A risk audit system for asset loss
By analyzing unstructured data through AI big data models and combining multi-source data fusion with the domestic environment, a risk warning and analysis model is constructed. This solves the problem of intelligent monitoring and early warning of the risk of "holding a controlling stake but not having control", improves audit efficiency and accuracy, adapts to dynamic policy iterations, and supports large-scale deployment in the domestic environment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CPI INFORMATION TECH CO LTD
- Filing Date
- 2026-02-03
- Publication Date
- 2026-05-29
AI Technical Summary
Existing technologies for identifying and auditing the risk of "holding but not controlling" state-owned enterprises suffer from several problems, including insufficient ability to analyze unstructured data, difficulty in integrating multi-source heterogeneous data, insufficient staticity and interpretability of risk models, and low adaptability of technological achievements to commercial products. These issues result in low audit efficiency and make it difficult to scale up applications.
The system employs a large AI model to analyze unstructured governance documents, performs data cleaning and standardization through a data governance module, and combines multi-source data fusion to build a risk warning and analysis model. This enables systematic risk monitoring and early warning in a domestically developed environment.
It enables intelligent parsing of unstructured data, efficient integration of multi-source data, and flexible expansion of risk models, thereby improving audit efficiency and accuracy, shortening audit cycles, reducing manual time consumption, adapting to dynamic policy iterations, and supporting large-scale deployment in a domestic environment.
Smart Images
Figure CN122114614A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of audit risk analysis technology, and in particular relates to a risk audit system for asset loss. Background Technology
[0002] In the field of state-owned asset supervision and auditing, with the deepening of mixed ownership reform of state-owned enterprises and the diversification of investment entities, "holding a controlling stake but not controlling the rights" has become one of the key hidden risks leading to the loss of state-owned assets. This risk usually manifests as follows: although state-owned shareholders legally hold a majority of the shares in the invested enterprise, due to special provisions in the company's articles of association (such as veto power, excessively high voting requirements), imbalance of board seats, key management personnel being appointed by non-state-owned shareholders, or ineffective governance mechanisms, their actual control is hollowed out, and they are unable to effectively exercise core shareholder rights such as major decision-making, asset supervision, and profit distribution.
[0003] Currently, the identification and auditing of such risks mainly rely on human experience, which is inefficient and difficult to systematize. Although there has been a great deal of research in the field of audit risk analysis (such as "Research on Audit Risk Assessment Model Based on Machine Learning" and "Audit Risk Identification and Quantitative Analysis in the Big Data Environment"), and progress has been made in general risk analysis techniques, the following significant bottlenecks still exist when applying them to the specific and complex business scenario of "holding but not controlling": 1. Lack of Unstructured Data Parsing Capabilities: The core audit documents, such as the Articles of Association, shareholder and board resolutions, and meeting minutes, are mostly in the form of unstructured text, images, or PDF documents. Current technology lacks efficient and accurate automated methods to extract and understand complex governance clauses (such as voting rights rules, authority agreements, and delegation relationships) from these documents. Auditing work still heavily relies on manual review and subjective judgment, becoming a major obstacle to large-scale auditing.
[0004] 2. Difficulty in Integrating Multi-Source Heterogeneous Data: Accurate assessment of the risk of "holding without control" requires cross-domain integration of internal data (consolidated financial statements, cash flow, contracts, master data) and external data (equity penetration diagrams, corporate registration information, related party networks). Existing audit systems are mostly siloed, with inconsistent data standards, hidden relationships, and a lack of a unified data governance and integration framework, making it difficult to form a comprehensive and dynamic profile of the company's control status.
[0005] 3. Insufficient Staticity and Interpretability of Risk Models: Existing risk analysis models are mostly built on static financial statements and structured financial indicators, failing to dynamically reflect changes in control caused by governance clauses and personnel changes. Furthermore, these models often operate as "black boxes," with their judgment logic difficult to directly link to specific violations. Auditors cannot quickly pinpoint the root cause of problems, reducing audit efficiency and the credibility of the results.
