Intelligent compliance test method

Abstract

The invention discloses an intelligent compliance test method. The test method comprises the following steps: receiving a to-be-tested item or question input by a user; entering a compliance test layer, and preliminarily matching task requirements with a built-in knowledge base; entering two calling agent layers and a calling large model layer in parallel, and respectively processing structured rule and non-structured complex problems; and outputting a comprehensive research and judgment result to a user. Through collaborative innovation of an MOE large model, multiple agents and a multi-source data platform, a new-generation compliance inspection system with specialty, intelligence, real-time performance and expandability is successfully constructed, the increasingly complex commercial compliance environment can be effectively dealt with, and powerful technical support is provided for risk management and control of enterprises in digital transformation.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of vertical domain large model construction of artificial intelligence, and particularly relates to an intelligent compliance verification method. BACKGROUND

[0002] The existing technologies disclosed in the bidding field are mainly concentrated in the electronicization and digitization of the bidding process, and a few technical inventions based on large models are also concentrated in individual bidding links and are oriented to specific subjects or fields. A power industry bidding large model construction system and method based on professional knowledge corpus focuses on the fusion of the professional knowledge of the power industry by the large model, and a supplier risk automatic evaluation control method and system in the bidding process mainly focuses on the method of using historical data to evaluate the risk of suppliers in the bid opening stage.

[0003] The bidding prices of multiple bidders are detected, the regularity of the bidding prices and the possibility of low-price bidding by several bidders are evaluated, and the bidding price curve of the subject matter is obtained through a model index constructed by historical big data to evaluate the abnormality of the bidding prices, so as to obtain a bidding price risk evaluation result. A supplier risk evaluation report is generated according to the credit risk and enterprise operation risk evaluation result of the bidder, the bidder correlation risk evaluation result, the bidding behavior correlation risk evaluation result, the bidding document correlation risk evaluation result and the bidding price risk evaluation result.

[0004] As described above, the detection method in the prior art has the risk of missed detection and false detection. SUMMARY

[0005] In view of the above problems, the present application is proposed to provide an intelligent compliance verification method to overcome the above problems or at least partially solve the above problems.

[0006] According to one aspect of the present application, an intelligent compliance verification method is provided, which comprises: receiving a user-inputted item or question to be verified; entering a compliance verification layer to preliminarily match the task requirement with an internal knowledge base; entering two calling intelligent agent layers and calling large model layers in parallel to process structured rules and unstructured complex problems, respectively; outputting the result of comprehensive research and judgment to the user.

[0007] Optionally, the knowledge base specifically includes policy documents, bidding and tendering laws, government procurement laws, company internal management systems, expert interpretations and related cases.

[0008] Optionally, the calling intelligent agent layer specifically includes: constructing a regulation and system expert model cluster to cover and iterate the professional field.

[0009] Optionally, the construction of a cluster of expert models for regulations and systems, and the coverage and iteration of professional fields, specifically includes: Deploy multiple specialized intelligent agents, including agent 1, agent 2, ..., agent n; The aforementioned expert model of regulations and systems is continuously learned and updated; By processing different compliance subdomains with intelligent agents, we have achieved in-depth vertical mining of knowledge and specialized division of labor.

[0010] Optionally, the compliance sub-fields specifically include: tax compliance, environmental compliance, and bidding compliance.

[0011] Optionally, the invocation of the large model layer specifically includes: adopting an advanced hybrid expert MOE large model architecture and distributed computation of complex problems.

[0012] Optionally, the advanced hybrid expert MOE large model architecture is adopted, and the distributed computation of complex problems specifically includes: Intelligent routing in gated networks: Gated networks first analyze and determine the type and complexity of the problem; Distributed processing of expert model pool: Based on the analysis results, the gating network dynamically selects one or more of the most relevant expert models from multiple expert models to collaboratively process the current task; Results Integration and Output: The selected expert models each contribute their professional analysis, and the results are finally summarized to form a comprehensive judgment.

[0013] Optionally, the results of the comprehensive analysis and evaluation output to the user specifically include: Integrate multi-dimensional data sources; A closed-loop verification mechanism is formed, consisting of rule-semantic-data triple verification.

[0014] Optionally, the closed-loop verification mechanism of rule-semantics-data triple verification specifically includes: Rule-driven verification of intelligent agents; MOE large-scale model semantic understanding and complex reasoning, handles ambiguity, performs case comparisons and provides in-depth risk insights; Fact verification of underlying data.

