Enterprise risk comprehensive assessment method and device

Abstract

The invention relates to the technical field of artificial intelligence, and particularly provides an enterprise risk comprehensive assessment method and device, and the method comprises the following steps: S1, obtaining enterprise related risk information from a multi-source heterogeneous data source; s2, performing cleaning, standardization and feature extraction on the collected original data so as to adapt to subsequent model training and analysis requirements; s3, performing automatic risk identification and classification labeling on the preprocessed enterprise multi-dimensional data; s4, carrying out risk conduction analysis from single enterprise risks to an industrial chain, a region and even a whole economic system by constructing an enterprise association map and quantifying a risk infection mechanism; s5, performing multi-dimensional aggregation analysis and visual presentation on the enterprise risk; s6, performing dynamic monitoring and future trend prediction on enterprise risks; and S7, a user interaction module realizes understandability and verifiability of a risk assessment result. Compared with the prior art, the business efficiency can be improved, the financial risk can be reduced, and the healthy development of the financial market is promoted.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and specifically provides a method and device for comprehensive enterprise risk assessment. Background Art

[0002] The essence of enterprise risk assessment is to predict a company's future solvency based on an evaluation of its past operations, finances, and creditworthiness. Enterprise risk assessment methods include traditional risk assessment methods and statistical models. Traditional risk assessment methods include expert judgment methods, such as the 5Cs approach, and comprehensive evaluation methods developed based on the 5Cs approach. Statistical models include linear discriminant models, linear probability models, and logistic regression models. The logistic regression model is the most commonly used machine learning model for enterprise risk rating. Compared to decision tree, ensemble learning, and neural network models, it offers greater stability, higher predictive accuracy, and improved interpretability.

[0003] Yan Anan et al. proposed "A Method and System for Enterprise Risk Assessment," which pre-processes government system data and stores it in a pre-created database. Using a two-dimensional list, they perform correlation analysis on the various government data stored in the database to obtain independent variables that affect enterprise risk. They then create a risk assessment model and train the independent variables that affect enterprise risk. However, the main limitations of this traditional enterprise risk assessment technology are: ① It is difficult to process massive amounts of unstructured text data (such as news, social media, and financial report notes), resulting in insufficient assessment of key dimensions such as public opinion risk and governance risk; ② It lacks dynamic analysis of complex inter-enterprise correlation networks (such as supply chains), making it impossible to quantify risk transmission effects; ③ Traditional model risk assessment results are output in the form of scores, lacking guidance and suggestions for financial institution business personnel; and ④ Manual analysis and evaluation methods based on preset rules suffer from low efficiency and slow update and iteration speeds.

[0004] With the widespread adoption of big data and artificial intelligence technologies, large-scale modeling techniques can deeply mine enterprise information and generate future predictions, enabling more scientific and timely assessments of enterprise risk profiles. In particular, the emergence of generative pre-trained models has made it possible to conduct enterprise risk analysis with both high accuracy and good interpretability. Furthermore, large-scale models can also enable dynamic analysis of complex enterprise networks and quantify the effects of risk transmission.

[0005] Compared with traditional methods, how to provide an enterprise risk assessment system based on a large model, which forms a complete intelligent risk assessment process from raw data to output risk analysis reports, provides financial institutions with complete data collection, quantification, and analysis services, and can comprehensively and efficiently analyze and evaluate enterprise risk information and risk transmission, and intelligently output multi-dimensional and well-interpretable risk analysis reports is an urgent problem that needs to be solved by technical personnel in this field. Summary of the Invention

[0006] The present invention aims to address the deficiencies of the above-mentioned prior art and provides a highly practical comprehensive enterprise risk assessment method.

[0007] A further technical task of the present invention is to provide a device for comprehensive enterprise risk assessment that is rationally designed, safe and applicable.

[0008] The technical solution adopted by the present invention to solve its technical problem is:

[0009] A comprehensive enterprise risk assessment method comprises the following steps:

[0010] S1, the data acquisition module obtains enterprise-related risk information from multiple heterogeneous data sources;

[0011] S2, the data preprocessing module cleans, standardizes and extracts features from the collected raw data to meet the needs of subsequent model training and analysis;

[0012] S3, the risk labeling module automatically identifies and classifies the pre-processed enterprise multi-dimensional data;

[0013] S4. The risk transmission module builds enterprise association maps and quantifies risk transmission mechanisms to analyze the risk transmission from individual enterprise risks to industrial chains, regions, and even the entire economic system.

[0014] S5. Risk statistics module provides multi-dimensional aggregate analysis and visual presentation of enterprise risks;

[0015] S6, scenario recognition and trend prediction module dynamically monitors enterprise risks and predicts future trends;

[0016] S7. The user interaction module makes the risk assessment results understandable and verifiable.