[0006] 4. Low degree of productization and adaptability of technological achievements: Many cutting-edge researches remain at the algorithm level and have failed to be packaged into configurable, scalable, and user-friendly productized systems. In particular, there is a lack of deep adaptation to the domestic information technology innovation environment (such as domestic chips, operating systems, and databases), which restricts the large-scale, secure, and controllable deployment and promotion within the state-owned asset system with strict regulatory requirements.
[0007] Therefore, there is an urgent need for an audit system that can intelligently parse unstructured governance documents, deeply integrate internal and external multi-source data, incorporate professional risk analysis models, and support domestic environments and flexible expansion. This system would enable effective monitoring and early warning of the risk of "holding a controlling stake but not controlling the rights" in an automated, precise, and scalable manner, filling the gap between current technical solutions and the actual needs of state-owned asset supervision. Summary of the Invention
[0008] The purpose of this invention is to overcome the shortcomings of the prior art and to disclose a risk auditing system for asset loss.
[0009] In view of this, the present invention discloses a risk auditing system for asset loss, comprising: The data acquisition module is used to acquire business bylaws documents, internal financial data, internal contract information, internal management data, and equity data. The data governance module is used to clean, extract key information, and standardize all acquired data. The identification and parsing module is used to store standardized key information into a pre-established audit intermediate table; and The audit output module is used to analyze the data in the audit intermediate table based on the pre-built risk warning analysis model group and output the audit results.
[0010] As an improvement to the above system, the data acquisition module acquires data in Word, image, PDF, and Excel formats.
[0011] As an improvement to the aforementioned system, the equity data is obtained through the Tianyancha interface and includes: information on all directly held units of the controlling entity, shareholder information, basic enterprise information, information on key personnel of the enterprise, and a list of companies controlled by individuals with equity or shareholding relationships within the internal organization.
[0012] As an improvement to the above system, the data governance module's processing procedure includes: An AI Agent is used to preprocess unstructured text, including text extraction and word segmentation, and then it is converted into a text format that the model can understand through intelligent compliance screening. Text recognition and extraction based on the PaddleOCR system; Using the Qwen-2.5-72B engine, the extracted text fields are defined as parameters through the Function Call mechanism, and structured JSON data is output.
[0013] As an improvement to the above system, the audit intermediate table includes: The shareholder information table stores information on all directly held entities of the controlling entity, with the data source being the Tianyancha API. The Holding Company Information Table stores information on directly held and jointly held organizations, with data sourced from the Tianyancha API. The Enterprise Basic Information Table is used to store basic enterprise information, and the data source comes from the Tianyancha interface. The executive information table is used to store information about the company's senior executives, and the data source comes from the Tianyancha interface; The personnel control information table includes a list of companies controlled by individuals with equity or shareholding relationships within the internal organization. The data source is Tianyancha API. The charter information table is used to store charter data information identified by the data governance module; The articles of association voting form is used to store data related to the voting ratios of the board of directors and shareholders' meetings in the business articles of association documents identified by the data governance module. The consolidated financial information table is used to store organizational information for all consolidated entities; The personnel master data table is used to store personnel information; The organization master data table is used to store organization information; A fund flow information table is used to store fund flow information; The Treasury Organization Table is used to store information on organizations included in the supervision. The contract information form is used to store information about contract signing; and The meeting information sheet is used to store meeting information records.
