A Business Intelligence Question Answering Method and System Based on Knowledge Graph

Abstract

This invention discloses a business intelligence question answering method and system based on knowledge graphs, comprising: collecting and preprocessing business data to form basic business semantic data; constructing a business knowledge graph; performing semantic parsing on natural business question answering requests to generate question semantic representations; improving the SubgraphRAG model with input constraints to generate candidate subgraphs of business semantics; performing query path reasoning to generate a set of candidate query paths; performing query executability verification to determine the target query path; generating an executable query task and a reverse tracing tree, executing the query to obtain indicator calculation results; and generating business intelligence question answering results. This invention uses subgraph reasoning with constraints to generate executable query paths, which has the advantages of high accuracy, strong executability, and good interpretability.

Description

Technical Field

[0001] This invention relates to the field of business intelligence question answering technology, and in particular to a business intelligence question answering method and system based on knowledge graphs. Background Technology

[0002] Currently, in the field of business intelligence question answering technology, existing technologies typically generate business question answering results based on natural language parsing, indicator library matching, and database queries, or utilize knowledge graphs to express the relationships between indicators, fields, data tables, and business rules to assist users in completing indicator queries and business analysis. In banking intelligent question data processing, a single question request often involves operational scope, regulatory statistical scope, accounting subjects, customer account relationships, data timing, and sensitive data authorization simultaneously. Existing node retrieval methods based on semantic similarity or fixed rules struggle to handle these constraints synchronously during the subgraph generation stage.

[0003] At the same time, when generating business intelligence Q&A results, existing technologies typically only return indicator calculation results, query statements, or simple path descriptions, lacking a structured reverse tracing mechanism corresponding to the target query path. For banking data Q&A, if the target query path does not verify regulatory standards, accounting subjects, customer account authorization, institutional permissions, and data time points, it is easy to encounter problems such as indicator results not conforming to regulatory reporting standards, unauthorized cross-institutional queries, or exposure of sensitive customer information. Summary of the Invention

[0004] One objective of this invention is to propose a business intelligence question answering method and system based on knowledge graphs. This invention uses caliber-constrained subgraph reasoning to generate executable query paths, which has the advantages of high accuracy, strong executability, and good interpretability.

[0005] A business intelligence question-answering method based on knowledge graphs according to an embodiment of the present invention includes:

[0006] Collect and preprocess business data to form basic business semantic data;

[0007] Building a business knowledge graph based on business semantic foundational data;

[0008] It receives natural language business question-answering requests from users and performs semantic parsing to generate bank question constraint information and question semantic representation;

[0009] The SubgraphRAG model is improved by inputting semantic representation of questions, business knowledge graph and bank question number constraint information into the caliber constraint. Joint attention is calculated based on the joint attention layer at the subgraph level to generate business semantic attention weights. Based on the business semantic attention weights, a set of candidate business triples is selected to form a business semantic candidate subgraph.

[0010] Perform query path reasoning within the business semantic candidate subgraph to generate a set of candidate query paths;

[0011] Perform query executability verification on the candidate query path set and determine the target query path;

[0012] An executable query task is generated based on the target query path. The target indicator node in the target query path is used as the root node for reverse traversal. Business knowledge graph nodes are extracted according to node type to generate traceability explanation branches. Business constraint information is embedded into the corresponding branches to form a reverse traceability tree of the query path. The executable query task is executed to obtain the indicator calculation results.

[0013] Business intelligence question-and-answer results are generated by reverse tracing the source tree based on the indicator calculation results and query path.

[0014] Optionally, the business data includes business intelligence data, indicator management data, data dictionary data, business rule data, and permission rule data. The preprocessing includes field cleaning, duplicate record removal, missing field completion, naming standardization, time format unification, and encoding unification. The business semantic basic data includes preprocessed business data and cross-semantic association records between various types of business data.

[0015] Optionally, the construction of the business knowledge graph includes:

[0016] Read the business semantic objects from the basic business semantic data, generate the corresponding business knowledge graph nodes, and configure the node identifier, node type and node attributes for each node;

[0017] Generate business knowledge graph relationships based on the associated records in the business semantic foundation data;

[0018] Write the business knowledge graph nodes into the entity node set, write the business knowledge graph relations into the relation edge set, and write the node attributes of the business knowledge graph nodes and the relation attributes of the business knowledge graph relations into the graph attribute set.

[0019] A business knowledge graph is constructed based on a set of entity nodes, a set of relation edges, and a set of graph attributes.

[0020] Optionally, the generation of the problem semantic representation includes:

[0021] It receives natural language business question-answering requests, performs text normalization, word segmentation, part-of-speech tagging, and syntactic dependency analysis, and obtains question-answering text analysis results;

[0022] Semantic elements are extracted based on the results of question-and-answer text analysis;

[0023] Based on semantic elements and business knowledge graph nodes, we obtain information on bank inquiry constraints, target indicators, business objects, time conditions, filtering conditions, aggregation methods, comparison relationships, and analysis intent.

[0024] Based on the constraints of bank question data, target indicators, business objects, time conditions, filtering conditions, aggregation methods, comparison relationships, and analysis intentions, a semantic representation of the question is generated.

[0025] Optionally, the formation of the business semantic candidate subgraph includes:

[0026] The problem semantic representation, business knowledge graph, and bank question number constraint information are input into the caliber constraint improved SubgraphRAG model. The caliber constraint improved SubgraphRAG model includes a candidate business triplet recall layer, a problem semantic encoding layer, a candidate business triplet encoding layer, a caliber constraint representation generation layer, a subgraph-level joint attention layer, and a business semantic candidate subgraph generation layer.

[0027] Input the semantic representation of the problem and the business knowledge graph into the candidate business triple recall layer to generate a set of candidate business triples;

[0028] The question semantic representation is encoded by the question semantic encoding layer to generate a question semantic vector, and the candidate business triplet set is encoded by the candidate business triplet encoding layer to generate a candidate business triplet representation.

[0029] In the caliber constraint representation generation layer, a general question number constraint check and a bank question number constraint check are performed on each candidate business triple to obtain the caliber constraint check result, and the caliber constraint check result is combined into a caliber constraint representation;

[0030] The problem semantic vector, candidate business triple representation, and caliber constraint representation are input into the subgraph-level joint attention layer. The degree of problem semantic matching, the degree of triple structure relevance, and the degree of caliber constraint satisfaction are jointly calculated to generate the business semantic attention weight for each candidate business triple.

[0031] The business semantic candidate subgraph generation layer filters the set of candidate business triples based on business semantic attention weights, and writes the business knowledge graph nodes and business knowledge graph relationships in the selected candidate business triples into the business semantic candidate subgraph.

[0032] Optionally, the generation of the candidate query path set includes:

[0033] Read the business knowledge graph nodes, business knowledge graph relationships and business semantic attention weights in the business semantic candidate subgraph, take the target indicator node in the question semantic representation as the path start node, and take the field node, data table node, dimension node, indicator calculation rule node, permission rule node and bank question number constraint node as the path constraint node.

