Auditing industry-oriented large model agent cooperation system

Abstract

The invention discloses an audit industry-oriented large model agent collaboration system, which is characterized by comprising a main control module, at least two special agent modules and an audit knowledge base, and is characterized in that the audit knowledge base stores a knowledge graph and a rule of an audit field; the main control module is used for receiving an audit task description, disassembling the audit task description into sub-audit tasks, and scheduling the corresponding special agent module to execute the sub-audit tasks according to task types of the sub-audit tasks; the at least two special agent modules comprise a document analysis agent configured with a document language model and used for analyzing a to-be-audited document in combination with the audit knowledge base; the data analysis agent is configured with a data language model and is used for analyzing the to-be-audited structured data in combination with the audit knowledge base; and the main control module is also used for performing conflict arbitration and fusion on the analysis results returned by the at least two special agent modules to generate an audit work result.

Description

Technical Field

[0001] This invention relates to a large-scale intelligent agent collaborative system for the auditing industry. Background Technology

[0002] In current auditing operations, automated auxiliary tools are mostly isolated task-oriented systems, such as rule-based single-point document screening tools or pre-scripted financial data analysis tools, which rely on manual arrangement of processes to connect various tools.

[0003] These traditional systems lack the ability to understand and dynamically plan complex audit tasks as a whole, making it difficult to coordinate the work between different tools. They also cannot effectively identify and reconcile potential contradictions and conflicts in the analysis results, resulting in limitations in the intelligence level, efficiency, and reliability of audit conclusions. Summary of the Invention

[0004] This invention provides a large-scale intelligent agent collaborative system for the auditing industry, which addresses the limitations of existing technologies in terms of intelligence, efficiency, and reliability of conclusions.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0006] This invention provides a large-scale intelligent agent collaborative system for the auditing industry, comprising a main control module, at least two dedicated intelligent agent modules, and an audit knowledge base, wherein:

[0007] The audit knowledge base stores knowledge graphs and rules in the audit field;

[0008] The main control module is used to receive the audit task description, break it down into sub-audit tasks, and schedule the corresponding dedicated intelligent agent module to execute according to the task type of the sub-audit task;

[0009] The at least two dedicated intelligent agent modules include:

[0010] The document analysis agent is equipped with a document language model, which is used to analyze the documents to be audited in conjunction with the audit knowledge base.

[0011] The data analysis intelligent agent is equipped with a data language model, which is used to analyze the structured data to be audited in conjunction with the audit knowledge base.

[0012] The main control module is also used to perform conflict arbitration and fusion on the analysis results returned by the at least two dedicated intelligent agent modules to generate audit results.

[0013] In one optional embodiment, the audit knowledge base includes:

[0014] The static knowledge layer is used to store entities, relationships, and rules extracted and structured from auditing standards, regulations, and / or historical audit case databases, forming the basic audit knowledge graph;

[0015] The dynamic task layer is used to store the task context, intermediate evidence, and / or the interaction records of the at least two dedicated intelligent agents derived from the audit task, forming a task execution graph.

[0016] In one optional embodiment, the nodes of the task execution graph include:

[0017] Audit task nodes, sub-task nodes, evidence document nodes, and / or data analysis nodes;

[0018] The edges of the task execution graph include:

[0019] Triggering relationships, generating relationships, supporting relationships, and / or contradictory relationships.

[0020] In one optional embodiment, the main control module is specifically used for:

[0021] The received audit task description is input into a pre-trained task parsing model, and the output is a hierarchical task tree, where the root node of the task tree is the total task and the leaf nodes are atomic subtasks.

[0022] Based on the objective of the atomic subtask, a dedicated intelligent agent module with the processing capability corresponding to the objective is matched from a preset intelligent agent capability registry.

[0023] Based on the dependencies in the task tree, a subtask scheduling sequence containing execution order and data dependencies is generated.

