Field alignment mapping method and system combined with third-party credit investigation data

Abstract

The application provides a field alignment mapping method and system combined with third-party credit investigation data, relates to the technical field of multi-modal data processing, and first acquires a multi-source heterogeneous credit investigation data set of a target enterprise; then field semantic units are respectively extracted from data of different structures to obtain an original field semantic unit group integrating different types of semantic units; a field semantic topology structure oriented to the field of enterprise credit investigation is constructed, the original field semantic units are topologically associated and mapped with standard field semantic core nodes in the topology to obtain a field alignment mapping scheme recording corresponding relationships and association paths; and finally, field fusion reconstruction is completed according to the mapping scheme, multi-source semantic units under the same standard field are merged into the same field container, and a unified credit investigation field structure of the target enterprise is generated. The application can effectively improve the accuracy of multi-source heterogeneous third-party credit investigation data field alignment mapping and adapt to the data processing requirements of subsequent enterprise credit evaluation.

Description

Technical Field

[0001] This application relates to the field of multimodal data processing technology, and more specifically, to a field alignment mapping method and system that combines third-party credit data. Background Technology

[0002] In corporate credit risk control and credit assessment scenarios, in order to fully understand the credit status of enterprises, it is often necessary to collect credit data from multiple different third-party credit reporting platforms. The data construction standards of different third-party credit reporting platforms are not uniform, and there are significant differences in field names, description methods, and storage structures for the same credit dimension. At the same time, different platforms will output credit data with different structural types. Therefore, it is necessary to perform field alignment mapping processing on credit data from multiple sources in order to integrate multi-source data into a unified credit data structure for credit assessment.

[0003] Most existing field alignment mapping methods only target structured numerical credit reporting data, matching original field names to standard fields through string matching. This approach fails to adapt to multi-source heterogeneous data scenarios that simultaneously contain unstructured text, semi-structured images, and structured numerical data. It also cannot handle alignment requirements for non-standardized original fields such as text fragments in unstructured text or visual regions in semi-structured images. Some solutions attempt to introduce semantic alignment to address the mapping problem of heterogeneous data, but most do not construct a dedicated semantic topology structure tailored to the characteristics of the enterprise credit reporting domain. They only use a general semantic space for matching, failing to reflect the contextual relationships between credit reporting fields, unable to find accurate corresponding standard nodes for original semantic units, and unable to record complete mapping paths. This easily leads to matching errors and low mapping accuracy. Furthermore, existing methods only perform simple data concatenation after mapping, without storing the merged multi-source semantic units into the same field container based on standard field nodes. This fails to generate a clear and reasonably merged unified credit reporting field structure, leading to low data retrieval efficiency in subsequent credit assessment processes and the potential for duplicate calculations of data in the same dimension. In other words, existing methods cannot adapt to the field alignment mapping requirements of multi-source heterogeneous third-party credit reporting data, resulting in low mapping accuracy and an unreasonable data structure after fusion, making it difficult to meet the needs of enterprises for high-quality unified credit reporting data in credit assessment. Summary of the Invention

[0004] In view of this, the purpose of this application is to provide a field alignment mapping method and system that combines third-party credit data.

[0005] According to a first aspect of this application, a field alignment mapping method combining third-party credit data is provided, the method comprising: Obtain a multi-source heterogeneous credit data set of the target enterprise, wherein the multi-source heterogeneous credit data set includes unstructured text credit reports, semi-structured image credit certificates, and structured numerical credit record tables collected from multiple third-party credit reporting platforms; The multi-source heterogeneous credit data set is processed by extracting field semantic units to obtain the original field semantic unit group of each credit data source. The original field semantic unit group includes text fragment semantic units segmented from the unstructured text credit report, visual region semantic units identified from the semi-structured image credit certificate, and indicator item semantic units parsed from the structured numerical credit record table. A field semantic topology structure for the enterprise credit investigation field is constructed. The field semantic topology structure includes standard field semantic core nodes and context co-occurrence association connections between the standard field semantic core nodes. Topological association mapping processing is performed on each semantic unit in the original field semantic unit group and the standard field semantic core node to obtain a field alignment mapping scheme. Each mapping record in the field alignment mapping scheme records the corresponding association path between a semantic unit and a standard field semantic core node. According to the field alignment mapping scheme, the original field semantic units in the multi-source heterogeneous credit data set are subjected to field fusion and reconstruction processing to generate a unified credit field structure for the target enterprise. In the unified credit field structure, the multi-source semantic units under the same standard field semantic core node are merged into the same field container.

[0006] According to a second aspect of this application, a field alignment mapping system combining third-party credit data is provided. The field alignment mapping system combining third-party credit data includes a machine-readable storage medium and a processor. The machine-readable storage medium stores machine-executable instructions. When the processor executes the machine-executable instructions, the field alignment mapping system combining third-party credit data implements the aforementioned field alignment mapping method combining third-party credit data.

[0007] Based on any of the above aspects, the technical effect of this application is as follows: By acquiring a multi-source heterogeneous credit data set of target enterprises, including unstructured text credit reports, semi-structured image credit certificates, and structured numerical credit record tables, this study extracts semantic units from the multi-source heterogeneous credit data set. Original semantic unit groups of corresponding types are extracted from the three different data structures, realizing the transformation of heterogeneous data with different structures into a unified semantic unit form. Furthermore, a field semantic topology structure for the enterprise credit reporting domain is constructed, containing standard field semantic core nodes and contextual co-occurrence associations between nodes. Based on this topology structure, each original semantic unit undergoes topological association mapping processing to obtain a field alignment mapping scheme recording corresponding relationships and association paths. This fully utilizes the semantic association characteristics of fields in the enterprise credit reporting domain, and improves the accuracy of mapping through the contextual association of the topology structure. This approach not only records the mapping path completely for easy tracing and verification, but also solves the problem of low accuracy in general semantic matching. Finally, based on the field alignment mapping scheme, field fusion and reconstruction processing is performed to generate a unified credit reporting field structure for the target enterprise, which merges multi-source semantic units under the same standard node into the same container. This achieves reasonable merging and reconstruction of multi-source heterogeneous credit reporting fields, generating a unified credit reporting data structure that is clear in structure and easy to call. It avoids fragmented storage and repeated calculation of data in the same dimension, and solves the technical defects of existing methods that cannot adapt to multi-source heterogeneous third-party credit reporting data, have low mapping accuracy, and have unreasonable fusion structures. This significantly improves the accuracy of third-party credit reporting data field alignment mapping, and the generated unified credit reporting field structure is more suitable for the subsequent processing needs of enterprise credit assessment, improving the overall accuracy and processing efficiency of enterprise credit assessment. Attached Figure Description

[0008] Figure 1 A flowchart illustrating the field alignment mapping method combining third-party credit data provided in an embodiment of this application is shown. Figure 2 This illustration shows a flowchart of the extraction of semantic units and construction of semantic topology of multi-source heterogeneous credit data fields in an embodiment of this application.

[0009] Figure 3 This paper illustrates a flowchart of the topology association mapping process and field alignment mapping scheme generation in an embodiment of this application.

[0010] Figure 4 The diagram illustrates the process flow of field fusion reconstruction processing and optimized verification closed loop in an embodiment of this application. Detailed Implementation

[0011] Figure 1 This paper illustrates a flowchart of a field alignment mapping method combining third-party credit data provided in an embodiment of this application. The detailed steps include: Step S110: Obtain a multi-source heterogeneous credit data set of the target enterprise. The multi-source heterogeneous credit data set includes unstructured text credit reports, semi-structured image credit certificates, and structured numerical credit record tables collected from multiple third-party credit reporting platforms.

[0012] After the target company submits a credit application, with its explicit authorization, data retrieval requests are initiated to multiple third-party credit reporting platforms via application programming interfaces (APIs). These requests include the target company's unique identifier and the hash value of its electronic authorization certificate. The resulting multi-source heterogeneous credit data set comprises three forms: The first is an unstructured text credit report, a plain text stream without a predefined format. The second is a semi-structured image credit certificate, an image pixel matrix containing visual layout information. The third is a structured numerical credit record table, a well-organized data table with clearly defined row and column separators. The data collection process adheres to legal and compliant principles, with the authorized scope limited to the purpose of this credit assessment. Privacy-sensitive information involving corporate trade secrets is anonymized and hashed using the SHA256 algorithm. The converted data is stored in a secure, isolated area, linked to the original data via a one-way mapping table.

[0013] Step S120: Perform field semantic unit extraction processing on the multi-source heterogeneous credit data set to obtain the original field semantic unit group for each credit data source. The original field semantic unit group includes text fragment semantic units segmented from the unstructured text credit report, visual region semantic units identified from the semi-structured image credit certificate, and indicator item semantic units parsed from the structured numerical credit record table.

[0014] Step S121: Obtain the full-text character stream of the unstructured text credit report, and perform continuous character sequence scanning processing on the full-text character stream through a sliding semantic window to generate multiple candidate character fragments.

[0015] The full-text character stream of the unstructured text credit report is completely read into memory using UTF-8 character encoding. The sliding semantic window is a character sequence observation box with a fixed length; its width parameter is set based on the average character length of terminology in the credit reporting field. Starting from the beginning of the full-text character stream, the sliding semantic window slides towards the end of the stream, incrementing by single characters. After each slide, a continuous character sequence within the sliding semantic window is truncated into a candidate character segment. The number of characters contained in a candidate character segment is determined by the width parameter of the sliding semantic window. By traversing the full-text character stream, a set of candidate character segments covering all character positions is generated.

[0016] Step S122: Perform boundary integrity assessment processing on each candidate character segment, and determine whether the starting character position and ending character position of the candidate character segment constitute a complete field semantic unit based on the preset field semantic boundary feature set. The field semantic boundary feature set includes prefix character features of credit reporting terms and suffix character features of credit reporting terms.

[0017] The pre-defined field semantic boundary feature set is a character feature library storing common start and end markers for credit reporting terminology. Prefix character features for credit reporting terminology include high-frequency starting character sequences such as "enterprise," "company," "tax," and "credit." Suffix character features for credit reporting terminology include high-frequency ending character sequences such as "level," "amount," "rating," and "status," as well as punctuation characters such as commas, periods, and colons. The boundary integrity assessment process iterates through each candidate character segment, checking whether the starting character position of the candidate character segment matches any prefix character feature of a credit reporting terminology in the field semantic boundary feature set, and simultaneously checking whether the ending character position of the candidate character segment matches any suffix character feature of a credit reporting terminology in the field semantic boundary feature set. When the starting character position matches the prefix and the ending character position matches the suffix, the candidate character segment is determined to constitute a complete field semantic unit.

[0018] Step S123: The candidate character segments that have undergone the boundary integrity assessment are identified as text segment semantic units, and the text segment semantic units are associated with and stored in the paragraph position index and sentence offset position in the unstructured text credit report to obtain text segment semantic unit records.

[0019] Candidate character segments that have passed boundary integrity assessment and are deemed to constitute complete field semantic units are officially marked as text segment semantic units. During the generation of each text segment semantic unit, its location metadata in the unstructured text credit report is extracted simultaneously. The location metadata includes a paragraph position index and an intra-sentence offset. The paragraph position index is the number of the paragraph containing the text segment semantic unit, with numbers sequentially increasing from the first paragraph of the entire text. The intra-sentence offset is the character position number of the first character of the text segment semantic unit within its sentence. The character sequence content, paragraph position index, and intra-sentence offset of a text segment semantic unit constitute a text segment semantic unit record, which is stored in a linked set of text segment semantic unit records.

[0020] Step S124: Obtain the pixel matrix of the semi-structured image credit certificate, perform visual region segmentation processing on the pixel matrix, and divide the title text region block, seal region block and data table region block based on the prior features of the credit certificate layout.