[0034] Starting from the path's starting node, the system extends in a directed manner along the relationships between indicator definitions, field mappings, data table affiliations, dimension hierarchies, indicator calculation rules, data lineage, permission constraints, and bank question constraints to generate an initial set of query paths.

[0035] For each initial query path in the initial query path set, node type and relationship type identification are performed. Initial query paths containing target indicator nodes, field nodes, data table nodes, dimension nodes, indicator calculation rule nodes, and bank question constraint nodes are retained to form a set of query paths to be scored.

[0036] The path relevance score for each query path to be scored is calculated based on business semantic attention weight, node type coverage, and relation continuity.

[0037] Sort the query paths to be scored according to their path-related scores from largest to smallest to obtain the sorted query path set, and write it into the candidate query path set.

[0038] Optionally, the generation of the target query path includes:

[0039] Read the business knowledge graph nodes, business knowledge graph relationships and path-related scores contained in each candidate query path in the candidate query path set, and perform node integrity identification on each candidate query path to obtain the node integrity identification value;

[0040] Perform consistency verification of indicator definitions for each candidate query path to obtain the consistency verification value of indicator definitions;

[0041] For each candidate query path, perform general query constraint verification and bank query constraint verification to obtain the following verification values: field mapping integrity verification value, dimension coverage integrity verification value, data lineage continuity verification value, indicator calculation rule availability verification value, permission rule matching verification value, regulatory statistical caliber consistency verification value, accounting subject mapping integrity verification value, customer account association integrity verification value, data time point consistency verification value, and sensitive data authorization matching verification value.

[0042] Calculate the query executability score for each candidate query path based on different verification values ​​and path-related scores;

[0043] All candidate query paths with valid verification values ​​are written into the executable path set, and any candidate query path with a non-valid verification value is written into the non-executable path set.

[0044] The executable paths are sorted from highest to lowest according to their query executability scores, and the executable path ranked first is determined as the target query path.

[0045] Optionally, the generation of the indicator calculation results includes:

[0046] Read the target indicator node, field node, data table node, dimension node, indicator calculation rule node, permission rule node, and bank question constraint node in the target query path, and generate query execution elements based on the business knowledge graph relationships in the target query path;

[0047] Generate an executable query task based on the query execution elements;

[0048] Using the target indicator node as the root node, we traverse backwards along the directed association links in the target query path to extract the business knowledge graph nodes associated with the target indicator node, and write the extracted business knowledge graph nodes into the source explanation branch according to the node type.

[0049] Embed business constraint information into the corresponding source explanation branch;

[0050] Based on the correspondence between the root node, the source interpretation branch, and the business constraint information, a reverse source tree for the query path is formed, and the reverse source tree for the query path is associated with the executable query task.

[0051] Execute executable query tasks, read target fields from the target data table, perform calculations according to dimension filtering conditions, aggregation methods, and indicator calculation rules, and output indicator calculation results when access permissions are met.

[0052] Optionally, the generation of the business intelligence question-answering results includes:

[0053] Read the indicator calculation results and the query path reverse tracing tree, and extract the target indicator node, tracing explanation branch, business constraint information and branch attributes from the query path reverse tracing tree;

[0054] Generate the main information of the result based on the target indicator nodes and indicator calculation results, generate the basis information of the result based on the traceability and interpretation branches, generate the caliber verification information based on the business constraint information, and write the main information of the result, the basis information of the result, and the caliber verification information into the business intelligence question and answer results.

[0055] According to an embodiment of the present invention, a business intelligence question-answering system based on a knowledge graph includes:

[0056] The Business Semantic Basic Data Formation Module is used to collect business data and preprocess it to form business semantic basic data.

[0057] The business knowledge graph construction module is used to build business knowledge graphs based on basic business semantic data.

[0058] The question semantic representation generation module is used to receive natural language business question-answering requests input by users and perform semantic parsing to generate bank question constraint information and question semantic representation;

[0059] The Business Semantic Candidate Subgraph Generation Module is used to input the semantic representation of the question, the business knowledge graph and the bank's question number constraint information into the caliber constraint to improve the SubgraphRAG model, generate a set of candidate business triples, perform joint attention calculation based on the subgraph-level joint attention layer, generate business semantic attention weights, and filter the set of candidate business triples based on the business semantic attention weights to form a business semantic candidate subgraph.

[0060] The query path reasoning module is used to perform query path reasoning within the business semantic candidate subgraph and generate a set of candidate query paths.

[0061] The query executability verification module is used to verify the executability of the candidate query path set and determine the target query path;

[0062] The query task and reverse tracing tree generation module is used to generate executable query tasks based on the target query path, and execute the executable query tasks to obtain the indicator calculation results.

[0063] The Business Intelligence Question Answering Result Generation Module is used to generate business intelligence question answering results based on the indicator calculation results and the reverse tracing tree of the query path.

[0064] The beneficial effects of this invention are:

[0065] The present invention provides a business intelligence question answering method and system based on knowledge graphs. By improving the SubgraphRAG model with question semantic representation and business knowledge graph input caliber constraints, a candidate business triple set is generated. Based on the subgraph-level joint attention layer, the degree of question semantic matching, the degree of triple structure relevance, and the degree of caliber constraint satisfaction are jointly calculated to generate business semantic attention weights. This enables the exclusion of candidate business triples with inconsistent regulatory calibers, incomplete accounting subject mapping, missing customer account associations, inconsistent data time points, and unauthorized sensitive data during the business semantic candidate subgraph generation stage, thereby improving the accuracy, compliance, and executability of target query path generation in bank intelligent question answering.

[0066] Furthermore, this invention uses the target indicator node in the target query path as the root node for reverse traversal, extracts business knowledge graph nodes according to node type to generate traceability explanation branches, and embeds business constraint information into the corresponding branches to form a query path reverse traceability tree. In intelligent question answering in the banking industry, the query path reverse traceability tree can simultaneously express regulatory basis, accounting basis, customer account basis, data time point basis, and permission anonymization basis, enabling the business intelligence question answering results to have the technical effects of traceability, auditability, and compliance review. Attached Figure Description

[0067] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0068] Figure 1 This is an overall flowchart of a business intelligence question answering method and system based on knowledge graphs proposed in this invention;

[0069] Figure 2 This is a schematic diagram illustrating the process of generating candidate subgraphs for business semantics using the improved SubgraphRAG model based on knowledge graph-based business intelligent question answering method and system proposed in this invention.

