vector in the domain word vector library into the domain word vector gating layer. The domain word vector gating layer is equipped with a gating coefficient generation unit. The gating coefficient generation unit takes the connection vector between the basic word vector and the domain word vector as input and performs linear transformation and normalization operations on the connection vector to generate variable gating coefficients.

[0019] S24. The domain word vector gating layer performs a weighted operation on the basic word vector and the domain word vector based on the variable gating coefficient to generate gated word vectors.

[0020] S25. The gated word vectors are input into the factorized embedding layer of the ALBERT model and then into the subsequent shared layer structure of the ALBERT model to perform forward propagation. The representation vectors output by the model are transformed into field semantic vectors.

[0021] Optionally, S3 specifically includes:

[0022] S31. Generate entity node data for each entity identifier based on the entity identifier, and perform uniqueness processing on the entity node data;

[0023] S32. Based on the relation type, source entity identifier and target entity identifier stored in the relation record, generate relation edge data, take the entity corresponding to the source entity identifier in the entity node data as the head entity, take the entity corresponding to the target entity identifier in the entity node data as the tail entity, take the relation type as the relation element, and combine the head entity, relation element and tail entity into a triple in a fixed order.

[0024] S33. Write the generated triples into the data storage structure of the knowledge graph.

[0025] Optionally, S4 specifically includes:

[0026] S41. Initialize the entity vector, relation vector, and hyperplane normal vector in the TransH model, set the learning rate, maximum number of iterations, and loss threshold; set the rotation coefficients as trainable scalar parameters and assign them initial values, set the linear transformation matrix and bias vector as trainable matrix parameters and trainable bias parameters, and add the entity vector, relation vector, hyperplane normal vector, rotation coefficients, linear transformation matrix, and bias vector to the trainable parameter set of the TransH model;

[0027] S42. Divide the triplet into multiple triplet batches, select the current triplet batch from the triplet batches, and map the head entity identifier and tail entity identifier in each triplet to the head entity vector and tail entity vector.

[0028] S43. Map the relation elements in each triplet to a relation vector and an initial hyperplane normal vector. Perform vector subtraction on the head entity vector and tail entity vector to generate a difference vector, and perform vector summation to generate a sum vector. Concatenate the difference vector and the sum vector in a fixed order to form a rotation input vector. Perform linear transformation and normalization operations on the rotation input vector to generate a rotation reference vector.

[0029] S44. Perform a rotation update on the initial hyperplane normal vector based on the rotation reference vector to obtain the rotated and updated hyperplane normal vector;

[0030] S45. Based on the rotated updated hyperplane normal vector and relation vector, perform positive and negative sample scoring on the current triplet batch, construct the loss function and perform gradient update;

[0031] S46. When the loss function is lower than the loss threshold or the number of iterations reaches the maximum number of iterations, the convergence condition is determined and the entity structure embedding vector is output; if the convergence condition is not met, return to S42 for the next round of iteration training.

[0032] Optionally, S44 specifically includes:

[0033] S442. Perform vector subtraction on the rotated reference vector and the hyperplane normal vector to generate a direction difference vector, and perform normalization on the direction difference vector to obtain a unit direction difference vector.

[0034] S443. Perform scalar multiplication on the rotation coefficient and the unit vector of the direction difference to generate the rotation update. The scalar multiplication operation is to multiply the values ​​of each dimension of the unit vector of the direction difference by the rotation coefficient.

[0035] S444. Perform vector addition on the rotation update amount and the hyperplane normal vector to generate the updated normal vector, and perform normalization operation on the updated normal vector to obtain the rotated and updated hyperplane normal vector.

[0036] Optionally, S5 specifically includes:

[0037] S51. Based on the entity structure embedding vector, match each entity structure embedding vector with the field semantic vector according to the entity identifier.

[0038] S52. Perform vector concatenation on each matched entity structure embedding vector and field semantic vector, perform linear transformation and normalization to generate entity embedding vector.

[0039] Optionally, S6 specifically includes:

[0040] S61. Based on the entity source information recorded in the entity section, group the entity embedding vectors according to the entity source information to obtain a set of entity embedding vectors from different business platforms.

