Intelligent compliance auditing method and system based on large model self-evolution

Abstract

The invention provides an intelligent compliance auditing method and system based on large model self-evolution, and relates to the technical field of data processing, and the method comprises the steps: constructing a compliance knowledge base; dynamically generating compliance judgment rules; obtaining multi-modal input content, wherein the multi-modal input content comprises material content, live broadcast content and chat content; multi-modal data features in the multi-modal input content are extracted; according to the multi-modal data features and compliance judgment rules, intelligent compliance auditing is carried out through a large vertical model; generating a structured auditing report according to the compliance auditing result; performing manual rechecking on the structured audit report; and an artificial review result is converted into a special training corpus, and the special training corpus is used for carrying out self-evolution optimization on the vertical class large model. According to the method, the model auditing capability is not reduced along with time, but is continuously enhanced along with the usage amount, and the compliance requirements of the financial industry on high sensitivity, high adaptability and continuous stability can be met.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a compliance intelligent auditing method and system based on large-scale model self-evolution. Background Technology

[0002] With the widespread adoption of online business models, marketing materials, live-streamed investor education, and customer communication generate massive amounts of unstructured data. Traditional review methods relying on manual or keyword matching are inefficient, while existing AI review systems suffer from problems such as "black box decision-making," "slow rule updates," and "models that do not evolve," making it difficult to meet the high-frequency, high-quality, and high-compliance regulatory requirements.

[0003] Chinese patent document CN120067278A discloses an intelligent content review system for electronic publications, which uses multimodal fusion coding and knowledge graph embedding technology to determine copyright infringement and generates review strategies through reinforcement learning.

[0004] However, this patent document focuses on copyright protection rather than regulatory compliance in financial compliance scenarios. It also lacks a self-evolutionary model mechanism, making it difficult to adapt to changes in scenarios and regulatory updates in compliance judgment. The quality of the review will gradually deteriorate over time, failing to meet the financial industry's requirements for high sensitivity, high adaptability, and continuous stability in compliance. Summary of the Invention

[0005] In view of the shortcomings of the prior art, the purpose of this invention is to provide a compliance intelligent audit method based on large model self-evolution, which can solve the technical problems of the prior art, such as the lack of a model self-evolution mechanism, the inability of compliance judgment capability to adapt to changes in scenarios and regulatory updates, the gradual deterioration of audit quality over time, and the inability to meet the financial industry's compliance requirements for high sensitivity, high adaptability and continuous stability.

[0006] A first aspect of this invention proposes a compliance intelligent auditing method based on large-scale model self-evolution, comprising: S1: Build a compliance knowledge base; S2: Dynamically generate compliance judgment rules; S3: Obtain multimodal input content, which includes: material-type content, live-stream content, and chat-type content; S4: Extract multimodal data features from the multimodal input content; S5: Based on the multimodal data features and the compliance judgment rules, conduct intelligent compliance review through a vertical category-based large model; S6: Generate a structured audit report based on the compliance audit results; S7: Manually review the structured audit report; S8: Convert the manual review results into a dedicated training corpus, and use the dedicated training corpus to perform self-evolutionary optimization of the vertical category large model.

[0007] A second aspect of this invention proposes a compliance intelligent audit system based on large-scale model self-evolution, comprising: a processor and a memory; The memory stores programs or instructions that can run on the processor, which, when executed by the processor, implement the steps of the compliance intelligent auditing method based on large model self-evolution as described in the first aspect.

[0008] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: (1) In this embodiment of the invention, by using a compliance knowledge base and dynamic rule generation, structured rules can be automatically parsed and formed when regulatory agencies issue new policies, thereby improving the ability to adapt to real-time scenarios in response to regulatory changes.

[0009] (2) In this embodiment of the invention, the manual review results are converted into special training corpus, and the special training corpus is used to self-evolve and optimize the vertical category model. The model's review capability will not decrease over time, but will continue to increase with usage, thus meeting the compliance requirements of the financial industry for high sensitivity, high adaptability and continuous stability. Attached Figure Description

[0010] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0011] Figure 1 This is a flowchart illustrating a compliance intelligent auditing method based on large-scale model self-evolution provided by an embodiment of the present invention.