[0070] Figure 3 This diagram illustrates the process of generating the executable query task and the reverse tracing tree of the query path in a business intelligent question-answering method and system based on knowledge graphs proposed in this invention. Detailed Implementation

[0071] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0072] refer to Figures 1-3 A knowledge graph-based business intelligence question answering method includes:

[0073] Collect and preprocess business data to form basic business semantic data;

[0074] Building a business knowledge graph based on business semantic foundational data;

[0075] It receives natural language business question-answering requests from users and performs semantic parsing to generate bank question constraint information and question semantic representation;

[0076] The SubgraphRAG model is improved by inputting semantic representation of questions, business knowledge graph and bank question number constraint information into the caliber constraint. Joint attention is calculated based on the joint attention layer at the subgraph level to generate business semantic attention weights. Based on the business semantic attention weights, a set of candidate business triples is selected to form a business semantic candidate subgraph.

[0077] Perform query path reasoning within the business semantic candidate subgraph to generate a set of candidate query paths;

[0078] Perform query executability verification on the candidate query path set and determine the target query path;

[0079] An executable query task is generated based on the target query path. The target indicator node in the target query path is used as the root node for reverse traversal. Business knowledge graph nodes are extracted according to node type to generate traceability explanation branches. Business constraint information is embedded into the corresponding branches to form a reverse traceability tree of the query path. The executable query task is executed to obtain the indicator calculation results.

[0080] Business intelligence question-and-answer results are generated by reverse tracing the source tree based on the indicator calculation results and query path.

[0081] In this implementation, business data includes business intelligence data, indicator management data, data dictionary data, business rule data, and permission rule data. Specifically, in banking intelligent data analysis, business intelligence data includes customer business data, customer identification data, account transaction data, account identification data, transaction occurrence time data, product business data, product identification data, institutional operating data, institutional identification data, account balance data, account identification data, account date data, and statistical period end-of-period data. Indicator management data includes operating indicator caliber data, regulatory indicator caliber data, regulatory caliber version data, regulatory statistical caliber data, and regulatory reporting time-of-period data. Data dictionary data includes field definition data, data table definition data, and accounting subject data. Business rule data includes regulatory statistical rules and product attribution rules; permission rule data includes job permission data, institutional permission data, customer sensitive information access rules, and regulatory reporting permission rules; preprocessing includes field cleaning, duplicate record removal, missing field completion, naming standardization, time format unification, and coding standardization. Specifically, in the banking intelligent data query, naming standardization and coding standardization include unifying customer identifiers, account identifiers, institutional identifiers, product identifiers, subject identifiers, and regulatory caliber versions; time format unification includes unifying transaction occurrence time, accounting date, statistical period end time, and regulatory reporting time; and commercial semantic basic data includes preprocessed commercial data and cross-semantic association records between various commercial data.

[0082] In this embodiment, the construction of the business knowledge graph includes:

[0083] Read the business semantic objects in the business semantic basic data, generate the corresponding business knowledge graph nodes, and configure the node identifier, node type and node attributes for each node. The business semantic objects include customers, accounts, products, institutions, transactions, loans, deposits, accounting subjects, regulatory indicators, data time points, indicators, indicator definitions, fields, data tables, dimensions, indicator calculation rules and permission rules.

[0084] Based on the associated records in the basic business semantic data, a business knowledge graph relationship is generated, which includes indicator definition relationship, field mapping relationship, data table ownership relationship, dimension hierarchy relationship, indicator calculation rule relationship, data lineage relationship, permission constraint relationship and bank question number constraint relationship;

[0085] Write the business knowledge graph nodes into the entity node set, write the business knowledge graph relations into the relation edge set, and write the node attributes of the business knowledge graph nodes and the relation attributes of the business knowledge graph relations into the graph attribute set.

[0086] A business knowledge graph is constructed based on a set of entity nodes, a set of relation edges, and a set of graph attributes.

[0087] In this embodiment, the generation of the problem semantic representation includes:

[0088] It receives natural language business question-answering requests, performs text normalization, word segmentation, part-of-speech tagging, and syntactic dependency analysis, and obtains question-answering text analysis results;

[0089] Semantic elements are extracted based on the results of question-and-answer text analysis. These semantic elements include indicator words, business object words, time condition words, filtering condition words, aggregation method words, comparison relationship words, and analysis intent words.

[0090] Based on semantic elements and business knowledge graph nodes, we obtain information on bank inquiry constraints, target indicators, business objects, time conditions, filtering conditions, aggregation methods, comparison relationships, and analysis intent.

[0091] The bank question-and-answer constraint information is a structured constraint data set dynamically generated based on the current natural language business question-and-answer request. The generation steps include: extracting customer identifier, account identifier, institution identifier, product identifier, subject identifier, regulatory definition version, transaction time, accounting date, statistical period end date, and regulatory reporting date based on the semantic elements in the semantic representation of the current question and the business knowledge graph nodes that match the semantic elements; combining customer identifier and account identifier to generate customer account constraint information corresponding to the current request; combining institution identifier and product identifier to generate institution product constraint information corresponding to the current request; combining subject identifier and regulatory definition version to generate definition version constraint information corresponding to the current request; combining transaction time, accounting date, statistical period end date, and regulatory reporting date to generate time consistency constraint information corresponding to the current request; and writing customer account constraint information, institution product constraint information, definition version constraint information, and time consistency constraint information into the bank question-and-answer constraint information corresponding to the current request.

[0092] The generation of target metrics, business objects, time conditions, filtering conditions, aggregation methods, comparison relationships, and analytical intent specifically includes: matching metric terms with metric nodes in the business knowledge graph to obtain target metrics; matching business object terms with business object nodes in the business knowledge graph to obtain business objects; and converting time condition terms, filtering condition terms, aggregation method terms, comparison relationship terms, and analytical intent terms into time conditions, filtering conditions, aggregation methods, comparison relationships, and analytical intent terms, respectively.

[0093] Based on the constraints of bank question data, target indicators, business objects, time conditions, filtering conditions, aggregation methods, comparison relationships, and analysis intentions, a semantic representation of the question is generated.

[0094] In this embodiment, the formation of the commercial semantic candidate subgraph includes:

[0095] The problem semantic representation, business knowledge graph, and bank question number constraint information are input into the caliber constraint improved SubgraphRAG model. The caliber constraint improved SubgraphRAG model includes a candidate business triplet recall layer, a problem semantic encoding layer, a candidate business triplet encoding layer, a caliber constraint representation generation layer, a subgraph-level joint attention layer, and a business semantic candidate subgraph generation layer.

[0096] The semantic representation of the problem and the business knowledge graph are input into the candidate business triplet recall layer to generate a set of candidate business triplets. The candidate business triplet recall layer recalls matching business knowledge graph nodes and business knowledge graph relationships in the business knowledge graph according to the target indicators, target indicator nodes, business objects, time conditions, filtering conditions, aggregation methods, comparison relationships and analysis intentions in the semantic representation of the problem. The candidate business triplet set is generated according to the combination structure of business knowledge graph node - business knowledge graph relationship - business knowledge graph node.

