An intelligent auditing method and system based on large model and rule engine fusion

Abstract

The application discloses an intelligent auditing method and system based on large model and rule engine fusion, relates to the technical field of information processing, acquires financial auditing business data to be audited, calls a multi-modal information extraction model to perform structured information extraction on the financial auditing business data, obtains structured auditing data containing key fields and corresponding values, calls a large model to perform semantic analysis on auditing basis documents and historical auditing data, generates structured rules, and deploys the structured rules to a rule engine, inputs the structured auditing data into the rule engine and the large model decision simultaneously, respectively obtains deterministic auditing results and confidence auditing results, generates a final auditing conclusion, matches a corresponding early warning level, executes a disposal action corresponding to the early warning level, and generates a structured auditing report. Through automatic understanding and learning of unstructured auditing basis, dynamic generation of auditing rules, and hierarchical processing of auditing results, the auditing efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of information processing technology, specifically to an intelligent auditing method and system based on the integration of large models and rule engines. Background Technology

[0002] As enterprises deepen their digital transformation, the auditing operations of financial shared service centers face the dual challenges of efficiency and quality. Currently, financial auditing mainly relies on manual checks of expense reports, invoice images, contract attachments, and other materials for compliance. This traditional model has the following significant technical problems: 1. Low review efficiency and high subjectivity: Manual review requires checking each order individually, which is time-consuming and labor-intensive. In addition, the experience and standards of different reviewers vary, resulting in poor consistency of review results and easy misjudgment and omission.

[0003] 2. Delayed rule updates and maintenance difficulties: Internal company rules and regulations, as well as external accounting regulations and tax policies, change frequently. Traditional automated auditing systems based on hard-coded or simple rules rely entirely on manual code or configuration modifications by technical personnel for rule updates, resulting in long response times and an inability to achieve real-time risk control.

[0004] 3. Inability to handle unstructured information: Existing automated auditing solutions excel at handling structured data and deterministic process operations. However, a large amount of auditing evidence (such as policy PDFs and documents) and content to be audited (such as reimbursement reasons and invoice remarks) is unstructured text. Current technology lacks a means to effectively understand this unstructured information and transform it into executable auditing logic.

[0005] 4. The review mechanism is too simplistic and lacks tiered early warning: Existing technologies often adopt a binary result of "pass / reject," which cannot differentiate the severity of non-compliance (such as exceeding the amount limit or typographical errors in attachment names), resulting in insufficiently refined review results that cannot meet the needs of enterprises for flexible management or strong control.

[0006] Therefore, there is an urgent need for an intelligent auditing method to overcome the shortcomings of existing technologies. Summary of the Invention

[0007] The purpose of this invention is to provide an intelligent review method and system based on the integration of large models and rule engines, which automatically understands and learns unstructured review criteria, dynamically generates review rules, and performs hierarchical processing of review results to improve review efficiency.

[0008] To achieve the above objectives, the following plan is adopted: On the one hand, this invention provides an intelligent review method based on the fusion of a large model and a rule engine, specifically including the following steps: S1. Obtain the financial audit business data to be audited, call the multimodal information extraction model to extract structured information from the financial audit business data, and obtain structured audit data containing key fields and their corresponding values. S2. Build and execute a fusion audit model to obtain unstructured audit basis documents and historical audit data for preprocessing; S3. Call the large model to perform semantic parsing on the preprocessed review basis documents and historical review data, generate structured rules, and deploy the structured rules to the rule engine; S4. Input the structured audit data into the rule engine and the big model decision simultaneously to obtain the deterministic audit results and the confidence audit results, respectively. S5. Generate the final audit conclusion based on the deterministic audit results and the confidence audit results, match the corresponding warning level, execute the corresponding handling actions, and generate a structured audit report.

[0009] In some specific implementation schemes, the process of generating structured rules is as follows: The large model is invoked to perform semantic understanding on unstructured audit basis documents, extract audit points, compliance standards and logical relationships, and convert natural language descriptions into structured intermediate representations; Obtain historical review business data, filter out the rejected abnormal data, use a large model to perform cluster analysis and pattern recognition on the abnormal data, and determine the high-frequency review risk themes and corresponding field association rules; The structured intermediate representation and field association rules are automatically compiled into structured rules that can be executed by the rule engine using a large model. The structured rules include condition judgment, threshold comparison and field association verification logic. The generated structured rules are validated and tested in an isolated environment, and then deployed to the rule engine after the validation test is passed.

