Multi-source data knowledge extraction method and system in auditing full-process scene

Abstract

The invention provides a multi-source data knowledge extraction method and system in an auditing full-process scene, and relates to the technical field of auditing, and the method comprises the steps: carrying out the entity recognition and relation extraction based on an auditing field semantic model, and extracting a candidate knowledge triple; constructing an audit process dependency graph for describing a time sequence dependency relationship and a data flow relationship of audit task nodes, mapping the candidate knowledge triples to corresponding nodes, and identifying a knowledge association path across process nodes; based on a knowledge association path, carrying out traceability aggregation on associated knowledge dispersed in different process nodes, carrying out consistency verification and conflict resolution through a back propagation constraint condition, and generating a knowledge expression subjected to process semantic enhancement; and establishing a bidirectional mapping relationship between the audit knowledge network and the audit process dependency graph, performing knowledge retrieval according to the audit query request, and outputting a result. According to the method, dispersed knowledge of the whole audit process can be effectively integrated, and the accuracy and integrity of knowledge extraction are improved.

Description

Technical Field

[0001] This invention relates to the field of auditing technology, and in particular to a method and system for extracting knowledge from multi-source data in the context of the entire audit process. Background Technology

[0002] With the rapid development of information technology and the deepening of enterprise digital transformation, auditing work has gradually shifted from the traditional manual auditing model to a digital and intelligent modern auditing model. Throughout the entire audit process, from audit planning, risk assessment, on-site audits, evidence collection to audit report preparation, a large amount of multi-source heterogeneous data is generated at each stage. This data includes various types such as financial statement data, business process data, electronic vouchers, audit working papers, and communication records, and is stored dispersedly across different systems and platforms.

[0003] Traditional audit data management primarily relies on document management systems and relational databases. Auditors need to manually organize and analyze this data, resulting in low efficiency and difficulty in discovering deep relationships between data. In recent years, knowledge graph technology has been widely applied in fields such as healthcare, finance, and law, providing new technical means for organizing and utilizing structured knowledge. Knowledge extraction, as a core component of knowledge graph construction, aims to automatically extract structured knowledge such as entities and relationships from unstructured or semi-structured data. However, the auditing field has its unique characteristics. Audit knowledge not only exists in a single data source but, more importantly, is embedded in the temporal evolution of the audit process and the data connections across stages. How to design specialized knowledge extraction methods for the entire audit process has become an important research topic for improving the intelligence level of auditing work. Summary of the Invention

[0004] This invention provides a method and system for extracting knowledge from multi-source data in the entire audit process, which can solve the problems in the prior art.

[0005] A first aspect of this invention provides a method for extracting knowledge from multi-source data in a full-process audit scenario, comprising: Collect multi-source heterogeneous data distributed across different stages of the entire audit process; perform entity recognition and relation extraction on the multi-source heterogeneous data based on a semantic model in the audit domain, and extract candidate knowledge triples; An audit process dependency graph is constructed, which describes the temporal dependencies and data flow relationships between audit task nodes. By mapping the candidate knowledge triples to the corresponding nodes in the audit process dependency graph, knowledge association paths across process nodes are identified. Based on the knowledge association path, the associated knowledge scattered across different process nodes is aggregated retrospectively. By backpropagating the constraints in the knowledge association path, the candidate knowledge triples are validated for consistency and conflict resolution is performed, generating a knowledge representation enhanced with process semantics. The knowledge representation is organized into an audit knowledge network, and a two-way mapping relationship is established between the audit knowledge network and the audit process dependency graph. According to the audit query request, knowledge is retrieved in the audit knowledge network, the associated process nodes are located using the two-way mapping relationship, and knowledge results that satisfy the audit query request are output.

[0006] Based on a semantic model from the auditing domain, entity recognition and relation extraction are performed on the multi-source heterogeneous data to extract candidate knowledge triples, including: Data pattern analysis is performed on the multi-source heterogeneous data to identify structured fields, semi-structured tags, and unstructured text fragments within the data. Based on the entity type ontology defined in the semantic model of the audit domain, entity boundary recognition is performed on the structured fields, semi-structured tags and unstructured text fragments respectively, candidate entities matching the entity type ontology are extracted, and each candidate entity is labeled with the corresponding entity type identifier and source data identifier. Based on the relational pattern library defined in the semantic model of the audit domain, pattern matching is performed on the semantic dependency relationships between the candidate entities to identify entity pairs that conform to the predefined relational types in the relational pattern library, and each entity pair is labeled with a relational type identifier; The candidate entities and the entity pairs are organized according to the triple structure of subject entity-relationship type-object entity to generate the candidate knowledge triples, and the source data identifier and the entity type identifier are retained in the candidate knowledge triples.

[0007] By mapping the candidate knowledge triples to corresponding nodes in the audit process dependency graph, knowledge association paths across process nodes are identified, including: Parse the source data identifiers in the candidate knowledge triples, and extract the audit stage information and audit task type information associated with the source data identifiers; Based on the audit stage information and the audit task type information, locate the corresponding audit task node in the audit process dependency graph; Extract the subject entity and object entity from the candidate knowledge triples, construct an entity identifier index, and record the set of audit task nodes associated with each entity in the audit process dependency graph; Traverse the entity identifier index to identify shared entities appearing in multiple audit task nodes. The shared entities represent audit objects passed across process nodes. For each shared entity, extract directed paths connecting multiple audit task nodes associated with the shared entity along the temporal dependencies and data flow relationships in the audit process dependency graph. The directed paths follow the temporal order and data transmission direction of the audit process. Mark the directed paths as the knowledge association paths, and record the sequence of audit task nodes traversed by the path, the types of temporal dependencies between nodes, and the types of data flow relationships in the knowledge association paths.

[0008] Based on the temporal dependencies and data flow relationships in the audit process dependency graph, the directed paths connecting multiple audit task nodes associated with the shared entity are extracted, including: Taking the audit task node where the shared entity first appears as the starting node, a forward traversal is performed along the temporal dependency relationship of the audit process dependency graph, recording each audit task node passed during the traversal and the dependency edges between the nodes; During the forward traversal, it is checked whether each audit task node encountered contains the shared entity. If it does, the audit task node is added to the node sequence of the directed path. When the forward traversal reaches the audit task node where the shared entity last appears, the traversal terminates and a complete node sequence is formed. For adjacent audit task nodes in the node sequence, the data flow relationship connecting the adjacent audit task nodes is extracted from the audit process dependency graph, and the data transmission method of the shared entity between the adjacent audit task nodes is identified; The node sequence, the temporal dependency, and the data flow relationship are combined to form a complete description of the directed path, and a unique path identifier is generated for the directed path.