[0024] In one alternative embodiment, the agent capability registry is stored in ontology form, defining the following capability description attributes for each dedicated agent module:

[0025] Data types that can be handled, areas of expertise in risk, applicable audit phases, and / or input / output interface specifications.

[0026] In one optional embodiment, the document analysis agent is specifically used for:

[0027] Receive sub-audit tasks and context information associated with the document to be audited, assigned by the main control module;

[0028] The document to be audited is segmented, and keywords, semantic roles, and risk statements are extracted from each segment based on the audit knowledge base.

[0029] The document language model is invoked, and based on the identified risk points and the contextual information, the analysis results are output.

[0030] In one optional embodiment, the data analysis agent is specifically used for:

[0031] Receive sub-audit tasks and analysis objectives assigned by the main control module and associated with the structured data to be audited;

[0032] Based on the risk rules and historical anomaly rules in the audit knowledge base, the structured data to be audited is analyzed;

[0033] The abnormal indicators obtained from the analysis are compared with the risk thresholds in the audit knowledge base, and the data language model is invoked to output the analysis results.

[0034] In one optional embodiment, the main control module is specifically used for:

[0035] Establish a rule set for conflict detection, which includes: logical contradictions in the statements of different agents regarding the same fact, differences in the severity assessment levels of different agents regarding the same risk point exceeding a preset threshold, and / or incomplete evidence chains;

[0036] When a conflict is detected, the conflict node and its associated raw and analytical data are analyzed to obtain the conflict analysis results.

[0037] An arbitration conclusion is generated based on the conflict analysis results using a decision fusion algorithm.

[0038] In an optional embodiment, the system further includes a human-machine collaboration interface for:

[0039] The audit results are presented to the user in the form of a visual interactive interface;

[0040] Receive the result processing instructions input by the user through the interface;

[0041] The data stored in the system is updated according to the result processing instructions.

[0042] This invention achieves an intelligent upgrade of auditing operations by constructing a system architecture in which a main control module, multiple specialized intelligent agents, and an audit knowledge base work collaboratively. Specifically, the system can automatically break down and plan complex audit processes based on audit task descriptions, invoke document analysis and data analysis intelligent agents configured with domain-specific large models to execute professional tasks in parallel, and ensure the professionalism and context awareness of the analysis by combining static and dynamic knowledge bases. Furthermore, the system introduces a conflict arbitration mechanism to automatically reconcile contradictions in multi-source analysis results, forming reliable and unified conclusions, significantly improving the execution efficiency, analysis depth, and consistency of audit tasks. Finally, through a human-machine collaborative interface, the system effectively combines intelligent analysis capabilities with the experience and judgment of human experts, constructing a virtuous cycle of continuous optimization and increasing intelligence with use. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a schematic diagram of the architecture of a large-scale intelligent agent collaborative system for the auditing industry provided in an embodiment of the present invention. Detailed Implementation

[0045] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. The following embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes, but the scope of protection of the present invention is not limited to the following embodiments.

[0046] In this invention, the terms "in one possible embodiment," "exemplary," or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "in one possible embodiment," "exemplary," or "for example" in this invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "in one possible embodiment," "exemplary," or "for example" is intended to present the relevant concepts in a specific manner.

[0047] Figure 1 This is a schematic diagram of the architecture of a large-scale intelligent agent collaborative system for the auditing industry provided in an embodiment of the present invention. Figure 1 As shown, a large-scale intelligent agent collaborative system 10 for the auditing industry includes a main control module 101, at least two dedicated intelligent agent modules 102 and 103, and an audit knowledge base 104, wherein:

[0048] The audit knowledge base stores knowledge graphs and rules in the audit field;

[0049] The main control module is used to receive the audit task description, break it down into sub-audit tasks, and schedule the corresponding dedicated intelligent agent module to execute according to the task type of the sub-audit task;

[0050] The at least two dedicated intelligent agent modules include:

[0051] The document analysis agent is equipped with a document language model, which is used to analyze the documents to be audited in conjunction with the audit knowledge base.