[0041] S62. Select target entity embedding vectors sequentially from the first business platform entity embedding vector set. While selecting target entity embedding vectors, select comparison entity embedding vectors corresponding to target entity embedding vectors from the second business platform entity embedding vector set. Combine target entity embedding vectors and comparison entity embedding vectors to form candidate entity pairs. Repeat the selection operation to generate a candidate entity pair set.

[0042] S63. For each pair of entity embedding vectors in the candidate entity pair set, perform the Euclidean distance algorithm to calculate the Euclidean distance value, the Manhattan distance algorithm to calculate the Manhattan distance value, and the cosine similarity algorithm to calculate the cosine similarity value. Record the Euclidean distance value, Manhattan distance value, and cosine similarity value in a fixed order as the multi-distance similarity result.

[0043] Optionally, S7 specifically includes:

[0044] S71. Sort the Euclidean distance and Manhattan distance values ​​of each candidate entity of the first business platform in ascending order, and sort the cosine similarity values ​​in descending order. Select the top five candidate entity pairs from the sorting results to generate the top five sets of Euclidean distance, Manhattan distance, and cosine similarity.

[0045] S72. Perform set cross-search on the top five sets of Euclidean distance, the top five sets of Manhattan distance, and the top five sets of cosine similarity. When a candidate entity pair appears in the top five sets of Euclidean distance, the top five sets of Manhattan distance, and the top five sets of cosine similarity, the candidate entity pair is judged as a duplicate entity.

[0046] S73. When there are no candidate entity pairs that appear simultaneously in the top five sets of Euclidean distance, the top five sets of Manhattan distance, and the top five sets of cosine similarity, perform average calculation of Euclidean distance, Manhattan distance, and cosine similarity values ​​for candidate entity pairs that appear simultaneously in any two of the top five sets of Euclidean distance, the top five sets of Manhattan distance, and the top five sets of cosine similarity to generate an average similarity value. Compare the average similarity value with a preset lenient threshold. If the lenient threshold condition is met, the candidate entity pair is determined to be a duplicate entity.

[0047] S74. When no candidate entity pair appears in any two of the top five sets of Euclidean distance, top five sets of Manhattan distance, and top five sets of cosine similarity, the Euclidean distance value, Manhattan distance value, and cosine similarity value of the candidate entity pair in the three top five sets are compared with the preset strict threshold. When all strict threshold conditions are met, the candidate entity pair is determined to be a similar entity.

[0048] S75. If a candidate entity pair does not satisfy S72, S73, or S74, the current candidate entity pair set is determined to have no similar entities, and cross-platform entity alignment processing is not performed.

[0049] The beneficial effects of this invention are:

[0050] (1) By introducing domain word vector library and ALBERT semantic enhancement mechanism into field text processing, a deep understanding of the semantics of business fields is realized, which effectively improves the accuracy of entity semantic representation and solves the problem of inconsistent field naming and large differences in semantic expression that lead to difficulty in alignment.

[0051] (2) By introducing a dynamic rotation update mechanism for the rotation reference vector and the hyperplane normal vector in the TransH relation embedding model, the entity structure embedding can be adaptively adjusted in combination with the business relationship environment, which improves the distinguishability of entity structure representation and makes up for the problem of the single structure representation and lack of dynamism in the traditional embedding model.

[0052] (3) By constructing a duplicate entity determination mechanism based on multi-distance sorting, multi-level screening of candidate entity pairs from multiple similarity indicators is realized, which effectively reduces the risk of misjudgment by a single distance, improves the accuracy of cross-platform entity matching and duplicate entity recognition, and enhances the intelligence and reliability of cross-platform business system integration. Attached Figure Description

[0053] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0054] Figure 1 This is a flowchart of a cross-platform business system integration method based on knowledge graphs proposed in this invention;

[0055] Figure 2 This is a flowchart of the ALBERT model training process for a cross-platform business system integration method based on knowledge graphs proposed in this invention.

[0056] Figure 3 This is a TransH training flowchart for a cross-platform business system integration method based on knowledge graphs proposed in this invention. Detailed Implementation

[0057] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0058] refer to Figure 1-3 A method for integrating cross-platform business systems based on knowledge graphs includes the following steps:

[0059] S1. Retrieve field names, field descriptions, entity identifiers, and relationship records from the business platform database. Perform preprocessing on the retrieved data to generate field text, entity identifiers, and relationship records.