[0021] The semi-structured image credit certificate is loaded as a three-channel color pixel matrix, with each pixel representing the brightness value of the red, green, and blue color channels. Visual region segmentation processing employs a region partitioning logic based on the prior features of the credit certificate layout. These prior features define the layout patterns of typical image credit certificates: the title text is usually located at the top of the pixel matrix, with dense strokes and a color darker than the background; the seal is usually located in the lower right area of ​​the pixel matrix, appearing as a cluster of circular red blocks; the data table is located in the middle area between the title and the seal, consisting of a grid formed by continuous horizontal and vertical lines. Through color space analysis and connected component detection, the pixel matrix is ​​divided into three non-overlapping pixel sub-matrices: the title text region, the seal region, and the data table region.

[0022] Step S125: Perform optical character recognition processing on the title text area block, convert the pixel array in the title text area block into a first text sequence, and match the first text sequence with a preset credit field title dictionary to obtain a visual region semantic unit.

[0023] The pixel submatrix of the title text region is input into the optical character recognition engine. The engine binarizes this submatrix, dividing the pixels into foreground text pixels and background blank pixels according to a brightness threshold. Then, connected component analysis is performed on the foreground text pixels to segment the glyph contours of individual characters. Each glyph contour is matched against a pre-trained glyph feature template library, outputting the corresponding character code. All recognized characters are arranged in their original pixel position order to form the first text sequence. The pre-defined credit report field title dictionary is a set of strings containing common title names for credit reports. The first text sequence is precisely or fuzzily matched against entries in the credit report field title dictionary; successfully matched text sequences are identified as visual region semantic units.

[0024] Step S126: Perform table structure parsing processing on the data table area block, extract the coordinates of the row and column separators and the cell boundary coordinates, crop the cell image block according to the cell boundary coordinates, perform the optical character recognition processing on the cell image block to obtain the second text sequence, and match the second text sequence with the credit field title dictionary to obtain the visual region semantic unit.

[0025] The pixel submatrix of the data table region block is first subjected to table line detection. Horizontal and vertical line segments in the pixel matrix are identified using the Hough transform line detection algorithm, and the coordinates of the endpoints of these segments are recorded. The coordinates of the intersection points of the horizontal and vertical line segments constitute the set of cell vertex coordinates for the table. A rectangular area enclosed by four adjacent vertices constitutes a cell. Each cell image block is cropped from the data table region block based on its cell boundary coordinates. Each cell image block is sequentially input into the optical character recognition engine, which outputs a second text sequence. This second text sequence is matched against a credit field title dictionary; successfully matched text sequences are identified as visual region semantic units.

[0026] Step S127: Extract visual region semantic units from the seal area block of the semi-structured image credit certificate, parse and standardize the structured numerical credit record table, and then group all extraction results according to source type.

[0027] After processing including seal circular outline detection, polar coordinate unfolding, and optical character recognition, visual region semantic units are extracted from the seal area block. Structured numerical credit record tables generate indicator item semantic units after header parsing and indicator item standardization. Text fragment semantic unit records, visual region semantic units, and indicator item semantic units are grouped and aggregated according to their respective credit data source types, forming text fragment semantic unit groups for unstructured text credit reports, visual region semantic unit groups for semi-structured image credit certificates, and indicator item semantic unit groups for structured numerical credit record tables.

[0028] Step S1271: Perform circular outline detection processing on the stamp area block, extract the circular pixel area of ​​text within the circular outline of the stamp, perform polar coordinate expansion processing on the circular pixel area of ​​text to obtain horizontally arranged text strips, perform optical character recognition processing on the horizontally arranged text strips to obtain a third text sequence, and match the third text sequence with a preset set of enterprise name suffix feature characters to obtain a visual region semantic unit.

[0029] A circular contour detection is first performed on the pixel submatrix of the seal area block. A Hough circle detection algorithm is adopted to search the pixel submatrix for a pixel edge set whose center coordinates and radius parameters conform to a circle equation, so as to determine the pixel position of the center of the circular seal contour and the pixel length of the radius. Taking the center of the circle as the pole, the annular pixel area inside the circular seal contour is extracted. A polar coordinate expansion process is performed on the text annularly arranged pixel area, and the annularly distributed text pixels are transformed into a text strip arranged at equal intervals in the horizontal direction according to a polar coordinate formula. The horizontally arranged text strip is input into an optical character recognition engine, which outputs a third text sequence. The preset enterprise name suffix feature character set includes legally required ending words of enterprise names such as "Limited Company", "Company Limited by Shares", and "Partnership Enterprise". The third text sequence is subjected to suffix matching with the enterprise name suffix feature character set, and the text sequence that is successfully matched is determined as a visual region semantic unit.

[0030] Step S1272: Acquire the header definition line and data record lines of the structured numerical credit reference record table, perform field delimiter scanning processing on the header definition line, and split the header definition line into an indicator item name sequence according to a preset field delimiter set, where the field delimiter set includes a tab character, a vertical line character, and a comma character.

[0031] The structured numerical credit reference record table is stored in the form of plain text or formatted text. First, the header definition line, which is usually the first line of text in the file, is located. A field delimiter scanning process is performed on the header definition line, the character sequence of the header definition line is traversed character by character, and it is detected whether a character matches any delimiter character in the preset field delimiter set. When a delimiter character is detected, the header definition line is split into a plurality of sub-strings with the position of the delimiter as a breakpoint. All the sub-strings are arranged in the splitting order to form the indicator item name sequence.

[0032] Step S1273: Perform semantic unit standardization processing on each indicator item name in the indicator item name sequence, compare and convert the indicator item name with a preset credit reference indicator synonym mapping dictionary, so as to generate an indicator item semantic unit with a unified expression form.

[0033] The preset credit reference indicator synonym mapping dictionary stores the correspondence between field names with the same meaning but different expression modes in the credit reference field. For example, "operating revenue" and "main business revenue" are synonyms of each other, and "asset-liability ratio" and "liability ratio" are synonyms of each other. Each indicator item name in the indicator item name sequence is used as a query key to perform a search in the credit reference indicator synonym mapping dictionary. If a matching entry is retrieved, the indicator item name is replaced with the standard expression form defined in the dictionary. If no matching entry is retrieved, the original expression form remains unchanged. The result after the comparison and conversion is the indicator item semantic unit with a unified expression form.

[0034] Step S1274: The text fragment semantic unit records, the visual region semantic units, and the indicator item semantic units are grouped and collected according to their respective credit data source types to form the text fragment semantic unit group corresponding to the unstructured text credit report, the visual region semantic unit group corresponding to the semi-structured image credit certificate, and the indicator item semantic unit group corresponding to the structured numerical credit record table.

[0035] All entries in the text fragment semantic unit record set are grouped into the text fragment semantic unit group, with the data source type of this group labeled as "unstructured text". All visual region semantic units extracted from the title text region block, data table region block, and seal region block are grouped into the visual region semantic unit group, with the data source type of this group labeled as "semi-structured image". All indicator item semantic units in the indicator item semantic unit sequence are grouped into the indicator item semantic unit group, with the data source type of this group labeled as "structured numerical". The three original field semantic unit groups constitute the original field semantic unit group set, which serves as the input for subsequent topological association mapping processing.

[0036] Step S130: Construct a field semantic topology structure for the enterprise credit investigation field. The field semantic topology structure includes standard field semantic core nodes and context co-occurrence association connections between the standard field semantic core nodes. Perform topological association mapping processing on each semantic unit in the original field semantic unit group and the standard field semantic core node to obtain a field alignment mapping scheme. Each mapping record in the field alignment mapping scheme records the corresponding association path between a semantic unit and a standard field semantic core node.

[0037] Step S131: Collect a large number of enterprise credit reporting business record documents, extract field names from the field definition section of the enterprise credit reporting business record documents, and extract standard field name entries from the field definition section to form a set of standard field name entries.

[0038] The massive enterprise credit reporting records are sourced from publicly available field standard specifications, financial institution loan approval form templates, and enterprise credit rating report templates within the credit reporting industry. Each enterprise credit reporting record contains a field definition section, which is marked with headings such as "Field Name," "Field Description," and "Field Definition." The field name extraction process locates the start and end range of the field definition section, scans line by line within this range, and extracts the text string located at the field name position in each line. After removing leading and trailing whitespace characters and standardizing full-width and half-width formats, the extracted text strings are stored as field standard name entries in a set of field standard name entries.

[0039] Step S132: Perform context co-occurrence frequency statistics on any two field standard name entries in the set of field standard name entries, traverse the full text of the enterprise credit reporting business record document, record the document frequency count of the two field standard name entries appearing together in the same credit reporting field description paragraph, and generate a field co-occurrence frequency record table.

[0040] Select any two field standard name entries from the set of field standard name entries, denoted as the first field standard name entry and the second field standard name entry. Iterate through the full text of all enterprise credit investigation business record documents, using natural paragraphs as the statistical window. When the text strings of the first and second field standard name entries appear simultaneously within the same natural paragraph, the document frequency count is incremented by one. After iterating and statistically analyzing all field standard name entry pairs, store each pair of field standard name entries and its corresponding document frequency count in the field co-occurrence frequency record table.

[0041] Step S133: Perform co-occurrence strength measurement processing on each pair of field standard name entries in the field co-occurrence frequency record table, and perform normalized association calculation on the document frequency count and the independent occurrence frequency of each of the two field standard name entries in the enterprise credit business record document to obtain the field co-occurrence association strength value.

[0042] For each pair of standard name entries in the field co-occurrence frequency record table, the co-occurrence strength measurement uses a point-to-point mutual information algorithm. The calculation process is as follows: Divide the document frequency count by the total number of natural paragraphs to obtain the joint occurrence probability of the first and second standard name entries. Divide the independent occurrence frequency of the first standard name entry by the total number of natural paragraphs to obtain the independent occurrence probability of the first standard name entry, and divide the independent occurrence frequency of the second standard name entry by the total number of natural paragraphs to obtain the independent occurrence probability of the second standard name entry. The field co-occurrence association strength value is equal to the natural logarithm of the joint occurrence probability minus the sum of the natural logarithms of the independent occurrence probabilities of the first and second standard name entries. A positive result for the field co-occurrence association strength value indicates that the two standard name entries have a positive co-occurrence association, while a negative result indicates that the two standard name entries have a mutually exclusive tendency to occur.

[0043] Step S134: Construct standard field semantic core nodes based on the set of standard field name entries, map each standard field name entry to a standard field semantic core node, and assign a node semantic description vector to each standard field semantic core node. The node semantic description vector is generated based on the contextual semantic environment of the standard field name entry.

[0044] Each field standard name entry in the set of field standard name entries is mapped to a standard field semantic core node. The node identifier of the standard field semantic core node uses the text string of the field standard name entry. The node semantic description vector is generated using a pre-trained word embedding model in the credit reporting domain. The text of the field standard name entry is used as input to the word embedding model. After word segmentation and stemming, the word embedding model compresses the contextual co-occurrence information of each word into a multi-dimensional floating-point vector with a fixed dimension. The standard field semantic core node stores the node identifier and the node semantic description vector as core attributes.

[0045] Step S135: Establish a context co-occurrence association connection between any two of the standard field semantic core nodes, and use the field co-occurrence association strength value as the connection weight of the context co-occurrence association connection. When the field co-occurrence association strength value is lower than the preset connection weight threshold, the context co-occurrence association connection is not established.

[0046] Iterate through each pair of standard field name entries in the field co-occurrence frequency record table to find the corresponding two standard field semantic core nodes. Establish an undirected contextual co-occurrence association between the two standard field semantic core nodes, assigning the connection weight the value of the field co-occurrence association strength. A preset connection weight threshold is the minimum threshold for determining whether a connection is meaningful. If the field co-occurrence association strength value is lower than the preset connection weight threshold, skip the connection establishment operation, and no contextual co-occurrence association is established between the two standard field semantic core nodes. The filtered contextual co-occurrence associations constitute a set of connection edges.

[0047] Step S136: Construct an initial field semantic network based on the standard field semantic core node and the context co-occurrence association connection, perform community structure discovery and semantic topic extraction on the initial field semantic network, and generate a field semantic topology with a hierarchical community organization structure.

[0048] All standard field semantic core nodes are used as a set of network nodes, and all context co-occurrence associations are used as a set of network edges to jointly construct the initial field semantic network, which is stored in an adjacency list format. Community structure discovery and semantic topic extraction are then performed on the initial field semantic network to generate the final field semantic topology.