[0097] The question semantic representation is encoded by the question semantic encoding layer to generate a question semantic vector, and the candidate business triplet set is encoded by the candidate business triplet encoding layer to generate a candidate business triplet representation.

[0098] In the caliber constraint representation generation layer, a general question number constraint check and a bank question number constraint check are performed on each candidate business triple to obtain the caliber constraint check result, and the caliber constraint check result is combined into a caliber constraint representation;

[0099] The general question-and-answer constraint verification specifically includes: indicator consistency verification, field mapping integrity verification, dimension coverage integrity verification, data lineage continuity verification, indicator calculation rule availability verification, and permission rule matching verification. Specifically, the indicator consistency verification result is obtained by the indicator constraint representation generation layer reading the indicator nodes, indicator definition information, and question semantic representation associated with the candidate business triples, and determining whether the statistical range, calculation boundaries, statistical period, and aggregation method recorded in the indicator definition node are consistent with the question semantic representation. The field mapping integrity verification result is obtained by the indicator node, field node, and data table node associated with the candidate business triples, and determining whether the target indicator is associated with the target field through field mapping relationships, and whether the target field is associated with the target data table through data table attribution relationships. The dimension coverage integrity verification result is obtained by the indicator constraint representation generation layer reading the business object, time condition, and filtering conditions in the question semantic representation, reading the dimension nodes associated with the candidate business triples, and determining whether the dimension nodes cover the requirements of the natural language business question-and-answer request. The data lineage continuity verification result is obtained from the business dimension, time dimension, and filtering dimension. The generation layer reads the field nodes, data table nodes, and data lineage relationships associated with the candidate business triples, and determines whether the source of the target field's field value, field processing relationship, and the data table relationship between the target field's data table and the associated data table are continuous. The indicator calculation rule availability verification result is obtained from the generation layer reading the indicator calculation rule nodes associated with the candidate business triples, and combining the target indicator, aggregation method, and comparison relationship in the problem semantic representation to determine whether the indicator calculation rule can generate the corresponding indicator calculation result. The permission rule matching result is obtained from the generation layer reading the permission rule nodes, the current user's permission information, and the indicator nodes, field nodes, data table nodes, and dimension nodes involved in the candidate business triples, and determining whether the current user has the permission to access the corresponding indicator, field, data table, and dimension range. All the above verification results are binary verification results, with a value of 1 for passing the verification and a value of 0 for failing the verification.

[0100] The bank's data query constraint verification specifically includes verification of regulatory statistical caliber consistency, accounting subject mapping integrity, customer account association integrity, data time point consistency, and sensitive data authorization matching. The regulatory statistical caliber consistency verification result is obtained by the caliber constraint representation generation layer reading the regulatory indicator caliber, regulatory caliber version, and regulatory statistical caliber associated with the candidate business triple, and combining this with the regulatory caliber version in the bank's data query constraint information to determine whether the regulatory statistical caliber used by the candidate business triple is consistent with the regulatory caliber version required by the current data query request. The accounting subject mapping integrity verification result is obtained by the caliber constraint representation generation layer reading the accounting subject node, field node, and indicator node associated with the candidate business triple, and determining whether the target indicator's corresponding field can be mapped to the corresponding accounting subject, and whether there is an association between the accounting subject and the indicator calculation rules. The customer account association integrity verification result is obtained by the caliber constraint representation generation layer reading... The data consistency verification result is obtained by retrieving customer node, account node, customer account relationship, and customer account constraint information from bank inquiry constraint information, and determining whether there is a traceable association between customer identifier and account identifier. The data time point consistency verification result is obtained by the caliber constraint representation generation layer reading the transaction occurrence time, accounting date, statistical period end time, and regulatory reporting time point, and combining it with the time point consistency constraint information in the bank inquiry constraint information to determine whether the data corresponding to the candidate business triplet belongs to the comparable time points under the same statistical caliber. The sensitive data authorization matching result is obtained by the caliber constraint representation generation layer reading the permission rule node, customer sensitive information access rule, regulatory reporting permission rule, and customer node, account node, and field node involved in the candidate business triplet, and determining whether the current user has the authorization to access customer sensitive information, account details information, or regulatory reporting data. The above verification results are binary verification results, with a value of 1 for passing the verification and a value of 0 for failing the verification.

[0101] The generation of the caliber constraint representation is specifically obtained by arranging and combining multiple caliber constraint verification results;

[0102] The problem semantic vector, candidate business triple representation, and caliber constraint representation are input into the subgraph-level joint attention layer. The degree of problem semantic matching, the degree of triple structure relevance, and the degree of caliber constraint satisfaction are jointly calculated to generate the business semantic attention weight for each candidate business triple.

[0103] The joint computation steps include generating a problem semantic matching representation, a triple structure-related representation, and a caliber constraint satisfaction representation for each candidate business triple. The problem semantic matching representation is obtained by performing linear mappings on the problem semantic vector and the candidate business triple representation, then weighting and summing the two linear mapping results, and finally performing a hyperbolic tangent function transformation on the weighted sum. The triple structure-related representation is obtained by performing linear mappings on the candidate business triple representation, and finally performing a hyperbolic tangent function transformation on the linear mapping result. The caliber constraint satisfaction representation is obtained by performing linear mappings on the caliber constraint representation, and finally performing a hyperbolic tangent function transformation on the linear mapping result.

[0104] The semantic matching representation of the problem, the structure-related representation of the triplet, and the caliber constraint satisfaction representation are mapped to the same attention space, and the three types of representations after mapping are weighted and fused to obtain the joint attention representation. The joint attention score is calculated based on the product between the joint attention representation and the joint attention mapping vector. The joint attention score of all candidate business triplets in the candidate business triplet set is subjected to exponential normalization to generate the business semantic attention weight of each candidate business triplet.

[0105] Through the business semantic candidate subgraph generation layer, the set of candidate business triples is filtered according to the business semantic attention weight, and the business knowledge graph nodes and business knowledge graph relationships in the filtered candidate business triples are written into the business semantic candidate subgraph.

[0106] The selection of candidate business triples specifically includes reading the business semantic attention weights corresponding to the candidate business triple set through the business semantic candidate subgraph generation layer, generating a dynamic selection threshold based on the mean, standard deviation, problem semantic complexity, and caliber constraint satisfaction ratio of the business semantic attention weights, and writing candidate business triples with business semantic attention weights greater than or equal to the dynamic selection threshold into the selection business triple set. The problem semantic complexity is the ratio of the number of semantic elements identified in the problem semantic representation to the value 7. The caliber constraint satisfaction ratio is the ratio of the number of candidate business triples in the candidate business triple set that pass the verification of indicator caliber consistency, field mapping integrity, dimension coverage integrity, data lineage continuity, indicator calculation rule availability, permission rule matching results, regulatory statistical caliber consistency, accounting subject mapping integrity, customer account association integrity, data time point consistency, and sensitive data authorization matching results to the total number of candidate business triples in the candidate business triple set.