[0060] S2. Construct the ALBERT model, load the domain word vector library, input the field text into the ALBERT language model one by one, and generate field semantic vectors through the domain word vector gating layer;

[0061] S3. Generate entity node data and relation edge data based on entity identifiers and relation records, organize them into triples, and write them into the knowledge graph;

[0062] S4. Construct the TransH model, set the relation vector and the hyperplane normal vector, generate the rotation reference vector based on the head entity vector and the tail entity vector when each triplet is input, perform rotation update on the hyperplane normal vector, and iteratively train to obtain the entity structure embedding vector.

[0063] S5. Based on the correspondence between entity identifiers and field texts, concatenate, linearly transform, and normalize the field semantic vector and the entity structure embedding vector to generate the entity embedding vector.

[0064] S6. Construct cross-platform candidate entity pairs from entity embedding vectors, calculate the distance for each pair of entity embedding vectors according to different distance formulas, and generate multi-distance similarity results.

[0065] S7. Based on the multi-distance similarity results, perform duplicate entity determination on the candidate entity pairs. If there are duplicate entities, write them into the duplicate entity list. If there are no duplicate entities, do not perform cross-platform entity alignment processing.

[0066] In this embodiment, S1 specifically includes:

[0067] S11. The business platform database provides field names, field descriptions, entity identifiers, and relationship records. It traverses the database table structure according to the preset field order, reads the field names and field descriptions in text form, reads the entity identifiers in unique number form, and reads and caches the relationship records in the form of a combination of relationship type, source entity number, and target entity number.

[0068] S12. Unify the character encoding format and remove redundant spaces, newlines and abnormal symbols from the original text; perform short text processing on the field descriptions, divide the long text field descriptions into processable units according to sentence boundaries, and combine the field names and the segmented field descriptions to form the field text.

[0069] S13. Input the field text into the word segmentation unit to perform word segmentation, input the segmentation result into the stop word filtering unit to delete noisy words, and write the filtered result as the final form of the field text into the field text cache; when processing entity identifiers, perform a uniqueness check on all numbers, remove duplicate numbers from the result set and retain unique numbers; when processing relation records, perform format validation on each record, remove records that do not conform to the format from the result set, and store records that conform to the format as relation records in the order of relation type, source entity number and target entity number.

[0070] In this embodiment, S2 specifically includes:

[0071] S21. Using the field text as input, construct a domain word vector library. The domain word vector library performs word segmentation, stop word removal, word frequency statistics, and co-occurrence statistics on historical business field text, industry terminology text, and entity name text. Then, it trains a word vector matrix based on the co-occurrence window and stores it in a fixed manner.

[0072] S22. Input the field text into the input embedding layer of the ALBERT model. The input embedding layer performs word vector mapping on each token in the field text to generate basic word vectors.

[0073] S23. Input the basic word vectors and the corresponding labeled domain word vectors from the domain word vector library into the domain word vector gating layer. The domain word vector gating layer is equipped with a gating coefficient generation unit. The gating coefficient generation unit takes the connection vector between the basic word vectors and the domain word vectors as input, performs linear transformation and normalization operations on the connection vector to generate variable gating coefficients. "Label" represents the smallest text processing unit generated by the word segmentation algorithm. In Chinese scenarios, it is usually the smallest trainable unit in the form of a single character or a word, used to represent the discretized structure of the field text before entering the model.

[0074] S24. The domain word vector gating layer performs a weighted operation on the basic word vector and the domain word vector based on the variable gating coefficient to generate gated word vectors.

[0075] S25. The gated word vectors are input into the factorized embedding layer of the ALBERT model and then into the subsequent shared layer structure of the ALBERT model to perform forward propagation. The representation vectors output by the model are transformed into field semantic vectors.

[0076] In this embodiment, S21 specifically includes:

[0077] S211. The word segmentation algorithm adopts a Chinese word segmentation algorithm based on statistical features. It performs maximum probability segmentation on the text character sequence through character frequency statistics and mutual information calculation to generate a labeled sequence. The labeled sequence serves as the basic input for training domain word vectors.

