Cross-platform business system integration method based on knowledge graph

Abstract

The invention discloses a cross-platform business system integration method based on a knowledge graph. The method comprises the following steps: S1, reading data from a business platform database and generating a field text, an entity identifier and a relation record; s2, inputting the field text into an ALBERT model, and generating a field semantic vector through a field word vector gating layer; s3, forming a triple according to the entity identifier and the relation record, and writing the triple into a knowledge graph; s4, inputting the triple into a TransH model, and generating an entity structure embedding vector through dynamic rotation updating of a relation vector and a hyperplane normal vector; s5, generating an entity embedding vector according to the field semantic vector and the entity structure embedding vector; s6, calculating a plurality of distances based on the entity embedding vectors to obtain the similarity of the cross-platform candidate entity pairs; and S7, performing repeated entity judgment on the similarity, and updating the knowledge graph. According to the invention, accurate alignment of cross-platform service entities is realized, the data fusion efficiency and consistency are improved, and the method is suitable for a multi-source service integration scene.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cross-platform business data integration, and particularly relates to a cross-platform business system integration method based on a knowledge graph. BACKGROUND

[0002] Existing cross-platform business data integration generally relies on rule configuration, field name comparison or manual mapping methods. When there are large differences in field naming, inconsistent field descriptions or ambiguous semantic expressions between different business platforms, traditional methods are difficult to accurately align. Entities of different platforms often have multiple expression methods, such as abbreviations, aliases or non-uniform naming. String-based matching methods cannot effectively identify synonymous entities, resulting in a large amount of manual confirmation work, low integration efficiency and errors.

[0003] Existing knowledge graph construction methods are mostly based on static field parsing and inherent relationship structure, lacking deep extraction of field text semantics, resulting in insufficient semantic dimensions of generated entity representations. Existing embedding models usually use fixed relationship constraints in relationship learning, without considering the dynamic changes of entities in different business relationship environments, making it difficult for entity structure embedding to accurately reflect business semantics. Cross-platform entity alignment often relies on a single distance formula to evaluate similarity, which cannot measure the closeness between entities from multiple angles, and is prone to similarity instability or misjudgment problems.

[0004] Existing duplicate entity determination methods generally lack a multi-level screening mechanism, neither can they make comprehensive judgments based on different distance results, nor do they lack explicit quantitative standards for duplicate entities, making cross-platform entity alignment lack reliable basis.

[0005] Therefore, how to provide a cross-platform business system integration method based on a knowledge graph is a problem that those skilled in the art need to solve. SUMMARY

[0006] One object of the present application is to provide a cross-platform business system integration method based on a knowledge graph. The present application combines ALBERT model and TransH model to uniformly model multi-platform business fields and entity relationships, realizes automatic matching and duplicate entity determination of cross-platform entities, and has the advantages of accurate semantic understanding, high entity alignment accuracy and high degree of data integration automation.

[0007] According to the cross-platform business system integration method based on the knowledge graph, the following steps are included: S1, obtaining field names, field descriptions, entity identifiers and relationship records from a business platform database, pre-processing the obtained data to generate field texts, entity identifiers and relationship records; S2, construct an ALBERT model, load a field word vector library, input field texts into the ALBERT language model piece by piece, and generate field semantic vectors through the field word vector gating layer; S3, generate entity node data and relationship edge data according to entity identification and relationship records, organize them into triples, and write them into a knowledge graph; S4, construct a TransH model, set a relationship vector and a hyperplane normal vector, generate a rotating reference vector according to the head entity vector and the tail entity vector during each triple batch input, perform rotating update on the hyperplane normal vector, and iteratively train to obtain entity structure embedding vectors; S5, according to the correspondence between entity identification and field text, perform splicing, linear transformation and normalization on the field semantic vector and the entity structure embedding vector to generate an entity embedding vector; S6, construct a cross-platform candidate entity pair from the entity embedding vector, calculate the distance according to different distance formulas for each pair of entity embedding vectors, and generate a multi-distance similarity result; S7, according to the multi-distance similarity result, perform duplicate entity determination on the candidate entity pair, if there is a duplicate entity, write it into the duplicate entity, if there is no duplicate entity, do not perform cross-platform entity alignment processing.