[0052] The data analysis intelligent agent is equipped with a data language model, which is used to analyze the structured data to be audited in conjunction with the audit knowledge base.

[0053] The main control module is also used to perform conflict arbitration and fusion on the analysis results returned by the at least two dedicated intelligent agent modules to generate audit results.

[0054] For example, the system can be deployed on a cloud server cluster. Users (auditors) input an audit task description via a web client, such as "Perform an audit of the revenue authenticity of Company A's 2023 financial statements." The main control module (centralized scheduling service) receives this description and breaks it down into multiple sub-audit tasks, such as "Examine sales contracts," "Analyze accounts receivable turnover," and "Perform revenue cutoff testing." Based on the task type, it schedules a document analysis agent (a container instance deploying a large, fine-tuned model for the financial sector) to process contract documents, and a data analysis agent (a container instance deploying a dedicated large model for data analysis) to process structured tables exported from the financial database. Both agents invoke their respective models and uniformly query the central audit knowledge base (a combination of graph database and rule engine) to obtain professional evidence. After analysis, the main control module finds inconsistencies in the two agents' judgments regarding the "confirmation date of a large transaction" and initiates conflict arbitration. Finally, it merges the evidence chains from both parties to generate a structured draft audit working paper as the audit result.

[0055] In one optional embodiment, the audit knowledge base includes:

[0056] The static knowledge layer is used to store entities, relationships, and rules extracted and structured from auditing standards, regulations, and / or historical audit case databases, forming the basic audit knowledge graph;

[0057] The dynamic task layer is used to store the task context, intermediate evidence, and / or the interaction records of the at least two dedicated intelligent agents derived from the audit task, forming a task execution graph.

[0058] For example, the construction of the static knowledge layer can be as follows: using natural language processing technology, the provisions of laws and regulations such as the Accounting Standards for Business Enterprises and the disclosure requirements of the China Securities Regulatory Commission are parsed, entities such as "revenue recognition" and "related party transactions" and their relationships such as "must meet" and "risks exist" are extracted and stored in the Neo4j graph database to form a basic audit knowledge graph.

[0059] The dynamic task layer can be constructed as follows: When the aforementioned "Revenue Authenticity Audit" task is executed, the system automatically creates a dedicated "Task Execution Graph." The initial node of this graph is the overall task "Company A Revenue Audit." When the document analysis agent uploads a sales contract as evidence, the system automatically creates an "Evidence File Node - Contract 001" and establishes a "support" relationship with the corresponding sub-task node. All intermediate analysis logs of all agents and discussion records between them regarding a certain point of doubt (such as "the data analysis agent questions the discrepancy between the amount in Contract 001 and the invoice") are updated in real time to this dynamic graph as attributes or new relational edges.

[0060] Furthermore, for example, the system analyzes the task execution graph in the dynamic task layer in real time during task execution. Graph algorithms (such as community discovery and influence propagation models) are used to identify the "critical path" and "weak evidence links" of the current audit task. For instance, the system might discover that multiple "supporting relationship" edges of the "Accounts Receivable Confirmation" subtask node are not connected to enough "evidence document nodes" (i.e., insufficient responses), and this node is located at the intersection of multiple analysis paths. The system can automatically issue a warning, pushing a message to the main control module and auditors: "The evidence chain for the core subtask 'Accounts Receivable Confirmation' is weak, which may affect the reliability of the conclusions of subsequent subtasks such as 'Bad Debt Provision.' It is recommended to prioritize supplementing and executing alternative procedures." This shifts the system from passive execution to proactive management.

[0061] In one optional embodiment, the nodes of the task execution graph include:

[0062] Audit task nodes, sub-task nodes, evidence document nodes, and / or data analysis nodes;

[0063] The edges of the task execution graph include:

[0064] Triggering relationships, generating relationships, supporting relationships, and / or contradictory relationships.