[0009] Based on the knowledge association path, the associated knowledge scattered across different process nodes is aggregated retrospectively. By backpropagating the constraints in the knowledge association path, the candidate knowledge triples are validated for consistency and conflict resolution is performed, generating a knowledge representation enhanced with process semantics, including: Starting from the endpoint audit task node of the knowledge association path, a node-by-node backtracking process is performed to extract the business constraints associated with each audit task node. The business constraints include the data integrity rules, logical consistency rules, and audit criteria requirements required by the audit task node. The business constraints are propagated along the knowledge association path to the preceding audit task nodes, forming a constraint propagation chain. The business constraints in the constraint propagation chain are applied one by one to the attribute values ​​of the candidate knowledge triples, and consistency verification is performed to check whether the attribute values ​​of the candidate knowledge triples violate the business constraints. When multiple candidate knowledge triples are detected that describe the same shared entity but have conflicting attribute values, the constraint priority is determined based on the position of the source node in the constraint propagation chain according to the business constraint. The attribute values ​​of candidate knowledge triples with constraint priorities higher than the preset constraint threshold are retained as attribute values. The attribute values ​​of candidate knowledge triples with constraint priorities lower than the preset constraint threshold are replaced or deleted to complete the conflict resolution. The candidate knowledge triples after consistency verification and conflict resolution are fused with the process semantic information of the knowledge association path to generate a knowledge representation enhanced with process semantics.

[0010] Based on the audit query request, knowledge is retrieved in the audit knowledge network, the associated process nodes are located using the bidirectional mapping relationship, and knowledge results that satisfy the audit query request are output, including: Extract the target entity conditions, target relationship conditions, and process scope conditions specified in the audit query request; perform a graph query operation in the audit knowledge network to match knowledge nodes that simultaneously satisfy the target entity conditions and the target relationship conditions, forming a preliminary set of candidate knowledge nodes; Using the positive mapping relationship in the bidirectional mapping relationship, the audit task node identifier corresponding to each knowledge node in the preliminary candidate knowledge node set is obtained; according to the process scope condition, audit task nodes that meet the process scope condition are selected in the audit process dependency graph, where the process scope condition limits the scope of the audit stage or the scope of the audit task type, to form the final candidate knowledge node set; For each knowledge node in the final candidate knowledge node set, its associated process node is located through the positive mapping relationship, and the temporal dependency relationship and data flow relationship information of the associated process node are obtained through the audit process dependency graph; the final candidate knowledge node set, the associated process node, and the temporal dependency relationship and data flow relationship information are combined and output as the knowledge retrieval result.

[0011] A second aspect of this invention provides a multi-source data knowledge extraction system for the entire audit process, comprising: The first unit is used to collect multi-source heterogeneous data distributed across different stages of the entire audit process; based on a semantic model in the audit domain, entity recognition and relation extraction are performed on the multi-source heterogeneous data to extract candidate knowledge triples; The second unit is used to construct an audit process dependency graph, which describes the temporal dependencies and data flow relationships between audit task nodes. By mapping the candidate knowledge triples to the corresponding nodes in the audit process dependency graph, knowledge association paths across process nodes are identified. The third unit is used to retrospectively aggregate related knowledge scattered across different process nodes based on the knowledge association path, and to perform consistency verification and conflict resolution on the candidate knowledge triples by backpropagating the constraints in the knowledge association path, thereby generating a knowledge representation enhanced with process semantics. The fourth unit is used to organize the knowledge representation into an audit knowledge network, establish a two-way mapping relationship between the audit knowledge network and the audit process dependency graph, perform knowledge retrieval in the audit knowledge network according to the audit query request, locate the associated process nodes using the two-way mapping relationship, and output the knowledge results that satisfy the audit query request.

[0012] A third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0013] Fourth aspect of the present invention, A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0014] The beneficial effects of this application are as follows: This invention constructs an audit process dependency graph to describe the temporal dependencies and data flow relationships between audit task nodes, effectively solving the problem of neglecting process semantic information in traditional audit knowledge extraction methods. By mapping candidate knowledge triples to corresponding nodes in the audit process dependency graph, cross-process node knowledge association paths can be identified, thereby achieving systematic integration of related knowledge scattered across different audit stages and improving the completeness and accuracy of knowledge extraction.

[0015] This invention employs a traceable aggregation mechanism based on knowledge association paths. By backpropagating constraints within the knowledge association paths, it performs consistency verification and conflict resolution on candidate knowledge triples, effectively addressing knowledge conflicts and inconsistencies in multi-source heterogeneous data. This process-enhanced knowledge representation method ensures the reliability and consistency of extracted knowledge, reduces the workload of manual verification, and improves the quality and efficiency of audit data analysis.

[0016] This invention establishes a bidirectional mapping relationship between an audit knowledge network and an audit process dependency graph, enabling the linkage between knowledge retrieval and process location. When responding to audit query requests, it can not only quickly retrieve relevant knowledge from the knowledge network but also simultaneously locate the process node from which the knowledge originates, providing auditors with a complete knowledge tracing path. This bidirectional mapping mechanism enhances the traceability and interpretability of audit knowledge, helping auditors understand the background and dependencies of knowledge generation, and improving the transparency and credibility of audit work. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the multi-source data knowledge extraction method in the full-process audit scenario of this invention. Figure 2 A flowchart illustrating the process of outputting knowledge results that satisfy audit query requests. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0020] refer to Figure 1 and Figure 2 The multi-source data knowledge extraction method for the entire audit process scenario of this invention includes: Collect multi-source heterogeneous data distributed across different stages of the entire audit process; perform entity recognition and relation extraction on the multi-source heterogeneous data based on a semantic model in the audit domain, and extract candidate knowledge triples; An audit process dependency graph is constructed, which describes the temporal dependencies and data flow relationships between audit task nodes. By mapping the candidate knowledge triples to the corresponding nodes in the audit process dependency graph, knowledge association paths across process nodes are identified. Based on the knowledge association path, the associated knowledge scattered across different process nodes is aggregated retrospectively. By backpropagating the constraints in the knowledge association path, the candidate knowledge triples are validated for consistency and conflict resolution is performed, generating a knowledge representation enhanced with process semantics. The knowledge representation is organized into an audit knowledge network, and a two-way mapping relationship is established between the audit knowledge network and the audit process dependency graph. According to the audit query request, knowledge is retrieved in the audit knowledge network, the associated process nodes are located using the two-way mapping relationship, and knowledge results that satisfy the audit query request are output.

[0021] In one optional implementation, entity recognition and relation extraction are performed on the multi-source heterogeneous data based on a semantic model from the auditing domain, and candidate knowledge triples are extracted, including: Data pattern analysis is performed on the multi-source heterogeneous data to identify structured fields, semi-structured tags, and unstructured text fragments within the data. Based on the entity type ontology defined in the semantic model of the audit domain, entity boundary recognition is performed on the structured fields, semi-structured tags and unstructured text fragments respectively, candidate entities matching the entity type ontology are extracted, and each candidate entity is labeled with the corresponding entity type identifier and source data identifier. Based on the relational pattern library defined in the semantic model of the audit domain, pattern matching is performed on the semantic dependency relationships between the candidate entities to identify entity pairs that conform to the predefined relational types in the relational pattern library, and each entity pair is labeled with a relational type identifier; The candidate entities and the entity pairs are organized according to the triple structure of subject entity-relationship type-object entity to generate the candidate knowledge triples, and the source data identifier and the entity type identifier are retained in the candidate knowledge triples.