[0008] Optionally, the S2 specifically comprises: S21, take the field text as input, construct a field word vector library, and train the word vector matrix based on the co-occurrence window and store it fixedly after performing word segmentation, stop word removal, word frequency statistics and co-occurrence statistics on historical business field texts, industry term texts and entity name texts; S22, input the field text into the input embedding layer of the ALBERT model, and the input embedding layer performs word vector mapping on each token in the field text to generate a basic word vector; S23, input the basic word vector and the field word vector corresponding to the token in the field word vector library into the field word vector gating layer at the same time, the field word vector gating layer is provided with a gating coefficient generation unit, and the gating coefficient generation unit takes the connection vector of the basic word vector and the field word vector as input, performs linear transformation and normalization operation on the connection vector to generate a variable gating coefficient; S24, the field word vector gating layer performs weighted operation on the basic word vector and the field word vector according to the variable gating coefficient to generate a gated word vector; S25, input the gated word vector into the factorization embedding layer of the ALBERT model and continue to input into the subsequent shared layer structure of the ALBERT model to perform forward propagation, and the representation vector output by the model forms a field semantic vector after output transformation processing.

[0009] Optionally, the S3 specifically comprises: S31, generating entity node data according to each entity identifier, and performing unique processing on the entity node data; S32, generating relationship edge data according to the relationship type, the source entity identifier and the target entity identifier stored in the relationship record, taking the entity corresponding to the source entity identifier in the entity node data as the head entity, taking the entity corresponding to the target entity identifier in the entity node data as the tail entity, taking the relationship type as the relationship element, and combining the head entity, the relationship element and the tail entity into a triple according to a fixed order; S33, writing the generated triple into the data storage structure of the knowledge graph.

[0010] Optionally, the S4 specifically comprises: S41, initializing the entity vector, the relationship vector and the hyperplane normal vector in the TransH model, setting the learning rate, the maximum iteration number and the loss threshold, setting the rotation coefficient as a trainable scalar parameter and assigning an initial value, setting the linear transformation matrix and the bias vector as trainable matrix parameters and trainable bias parameters, and adding the entity vector, the relationship vector, the hyperplane normal vector, the rotation coefficient, the linear transformation matrix and the bias vector to the trainable parameter set of the TransH model; S42, dividing the triples into multiple triple batches, selecting a current triple batch from the triple batches, and mapping the head entity identifier and the tail entity identifier in each triple into a head entity vector and a tail entity vector; S43, mapping the relationship element in each triple into a relationship vector and an initial hyperplane normal vector, performing vector subtraction operation on the head entity vector and the tail entity vector to generate a difference vector, and performing vector summation operation to generate a summation vector, and splicing the difference vector and the summation vector into a rotation input vector according to a fixed order; performing linear transformation operation and normalization operation on the rotation input vector to generate a rotation reference vector; S44, performing rotation update on the initial hyperplane normal vector according to the rotation reference vector to obtain a rotation updated hyperplane normal vector; S45, performing positive sample and negative sample scoring on the current triple batch based on the rotation updated hyperplane normal vector and the relationship vector, constructing a loss function and performing gradient update; S46, determining whether the convergence condition is reached when the loss function is lower than the loss threshold or the iteration number reaches the maximum iteration number, and outputting the entity structure embedding vector; if the convergence condition is not reached, returning to S42 for next round of iteration training.

[0011] Optionally, the S44 specifically comprises: S442, performing vector subtraction operation on the rotation reference vector and the hyperplane normal vector to generate a direction difference vector, and performing normalization operation on the direction difference vector to obtain a direction difference unit vector; S443, perform scalar multiplication operation on the rotation coefficient and the direction difference unit vector to generate a rotation update amount, the scalar multiplication operation being multiplication of each dimension value of the direction difference unit vector by the rotation coefficient; S444, perform vector addition operation on the rotation update amount and the hyperplane normal vector to generate an update normal vector, and perform normalization operation on the update normal vector to obtain the hyperplane normal vector after rotation update.

[0012] Optionally, the S5 specifically includes: S51, according to the entity structure embedding vector, correspondingly match each entity structure embedding vector with the field semantic vector according to the entity identifier; S52, perform vector splicing on each matched entity structure embedding vector and field semantic vector, perform linear transformation and normalization to generate an entity embedding vector.

[0013] Optionally, the S6 specifically includes: S61, according to the entity source information recorded in the entity section, group the entity embedding vectors according to the entity source information to obtain entity embedding vector sets from different business platforms; S62, sequentially select a target entity embedding vector from the first business platform entity embedding vector set, and simultaneously select a comparison entity embedding vector corresponding to the target entity embedding vector from the second business platform entity embedding vector set, to form a candidate entity pair, and repeat the selection operation to generate a candidate entity pair set; S63, perform Euclidean distance algorithm on each pair of entity embedding vectors in the candidate entity pair set to calculate Euclidean distance values, perform Manhattan distance algorithm to calculate Manhattan distance values, perform cosine similarity algorithm to calculate cosine similarity values, and record the Euclidean distance values, Manhattan distance values and cosine similarity values in a fixed order as multi-distance similarity results.