[0065] For example, in the dynamic task layer, the created nodes may specifically include:

[0066] Audit task node: Tag is “AuditTask”, attributes include {Task ID: T001, description: “A Company’s revenue audit”}.

[0067] Subtask nodes: labeled "SubTask", for example {Subtask ID: ST001, description: "Analyze sales contract matching"}.

[0068] Evidence file node: Tag is "EvidenceDoc", for example {File ID: D001, Name: "Sales Contract_2023-05.pdf", Hash value: xxxx}.

[0069] Data Analysis Node: Tag is "DataAnalysis", for example {Analysis ID: A001, Method: "Trend Analysis", Target: "Monthly Income Fluctuation"}.

[0070] The edges created can specifically include:

[0071] Triggering relationship: (AuditTask:T001)-[:TRIGGERS]->(SubTask:ST001)

[0072] Generate relation: (SubTask:ST001)-[:GENERATES]->(EvidenceDoc:D001)

[0073] Support relationship: (DataAnalysis:A001)-[:SUPPORTS]->(SubTask:ST001)

[0074] Conflicting Relationship: When the conclusions of two analysis nodes conflict, establish (DataAnalysis:A001)-[:CONTRADICTS]->(DataAnalysis:A002).

[0075] In one optional embodiment, the main control module is specifically used for:

[0076] The received audit task description is input into a pre-trained task parsing model, and the output is a hierarchical task tree, where the root node of the task tree is the total task and the leaf nodes are atomic subtasks.

[0077] Based on the objective of the atomic subtask, a dedicated intelligent agent module with the processing capability corresponding to the objective is matched from a preset intelligent agent capability registry.

[0078] Based on the dependencies in the task tree, a subtask scheduling sequence containing execution order and data dependencies is generated.

[0079] For example, the task parsing model is a large language model fine-tuned from audit workflow text (such as audit plan templates and past task logs). It parses the description "Perform a revenue authenticity audit on Company A's 2023 financial statements" into a task tree. The agent capability registry can be a configuration file in JSON or XML format. The main control module matches DocAnalyzer_01 based on the data_types (documents) and risk_domains (revenue recognition) of the subtask "Check sales contracts". Then, based on the task tree's requirement that "Analyze accounts receivable" must follow "Check sales contracts" (dependency), a scheduling sequence is generated: the document analysis agent is scheduled first, and after it returns the results, the data analysis agent is scheduled.

[0080] In one alternative embodiment, the agent capability registry is stored in ontology form, defining the following capability description attributes for each dedicated agent module:

[0081] Data types that can be handled, areas of expertise in risk, applicable audit phases, and / or input / output interface specifications.

[0082] For example, the ontology uses OWL (Web Ontology Language) to define core concepts. For instance, a class `AuditAgent` is defined with attributes including `canProcessDataType` (associated with the `Document` or `StructuredData` class), `expertInRiskDomain` (associated with the `FinancialRisk` or `OperationalRisk` class), and `applicableInStage` (associated with the `PlanningStage` or `ExecutionStage` class). Each specific dedicated agent module, such as `DocumentAnalyzer01`, is declared as an instance of `AuditAgent` and associated with specific values ​​through these attributes. The master module matches agents using the SPARQL query language, for example, querying all `AuditAgent` instances with `canProcessDataType` and `expertInRiskDomain` revenue recognition.

[0083] In one optional embodiment, the document analysis agent is specifically used for:

[0084] Receive sub-audit tasks and context information associated with the document to be audited, assigned by the main control module;

[0085] The document to be audited is segmented, and keywords, semantic roles, and risk statements are extracted from each segment based on the audit knowledge base.

[0086] The document language model is invoked, and based on the identified risk points and the contextual information, the analysis results are output.