[0014] As an improvement to the above system, the risk warning analysis model group includes: The power exercise judgment model is used to determine whether there are situations where the shareholders' meeting or the board of directors is unable to exercise or neglects to exercise its powers by reading the holding company information table, articles of association information table, basic enterprise information table, and consolidated financial information table; it belongs to the governance model and equity model. The accounting voucher query judgment model is used to read the holding company information table, enterprise basic information table, and treasury organization table to determine whether there are situations where major shareholders cannot access the company's and its wholly-owned subsidiaries' accounting books and vouchers, and are unable to supervise the company's operations; it belongs to the governance model. The delegation-related judgment model is used to read the holding company information table, personnel master data table, senior management information table, and shareholder information table to determine whether there are situations where the chairman is appointed by other shareholders or has a specific relationship with other shareholders such as kinship or common interests, or whether the general manager and financial officer are simultaneously appointed by other shareholders or have a specific relationship with other shareholders; it belongs to the personnel category model. The overreach judgment model is used to read the holding company's information table, articles of association information table, and meeting information table to determine whether there are situations where the management fails to implement the resolutions of the shareholders' meeting or the board of directors, or exercises the powers of the shareholders' meeting or the board of directors without authorization; it belongs to the governance model. The seat-to-equity ratio matching judgment model is used to read the personnel master data table, holding company information table, senior management information table, holding company information table, and articles of association voting information table to determine whether there are situations where the shareholding is more than 50% and the total number of director seats appointed by other shareholders exceeds half of the total number of seats on the board of directors; it belongs to the governance model and personnel model; The veto power assessment model is used to read the holding company's information table and articles of association information table to determine whether there are other shareholders with a shareholding ratio of less than one-third who have veto power over multiple key matters or agree that all shareholders must agree, which would substantially damage the controlling shareholder's control; it belongs to the governance model category. The related-party transaction assessment model is used to read information tables such as holding company information, personnel holding information, shareholder information, cash flow information, and contract information to determine whether there is a risk of transferring benefits to other shareholders through related-party transactions; it belongs to the transaction-related model category. The Party building work judgment module is used to read the holding company's information table and articles of association information table to determine whether there are any situations where Party building work is not clearly defined or is neglected; it belongs to the governance model. The model for determining the exercise of rights by overseas companies is used to read information tables of holding companies, consolidated financial statements, and articles of association to determine whether there are situations where major shareholders are unable to exercise corresponding shareholder rights over overseas companies; it belongs to the governance category of models; and Other behavioral judgment models are used to read the holding company information table, articles of association information table, personnel master data table and shareholder information table to determine whether there are other situations that cause serious damage to the rights and interests of major shareholders. They belong to the governance model category.
[0015] As an improvement to the above system, the processing procedure of the audit output module includes: The audit intermediate table is invoked to achieve cross-table data association through pre-defined key fields, and to filter out enterprises with deregistration or revocation status, as well as duplicate and invalid data. Call the risk warning analysis model group, screen according to model type, and output risk records, including original data source, enterprise information, specific violation clauses, and construct a structured encapsulated result that includes at least the group organization code, violation type and original data source for supporting evidence; Output a problem list, a risk statistics report, and a structured encapsulation result; wherein, the problem list is used to list the problematic enterprises and violations according to the model; the risk statistics report is used for risk distribution and verification; and the encapsulation result is in JSON or tabular format.
[0016] As an improvement to the above system, the step of calling the risk warning analysis model group and performing screening based on the model type specifically includes: For equity-based models, the controlling stake is calculated using a look-through algorithm, and issues are identified by comparing consolidated financial information. The governance model checks compliance by verifying the authority reflected in the charter and the number of meetings held as indicated in the forms; The personnel model identifies the risk of non-internal employees in key positions by linking the master data of senior executives and personnel in the form; The transaction model identifies abnormal transactions by analyzing fund flows and contracts.
[0017] As an improvement to the above system, the system also includes a data storage module, which is used to store various types of data acquired by the data acquisition module, to save the intermediate results output by the data governance module and the identification and parsing module, as well as the audit results output by the audit output module, and to store a pre-established audit intermediate table.
[0018] As an improvement to the aforementioned system, the front end of the system is built on Vue2 to create a cockpit and data management interaction page, enabling user operation visualization; the back end is based on the Java technology ecosystem, uses the Spring Boot framework to implement business logic processing, and adopts KingBase, which is deeply adapted to the domestic information technology innovation environment, as the main database.
[0019] Compared with the prior art, the advantages of the present invention are: (i) Automated data acquisition and processing to reduce manual time consumption. Intelligent parsing of unstructured data: Automatically parses documents such as company bylaws using AI big data models to extract key information (such as equity structure and voting rights clauses), replacing the traditional process of manually reviewing documents and improving efficiency by more than 90%.
[0020] Multi-source data integration: Integrating internal and external data sources such as Tianyancha, Treasury System, and Three Major Decisions, achieving a structured data coverage rate of ≥95%, and shortening the audit data preparation time from weeks to hours.