[0015] This invention provides an intelligent compliance verification method, which includes: receiving user input of items or questions to be verified; entering a compliance verification layer to initially match task requirements with a built-in knowledge base; parallelly entering two agent invocation layers and a large model invocation layer to handle structured rules and unstructured complex problems respectively; and outputting the comprehensive judgment result to the user. Through the collaborative innovation of the MOE large model, multiple agents, and a multi-source data platform, a new generation of compliance verification system with professionalism, intelligence, real-time performance, and scalability has been successfully constructed. This system can effectively cope with the increasingly complex business compliance environment and provides strong technical support for enterprises' risk management in digital transformation.

[0016] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the 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.

[0018] Figure 1 A flowchart of an intelligent compliance verification method provided in an embodiment of the present invention. Detailed Implementation

[0019] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0020] The terms "comprising" and "having," and any variations thereof, in the specification, embodiments, claims, and drawings of this invention are intended to cover non-exclusive inclusion, such as including a series of steps or units.

[0021] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0022] Example 1 This invention constructs an efficient, accurate, and scalable intelligent compliance verification method that integrates the decision-making capabilities of a Mixture of Experts (MOE) model with the collaborative workflow of multiple agents. By performing real-time analysis of massive amounts of multi-source data, it provides users with a one-stop procurement model service.

[0023] like Figure 1 As shown, the overall architecture adopts layered processing and intelligent routing, and the workflow is divided into four core stages, forming a complete processing loop.

[0024] First, the system receives the items or questions to be tested from the user's input.

[0025] Next, the system enters the "compliance verification layer," which acts as the "brain" of the system and is responsible for overall coordination. It performs an initial match between task requirements and the built-in "knowledge base" (including policy documents, bidding laws, government procurement laws, internal management systems, expert interpretations, and relevant cases).

[0026] Next, the process proceeds in parallel to two core computing layers: the "calling the agent" layer and the "calling the large model" layer, which respectively handle structured rules and unstructured complex problems.

[0027] Finally, the results, after comprehensive analysis, are output to the user. This layered architecture ensures the orderly, modular, and highly cohesive yet loosely coupled nature of the processing.

[0028] Construct a cluster of expert models for laws and regulations to achieve coverage of professional fields and rapid iteration, including: In the process of invoking intelligent agents, multiple specialized intelligent agents (Agent 1, Agent 2, ... Agent n) are deployed. Among them, the "Legal and Regulatory Expert Model" is particularly emphasized as a key intelligent agent, possessing the ability to update rapidly. The expert model can continuously learn the latest laws, regulations, policy documents, and judicial interpretations, ensuring the timeliness and accuracy of its knowledge base. By entrusting different compliance sub-domains (such as tax compliance, environmental compliance, and bidding compliance) to the most specialized intelligent agents, in-depth vertical knowledge mining and specialized division of labor are achieved, enhancing the authority of the consultation conclusions.

[0029] The large model layer is invoked, and the advanced MOE (Hybrid Expert) architecture is adopted. By introducing the MOE large model architecture, accurate and efficient distributed computing of complex problems can be achieved.

[0030] Its working mechanism is as follows: 1. Gating Network Intelligent Routing: When input enters this layer, the gating network first analyzes it to determine the type and complexity of the problem.

[0031] 2. Distributed processing of expert model pool (ExpertsPool): Based on the analysis results, the gating network dynamically selects one or several most relevant expert models (such as experts focusing on contract review, experts focusing on risk identification, etc.) from the huge pool of expert models 1, 2, 3...n to collaboratively process the current task.

[0032] 3. Results Integration and Output: Each selected expert model contributes its professional analysis, and their results are ultimately aggregated to form a comprehensive judgment. This "divide and conquer" strategy enables the system to mobilize the professional capabilities of models with hundreds of billions of parameters at a computational cost far lower than activating all parameters, achieving a balance between high efficiency and high accuracy in handling complex, unstructured compliance issues.

[0033] Integrating multi-dimensional data sources provides a solid data foundation for compliant decision-making.

[0034] The system explicitly integrates judicial data, business registration data, bidding data, credit data, and patent data. These data collectively constitute the system's "fact base." For example, when verifying a supplier's compliance, the system can cross-verify whether their business registration information is normal, whether they have any legal litigation records, whether there were any violations in their historical bidding processes, and their credit rating. This multi-dimensional data fusion analysis can reveal risk points that cannot be detected by a single dimension, making compliance verification conclusions more comprehensive and reliable.

[0035] This invention integrates three technical paths to form a closed-loop verification mechanism with "rules-semantics-data" triple verification, creating a powerful verification loop.

[0036] User input proceeds sequentially as follows: 1. Rule-driven verification of intelligent agents: logical reasoning based on explicit legal provisions and systems.

[0037] 2. Semantic understanding and complex reasoning of the MOE large model: handling ambiguities, conducting case comparisons, and providing in-depth risk insights.

[0038] 3. Factual verification of underlying data: providing data support for the analytical conclusions of the previous two steps.