[0049] Step S1361: Perform community structure discovery processing on the initial field semantic network, identify the node clustering areas in the initial field semantic network where the connection density exceeds the preset community density, and classify the standard field semantic core nodes in the node clustering areas into the same field semantic function community.

[0050] The community structure discovery process employs the Louvain algorithm. The algorithm first initializes each standard field semantic core node as an independent community. Then, it sequentially attempts to move each node into the community of its neighboring nodes, calculating the modularity increment of the entire network before and after the move. The modularity increment is equal to the difference between the sum of edge weights within the new community and the expected sum of edge weights. A positive modularity increment is accepted, while a negative increment is reversed. This process is iteratively executed until the modularity no longer increases. Upon convergence, standard field semantic core nodes assigned to the same community identifier constitute a field semantic functional community.

[0051] Step S1362: Perform community semantic topic extraction processing on each of the field semantic function communities, extract the average value of the node semantic description vectors of all standard field semantic core nodes in the field semantic function community as the community semantic topic representation vector, and attach the community semantic topic representation vector to each standard field semantic core node contained in the field semantic function community.

[0052] Each field semantic function community contains several standard field semantic core nodes. The node semantic description vectors of all standard field semantic core nodes within this community are extracted. The floating-point values ​​in each vector dimension are then arithmetically averaged, resulting in a community semantic topic representation vector with the same dimensions as the node semantic description vectors. This community semantic topic representation vector represents the overall semantic tendency of the field semantic function community. The community semantic topic representation vector is then written as an additional attribute into the data structure of each standard field semantic core node within this community.

[0053] Step S1363: Based on the cross-community connection density between the field semantic function communities, establish a community-level semantic association edge between any two field semantic function communities. The community-level semantic association edge is used to represent the semantic mapping relationship between different credit reporting field subject domains.

[0054] Calculate the cross-community connection density between any two field semantic function communities. The cross-community connection density equals the sum of the connection weights of the context co-occurrence associations connecting the standard field semantic core nodes within the two communities, divided by the product of the number of standard field semantic core nodes within each community. When the cross-community connection density exceeds a preset cross-community connection density threshold, a community-level semantic association edge is established between the two field semantic function communities, and the weight of this edge is the cross-community connection density. The community-level semantic association edge connects independent field semantic function communities into a higher-level semantic network structure.

[0055] Step S1364: Integrate and encapsulate the standard field semantic core node with the attached community semantic topic representation vector, the context co-occurrence association connection, and the community hierarchical semantic association edge to generate a field semantic topology structure with a hierarchical community organizational structure.

[0056] The integration and encapsulation operation packages three types of data objects into a unified data structure. The first type of object consists of standard field semantic core nodes, carrying a node identifier, a node semantic description vector, and a community semantic topic representation vector. The second type of object consists of context co-occurrence associations, carrying a starting node identifier, a target node identifier, and connection weights. The third type of object consists of community-level semantic association edges, carrying a starting community identifier, a target community identifier, and cross-community connection density. These three types of objects are serialized and stored according to a predefined graph structure file format, generating a field semantic topology with a hierarchical community organizational structure. This field semantic topology is persistently stored and can be loaded and used multiple times in subsequent steps.

[0057] Step S140: Perform topological association mapping processing on each semantic unit in the original field semantic unit group and the standard field semantic core node to obtain a field alignment mapping scheme. Each mapping record in the field alignment mapping scheme records the corresponding association path between a semantic unit and a standard field semantic core node.

[0058] Step S141: Expand all semantic units in the original field semantic unit group into a list of semantic units to be mapped, perform semantic vector initialization processing on each semantic unit in the list of semantic units to be mapped, and call the pre-trained credit scoring domain semantic encoder to convert the semantic unit into a semantic unit vector.

[0059] The original field semantic unit group includes text fragment semantic unit groups, visual region semantic unit groups, and indicator item semantic unit groups. All semantic units in these three groups are expanded into a one-dimensional list of semantic units to be mapped. The pre-trained credit reporting domain semantic encoder is a deep bidirectional Transformer encoder based on the BERT architecture. Its pre-training corpus consists of large-scale corporate credit reporting report texts, and the training tasks are masked language modeling and next-sentence prediction. The text string of the semantic unit is taken as input, processed through multiple encoding layers of the semantic encoder, and the output vector at the first position of the last encoding layer is taken as the semantic unit vector. The dimension of the semantic unit vector is consistent with the dimension of the node semantic description vector.

[0060] Step S142: Extract the node semantic description vectors of all the standard field semantic core nodes from the field semantic topology structure, input the node semantic description vectors and the semantic unit vectors into a vector similarity comparator, and calculate the cosine value of the vector space angle between the semantic unit vector and each node semantic description vector as the semantic basic similarity.

[0061] The semantic topology of the loaded fields is traversed, and all standard field semantic core nodes are extracted to obtain the node semantic description vector for each standard field semantic core node. A vector similarity comparator performs cosine similarity calculation on the semantic unit vector and the node semantic description vector. The calculation process is as follows: the semantic unit vector and the node semantic description vector are multiplied element-wise along each dimension, and the results of all dimension multiplications are summed to obtain the vector dot product. The vector magnitudes of the semantic unit vector and the node semantic description vector are calculated separately; the vector magnitude is equal to the square root of the sum of the squares of the elements in each dimension. The semantic basic similarity is equal to the vector dot product divided by the product of the semantic unit vector magnitude and the node semantic description vector magnitude.

[0062] Step S143: Obtain the community semantic topic representation vector of the field semantic function community to which the standard field semantic core node belongs in the field semantic topology, and calculate the cosine value of the vector space angle between the semantic unit vector and the community semantic topic representation vector as the community semantic similarity.

[0063] For each standard field semantic core node, obtain the community semantic topic representation vector of its respective field semantic function community. Using the same cosine similarity calculation method as in step S142, calculate the cosine value of the vector space angle between the semantic unit vector and the community semantic topic representation vector. The calculation result is used as the community semantic similarity. The community semantic similarity measures the overall semantic fit between the semantic unit and the topic community to which the standard field semantic core node belongs.

[0064] Step S144: Obtain the credit data source type tag corresponding to the semantic unit, find the corresponding source confidence adjustment coefficient in the preset data source confidence configuration table according to the credit data source type tag, and perform weighted processing on the semantic basic similarity and the source confidence adjustment coefficient to obtain the adjusted semantic basic similarity.

[0065] Each semantic unit carries a credit data source type tag during generation. The preset data source confidence configuration table is a key-value mapping table, where the key is the credit data source type tag and the value is the source confidence adjustment coefficient. The confidence adjustment coefficient for structured numerical sources is usually higher than that for unstructured text sources. The semantic basic similarity is multiplied and weighted with the source confidence adjustment coefficient obtained from the query, and the product is used as the adjusted semantic basic similarity.

[0066] Step S145: The semantic basic similarity and the community semantic similarity are fused. The semantic basic similarity and the community semantic similarity are linearly superimposed according to a preset similarity fusion ratio to generate a comprehensive semantic matching score. The standard field semantic core nodes whose comprehensive semantic matching scores exceed a preset matching threshold are determined as candidate matching nodes. The candidate matching nodes are sorted from high to low according to the comprehensive semantic matching scores. The candidate matching node with the first position in the sorting is selected as the alignment target node of the semantic unit.

[0067] The preset similarity fusion ratio is a weighting coefficient between zero and one. The comprehensive semantic matching score equals the adjusted basic semantic similarity multiplied by the similarity fusion ratio, plus the difference between the community semantic similarity multiplied by one and the similarity fusion ratio. All standard field semantic core nodes are traversed, and the comprehensive semantic matching score between the semantic unit and each standard field semantic core node is calculated. Standard field semantic core nodes with comprehensive semantic matching scores exceeding a preset matching threshold are added to the candidate matching node list. The candidate matching node list is sorted in descending order of comprehensive semantic matching score, and the candidate matching node at the top of the list is determined as the alignment target node for the semantic unit.

[0068] Step S146: After determining the alignment target node of the semantic unit based on the comprehensive semantic matching score, perform path penalty verification and mapping record encapsulation processing on the alignment target node to generate a field alignment mapping scheme.

[0069] Step S1461: Obtain the topological path distance from the semantic unit to the alignment target node in the field semantic topology structure, perform shortest topological path search processing on the semantic unit in the field semantic topology structure, and calculate the path jump steps based on the number of context co-occurrence association connections passed through the path.

[0070] In the field semantic topology, the semantic unit is not yet a network node. It needs to be temporarily connected as a node to the top few standard field semantic core nodes with the highest comprehensive semantic matching score. Then, Dijkstra's shortest path search algorithm is used to calculate the shortest topological path from the semantic unit to the alignment target node. The number of context co-occurrence associations along the path is the path jump number. A smaller path jump number indicates a closer semantic distance between the semantic unit and the alignment target node in the topology.

[0071] Step S1462: Adjust the comprehensive semantic matching score according to the number of path jumps. When the number of path jumps exceeds the preset upper limit of path jumps, perform a reduction and attenuation process on the comprehensive semantic matching score according to the preset path penalty coefficient, and redetermine the alignment target node of the semantic unit.

[0072] The preset path jump limit is a threshold used to assess whether the mapped path is too long. When the path jump limit exceeds the preset limit, the overall semantic matching score is multiplied by a preset path penalty coefficient to obtain a decayed overall semantic matching score. The candidate matching node with the highest decayed overall semantic matching score is then selected from the candidate matching node list as the new alignment target node, and the path jump step calculation and penalty adjustment are repeated until the path jump step requirement is met or the candidate matching node list is empty.

[0073] Step S1463: Establish the mapping correspondence between the semantic unit and the finally determined alignment target node, and record the sequence of all intermediate standard field semantic core nodes traversed by the shortest topological path from the semantic unit to the alignment target node, as the mapping propagation path.

[0074] The final alignment target nodes are then mapped one-to-one with the semantic units. The shortest topological path search process is traced back, and all intermediate standard field semantic core nodes traversed along the path are sequentially arranged into a node sequence, which constitutes the mapping propagation path. This path records the complete trajectory of a semantic unit as it hops from the edge of the semantic space to the core standard field.

[0075] Step S1464: Encapsulate the semantic unit, the alignment target node, the mapping propagation path, and the comprehensive semantic matching score into a field alignment mapping record, and summarize the field alignment mapping records of all the semantic units to form the field alignment mapping scheme.

[0076] Each field alignment mapping record contains four data items: a semantic unit data item storing the text content or identifier of the semantic unit; an alignment target node data item storing the node identifier of the target standard field semantic core node; a mapping propagation path data item storing the sequence of intermediate standard field semantic core nodes; and a comprehensive semantic matching score data item storing the final comprehensive semantic matching score value. The field alignment mapping records of all semantic units in the list of semantic units to be mapped are aggregated into a mapping record set, which is then serialized into a field alignment mapping scheme file.

[0077] Step S150: Perform field fusion and reconstruction processing on each original field semantic unit in the multi-source heterogeneous credit data set according to the field alignment mapping scheme to generate a unified credit field structure for the target enterprise. In the unified credit field structure, multi-source semantic units under the same standard field semantic core node are merged into the same field container.

[0078] Step S151: Traverse each field alignment mapping record in the field alignment mapping scheme, extract the semantic units and the alignment target nodes in the field alignment mapping record, take the semantic units as semantic units to be merged, and take the alignment target nodes as merge container identifiers.

[0079] Read each field alignment mapping record in the field alignment mapping scheme one by one. Extract the semantic unit data item and the alignment target node data item from each record. Mark the semantic unit data item as a semantic unit to be merged, and mark the node identifier in the alignment target node data item as a merge container identifier.

[0080] Step S152: Using the merge container identifier as the key, search in the preset field container dictionary whether there is an existing field container corresponding to the merge container identifier. If there is no existing field container in the field container dictionary, create a new field container for the merge container identifier and initialize the new field container to an empty set state.