[0107] The training process of the caliber constraint-improved SubgraphRAG model includes: constructing a business intelligence question-answering training sample set, which includes historical natural language business question-answering requests, corresponding question semantic representations, labeled query paths, labeled business semantic subgraphs, labeled executable query tasks, and labeled metric calculation results; setting model training parameters, including learning rate, batch size, training epochs, subgraph filtering loss weights, path matching loss weights, caliber constraint loss weights, dynamic filtering threshold adjustment coefficients, and convergence thresholds; and inputting the training samples into the caliber constraint-improved SubgraphRAG model to generate training business semantic attention weights, training... The process involves training a business semantic candidate subgraph and a set of training candidate query paths. A subgraph filtering loss is generated based on the difference between the trained and labeled business semantic candidate subgraphs. A path matching loss is generated based on the difference between the trained candidate query path set and the labeled target query path. A caliber constraint loss is generated based on the difference between the caliber constraint verification results corresponding to the trained candidate query paths and the constraint states corresponding to the labeled executable query tasks. A joint training loss is generated based on the subgraph filtering loss, path matching loss, caliber constraint loss, and their corresponding loss weights. Parameter iteration is stopped when the joint training loss meets the convergence threshold, resulting in the trained caliber constraint improved SubgraphRAG model.

[0108] Compared to existing business intelligence question answering methods that directly retrieve graph nodes based on natural language semantic similarity, and those that first retrieve question-related subgraphs based on the pre-improved SubgraphRAG model before performing question-answering reasoning, the caliber constraint-improved SubgraphRAG model in this invention introduces caliber constraint representation and a subgraph-level joint attention layer during the subgraph generation stage. This allows the generation of business semantic candidate subgraphs to no longer rely solely on question semantic relevance and graph structure relevance, but rather on a joint calculation based on the degree of question semantic matching, the degree of triple structure relevance, and the degree of caliber constraint satisfaction. The caliber constraint representation is formed by the consistency of indicator caliber, field mapping completeness, dimension coverage completeness, data lineage continuity, indicator calculation rule availability, permission rule matching results, as well as the consistency of regulatory statistics caliber, accounting subject mapping completeness, customer account association completeness, data time point consistency, and sensitive data authorization matching results corresponding to the candidate business triples. This representation participates in the generation of business semantic attention weights, ensuring that the selected candidate business triples simultaneously meet the semantic requirements of natural language business question answering requests and the caliber requirements for executable business intelligence queries.

[0109] In this embodiment, the generation of the candidate query path set includes:

[0110] Read the business knowledge graph nodes, business knowledge graph relationships, and business semantic attention weights in the business semantic candidate subgraph. Take the target indicator node in the question semantic representation as the path starting node, and take the field node, data table node, dimension node, indicator calculation rule node, permission rule node, and bank question data constraint node as the path constraint node. The bank question data constraint node includes customer node, account node, product node, institution node, accounting subject node, regulatory caliber node, and data time point node.

[0111] Starting from the path's initial node, the system extends in a directed manner along the relationships of indicator scope, field mapping, data table affiliation, dimension hierarchy, indicator calculation rules, data lineage, access constraints, and bank query constraints to generate an initial query path set. Indicator scope relationships characterize the association between the target indicator and its statistical range, calculation boundaries, business scope, and statistical period. Field mapping relationships characterize the calculation mapping between the target indicator and the target field. Data table affiliation relationships characterize the affiliation between the target field and the target data table. Dimension hierarchy relationships characterize the hierarchical association between time, region, product, customer, and organizational dimensions. Indicator calculation rules characterize the association between the target indicator and the aggregation method and calculation expression. Data lineage relationships characterize the source of the target field's value in business data, field processing relationships, and the data table association between the target field's data table and related data tables. Access constraints characterize user access restrictions on indicators, fields, data tables, and dimension ranges. Bank query constraints determine the customer account range, product affiliation range, accounting subject mapping, regulatory statistical scope, and data point-in-time conditions corresponding to the bank query request.

[0112] For each initial query path in the initial query path set, node type and relationship type identification are performed. Initial query paths containing target indicator nodes, field nodes, data table nodes, dimension nodes, indicator calculation rule nodes, and bank question constraint nodes are retained to form a set of query paths to be scored.

[0113] The path-related score for each query path to be scored is calculated based on the business semantic attention weight, node type coverage, and relation continuity. Specifically, the path-related score is obtained by weighted summation of the average business semantic attention weight, node type coverage, and relation continuity within the query path to be scored. The average business semantic attention weight is the average business semantic attention weight of each candidate business triple in the query path to be scored. The node type coverage is the ratio of the number of path constraint node types covered by the query path to the total number of path constraint node types. The relation continuity is the ratio of the number of business knowledge graph relations that satisfy the directed connection order in the query path to the total number of business knowledge graph relations in the query path to be scored.

[0114] The query paths to be scored are sorted in descending order of their path-related scores to obtain a sorted query path set. The number of candidate query paths to be retained is determined based on the number of semantic elements in the question semantic representation and the number of paths in the query path set to be scored. Starting from the first element in the sorted query path set, the query paths to be scored are read sequentially according to the sorting order until the number of read paths reaches the candidate query path retention number. The read query paths to be scored are then written into the candidate query path set. The candidate query path retention number is determined as follows: an initial retention number is obtained based on the number of semantic elements in the question semantic representation and the number of paths in the query path set to be scored. The initial retention number is compared with the minimum retention number, and the larger value is taken. Then, the comparison result is compared with the maximum retention number, and the smaller value is taken to obtain the candidate query path retention number.

[0115] In this embodiment, the generation of the target query path includes:

[0116] Read the business knowledge graph nodes, business knowledge graph relationships and path-related scores contained in each candidate query path in the candidate query path set, and perform node integrity identification on each candidate query path. The objects of node integrity identification include indicator caliber, field mapping, dimension coverage, data lineage, indicator calculation rules and permission rules.

[0117] Perform consistency verification of indicator definitions for each candidate query path to obtain the consistency verification value of indicator definitions;

[0118] For each candidate query path, perform general query constraint verification and bank query constraint verification to obtain the following verification values: field mapping integrity verification value, dimension coverage integrity verification value, data lineage continuity verification value, indicator calculation rule availability verification value, permission rule matching verification value, regulatory statistical caliber consistency verification value, accounting subject mapping integrity verification value, customer account association integrity verification value, data time point consistency verification value, and sensitive data authorization matching verification value.

[0119] Based on different verification values ​​and path-related scores, the query executability score of each candidate query path is calculated. The query executability score is obtained by weighted summation of different verification values ​​and path-related scores.