[0022] When implementing this solution, multi-source heterogeneous data is received from the audit business system. This data includes financial statement data stored in relational databases, audit working papers saved in XML format, and audit report texts in PDF and DOCX formats. By scanning the data's storage format identifiers and file header information, fields such as "account code," "debit amount," and "credit amount" in the financial statement data are identified as structured fields. Tag nodes in the XML working papers, such as "audit project," "risk level," and "auditor," are identified as semi-structured tags. Continuous text paragraphs in the PDF audit report, after optical character recognition conversion, are identified as unstructured text fragments.

[0023] Load a pre-built semantic model for the audit domain. The entity type ontology in this model includes entity types such as "Audit Entity," "Audit Project," "Auditor," "Accounting Subject," "Voucher Number," "Amount," "Date," and "Audit Opinion." For structured fields, read the value "A Manufacturing Enterprise Co., Ltd." from the "Entity Name" column in the financial statement data table. Using the field semantic mapping rules, directly map this field to the "Audit Entity" entity type, extract this value as a candidate entity, and label it with the entity type identifier: "AuditEntity_001". The source data is identified as follows: "DataSource_DB_FinancialReport_Table01_Row05_Column02". For the semi-structured tags, parse the tag path " / audit working papers / basic information / project name" in the XML working papers, extract the tag content "fixed asset audit", identify it as conforming to the entity type characteristics of "audit project", and label the entity type identifier as: "AuditProject_001". The source data is identified as follows: "DataSource_XML_Worksheet_Node156". For unstructured text fragments, a combined approach based on dictionary matching and contextual analysis was used. In the audit report text paragraph "After verification by auditors Wang and Li, it was found that accounts receivable exceeding three years amounted to RMB 2 million," the prefix "Auditor" was matched using a feature dictionary associated with the "Auditor" type in the entity type ontology. "Wang" and "Li" were identified as candidate entities, with boundaries determined by detecting name characteristic characters and stop words. The entity type identifiers were designated "Auditor_001" and "Auditor_002," respectively, and the source data identifier was: "DataSource_PDF_Report_Page12_Paragraph03_Offset25". Simultaneously, within this text fragment, "Accounts Receivable" is identified as matching the "Accounting Subject" entity type, and "Two Million Yuan" is identified as matching the "Amount" entity type. The corresponding entity type identifier and source data identifier are then labeled accordingly.

[0024] Access the relation schema library in the semantic model of the audit domain. This library defines relation types such as "Performing Audit", "Accounting", "Risk", "Issuing Opinion", and "Responsible Project" and their constraint rules. The schema definition for the relation type "Performing Audit" is that the subject entity type must be "Auditor", the object entity type must be "Audit Project", and both must be in the same data source or have cross-data source association identifiers. A Cartesian product traversal is performed on the extracted candidate entities to examine the relationship between the candidate entities "Wang Mou" (Auditor_001) and "Fixed Asset Audit" (AuditProject_001). By analyzing the source data identifiers of the two entities, it is found that they come from the audit report and working papers, respectively. Further retrieval of the association table between data sources confirms that there is a foreign key association between the document number of the audit report and the project number of the working paper, and that there is an explicit association record from "Wang Mou" to "Fixed Asset Audit" in the "Personnel Assignment" node of the working paper.

[0025] The system determined that the entity pair conforms to the "Execute Audit" relation pattern and labeled the relation type for this entity pair as "Relation_ExecuteAudit_001". For the candidate entities "Accounts Receivable" (AccountSubject_003) and "Two Million Yuan" (Amount_007), a co-occurrence distance of five tokens was detected in the unstructured text, and semantic dependency conjunctions such as "existence" and "funds" were present. The system matched the "Subject Amount" relation type in the relation pattern library. This relation type defines the subject as "account subject" and the object as "amount", and labels the relation type as "Relation_SubjectAmount_005".

[0026] The candidate entity "Wang Mou" is read as the subject entity, "Relation_ExecuteAudit_001" as the relation type, and "Fixed Asset Audit" as the object entity. A candidate knowledge triple "(Auditor_001, Execution Audit, AuditProject_001)" is generated according to the structure of subject entity-relation type-object entity. This triple retains attribute information including the source data identifier of the subject entity. "DataSource_PDF_Report_Page12_Paragraph03_Offset25", Object entity source data identifier: The database name is "DataSource_XML_Worksheet_Node156", with the subject entity type identifier "Auditor", the object entity type identifier "Audit Project", and the relation type identifier "Relation_ExecuteAudit_001". Similarly, a candidate knowledge triple "(AccountSubject_003, Account Amount, Amount_007)" is generated, retaining its source data identifier as "DataSource_PDF_Report_Page12_Paragraph03", and the entity types as "Accounting Subject" and "Amount" respectively. All generated candidate knowledge triples are serialized and stored in the intermediate result library, awaiting subsequent knowledge fusion and verification processing.

[0027] In one optional implementation, knowledge association paths across process nodes are identified by mapping the candidate knowledge triples to corresponding nodes in the audit process dependency graph, including: Parse the source data identifiers in the candidate knowledge triples, and extract the audit stage information and audit task type information associated with the source data identifiers; Based on the audit stage information and the audit task type information, locate the corresponding audit task node in the audit process dependency graph; Extract the subject entity and object entity from the candidate knowledge triples, construct an entity identifier index, and record the set of audit task nodes associated with each entity in the audit process dependency graph; Traverse the entity identifier index to identify shared entities appearing in multiple audit task nodes. The shared entities represent audit objects passed across process nodes. For each shared entity, extract directed paths connecting multiple audit task nodes associated with the shared entity along the temporal dependencies and data flow relationships in the audit process dependency graph. The directed paths follow the temporal order and data transmission direction of the audit process. Mark the directed paths as the knowledge association paths, and record the sequence of audit task nodes traversed by the path, the types of temporal dependencies between nodes, and the types of data flow relationships in the knowledge association paths.

[0028] The candidate knowledge triple contains basic elements such as the subject entity, relation type, object entity, and source data identifier. The parsing engine reads the source data identifier field from the triple; this identifier uses structured encoding to record audit process information generated by the data. The parsing engine segments the identifier according to predefined encoding rules, extracting audit stage information and audit task type information. Audit stage information includes time-series markers such as planning, implementation, and reporting stages, while audit task type information covers specific task categories such as risk assessment, internal control testing, and substantive testing.

[0029] The audit process relies on a knowledge graph as a global knowledge structure framework. This graph uses audit task nodes as basic units, and nodes are connected through temporal dependencies and data flow relationships. Each audit task node stores attribute information such as task identifier, stage, task type, and execution sequence. After parsing and obtaining the audit stage information and audit task type information, a node location operation is performed in the graph. The location algorithm traverses all audit task nodes in the graph, comparing the node's audit stage attribute with the extracted audit stage information, and simultaneously comparing the node's task type attribute with the extracted audit task type information. When both attributes of a node match the extracted information, the node is marked as the target node for the candidate triplet mapping.