[0078] S212, the word vector training method adopts the Skip-gram model, which constructs a set of context tags for each center tag within a context range with a window size of 5, trains the model to predict context tags for the center tags, and generates a domain word vector matrix;

[0079] S213. The training parameters are set to fixed values, the number of training iterations is set to 30 rounds, the learning rate is set to 0.01, the embedding dimension is set to 128, which is consistent with the word vector dimension of the ALBERT input. The word vector matrix is ​​constructed according to the embedding dimension, and each token is mapped to a dense vector of length 128. The training process uses stochastic gradient descent to update the parameters.

[0080] S214. After training, the generated word vector matrix is ​​stored in the domain word vector library.

[0081] In this embodiment, S23 specifically includes:

[0082] S231. The basic word vectors and the domain word vectors are concatenated in dimensional order, and the linear transformation unit performs an affine transformation on the concatenated vectors.

[0083] S232. Perform element-wise numerical compression on the vector after affine transformation, and use the element-wise Sigmoid function to generate variable gating coefficients.

[0084] S233. Variable gating coefficients are used to control the weighting ratio of base word vectors to domain word vectors.

[0085] In this embodiment, S3 specifically includes:

[0086] S31. Generate entity node data for each entity identifier based on the entity identifier, and perform uniqueness processing on the entity node data;

[0087] S32. Based on the relation type, source entity identifier and target entity identifier stored in the relation record, generate relation edge data, take the entity corresponding to the source entity identifier in the entity node data as the head entity, take the entity corresponding to the target entity identifier in the entity node data as the tail entity, take the relation type as the relation element, and combine the head entity, relation element and tail entity into a triple in a fixed order.

[0088] S33. Write the generated triples into the data storage structure of the knowledge graph.

[0089] In this embodiment, S4 specifically includes:

[0090] S41. Initialize the entity vector, relation vector, and hyperplane normal vector in the TransH model, set the learning rate, maximum number of iterations, and loss threshold; set the rotation coefficients as trainable scalar parameters and assign them initial values, set the linear transformation matrix and bias vector as trainable matrix parameters and trainable bias parameters, and add the entity vector, relation vector, hyperplane normal vector, rotation coefficients, linear transformation matrix, and bias vector to the trainable parameter set of the TransH model;

[0091] S42. Divide the triplet into multiple triplet batches, select the current triplet batch from the triplet batches, and map the head entity identifier and tail entity identifier in each triplet to the head entity vector and tail entity vector.

[0092] S43. Map the relation elements in each triplet to a relation vector and an initial hyperplane normal vector. Perform vector subtraction on the head entity vector and tail entity vector to generate a difference vector, and perform vector summation to generate a sum vector. Concatenate the difference vector and the sum vector in a fixed order to form a rotation input vector. Perform linear transformation and normalization operations on the rotation input vector to generate a rotation reference vector.

[0093] S44. Perform a rotation update on the initial hyperplane normal vector based on the rotation reference vector to obtain the rotated and updated hyperplane normal vector;

[0094] S45. Based on the rotated updated hyperplane normal vector and relation vector, perform positive and negative sample scoring on the current triplet batch, construct the loss function and perform gradient update;

[0095] S46. When the loss function is lower than the loss threshold or the number of iterations reaches the maximum number of iterations, the convergence condition is determined and the entity structure embedding vector is output. If the convergence condition is not met, return to S42 for the next round of iteration training. In this embodiment, the loss threshold is 0.001 and the maximum number of iterations is 500.

[0096] In this embodiment, S41 specifically includes:

[0097] The entity vector initialization unit creates entity vectors one by one according to the entity identifier list, with the entity vector dimension set to a fixed length. The initialization method adopts uniform distribution random initialization, generating values ​​within the range of each dimension of the vector to form the initial state of the entity vector. The relation vector initialization unit creates relation vectors one by one according to the relation element list, with the relation vector dimension consistent with the entity vector. The initialization method is consistent with the entity vector, generating values ​​according to the same random distribution rule. The hyperplane normal vector initialization unit generates a corresponding initial normal vector for each relation vector, with the initial normal vector dimension consistent with the relation vector. The normal vector initialization method adopts random sampling within a fixed numerical range to generate the original normal vector, and performs normalization processing on the original normal vector to maintain the normal vector length at unit length.

[0098] In this embodiment, S44 specifically includes:

[0099] S442. Perform vector subtraction on the rotated reference vector and the hyperplane normal vector to generate a direction difference vector, and perform normalization on the direction difference vector to obtain a unit direction difference vector.