[0012] Figure 2 This is a schematic diagram of the structure of a compliance intelligent audit system based on large model self-evolution provided by an embodiment of the present invention. Detailed Implementation

[0013] To enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, the technical solutions 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, not all embodiments. It should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0014] The following description, in conjunction with the accompanying drawings, details the compliance intelligent audit method based on large model self-evolution provided by the embodiments of the present invention through specific implementations and application scenarios.

[0015] Reference manual attached Figure 1 The diagram illustrates a flowchart of a compliance intelligent auditing method based on large model self-evolution provided by an embodiment of the present invention.

[0016] This invention provides a compliance intelligent auditing method based on large-scale model self-evolution, which may include the following steps: S1: Build a compliance knowledge base.

[0017] Optionally, the compliance knowledge base stores industry regulations, company internal policies, and a database of historical penalty cases.

[0018] Optionally, industry regulations include regulatory documents issued by the China Securities Regulatory Commission (CSRC), the China Futures Association (CFA), and other relevant bodies.

[0019] Optionally, internal company policies may include marketing guidelines, prohibited phrases, and performance review standards.

[0020] Optionally, the historical penalty case database includes internal and industry violation cases, labeled with the violation type and the handling result.

[0021] Furthermore, all entries are stored in a vector database (such as Milvus) to support semantic retrieval.

[0022] In this embodiment of the invention, a knowledge system integrating laws, regulations, and cases is used to support the automatic parsing of regulatory policies and the generation of rules.

[0023] S2: Dynamically generate compliance judgment rules.

[0024] In one possible implementation, S2 specifically includes sub-steps S201 to S203: S201: Actively crawl policy data from the official websites of regulatory agencies and industry reports.

[0025] S202: Use NLP technology to analyze new policies from policy data and extract key requirements.

[0026] S203: Dynamically generate compliance judgment rules based on new policies and key requirements.

[0027] Furthermore, the compliance judgment rules are output in JSON format for subsequent input into the large model.

[0028] In this embodiment of the invention, by using a compliance knowledge base and dynamic rule generation, structured rules can be automatically parsed and formed when regulatory agencies issue new policies, thereby improving the ability to adapt to real-time scenarios in response to regulatory changes.

[0029] S3: Obtain multimodal input content.

[0030] The multimodal input content includes: content-based materials, live streaming content, and chat content. It covers all scenarios such as content, chat, and live streaming, and supports cross-modal comprehensive judgment.

[0031] S4: Extract multimodal data features from multimodal input content.

[0032] Furthermore, different types of input content are standardized to extract multimodal features that can be used for analysis.

[0033] Optionally, for material-based content, the entire text / image / video input is used: This involves acquiring the material, including complete article images, promotional posters, or video content. Text extraction + OCR recognition: The text content is extracted, and optical character recognition (OCR) technology is used to recognize the text information from the image. Keyframe extraction: For video material, representative frames are selected as keyframes for subsequent analysis to reduce computational load and focus on important content.

[0034] Optionally, for live streaming content, the input is: Live stream / replay: The content of the live stream or recorded video is used as input. Segment ASR to Text Conversion: Automatic speech recognition (ASR) technology is used to convert audio segments in the video into text for easier understanding. Simultaneous Screenshotting: While performing speech-to-text conversion, video frames are captured at regular time intervals to ensure consistency between audio and video. Audio-Visual Consistency Analysis: The converted text content is compared with the corresponding screenshots to check for consistency, ensuring the accuracy and completeness of the information. Generation of Live Stream Segment Feature Sequence: Information from each segment processed through the above steps is summarized to construct a temporal feature sequence for subsequent risk assessment and compliance review. Generation of Material-Level Feature Package: The text and image features obtained in the above processes are integrated to form a comprehensive feature set for the material, used for further analysis and processing.