[0030] For a specific case, a candidate knowledge triple represents "Revenue recognition process _exists _ risk of overstatement", with its source data identified as "P2_T5_D20231015". After parsing, it is found that the audit stage is the implementation stage and the audit task type is revenue substantive testing. The node identified as "Task_0025" is located in the audit process dependency graph. The node attribute shows that it is a revenue substantive testing task in the implementation stage, thus completing the node mapping.

[0031] After node localization, extract the subject and object entity information from the candidate knowledge triples. Subject and object entities are the specific audit objects involved in the audit process, including business processes, accounting subjects, internal control activities, and risk factors. Create an entity identifier index data structure, using the entity's unique identifier as the key and the set of audit task nodes associated with that entity as the mapping value. Traverse all candidate knowledge triples. Whenever processing a triple, extract its subject and object entities and check if these entities already exist in the index. If the entity appears for the first time, create a new entry in the index and add the currently mapped audit task node to the entity's associated node set; if the entity already exists in the index, append the current audit task node to the entity's existing associated node set.

[0032] Continuing with the above example, the main entity "Revenue Recognition Process" is added to the entity identifier index, and its associated node set contains "Task_0025". When processing another triple "Revenue Recognition Process_Impact_Financial Statement Accuracy", this triple is mapped to the audit opinion in the reporting stage to form the task node "Task_0048". The system updates the associated node set of the "Revenue Recognition Process" entity, at which point the set contains both "Task_0025" and "Task_0048".

[0033] Iterate through all entries in the entity identifier index and check the number of elements in the set of audit task nodes associated with each entity. When an entity's associated node set contains two or more audit task nodes, the entity is identified as a shared entity. A shared entity represents an audit object that is repeatedly processed, passed, or referenced in multiple stages of the audit process and has the characteristic of being passed across process nodes. Create a dedicated analysis record for each shared entity, containing the entity identifier, a list of all associated audit task node identifiers, and the path information to be extracted.

[0034] For the identified shared entity "Revenue Confirmation Process," a path extraction algorithm is initiated. The algorithm reads all audit task nodes associated with this entity from the audit process dependency graph, sorts them according to their execution time sequence, forming a chronologically ordered sequence of nodes. The path extraction process starts with the earliest node and ends with the latest node. Directed paths connecting these nodes are searched in the graph; these paths must adhere to the time-series dependencies and data flow relationships defined in the audit process dependency graph. Time-series dependencies represent the order in which tasks are executed, while data flow relationships represent the direction in which audit data, audit evidence, or audit conclusions are transmitted between tasks.

[0035] The path search employs a depth-first strategy, starting from the initial node and exploring reachable next-level nodes along outgoing edges. The algorithm examines the relationship type of edges, retaining only edges with temporal dependencies and data flow relationships, filtering out other types of associations. When the explored path reaches a target node associated with a shared entity, the algorithm records all intermediate nodes traversed from the initial node to the target node, as well as the connecting edges between nodes. If the shared entity is associated with multiple nodes, the algorithm needs to extract a complete path network covering all associated nodes.

[0036] The extracted directed paths are marked as knowledge association paths, and complete structured information is recorded in the path object. The path object contains an audit task node sequence field, storing the identifiers of all nodes traversed by the path in chronological order; in this case, it stores "Task_0025, Task_0033, Task_0048". The path object also contains a node relationship type field, recording the connection relationships between adjacent nodes. Task_0025 and Task_0033 have a data flow relationship, identified as "data_flow", indicating that audit findings data flows from the revenue testing task to the findings aggregation task; Task_0033 and Task_0048 have a temporal dependency relationship, identified as "temporal_dependency", indicating that the audit opinion formation task must be executed after the findings aggregation task is completed. This information is serialized and stored as follows: "data_flow,temporal_dependency" corresponds to the node sequence.

[0037] Once the knowledge association path is constructed, the path information is associated with the corresponding shared entity record. At the same time, the knowledge transfer path of these cross-process nodes is visualized in the form of virtual edges in the audit process dependency graph, providing structured support for subsequent audit knowledge reasoning and process optimization.

[0038] In one optional implementation, along the temporal dependencies and data flow relationships in the audit process dependency graph, directed paths connecting multiple audit task nodes associated with the shared entity are extracted, including: Taking the audit task node where the shared entity first appears as the starting node, a forward traversal is performed along the temporal dependency relationship of the audit process dependency graph, recording each audit task node passed during the traversal and the dependency edges between the nodes; During the forward traversal, it is checked whether each audit task node encountered contains the shared entity. If it does, the audit task node is added to the node sequence of the directed path. When the forward traversal reaches the audit task node where the shared entity last appears, the traversal terminates and a complete node sequence is formed. For adjacent audit task nodes in the node sequence, the data flow relationship connecting the adjacent audit task nodes is extracted from the audit process dependency graph, and the data transmission method of the shared entity between the adjacent audit task nodes is identified; The node sequence, the temporal dependency, and the data flow relationship are combined to form a complete description of the directed path, and a unique path identifier is generated for the directed path.

[0039] After the audit process dependency graph is established, it is necessary to extract directed paths related to shared entities from the graph. At this point, a scan operation is performed on the audit process dependency graph, traversing all audit task nodes and identifying the data entities involved in each node. When a data entity is detected to appear repeatedly in multiple audit task nodes, that data entity is marked as a shared entity, and its position information in each audit task node is recorded.

[0040] Taking a corporate financial audit scenario as an example, the data entity "accounts receivable balance" was identified as appearing in five different audit task nodes: account balance verification node, aging analysis node, bad debt provision node, asset impairment test node, and financial statement preparation node. "Accounts receivable balance" was marked as a shared entity, and its specific data field names and data types in each node were recorded.

[0041] All audit task nodes in the audit process dependency graph are sorted according to their execution sequence, and the timestamp information of each node is obtained. By comparing the timestamps of the shared entity "Accounts Receivable Balance" in each node, it is determined that the shared entity first appears at the account balance reconciliation node, and the execution time of this node is marked as the first time point. The account balance reconciliation node is set as the starting node, and an access marker is created for this node in the graph data structure to indicate that the path extraction process starts from this node.

[0042] Starting from the initial node, a forward traversal is performed along the directed edges of the audit process dependency graph. Each directed edge in the graph represents the temporal dependency between audit tasks, and the direction of the edge indicates the order in which the tasks are executed. Starting from the account balance verification node, all outgoing edge information is read. It is found that this node has three outgoing edges, pointing to the aging analysis node, the customer credit assessment node, and the sales data statistics node, respectively. These three successor nodes are visited sequentially. When visiting each node, the data entity list of that node is read, and it is checked whether the list contains the shared entity "accounts receivable balance".