[0120] Candidate query paths with all valid verification values ​​are written to the executable path set, and candidate query paths with any valid verification value are written to the non-executable path set. Among them, the node integrity identification value, indicator caliber consistency verification value, field mapping integrity verification value, dimension coverage integrity verification value, data lineage continuity verification value, indicator calculation rule availability verification value, permission rule matching verification value, regulatory statistical caliber consistency verification value, accounting subject mapping integrity verification value, customer account association integrity verification value, data time point consistency verification value, and sensitive data authorization matching verification value are all binary verification values. The value is 1 when the verification passes and 0 when the verification fails.

[0121] The executable paths are sorted from highest to lowest according to their query executability scores, and the executable path ranked first is determined as the target query path.

[0122] In this embodiment, the generation of the index calculation results includes:

[0123] Read the target indicator nodes, field nodes, data table nodes, dimension nodes, indicator calculation rule nodes, permission rule nodes, regulatory indicator nodes, accounting subject nodes, customer account nodes, institution nodes, and data time point nodes in the target query path, and generate query execution elements based on the business knowledge graph relationships in the target query path. The query execution elements include the target data table, target field, dimension filtering conditions, indicator calculation rules, aggregation method, and permission access conditions.

[0124] An executable query task is generated based on the query execution elements. The executable query task specifically includes the target data table, target fields, dimension filtering conditions, indicator calculation rules, aggregation methods, and access permission conditions.

[0125] Using the target indicator node as the root node, a reverse traversal is performed along the directed association links in the target query path to extract business knowledge graph nodes associated with the target indicator node. The extracted business knowledge graph nodes are then written into the source interpretation branch according to the node type. The source interpretation branch includes the caliber branch, field branch, data source branch, dimension constraint branch, calculation rule branch, permission constraint branch, regulatory caliber branch, accounting subject branch, customer account branch, and point-in-time branch. The reverse traversal specifically refers to taking the business knowledge graph nodes that directly point to the target indicator node in the target query path as the first-level reverse nodes, taking the business knowledge graph nodes that directly point to the first-level reverse nodes as the second-level reverse nodes, and traversing layer by layer in the reverse direction of the business knowledge graph relationship until the field node, data table node, dimension node, indicator calculation rule node, and permission rule node are traversed.

[0126] The business constraint information is embedded into the corresponding traceability and interpretation branch. The business constraint information includes the consistency verification value of indicator caliber, the integrity verification value of field mapping, the integrity verification value of dimension coverage, the continuity verification value of data lineage, the availability verification value of indicator calculation rules, the matching verification value of permission rules, the consistency verification value of regulatory statistical caliber, the integrity verification value of accounting subject mapping, the integrity verification value of customer account association, the consistency verification value of data time point, and the sensitive data authorization matching verification value.

[0127] The embedding of business constraint information specifically includes writing the consistency verification value of indicator caliber into the caliber class traceability and explanation branch, the integrity verification value of field mapping into the field class traceability and explanation branch, the integrity verification value of dimension coverage into the dimension class traceability and explanation branch, the data lineage continuity verification value into the data source class traceability and explanation branch, the availability verification value of indicator calculation rules into the indicator calculation rule class traceability and explanation branch, the permission rule matching verification value and the sensitive data authorization matching verification value into the permission class traceability and explanation branch, the consistency verification value of regulatory statistical caliber into the regulatory caliber class traceability and explanation branch, the integrity verification value of accounting subject mapping into the accounting subject class traceability and explanation branch, and the integrity verification value of customer account association into the customer account class. The source tracing explanation branch writes the data time point consistency verification value into the time point class source tracing explanation branch. At the same time, a branch constraint score is generated for each source tracing explanation branch. The branch constraint score is obtained by weighted summing of the business constraint verification value, the path related score normalization value, and the matching value between the source tracing explanation branch and the query execution elements in the executable query task. The business constraint verification value is the verification value corresponding to the source tracing explanation branch. The path related score normalization value is the result of normalizing the path related score of the target query path where the source tracing explanation branch is located. The matching value is used to characterize whether the business knowledge graph nodes contained in the source tracing explanation branch match the query execution elements in the executable query task.

[0128] Based on the correspondence between the root node, the source interpretation branch, and the business constraint information, a reverse source tree for the query path is formed, and the reverse source tree for the query path is associated with the executable query task.

[0129] Execute executable query tasks, read target fields from the target data table, perform calculations according to dimension filtering conditions, aggregation methods, and indicator calculation rules, and output indicator calculation results when access permissions are met.

[0130] In this embodiment, the generation of business intelligence question-answering results includes:

[0131] Read the indicator calculation results and the query path reverse tracing tree, and extract the target indicator node, tracing explanation branch, business constraint information and branch attributes from the query path reverse tracing tree;

[0132] Generate the main information of the result based on the target indicator nodes and indicator calculation results, generate the basis information of the result based on the traceability and interpretation branches, generate the caliber verification information based on the business constraint information, and write the main information of the result, the basis information of the result, and the caliber verification information into the business intelligence question and answer results.

[0133] A knowledge graph-based business intelligence question answering system includes:

[0134] The Business Semantic Basic Data Formation Module is used to collect business data and preprocess it to form business semantic basic data.

[0135] The business knowledge graph construction module is used to build business knowledge graphs based on basic business semantic data.

[0136] The question semantic representation generation module is used to receive natural language business question-answering requests input by users and perform semantic parsing to generate bank question constraint information and question semantic representation;

[0137] The Business Semantic Candidate Subgraph Generation Module is used to input the semantic representation of the question, the business knowledge graph and the bank's question number constraint information into the caliber constraint to improve the SubgraphRAG model, generate a set of candidate business triples, perform joint attention calculation based on the subgraph-level joint attention layer, generate business semantic attention weights, and filter the set of candidate business triples based on the business semantic attention weights to form a business semantic candidate subgraph.

[0138] The query path reasoning module is used to perform query path reasoning within the business semantic candidate subgraph and generate a set of candidate query paths.

[0139] The query executability verification module is used to verify the executability of the candidate query path set and determine the target query path;

[0140] The query task and reverse tracing tree generation module is used to generate executable query tasks based on the target query path, and execute the executable query tasks to obtain the indicator calculation results.

[0141] The Business Intelligence Question Answering Result Generation Module is used to generate business intelligence question answering results based on the indicator calculation results and the reverse tracing tree of the query path.

[0142] Example 1: To verify the feasibility of this invention in intelligent question answering within the banking industry, it was applied to the operational analysis and risk compliance question answering environment of a commercial bank's head office and branches. The bank's business data is distributed across its core accounting system, credit management system, customer relationship management system, payment and settlement system, bank card system, risk management system, regulatory reporting platform, data warehouse, data dictionary, and access control system. Business personnel frequently submit natural language business question answering requests such as "What are the reasons for the change in the balance of inclusive loans in the East China branch last month compared to the previous period?", "Is the non-performing loan ratio this quarter affected by manufacturing clients?", and "Which customer segments are involved in the decline in the average daily deposits of a certain branch?".