[0081] The default field container dictionary is a key-value pair data structure, where the key is the merge container identifier and the value is the field container object. The merge container identifier is used as the lookup key to search the field container dictionary. If the corresponding field container object is found, a reference to that object is returned. If not found, a new field container object is created, its internal data set is initialized to an empty set, and then the key-value pair is inserted into the field container dictionary using the merge container identifier as the key and the new field container object as the value.

[0082] Step S153: Place the semantic unit to be merged into the field container corresponding to the merge container identifier, and at the same time, attach the credit data source type mark corresponding to the semantic unit to be merged and the source field location information in the original credit data as merge metadata to the entry of the semantic unit to be merged in the field container.

[0083] The semantic units to be merged are added to the data set of the field container object. Simultaneously, the credit data source type marker and source field location information recorded in step S120 for the semantic units to be merged are extracted. The source field location information includes paragraph position index, sentence offset position, or cell row and column coordinates, etc. The credit data source type marker and source field location information are packaged into merge metadata and appended to the entry record of the semantic unit to be merged within the field container.

[0084] Step S154: Perform semantic consistency verification and conflict detection on all semantic units to be merged within each field container, and determine and standardize the field values ​​of the field containers to generate a unified credit field structure for the target enterprise.

[0085] Step S1541: Perform semantic consistency verification on all semantic units to be merged in each field container, extract the semantic unit vectors of all semantic units to be merged in the field container, calculate the cosine of the vector space angle between any two semantic unit vectors as the semantic deviation degree in the container, and count the proportion of deviation entries whose semantic deviation degree in the container exceeds the preset deviation threshold.

[0086] Iterate through all semantic units to be merged within the field container, extracting the semantic unit vector for each unit. Pair all semantic units to be merged within the field container and calculate the cosine of the vector space angle between the semantic unit vectors of each pair. This cosine value is used as the semantic deviation degree within the container for that pair. Compare the semantic deviation degrees within all containers one by one with a preset deviation threshold, and record the number of deviation entries exceeding the preset deviation threshold. The percentage of deviation entries is equal to the number of deviation entries divided by the total number of pairwise pairwise pairings.

[0087] Step S1542: When the proportion of the deviation entries exceeds the preset fault tolerance ratio, the semantic units in the field container are subjected to outlier entry removal processing. The semantic unit with the largest average semantic deviation from other semantic units is removed from the field container, and the removed semantic unit is added back to the list of semantic units to be mapped.

[0088] The preset fault tolerance ratio is the upper limit for judging semantic consistency within the field container. When the proportion of deviating entries exceeds the preset fault tolerance ratio, it is determined that there are outliers that are incorrectly merged within the field container. The arithmetic mean of the semantic deviation between each semantic unit to be merged and all other semantic units to be merged within the field container is calculated as the average semantic deviation of that semantic unit. The semantic unit to be merged with the largest average semantic deviation is removed from the field container, and the removed semantic unit is put back into the list of semantic units to be mapped, waiting for the topological association mapping process to be repeated.

[0089] Step S1543: Perform source field value content conflict detection processing on the remaining semantic units in the field container after the outlier entry removal process. When the original field value content corresponding to the semantic units from different credit data sources in the field container is inconsistent, perform weighted voting processing on the original field value content according to the source confidence adjustment coefficient of each credit data source in the data source confidence configuration table, and take the original field value content with the highest confidence after weighted voting as the fused field value of the field container.

[0090] For all remaining semantic units to be merged within the field container, extract the original field value content corresponding to each semantic unit. The original field value content is the specific numerical value or text carried by the semantic unit in the original credit data. Compare the original field value contents of all semantic units to be merged. When inconsistencies exist, look up the source confidence adjustment coefficient in the data source confidence configuration table using the credit data source type tag as the key. The weighted voting process sums the source confidence adjustment coefficients of semantic units with the same original field value content, and the original field value content with the highest sum becomes the merged field value.

[0091] Step S1544: When the original field value content corresponding to the semantic unit of different credit data sources is numerical and the numerical difference exceeds the preset numerical deviation tolerance, the weighted voting process is not performed, and the original field value content is retained in the fused field value of the field container in the form of an array of data source labels.

[0092] When the original field value is numeric and the differences between the values ​​from different sources exceed a preset tolerance for numerical deviation, it indicates a substantial conflict between the values ​​from different data sources. Weighted voting is no longer applicable. The merged field value is stored using an array data structure, where each element contains the original field value and its corresponding credit data source type flag.

[0093] Step S1545: When the field container contains only semantic units from a single credit data source, the original field value content corresponding to the semantic unit is directly used as the fused field value of the field container, and a single data source marker is added to the fused field value of the field container.

[0094] When all semantic units to be merged within a field container have the same credit data source type tag, conflict detection and weighted voting are unnecessary. The original field value of the unique semantic unit is directly assigned to the merged field value, and a single source tag is appended to the metadata of the merged field value to indicate that the field value has not undergone multi-source cross-validation.

[0095] Step S1546: Standardize the field names of each field container, use the merge container identifier corresponding to the field container as the standard field name of the field container, combine the standard field name with the fusion field value of the field container and the merge metadata to form a standard credit field record, summarize the standard credit field records generated by all the field containers, sort and combine the standard credit field records according to the standard field arrangement order in the enterprise credit field to generate the unified credit field structure of the target enterprise.

[0096] The merge container identifier of the field container is directly used as the standard field name. Each field container generates a standard credit reporting field record, which contains three data blocks: the standard field name, the merged field value, and the merged metadata. After all standard credit reporting field records are aggregated, they are sorted according to the standard field arrangement order in the enterprise credit reporting field to form a structured unified credit reporting field structure. The unified credit reporting field structure is stored in a key-value pair document format, where the key is the standard field name, the value is the merged field value, and the metadata area stores the merged metadata.

[0097] For example, step S210: Construct a multi-level semantic alignment verification mechanism, perform mapping accuracy verification processing on the field alignment mapping records in the field alignment mapping scheme, extract the original field semantic context window corresponding to the semantic unit from the field alignment mapping record, and the original field semantic context window contains at least two semantic units that are adjacent to the semantic unit in the original credit data as context semantic units.

[0098] After the field alignment mapping scheme is generated, the semantic coherence of the mapping results needs to be verified. For each field alignment mapping record, the source field position information of the semantic unit is used to locate it in the original credit data. Taking the semantic unit position as the center, two adjacent semantic units are taken forward and backward to form an original field semantic context window containing five semantic units. The other four semantic units in the window are used as context semantic units.

[0099] Step S220: Input the semantic units in the original field semantic context window and the context semantic units into the context semantic coherence analyzer, extract the semantic transition probability features between the semantic units and the context semantic units, and generate the original context semantic coherence vector.

[0100] The context semantic coherence analyzer employs a corpus-based bigram model. This model pre-calculates the transition probabilities between adjacent semantic units of any two fields on a large number of credit investigation texts. The sequence of semantic units within the original field semantic context window is input into the bigram model, and the semantic transition probability between each pair of adjacent semantic units is calculated sequentially. All transition probability values ​​within the context window are then concatenated in order into a fixed-length original context semantic coherence vector.

[0101] Step S230: Based on the mapping propagation path in the field alignment mapping record, extract the sequence of intermediate standard field semantic core nodes that the mapping propagation path passes through from the field semantic topology, and combine the sequence of intermediate standard field semantic core nodes with the mapping propagation path to form a path semantic chain.

[0102] The mapping propagation path is read from the field alignment mapping record. This path contains a series of intermediate standard field semantic core nodes. The node identifiers of these intermediate standard field semantic core nodes are arranged in the order of the path, and together with the alignment start node at the beginning of the semantic unit and the alignment target node at the end of the path, they form the path semantic chain.

[0103] Step S240: Extract the alignment target node and its adjacent standard field semantic core node in the field semantic function community from the field semantic topology, and combine the node semantic description vectors of the alignment target node and the adjacent standard field semantic core node to generate the target context semantic coherence vector.

[0104] Locate the alignment target node in the field semantic topology and obtain its corresponding field semantic function community. Within this field semantic function community, select the adjacent standard field semantic core nodes that are directly connected to the alignment target node. Combine the node semantic description vector of the alignment target node with the node semantic description vectors of all adjacent standard field semantic core nodes according to topological distance to generate the target context semantic coherence vector.

[0105] Step S250: Perform vector similarity comparison processing between the original context semantic coherence vector and the target context semantic coherence vector, calculate the spatial distance as the semantic coherence deviation value, and when the semantic coherence deviation value exceeds the preset coherence deviation tolerance threshold, determine that the field alignment mapping record has a potential mapping deviation, and mark the field alignment mapping record with the potential mapping deviation as a mapping record to be reviewed.

[0106] The Euclidean distance between the original context semantic coherence vector and the target context semantic coherence vector is calculated using the square root of the sum of the squares of the differences in each dimension of the two vectors. This distance value is used as the semantic coherence deviation value. The semantic coherence deviation value is compared with a preset coherence deviation tolerance threshold. If the value exceeds the threshold, the corresponding field alignment mapping record is marked as a mapping record to be reviewed.

[0107] Step S260: Perform mapping path backtracking correction processing on the mapping record to be reviewed. Extract multiple candidate standard field semantic core nodes with the highest semantic basic similarity to the semantic unit from the field semantic topology as candidate backtracking nodes. Substitute the candidate backtracking nodes into the original field semantic context window respectively, and calculate the simulated context semantic coherence deviation value corresponding to each candidate backtracking node.

[0108] For the semantic units in the mapping record to be reviewed, candidate standard field semantic core nodes are searched again from the field semantic topology, sorted in descending order of semantic basic similarity, and the top few are selected as candidate backtracking nodes. Each candidate backtracking node is assumed to be the alignment target node, and the corresponding simulated context semantic coherence deviation value is calculated according to the process of steps S230 to S250.

[0109] Step S270: Select the candidate backtracking node with the smallest simulated context semantic coherence deviation value as the correction alignment target node, replace the original alignment target node in the mapping record to be reviewed with the correction alignment target node, update the mapping propagation path, re-incorporate the corrected mapping record to be reviewed into the field alignment mapping scheme, and trigger the re-execution operation of the field fusion reconstruction process based on the updated field alignment mapping scheme to generate the updated unified credit field structure.

[0110] Compare the simulated context semantic coherence deviation values ​​corresponding to all candidate backtracking nodes, and select the candidate backtracking node with the smallest deviation value as the corrected alignment target node. Replace the alignment target node in the mapping record to be reviewed with the corrected alignment target node, and recalculate the mapping propagation path. Replace the original record with the corrected field alignment mapping record and incorporate it into the field alignment mapping scheme. Based on the updated field alignment mapping scheme, trigger the re-execution of step S150 to generate a unified credit reporting field structure with better semantic coherence.

[0111] Step S310: Obtain a historical version credit data snapshot for each credit data source in the multi-source heterogeneous credit data set. The historical version credit data snapshot contains credit data records collected in the past time period corresponding to the current version credit data in the multi-source heterogeneous credit data set.

[0112] Credit data providers may periodically adjust the field structure and naming of their credit reports. To detect version changes in field structure, snapshots of historical versions of credit data collected by each credit data source at certain points in the past are obtained. The field structure of historical versions of credit data snapshots corresponds to, but may differ from, the current version of credit data.

[0113] Step S320: Perform the field semantic unit extraction process on the historical version credit data snapshot to obtain the historical version original field semantic unit group. Perform field structure change detection processing on the historical semantic units in the historical version original field semantic unit group and the corresponding current semantic units in the original field semantic unit group to generate field structure change records. Parse the field structure change records to determine the field splitting change path between the historical semantic unit and the current semantic unit. The field splitting change path records the splitting relationship of a historical field semantic unit being split into multiple current field semantic units, or the merging relationship of a current field semantic unit being merged from multiple historical field semantic units.

[0114] The historical version of the credit data snapshot undergoes step S120's field semantic unit extraction process to obtain the original historical version field semantic unit group. The historical semantic unit set is then compared with the current semantic unit set. When the text content of one historical semantic unit overlaps with the text content of multiple current semantic units, it is recorded as a split relationship. When the text content of multiple historical semantic units merges into the text content of one current semantic unit, it is recorded as a merge relationship. The split and merge relationships are collectively referred to as the field split change path.