[0014] Optionally, the S7 specifically includes: S71, perform ascending order sorting on the Euclidean distance values and Manhattan distance values of each first business platform candidate entity, and perform descending order sorting on the cosine similarity values, and select the top five candidate entity pairs from the sorting results respectively to generate a Euclidean distance top five set, a Manhattan distance top five set and a cosine similarity top five set; S72, perform set intersection retrieval on the Euclidean distance top five set, the Manhattan distance top five set and the cosine similarity top five set, and when a candidate entity pair appears in the Euclidean distance top five set, the Manhattan distance top five set and the cosine similarity top five set at the same time, the candidate entity pair is determined as a duplicate entity; S73、when there is no candidate entity pair appearing in the Euclidean distance top five set, Manhattan distance top five set and cosine similarity top five set at the same time, the average calculation of the Euclidean distance value, Manhattan distance value and cosine similarity value is performed on the candidate entity pair appearing in any two sets of the Euclidean distance top five set, Manhattan distance top five set and cosine similarity top five set at the same time, the average similarity value is generated, the average similarity value is compared with the preset loose threshold, and when the loose threshold condition is met, the candidate entity pair is determined as a duplicate entity; S74、when there is no candidate entity pair appearing in any two sets of the Euclidean distance top five set, Manhattan distance top five set and cosine similarity top five set at the same time, the Euclidean distance value, Manhattan distance value and cosine similarity value are compared with the preset strict threshold for the candidate entity pair in the three top five sets, and when all the strict threshold conditions are met, the candidate entity pair is determined as a similar entity; S75、the candidate entity pair that does not meet S72, S73 and S74 is determined as no similar entity, and the cross-platform entity alignment processing is not performed.

[0015] The beneficial effects of the present application are: (1) By introducing the field word vector library and the ALBERT semantic enhancement mechanism in the field text processing, the deep understanding of the business field semantics is realized, the accuracy of the entity semantic representation is effectively improved, and the problem that the existing cross-platform field naming is inconsistent and the semantic expression difference is large, resulting in alignment difficulty, is solved.

[0016] (2) By introducing the dynamic rotation update mechanism of the rotation reference vector and the hyperplane normal vector in the TransH relationship embedding model, the entity structure embedding can be adaptively adjusted in combination with the business relationship environment, the discrimination ability of the entity structure representation is improved, and the problem of single structure representation and lack of dynamics of the traditional embedding model is supplemented.

[0017] (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, the risk of single distance misjudgment is effectively reduced, the precision of cross-platform entity matching and duplicate entity recognition is improved, and the intelligence and reliability of cross-platform business system integration are improved. BRIEF DESCRIPTION OF DRAWINGS

[0018] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation on the present application. In the drawings:

[0019] Fig. 1 A flowchart of a cross-platform business system integration method based on a knowledge graph is proposed for the present application; Fig. 2 An ALBERT model training flowchart of a cross-platform business system integration method based on a knowledge graph according to the present application; Fig. 3 A TransH training flowchart of a cross-platform business system integration method based on a knowledge graph according to the present application. DETAILED DESCRIPTION

[0020] The application will now be described in further detail with reference to the drawings. These drawings are simplified schematic diagrams, which only show the basic structure of the application in a schematic manner, and thus only show the components relevant to the application.

[0021] REFERENCE Figs. 1-3 A cross-platform business system integration method based on a knowledge graph, comprising the following steps: S1. Obtain field names, field descriptions, entity identifiers, and relationship records from a business platform database, perform preprocessing on the obtained data, and generate field texts, entity identifiers, and relationship records; S2. Construct an ALBERT model, load a domain word vector library, input the field texts 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 relationship edge data according to the entity identifiers and the relationship records, organize them into triples, and write them into a knowledge graph; S4. Construct a TransH model, set relationship vectors and hyperplane normal vectors, generate a rotation reference vector according to the head entity vector and the tail entity vector at each triple batch input, perform rotation update on the hyperplane normal vector, and iteratively train to obtain entity structure embedding vectors; S5. According to the correspondence between the entity identifiers and the field texts, perform splicing, linear transformation, and normalization on the field semantic vectors and the entity structure embedding vectors to generate entity embedding vectors; S6. Construct cross-platform candidate entity pairs from the entity embedding vectors, calculate the distance according to different distance formulas for each pair of entity embedding vectors, and generate multi-distance similarity results; S7. Perform duplicate entity determination on the candidate entity pairs according to the multi-distance similarity results, if there are duplicate entities, write them into the duplicate entities, and if there are no duplicate entities, do not perform cross-platform entity alignment processing.