[0035] Optionally, for chat-related content, the following steps are included: Enterprise WeChat / Outbound Call Record Input: Collect chat logs from Enterprise WeChat or other communication tools, including text, voice messages, and images. Text + Voice + Image Integration: Integrate various types of information from the chat logs to form a complete information flow. Voice-to-Text: For voice messages in the chat logs, use voice-to-text technology to convert them into text format for easier understanding and processing. Contextual Semantic Understanding: Based on the converted text content and the context of the dialogue, perform semantic analysis to understand user intent and sentiment. Generation of Session-Level Feature Stream: Organize all information from the entire chat session and its semantic understanding results into a feature stream, providing it for subsequent risk assessment and compliance checks.

[0036] In this embodiment of the invention, heterogeneous data such as images, text, audio, and video can be uniformly converted into a structured multimodal feature stream, achieving cross-modal and consistent information expression. This not only significantly improves the system's ability to understand different content formats, but also accurately captures risk factors that traditional text review cannot identify, such as misleading combinations of images and text, live broadcast promises, and implicit expressions in chat. This provides a complete, accurate, and associative input basis for subsequent large-scale model compliance judgments, thereby significantly improving the coverage, contextual coherence, and recognition effect of review in complex compliance scenarios.

[0037] S5: Based on multimodal data characteristics and compliance judgment rules, conduct intelligent compliance audits through large vertical models.

[0038] Alternatively, the specific large-scale model for the vertical category can be either Qwen or DeepSeek, which are fine-tuned based on LoRA.

[0039] By using compliance judgment rules as contextual input to the large model, semantic-level compliance understanding can be achieved, avoiding the separation of rules and models.

[0040] In one possible implementation, S5 specifically includes sub-steps S501 and S502: S501: Combine multimodal data features with compliance judgment rules into a single prompt word.

[0041] S502: Based on the prompts, conduct intelligent compliance review using a large vertical model and output structured results.

[0042] Optionally, the structured results include: whether a violation occurred, the type of violation, the risk level, and the segment involved.

[0043] Furthermore, violation types could include, for example, "incomplete performance review" or "inclusion of expected returns." Risk levels could be categorized as high, medium, and low. The relevant segments could include timestamps.

[0044] In this embodiment of the invention, by concatenating multimodal data features with dynamically generated compliance judgment rules into a unified prompt word input to a vertical category model, the model can perform semantic-level reasoning on cross-modal content within the context of the rules. This not only enables the model to understand the actual meaning and applicable conditions of regulatory rules, avoiding the problem of rules and models being disconnected, but also enables the accurate positioning and classification of specific risk segments. This significantly improves the accuracy, interpretability, and enforceability of compliance audit results, providing clear and fine-grained decision-making basis for subsequent handling processes.

[0045] S6: Generate a structured audit report based on the compliance audit results.

[0046] Optionally, the structured audit report includes: highlighted violation content, associated specific legal provisions, matching internal policy entries, and similar historical cases.

[0047] S7: Manually review the structured audit report.

[0048] Furthermore, a visual interface can be provided to support manual review and achieve accurate traceability.

[0049] S8: Convert the results of manual review into a dedicated training corpus, and use the dedicated training corpus to perform self-evolutionary optimization of the large vertical model.

[0050] Specifically, the system automatically cleans the manually reviewed data to create structured standard samples. These samples are then stored in a dedicated corpus and managed according to tags. The automatically cleaned, structured corpus of manually reviewed results requires no additional annotation. Accumulated corpus data triggers automatic fine-tuning, and the system goes live after passing evaluation, enabling the model to self-evolve.

[0051] Furthermore, the self-evolutionary optimization supports A / B testing to ensure online stability.

[0052] In this embodiment of the invention, the results of manual review are converted into dedicated training corpus. The dedicated training corpus is used to self-evolve and optimize the large vertical model. The model's review capability will not decrease over time, but will continue to increase with usage, thus meeting the compliance requirements of the financial industry for high sensitivity, high adaptability and continuous stability.

[0053] In one possible implementation, the compliance intelligent auditing method based on large-scale model self-evolution also includes: S9: Implement tiered management based on the risk level identified in the compliance audit results.