[0100] S443. Perform scalar multiplication on the rotation coefficient and the unit vector of the direction difference to generate the rotation update. The scalar multiplication operation is to multiply the values ​​of each dimension of the unit vector of the direction difference by the rotation coefficient.

[0101] S444. Perform vector addition on the rotation update amount and the hyperplane normal vector to generate the updated normal vector, and perform normalization operation on the updated normal vector to obtain the rotated and updated hyperplane normal vector.

[0102] In this embodiment, S45 specifically includes:

[0103] S451. The updated head entity vector and tail entity vector are fed into the TransH projection operator, which projects according to the standard orthogonal projection algorithm.

[0104] S452. The distance calculation uses the L2 norm distance formula to calculate the score;

[0105] S453. The loss calculation uses the maximum boundary loss formula based on the Hinge loss function, and the loss value is passed into the backpropagation link; the gradient calculation process uses the chain rule to generate gradients for entity vectors, relation vectors and rotation coefficients; the gradient update method uses the stochastic gradient descent algorithm to perform the update.

[0106] In this embodiment, S5 specifically includes:

[0107] S51. Based on the entity structure embedding vector, match each entity structure embedding vector with the field semantic vector according to the entity identifier.

[0108] S52. Perform vector concatenation on each matched entity structure embedding vector and field semantic vector, perform linear transformation and normalization to generate entity embedding vector.

[0109] In this embodiment, S6 specifically includes:

[0110] S61. Based on the entity source information recorded in the entity section, group the entity embedding vectors according to the entity source information to obtain a set of entity embedding vectors from different business platforms.

[0111] S62. Select target entity embedding vectors sequentially from the first business platform entity embedding vector set. While selecting target entity embedding vectors, select comparison entity embedding vectors corresponding to target entity embedding vectors from the second business platform entity embedding vector set. Combine target entity embedding vectors and comparison entity embedding vectors to form candidate entity pairs. Repeat the selection operation to generate a candidate entity pair set.

[0112] S63. For each pair of entity embedding vectors in the candidate entity pair set, perform the Euclidean distance algorithm to calculate the Euclidean distance value, the Manhattan distance algorithm to calculate the Manhattan distance value, and the cosine similarity algorithm to calculate the cosine similarity value. Record the Euclidean distance value, Manhattan distance value, and cosine similarity value in a fixed order as the multi-distance similarity result.

[0113] In this embodiment, S7 specifically includes:

[0114] S71. Sort the Euclidean distance and Manhattan distance values ​​of each candidate entity of the first business platform in ascending order, and sort the cosine similarity values ​​in descending order. Select the top five candidate entity pairs from the sorting results to generate the top five sets of Euclidean distance, Manhattan distance, and cosine similarity.

[0115] S72. Perform set cross-search on the top five sets of Euclidean distance, the top five sets of Manhattan distance, and the top five sets of cosine similarity. When a candidate entity pair appears in the top five sets of Euclidean distance, the top five sets of Manhattan distance, and the top five sets of cosine similarity, the candidate entity pair is judged as a duplicate entity.

[0116] S73. When there are no candidate entity pairs that appear simultaneously in the top five sets of Euclidean distance, Manhattan distance, and cosine similarity, the average of the Euclidean distance, Manhattan distance, and cosine similarity values ​​is calculated for any two of the candidate entity pairs that appear simultaneously in any two of the top five sets of Euclidean distance, Manhattan distance, and cosine similarity to generate an average similarity value. The average similarity value is compared with a preset lenient threshold. If the lenient threshold condition is met, the candidate entity pair is determined to be a duplicate entity. In this embodiment, the lenient threshold is 0.85.

[0117] S74. When no candidate entity pair appears in any two of the top five sets of Euclidean distance, top five sets of Manhattan distance, and top five sets of cosine similarity, the Euclidean distance, Manhattan distance, and cosine similarity values ​​of the candidate entity pairs in the three top five sets are compared with preset strict thresholds. If all strict threshold conditions are met, the candidate entity pair is determined to be a similar entity. In this embodiment, the strict threshold is 0.85, and the value of all distance formulas is greater than the strict threshold to meet the conditions.