[0022] In this embodiment, the S1 specifically comprises: S11, the business platform database provides field name, field description, entity identification and relationship record, traverses the database table structure according to the preset field order, reads the field name and field description in text form, reads the entity identification in the form of unique number, and reads and caches the relationship record in the form of combination of relationship type, source entity number and target entity number; S12, unify the character encoding format, remove redundant spaces, line breaks and abnormal symbols from the original text; perform short text processing on the field description, divide the long text field description into processable units according to the sentence boundary, combine the field name and the divided field description to form the field text; 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 noise words, and write the filtered result as the final form of the field text into the field text cache; when processing entity identification, perform uniqueness check on all numbers, remove duplicate numbers from the result set and keep unique numbers; when processing relationship records, perform format checking on each record, remove records that do not meet the format from the result set, and store the records that meet the format as relationship records according to the order of relationship type, source entity number and target entity number.

[0023] In this embodiment, S2 specifically includes: S21, input the field text as input, and build a domain word vector library, which is trained by performing word segmentation, stop word removal, word frequency statistics and co-occurrence statistics on historical business field text, industry term text and entity name text, and stored fixedly based on a co-occurrence window to form a word vector matrix; S22, input the field text into the input embedding layer of the ALBERT model, and the input embedding layer performs word vector mapping on each token in the field text to generate a basic word vector; S23, input the basic word vector and the domain word vector corresponding to the token in the domain word vector library into the domain word vector gating layer at the same time, the domain word vector gating layer is provided with a gating coefficient generation unit, the gating coefficient generation unit takes the connection vector of the basic word vector and the domain word vector as input, performs linear transformation and normalization operation on the connection vector to generate a variable gating coefficient; "token" represents the smallest text processing unit generated by the word segmentation algorithm, which is usually a single character or the smallest trainable unit in the form of a word in the Chinese scenario, used to represent the discretization structure of the field text before entering the model.

[0024] S24, the domain word vector gating layer performs weighted operation on the basic word vector and the domain word vector according to the variable gating coefficient to generate a gated word vector; S25, input the gating word vector into the factorization embedding layer of the ALBERT model and continue to input into the subsequent shared layer structure of the ALBERT model to perform forward propagation, and the representation vector output by the model is formed into a field semantic vector after output transformation processing.

[0025] In the embodiment, the S21 specifically includes: S211, the word segmentation algorithm adopts a Chinese word segmentation algorithm based on statistical characteristics, performs maximum probability segmentation on the text character sequence through word frequency statistics and mutual information calculation to generate a token sequence; the token sequence is used as the basic input for domain word vector training; S212, the word vector training method adopts a Skip-gram model, constructs a context token set within a context range of a window size of 5 for each center token, trains the model to predict the context token for the center token, and generates a domain word vector matrix; S213, the training parameters adopt fixed numerical 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 end; the word vector matrix is constructed according to the embedding dimension, and each token is mapped to a dense vector with a length of 128; the training process adopts a stochastic gradient descent to update the parameters; S214, the word vector matrix generated after the training is stored in the domain word vector library.

[0026] In the embodiment, the S23 specifically includes: S231, the basic word vector and the domain word vector perform vector connection according to the dimension order, and the linear transformation unit performs affine transformation on the connected vector; S232, element-level numerical compression is performed on the vector after the affine transformation, and a variable gating coefficient is generated by using an element-wise Sigmoid function; S233, the variable gating coefficient is used to control the weighting ratio of the basic word vector and the domain word vector.

[0027] In the embodiment, the S3 specifically includes: S31, generating entity node data for each entity identifier according to the entity identifier, and performing unique processing on the entity node data; S32, generating relationship edge data according to the relationship type, the source entity identifier and the target entity identifier stored in the relationship record, taking the corresponding entity of the source entity identifier in the entity node data as the head entity, taking the corresponding entity of the target entity identifier in the entity node data as the tail entity, taking the relationship type as the relationship element, and combining the head entity, the relationship element and the tail entity into a triple according to a fixed order; S33, writing the generated triple into the data storage structure of the knowledge graph.

[0028] In this embodiment, the S4 specifically includes: S41, initialize the entity vector, the relationship vector and the hyperplane normal vector in the TransH model, set the learning rate, the maximum iteration number and the loss threshold; set the rotation coefficient as a trainable scalar parameter and assign an initial value, set the linear transformation matrix and the bias vector as trainable matrix parameters and trainable bias parameters, and add the entity vector, the relationship vector, the hyperplane normal vector, the rotation coefficient, the linear transformation matrix and the bias vector to the trainable parameter set of the TransH model; S42, divide the triplets into multiple triplet batches, select a current triplet batch from the triplet batches, and map the head entity identifier and the tail entity identifier in each triplet to the head entity vector and the tail entity vector; S43, map the relationship element in each triplet to the relationship vector and the initial hyperplane normal vector, perform vector subtraction operation on the head entity vector and the tail entity vector to generate a difference vector, and perform vector summation operation to generate a summation vector, and concatenate the difference vector and the summation vector in a fixed order to generate a rotation input vector; perform linear transformation operation and normalization operation on the rotation input vector to generate a rotation reference vector; S44, perform rotation update on the initial hyperplane normal vector according to the rotation reference vector to obtain the rotation updated hyperplane normal vector; S45, perform positive sample and negative sample scoring on the current triplet batch based on the rotation updated hyperplane normal vector and the relationship vector, construct a loss function and perform gradient update; S46, when the loss function is lower than the loss threshold or the iteration number reaches the maximum iteration number, it is determined that the convergence condition is reached, and the entity structure embedding vector is output; if the convergence condition is not reached, return to S42 for the next iteration training; in this embodiment, the loss threshold is 0.001 and the maximum iteration number is 500.