[0021] (ii) Shorten the audit cycle Precisely identify high-risk entities: Through models such as equity penetration analysis and governance mechanism failure detection, high-risk subsidiaries are automatically marked, helping the audit team quickly focus on key objectives and reduce the workload of investigation by 80%.
[0022] (iii) The model can be flexibly expanded to adapt to dynamic policy iterations. Low-code configuration platform: Auditors can extend risk indicators (such as new compliance clauses required by State-owned Assets Supervision and Administration Commission policies), and can quickly configure and update models, improving the efficiency of responding to policy changes by 70%.
[0023] Continuously optimize the early warning mechanism: reduce false alarm rate and reduce invalid verification by dynamically adjusting risk thresholds (such as triggering an early warning when the board attendance rate is below 60%).
[0024] (iv) Expanding the value of large-scale promotion and improving the efficiency of industry auditing Standardized risk database sharing: The risk assessment standards output by the model can be replicated to other central enterprises, avoiding redundant development and realizing the accumulation of cross-enterprise audit experience.
[0025] Domestic adaptation reduces deployment costs: It supports domestically developed environments such as Kylin system and Kingbase database, providing technical compatibility assurance for large-scale promotion. Attached Figure Description
[0026] Figure 1 This is a system architecture diagram of the present invention; Figure 2 This is a flowchart of the business execution process of the present invention; Figure 3 This is the data architecture of the present invention. Detailed Implementation
[0027] System Architecture: like Figure 1 As shown, the front-end is built on Vue2 to create interactive pages such as a dashboard and data management interface, enabling visualized user operations. The back-end relies on the Java technology ecosystem, using frameworks such as Spring Boot to efficiently support business logic processing. KingBase (Renmin University's Kingbase database), which is deeply adapted to the domestic IT innovation environment, is used as the main database. Each layer of technology is tailored to the specific needs of the scenario, collaboratively ensuring the implementation of system data management, risk analysis, and other functions, while balancing security and operational efficiency.
[0028] The specific business execution process is as follows: Figure 2As shown, in the "holding but not controlling" business execution process, the front-end page serves as the entry point, uploading the articles of association documents to the AI application. Through intelligent collaboration between the AI application and the AI model, the articles of association data is accurately analyzed and deeply processed. Simultaneously, Tianyancha injects key data such as equity into the back-end service, enabling efficient interaction between the front-end and back-end and streamlining the business data chain. The back-end service not only drives the model to conduct professional risk analysis but also relies on MiniIO for secure file storage. By retaining all business data through Renmin University's Kingbase, a complete closed loop is built, from data collection and intelligent processing to risk assessment and secure storage. This digital and intelligent approach significantly improves the efficiency and accuracy of the business process, laying a solid foundation for risk prevention and decision support in the "holding but not controlling" business.
[0029] Data source tracing and wide table design: A thorough analysis of 17 model scenarios was conducted, and the final data source systems (such as major decision-making, legal system, financial shared services, treasury system, etc.) were identified. By combining external data (Tianyan Check in one embodiment) with the correlation and integration, 14 audit intermediate tables for the "holding but not controlling" supervision model were designed, as shown in Table 1.
[0030] Table 1
[0031] Table 2 shows the audit intermediate tables used for modeling the 17 scenarios: Table 2
[0032]
[0033] Data architecture: like Figure 3 As shown, this system achieves efficient integration and intelligent analysis of state-owned asset supervision data by constructing a closed-loop system encompassing "data collection → governance → development → modeling → application." Specifically, the data collection layer aggregates multi-source data (financial, management, unstructured documents, etc.) covering all aspects of enterprise operations; the governance layer addresses data quality issues and ensures high availability through AI cleaning and standardization; the development layer breaks down system barriers, unifies indicators and dimensions, and constructs standardized data assets; the detailed broad-surface layer integrates multi-dimensional business data such as equity and capital to support precise analysis; and the model layer relies on AI technology (OCR, intelligent parsing) to build risk warning and audit models, outputting penetrating regulatory results. Ultimately, this forms a closed loop from data collection to business value, empowering the intelligent transformation of state-owned asset supervision. OCR: Optical Character Recognition.