[0118] S75. If a candidate entity pair does not satisfy S72, S73, or S74, the current candidate entity pair set is determined to have no similar entities, and cross-platform entity alignment processing is not performed.

[0119] Example 1:

[0120] To verify the feasibility of this invention in practice, it was applied to a cross-platform information fusion scenario involving multiple independent business platforms with significantly different data structures. Field naming conventions, field descriptions, entity identifier structures, and relational record formats are all inconsistent, and there is a lack of unified entities across platforms. When business processes are transferred between multiple systems, problems arise such as duplicate data entry, entity redundancy, entity inconsistencies, and reliance on manual maintenance for mapping. Traditional methods often rely on manual mapping or string similarity matching based on fixed rules, resulting in slow processing speeds, low accuracy, and difficulty maintaining stable performance in database environments with chaotic structures and non-standard descriptions.

[0121] To address these issues, the cross-platform business system integration method of this invention is applied to a multi-source business database environment. Multiple business platforms contain entities such as user information, order information, warehousing information, logistics information, and product information, but their field descriptions differ. For example, a user account might be "login_id" on one platform and "usr_code" on another; a product SKU might be "product_SKU" on platform A and "sku_id" on platform B; and warehouse codes might use English abbreviations, numerical codes, and mixed formats. This invention utilizes a domain-specific word vector library to enhance the semantic expressiveness of field text, employs TransH structure embedding to reduce cross-platform structural offsets, and identifies duplicate entities through a multi-indicator fusion mechanism based on distance calculation, achieving high-precision cross-platform alignment.

[0122] In this embodiment, cross-platform matching experiments were conducted on six types of entities, with a total of 3821 field texts from all platforms. The domain word vector library was established using methods such as word segmentation, word frequency statistics, and co-occurrence statistics, achieving a stable semantic distribution after embedding training. The ALBERT model semantically encoded the field texts to output field semantic vectors, while the TransH model generated entity structure embedding vectors. The two vectors were concatenated and normalized to form a unified entity embedding space. During the entity alignment stage, multiple ranking criteria were applied based on Euclidean distance, Manhattan distance, and cosine similarity, and entities with higher rankings were selected for duplicate entity detection.

[0123] In the experiment, manual annotation was selected as the actual alignment benchmark. This invention significantly outperforms traditional rule-based matching methods across multiple entity types. Taking user account entities as an example, traditional methods, limited by fixed rules and highly ambiguous fields, achieve an accuracy of less than 80%; this invention, through semantic enhancement, structural embedding, and multi-distance fusion, achieves an accuracy of over 95%. For product SKU entities, traditional methods struggle to identify differences in capitalization, hyphens, and naming formats; this invention, through the fusion of semantic and structural representations, significantly improves accuracy. Duplicate entity recognition is particularly crucial in multi-source platforms; this invention achieves a substantial increase in the number of duplicate entities through stable fusion of three distances, reducing missed identifications.

[0124] To verify the feasibility of this invention in practice, it was applied to a business database environment with mixed multi-source field descriptions, and cross-platform alignment analysis was performed on six types of business entities. Experimental data are shown in Table 1 below. This invention improves accuracy and increases the number of duplicate entities identified.

[0125] Table 1: Comparison of Entity Matching Results

[0126]

[0127] Table 1 shows the comparison results of six key business entities under traditional methods and the method of this invention. Traditional rule matching has low accuracy for entities with significant differences in field descriptions and inconsistent naming, ranging from approximately 66% to 78%. This invention, through the fusion of semantic enhancement and structural embedding, enables similar entities to have a centralized and distinguishable representation in the vector space, improving accuracy to between 91% and 96%. The number of duplicate entities identified is significantly improved in this invention's method, with an increase of over 30%, indicating that this invention can effectively reduce entity redundancy and improve cross-platform consistency during multi-platform data fusion. The accuracy improvement is most prominent for order numbers, user accounts, and product SKUs among the six entity types, demonstrating the stability and scalability of the semantic enhancement mechanism and TransH structural embedding in structurally chaotic database environments.