[0029] In this embodiment, the S41 specifically includes: The entity vector initialization unit creates an entity vector according to each entity identifier in the entity identifier list, and the dimension of the entity vector is set to a fixed length; the initialization method adopts uniform distribution random initialization to generate values in the range of each dimension of the vector, forming the initial state of the entity vector; the relationship vector initialization unit creates a relationship vector according to each relationship element in the relationship element list, and the dimension of the relationship vector is consistent with that of the entity vector; the initialization method is consistent with that of the entity vector, and values are generated according to the same random distribution rule; the hyperplane normal vector initialization unit generates an initial normal vector corresponding to each relationship vector, and the dimension of the initial normal vector is consistent with that of the relationship vector; the normal vector initialization method adopts random sampling to generate an original normal vector in a fixed value interval, and performs normalization processing on the original normal vector to keep the length of the normal vector as a unit length.

[0030] In this embodiment, the S44 specifically includes: S442, performing vector subtraction operation on the rotation reference vector and the hyperplane normal vector to generate a direction difference vector, and performing normalization operation on the direction difference vector to obtain a direction difference unit vector; S443, performing scalar multiplication operation on the rotation coefficient and the direction difference unit vector to generate a rotation update amount, the scalar multiplication operation being multiplication of each dimension value of the direction difference unit vector and the rotation coefficient; S444, performing vector addition operation on the rotation update amount and the hyperplane normal vector to generate an updated normal vector, and performing normalization operation on the updated normal vector to obtain the hyperplane normal vector after rotation update.

[0031] In this embodiment, the S45 specifically includes: S451, the head entity vector and the tail entity vector after normal vector update enter the TransH projection operator, and the projection operator performs projection according to the standard orthogonal projection algorithm; S452, the score is calculated by using the L2 norm distance formula; S453, the loss value is calculated by using the maximum margin loss formula based on the Hinge loss function, and is transmitted to the back propagation link; the gradient calculation process generates gradients for the entity vector, the relation vector and the rotation coefficient by using the chain rule; and the gradient update method updates by using the stochastic gradient descent algorithm.

[0032] In this embodiment, the S5 specifically includes: S51, according to the entity structure embedding vector, each entity structure embedding vector and field semantic vector are matched according to the entity identifier; S52, each matched entity structure embedding vector and field semantic vector are spliced, linearly transformed and normalized to generate an entity embedding vector.

[0033] In this embodiment, the S6 specifically includes: S61, according to the entity source information recorded in the entity section record, the entity embedding vectors are grouped according to the entity source information to obtain entity embedding vector sets from different business platforms; S62, a target entity embedding vector is selected from the first business platform entity embedding vector set, and at the same time, a comparison entity embedding vector corresponding to the target entity embedding vector is selected from the second business platform entity embedding vector set, the target entity embedding vector and the comparison entity embedding vector are combined to form a candidate entity pair, and the selection operation is repeated to generate a candidate entity pair set; S63, performing Euclidean distance algorithm to calculate Euclidean distance value, performing Manhattan distance algorithm to calculate Manhattan distance value, performing cosine similarity algorithm to calculate cosine similarity value on each pair of entity embedding vectors in the candidate entity pair set, and recording the Euclidean distance value, Manhattan distance value and cosine similarity value in a fixed order as a multi-distance similarity result.