[0143] The system first collects core accounting data, account transaction data, customer master data, loan contract data, deposit balance data, payment and settlement data, bank card transaction data, accounting subject data, regulatory reporting data, and institutional branch operation data. It then performs consistency processing on customer identifiers, account identifiers, institutional identifiers, product identifiers, subject identifiers, regulatory versions, and data time points to form foundational business semantic data. Based on this, the system constructs a business knowledge graph, enabling customers, accounts, products, institutions, transactions, loans, deposits, accounting subjects, regulatory indicators, indicator definitions, fields, data tables, dimensions, indicator calculation rules, and permission rules to form reasonable business knowledge graph nodes and relationships.

[0144] When a business user inputs the question, "What are the main reasons for the decrease in the balance of inclusive loans in the East China branch last month compared to the previous period?", the system generates a semantic representation of the question. This representation includes the target indicator "balance of inclusive loans", the business object "East China branch", the time condition "last month", the comparison relationship "compared to the previous period", and the analysis intent "analysis of the reasons for the decrease". The system also identifies bank-related constraints such as the definition of inclusive loans, loan balance categories, data point in time, and customer account authorization. The improved SubgraphRAG model incorporates these constraints along with the semantic representation and business knowledge graph. A subgraph-level joint attention layer filters candidate business triples related to the definition of inclusive loan balance, loan contract balance field, accounting category, five-level loan classification, and permission rules, forming a candidate business semantic subgraph.

[0145] Subsequently, the system performs query path reasoning within the business semantic candidate subgraph and eliminates candidate query paths lacking generalization scope, mismatched accounting subjects, unmet customer account authorization, or inconsistent data timing through query executability checks. After determining the target query path, the system generates an executable query task and forms a reverse tracing tree of the query path. This reverse tracing tree includes branches for regulatory scope, accounting subjects, customer accounts, data timing, data sources, and access control anonymization, enabling business personnel to simultaneously verify indicator results, regulatory scope, subject basis, field source, institutional scope, customer authorization, and access control anonymization status within the query results.

[0146] To illustrate the implementation results, the processing records of the bank's business analysis system for the same batch of natural language question-answering requests are summarized. Traditional methods refer to generating question-answering results based on keyword matching, fixed indicator libraries, and rule-based queries. The method of this invention refers to generating business intelligence question-answering results by using a caliber constraint-improved SubgraphRAG model, query path reasoning, query executability verification, and query path reverse tracing tree.

[0147] Table 1. Comparison of Business Intelligence Question Answering Path Generation and Explanation Effects

[0148] Average time / second to generate candidate paths 5.2 2.3 Executable query path hit rate / percentage 76.9 94.2 Accuracy rate / percentage of bank indicator matching 80.8 96.4 Accuracy rate / percentage of matching regulatory standards with accounting subjects 79.6 95.8 Customer authorization and institutional permission verification accuracy rate / percentage 83.5 97.1 Invalid path percentage 34.6 10.5

[0149] As shown in Table 1, the method of this invention reduces the average time for generating candidate paths from 5.2 seconds to 2.3 seconds. This is because the system first generates a business semantic candidate subgraph through business semantic attention weights, and then performs path reasoning within the subgraph, reducing invalid expansions in the complete business knowledge graph. The executable query path hit rate increases from 76.9% to 94.2%, indicating that the target query path no longer depends solely on semantic similarity, but is simultaneously controlled by indicator definitions, field mapping, dimension coverage, data lineage, indicator calculation rules, permission rules, and bank query constraints during the generation process.

[0150] The accuracy rate of matching bank indicator standards improved from 80.8% to 96.4%, the accuracy rate of matching regulatory standards with accounting subjects improved from 79.6% to 95.8%, and the accuracy rate of verifying customer authorization and institutional permissions improved from 83.5% to 97.1%. This indicates that the present invention can better adapt to the regulatory statistical standards, accounting subject mapping, and customer account authorization in intelligent bank data queries. The proportion of invalid paths decreased from 34.6% to 10.5%. This is because the reverse tracing tree of the query path organizes the desensitized basis of regulatory standards, accounting subjects, customer accounts, data time points, and permissions into tracing and interpretation branches, making the data query results have a traceable, auditable, and compliant review basis structure.

[0151] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A business intelligence question-answering method based on knowledge graphs, characterized in that, include: Collect and preprocess business data to form basic business semantic data; Building a business knowledge graph based on business semantic foundational data; It receives natural language business question-answering requests from users and performs semantic parsing to generate bank question constraint information and question semantic representation; The SubgraphRAG model is improved by inputting semantic representation of questions, business knowledge graph and bank question number constraint information into the caliber constraint. Joint attention is calculated based on the joint attention layer at the subgraph level to generate business semantic attention weights. Based on the business semantic attention weights, a set of candidate business triples is selected to form a business semantic candidate subgraph. Perform query path reasoning within the business semantic candidate subgraph to generate a set of candidate query paths; Perform query executability verification on the candidate query path set and determine the target query path; An executable query task is generated based on the target query path. The target indicator node in the target query path is used as the root node for reverse traversal. Business knowledge graph nodes are extracted according to node type to generate traceability explanation branches. Business constraint information is embedded into the corresponding branches to form a reverse traceability tree of the query path. The executable query task is executed to obtain the indicator calculation results. Business intelligence question-and-answer results are generated by reverse tracing the source tree based on the indicator calculation results and query path.

2. The business intelligence question answering method based on knowledge graphs according to claim 1, characterized in that, The business data includes business intelligence data, indicator management data, data dictionary data, business rule data, and permission rule data. The preprocessing includes field cleaning, duplicate record removal, missing field completion, naming standardization, time format unification, and encoding unification. The business semantic basic data includes preprocessed business data and cross-semantic association records between various types of business data.

3. The business intelligence question answering method based on knowledge graphs according to claim 1, characterized in that, The construction of the business knowledge graph includes: Read the business semantic objects from the basic business semantic data, generate the corresponding business knowledge graph nodes, and configure the node identifier, node type and node attributes for each node; Generate business knowledge graph relationships based on the associated records in the business semantic foundation data; Write the business knowledge graph nodes into the entity node set, write the business knowledge graph relations into the relation edge set, and write the node attributes of the business knowledge graph nodes and the relation attributes of the business knowledge graph relations into the graph attribute set. A business knowledge graph is constructed based on a set of entity nodes, a set of relation edges, and a set of graph attributes.

4. The business intelligence question answering method based on knowledge graphs according to claim 1, characterized in that, The generation of the semantic representation of the problem includes: It receives natural language business question-answering requests, performs text normalization, word segmentation, part-of-speech tagging, and syntactic dependency analysis, and obtains question-answering text analysis results; Semantic elements are extracted based on the results of question-and-answer text analysis; Based on semantic elements and business knowledge graph nodes, we obtain information on bank inquiry constraints, target indicators, business objects, time conditions, filtering conditions, aggregation methods, comparison relationships, and analysis intent. Based on the constraints of bank question count, target indicators, business objects, time conditions, filtering conditions, aggregation methods, comparison relationships, and analysis intentions, a semantic representation of the question is generated.