[0043] When traversing the aging analysis node, it was detected that the input data field of this node contains a "Accounts Receivable Balance" field. The data type of this field is a decimal numeric type, the field length is 15 digits, and the precision is two decimal places. It is determined that this node contains the target shared entity, and the aging analysis node is added to the path node sequence. This sequence now contains two elements: the account balance reconciliation node and the aging analysis node. At the same time, the dependency edge information connecting these two nodes is recorded. The attributes of this dependency edge include edge type as temporal dependency, edge weight as 1.0, and edge label as data transfer.

[0044] Continue traversing backwards from the aging analysis node, reading all outgoing edges of that node. The aging analysis node has two outgoing edges, pointing to the bad debt provision node and the aging distribution report generation node, respectively. Access the bad debt provision node and examine its data entity list; it contains the "Accounts Receivable Balance" field, which serves as the basis for calculating the bad debt provision amount in this node. Add the bad debt provision node to the path node sequence; the sequence now contains three nodes. Record the dependency edge from the aging analysis node to the bad debt provision node; the data flow volume of this edge is marked as a complete accounts receivable dataset, and the data transfer mode is full transfer.

[0045] The traversal process continues, searching backwards from the bad debt provision node. The outgoing edges of this node point to the asset impairment testing node and the profit and loss calculation node. Inspecting the asset impairment testing node reveals that it uses "accounts receivable balance" for asset impairment assessment, thus adding this node to the path sequence. At this point, the path node sequence expands to four nodes, recording the dependency edges between the bad debt provision node and the asset impairment testing node. The data flow relationship associated with these edges shows that the "accounts receivable balance" data is transmitted in read-only mode and does not undergo numerical changes at this stage.

[0046] Continuing the forward traversal to the financial statement preparation node, it is detected that "Accounts Receivable Balance" appears as a current asset item in the balance sheet. The financial statement preparation node is added to the path sequence, and it is checked whether there are any subsequent nodes containing "Accounts Receivable Balance" after this node. By traversing all outgoing edges of the financial statement preparation node and its subsequent nodes, it is confirmed that this shared entity does not appear in subsequent audit task nodes. The financial statement preparation node is determined to be the last occurrence of "Accounts Receivable Balance," and this node is marked as the termination node. The forward traversal operation is stopped, forming a complete node sequence containing five nodes.

[0047] For each pair of adjacent nodes in the path node sequence, the data flow relationship between them is extracted one by one. For the adjacent node pair of the account balance reconciliation node and the aging analysis node, the data flow information is read from the edge attributes of the audit process dependency graph. This information shows that the "accounts receivable balance" is passed from the account balance reconciliation node to the aging analysis node through a database table join. Specifically, the data transfer path is that the field balance_amount in the output table ARBalanceOutput of the account balance reconciliation node is mapped to the field receivable_balance in the input table AgeAnalysisInput of the aging analysis node. The data transfer time interval is 0.5 hours, the data transfer method is incremental synchronization, and the number of data records transferred is 3,200.

[0048] For the data flow between the aging analysis node and the bad debt provision node, the "accounts receivable balance" is transferred between these two nodes via an intermediate data file. The aging analysis node outputs the processed accounts receivable data to an XML file named AgeAnalysisResult.xml in a temporary storage area. This file contains accounts receivable balance information stratified by aging. The bad debt provision node reads this XML file, parses the accounts receivable balance field, and loads it into an in-memory data structure for subsequent calculations. This data transfer method is recorded as a file transfer mode, with a file size of 52 kilobytes.

[0049] The extracted node sequences, temporal dependencies, and data flow relationships are structurally combined. The node sequences are stored as arrays, with each element being a node identifier: NodeID_A_001, NodeID_B_002, NodeID_C_003, NodeID_D_004, and NodeID_E_005. Temporal dependencies are stored as edge lists, with each edge record containing the source node identifier, target node identifier, dependency type, and temporal order. Data flow relationships are stored as mapping tables, with keys representing the combined identifiers of adjacent node pairs and values ​​representing detailed data transfer description objects. These objects contain attributes such as data field mappings, transfer methods, transfer time, and data volume.

[0050] A unique identifier is created for each generated directed path, using a combined encoding method. The first part of the identifier is the hash value of the shared entity; the string "Accounts Receivable Balance" is hashed to obtain an eight-digit hexadecimal code A3F7B2E9. The second part is a simplified encoding of the starting node identifier, obtained by taking the last four digits of NodeID_A_001. The third part is a simplified encoding of the ending node identifier, 005. The fourth part is a path generation timestamp, a fourteen-digit string in the format of year, month, day, hour, minute, and second.

[0051] In one optional implementation, based on the knowledge association path, the associated knowledge scattered across different process nodes is retrospectively aggregated. By backpropagating the constraints in the knowledge association path, the candidate knowledge triples are subjected to consistency verification and conflict resolution to generate a knowledge representation enhanced with process semantics, including: Starting from the endpoint audit task node of the knowledge association path, a node-by-node backtracking process is performed to extract the business constraints associated with each audit task node. The business constraints include the data integrity rules, logical consistency rules, and audit criteria requirements required by the audit task node. The business constraints are propagated along the knowledge association path to the preceding audit task nodes, forming a constraint propagation chain. The business constraints in the constraint propagation chain are applied one by one to the attribute values ​​of the candidate knowledge triples, and consistency verification is performed to check whether the attribute values ​​of the candidate knowledge triples violate the business constraints. When multiple candidate knowledge triples are detected that describe the same shared entity but have conflicting attribute values, the constraint priority is determined based on the position of the source node in the constraint propagation chain according to the business constraint. The attribute values ​​of candidate knowledge triples with constraint priorities higher than the preset constraint threshold are retained as attribute values. The attribute values ​​of candidate knowledge triples with constraint priorities lower than the preset constraint threshold are replaced or deleted to complete the conflict resolution. The candidate knowledge triples after consistency verification and conflict resolution are fused with the process semantic information of the knowledge association path to generate a knowledge representation enhanced with process semantics.

[0052] The knowledge association path, in its implementation of traceable aggregation of related knowledge scattered across different process nodes, constructs a complete backtracking chain from the endpoint audit task node to the starting audit task node. Starting from the endpoint audit task node of the knowledge association path (which could be a financial statement review node), the constraint storage structure of that node is accessed. This storage structure stores business constraints using key-value pairs. Data integrity rules are recorded as a list of required fields, including fields such as amount, date, and voucher number that must exist and cannot be null. Logical consistency rules are recorded as the verification conditions for the association between fields, including logical judgment rules such as debit / credit balance relationships, amount range limits, and date order. Audit standards requirements are recorded as specific audit standard numbers and clause contents, such as accounting subject usage specifications, voucher attachment requirements, and approval process integrity standards.