[0010] In some specific implementation schemes, step S4 is performed as follows: The rules engine is used to process quantitative audit rules, perform precise field comparisons, numerical range checks, logical consistency checks, and format checks on structured audit data, and output deterministic audit results. The large model decision engine is invoked to process qualitative review rules. Based on the context of business documents and extracted attachment information, the compliance of principle clauses in structured review data is evaluated, and review results with confidence scores are output.

[0011] In some specific implementation plans, the process of generating the final audit conclusion is as follows: When the deterministic audit result is non-compliant, the final audit conclusion is output as non-compliant, and a strong audit warning is triggered; When the deterministic audit result is compliant or uncertain, the confidence audit result will be used as the final audit conclusion and a graded decision will be made.

[0012] In some specific implementation plans, the tiered adjudication process is as follows: If the confidence level review result is greater than or equal to the first threshold, the final review conclusion is compliance and no warning will be triggered; If the confidence level audit result is greater than or equal to the second threshold and less than the first threshold, the final audit conclusion is compliant, triggering a weak audit warning. If the confidence level audit result is less than the second threshold, the final audit conclusion is non-compliance, triggering a strong audit warning.

[0013] In some specific implementation plans, the warning levels include strong review warnings and weak review warnings. The specific process for executing the corresponding actions for each warning level is as follows: The corresponding action for a strong review warning is to automatically reject the current review application and indicate the reason for rejection and the specific rules triggered in the generated structured review report; The corresponding action for a weak review warning is to generate a prompt message in the current review process and record the violation, but without interrupting the review process.

[0014] In some specific implementations, the multimodal information extraction model integrates at least one of the following extraction methods: Template extraction: Invokes a pre-configured extraction template to extract specified fields from an image or document; Custom extraction: Receives user-defined field configuration information online and extracts it; Intelligent extraction: The large model intelligent agent is invoked to automatically identify and extract key information from images or documents.

[0015] In some specific implementation plans, the financial audit data to be audited is obtained through at least one of the following methods: Obtained from the Enterprise Resource Planning (ERP) system via the data feedback interface; The intelligent RPA robot captures data and image attachments from the financial system interface in real time. The data entry and reporting tool is used to collect structured data entered by the business departments.

[0016] Secondly, this application provides an intelligent review system based on the fusion of a large model and a rule engine, including: The information extraction module is used to obtain the financial audit business data to be audited, and call the multimodal information extraction model to extract structured information from the financial audit business data to obtain structured audit data containing key fields and their corresponding values. The review module is used to build and execute the integrated review model. The integrated review model adopts a four-layer architecture: rule input layer, rule parsing and enhancement layer, intelligent review execution layer, and result output layer. The rule input layer is used to preprocess the unstructured audit basis documents and historical audit data. The rule parsing and enhancement layer calls the large model to perform semantic parsing on the preprocessed review basis documents and historical review data, generate structured rules, and deploy the structured rules to the rule engine; The intelligent audit execution layer simultaneously inputs structured audit data into the rule engine and the large model decision engine to obtain deterministic audit results and confidence audit results, respectively. The results output layer is used to generate a final audit conclusion based on the deterministic audit results and the confidence audit results, match the corresponding warning level, execute the handling action corresponding to the warning level, and generate a structured audit report.

[0017] In some specific implementations, the rule parsing and enhancement layer specifically includes a rule semantic parser, a rule code generator, and a rule sandbox, wherein: The rule semantic parser calls the large model to perform semantic understanding on unstructured audit basis documents, extract audit points, compliance standards and logical relationships, and convert natural language descriptions into structured intermediate representations; it obtains historical audit business data, filters out the rejected abnormal data, and uses the large model to perform cluster analysis and pattern recognition on the abnormal data to determine high-frequency audit risk themes and corresponding field association rules; The rule code generator is used to automatically compile the structured intermediate representation and field association rules into structured rules that can be executed by the rule engine using a large model. The structured rules include condition judgment, threshold comparison and field association verification logic. The rule sandbox is used to validate and test the generated structured rules in an isolated environment. Once the validation test is passed, the rules are deployed to the rule engine.