[0115] Step S330: Update the field semantic topology according to the field splitting change path. When it is detected that the field splitting change path belongs to a splitting relationship, add the context co-occurrence association connection to the standard field semantic core node corresponding to the current field semantic unit obtained by splitting in the field semantic topology.

[0116] When a field splitting change path involves a splitting relationship, the standard field semantic core node corresponding to each of the multiple current field semantic units split out is located in the field semantic topology. Contextual co-occurrence associations are added between these standard field semantic core nodes, with high initial weights assigned to the connections, indicating a strong association derived from the same historical field.

[0117] Step S340: When it is detected that the field splitting change path belongs to a merging relationship, a merging type semantic association edge is established between the standard field semantic core nodes corresponding to the multiple historical field semantic units of the merging source and the standard field semantic core nodes corresponding to the current field semantic unit of the merging target in the field semantic topology.

[0118] When a field splitting change path involves a merge relationship, the standard field semantic core node corresponding to the historical field semantic unit of the merge source and the standard field semantic core node corresponding to the current field semantic unit of the merge target are located in the field semantic topology. A directed merge-type semantic association edge is established between them, with the edge direction pointing from the source node to the target node, and the edge type attribute is marked as merge.

[0119] Step S350: Store the updated field semantic topology structure as a new version of the field semantic topology structure, and re-execute the topology association mapping process on the current version of the credit data in the multi-source heterogeneous credit data set to generate an updated version of the field alignment mapping scheme that is compatible with the new version of the field semantic topology structure.

[0120] The updated field semantic topology is saved with the new version number. The entire topology association mapping process of step S140 is re-executed on the current version of the credit data to generate an updated version of the field alignment mapping scheme that is compatible with the new version of the field semantic topology.

[0121] Step S360: According to the updated version of the field alignment mapping scheme, perform field structure change synchronization processing on the unified credit reporting field structure, identify the standard field name in the unified credit reporting field structure that has changed the alignment target node due to the update of the field semantic topology structure, perform content migration operation on the field container under the standard field name, and record the content change information of the field container involved in the content migration operation as a field version evolution log. The field version evolution log includes the field change timestamp, the standard field name before the change, the standard field name after the change, and the migrated fused field value content.

[0122] Compare the original and updated versions of the field alignment mapping scheme to identify semantic units where the alignment target node has changed. Perform content migration operations on the affected field containers in the unified credit reporting field structure, migrating the fused field values ​​within the field containers from the old standard field names to the new standard field names. All change information during the migration process is written to the field version evolution log in chronological order.

[0123] Step S410: In the field semantic topology, establish a field usage frequency counter for each standard field semantic core node, and record the cumulative number of times the standard field semantic core node is determined as an alignment target node during the topology association mapping process.

[0124] The field usage frequency counter is an integer variable bound to each standard field semantic core node, with an initial value of zero. During each round of topology association mapping, whenever a standard field semantic core node is determined to be the alignment target node for a certain semantic unit, its field usage frequency counter is incremented by one.

[0125] Step S420: Obtain the count value of the field usage frequency counter of all standard field semantic core nodes contained in each field semantic function community in the field semantic topology structure, sort the count value within the community, and determine the standard field semantic core node with the highest usage frequency in each field semantic function community as the high-frequency core node of the community.

[0126] Iterate through each field semantic function community and collect the field usage frequency counter values ​​of all standard field semantic core nodes within the community. Sort the counter values ​​in descending order, and mark the standard field semantic core node at the top of the sorted position as a high-frequency core node in the community.

[0127] Step S430: Extract the node semantic description vector of the high-frequency core node of the community, perform semantic gravity propagation processing on the node semantic description vector of the high-frequency core node of the community to other standard field semantic core nodes in the field semantic function community, and adjust the node semantic description vector of the other standard field semantic core nodes according to the topological distance attenuation coefficient between nodes in the field semantic topology.

[0128] Semantic gravity propagation processing uses the semantic description vectors of high-frequency core nodes in the community as the gravity source. For other standard field semantic core nodes within the community, the shortest path steps between them and the high-frequency core nodes in the field semantic topology are calculated as the topological distance. The topological distance decay coefficient is a negative topological distance raised to the power of the natural constant. The semantic description vector of each other standard field semantic core node is adjusted to the original node semantic description vector plus the sum of the node semantic description vector of the high-frequency core nodes in the community multiplied by the topological distance decay coefficient.

[0129] Step S440: Perform topology adaptive adjustment processing on the field semantic topology structure after the semantic gravity propagation processing. When the semantic similarity between the high-frequency core nodes of the two field semantic function communities exceeds the preset community fusion threshold, merge the two field semantic function communities into a fused field semantic function community, and recalculate the community semantic topic representation vector of the fused field semantic function community.

[0130] Calculate the semantic similarity between the high-frequency core nodes of any two field semantic function communities, using the cosine similarity of the semantic description vectors of the two nodes. When the semantic similarity exceeds a preset community fusion threshold, merge the two field semantic function communities. The merged field semantic function community contains all the standard field semantic core nodes of the original two communities, and recalculate the community semantic topic representation vector.

[0131] Step S450: When a low-frequency standard field semantic core node appears within the field semantic function community, and the count value of the field usage frequency counter is continuously lower than a preset low-frequency threshold, and the semantic similarity of the low-frequency standard field semantic core node with the community semantic topic representation vector of other field semantic function communities exceeds a preset cross-community migration threshold, the low-frequency standard field semantic core node is migrated from the current field semantic function community to the target field semantic function community with the highest semantic similarity.

[0132] The frequency of field usage for each standard field semantic core node is counted over multiple consecutive topological association mapping processing cycles. When the count value consistently falls below a preset low-frequency threshold, it is marked as a low-frequency standard field semantic core node. The semantic similarity between the node semantic description vector of the low-frequency standard field semantic core node and the community semantic topic representation vector of other field semantic function communities is calculated. The field semantic function community with the highest semantic similarity and exceeding a preset cross-community migration threshold is selected as the migration destination. The community affiliation of the low-frequency standard field semantic core node is then migrated from its original community to the target community.

[0133] Step S460: Perform community semantic topic representation vector recalculation processing on the original field semantic function community and the target field semantic function community after the migration of the low-frequency standard field semantic core node, and update the connection weight of the community-level semantic association edge between the original field semantic function community and the target field semantic function community.

[0134] After migrating the semantic core nodes of low-frequency standard fields, the members of both the original and target communities change, and the community semantic topic representation vectors of the two communities are recalculated. The connection weights of the community-level semantic association edges between the original and target communities are recalculated and updated based on the cross-community connection density after the migration.

[0135] Step S470: Based on the adaptively adjusted field semantic topology, the topology association mapping process is re-executed on the multi-source heterogeneous credit data set of the target enterprise subsequently acquired. The updated standard field semantic core nodes and the context co-occurrence association connection are used to generate a field alignment mapping scheme with dynamic adaptive characteristics.

[0136] The adaptively adjusted field semantic topology is persistently saved as the baseline topology for all subsequent mapping processes. When a new multi-source heterogeneous credit data set is input, step S140 is executed using this updated field semantic topology to generate a field alignment mapping scheme that can adapt to changes in field usage frequency and community structure evolution.

[0137] Step S510: Obtain the external business scenario identification code of the target enterprise. The external business scenario identification code is used to indicate the credit assessment business scenario, supply chain finance business scenario, or insurance underwriting business scenario that the target enterprise is currently participating in.

[0138] The external business scenario identifier is passed in by the business system that calls this method. Its value is an enumeration value, including credit assessment identifier, supply chain finance identifier, insurance underwriting identifier, etc.

[0139] Step S520: Extract the list of necessary credit fields corresponding to the business scenario from the preset business scenario requirement mapping table according to the external business scenario identifier code. The list of necessary credit fields contains a sequence of standard names of credit fields that must be filled in under the business scenario.

[0140] A pre-defined business scenario requirement mapping table establishes a mapping relationship between business scenario identifiers and a list of necessary credit information fields. A query in the table using the external business scenario identifier as the key returns a sequence of standard names of the credit information fields that must be filled in under that scenario.

[0141] Step S530: Extract standard credit field records corresponding to the standard name of each credit field in the list of necessary credit fields from the unified credit field structure, and combine the extracted standard credit field records into a subset of business scenario fields.

[0142] Iterate through the list of necessary credit reporting fields, searching for the standard credit reporting field record in the unified credit reporting field structure using the standard field name as the key. Store the retrieved records into a subset of business scenario fields.

[0143] Step S540: Detect whether there are standard credit reporting field records in the business scenario field subset where the field container is empty or the fusion field value of the field container is accompanied by the single data source marker and the source confidence adjustment coefficient is lower than the preset business scenario confidence requirement, and determine the detected standard credit reporting field records as credit reporting field missing items.

[0144] Each standard credit reporting field record in the business scenario field subset is examined one by one. When the data set of the field container is empty, or when the fused field value carries a single data source marker and the corresponding source confidence adjustment coefficient is lower than the preset business scenario confidence requirement, the standard credit reporting field record is marked as a missing credit reporting field.

[0145] Step S550: Perform missing field tracing processing on the alignment target node corresponding to the missing item of the credit field, and search for the credit data source identifier and corresponding semantic unit that the alignment target node has been successfully mapped in the historical execution record of the field alignment mapping scheme.

[0146] Based on the standard field names of missing items in the credit reporting field, the target node is aligned by tracing back to the semantic topology of the field. The historical execution records of the field alignment mapping scheme are searched for semantic unit records of the target node that were successfully mapped in previous mapping cycles, and the corresponding credit reporting data source identifiers are extracted.

[0147] Step S560: Generate a credit data supplementation instruction based on the credit data source identifier. The credit data supplementation instruction includes the credit data source identifier, the standard field name corresponding to the alignment target node, and the unique enterprise identifier of the target enterprise. Send the credit data supplementation instruction to the corresponding third-party credit reporting platform interface.

[0148] The credit data supplementation instruction is encapsulated into a request data packet of a specified format. The data packet contains the credit data source identifier corresponding to the third-party credit reporting platform's interface address, the standard field names of the fields to be supplemented, and the target company's unique identifier. This instruction is sent to the corresponding third-party credit reporting platform interface via a secure network channel.

[0149] Step S570: Receive the credit supplementary data returned by the third-party credit reporting platform interface, perform the field semantic unit extraction processing and the topological association mapping processing on the credit supplementary data to obtain supplementary field alignment mapping records for the missing credit fields, input the supplementary field alignment mapping records into the field fusion and reconstruction processing process to generate supplementary standard credit field records for the missing credit fields, and merge the supplementary standard credit field records into the business scenario field subset to form a complete business scenario credit field set.

[0150] The credit data collected from the third-party credit reporting platform interface is processed through step S120 (field semantic unit extraction) and step S140 (topology association mapping) to generate a supplementary field alignment mapping record. This record is then processed through step S150 (field fusion and reconstruction) to generate a supplementary standard credit reporting field record. This record is then merged into the corresponding position of the business scenario field subset to form a complete set of business scenario credit reporting fields.

[0151] Step S580: Generate a compliance credit report for the target enterprise in the business scenario based on the complete set of credit fields for the business scenario. Each field item in the compliance credit report corresponds one-to-one with the standard name of the credit field in the list of necessary credit fields.

[0152] The complete set of credit reporting fields for business scenarios is arranged in the order of the necessary credit reporting field list to generate a formatted compliant credit reporting report document. The field items in the report correspond exactly to the standard names of the credit reporting fields in the list, and all field values ​​have been verified by multi-source fusion or supplemented by data collection.

[0153] Step S610: Obtain the data storage certificate issued by the third-party credit reporting platform for each credit data record in the multi-source heterogeneous credit data set. The data storage certificate includes the credit data source identifier, the data collection timestamp, and the data content hash value.

[0154] The data storage certificate is issued simultaneously by the third-party credit reporting platform when the data is retrieved, and digital signature technology is used to ensure that the certificate cannot be tampered with. The certificate records the credit data source identifier, the data collection timestamp, and the hash value of the data content calculated using the SM3 hash algorithm.