[0034] (1) Theoretical innovation We pioneered a systematic analytical framework for "controlling but not controlling" situations, aiming to drive the transformation of state-owned asset supervision from an experience-driven to a data-driven capital governance model. The core of this framework lies in constructing a unified indicator system and breaking down 10 key risk scenarios into 17 precise analytical models. By integrating multi-dimensional data such as equity structure, board seats, and voting rights, we have achieved algorithmic identification of risks related to control imbalances (such as minority shareholders controlling key positions). Simultaneously, we have integrated more than 10 types of data sources, including financial, contractual, and external platform data, to construct analytical models for complex risks such as related-party transactions and overseas SPV penetration. (2) Technological innovation ① Integration of OCR and traditional text extraction: Enhanced image / table recognition capabilities based on PaddleOCR (Baidu's open-source deep learning platform) to ensure document structural integrity; ② Accurate and structured output for large models: The Qwen-2.5-72B engine is used, and the extracted fields are defined as parameters through the Function Call mechanism to directly output structured JSON data, reducing post-processing costs; ③ Low-code engineering implementation: Leveraging a drag-and-drop development platform to build end-to-end processes, encapsulating the entire chain from OCR recognition to LLM parsing to data integration, supporting zero-code user operation. LLM: Large Language Model.
[0035] (3) Model innovation By integrating data on equity, board seats, and voting rights, algorithms can identify imbalances in control (such as minority shareholders controlling key positions).
[0036] Integrate 10+ types of data sources (financial / contractual / external platforms) to build risk models for related-party transactions, overseas SPV penetration, etc.
[0037] Localized large-scale model + OCR + AI Agent enables intelligent compliance screening of unstructured texts (formalities, meeting minutes).
[0038] (4) Original design across the entire chain From theoretical framework (unified indicator system for central and state-owned enterprises), model construction (17 risk scenarios) to system implementation (full-process closed-loop platform), it has achieved the first penetrating supervision solution in China that adapts to diversified equity structures, and solved the structural contradiction between the traditional model and the State-owned Assets Supervision and Administration Commission's (SASAC) need for "classified authorization and precise control".
[0039] It fills the gap in domestic intelligent early warning for "holding without controlling rights" and provides the State-owned Assets Supervision and Administration Commission with real-time monitoring tools. Case studies show that its risk identification accuracy rate is on average over 85%.
[0040] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and embodiments.
[0041] Example Embodiments of the present invention propose a risk auditing system for product loss, the system comprising: 1. Data Acquisition Module: This module acquires business bylaws, internal financial data, internal contract information, internal management data, and equity data. The data types include Word, images, PDF, and Excel formats. Equity data is obtained through the Tianyancha API, including: information on all directly held units of the controlling entity, shareholder information, basic enterprise information, information on key personnel, and a list of companies controlled by individuals with equity or shareholding relationships within the internal organization. It should be noted that while the Tianyancha API is used in this embodiment, it is not a limitation; other third-party data sources capable of acquiring equity data are also acceptable.
[0042] 2. Data governance module, used to clean, extract key information and standardize all acquired data; specifically, it includes: using an AI Agent to preprocess unstructured text including text extraction and word segmentation, and converting it into a text format that the model can understand through intelligent compliance screening; Text recognition and extraction based on the PaddleOCR system; Using the Qwen-2.5-72B engine, the extracted text fields are defined as parameters through the Function Call mechanism, and structured JSON data is output.