[0018] The beneficial effects of this invention are as follows: This application utilizes a large model-driven rule semantic parser and code generator to automatically transform unstructured systems, regulations, and cases into executable structured rule code, enabling business personnel to create, sandbox test, and deploy rules without programming, compressing the update cycle of audit rules from weeks to minutes, and achieving real-time response of risk control strategies. Based on this, a dual-engine execution architecture was constructed, in which the rule engine and the big model decision engine run in parallel. The rule engine is responsible for the deterministic and quantitative review of amounts and fields, while the big model decision engine is responsible for the qualitative and semantic judgments such as the rationality of the cause and the semantic consistency of the attachments. The two complement each other and work together, which not only avoids the shortcomings of the pure rule engine in handling semantic problems, but also overcomes the disadvantages of the pure big model being costly and having uncontrollable results. Furthermore, a rule-first, confidence-level integrated adjudication mechanism is adopted. When the rule engine determines non-compliance, it directly rejects the decision to ensure rigid constraints. When the rule is compliant or missing, the result is divided into three levels: no warning, weak review, and strong review, based on the confidence threshold. The threshold can be dynamically configured to achieve refined and differentiated handling of different risk levels. Strong reviews are automatically rejected with the reason marked, while weak reviews are handled with reminders to ensure the process is not interrupted. This effectively controls core financial risks while ensuring business efficiency and avoids a one-size-fits-all, extensive management approach. Attached Figure Description

[0019] Figure 1 A flowchart illustrating the intelligent extraction method for multimodal image information provided in this embodiment of the invention. Detailed Implementation

[0020] 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. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. 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.

[0021] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention.

[0022] Furthermore, for clarity and brevity, descriptions of well-known functions and methods may have been omitted. Those skilled in the art will recognize that various changes and modifications can be made to the examples described herein without departing from the spirit and scope of this disclosure.

[0023] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0024] In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0025] Example 1 like Figure 1 As shown in the figure, this embodiment 1 provides an intelligent review method based on the fusion of a large model and a rule engine, which specifically includes the following steps: S1. Obtain the financial audit business data to be audited, call the multimodal information extraction model to extract structured information from the financial audit business data, and obtain structured audit data containing key fields and their corresponding values. Because corporate financial data is scattered across different systems, and some systems (such as outdated unified systems) do not provide data interfaces, data can only be obtained through RPA to simulate manual operations. Furthermore, certain business indicators (such as budget adjustment explanations) may not have been entered into any system at all, requiring manual input by business personnel. Therefore, multiple data acquisition methods need to work together to build a complete data foundation. Financial audit business data awaiting review should be obtained through at least one of the following methods: Obtained from the Enterprise Resource Planning (ERP) system via the data feedback interface; The intelligent RPA robot captures data and image attachments from the financial system interface in real time. The data entry and reporting tool is used to collect structured data entered by the business departments.

[0026] The three methods complement each other, ensuring that the audit system can obtain complete and accurate data to be audited.

[0027] Furthermore, the types of attachments involved in financial audits are extremely diverse: VAT invoices have a fixed format, but contracts, acceptance forms, and project progress reports vary in format. Therefore, to avoid a single extraction method failing to meet all scenarios, the multimodal information extraction model integrates at least one of the following extraction methods: Template extraction: Invokes a pre-configured extraction template to extract specified fields from an image or document; Custom extraction: Receives user-defined field configuration information online and extracts it; Intelligent extraction: The large model intelligent agent is invoked to automatically identify and extract key information from images or documents.

[0028] Template extraction is suitable for high-frequency, fixed-format documents (such as invoices) and is highly efficient; custom extraction is suitable for temporary and personalized extraction tasks; and intelligent extraction is suitable for documents with unknown formats or complex layouts and offers the greatest flexibility.