[0034] In the embodiment, the S7 specifically includes: S71, performing ascending order sorting on the Euclidean distance value and the Manhattan distance value of each first business platform candidate entity, performing descending order sorting on the cosine similarity value, and selecting the top five candidate entity pairs from the sorting results respectively to generate a Euclidean distance top five set, a Manhattan distance top five set and a cosine similarity top five set; S72, performing set intersection retrieval on the Euclidean distance top five set, the Manhattan distance top five set and the cosine similarity top five set, and determining a candidate entity pair as a duplicate entity when the candidate entity pair appears in the Euclidean distance top five set, the Manhattan distance top five set and the cosine similarity top five set at the same time; S73, when there is no candidate entity pair appearing in the Euclidean distance top five set, the Manhattan distance top five set and the cosine similarity top five set at the same time, performing average calculation of the Euclidean distance value, the Manhattan distance value and the cosine similarity value on the candidate entity pairs appearing in any two sets of the Euclidean distance top five set, the Manhattan distance top five set and the cosine similarity top five set at the same time to generate an average similarity value, comparing the average similarity value with a preset loose threshold, and determining the candidate entity pair as a duplicate entity when the loose threshold condition is met; the loose threshold is 0.85 in the embodiment; S74, when there is no candidate entity pair appearing in any two sets of the Euclidean distance top five set, the Manhattan distance top five set and the cosine similarity top five set at the same time, comparing the Euclidean distance value, the Manhattan distance value and the cosine similarity value of the candidate entity pairs in the three top five sets with a preset strict threshold respectively, and determining the candidate entity pair as a similar entity when all the strict threshold conditions are met; the strict threshold is 0.85 in the embodiment, and the values of all distance formulas are greater than the strict threshold to meet the condition; S75, the candidate entity pairs not meeting S72, S73 and S74 are determined as no similar entities, and no cross-platform entity alignment processing is performed.

[0035] Embodiment 1: In order to verify the feasibility of the application in implementation, the application is applied to a cross-platform information fusion scene where multiple independent business platforms exist, and the data structures between the platforms are quite different. The field naming method, field description content, entity identification structure, and relationship record format are all inconsistent, and there is a lack of unified entity between the platforms. When the business process is transmitted between multiple systems, problems such as data re-entry, entity redundancy, entity inconsistency, and mapping dependence on manual maintenance occur. Traditional methods rely more on manual mapping or string similarity matching based on fixed rules, and the processing speed is slow, the accuracy is not high, and it is difficult to maintain stable performance in a database environment with chaotic structure and non-standard description.

[0036] To solve these problems, the cross-platform business system integration method of the application is applied to a multi-source business database environment. Multiple business platforms exist user information, order information, warehouse information, logistics information, commodity information, and other entities, but the field description methods are different. For example, the user account is "login_id" in a certain platform and "usr_code" in another platform; the product SKU is "product_SKU" in platform A and "sku_id" in platform B; the warehouse code exists in English abbreviations, numerical codes, and mixed expression formats. The application uses a field word vector library to improve the field text semantic expression ability, uses a TransH structure embedding to reduce the cross-platform structure deviation, and identifies repeated entities through a multi-index fusion mechanism of distance calculation to achieve high-precision alignment across platforms.

[0037] In this embodiment, six types of entities are matched across platforms, and the total number of field texts of all platforms is 3821. The field word vector library is established in a way of word segmentation, word frequency statistics, and co-occurrence statistics, and achieves stable semantic distribution after embedding training. The ALBERT model encodes the semantic of the field text and outputs the field semantic vector, and the TransH model generates the entity structure embedding vector. After the two vectors are spliced and normalized, a unified entity embedding space is formed. In the entity alignment stage, the Euclidean distance, Manhattan distance, and cosine similarity are ranked by multiple indexes, and the top-ranked entities are selected for repeated entity detection.

[0038] In the experiment, manual annotation is selected as the actual alignment benchmark result. The performance of the application on multiple types of entities is significantly better than that of the traditional rule matching method. Taking the user account type entity as an example, the traditional method is limited by fixed rules and strong field ambiguity, and the accuracy is less than 80%; the accuracy of the application method is more than 95% under the conditions of semantic enhancement, structure embedding, and multi-distance fusion. The commodity SKU type entity is difficult to identify the differences in letter case, hyphen, and different naming formats in the traditional way, and the accuracy of the application method is significantly improved through the fusion of semantic representation and structure representation. Repeated entity recognition is particularly important in multi-source platforms, and the application realizes a significant increase in the number of repeated entities through the stable fusion of three distances, reducing the missed recognition.

[0039] In order to verify the feasibility of the application in implementation, the application is applied to a multi-source field description mixed business database environment, and six types of business entities are analyzed for cross-platform alignment. The experimental data are shown in Table 1, the application improves the accuracy and increases the number of repeated entity recognition.

[0040] Table 1: Entity matching effect comparison table

[0041] Table 1 shows the comparison results of six types of key business entities under the traditional method and the application method. The traditional rule matching has a low accuracy on entities with large field description differences and inconsistent naming, and the accuracy is concentrated in 66% to 78%. The application fuses semantic enhancement and structure embedding, so that entities of the same type have concentrated and distinguishable representation in the vector space, and the accuracy is improved to 91% to 96%. The number of repeated entity recognition is significantly improved in the application method, with a growth rate of more than 30%, which shows that the application can effectively reduce entity redundancy and improve cross-platform consistency in the process of multi-platform data fusion. The accuracy of order number, user account and commodity SKU in the six types of entities is improved most prominently, which shows that the semantic enhancement mechanism and TransH structure embedding have stability and scalability in the structure chaotic database environment.