[0155] Step S620: Perform credential verification processing on the data storage credential, extract the data content hash value from the data storage credential, compare the data content hash value with the real-time content hash value of the corresponding credit data record in the multi-source heterogeneous credit data set, and generate a data integrity anomaly alarm message if the two are inconsistent.

[0156] For each credit data record, the real-time content hash value is recalculated using the same algorithm as that in the data storage certificate. The real-time content hash value is then compared bit-by-bit with the data content hash value in the data storage certificate. If the comparison results are inconsistent, a data integrity anomaly alarm is generated, preventing that credit data record from entering the subsequent processing flow.

[0157] Step S630: When the data integrity abnormality alarm information is not generated, extract the fusion field value and its corresponding merged metadata of each field container from the field fusion reconstruction process, package the fusion field value and the merged metadata into a field fusion traceability data package, perform traceability hash calculation on the field fusion traceability data package to generate the fusion traceability hash value corresponding to the field fusion traceability data package, and perform Merkle tree construction processing on the fusion traceability hash value and the data content hash values ​​of all credit data records participating in the field fusion reconstruction process to generate the root value of the field fusion traceability Merkle tree.

[0158] Once all data storage credentials have passed integrity verification, a field fusion traceability data package is constructed for each field container. This data package contains the fused field values ​​and merged metadata. A fusion traceability hash value is calculated for the data package. Using all fusion traceability hash values ​​of the same target enterprise and the data content hash values ​​of all original credit data records as leaf nodes of a Merkle tree, the parent node hash values ​​are calculated layer by layer upwards in pairs, ultimately yielding a unique field fusion traceability Merkle tree root value.

[0159] Step S640: Write the Merkel root value of the field fusion traceability and the standard field name of each standard credit field record in the unified credit field structure into the extended storage record area of ​​the data storage certificate to generate an enhanced data storage certificate. Synchronize the enhanced data storage certificate to multiple storage consensus nodes of the distributed storage network. Perform network consensus confirmation processing on the enhanced data storage certificate through the consensus protocol of the distributed storage network. When more than the preset number of consensus confirmation nodes are obtained, generate a field fusion traceability storage block.

[0160] The extended evidence record area of ​​the data evidence certificate allows the writing of custom key-value pairs. The field fusion traceability Merkle tree root value and a list of standard field names are written into this area to generate an enhanced data evidence certificate. This enhanced data evidence certificate is broadcast to multiple consensus nodes in the distributed evidence network. The consensus nodes run a Byzantine fault-tolerant consensus protocol to vote on and confirm the enhanced data evidence certificate. When the number of confirmed signatures collected exceeds the preset number of consensus confirmation nodes, the enhanced data evidence certificate and its confirmed signatures are packaged into a field fusion traceability evidence block and appended to the blockchain interface of the distributed evidence network.

[0161] Step S650: When querying the fusion field value of any standard credit field record in the unified credit field structure, the field fusion traceability evidence storage block is obtained synchronously from the distributed evidence storage network. Field value traceability verification is performed based on the field fusion traceability Merkle root value in the field fusion traceability evidence storage block and the current fusion traceability hash value of the standard credit field record. When the field value traceability verification fails, a field value traceability verification failure notification is output. The field value traceability verification failure notification includes the standard field name of the standard credit field record that failed verification and a list of all credit data source identifiers participating in the fusion of the standard credit field record.

[0162] The query request triggers the source tracing verification process. The latest field-fused source tracing evidence block corresponding to the target enterprise is retrieved from the distributed evidence storage network. The fused field value and merged metadata of the currently queried standard credit field record are recalculated to obtain the fused source tracing hash value. This hash value is then compared with the Merkel root value of the field fused source tracing tree for Merkel verification path comparison. If verification fails, a field value source tracing verification failure notification is output, listing the names of the standard fields that failed verification and a list of all credit data source identifiers involved in the fusion.

[0163] Step S710: Obtain the credit reporting objection appeal request submitted by the target enterprise. The credit reporting objection appeal request includes the standard name of the objection field and a statement of correction of the original value of the objection field.

[0164] Credit reporting objection appeals can be submitted via the front-end interface or application programming interface (API). The request body should include the standard name of the objected field and a statement from the target company correcting the original value of the objected field.

[0165] Step S720: Locate the target standard credit field record in the unified credit field structure according to the objection field standard name, and extract the merged metadata of the target standard credit field record. The merged metadata records all the original field semantic units and their credit data source identifiers that participated in generating the fused field value of the target standard credit field record.

[0166] Key-value matching is performed within the unified credit reporting field structure using the standard name of the objection field to locate the record of the target standard credit reporting field. All original field semantic units and their corresponding credit reporting data source identifier lists are extracted from the merged metadata of this record.

[0167] Step S730: Extract all field alignment mapping records associated with the target standard credit reporting field record from the field alignment mapping scheme, parse the mapping propagation path in each field alignment mapping record one by one, and obtain the sequence of intermediate standard field semantic core nodes through which the mapping propagation path passes.

[0168] Search the field alignment mapping scheme for all field alignment mapping records where the alignment target node is equal to the standard field semantic core node corresponding to the standard name of the objection field. Parse the mapping propagation paths in these records one by one, and expand the intermediate standard field semantic core nodes along each path in the order of the path.

[0169] Step S740: Perform mapping divergence analysis on the intermediate standard field semantic core node sequence. Mark the standard field semantic core node where the alignment mapping record of fields from different credit data sources first appears on the mapping propagation path as the mapping divergence node. Extract the field semantic function community to which the mapping divergence node belongs in the field semantic topology. Obtain the community semantic theme representation vector of the field semantic function community. Use the community semantic theme representation vector to perform semantic stance verification on the original value correction statement of the objection field to determine whether there is a semantic stance conflict between the original value correction statement of the objection field and the community semantic theme representation vector.

[0170] Starting from the beginning of the field semantic topology, the mapping propagation paths of different credit data sources are compared hop-by-hop. The first standard field semantic core node where a path fork occurs is marked as a mapping divergence node. The community semantic topic representation vector of the field semantic function community to which the mapping divergence node belongs is extracted. The text of the correction statement of the original value of the objection field is input into the pre-trained credit domain semantic encoder to generate a correction statement semantic vector. The cosine similarity between the correction statement semantic vector and the community semantic topic representation vector is calculated. If the cosine similarity is lower than a preset stance consistency threshold, a semantic stance conflict is determined, and the verification fails.

[0171] Step S750: When the original value correction statement of the objection field passes the semantic position verification process, perform a mapping path re-exploration process along the topological edge direction of the field semantic topology structure from the mapping divergence node, generate an alternative mapping propagation path from the corrected semantic unit corresponding to the original value correction statement of the objection field to the standard field semantic core node, replace the original mapping propagation path corresponding to the target standard credit field record with the alternative mapping propagation path, and update the affected field alignment mapping records in the field alignment mapping scheme based on the alternative mapping propagation path.

[0172] After the semantic stance verification is passed, starting from the mapping divergence point node, the standard field semantic core nodes are explored outward along the topological edge direction. The standard field semantic core node that best matches the semantics of the correction statement semantic vector is found, and an alternative mapping propagation path is generated. The original mapping propagation path of the target standard credit field record is replaced with the alternative mapping propagation path, and the downstream affected field alignment mapping records in the field alignment mapping scheme are updated in a cascading manner.

[0173] Step S760: Trigger the targeted re-execution operation of the field fusion reconstruction process for the updated field alignment mapping scheme, and regenerate the fusion field value only for the field container corresponding to the target standard credit field record, while keeping the field container content of other standard credit field records unchanged.

[0174] The targeted re-execution operation locks the field container corresponding to the target standard credit field record, and only re-executes the fusion field value generation process of steps S153 to S1545 for that field container. The contents of other field containers in the unified credit field structure remain unchanged, avoiding the performance overhead of full reconstruction.

[0175] Step S770: Update the target standard credit field record in the unified credit field structure according to the regenerated fusion field value, and write the original mapping propagation path, the alternative mapping propagation path and the original value correction statement of the objection field into the objection correction log. Re-output the credit report document of the target enterprise based on the updated unified credit field structure. The field value of the target standard credit field record in the credit report document comes from the regenerated fusion field value.

[0176] In the unified credit reporting field structure, the merged field values ​​of the target standard credit reporting field records are overwritten by the regenerated merged field values. The original mapping propagation path, the alternative mapping propagation path, and the original value correction statement for the objection field are written as audit records to the objection correction log. In the re-output credit report document, the corrected field values ​​are displayed in the corresponding positions of the objection fields.

[0177] Step S810: Obtain a list of upstream and downstream related enterprises in the supply chain network of the target enterprise, and perform the operation of obtaining the multi-source heterogeneous credit data set of the target enterprise on each upstream and downstream related enterprise in the list to obtain the multi-source heterogeneous credit data set of each upstream and downstream related enterprise.

[0178] Obtain a list of upstream and downstream related enterprises of the target enterprise from the supply chain management platform. Iterate through each upstream and downstream related enterprise in the list, and after obtaining authorization and consent from the related enterprise, execute the data retrieval process in step S110 to obtain a multi-source heterogeneous credit data set of each related enterprise.

[0179] Step S820: Perform the field semantic unit extraction process independently on the multi-source heterogeneous credit data set of each of the upstream and downstream related enterprises to obtain the original field semantic unit group of each of the upstream and downstream related enterprises.

[0180] Each related enterprise's multi-source heterogeneous credit data set independently executes the field semantic unit extraction processing flow of step S120 to generate its own original field semantic unit group for related enterprises.

[0181] Step S830: Compare the original field semantic unit groups of each of the upstream and downstream related enterprises with the original field semantic unit groups of the target enterprise, identify the shared field semantic units that appear in the target enterprise and more than a preset number of upstream and downstream related enterprises, and mark the shared field semantic units as supply chain common field semantic units.

[0182] Extract the text strings of all semantic units from the target company's original field semantic unit group as a baseline set. Iterate through the original field semantic unit group of each associated company and record the semantic units that match the baseline set. Count the total number of times each matching semantic unit appears in the associated company list. Semantic units that appear more than a preset number of times are marked as supply chain common field semantic units.

[0183] Step S840: Add a supply chain shared node to the field semantic topology for the common field semantic unit of the supply chain. The supply chain shared node inherits the existing context co-occurrence association connection of the common field semantic unit of the supply chain in the field semantic topology and attaches a supply chain level attribute tag.

[0184] In the field semantic topology, a new supply chain shared node is created for each common field semantic unit of the supply chain. The supply chain shared node replicates all context co-occurrence associations of the original standard field semantic core node and adds a supply chain level attribute tag, the tag value of which is an identifier representing its belonging to the supply chain level dimension.

[0185] Step S850: Obtain the unified credit field structure of each of the upstream and downstream related enterprises, extract the standard credit field records corresponding to the common field semantic units of the supply chain in the unified credit field structure of each of the upstream and downstream related enterprises, form a supply chain field value distribution set, perform field value deviation pattern mining processing on the supply chain field value distribution set, use the fusion field value of the target enterprise as the benchmark center value, calculate the field value offset direction and offset magnitude between the corresponding fusion field value of each of the upstream and downstream related enterprises and the benchmark center value, and generate a field value deviation pattern vector.

[0186] Perform the complete process from steps S130 to S150 on each related enterprise to obtain their respective unified credit information field structures. Extract the fused field values ​​of the standard credit information field records corresponding to the semantic units of the common fields in the supply chain. Using the fused field value of the target enterprise as the benchmark center value, calculate the difference between the fused field value of each related enterprise and the benchmark center value. The sign of the difference indicates the offset direction, and the absolute value of the difference indicates the offset magnitude. Combine the offset direction and offset magnitude into a field value deviation pattern vector.