[0128] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for integrating cross-platform business systems based on knowledge graphs, characterized in that, Includes the following steps: S1. Retrieve field names, field descriptions, entity identifiers, and relationship records from the business platform database. Perform preprocessing on the retrieved data to generate field text, entity identifiers, and relationship records. S2. Construct the ALBERT model, load the domain word vector library, input the field text into the ALBERT language model one by one, and generate field semantic vectors through the domain word vector gating layer; S3. Generate entity node data and relation edge data based on entity identifiers and relation records, organize them into triples, and write them into the knowledge graph; S4. Construct the TransH model, set the relation vector and the hyperplane normal vector, generate the rotation reference vector based on the head entity vector and the tail entity vector when each triplet is input, perform rotation update on the hyperplane normal vector, and iteratively train to obtain the entity structure embedding vector. S5. Based on the correspondence between entity identifiers and field texts, concatenate, linearly transform, and normalize the field semantic vector and the entity structure embedding vector to generate the entity embedding vector. S6. Construct cross-platform candidate entity pairs from entity embedding vectors, calculate the distance for each pair of entity embedding vectors according to different distance formulas, and generate multi-distance similarity results. S7. Based on the multi-distance similarity results, perform duplicate entity determination on the candidate entity pairs. If there are duplicate entities, write them into the duplicate entity list. If there are no duplicate entities, do not perform cross-platform entity alignment processing.

2. The method for cross-platform business system integration based on knowledge graphs according to claim 1, characterized in that, S2 specifically includes: S21. Using the field text as input, construct a domain word vector library. The domain word vector library performs word segmentation, stop word removal, word frequency statistics, and co-occurrence statistics on historical business field text, industry terminology text, and entity name text. Then, it trains a word vector matrix based on the co-occurrence window and stores it in a fixed manner. S22. Input the field text into the input embedding layer of the ALBERT model. The input embedding layer performs word vector mapping on each token in the field text to generate basic word vectors. S23. Input the basic word vector and the corresponding domain word vector in the domain word vector library into the domain word vector gating layer. The domain word vector gating layer is equipped with a gating coefficient generation unit. The gating coefficient generation unit takes the connection vector between the basic word vector and the domain word vector as input and performs linear transformation and normalization operations on the connection vector to generate variable gating coefficients. S24. The domain word vector gating layer performs a weighted operation on the basic word vector and the domain word vector based on the variable gating coefficient to generate gated word vectors. S25. The gated word vectors are input into the factorized embedding layer of the ALBERT model and then into the subsequent shared layer structure of the ALBERT model to perform forward propagation. The representation vectors output by the model are transformed into field semantic vectors.

3. The method for cross-platform business system integration based on knowledge graphs according to claim 2, characterized in that, S3 specifically includes: S31. Generate entity node data for each entity identifier based on the entity identifier, and perform uniqueness processing on the entity node data; S32. Based on the relation type, source entity identifier and target entity identifier stored in the relation record, generate relation edge data, take the entity corresponding to the source entity identifier in the entity node data as the head entity, take the entity corresponding to the target entity identifier in the entity node data as the tail entity, take the relation type as the relation element, and combine the head entity, relation element and tail entity into a triple in a fixed order. S33. Write the generated triples into the data storage structure of the knowledge graph.

4. The method for cross-platform business system integration based on knowledge graphs according to claim 3, characterized in that, S4 specifically includes: S41. Initialize the entity vector, relation vector, and hyperplane normal vector in the TransH model, set the learning rate, maximum number of iterations, and loss threshold; set the rotation coefficients as trainable scalar parameters and assign them initial values, set the linear transformation matrix and bias vector as trainable matrix parameters and trainable bias parameters, and add the entity vector, relation vector, hyperplane normal vector, rotation coefficients, linear transformation matrix, and bias vector to the trainable parameter set of the TransH model; S42. Divide the triplet into multiple triplet batches, select the current triplet batch from the triplet batches, and map the head entity identifier and tail entity identifier in each triplet to the head entity vector and tail entity vector. S43. Map the relation elements in each triplet to a relation vector and an initial hyperplane normal vector. Perform vector subtraction on the head entity vector and tail entity vector to generate a difference vector, and perform vector summation to generate a sum vector. Concatenate the difference vector and the sum vector in a fixed order to form a rotation input vector. Perform linear transformation and normalization operations on the rotation input vector to generate a rotation reference vector. S44. Perform a rotation update on the initial hyperplane normal vector based on the rotation reference vector to obtain the rotated and updated hyperplane normal vector; S45. Based on the rotated updated hyperplane normal vector and relation vector, perform positive and negative sample scoring on the current triplet batch, construct the loss function and perform gradient update; S46. When the loss function is lower than the loss threshold or the number of iterations reaches the maximum number of iterations, the convergence condition is determined and the entity structure embedding vector is output; if the convergence condition is not met, return to S42 for the next round of iteration training.