[0042] The above is only the preferred specific embodiment of the application, but the protection scope of the application is not limited thereto, any skilled person in the art can make equivalent replacement or change according to the technical solution and the inventive concept of the application within the technical range disclosed by the application, which should be covered in the protection scope of the application.

Claims

1. A knowledge graph-based cross-platform business system integration method, characterized in that, The method comprises the following steps: S1, obtaining field name, field description, entity identifier and relationship record from a service platform database, performing preprocessing on the obtained data to generate field text, entity identifier and relationship record; S2, constructing an ALBERT model, loading a field word vector library, inputting the field text into the ALBERT language model piece by piece, and generating field semantic vectors through a field word vector gating layer; S3, generating entity node data and relationship edge data according to the entity identifier and the relationship record, organizing the data into triples, and writing the triples into a knowledge graph; S4, constructing a TransH model, setting a relationship vector and a hyperplane normal vector, generating a rotating reference vector according to a head entity vector and a tail entity vector when each triple batch is input, performing rotating update on the hyperplane normal vector, and iteratively training to obtain an entity structure embedding vector; S5, according to the correspondence between the entity identifier and the field text, performing splicing, linear transformation and normalization on the field semantic vector and the entity structure embedding vector to generate an entity embedding vector; S6, constructing a cross-platform candidate entity pair from the entity embedding vector, calculating the distance according to different distance formulas for each pair of entity embedding vectors to generate a multi-distance similarity result; S7, performing duplicate entity determination on the candidate entity pair according to the multi-distance similarity result, if there is a duplicate entity, writing the duplicate entity into a duplicate entity, if there is no duplicate entity, not performing cross-platform entity alignment processing. 2.The knowledge graph-based cross-platform business system integration method of claim 1, wherein, The S2 specifically comprises: S21, taking the field text as input, constructing a field word vector library, and training a word vector matrix based on a co-occurrence window and fixed storage after performing word segmentation, stop word removal, word frequency statistics and co-occurrence statistics on historical business field text, industry term text and entity name text; S22, inputting the field text into the input embedding layer of the ALBERT model, and the input embedding layer performing word vector mapping on each token in the field text to generate a basic word vector; S23, inputting the basic word vector and the field word vector corresponding to the token in the field word vector library into the field word vector gating layer at the same time, the field word vector gating layer being provided with a gating coefficient generation unit, the gating coefficient generation unit taking the connection vector of the basic word vector and the field word vector as input, performing linear transformation and normalization operation on the connection vector to generate a variable gating coefficient; S24, the field word vector gating layer performing weighting operation on the basic word vector and the field word vector according to the variable gating coefficient to generate a gated word vector; S25, inputting the gated word vector into the factorization embedding layer of the ALBERT model and continuing to input into the subsequent shared layer structure of the ALBERT model to perform forward propagation, and the representation vector output by the model forms the field semantic vector after output transformation processing. 3.The knowledge graph-based cross-platform business system integration method of claim 2, wherein, The S3 specifically comprises: S31, generating entity node data for each entity identifier according to the entity identifier, and performing unique processing on the entity node data; S32, generating a relation edge data according to the relation type stored in the relation record, the source entity identifier and the target entity identifier, taking the entity corresponding to the source entity identifier in the entity node data as a head entity, taking the entity corresponding to the target entity identifier in the entity node data as a tail entity, taking the relation type as a relation element, and combining the head entity, the relation element and the tail entity into a triple according to a fixed order; S33, writing the generated triple into a data storage structure of the knowledge graph.