[0053] The extracted business constraints are loaded into a constraint container, which records the source node identifier, constraint type identifier, constraint content description, and verification method identifier for each constraint. As constraints propagate along the knowledge association path to preceding audit task nodes, each preceding node is visited sequentially according to the node order in the path. The content in the current constraint container is merged with the constraints of the preceding nodes to form a cumulative constraint set. During propagation, a propagation depth marker is attached to each constraint, recording the number of hops from the starting node, used for subsequent constraint priority determination. When constraints propagate to the purchase order approval node, which itself has supplier qualification verification rules and purchase amount approval authority rules, these rules are merged with the constraints propagated from subsequent nodes to form a constraint propagation chain containing seven constraints.

[0054] During the consistency verification of candidate knowledge triples, the business constraints in the constraint propagation chain are read one by one, and the verification method identifier of each constraint is parsed. When the constraint is a data integrity rule, the candidate knowledge triple set is traversed, the attribute values ​​in the triples are extracted, and it is checked whether the attribute value is null or does not exist. Assuming the candidate knowledge triple is "Purchase Order 001, contains, Supplier Name", where the Supplier Name attribute value is an empty string, this triple is marked as violating the integrity rule, and the violation constraint identifier and violation details are recorded.

[0055] When the constraint is a logical consistency rule, candidate knowledge triples involving multiple attribute values ​​are extracted, and logical operations and comparisons are performed between the attribute values. Assuming the rule requires the invoice date to be later than the order date, the system extracts the triple "Purchase Order 001, with Order Date: 20230315" and the triple "Invoice 002, with Invoice Date: 20230310". The value of the order date (20230315) is compared with the value of the invoice date (20230310). If the system determines that the invoice date is earlier than the order date, it violates the logical consistency rule, and the relevant triple is marked as inconsistent.

[0056] When the conflict detection mechanism is activated, a shared entity index table is established. The key of this index table is the entity identifier, and the value is a set of all candidate knowledge triples describing that entity. The set of candidate knowledge triples is traversed, and the main entity of each triple is extracted. Triples with the same main entity are aggregated into the same entry in the index table. For each entity in the index table, all its associated triples are extracted, and the attribute values ​​describing the same attribute in these triples are compared to see if they are consistent. Assuming the shared entity is Supplier A, there exists a triple "Supplier A, Registered Capital, 5 million RMB" originating from the contract review node, and simultaneously, there exists a triple "Supplier A, Registered Capital, 8 million RMB" originating from the qualification review node. If two triples describing the same attribute, registered capital, are detected but their attribute values ​​differ, the conflict marking process is triggered.

[0057] During the constraint priority determination process, the constraints associated with each access conflict triple are determined, and the priority value is calculated based on the position of the source node in the constraint propagation chain. The position of the source node is represented by the depth value of the node in the knowledge association path; the larger the depth value, the closer the node is to the endpoint audit task node, and the higher the priority of the corresponding constraint.

[0058] The path depth value of the qualification review node is 5, and the path depth value of the contract review node is 3. The system calculates that the constraint condition from the qualification review node has a higher priority than that from the contract review node. The preset constraint threshold is set to a path depth value of 4. The priority value 5 is compared with the threshold 4, and the triplet from the qualification review node is determined to have a higher priority than the threshold, so its attribute value of 8 million yuan is retained. The priority value 3 from the contract review node is lower than the threshold 4, so a replacement operation is performed on this triplet, replacing its registered capital attribute value from 5 million yuan to 8 million yuan, thus resolving the conflict.

[0059] In the process semantic information fusion stage, process semantic metadata is extracted from the knowledge association path. This metadata includes the sequence of audit task node names involved in the path, the temporal relationship identifiers between nodes, and the types of business operations performed by the nodes. Process semantic annotations are added to each candidate knowledge triple that has undergone consistency verification and conflict resolution. These annotations record the audit task node from which the triple originates, the node's position number in the path, and the node's dependencies on other nodes. For example, assuming the triple "Payment Application 003, Amount, 120,000 RMB" originates from the payment approval node, with position number 6 in the path, and depends on the preceding contract review and invoice acceptance nodes, a process semantic annotation structure is generated for this triple. This structure includes the source node identifier as payment approval, position number 6, and a list of preceding dependent node identifiers including contract review and invoice acceptance. The original triples are combined with the process semantic annotation structure to form an enhanced knowledge representation. This representation retains the relationships between entities and attributes while embedding process context information, enabling the knowledge representation to reflect the execution order of audit tasks and business logic constraints, thus completing the generation of enhanced process semantic knowledge representation.

[0060] In one optional implementation, knowledge retrieval is performed in the audit knowledge network according to the audit query request, the associated process nodes are located using the bidirectional mapping relationship, and knowledge results that satisfy the audit query request are output, including: Extract the target entity conditions, target relationship conditions, and process scope conditions specified in the audit query request; perform a graph query operation in the audit knowledge network to match knowledge nodes that simultaneously satisfy the target entity conditions and the target relationship conditions, forming a preliminary set of candidate knowledge nodes; Using the positive mapping relationship in the bidirectional mapping relationship, the audit task node identifier corresponding to each knowledge node in the preliminary candidate knowledge node set is obtained; according to the process scope condition, audit task nodes that meet the process scope condition are selected in the audit process dependency graph, where the process scope condition limits the scope of the audit stage or the scope of the audit task type, to form the final candidate knowledge node set; For each knowledge node in the final candidate knowledge node set, its associated process node is located through the positive mapping relationship, and the temporal dependency relationship and data flow relationship information of the associated process node are obtained through the audit process dependency graph; the final candidate knowledge node set, the associated process node, and the temporal dependency relationship and data flow relationship information are combined and output as the knowledge retrieval result.

[0061] When the audit system receives an audit query request, it needs to parse and decompose the request content. Audit query requests typically contain multiple dimensions of limiting information. The natural language processing module performs lexical and semantic analysis on the request text to identify key information components. For extracting target entity conditions, the specific audit objects involved in the request are identified, such as an auditee's asset items, liability items, or specific financial indicators. Assuming the request content is "Query audit findings related to supplier qualification review in the fixed asset procurement process," the target entities identified include three entity nodes: fixed assets, procurement process, and supplier qualifications. For extracting target relationship conditions, the types of relationships between entities are identified. In the above example, this includes the business affiliation relationship between fixed assets and the procurement process, and the audit dependency relationship between the procurement process and supplier qualifications. For extracting process scope conditions, explicit or implicit audit stage limitations in the request are identified. For example, this request implicitly limits the scope of the pre-audit stage because supplier qualification review is a pre-procurement process.

[0062] When performing graph query operations within the audit knowledge network, a query statement is constructed to traverse and match the knowledge graph. Each knowledge node in the knowledge network has a structured representation of node type, node attributes, and relational edges. Knowledge nodes for all fixed asset types are located from the entity type index, which records the mapping relationship between node identifiers and entity types. Taking node identifier KN20230501 as an example, the entity type attribute value of this node is fixed asset, and the node attributes include asset name "office equipment," original asset value of 150,000 yuan, and purchase date of a specific year and month.