[0029] S2. Build and execute a fusion audit model to obtain unstructured audit basis documents and historical audit data for preprocessing; S3. Call the large model to perform semantic parsing on the preprocessed review basis documents and historical review data, generate structured rules, and deploy the structured rules to the rule engine; Since financial audit rules typically exist in the form of natural language within regulations, laws, and case studies, directly converting them into code manually is inefficient and prone to errors. A two-step process using a large model (first parsing into an intermediate representation, then generating code) leverages the model's semantic understanding capabilities while ensuring the structured and executable nature of the output code. The process of generating structured rules is as follows: S31. Call the large model to perform semantic understanding on unstructured audit basis documents, extract audit points, compliance standards and logical relationships, and convert natural language descriptions into structured intermediate representations; S32. Obtain historical review business data, filter out the rejected abnormal data, use a large model to perform cluster analysis and pattern recognition on the abnormal data, and determine the high-frequency review risk themes and corresponding field association rules. S33. The structured intermediate representation and field association rules are automatically compiled into structured rules that can be executed by the rule engine using the large model. The structured rules include condition judgment, threshold comparison and field association verification logic. S34. Perform verification tests on the generated structured rules in an isolated environment, and deploy them to the rule engine after the verification tests are passed.

[0030] S4. Simultaneously input the structured audit data into the rule engine and the large model decision-making process to obtain deterministic audit results and confidence audit results, where: S41, the rule engine is used to process quantitative audit rules, perform precise field comparison, numerical range verification, logical consistency check and format verification on structured audit data, and output deterministic audit results; S42. Call the large model decision engine to process qualitative review rules. Based on the business document context and extracted attachment information, evaluate the compliance of the principle clauses in the structured review data and output the review results with confidence scores.

[0031] S5. Generate the final audit conclusion based on the deterministic audit results and the confidence audit results, match the corresponding warning level, execute the corresponding handling actions, and generate a structured audit report.

[0032] The process of generating the final review conclusion is as follows: S51. When the deterministic audit result is non-compliant, output the final audit conclusion as non-compliant and trigger a strong audit warning; S52. When the deterministic audit result is compliant or uncertain, the confidence level audit result shall be used as the final audit conclusion and a tiered decision shall be made. The tiered decision process is as follows: If the confidence level audit result is greater than or equal to the first threshold, the final audit conclusion is compliant and no warning is triggered; if the confidence level audit result is greater than or equal to the second threshold and less than the first threshold, the final audit conclusion is compliant and a weak audit warning is triggered; if the confidence level audit result is less than the second threshold, the final audit conclusion is non-compliant and a strong audit warning is triggered.

[0033] The warning levels include strong review warnings and weak review warnings. The specific process for implementing the corresponding actions for each warning level is as follows: The corresponding action for a strong review warning is to automatically reject the current review application and indicate the reason for rejection and the specific rules triggered in the generated structured review report; The corresponding action for a weak review warning is to generate a prompt message in the current review process and record the violation, but without interrupting the review process.

[0034] Understandably, in existing technologies, review rules rely on manual coding, resulting in slow response times after policy updates; furthermore, traditional rule engines cannot handle unstructured semantic judgments. Through a four-layer architecture design, a large model is introduced into the rule generation and qualitative judgment stages, integrating adjudication rules (rule priority, confidence level). Quantitative rules processed by the rule engine (such as exceeding the limit) are mandatory and cannot be overridden by a high model confidence level; while qualitative rules processed by the large model have uncertainty, requiring confidence thresholds to distinguish between three states: certain compliance, questionable, and certain non-compliance, thus achieving flexible control. A fallback adjudication mechanism for missing rules ensures the system can still operate even when the rule base is incomplete.

[0035] This embodiment achieves both the accuracy of quantitative review and the semantic understanding capability of qualitative review through the parallel execution of the rule engine and the large model decision engine. Employing a multi-level confidence threshold and rule-priority fusion adjudication mechanism, it can finely classify review results (no warning / weak warning / strong warning) and automatically match corresponding actions (approval / reminder / rejection), significantly improving the automation level of review and the differentiated risk control capabilities. It realizes fully automated generation from unstructured review basis to executable rules, establishing a closed loop from system → rule → review.