[0039] Finally, the results from the three aspects were weighted, fused, and cross-validated to complete the fourth and most comprehensive compliance check. The final conclusions and detailed analysis process were then returned to the user. This reduced the probability of missed and false positives, improving the system's robustness and decision-making quality.

[0040] Through collaborative innovation of the MOE large model, multi-agent and multi-source data platform, a new generation of compliance verification system with professionalism, intelligence, real-time and scalability has been successfully built. It can effectively cope with the increasingly complex business compliance environment and provide strong technical support for enterprises to manage risks in digital transformation.

[0041] Example 2 This invention provides a large-scale procurement model for natural language question answering, case consultation, tender document generation, bid document compliance review, and compliance review report generation.

[0042] The following uses a natural language question-answering scenario as an example to illustrate the operation of this architecture: 1. Input intent recognition. Users ask questions in natural language (e.g., "What qualifications are required for a 'consortium bid'?"). The system performs semantic understanding and intent recognition on the query, determining it to be a factual question.

[0043] 2. Compliance review of "input" When a user submits original materials (such as a draft of a tender document to be reviewed, a tender document, or a specific consultation question), the system will first call the legal and regulatory expert model to conduct a preliminary review of the input content itself while parsing and understanding it.

[0044] The review mainly includes: Content legality review: Compliance of Questions / Requests: For user inquiries, the model will determine whether they involve illegal or irregular activities or an attempt to exploit policy loopholes. For example, if a user asks "how to bypass public bidding," the model will identify the intent to violate regulations, refuse to provide advice, and issue a risk warning.

[0045] Basic compliance of input documents: Conduct a basic and explicit violation scan of the uploaded bidding / tender documents. For example, check whether the bidding documents contain explicitly prohibited restrictive or discriminatory clauses (such as specifying specific brands or patents that are irrelevant to the actual needs of the project); check whether the tender documents have obvious signs of falsified qualifications (such as formatting that is seriously inconsistent with the official version).

[0046] Sensitive Information and Trade Secret Identification: The system reviews input content to determine if it contains state secrets, trade secrets, or personal privacy information. For example, if a tender document inadvertently includes un-anonymized ID numbers, bank account information, etc., the model will identify this and prompt the user to anonymize the information to comply with data security regulations.

[0047] 3. Gated networks route to question-answering experts. Based on the identified "fact-based question-answering" intent, the gated network primarily routes the question to expert models that specialize in information retrieval and extraction, rather than experts who require complex reasoning.

[0048] 4. Intelligent Agent Collaborative Retrieval. The system activates the "Legal and Regulatory Expert" intelligent agent to perform precise searches in policy documents and legal databases. Simultaneously, it may invoke the "Case Query" intelligent agent to search for relevant supporting evidence in judicial and bidding case databases.

[0049] 5. Information Synthesis and Refinement. The activated expert models and agents submit the retrieved fragmented information (such as legal provisions, policy texts, and expert interpretations) to the core large model for synthesis, refinement, and natural language recombination.

[0050] 6. Organize information to form output. The system generates a concise, accurate, and conversational answer.

[0051] 7. Conduct compliance audits on the output.

[0052] Specifically, it includes: Review of the accuracy and sufficiency of the conclusions Verify legal citations: Check whether the laws, regulations, and policy provisions cited in the system-generated answers, reports, or review comments are accurate, currently valid, and highly relevant to the conclusions. Prevent misattribution or citation of outdated regulations.

[0053] Verify logical consistency: Examine whether the reasoning process from "facts" to "conclusion" conforms to legal logic. For example, if the system determines that "a bid is invalid because it failed to provide proof of social security contributions," the regulatory model needs to verify whether the bidding documents explicitly list this as a mandatory clause.

[0054] (2) Fairness and risk review of generated content For the generated tender documents: review the qualification requirements, technical parameters, scoring methods, etc., automatically generated by the system to see if there are any hidden risks of discrimination or exclusion. This is a crucial review of the "creative" output.

[0055] For the provided advice: review the compliance risk level of the advice. For complex issues in a legal gray area, the model will require the addition of a warning statement to the output, such as "This situation is controversial; it is recommended to consult the competent authority or legal counsel," to prevent users from incurring risks due to blindly following advice.

[0056] (3) Linguistic precision and elimination of "illusion" Review the generated text for any imprecise, ambiguous, or AI-generated "illusions" or "fictitious clauses" that do not exist in any regulations. Ensure that every legally significant statement in the output is verifiable.

[0057] (4) Conflict detection with internal knowledge base The system output is compared with the built-in "expert interpretations" and "typical cases" to check for any conclusions that contradict authoritative interpretations or mainstream precedents. If a conflict is found, the system will be triggered to re-evaluate or highlight the conflict.