[0087] For example, the document analysis agent receives a subtask package from the main control module (containing the path to the document to be audited and the context "Check revenue recognition conditions"). It first uses the Apache PDFBox library to segment the PDF contract document (by chapter or paragraph). For each segment, it calls Stanford CoreNLP for semantic role labeling, identifying roles such as "Party A," "Party B," "Delivery," and "Payment." Simultaneously, based on rules related to "revenue recognition" in the static knowledge layer (such as "Risk: The contract contains return clauses"), it identifies risk statements in the segmented text, marking sentences like "Party B has the right to unconditionally return goods within 30 days." Next, it calls a document language model (such as a version of ChatGLM3 fine-tuned with contract QA data) and uses the prompt "An unconditional return clause has been identified in the contract, which may affect revenue recognition. Please analyze the specific risks of this clause to revenue recognition this year, considering the context," to generate an analysis result text containing risk assessment, criterion citations, and audit recommendations.

[0088] In one optional embodiment, the data analysis agent is specifically used for:

[0089] Receive sub-audit tasks and analysis objectives assigned by the main control module and associated with the structured data to be audited;

[0090] Based on the risk rules and historical anomaly rules in the audit knowledge base, the structured data to be audited is analyzed;

[0091] The abnormal indicators obtained from the analysis are compared with the risk thresholds in the audit knowledge base, and the data language model is invoked to output the analysis results.

[0092] For example, the data analysis agent receives a subtask package (containing database connection information and the objective "Analyze abnormal monthly revenue fluctuations"). It calls a predefined "revenue fluctuation analysis rule set" from the rule base of the static knowledge layer, including rules such as "calculate the month-on-month growth rate" and "identify data points exceeding the historical mean ± 2 standard deviations." It connects to the database, executes the corresponding SQL queries or Python scripts, and performs the analysis. For example, it discovers a 300% month-on-month revenue increase in December 2023. Then, it compares this abnormal indicator with risk thresholds in the knowledge base (such as "monthly growth exceeding 200% requires close investigation"). Finally, it calls a data language model (such as a version fine-tuned with financial analysis instructions), inputting the prompt: "Data shows an abnormal 300% surge in revenue in December. What are the possible reasons? Please sort by probability." The model outputs a list of possible reasons, such as "year-end sales push," "early revenue recognition," and "related party transactions," as the analysis result.

[0093] In one optional embodiment, the main control module is specifically used for:

[0094] Establish a rule set for conflict detection, which includes: logical contradictions in the statements of different agents regarding the same fact, differences in the severity assessment levels of different agents regarding the same risk point exceeding a preset threshold, and / or incomplete evidence chains;

[0095] When a conflict is detected, the conflict node and its associated raw and analytical data are analyzed to obtain the conflict analysis results.

[0096] An arbitration conclusion is generated based on the conflict analysis results using a decision fusion algorithm.

[0097] For example, the main control module initiates conflict detection after receiving reports from both the document analysis agent (which believes "contract terms support revenue recognition") and the data analysis agent (which believes "the surge in revenue in December is questionable").

[0098] (1) Establish a set of rules: For example, rule 1: If the document analysis conclusion is "supports confirmation", but the data analysis conclusion is "there is an anomaly and the risk level is 'high'", then a "logical contradiction" conflict is triggered.

[0099] (2) Detection and Review: The above situation triggers rule 1. The main control module will repackage the conflict point ("revenue recognition time"), contract-related fragments, and December revenue data and send them to the two agents, requiring them to conduct cross-review. The document analysis agent needs to re-examine the contract in conjunction with the abnormal revenue data, and the data analysis agent needs to re-evaluate the cause of the abnormality in conjunction with the specific terms of the contract.

[0100] (3) Decision Fusion: Assuming that after review, the document analysis agent adds "the contract has special acceptance clauses", the data analysis agent adjusts the risk level from "high" to "medium". The main control module adopts a confidence-based fusion algorithm to give higher weight to the document analysis agent with the original evidence (contract) and finally generates the arbitration conclusion: "The income can be confirmed, but the risk of abnormal fluctuations and special contract clauses should be emphasized in the working paper".