5. The method for cross-platform business system integration based on knowledge graphs according to claim 4, characterized in that, S44 specifically includes: S442. Perform vector subtraction on the rotated reference vector and the hyperplane normal vector to generate a direction difference vector, and perform normalization on the direction difference vector to obtain a unit direction difference vector. S443. Perform scalar multiplication on the rotation coefficient and the unit vector of the direction difference to generate the rotation update. The scalar multiplication operation is to multiply the values ​​of each dimension of the unit vector of the direction difference by the rotation coefficient. S444. Perform vector addition on the rotation update amount and the hyperplane normal vector to generate the updated normal vector, and perform normalization operation on the updated normal vector to obtain the rotated and updated hyperplane normal vector.

6. The method for cross-platform business system integration based on knowledge graphs according to claim 5, characterized in that, S5 specifically includes: S51. Based on the entity structure embedding vector, match each entity structure embedding vector with the field semantic vector according to the entity identifier. S52. Perform vector concatenation on each matched entity structure embedding vector and field semantic vector, perform linear transformation and normalization to generate entity embedding vector.

7. The method for cross-platform business system integration based on knowledge graphs according to claim 6, characterized in that, S6 specifically includes: S61. Based on the entity source information recorded in the entity section, group the entity embedding vectors according to the entity source information to obtain a set of entity embedding vectors from different business platforms. S62. Select target entity embedding vectors sequentially from the first business platform entity embedding vector set. While selecting target entity embedding vectors, select comparison entity embedding vectors corresponding to target entity embedding vectors from the second business platform entity embedding vector set. Combine target entity embedding vectors and comparison entity embedding vectors to form candidate entity pairs. Repeat the selection operation to generate a candidate entity pair set. S63. For each pair of entity embedding vectors in the candidate entity pair set, perform the Euclidean distance algorithm to calculate the Euclidean distance value, the Manhattan distance algorithm to calculate the Manhattan distance value, and the cosine similarity algorithm to calculate the cosine similarity value. Record the Euclidean distance value, Manhattan distance value, and cosine similarity value in a fixed order as the multi-distance similarity result.

8. The method for cross-platform business system integration based on knowledge graphs according to claim 7, characterized in that, Specifically, S7 includes: S71. Sort the Euclidean distance and Manhattan distance values ​​of each candidate entity of the first business platform in ascending order, and sort the cosine similarity values ​​in descending order. Select the top five candidate entity pairs from the sorting results to generate the top five sets of Euclidean distance, Manhattan distance, and cosine similarity. S72. Perform set cross-search on the top five sets of Euclidean distance, the top five sets of Manhattan distance, and the top five sets of cosine similarity. When a candidate entity pair appears in the top five sets of Euclidean distance, the top five sets of Manhattan distance, and the top five sets of cosine similarity, the candidate entity pair is judged as a duplicate entity. S73. When there are no candidate entity pairs that appear simultaneously in the top five sets of Euclidean distance, the top five sets of Manhattan distance, and the top five sets of cosine similarity, perform average calculation of Euclidean distance, Manhattan distance, and cosine similarity values ​​for candidate entity pairs that appear simultaneously in any two of the top five sets of Euclidean distance, the top five sets of Manhattan distance, and the top five sets of cosine similarity to generate an average similarity value. Compare the average similarity value with a preset lenient threshold. If the lenient threshold condition is met, the candidate entity pair is determined to be a duplicate entity. S74. When no candidate entity pair appears in any two of the top five sets of Euclidean distance, top five sets of Manhattan distance, and top five sets of cosine similarity, the Euclidean distance value, Manhattan distance value, and cosine similarity value of the candidate entity pair in the three top five sets are compared with the preset strict threshold. When all strict threshold conditions are met, the candidate entity pair is determined to be a similar entity. S75. If a candidate entity pair does not satisfy S72, S73, or S74, the current candidate entity pair set is determined to have no similar entities, and cross-platform entity alignment processing is not performed.

Images