[0054] In one possible implementation, S9 specifically includes sub-steps S901 to S903: S901: When the risk level in the compliance audit result is high risk, immediately intercept the input content and notify the business, risk control and compliance teams.

[0055] S902: When the risk level in the compliance audit results is medium risk, add a watermark to notify the compliance team.

[0056] S903: When the risk level in the compliance audit results is low risk, record and archive the data, and generate reports periodically.

[0057] In this embodiment of the invention, by implementing differentiated control based on the risk level of the review results, the system can adopt graded response strategies such as immediate interception, prompting reminders, or recording and archiving for high, medium, and low risk levels, thereby achieving automation and refinement of the review and handling process. This approach can not only promptly block high-risk content and simultaneously notify relevant responsible departments to prevent the leakage or spread of illegal information, but also effectively prompt and traceable manage medium and low-risk content, taking into account both business continuity and compliance control, and significantly improving the response efficiency, risk control accuracy, and overall operational stability of the compliance management system.

[0058] Reference manual attached Figure 2 The diagram shows a structural schematic of a compliance intelligent audit system based on large model self-evolution provided by an embodiment of the present invention.

[0059] This invention provides a compliance intelligent audit system 20 based on large model self-evolution, including: processor 201 and memory 202; The memory 202 stores programs or instructions that can run on the processor 201. When the program or instructions are executed by the processor 201, they implement the steps of the above-mentioned compliance intelligent audit method based on large model self-evolution and achieve the same technical effect. To avoid repetition, the present invention will not elaborate further.

[0060] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the embodiments of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the protection scope of the present invention.

Claims

1. A method for compliance intelligent auditing based on large model self-evolution, characterized in that, Comprise: S1: build a compliance knowledge base; S2: dynamically generate compliance judgment rules; S3: obtain multi-modal input content, including: material content, live content, and chat content; S4: extract multi-modal data features from the multi-modal input content; S5: according to the multi-modal data features and the compliance judgment rules, through the vertical large model, conduct compliance intelligent audit; S6: generate a structured audit report according to the compliance audit result; S7: manually review the structured audit report; S8: convert the manual review result into special training corpus, and use the special training corpus to optimize the vertical large model.

2. The method of claim 1, wherein, The compliance knowledge base stores industry regulations, company internal systems, and historical penalty case library.

3. The method of claim 1, wherein, The S2 specifically comprises: S201: actively crawl policy data on the official website of the regulatory agency and industry bulletins; S202: use NLP technology to parse new policies from the policy data and extract key requirements; S203: dynamically generate the compliance judgment rules according to the new policies and the key requirements.

4. The method of claim 1, wherein, The S5 specifically comprises: S501: splice the multi-modal data features and the compliance judgment rules into a prompt word; S502: according to the prompt word, through the vertical large model, conduct compliance intelligent audit, and output structured results.

5. The method of claim 4, wherein, The structured results include: whether it violates the rules, the type of violation, the risk level, and the involved segment.

6. The method of claim 1, wherein, The vertical large model is specifically: LoRA fine-tuned Qwen or DeepSeek.

7. The method of claim 1, wherein, The structured audit report includes: violation segment highlighting, associated specific regulation clauses, matched internal system items, and pushed similar historical cases.

8. The method of claim 1, wherein, Also include: S9: according to the risk level in the compliance audit result, conduct hierarchical control.

9. The method of claim 8, wherein, The S9 specifically comprises: S901: when the risk level in the compliance audit result is high risk, immediately intercept the input content, and notify the business, risk control, and compliance three-party teams; S902: when the risk level in the compliance audit result is medium risk, add a watermark prompt and notify the compliance team; S903: when the risk level in the compliance audit result is low risk, record and archive, and generate reports regularly.

10. A compliance intelligent auditing system based on large model self-evolution, characterized in that, Comprise: A processor and a memory; The memory stores programs or instructions that can be run on the processor, which are executed by the processor to realize the steps of the compliance intelligent audit method based on large model self-evolution as claimed in any one of claims 1 to 9.

Images