4. The knowledge graph-based cross-platform business system integration method of claim 3, wherein, The S4 specifically comprises: S41, initializing entity vectors, relation vectors and hyperplane normal vectors in the TransH model, setting a learning rate, a maximum iteration number and a loss threshold, setting a rotation coefficient as a trainable scalar parameter and assigning an initial value to the rotation coefficient, setting a linear transformation matrix and a bias vector as trainable matrix parameters and trainable bias parameters, and adding the entity vectors, the relation vectors, the hyperplane normal vectors, the rotation coefficient, the linear transformation matrix and the bias vector to a trainable parameter set of the TransH model; S42, dividing the triples into multiple triple batches, selecting a current triple batch from the triple batches, and mapping the head entity identifier and the tail entity identifier in each triple into a head entity vector and a tail entity vector; S43, mapping the relation element in each triple into a relation vector and an initial hyperplane normal vector, performing a vector subtraction operation on the head entity vector and the tail entity vector to generate a difference vector, performing a vector summation operation to generate a summation vector, and splicing the difference vector and the summation vector into a rotation input vector according to a fixed order; performing a linear transformation operation and a normalization operation on the rotation input vector to generate a rotation reference vector; S44, performing a rotation update on the initial hyperplane normal vector according to the rotation reference vector to obtain a rotation-updated hyperplane normal vector; S45, performing positive sample and negative sample scoring on the current triple batch based on the rotation-updated hyperplane normal vector and the relation vector, constructing a loss function and performing gradient update; S46, determining whether a convergence condition is reached when the loss function is lower than the loss threshold or the iteration number reaches the maximum iteration number, and outputting the entity structure embedding vector; if the convergence condition is not reached, returning to S42 for next round of iteration training. 5.The knowledge graph-based cross-platform business system integration method of claim 4, wherein, The S44 specifically comprises: S442, performing a vector subtraction operation on the rotation reference vector and the hyperplane normal vector to generate a direction difference vector, and performing a normalization operation on the direction difference vector to obtain a direction difference unit vector; S443, performing a scalar multiplication operation on the rotation coefficient and the direction difference unit vector to generate a rotation update amount, the scalar multiplication operation being multiplication of each dimension value of the direction difference unit vector and the rotation coefficient; S444, performing a vector addition operation on the rotation update amount and the hyperplane normal vector to generate an updated normal vector, and performing a normalization operation on the updated normal vector to obtain the rotation-updated hyperplane normal vector.

6. The knowledge graph-based cross-platform business system integration method of claim 5, wherein, The S5 specifically comprises: S51, according to the entity structure embedding vector, corresponding matching each entity structure embedding vector and a field semantic vector according to an entity identifier; S52, perform vector splicing on the matched entity structure embedding vector and the field semantic vector, perform linear transformation and normalization to generate an entity embedding vector.

7. The knowledge graph-based cross-platform business system integration method of claim 6, wherein, The S6 specifically comprises: S61, group the entity embedding vectors according to the entity source information according to the entity source information recorded in the entity section record, to obtain entity embedding vector sets from different business platforms; S62, sequentially select a target entity embedding vector from the first business platform entity embedding vector set, while selecting the target entity embedding vector, select a comparison entity embedding vector corresponding to the target entity embedding vector from the second business platform entity embedding vector set, compose a candidate entity pair from the target entity embedding vector and the comparison entity embedding vector, and repeatedly perform the selection operation to generate a candidate entity pair set; S63, calculate the Euclidean distance value by performing the Euclidean distance algorithm on each pair of entity embedding vectors in the candidate entity pair set, calculate the Manhattan distance value by performing the Manhattan distance algorithm, calculate the cosine similarity value by performing the cosine similarity algorithm, and record the Euclidean distance value, the Manhattan distance value and the cosine similarity value in a fixed order as a multi-distance similarity result. 8.The knowledge graph-based cross-platform business system integration method of claim 7, wherein, The S7 specifically comprises: S71, sort the Euclidean distance values and the Manhattan distance values of each first business platform candidate entity in ascending order, sort the cosine similarity values in descending order, select the top five candidate entity pairs from the sorting results respectively, and generate a Euclidean distance top five set, a Manhattan distance top five set and a cosine similarity top five set; S72, perform set intersection retrieval on the Euclidean distance top five set, the Manhattan distance top five set and the cosine similarity top five set, and when a candidate entity pair appears in the Euclidean distance top five set, the Manhattan distance top five set and the cosine similarity top five set at the same time, determine the candidate entity pair as a duplicate entity; S73, when there is no candidate entity pair appearing in the Euclidean distance top five set, the Manhattan distance top five set and the cosine similarity top five set at the same time, perform average calculation on the Euclidean distance value, the Manhattan distance value and the cosine similarity value of the candidate entity pair appearing in any two sets of the Euclidean distance top five set, the Manhattan distance top five set and the cosine similarity top five set at the same time, to generate an average similarity value, compare the average similarity value with a preset loose threshold, and when the loose threshold condition is met, determine the candidate entity pair as a duplicate entity; S74, when there is no candidate entity pair appearing in any two sets of the Euclidean distance top five set, the Manhattan distance top five set and the cosine similarity top five set at the same time, compare the Euclidean distance value, the Manhattan distance value and the cosine similarity value of the candidate entity pair in the three top five sets with a preset strict threshold, and when all the strict threshold conditions are met, determine the candidate entity pair as a similar entity; S75, the candidate entity pair that does not meet S72, S73 and S74, determines the current candidate entity pair set as having no similar entity, and does not perform cross-platform entity alignment processing.

Images