[0101] In an optional embodiment, the system further includes a human-machine collaboration interface for:

[0102] The audit results are presented to the user in the form of a visual interactive interface;

[0103] Receive the result processing instructions input by the user through the interface;

[0104] The data stored in the system is updated according to the result processing instructions.

[0105] For example, auditors log into the system's web interface and see the audit results (a preliminary working paper) generated by the main control module in the "Audit Workbench." The sidebar highlights the "Concerns" area automatically marked by the system, which represents the conclusion after the aforementioned conflict arbitration. Auditors can click on this area to expand and view the complete conflict process, evidence from both parties, and the system's arbitration logic. Based on experience, auditors believe that the system's emphasis on the risks of special contractual clauses is insufficient, so they input an instruction through the interface: manually add an emphasis item to the "Audit Conclusion" section of the working paper. This instruction is captured by the system, which updates the data: first, it updates the final audit results (working paper); second, it stores this manual correction as a feedback sample, along with the corresponding task context, in the dynamic task layer, labeled as "Expert Amendment Example," for subsequent optimization of the conflict arbitration rules or training of the large model.

[0106] In addition, for example, after each audit task (especially complex tasks involving manual corrections) is completed, the system initiates an offline knowledge evolution process.

[0107] (1) Case accumulation: The entire dynamic task map, the final audit results and human feedback are packaged into a structured "audit case package".

[0108] (2) Pattern mining: Using graph pattern mining algorithms, we search for frequently occurring combinations of "risk pattern - audit procedure - valid evidence" in a large number of historical case packages. For example, when we find that "pattern X: sales contract contains ambiguous terms & revenue increases significantly at the end of the quarter", executing "procedure Y: check delivery note and customer signature record" and obtaining "evidence Z: third-party logistics signature" can most effectively confirm or eliminate the risk.

[0109] (3) Knowledge Feedback: The effective patterns discovered are transformed into new structured rules or entity relationships after being reviewed by audit experts and actively injected into the static knowledge layer. For example, a new rule is added: "If 'pattern X' is detected, it is recommended to automatically create and schedule the execution of 'program Y'." This allows the system's static knowledge base to be continuously enriched and refined automatically with the accumulation of practical experience.

[0110] Based on the above structure, the large-scale intelligent agent collaborative system for the auditing industry provided by this invention achieves an intelligent upgrade of auditing operations by constructing a system architecture in which a main control module, multiple specialized intelligent agents, and an audit knowledge base work collaboratively. Specifically, the system can automatically decompose and plan complex audit processes based on audit task descriptions, call upon document analysis and data analysis intelligent agents configured with domain-specific large models to execute professional tasks in parallel, and ensure the professionalism and context awareness of the analysis by combining static and dynamic knowledge bases. In addition, the system introduces a conflict arbitration mechanism to automatically reconcile contradictions in multi-source analysis results, forming reliable unified conclusions, significantly improving the execution efficiency, analysis depth, and consistency of audit tasks. Finally, through a human-machine collaborative interface, the system effectively combines intelligent analysis capabilities with the experience and judgment of human experts, constructing a virtuous cycle of continuous optimization and increasing intelligence with use.

[0111] Furthermore, it should be noted that the present invention can be provided as a method, apparatus, or computer program product. Therefore, embodiments of the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, embodiments of the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code.

[0112] Embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0113] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing terminal equipment to cause a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0114] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0115] Finally, it should be noted that the above description represents a preferred embodiment of the present invention. It should be pointed out that although preferred embodiments have been described, those skilled in the art, once they understand the basic inventive concept of the present invention, can make various improvements and modifications without departing from the principles described herein. These improvements and modifications should also be considered within the scope of protection of the present invention. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the embodiments of the present invention.