[0063] Further searching for adjacent nodes related to this node, filtering by the type label of the relationship edge, reveals relationship edges of type "business affiliation" pointing to the entity in the procurement process. This relationship edge connects to the procurement node with node identifier KN20230612, whose attributes include procurement batch number, procurement amount, and procurement method. Continuing from this procurement node, searching for relationship edges of type "audit dependency" pointing to the entity in the supplier qualification process, locates the supplier qualification node with node identifier KN20230738, which records attribute information such as supplier name, qualification certificate number, and qualification validity period. Through this layer-by-layer graph traversal, all knowledge node paths that simultaneously satisfy the entity and relationship conditions are found. All nodes on these paths are aggregated to form a preliminary candidate knowledge node set, which includes nodes KN20230501, KN20230612, KN20230738, and several others.

[0064] When retrieving process node information using the forward mapping relationship in the bidirectional mapping relationship, the mapping relationship table is queried to obtain the audit task node identifier corresponding to each knowledge node. The mapping relationship table uses the knowledge node identifier as the primary key and records one or more audit task node identifiers corresponding to it. Querying the mapping record for node KN20230612 yields its corresponding audit task node identifier TN40050, which represents a procurement compliance audit task. Querying the mapping record for node KN20230738 yields its corresponding audit task node identifier TN40051, which represents a supplier qualification verification task. The audit task node identifiers associated with all knowledge nodes in the preliminary candidate set are extracted to form a task node identifier set.

[0065] When filtering within the audit process dependency graph based on process scope conditions, the node and edge data of the audit process dependency graph are loaded. Each task node in the audit process dependency graph contains attributes such as task type, audit stage, and execution sequence. The attribute information of task node TN40050 is read; its audit stage attribute is "pre-audit stage," and its task type attribute is "compliance audit." The attribute information of task node TN40051 is read; its audit stage attribute is "pre-audit stage," and its task type attribute is "qualification verification."

[0066] Based on the scope of the pre-audit phase defined by the process scope conditions, both task nodes meet the screening criteria and are retained in the candidate set. Assuming the initial candidate set also includes node KN20230855, whose corresponding audit task node identifier is TN40188, and whose audit phase attribute is in-process audit, it does not meet the process scope conditions; therefore, node KN20230855 is removed from the candidate set. After screening, a final candidate knowledge node set is formed, which includes nodes KN20230501, KN20230612, and KN20230738.

[0067] When locating associated process nodes for each knowledge node in the final candidate knowledge node set, the forward mapping table is queried again. For node KN20230612, its associated audit task node TN40050 is obtained, and its complete information is located in the audit process dependency graph. Reading the adjacent edge data of this task node reveals a temporal dependency edge from task node TN40048 to task node TN40050. The type label of this edge is "pre-dependency," indicating that task TN40048 must be executed before task TN40050. Task node TN40048 represents the procurement requirement confirmation task, with an execution sequence number of three, while task node TN40050 has an execution sequence number of four. A data flow edge is also found from task node TN40050 to task node TN40051. The type label of this edge is "data output," and the attribute information it carries indicates that the supplier list data produced by task TN40050 is passed as input to task TN40051. The data flow edge records detailed information such as the data field name "Supplier Basic Information Table", the data format "structured table", and the transmission timestamp.

[0068] Continuing with node KN20230738, we obtain its associated audit task node TN40051. Querying the dependency graph of this task node reveals that it receives data input from task node TN40050 and outputs qualification verification result data to task node TN40052. Task node TN40052 represents the qualified supplier list confirmation task, and its execution sequence number is five. We extract all associated edge information for task node TN40051, including input edges, output edges, and temporal dependency edges, to construct the complete relationship network of this task node.

[0069] When the system outputs the final set of candidate knowledge nodes, associated process nodes, and dependency information, it constructs a structured retrieval result object. This result object contains three main parts: a knowledge node list, a process node list, and a relationship information list. The knowledge node list records the identifier, entity type, node attributes, and knowledge content of each candidate knowledge node. Taking node KN20230612 as an example, the output includes the node identifier, entity type as procurement stage, procurement batch number PO2023Q2B08 in the node attributes, procurement amount of 150,000 yuan, and a text description of the specific business scenario involved in this batch of procurement. The process node list records the identifier, task type, stage, and execution order of each associated audit task node. The output of task node TN40050 includes the node identifier, task type as compliance audit, stage as pre-audit stage, and execution order number four. The relationship information list records detailed information on temporal dependencies and data flow relationships. Each relationship record includes the source node identifier, target node identifier, relationship type, and relationship attributes. The temporal dependency record shows the source node TN40048, the target node TN40050, and the relationship type is prerequisite dependency. The relationship attribute states that the completion status of task TN40048 is a prerequisite for starting task TN40050. The data flow relationship record shows the source node TN40050, the target node TN40051, and the relationship type is data output. The relationship attributes include a list of data fields and the data transfer interface address. The system serializes this structured information into a standard data format and returns it to the query initiator through the interface, completing the response output of the entire knowledge retrieval process.

[0070] The multi-source data knowledge extraction system for the entire audit process scenario of this invention includes: The first unit is used to collect multi-source heterogeneous data distributed across different stages of the entire audit process; based on a semantic model in the audit domain, entity recognition and relation extraction are performed on the multi-source heterogeneous data to extract candidate knowledge triples; The second unit is used to construct an audit process dependency graph, which describes the temporal dependencies and data flow relationships between audit task nodes. By mapping the candidate knowledge triples to the corresponding nodes in the audit process dependency graph, knowledge association paths across process nodes are identified. The third unit is used to retrospectively aggregate related knowledge scattered across different process nodes based on the knowledge association path, and to perform consistency verification and conflict resolution on the candidate knowledge triples by backpropagating the constraints in the knowledge association path, thereby generating a knowledge representation enhanced with process semantics. The fourth unit is used to organize the knowledge representation into an audit knowledge network, establish a two-way mapping relationship between the audit knowledge network and the audit process dependency graph, perform knowledge retrieval in the audit knowledge network according to the audit query request, locate the associated process nodes using the two-way mapping relationship, and output the knowledge results that satisfy the audit query request.

[0071] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0072] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0073] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0074] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for extracting knowledge from multi-source data in the entire audit process, characterized in that: include: Collect multi-source heterogeneous data distributed across different stages of the entire audit process; perform entity recognition and relation extraction on the multi-source heterogeneous data based on a semantic model in the audit domain, and extract candidate knowledge triples; An audit process dependency graph is constructed, which describes the temporal dependencies and data flow relationships between audit task nodes. By mapping the candidate knowledge triples to the corresponding nodes in the audit process dependency graph, knowledge association paths across process nodes are identified. Based on the knowledge association path, the associated knowledge scattered across different process nodes is aggregated retrospectively. By backpropagating the constraints in the knowledge association path, the candidate knowledge triples are validated for consistency and conflict resolution is performed, generating a knowledge representation enhanced with process semantics. The knowledge representation is organized into an audit knowledge network, and a two-way mapping relationship is established between the audit knowledge network and the audit process dependency graph. According to the audit query request, knowledge is retrieved in the audit knowledge network, the associated process nodes are located using the two-way mapping relationship, and knowledge results that satisfy the audit query request are output.