[0036] Example 2 This embodiment provides an intelligent review system based on the fusion of a large model and a rule engine, implementing the method of Embodiment 1, including: The information extraction module is used to obtain the financial audit business data to be audited, and call the multimodal information extraction model to extract structured information from the financial audit business data to obtain structured audit data containing key fields and their corresponding values. The review module is used to build and execute the integrated review model. The integrated review model adopts a four-layer architecture: rule input layer, rule parsing and enhancement layer, intelligent review execution layer, and result output layer. The rule input layer is used to preprocess the unstructured audit basis documents and historical audit data. The rule parsing and enhancement layer calls the large model to perform semantic parsing on the preprocessed review basis documents and historical review data, generate structured rules, and deploy the structured rules to the rule engine; The rule parsing and enhancement layer specifically includes a rule semantic parser, a rule code generator, and a rule sandbox, among which: The rule semantic parser calls the large model to perform semantic understanding on unstructured audit basis documents, extract audit points, compliance standards and logical relationships, and convert natural language descriptions into structured intermediate representations; it obtains historical audit business data, filters out the rejected abnormal data, and uses the large model to perform cluster analysis and pattern recognition on the abnormal data to determine high-frequency audit risk themes and corresponding field association rules; The rule code generator is used to automatically compile the structured intermediate representation and field association rules into structured rules that can be executed by the rule engine using a large model. The structured rules include condition judgment, threshold comparison and field association verification logic. The rule sandbox is used to validate and test the generated structured rules in an isolated environment. Once the validation test is passed, the rules are deployed to the rule engine.

[0037] The intelligent audit execution layer simultaneously inputs structured audit data into the rule engine and the large model decision engine to obtain deterministic audit results and confidence audit results, respectively. The results output layer is used to generate a final audit conclusion based on the deterministic audit results and the confidence audit results, match the corresponding warning level, execute the handling action corresponding to the warning level, and generate a structured audit report.

[0038] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Based on the technical essence of the present invention, any simple modifications, equivalent substitutions, and improvements made to the above embodiments within the spirit and principles of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An intelligent review method based on the fusion of large-scale models and rule engines, characterized in that, Specifically, the following steps are included: S1. Obtain the financial audit business data to be audited, call the multimodal information extraction model to extract structured information from the financial audit business data, and obtain structured audit data containing key fields and their corresponding values. S2. Build and execute a fusion audit model to obtain unstructured audit basis documents and historical audit data for preprocessing; S3. Call the large model to perform semantic parsing on the preprocessed review basis documents and historical review data, generate structured rules, and deploy the structured rules to the rule engine; S4. Input the structured audit data into the rule engine and the big model decision simultaneously to obtain the deterministic audit results and the confidence audit results, respectively. S5. Generate the final audit conclusion based on the deterministic audit results and the confidence audit results, match the corresponding warning level, execute the corresponding handling actions, and generate a structured audit report.

2. The intelligent review method based on the fusion of a large model and a rule engine according to claim 1, characterized in that, The process of generating structured rules is as follows: The large model is invoked to perform semantic understanding on unstructured audit basis documents, extract audit points, compliance standards and logical relationships, and convert natural language descriptions into structured intermediate representations; Obtain historical review business data, filter out the rejected abnormal data, use a large model to perform cluster analysis and pattern recognition on the abnormal data, and determine the high-frequency review risk themes and corresponding field association rules; The structured intermediate representation and field association rules are automatically compiled into structured rules that can be executed by the rule engine using a large model. The structured rules include condition judgment, threshold comparison and field association verification logic. The generated structured rules are validated and tested in an isolated environment, and then deployed to the rule engine after the validation test is passed.

3. The intelligent review method based on the fusion of a large model and a rule engine according to claim 1, characterized in that, The specific process of step S4 is as follows: The rules engine is used to process quantitative audit rules, perform precise field comparisons, numerical range checks, logical consistency checks, and format checks on structured audit data, and output deterministic audit results. The large model decision engine is invoked to process qualitative review rules. Based on the context of business documents and extracted attachment information, the compliance of principle clauses in structured review data is evaluated, and review results with confidence scores are output.