Claims

1. A large-scale intelligent agent collaborative system for the auditing industry, characterized in that, It includes a main control module, at least two dedicated intelligent agent modules, and an audit knowledge base, among which: The audit knowledge base stores knowledge graphs and rules in the audit field; The main control module is used to receive the audit task description, break it down into sub-audit tasks, and schedule the corresponding dedicated intelligent agent module to execute according to the task type of the sub-audit task; The at least two dedicated intelligent agent modules include: The document analysis agent is equipped with a document language model, which is used to analyze the documents to be audited in conjunction with the audit knowledge base. The data analysis intelligent agent is equipped with a data language model, which is used to analyze the structured data to be audited in conjunction with the audit knowledge base. The main control module is also used to perform conflict arbitration and fusion on the analysis results returned by the at least two dedicated intelligent agent modules to generate audit results.

2. The large-scale intelligent agent collaborative system for the auditing industry as described in claim 1, characterized in that, The audit knowledge base includes: The static knowledge layer is used to store entities, relationships, and rules extracted and structured from auditing standards, regulations, and / or historical audit case databases, forming the basic audit knowledge graph; The dynamic task layer is used to store the task context, intermediate evidence, and / or the interaction records of the at least two dedicated intelligent agents derived from the audit task, forming a task execution graph.

3. The large-scale intelligent agent collaborative system for the auditing industry as described in claim 2, characterized in that, The nodes of the task execution graph include: Audit task nodes, sub-task nodes, evidence document nodes, and / or data analysis nodes; The edges of the task execution graph include: Triggering relationships, generating relationships, supporting relationships, and / or contradictory relationships.

4. The large-scale intelligent agent collaborative system for the auditing industry as described in claim 1, characterized in that, The main control module is specifically used for: The received audit task description is input into a pre-trained task parsing model, and the output is a hierarchical task tree, where the root node of the task tree is the total task and the leaf nodes are atomic subtasks. Based on the objective of the atomic subtask, a dedicated intelligent agent module with the processing capability corresponding to the objective is matched from a preset intelligent agent capability registry. Based on the dependencies in the task tree, a subtask scheduling sequence containing execution order and data dependencies is generated.

5. The large-scale intelligent agent collaborative system for the auditing industry as described in claim 4, characterized in that, The agent capability registry is stored in ontology form, defining the following capability description attributes for each dedicated agent module: Data types that can be handled, areas of expertise in risk, applicable audit phases, and / or input / output interface specifications.

6. The large-scale intelligent agent collaborative system for the auditing industry as described in claim 1, characterized in that, The document analysis agent is specifically used for: Receive sub-audit tasks and context information associated with the document to be audited, assigned by the main control module; The document to be audited is segmented, and keywords, semantic roles, and risk statements are extracted from each segment based on the audit knowledge base. The document language model is invoked, and based on the identified risk points and the contextual information, the analysis results are output.

7. The large-scale intelligent agent collaborative system for the auditing industry as described in claim 1, characterized in that, The data analysis intelligent agent is specifically used for: Receive sub-audit tasks and analysis objectives assigned by the main control module and associated with the structured data to be audited; Based on the risk rules and historical anomaly rules in the audit knowledge base, the structured data to be audited is analyzed; The abnormal indicators obtained from the analysis are compared with the risk thresholds in the audit knowledge base, and the data language model is invoked to output the analysis results.

8. The large-scale intelligent agent collaborative system for the auditing industry as described in claim 1, characterized in that, The main control module is specifically used for: Establish a rule set for conflict detection, which includes: logical contradictions in the statements of different agents regarding the same fact, differences in the severity assessment levels of different agents regarding the same risk point exceeding a preset threshold, and / or incomplete evidence chains; When a conflict is detected, the conflict node and its associated raw and analytical data are analyzed to obtain the conflict analysis results. An arbitration conclusion is generated based on the conflict analysis results using a decision fusion algorithm.

9. The large-scale intelligent agent collaborative system for the auditing industry as described in claim 1, characterized in that, The system also includes a human-machine collaboration interface for: The audit results are presented to the user in the form of a visual interactive interface; Receive the result processing instructions input by the user through the interface; The data stored in the system is updated according to the result processing instructions.

Images