5. The business intelligence question answering method based on knowledge graphs according to claim 1, characterized in that, The formation of the commercial semantic candidate subgraph includes: The problem semantic representation, business knowledge graph, and bank question number constraint information are input into the caliber constraint improved SubgraphRAG model. The caliber constraint improved SubgraphRAG model includes a candidate business triplet recall layer, a problem semantic encoding layer, a candidate business triplet encoding layer, a caliber constraint representation generation layer, a subgraph-level joint attention layer, and a business semantic candidate subgraph generation layer. Input the semantic representation of the problem and the business knowledge graph into the candidate business triple recall layer to generate a set of candidate business triples; The question semantic representation is encoded by the question semantic encoding layer to generate a question semantic vector, and the candidate business triplet set is encoded by the candidate business triplet encoding layer to generate a candidate business triplet representation. In the caliber constraint representation generation layer, a general question number constraint check and a bank question number constraint check are performed on each candidate business triple to obtain the caliber constraint check result, and the caliber constraint check result is combined into a caliber constraint representation; The problem semantic vector, candidate business triple representation, and caliber constraint representation are input into the subgraph-level joint attention layer. The degree of problem semantic matching, the degree of triple structure relevance, and the degree of caliber constraint satisfaction are jointly calculated to generate the business semantic attention weight for each candidate business triple. The business semantic candidate subgraph generation layer filters the set of candidate business triples based on business semantic attention weights, and writes the business knowledge graph nodes and business knowledge graph relationships in the selected candidate business triples into the business semantic candidate subgraph.

6. The business intelligence question answering method based on knowledge graphs according to claim 1, characterized in that, The generation of the candidate query path set includes: Read the business knowledge graph nodes, business knowledge graph relationships and business semantic attention weights in the business semantic candidate subgraph, take the target indicator node in the question semantic representation as the path start node, and take the field node, data table node, dimension node, indicator calculation rule node, permission rule node and bank question number constraint node as the path constraint node. Starting from the path's starting node, the system extends in a directed manner along the relationships between indicator definitions, field mappings, data table affiliations, dimension hierarchies, indicator calculation rules, data lineage, permission constraints, and bank question constraints to generate an initial set of query paths. For each initial query path in the initial query path set, node type and relationship type identification are performed. Initial query paths containing target indicator nodes, field nodes, data table nodes, dimension nodes, indicator calculation rule nodes, and bank question constraint nodes are retained to form a set of query paths to be scored. The path relevance score for each query path to be scored is calculated based on business semantic attention weight, node type coverage, and relation continuity. Sort the query paths to be scored according to their path-related scores from largest to smallest to obtain the sorted query path set, and write it into the candidate query path set.

7. The business intelligence question answering method based on knowledge graphs according to claim 1, characterized in that, The generation of the target query path includes: Read the business knowledge graph nodes, business knowledge graph relationships and path-related scores contained in each candidate query path in the candidate query path set, and perform node integrity identification on each candidate query path to obtain the node integrity identification value; Perform consistency verification of indicator definitions for each candidate query path to obtain the consistency verification value of indicator definitions. For each candidate query path, perform general query constraint verification and bank query constraint verification to obtain the following verification values: field mapping integrity verification value, dimension coverage integrity verification value, data lineage continuity verification value, indicator calculation rule availability verification value, permission rule matching verification value, regulatory statistical caliber consistency verification value, accounting subject mapping integrity verification value, customer account association integrity verification value, data time point consistency verification value, and sensitive data authorization matching verification value. Calculate the query executability score for each candidate query path based on different verification values ​​and path-related scores; All candidate query paths with valid verification values ​​are written into the executable path set, and any candidate query path with a non-valid verification value is written into the non-executable path set. The executable paths are sorted from highest to lowest according to their query executability scores, and the executable path ranked first is determined as the target query path.

8. The business intelligence question answering method based on knowledge graphs according to claim 1, characterized in that, The generation of the indicator calculation results includes: Read the target indicator node, field node, data table node, dimension node, indicator calculation rule node, permission rule node, and bank question constraint node in the target query path, and generate query execution elements based on the business knowledge graph relationships in the target query path; Generate an executable query task based on the query execution elements; Using the target indicator node as the root node, we traverse backwards along the directed association links in the target query path to extract the business knowledge graph nodes associated with the target indicator node, and write the extracted business knowledge graph nodes into the source explanation branch according to the node type. Embed business constraint information into the corresponding source explanation branch; Based on the correspondence between the root node, the source interpretation branch, and the business constraint information, a reverse source tree for the query path is formed, and the reverse source tree for the query path is associated with the executable query task. Execute executable query tasks, read target fields from the target data table, perform calculations according to dimension filtering conditions, aggregation methods, and indicator calculation rules, and output indicator calculation results when access permissions are met.

9. A business intelligence question-answering method based on knowledge graphs according to claim 1, characterized in that, The generation of the business intelligence question-and-answer results includes: Read the indicator calculation results and the query path reverse tracing tree, and extract the target indicator node, tracing explanation branch, business constraint information and branch attributes from the query path reverse tracing tree; Generate the main information of the result based on the target indicator nodes and indicator calculation results, generate the basis information of the result based on the traceability and interpretation branches, generate the caliber verification information based on the business constraint information, and write the main information of the result, the basis information of the result, and the caliber verification information into the business intelligence question and answer results.

10. A knowledge graph-based business intelligence question answering system, executing the knowledge graph-based business intelligence question answering method according to any one of claims 1 to 9, characterized in that, include: The Business Semantic Basic Data Formation Module is used to collect business data and preprocess it to form business semantic basic data. The business knowledge graph construction module is used to build business knowledge graphs based on basic business semantic data. The question semantic representation generation module is used to receive natural language business question-answering requests input by users, perform semantic parsing, and generate bank question constraint information and question semantic representation; The Business Semantic Candidate Subgraph Generation Module is used to input the semantic representation of the question, the business knowledge graph and the bank's question number constraint information into the caliber constraint to improve the SubgraphRAG model, generate a set of candidate business triples, perform joint attention calculation based on the subgraph-level joint attention layer, generate business semantic attention weights, and filter the set of candidate business triples based on the business semantic attention weights to form a business semantic candidate subgraph. The query path reasoning module is used to perform query path reasoning within the business semantic candidate subgraph and generate a set of candidate query paths. The query executability verification module is used to verify the executability of the candidate query path set and determine the target query path; The query task and reverse tracing tree generation module is used to generate executable query tasks based on the target query path, and execute the executable query tasks to obtain the indicator calculation results. The Business Intelligence Question Answering Result Generation Module is used to generate business intelligence question answering results based on the indicator calculation results and the reverse tracing tree of the query path.

Images