[0058] In the entire MOE-based agent collaboration architecture, the regulatory expert model plays a dual role: Pre-filter: During the input review process, it ensures that the "source water" flowing into the system is clean.

[0059] Post-inspection: In the output review stage, it ensures that the "products" leaving the system are qualified.

[0060] Through these two closed-loop reviews, the model elevates compliance from an isolated function to a fundamental capability and core design principle that runs throughout the system, thereby truly achieving the ultimate goal of intelligent systems being "both intelligent and reliable" in professional fields.

[0061] Beneficial effects: The existing MOE large model + intelligent agent architecture places greater emphasis on the professionalism of the large model at the gating network level and the compliance verification before returning to the user. It realizes the external service of the large model under unified management and system guidance, and makes up for the three major shortcomings of the current large model technology in the bidding field: The procurement field encompasses a wide range of scenarios, from compliance review of tender documents and intelligent generation of bid documents to intelligent assisted bid evaluation. Current technologies primarily train different large models for different scenarios, resulting in varying levels of "regulatory content" within these models and making it difficult to guarantee the compliance of the outputs from these models across different scenarios. This invention constructs a unified regulatory model for use by intelligent agents and expert models within the MOE (Mechanical Engineering Environment), achieving a unified understanding of regulations from functional processes to compliance content.

[0062] With the continuous introduction of national and local regulations and systems in the current procurement field, it is difficult to quickly update and abolish the existing knowledge base in the large model using current technical solutions, making it difficult to ensure the timeliness of the regulations output by the large model. This invention constructs a knowledge graph based on a unified regulatory large model to achieve the management, updating and abolishment of data such as cases, regulations, and systems, effectively ensuring the professionalism and compliance of the entire large model.

[0063] Before the large model is output to the user, this invention performs a compliance and anti-conflict detection on the large model to conduct a final check and confirmation on the compliance and regulatory compliance of the large model output, filling the gap in the current large model output that does not have compliance detection.

[0064] The above specific embodiments further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An intelligent compliance verification method, characterized in that, The testing method includes: Receive user input for items or questions to be tested; Upon entering the compliance verification layer, the task requirements are initially matched with the built-in knowledge base; The system enters two agent layers and a large model layer in parallel to handle structured rules and unstructured complex problems, respectively. The results of comprehensive analysis are then output to the user.

2. The intelligent compliance verification method according to claim 1, characterized in that, The knowledge base specifically includes policy documents, bidding and tendering laws, government procurement laws, company internal management systems, expert interpretations, and relevant cases.

3. The intelligent compliance verification method according to claim 1, characterized in that, The invocation of the intelligent agent layer specifically includes: constructing a cluster of expert models for laws and regulations, and carrying out professional field coverage and iteration.

4. The intelligent compliance verification method according to claim 3, characterized in that, The construction of a cluster of expert models for laws and regulations, and the coverage and iteration of professional fields, specifically includes: Deploy multiple specialized intelligent agents, including agent 1, agent 2, ..., agent n; The aforementioned expert model of regulations and systems is continuously learned and updated; By processing different compliance subdomains with intelligent agents, we have achieved in-depth vertical mining of knowledge and specialized division of labor.

5. The intelligent compliance verification method according to claim 4, characterized in that, The specific compliance sub-areas include: tax compliance, environmental compliance, and bidding compliance.

6. The intelligent compliance verification method according to claim 1, characterized in that, The invocation of the large model layer specifically includes: the adoption of an advanced hybrid expert MOE large model architecture and distributed computation of complex problems.

7. The intelligent compliance verification method according to claim 6, characterized in that, The aforementioned advanced hybrid expert MOE large model architecture, with distributed computation of complex problems specifically including: Intelligent routing in gated networks: Gated networks first analyze and determine the type and complexity of the problem; Distributed processing of expert model pool: Based on the analysis results, the gating network dynamically selects one or more of the most relevant expert models from multiple expert models to collaboratively process the current task; Results Integration and Output: The selected expert models each contribute their professional analysis, and the results are finally summarized to form a comprehensive judgment.

8. The intelligent compliance verification method according to claim 1, characterized in that, The results of the comprehensive analysis and evaluation output to the user specifically include: Integrate multi-dimensional data sources; A closed-loop verification mechanism is formed, consisting of rule-semantic-data triple verification.

9. The intelligent compliance verification method according to claim 8, characterized in that, The closed-loop verification mechanism of rule formation, semantics, and data triple verification specifically includes: Rule-driven verification of intelligent agents; MOE large-scale model semantic understanding and complex reasoning, handles ambiguity, performs case comparisons and provides in-depth risk insights; Fact verification of underlying data.

Images