[0043] 3. The identification and parsing module is used to store the standardized key information into a pre-established audit intermediate table; 4. Audit Output Module: This module analyzes the data in the audit intermediate tables based on a pre-built risk warning analysis model group and outputs the audit results. Specifically, it includes: (1) Call the audit intermediate table, realize cross-table data association through the pre-defined key fields, and filter out enterprises with cancellation or revocation status and duplicate invalid data; (2) Call the risk warning analysis model group, screen according to the model type, and output the risk records, including the original data source, enterprise information, specific violation clauses, and construct a structured encapsulated result that includes at least the group organization code, violation type and original data source used for evidence; The risk warning analysis model group, based on Table 2, includes: The power exercise judgment model is used to determine whether there are situations where the shareholders' meeting or the board of directors is unable to exercise or neglects to exercise its powers by reading the holding company information table, articles of association information table, basic enterprise information table, and consolidated financial information table; it belongs to the governance model and equity model. The accounting voucher query judgment model is used to read the holding company information table, enterprise basic information table, and treasury organization table to determine whether there are situations where major shareholders cannot access the company's and its wholly-owned subsidiaries' accounting books and vouchers, and are unable to supervise the company's operations; it belongs to the governance model. The delegation-related judgment model is used to read the holding company information table, personnel master data table, senior management information table, and shareholder information table to determine whether there are situations where the chairman is appointed by other shareholders or has a specific relationship with other shareholders such as kinship or common interests, or whether the general manager and financial officer are simultaneously appointed by other shareholders or have a specific relationship with other shareholders; it belongs to the personnel category model. The overreach judgment model is used to read the holding company's information table, articles of association information table, and meeting information table to determine whether there are situations where the management fails to implement the resolutions of the shareholders' meeting or the board of directors, or exercises the powers of the shareholders' meeting or the board of directors without authorization; it belongs to the governance model. The seat-to-equity ratio matching judgment model is used to read the personnel master data table, holding company information table, senior management information table, holding company information table, and articles of association voting information table to determine whether there are situations where the shareholding is more than 50% and the total number of director seats appointed by other shareholders exceeds half of the total number of seats on the board of directors; it belongs to the governance model and personnel model; The veto power assessment model is used to read the holding company's information table and articles of association information table to determine whether there are other shareholders with a shareholding ratio of less than one-third who have veto power over multiple key matters or agree that all shareholders must agree, which would substantially damage the controlling shareholder's control; it belongs to the governance model category. The related-party transaction assessment model is used to read information tables such as holding company information, personnel holding information, shareholder information, cash flow information, and contract information to determine whether there is a risk of transferring benefits to other shareholders through related-party transactions; it belongs to the transaction-related model category. The Party building work judgment module is used to read the holding company's information table and articles of association information table to determine whether there are any situations where Party building work is not clearly defined or is neglected; it belongs to the governance model. The model for judging the exercise of power over overseas enterprises is used to read the information table of the holding company, the consolidated financial statements, and the articles of association to determine whether there are situations where the major shareholder cannot exercise the corresponding shareholder rights over the overseas enterprise; it belongs to the governance model. Other behavioral judgment models are used to read the holding company information table, articles of association information table, personnel master data table and shareholder information table to determine whether there are other situations that cause serious damage to the rights and interests of major shareholders. They belong to the governance model category.
[0044] Screening is performed based on model type, including: For equity-based models, the controlling stake is calculated using a look-through algorithm, and issues are identified by comparing consolidated financial information. The governance model checks compliance by verifying the authority reflected in the charter and the number of meetings held as indicated in the forms; The personnel model identifies the risk of non-internal employees in key positions by linking the master data of senior executives and personnel in the form; The transaction model identifies abnormal transactions by analyzing fund flows and contracts.
[0045] (3) Output a problem list, risk statistics report and structured encapsulation results; wherein, the problem list is used to list the problem enterprises and violations according to the model; the risk statistics report is used for risk distribution and verification; the encapsulation results are in JSON or table format.
[0046] 5. Data storage module, used to store various types of data acquired by the data acquisition module, to save the intermediate results output by the data governance module and the identification and parsing module, as well as the audit results output by the audit output module, and also to store pre-established audit intermediate tables.
[0047] The front end of this system is built on Vue2 to create a dashboard and data management interaction page, enabling user operation visualization; the back end is based on the Java technology ecosystem, using the Spring Boot framework to handle business logic, and adopts KingBase, which is deeply adapted to the domestic information technology innovation environment, as the main database.
[0048] It is worth noting that in the embodiments of the above system, the modules included are divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional module are only for easy differentiation and are not used to limit the scope of protection of the present invention.