[0187] Step S860: Perform supply chain transmission path analysis on the field value deviation pattern vector. Divide the upstream and downstream related enterprises into direct transaction level and indirect transaction level according to their supply chain transaction distance with the target enterprise. Compare the aggregation characteristics of the field value deviation pattern vectors of enterprises in the direct transaction level and the indirect transaction level to identify the transmission law of field value deviation along the supply chain. Construct a supply chain field value deviation transmission graph based on the transmission law. In the supply chain field value deviation transmission graph, the shared nodes of the supply chain are used as node units, the supply chain transaction relationship between the upstream and downstream related enterprises is used as the transmission edge, and the field value deviation pattern vector is used as the edge weight of the transmission edge.

[0188] Based on the distance of supply chain transaction relationships, related enterprises are divided into direct transaction levels and indirect transaction levels. The statistical characteristics of the field value deviation pattern vectors of enterprises in the direct transaction level are compared with those of enterprises in the indirect transaction level to identify the progressive amplification or reduction of deviation magnitudes. A supply chain field value deviation transmission graph is constructed using shared nodes in the supply chain as graph nodes, inter-enterprise supply chain transaction relationships as directed transmission edges, and field value deviation pattern vectors as edge weights.

[0189] Step S870: When a newly added enterprise to the supply chain network submits credit data, the original field semantic unit group of the enterprise to be evaluated is matched with the common field semantic unit of the supply chain. After successful matching, the comprehensive semantic matching score in the field alignment mapping scheme of the enterprise to be evaluated is adjusted for supply chain context bias based on the field value deviation pattern vector of the adjacent enterprises in the supply chain position of the enterprise to be evaluated in the supply chain field value deviation transmission map. A unified credit field structure of the enterprise to be evaluated is generated based on the field alignment mapping scheme after the supply chain context bias adjustment, and the unified credit field structure of the enterprise to be evaluated is incorporated into the supply chain field value distribution set to update the supply chain field value deviation transmission map.

[0190] After a newly added enterprise submits its credit data, step S120 extracts the original semantic unit group of the field to be evaluated and matches it with the common semantic units of the fields in the supply chain. Upon successful matching, the adjacent enterprise nodes of the enterprise to be evaluated are located in the supply chain field value deviation transmission graph, and the field value deviation pattern vectors of the adjacent enterprises are obtained as supply chain context deviation adjustment parameters. The comprehensive semantic matching score of the enterprise to be evaluated is weighted and adjusted by adding the product of the supply chain context deviation adjustment parameter and the supply chain adjustment coefficient to the comprehensive semantic matching score. Based on the adjusted comprehensive semantic matching score, a field alignment mapping scheme and a unified credit field structure are generated for the enterprise to be evaluated. The newly generated unified credit field structure is incorporated into the supply chain field value distribution set, the field value deviation pattern vector is recalculated, and the node and edge data of the supply chain field value deviation transmission graph are updated.

[0191] In the specific implementation process, combined with Figure 2 The flowchart shown illustrates the process flow. The process first executes step S120, which involves parallel extraction of semantic units from multi-source heterogeneous data. Specifically, the process aggregates the three data sources provided on the left—unstructured text credit reports, semi-structured image credit certificates, and structured numerical credit record tables—into a multi-source heterogeneous credit data set. Subsequently, following steps S121 to S123, the method performs a sliding semantic window scan and boundary integrity assessment on the full-text character stream of the unstructured text credit report. Based on the determination of the field semantic boundary feature set, it generates and outputs semantic units for text fragments. Simultaneously, following steps S124 to S127, the method first performs visual region segmentation (dividing into title region, seal region, and data table region) on the semi-structured image credit certificate. Then, it applies optical character recognition (OCR) to each region and matches it with a credit field title dictionary to generate visual region semantic units. For structured numerical credit record tables, this method, following steps S1272 to S1273, performs field separator scanning and thesaurus matching to generate semantic units for indicator items. The outputs of the above three processing steps are... Figure 2 The text segments are grouped together with dashed border boxes to generate the text fragment semantic unit group, visual region semantic unit group, and index item semantic unit group as described in step S1274.

[0192] Figure 2The lower half of the diagram illustrates the construction process of the field semantic topology structure for the enterprise credit reporting field in step S130. Following the description in step S131, this method extracts standard field name entries from a massive amount of enterprise credit reporting business record documents, forming a set of standard field name entries. Subsequently, according to the descriptions in steps S132 to S133, the document co-occurrence frequency of any two standard field name entries within the same paragraph is statistically analyzed, and the field co-occurrence association strength value is obtained through normalized association calculation. According to the description in step S134, the aforementioned standard field name entries are mapped to standard field semantic core nodes, and based on the description in step S135, a contextual co-occurrence association connection is established between the two nodes; the weight of the connection is the field co-occurrence association strength value. Further, according to the descriptions in steps S1361 to S1364, the process performs community structure discovery processing on this initial network, identifying densely connected node clusters as field semantic function communities, extracting community semantic topic representation vectors, and establishing community-level semantic association edges between communities. The aforementioned nodes, connections, and topic vectors are integrated and encapsulated to form... Figure 2 The hierarchical field semantic topology structure enclosed by the large dashed circle on the far right serves as the target network benchmark for the subsequent topology association mapping process in step S140.

[0193] Figure 3 This demonstrates the complete logic of topological association mapping processing and field alignment mapping scheme generation in step S140. Based on the description of step S141, the method first expands the original field semantic unit group generated in the previous step S120 into a list of semantic units to be mapped, and then converts them into semantic unit vectors using a pre-trained credit scoring domain semantic encoder. Figure 3In the second-layer similarity calculation structure, according to the descriptions in steps S142 and S143, on the one hand, the semantic unit vector is compared with the semantic description vector of the standard field semantic core node in the field semantic topology, and the cosine value of the angle between the vector spaces is calculated as the basic semantic similarity; on the other hand, it is compared with the community semantic topic representation vector of the field semantic function community to which the core node belongs, and the community semantic similarity is calculated. Simultaneously, according to the description in step S144, this method combines the source confidence adjustment coefficient in the data source confidence configuration table to perform weighted adjustment of the basic semantic similarity. Subsequently, according to the description in step S145, the two similarities are linearly superimposed according to a preset ratio to generate a comprehensive semantic matching score, and candidate matching nodes with scores exceeding a threshold are sorted to select the alignment target node. When performing steps S1461 to S1464, the method performs path penalty verification on the selected target node, calculates the number of jump steps for the shortest topological path, and when the number of jump steps exceeds the limit, reduces the score based on the path penalty coefficient and re-determines the alignment target node. Finally, it establishes an alignment mapping record containing semantic units, alignment target nodes, mapping propagation paths, and scores. The summary result of the above records is... Figure 3 Field alignment mapping scheme enclosed in bottom curly braces.

[0194] Figure 4 This demonstrates the complete process of field fusion and reconstruction in step S150, as well as subsequent dynamic validation and optimization. Figure 4 In the field container merging logic on the left, according to the description of step S151, the method traverses each record in the field alignment mapping scheme and extracts the merge container identifier. According to the descriptions of steps S152 to S153, the field container is searched or created in the field container dictionary using this identifier, and semantic units from multi-source heterogeneous data are placed in the same container, along with merge metadata. According to the descriptions of steps S1541 to S1545, the method calculates the semantic deviation within the container for the semantic units, removes outliers exceeding the tolerance ratio, and, when a conflict is detected between the original field values ​​from different sources, determines the fused field value by weighted voting based on source confidence (if the numerical difference exceeds the limit, it is retained in array form). According to the description of step S1546, the field container uses the merge container identifier as the standard field name, and together with the fused field value and merge metadata, forms a standard credit reporting field record, which is then aggregated and generated according to the standard field arrangement order. Figure 4 The unified credit reporting field structure is shown in the lower left corner.

[0195] Figure 4The intermediate verification and evolution logic safeguards and optimizes the aforementioned steps. Specifically, the contextual semantic coherence verification, based on steps S210 to S270, extracts the original field semantic context window, combines the path semantic chain with the target contextual coherence vector, and calculates the semantic coherence deviation value. When the deviation value exceeds the limit, it is marked as a mapping record to be reviewed, and candidate backtracking node search and mapping path correction are performed. The following circular dynamic evolution logic, based on steps S410 to S470, establishes a field usage frequency counter, identifies high-frequency core nodes in the community, performs semantic gravity propagation, and performs cross-community fusion and node migration operations. This allows the field semantic topology to dynamically and adaptively adjust with actual applications, and the update results are fed back to steps S140 and S150 through a feedback loop, forming a continuous optimization closed loop.

[0196] Figure 4 The five extended modules on the right correspond to specific application scenarios addressed in this application. The Business Scenario Field Supplementation Module, based on steps S510 to S580, extracts a list of necessary fields from the external business scenario identifier, automatically identifies missing items, and generates supplementation instructions. The Evidence Preservation and Traceability Verification Module, based on steps S610 to S650, performs data evidence preservation verification, merges traceability hash calculations, and constructs a Merkle tree, synchronizing the traceability block to the distributed evidence preservation network. The Data Correction Path Adjustment Module, based on steps S710 to S770, performs semantic position verification and mapping divergence analysis on credit dispute appeal requests, generates alternative propagation paths, and triggers targeted re-execution. The Related Enterprise Deviation Transmission Graph Module, based on steps S810 to S870, extracts common fields from the supply chain, constructs a deviation transmission graph, and performs supply chain context deviation adjustments for newly added enterprises. The historical version evolution log module, based on the descriptions in steps S310 to S360, identifies field splitting or merging paths based on version snapshot comparison and updates the field semantic topology to a new version accordingly. The processing results of the aforementioned extended module are then processed through... Figure 4 The horizontal continuous optimization loop arrow at the bottom feeds back to the field fusion and reconstruction process, ensuring that the unified credit reporting field structure can continuously adapt to the ever-changing business scenarios.

[0197] Furthermore, the field alignment mapping system combining third-party credit data provided in this application embodiment includes a processor and a memory. The processor and memory are connected, for example, via a bus. Optionally, the field alignment mapping system combining third-party credit data may further include a transceiver, which can be used for data interaction between this field alignment mapping system and other field alignment mapping systems combining third-party credit data, such as sending and / or receiving data. It should be noted that in actual scheduling, the transceiver is not limited to one, and the structure of this field alignment mapping system combining third-party credit data does not constitute a limitation on the embodiments of this application.

[0198] The memory stores program code for executing the embodiments of this application, and its execution is controlled by a processor. The processor executes the program code stored in the memory to implement the steps shown in the foregoing method embodiments.

[0199] The above description is only an optional implementation method for some implementation scenarios of this application. It should be noted that for those skilled in the art, other similar implementation methods based on the technical concept of this application, without departing from the technical concept of this application, also fall within the protection scope of the embodiments of this application.

Claims

1. A field alignment mapping method combining third-party credit data, characterized in that, The method includes: Obtain a multi-source heterogeneous credit data set of the target enterprise, wherein the multi-source heterogeneous credit data set includes unstructured text credit reports, semi-structured image credit certificates, and structured numerical credit record tables collected from multiple third-party credit reporting platforms; The multi-source heterogeneous credit data set is processed by extracting field semantic units to obtain the original field semantic unit group of each credit data source. The original field semantic unit group includes text fragment semantic units segmented from the unstructured text credit report, visual region semantic units identified from the semi-structured image credit certificate, and indicator item semantic units parsed from the structured numerical credit record table. A field semantic topology structure for the enterprise credit investigation field is constructed. The field semantic topology structure includes standard field semantic core nodes and context co-occurrence association connections between the standard field semantic core nodes. Topological association mapping processing is performed on each semantic unit in the original field semantic unit group and the standard field semantic core node to obtain a field alignment mapping scheme. Each mapping record in the field alignment mapping scheme records the corresponding association path between a semantic unit and a standard field semantic core node. According to the field alignment mapping scheme, the original field semantic units in the multi-source heterogeneous credit data set are subjected to field fusion and reconstruction processing to generate a unified credit field structure for the target enterprise. In the unified credit field structure, the multi-source semantic units under the same standard field semantic core node are merged into the same field container.