2. The method according to claim 1, characterized in that, Based on a semantic model from the auditing domain, entity recognition and relation extraction are performed on the multi-source heterogeneous data to extract candidate knowledge triples, including: Data pattern analysis is performed on the multi-source heterogeneous data to identify structured fields, semi-structured tags, and unstructured text fragments within the data. Based on the entity type ontology defined in the semantic model of the audit domain, entity boundary recognition is performed on the structured fields, semi-structured tags and unstructured text fragments respectively, candidate entities matching the entity type ontology are extracted, and each candidate entity is labeled with the corresponding entity type identifier and source data identifier. Based on the relational pattern library defined in the semantic model of the audit domain, pattern matching is performed on the semantic dependency relationships between the candidate entities to identify entity pairs that conform to the predefined relational types in the relational pattern library, and each entity pair is labeled with a relational type identifier; The candidate entities and the entity pairs are organized according to the triple structure of subject entity-relationship type-object entity to generate the candidate knowledge triples, and the source data identifier and the entity type identifier are retained in the candidate knowledge triples.

3. The method according to claim 1, characterized in that, By mapping the candidate knowledge triples to corresponding nodes in the audit process dependency graph, knowledge association paths across process nodes are identified, including: Parse the source data identifiers in the candidate knowledge triples, and extract the audit stage information and audit task type information associated with the source data identifiers; Based on the audit stage information and the audit task type information, locate the corresponding audit task node in the audit process dependency graph; Extract the subject entity and object entity from the candidate knowledge triples, construct an entity identifier index, and record the set of audit task nodes associated with each entity in the audit process dependency graph; Traverse the entity identifier index to identify shared entities appearing in multiple audit task nodes. The shared entities represent audit objects passed across process nodes. For each shared entity, extract directed paths connecting multiple audit task nodes associated with the shared entity along the temporal dependencies and data flow relationships in the audit process dependency graph. The directed paths follow the temporal order and data transmission direction of the audit process. Mark the directed paths as the knowledge association paths, and record the sequence of audit task nodes traversed by the path, the types of temporal dependencies between nodes, and the types of data flow relationships in the knowledge association paths.

4. The method according to claim 3, characterized in that, Based on the temporal dependencies and data flow relationships in the audit process dependency graph, the directed paths connecting multiple audit task nodes associated with the shared entity are extracted, including: Taking the audit task node where the shared entity first appears as the starting node, a forward traversal is performed along the temporal dependency relationship of the audit process dependency graph, recording each audit task node passed during the traversal and the dependency edges between the nodes; During the forward traversal, it is checked whether each audit task node encountered contains the shared entity. If it does, the audit task node is added to the node sequence of the directed path. When the forward traversal reaches the audit task node where the shared entity last appears, the traversal terminates and a complete node sequence is formed. For adjacent audit task nodes in the node sequence, the data flow relationship connecting the adjacent audit task nodes is extracted from the audit process dependency graph, and the data transmission method of the shared entity between the adjacent audit task nodes is identified; The node sequence, the temporal dependency, and the data flow relationship are combined to form a complete description of the directed path, and a unique path identifier is generated for the directed path.

5. The method according to claim 1, characterized in that, Based on the knowledge association path, the associated knowledge scattered across different process nodes is aggregated retrospectively. By backpropagating the constraints in the knowledge association path, the candidate knowledge triples are validated for consistency and conflict resolution is performed, generating a knowledge representation enhanced with process semantics, including: Starting from the endpoint audit task node of the knowledge association path, a node-by-node backtracking process is performed to extract the business constraints associated with each audit task node. The business constraints include the data integrity rules, logical consistency rules, and audit criteria requirements required by the audit task node. The business constraints are propagated along the knowledge association path to the preceding audit task nodes, forming a constraint propagation chain. The business constraints in the constraint propagation chain are applied one by one to the attribute values ​​of the candidate knowledge triples, and consistency verification is performed to check whether the attribute values ​​of the candidate knowledge triples violate the business constraints. When multiple candidate knowledge triples are detected that describe the same shared entity but have conflicting attribute values, the constraint priority is determined based on the position of the source node in the constraint propagation chain according to the business constraint. The attribute values ​​of candidate knowledge triples with constraint priorities higher than the preset constraint threshold are retained as attribute values. The attribute values ​​of candidate knowledge triples with constraint priorities lower than the preset constraint threshold are replaced or deleted to complete the conflict resolution. The candidate knowledge triples after consistency verification and conflict resolution are fused with the process semantic information of the knowledge association path to generate a knowledge representation enhanced with process semantics.

6. The method according to claim 1, characterized in that, Based on the audit query request, knowledge is retrieved in the audit knowledge network, the associated process nodes are located using the bidirectional mapping relationship, and knowledge results that satisfy the audit query request are output, including: Extract the target entity conditions, target relationship conditions, and process scope conditions specified in the audit query request; perform a graph query operation in the audit knowledge network to match knowledge nodes that simultaneously satisfy the target entity conditions and the target relationship conditions, forming a preliminary set of candidate knowledge nodes; Using the positive mapping relationship in the bidirectional mapping relationship, the audit task node identifier corresponding to each knowledge node in the preliminary candidate knowledge node set is obtained; according to the process scope condition, audit task nodes that meet the process scope condition are selected in the audit process dependency graph, where the process scope condition limits the scope of the audit stage or the scope of the audit task type, to form the final candidate knowledge node set; For each knowledge node in the final candidate knowledge node set, its associated process node is located through the positive mapping relationship, and the temporal dependency relationship and data flow relationship information of the associated process node are obtained through the audit process dependency graph; the final candidate knowledge node set, the associated process node, and the temporal dependency relationship and data flow relationship information are combined and output as the knowledge retrieval result.

7. A multi-source data knowledge extraction system for the entire audit process, used to implement the method as described in any one of claims 1-6, characterized in that, include: The first unit is used to collect multi-source heterogeneous data distributed across different stages of the entire audit process; based on a semantic model in the audit domain, entity recognition and relation extraction are performed on the multi-source heterogeneous data to extract candidate knowledge triples; The second unit is used to construct an audit process dependency graph, which describes the temporal dependencies and data flow relationships between audit task nodes. By mapping the candidate knowledge triples to the corresponding nodes in the audit process dependency graph, knowledge association paths across process nodes are identified. The third unit is used to retrospectively aggregate related knowledge scattered across different process nodes based on the knowledge association path, and to perform consistency verification and conflict resolution on the candidate knowledge triples by backpropagating the constraints in the knowledge association path, thereby generating a knowledge representation enhanced with process semantics. The fourth unit is used to organize the knowledge representation into an audit knowledge network, establish a two-way mapping relationship between the audit knowledge network and the audit process dependency graph, perform knowledge retrieval in the audit knowledge network according to the audit query request, locate the associated process nodes using the two-way mapping relationship, and output the knowledge results that satisfy the audit query request.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images