4. The intelligent review method based on the fusion of a large model and a rule engine according to claim 1, characterized in that, The process of generating the final review conclusion is as follows: When the deterministic audit result is non-compliant, the final audit conclusion is output as non-compliant, and a strong audit warning is triggered; When the deterministic audit result is compliant or uncertain, the confidence audit result will be used as the final audit conclusion and a graded decision will be made.

5. The intelligent review method based on the fusion of a large model and a rule engine according to claim 4, characterized in that, The process of tiered adjudication is as follows: If the confidence level review result is greater than or equal to the first threshold, the final review conclusion is compliance and no warning will be triggered; If the confidence level audit result is greater than or equal to the second threshold and less than the first threshold, the final audit conclusion is compliant, triggering a weak audit warning. If the confidence level audit result is less than the second threshold, the final audit conclusion is non-compliance, triggering a strong audit warning.

6. The intelligent review method based on the fusion of a large model and a rule engine according to claim 5, characterized in that, The warning levels include strong review warnings and weak review warnings. The specific process for implementing the corresponding actions for each warning level is as follows: The corresponding action for a strong review warning is to automatically reject the current review application and indicate the reason for rejection and the specific rules triggered in the generated structured review report; The corresponding action for a weak review warning is to generate a prompt message in the current review process and record the violation, but without interrupting the review process.

7. The intelligent review method based on the fusion of a large model and a rule engine according to claim 1, characterized in that, The multimodal information extraction model integrates at least one of the following extraction methods: Template extraction: Invokes a pre-configured extraction template to extract specified fields from an image or document; Custom extraction: Receives user-defined field configuration information online and extracts it; Intelligent extraction: The large model intelligent agent is invoked to automatically identify and extract key information from images or documents.

8. The intelligent review method based on the fusion of a large model and a rule engine according to claim 1, characterized in that, Financial audit data pending review is obtained through at least one of the following methods: Obtained from the Enterprise Resource Planning (ERP) system via the data feedback interface; The intelligent RPA robot captures data and image attachments from the financial system interface in real time. The data entry and reporting tool is used to collect structured data entered by the business departments.

9. An intelligent review system based on the integration of large-scale models and rule engines, characterized in that, include: The information extraction module is used to obtain the financial audit business data to be audited, and call the multimodal information extraction model to extract structured information from the financial audit business data to obtain structured audit data containing key fields and their corresponding values. The review module is used to build and execute the integrated review model. The integrated review model adopts a four-layer architecture: rule input layer, rule parsing and enhancement layer, intelligent review execution layer, and result output layer. The rule input layer is used to preprocess the unstructured audit basis documents and historical audit data. The rule parsing and enhancement layer calls the large model to perform semantic parsing on the preprocessed review basis documents and historical review data, generate structured rules, and deploy the structured rules to the rule engine; The intelligent audit execution layer simultaneously inputs structured audit data into the rule engine and the large model decision engine to obtain deterministic audit results and confidence audit results, respectively. The results output layer is used to generate a final audit conclusion based on the deterministic audit results and the confidence audit results, match the corresponding warning level, execute the handling action corresponding to the warning level, and generate a structured audit report.

10. The intelligent review system based on the fusion of a large model and a rule engine according to claim 9, characterized in that, The rule parsing and enhancement layer specifically includes a rule semantic parser, a rule code generator, and a rule sandbox, among which: The rule semantic parser calls the large model to perform semantic understanding on unstructured audit basis documents, extract audit points, compliance standards and logical relationships, and convert natural language descriptions into structured intermediate representations; it obtains historical audit business data, filters out the rejected abnormal data, and uses the large model to perform cluster analysis and pattern recognition on the abnormal data to determine high-frequency audit risk themes and corresponding field association rules; The rule code generator is used to automatically compile the structured intermediate representation and field association rules into structured rules that can be executed by the rule engine using a large model. The structured rules include condition judgment, threshold comparison and field association verification logic. The rule sandbox is used to validate and test the generated structured rules in an isolated environment. Once the validation test is passed, the rules are deployed to the rule engine.

Images