2. The field alignment mapping method combining third-party credit data according to claim 1, characterized in that, The step of extracting semantic units from the multi-source heterogeneous credit data set to obtain the original semantic unit group for each credit data source includes: The full-text character stream of the unstructured text credit report is obtained, and the full-text character stream is scanned continuously through a sliding semantic window to generate multiple candidate character fragments; Each candidate character segment undergoes boundary integrity assessment. Based on a preset set of field semantic boundary features, it is determined whether the starting and ending character positions of the candidate character segment constitute a complete field semantic unit. The set of field semantic boundary features includes prefix character features and suffix character features of credit reporting terminology. Candidate character segments that have undergone the boundary integrity assessment are identified as text segment semantic units, and the text segment semantic units are associated with and stored in the paragraph position index and sentence offset position in the unstructured text credit report to obtain text segment semantic unit records. Obtain the pixel matrix of the semi-structured image credit certificate, perform visual region segmentation processing on the pixel matrix, and divide the title text area block, seal area block and data table area block based on the prior features of the credit certificate layout. Optical character recognition processing is performed on the title text area block to convert the pixel array within the title text area block into a first text sequence. The first text sequence is then matched with a preset credit field title dictionary to obtain a visual region semantic unit. The data table area block is parsed to extract the table's row and column separator coordinates and cell boundary coordinates. Cell image blocks are cropped based on the cell boundary coordinates. The optical character recognition process is performed on the cell image blocks to obtain a second text sequence. The second text sequence is matched with the credit field title dictionary to obtain a visual region semantic unit. Visual region semantic units are extracted from the seal area block of the semi-structured image credit certificate, and the structured numerical credit record table is parsed and standardized. Then, all extraction results are grouped and collected according to source type.

3. The field alignment mapping method combining third-party credit data according to claim 2, characterized in that, The process involves extracting visual region semantic units from the seal area block of the semi-structured image credit certificate, parsing and standardizing the structured numerical credit record table, and then grouping all extraction results by source type, including: The circular outline of the stamp area is detected, and the circular arrangement of text pixels within the circular outline of the stamp is extracted. The circular arrangement of text pixels is expanded in polar coordinates to obtain horizontal text strips. The optical character recognition process is performed on the horizontal text strips to obtain a third text sequence. The third text sequence is matched with a preset set of enterprise name suffix feature characters to obtain a visual region semantic unit. Obtain the header definition row and data record row of the structured numerical credit record table, perform field separator scanning processing on the header definition row, and split the header definition row into a sequence of indicator item names according to a preset set of field separators. The set of field separators includes tab characters, vertical bar characters, and comma characters. Semantic unit standardization processing is performed on each indicator item name in the indicator item name sequence. The indicator item name is compared and converted with a preset credit indicator synonym mapping dictionary to generate indicator item semantic units with a unified expression form. The text fragment semantic unit records, the visual region semantic units, and the indicator item semantic units are grouped and collected according to their respective credit data source types to form the text fragment semantic unit group corresponding to the unstructured text credit report, the visual region semantic unit group corresponding to the semi-structured image credit certificate, and the indicator item semantic unit group corresponding to the structured numerical credit record table.

4. The field alignment mapping method combining third-party credit data according to claim 1, characterized in that, The construction of the field semantic topology structure for the enterprise credit investigation field includes: Collect a large number of enterprise credit investigation business record documents, extract field names from the field definition section of the enterprise credit investigation business record documents, extract standard field name entries from the field definition section, and form a set of standard field name entries; Perform context co-occurrence frequency statistics on any two field standard name entries in the set of field standard name entries, traverse the full text of the enterprise credit business record document, record the document frequency count of the two field standard name entries appearing together in the same credit field description paragraph, and generate a field co-occurrence frequency record table; For each pair of field standard name entries in the field co-occurrence frequency record table, a co-occurrence strength measurement process is performed. The document frequency count and the independent occurrence frequency of the two field standard name entries in the enterprise credit business record document are normalized and correlated to obtain the field co-occurrence correlation strength value. Standard field semantic core nodes are constructed based on the set of standard field name entries. Each standard field name entry is mapped to a standard field semantic core node, and a node semantic description vector is assigned to each standard field semantic core node. The node semantic description vector is generated based on the contextual semantic environment of the standard field name entry. Establish a context co-occurrence association connection between any two of the standard field semantic core nodes, and use the field co-occurrence association strength value as the connection weight of the context co-occurrence association connection. When the field co-occurrence association strength value is lower than a preset connection weight threshold, the context co-occurrence association connection is not established. An initial field semantic network is constructed based on the standard field semantic core nodes and the context co-occurrence association connections. The initial field semantic network is then subjected to community structure discovery and semantic topic extraction to generate a field semantic topology with a hierarchical community organizational structure.

5. The field alignment mapping method combining third-party credit data according to claim 4, characterized in that, The process of performing community structure discovery and semantic topic extraction on the initial field semantic network to generate a field semantic topology with a hierarchical community organization structure includes: The initial field semantic network is subjected to community structure discovery processing to identify node clusters in the initial field semantic network whose connection density exceeds the preset community density, and the standard field semantic core nodes in the node clusters are assigned to the same field semantic function community. For each of the field semantic function communities, a community semantic theme extraction process is performed. The average value of the node semantic description vectors of all standard field semantic core nodes in the field semantic function community is extracted as the community semantic theme representation vector. The community semantic theme representation vector is then appended to each standard field semantic core node contained in the field semantic function community. Based on the cross-community connection density between the semantic function communities of the fields, a community-level semantic association edge is established between any two semantic function communities of the fields. The community-level semantic association edge is used to represent the semantic mapping relationship between different credit field subject domains. The standard field semantic core node with the attached community semantic topic representation vector, the context co-occurrence association connection, and the community hierarchical semantic association edge are integrated and encapsulated to generate a field semantic topology structure with a hierarchical community organizational structure.

6. The field alignment mapping method combining third-party credit data according to claim 1, characterized in that, The step of performing topological association mapping processing on each semantic unit in the original field semantic unit group and the standard field semantic core node to obtain a field alignment mapping scheme includes: All semantic units in the original field semantic unit group are expanded into a list of semantic units to be mapped. Semantic vector initialization processing is performed on each semantic unit in the list of semantic units to be mapped. A pre-trained credit scoring domain semantic encoder is called to convert the semantic unit into a semantic unit vector. Extract the node semantic description vectors of all the standard field semantic core nodes from the field semantic topology, input the node semantic description vectors and the semantic unit vectors into a vector similarity comparator, and calculate the cosine value of the vector space angle between the semantic unit vector and each node semantic description vector as the semantic basic similarity. Obtain the community semantic topic representation vector of the field semantic function community to which the standard field semantic core node belongs in the field semantic topology, and calculate the cosine value of the vector space angle between the semantic unit vector and the community semantic topic representation vector as the community semantic similarity; Obtain the credit data source type tag corresponding to the semantic unit, look up the corresponding source confidence adjustment coefficient in the preset data source confidence configuration table according to the credit data source type tag, and perform weighted processing on the semantic basic similarity and the source confidence adjustment coefficient to obtain the adjusted semantic basic similarity; The semantic base similarity and the community semantic similarity are fused together. The semantic base similarity and the community semantic similarity are linearly superimposed according to a preset similarity fusion ratio to generate a comprehensive semantic matching score. The standard field semantic core nodes whose comprehensive semantic matching scores exceed a preset matching threshold are determined as candidate matching nodes. The candidate matching nodes are sorted from high to low according to the comprehensive semantic matching scores. The candidate matching node with the first position in the sorting is selected as the alignment target node of the semantic unit. After determining the alignment target node of the semantic unit based on the comprehensive semantic matching score, the alignment target node is subjected to path penalty verification and mapping record encapsulation processing to generate a field alignment mapping scheme.

7. The field alignment mapping method combining third-party credit data according to claim 6, characterized in that, After determining the alignment target node of the semantic unit based on the comprehensive semantic matching score, the alignment target node is subjected to path penalty verification and mapping record encapsulation processing to generate a field alignment mapping scheme, including: Obtain the topological path distance from the semantic unit to the alignment target node in the semantic topology of the field, perform shortest topological path search processing on the semantic unit in the semantic topology of the field, and calculate the path jump steps based on the number of context co-occurrence association connections traversed by the path. The comprehensive semantic matching score is adjusted by path penalty based on the number of path jumps. When the number of path jumps exceeds the preset path jump limit, the comprehensive semantic matching score is reduced and attenuated according to the preset path penalty coefficient, and the alignment target node of the semantic unit is re-determined. Establish a mapping relationship between the semantic unit and the finally determined alignment target node, and record the sequence of all intermediate standard field semantic core nodes traversed by the shortest topological path from the semantic unit to the alignment target node, as the mapping propagation path; The semantic unit, the alignment target node, the mapping propagation path, and the comprehensive semantic matching score are encapsulated into a field alignment mapping record. The field alignment mapping records of all the semantic units are summarized to form the field alignment mapping scheme.

8. The field alignment mapping method combining third-party credit data according to claim 7, characterized in that, The step of performing field fusion and reconstruction processing on the semantic units of each original field in the multi-source heterogeneous credit data set according to the field alignment mapping scheme to generate a unified credit field structure for the target enterprise includes: Traverse each field alignment mapping record in the field alignment mapping scheme, extract the semantic units and the alignment target nodes in the field alignment mapping record, and use the semantic units as semantic units to be merged and the alignment target nodes as merge container identifiers. Using the merge container identifier as the key, search in the preset field container dictionary whether there is an existing field container corresponding to the merge container identifier. If there is no existing field container in the field container dictionary, create a new field container for the merge container identifier and initialize the new field container to an empty set state. The semantic units to be merged are placed into the field container corresponding to the merge container identifier. At the same time, the credit data source type mark corresponding to the semantic units to be merged and the source field location information in the original credit data are used as merge metadata and appended to the entries of the semantic units to be merged in the field container. Semantic consistency verification and conflict detection are performed on all semantic units to be merged within each field container, and field values ​​are determined and standardized for the field containers to generate a unified credit field structure for the target enterprise.

9. The field alignment mapping method combining third-party credit data according to claim 8, characterized in that, The process of performing semantic consistency verification and conflict detection on all semantic units to be merged within each field container, and determining and standardizing field values ​​for the field containers to generate a unified credit field structure for the target enterprise includes: For all semantic units to be merged in each field container, perform semantic consistency verification processing, extract the semantic unit vectors of all semantic units to be merged in the field container, calculate the cosine of the vector space angle between any two semantic unit vectors as the semantic deviation degree in the container, and count the proportion of deviation entries whose semantic deviation degree in the container exceeds the preset deviation threshold. When the proportion of the deviation entries exceeds the preset fault tolerance ratio, the semantic units in the field container are subjected to outlier entry removal processing. The semantic unit with the largest average semantic deviation from other semantic units is removed from the field container, and the removed semantic unit is added back to the list of semantic units to be mapped. After the outlier item removal process is performed, the remaining semantic units in the field container are subjected to source field value content conflict detection processing. When the original field value content corresponding to the semantic units from different credit data sources in the field container is inconsistent, the original field value content is weighted and voted according to the source confidence adjustment coefficient of each credit data source in the data source confidence configuration table. The original field value content with the highest confidence after weighted voting is taken as the fused field value of the field container. When the original field value content corresponding to the semantic unit of different credit data sources is numerical and the numerical difference exceeds the preset numerical deviation tolerance, the weighted voting process is not performed, and the original field value content is retained in the fused field value of the field container in the form of an array of data source labels. When the field container contains only semantic units from a single credit data source, the original field value content corresponding to the semantic unit is directly used as the fused field value of the field container, and a single data source marker is added to the fused field value of the field container. Each field container is processed to standardize its field name. The merged container identifier corresponding to the field container is used as the standard field name of the field container. The standard field name is combined with the merged field value of the field container and the merged metadata to form a standard credit field record. All standard credit field records generated by the field containers are summarized. The standard credit field records are sorted and combined according to the standard field arrangement order in the enterprise credit field to generate the unified credit field structure of the target enterprise.

10. A field alignment mapping system combining third-party credit data, characterized in that, The invention includes a processor and a computer-readable storage medium storing machine-executable instructions that, when executed by the processor, implement the field alignment mapping method combining third-party credit data as described in any one of claims 1-9.

Images