[0017] Furthermore, in step S1, a hierarchical update strategy is adopted, with industrial and commercial information updated in full daily, financial data synchronized quarterly, and news and public opinion data streamed in real time. At the same time, missing values, outliers, and duplicate data are automatically cleaned and alarmed to monitor data quality.

[0018] Furthermore, in step S2, during the data cleaning, a rule engine is used to perform logical verification on the structured data; for the unstructured text data, word segmentation and stop word removal are first performed, and then named entity recognition is performed through the BERT-CRF model.

[0019] Furthermore, in step S3, a hybrid architecture of rule engine + machine learning model is adopted to build an expert knowledge base. For news and public opinion data, a fine-tuned Llama-3 model is used to perform multi-dimensional analysis. First, the news subject is identified, then the risk type is determined, and finally the severity of the risk is evaluated. The risk label is generated using a dynamic weighting algorithm.

[0020] Furthermore, in step S4, based on the graph neural network GNN, first, a dynamic enterprise association graph is constructed, and the graph data is updated in real time;

[0021] Develop an incremental graph learning algorithm to automatically identify three types of transmission paths when new risk events occur: direct transmission, indirect transmission, and industry resonance.

[0022] Furthermore, in step S5, including risk dimension statistics, risk classification and label management,

[0023] The risk dimension statistics include 10 first-level risk dimensions and 38 second-level sub-dimensions, and adopt a four-level risk classification system;

[0024] The four-level risk classification system includes:

[0025] (1) Serious warning: three or more high-risk indicators are triggered simultaneously;

[0026] (2) Warning: Combination of two medium- to high-risk indicators;

[0027] (3) Concern: Continuous deterioration of a single risk indicator;

[0028] (4) Tips: potential risk signals;

[0029] In label management, a dynamic threshold method is adopted, and the thresholds of various industries are adjusted automatically.

[0030] Furthermore, in step S6, a time series anomaly detection model is constructed to process the time series data of the enterprise's multidimensional risk indicators, capture the nonlinear relationship between indicators through the self-attention mechanism, and automatically trigger an early warning when abnormal fluctuations are detected.

[0031] Furthermore, in step S7, an intelligent question-answering system, natural language query and traceability query are included;

[0032] The intelligent question-answering system generates enterprise risk reports with one click based on a large language model in the financial field with 70B parameters;

[0033] The traceability query adopts the RAG retrieval enhanced generation architecture, and the answer is based on the latest data and internal risk control strategies.

[0034] An enterprise risk comprehensive assessment device comprises: at least one memory and at least one processor;

[0035] The at least one memory is configured to store a machine-readable program;

[0036] The at least one processor is configured to call the machine-readable program to execute a method for comprehensive enterprise risk assessment.

[0037] Compared with the prior art, the enterprise risk comprehensive assessment method and device of the present invention have the following outstanding beneficial effects:

[0038] (1) This invention integrates a large language model (LLM), a graph neural network (GNN), and multi-source heterogeneous data to achieve a comprehensive assessment of 10+ risk dimensions, including corporate finance, public opinion, judicial affairs, supply chain, and industry cycles. Experimental data shows that the system can simultaneously monitor 128 risk factors and automatically screen key influencing factors through a dynamic weighting mechanism, thereby improving the comprehensiveness of risk assessment.

[0039] Unstructured data processing capabilities: Traditional methods have difficulty in effectively parsing unstructured data such as financial report notes, news and public opinion, and social media, resulting in the omission of key risk signals. The present invention uses a domain-optimized LLM to improve the accuracy of financial report text analysis tasks and public opinion sentiment analysis, and can capture negative corporate events in real time; Risk transmission quantification accuracy: Traditional statistical models cannot accurately characterize the risk contagion effect between companies. The present invention constructs a dynamic corporate association map based on GNN to quantify three types of transmission paths: direct association (such as guarantee chain), indirect association (such as common suppliers) and industry resonance. In the bank stress test, it successfully warned of multiple "gray rhino" risks hidden in the high-guarantee network; Improved prediction timeliness: The system can use a large model to predict financial deterioration trends and deduce industry systemic risks in advance.

[0040] (2) The present invention realizes fully automated analysis, shortening the evaluation time for a single enterprise to 15 minutes and increasing the speed of report generation. Through time series anomaly detection and event chain reasoning, the system can detect early risk signals that are difficult to capture with traditional methods, thus gaining a critical adjustment window for enterprises.