[0049] Manual verification and model validation: This verification process collected 2,539 articles of association documents from the group, and successfully identified 2,072 valid documents using the AI intelligent model. We then conducted a sample of manual verification of the model's identification results; the specific results are shown in Table 3. Table 3
[0050] Model result validation: Based on internal test data, the model results were manually verified as shown in Table 4: Table 4
[0051] Model scalability: This system creates a modeling environment with "configurable extension + elastic architecture", achieving scalability at three levels: logic, dimensions, and tools, and meeting the agile iteration needs of complex regulatory scenarios.
[0052] First, the logic layer uses a rule engine to support dynamic modeling, eliminating the need for large-scale code refactoring.
[0053] Second, the dimension layer is compatible with new data (such as supply chain finance) through a hierarchical data architecture, and can use graph databases (Neo4j) to deepen stakeholder analysis; Third, an AI model management platform is built at the tool layer to support OCR / NLP model replacement and graph computing tool integration, enabling unstructured data parsing and equity correlation mining.
[0054] Method indicators: This project has been validated, and the core indicators have been largely achieved, as detailed in Table 5: Table 5
[0055] It should be noted that the risk identification result of this invention is a "precautionary warning" rather than a "conclusion".
[0056] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to the embodiments, those skilled in the art should understand that modifications or equivalent substitutions to the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A risk auditing system for asset loss, characterized in that, include: The data acquisition module is used to acquire business bylaws documents, internal financial data, internal contract information, internal management data, and equity data. The data governance module is used to clean, extract key information, and standardize all acquired data. The identification and parsing module is used to store standardized key information into a pre-established audit intermediate table; and The audit output module is used to analyze the data in the audit intermediate table based on the pre-built risk warning analysis model group and output the audit results.
2. The asset loss risk audit system according to claim 1, characterized in that, The data acquisition module acquires data in Word, image, PDF, and Excel formats.
3. The asset loss risk audit system according to claim 1, characterized in that, The equity data is obtained through the Tianyancha interface and includes: information on all directly held units of the controlling entity, shareholder information, basic enterprise information, information on key personnel of the enterprise, and a list of companies controlled by individuals with equity or shareholding relationships within the internal organization.
4. The asset loss risk audit system according to claim 1, characterized in that, The data governance module's processing steps include: An AI Agent is used to preprocess unstructured text, including text extraction and word segmentation, and then it is converted into a text format that the model can understand through intelligent compliance screening. Text recognition and extraction based on the PaddleOCR system; Using the Qwen-2.5-72B engine, the extracted text fields are defined as parameters through the Function Call mechanism, and structured JSON data is output.
5. The asset loss risk audit system according to claim 1, characterized in that, The audit intermediate table includes: The shareholder information table stores information on all directly held entities of the controlling entity, with the data source being the Tianyancha API. The Holding Company Information Table stores information on directly held and jointly held organizations, with data sourced from the Tianyancha API. The Enterprise Basic Information Table is used to store basic enterprise information, and the data source comes from the Tianyancha interface. The executive information table is used to store information about the company's senior executives, and the data source comes from the Tianyancha interface; The personnel control information table includes a list of companies controlled by individuals with equity or shareholding relationships within the internal organization. The data source is Tianyancha API. The charter information table is used to store charter data information identified by the data governance module; The articles of association voting form is used to store data related to the voting ratios of the board of directors and shareholders' meetings in the business articles of association documents identified by the data governance module. The consolidated financial information table is used to store organizational information for all consolidated entities; The personnel master data table is used to store personnel information; The organization master data table is used to store organization information; A fund flow information table is used to store fund flow information; The Treasury Organization Table is used to store information on organizations included in the supervision. The contract information form is used to store information about contract signing; and The meeting information sheet is used to store meeting information records.