[0041] (3) This invention introduces a reinforcement learning framework to automatically optimize the weight allocation strategy for each industry every quarter; it innovatively introduces the industry beta coefficient to quantify the differences in risk contagion across industries. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0043] Attachment Figure 1 It is a flow chart of a comprehensive enterprise risk assessment method. DETAILED DESCRIPTION

[0044] In order to enable those skilled in the art to better understand the solutions of the present invention, the present invention will be further described in detail below in conjunction with specific embodiments. Obviously, the embodiments described are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0045] A best embodiment is given below:

[0046] like Figure 1 As shown, a comprehensive enterprise risk assessment method in this embodiment has the following steps:

[0047] S1, the data acquisition module obtains enterprise-related risk information from multiple heterogeneous data sources;

[0048] It supports multiple data access methods such as API docking, Web crawling, and direct database connection. In terms of structured data collection, the system connects to industrial and commercial data platforms such as the National Enterprise Credit Information Publicity System, Tianyancha, and Qichacha, and obtains real-time data such as basic enterprise information, shareholder structure, and business scope, and establishes a unique enterprise identifier (unified social credit code) for data association. For financial data, the system connects to financial databases such as Wind and Tonghuashun, providing core financial indicators such as corporate balance sheets, income statements, and cash flow statements. In terms of unstructured data collection, the system uses NLP-enhanced crawler technology, focusing on capturing corporate public opinion information from channels such as news media, social media, and industry forums. The accuracy of data collection is improved through keyword expansion and sentiment analysis pre-filtering. In addition, the system also collects official data sources such as the Intellectual Property Office, the Judgment Document Network, and the tax system to ensure comprehensive coverage of key risk indicators such as judicial litigation, execution of dishonesty, and tax anomalies.

[0049] To ensure real-time data availability, the system adopts a tiered update strategy: daily updates of industrial and commercial information, quarterly synchronization of financial data, and real-time streaming of news and public opinion data. Furthermore, the system incorporates a data quality monitoring mechanism that automatically cleans and alerts missing values, outliers, and duplicate data, ensuring the accuracy of subsequent analysis.

[0050] S2, the data preprocessing module cleans, standardizes and extracts features from the collected raw data to meet the needs of subsequent model training and analysis;

[0051] A multi-stage processing flow is adopted, including data cleaning, entity recognition, relationship extraction, time series alignment, etc.

[0052] During the data cleaning phase, the system develops differentiated processing strategies for different data types: Structured data (such as financial statements) uses a rule engine for logical validation, for example, checking the balance of "assets = liabilities + owner's equity" and interpolating or marking outliers. Unstructured text data (such as news and financial report notes) is first segmented (using Jieba and a domain dictionary) and stop words are removed. Then, named entity recognition (NER) is performed using the BERT-CRF model to extract key information such as company names, executives, amounts, and time periods.

[0053] During the feature engineering phase, the system constructs a multi-level feature system: Financial features include over 10 indicators such as solvency, profitability, and operational capacity. Public opinion features combine dissemination data (number of reposts and views) to construct a comprehensive public opinion index. Judicial features count the number of cases involving companies, the amount involved, and the verdicts, and calculate a judicial risk score. Finally, the pre-processed data is stored in a feature library, providing a unified data view for risk assessment.

[0054] The S3 risk labeling module automatically identifies and categorizes pre-processed enterprise multi-dimensional data for risk. This module utilizes a hybrid architecture of "rule engine + machine learning model" to accurately identify 10 major risk dimensions, including change risk, operational risk, and legal proceedings, and outputs four levels of risk labels (serious warning, warning, concern, and prompt). Technically, the module first constructed an expert knowledge base containing over 50,000 industry rules. For example, for the risk of "dishonest execution," the rule base includes matching rules for the list of dishonest persons subject to execution published by the Supreme People's Court and rules for the execution target amount threshold. For the risk of "tax information," it integrates more than 20 judgment criteria, including tax ratings and the amount of tax owed.

[0055] This module innovatively introduces a Large Language Model (LLM) to label risks in complex scenarios. Specifically, for news and public opinion data, a fine-tuned Llama-3 model is used to perform multi-dimensional analysis: first, identifying the news subject (company / individual), then determining the risk type (such as product quality issues, executive changes), and finally assessing the risk severity (based on sentiment analysis and event impact prediction).

[0056] Risk tags are generated using a dynamic weighting algorithm, taking into account factors such as rule matching results, model confidence, and data freshness. For example, for a manufacturing company, if both financial risk and supply chain risk are triggered simultaneously, the system will automatically increase its overall risk level and generate a combined risk alert.

[0057] S4. The risk transmission module builds enterprise association maps and quantifies risk transmission mechanisms to analyze the risk transmission from individual enterprise risks to industrial chains, regions, and even the entire economic system.

[0058] This module, based on graph neural networks (GNNs) and complex network theory, includes three core functions: First, it constructs a dynamic enterprise relationship graph. Graph nodes include business entities (registered entities), individuals (legal representatives, actual controllers), and financial institutions. Edge relationships encompass 15 types of relationships, including equity control (shareholding ratio), guarantee relationships (guaranteed amount), and supply chain transactions (transaction frequency and scale). Graph data is updated in real time, supporting millisecond-level queries up to 1,000 queries.