6. The asset loss risk audit system according to claim 5, characterized in that, The risk warning analysis model group includes: The power exercise judgment model is used to determine whether there are situations where the shareholders' meeting or the board of directors is unable to exercise or neglects to exercise its powers by reading the holding company information table, articles of association information table, basic enterprise information table, and consolidated financial information table; it belongs to the governance model and equity model. The accounting voucher query judgment model is used to read the holding company information table, enterprise basic information table, and treasury organization table to determine whether there are situations where major shareholders cannot access the company's and its wholly-owned subsidiaries' accounting books and vouchers, and are unable to supervise the company's operations; it belongs to the governance model. The delegation-related judgment model is used to read the holding company information table, personnel master data table, senior management information table, and shareholder information table to determine whether there are situations where the chairman is appointed by other shareholders or has a specific relationship with other shareholders such as kinship or common interests, or whether the general manager and financial officer are simultaneously appointed by other shareholders or have a specific relationship with other shareholders; it belongs to the personnel category model. The overreach judgment model is used to read the holding company's information table, articles of association information table, and meeting information table to determine whether there are situations where the management fails to implement the resolutions of the shareholders' meeting or the board of directors, or exercises the powers of the shareholders' meeting or the board of directors without authorization; it belongs to the governance model. The seat-to-equity ratio matching judgment model is used to read the personnel master data table, holding company information table, senior management information table, holding company information table, and articles of association voting information table to determine whether there are situations where the shareholding is more than 50% and the total number of director seats appointed by other shareholders exceeds half of the total number of seats on the board of directors; it belongs to the governance model and personnel model; The veto power assessment model is used to read the holding company's information table and articles of association information table to determine whether there are other shareholders with a shareholding ratio of less than one-third who have veto power over multiple key matters or agree that all shareholders must agree, which would substantially damage the controlling shareholder's control; it belongs to the governance model category. The related-party transaction assessment model is used to read information tables such as holding company information, personnel holding information, shareholder information, cash flow information, and contract information to determine whether there is a risk of transferring benefits to other shareholders through related-party transactions; it belongs to the transaction-related model category. The Party building work judgment module is used to read the holding company's information table and articles of association information table to determine whether there are any situations where Party building work is not clearly defined or is neglected; it belongs to the governance model. The model for determining the exercise of rights by overseas companies is used to read information tables of holding companies, consolidated financial statements, and articles of association to determine whether there are situations where major shareholders are unable to exercise corresponding shareholder rights over overseas companies; it belongs to the governance category of models; and Other behavioral judgment models are used to read the holding company information table, articles of association information table, personnel master data table and shareholder information table to determine whether there are other situations that cause serious damage to the rights and interests of major shareholders. They belong to the governance model category.
7. The asset loss risk audit system according to claim 6, characterized in that, The processing procedure of the audit output module includes: The audit intermediate table is invoked to achieve cross-table data association through pre-defined key fields, and to filter out enterprises with deregistration or revocation status, as well as duplicate and invalid data. Call the risk warning analysis model group, screen according to model type, and output risk records, including original data source, enterprise information, specific violation clauses, and construct a structured encapsulated result that includes at least the group organization code, violation type and original data source for supporting evidence; Output a problem list, a risk statistics report, and a structured encapsulation result; wherein, the problem list is used to list the problematic enterprises and violations according to the model; the risk statistics report is used for risk distribution and verification; and the encapsulation result is in JSON or tabular format.
8. The asset loss risk audit system according to claim 7, characterized in that, The process of calling the risk warning analysis model group and screening based on model type specifically includes: For equity-based models, the controlling stake is calculated using a look-through algorithm, and issues are identified by comparing consolidated financial information. The governance model checks compliance by verifying the authority reflected in the charter and the number of meetings held as indicated in the forms; The personnel model identifies the risk of non-internal employees in key positions by linking the master data of senior executives and personnel in the form; The transaction model identifies abnormal transactions by analyzing fund flows and contracts.
9. The asset loss risk audit system according to claim 1, characterized in that, The system also includes a data storage module, which stores various types of data acquired by the data acquisition module, saves intermediate results output by the data governance module and the identification and parsing module as well as audit results output by the audit output module, and stores pre-established audit intermediate tables.
10. The asset loss risk audit system according to claim 1, characterized in that, The system's front end is built on Vue2 to create a cockpit and data management interaction page, enabling user operation visualization; the back end is based on the Java technology ecosystem, using the Spring Boot framework to handle business logic, and adopts KingBase, which is deeply adapted to the domestic information technology innovation environment, as the main database.