[0059] A heterogeneous graph neural network (HGNN) is used to process multivariate relationships, and a specialized edge-type attention mechanism is designed to automatically learn the importance weights of different association types. An incremental graph learning algorithm has been developed. When a new risk event, such as a legal lawsuit, is added, the risk value of the entire graph can be recalculated in 5 minutes, compared to hours required by traditional methods. The system automatically identifies three types of transmission paths: ① Direct transmission: For example, when a parent company provides a guarantee to a subsidiary, risk is directly transmitted (transmission efficiency >80%); ② Indirect transmission: Secondary transmission through common suppliers (efficiency 30-50%); ③ Industry resonance: Companies in the same industry are affected by systemic risks (quantified by industry beta coefficients). Finally, the module provides a visual transmission analysis interface. It supports risk tracing (tracing the source of risk upward) and impact deduction (simulating the spread of risk downward), and generates reports on key nodes in the transmission chain.

[0060] S5. Risk statistics module provides multi-dimensional aggregate analysis and visual presentation of enterprise risks;

[0061] Supports comprehensive risk insights from macro-industries to micro-enterprises. The module includes two core sub-modules: risk dimension statistics, and risk classification and tag management.

[0062] The system includes 10 primary risk dimensions and 38 secondary sub-dimensions. Each dimension provides: A dynamic risk score (0-100 points): This uses the Analytic Hierarchy Process (AHP) to construct an indicator weighting system. For example, the "Judicial Litigation" dimension includes sub-indicators such as the number of cases (weighted 30%), the amount involved (40%), and the type of case (30%). Industry Relative Percentile: This shows a company's ranking in the manufacturing industry for its "Tax Information Risk" score. Trend Change Rate: This shows the magnitude and acceleration of the change in the score compared to the previous period.

[0063] The system uses a four-level classification system: Severe Warning (simultaneous triggering of three or more high-risk indicators); Warning (a combination of two medium- and high-risk indicators); Concern (continuous deterioration of a single risk indicator); and Warning (a potential risk signal). Label generation utilizes a dynamic threshold method, with thresholds automatically adjusted for each industry. The system automatically evaluates label accuracy monthly and iterates the model to address mislabeled cases.

[0064] S6, scenario recognition and trend prediction module dynamically monitors enterprise risks and predicts future trends;

[0065] By integrating time series analysis, event reasoning, and the causal inference capabilities of large language models, this module enables dynamic monitoring of enterprise risks and forecasts of future trends, supporting risk forecasting needs for time spans ranging from 30 days to three years. The module also builds a time series anomaly detection model.

[0066] The model processes the time series data of the enterprise's multi-dimensional risk indicators (such as monthly financial data and daily public opinion index), captures the nonlinear relationship between indicators through the self-attention mechanism, and automatically triggers an early warning when abnormal fluctuations are detected.

[0067] S7. The user interaction module makes the risk assessment results understandable and verifiable;

[0068] The user interaction module uses natural language interaction and visual analysis technology to make risk assessment results understandable and verifiable. This module includes three innovative designs:

[0069] (1) Intelligent Question Answering System: Based on the 70B parameter large language model in the financial field (RiskGPT), it supports multi-round and multi-modal risk queries and generates enterprise risk reports with one click;

[0070] (2) Natural language query: For example, “show the companies among new energy vehicle companies with increased financial risks but decreased public opinion risks in the past three years.”

[0071] (3) Traceability query: For any risk assessment result, you can ask "Why is the company marked as a serious warning?" The system adopts the RAG (Retrieval Enhanced Generation) architecture to ensure that the answer is based on the latest data and internal risk control policies.

[0072] Based on the above method, an enterprise risk comprehensive assessment device in this embodiment includes: at least one memory and at least one processor;

[0073] The at least one memory is configured to store a machine-readable program;

[0074] The at least one processor is configured to call the machine-readable program to execute a method for comprehensive enterprise risk assessment.

[0075] The above-mentioned specific implementation methods are only specific cases of the present invention. The patent protection scope of the present invention includes but is not limited to the above-mentioned specific implementation methods. Any technical solutions that conform to the above-mentioned specific implementation methods of the present invention and any appropriate changes or substitutions made thereto by ordinary technicians in the relevant technical field shall fall within the patent protection scope of the present invention.

[0076] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A comprehensive enterprise risk assessment method, characterized in that: The steps are as follows: S1, the data acquisition module obtains enterprise-related risk information from multiple heterogeneous data sources; S2, the data preprocessing module cleans, standardizes and extracts features from the collected raw data to meet the needs of subsequent model training and analysis; S3, the risk labeling module automatically identifies and classifies the pre-processed enterprise multi-dimensional data; S4. The risk transmission module builds enterprise association maps and quantifies risk transmission mechanisms to analyze the risk transmission from individual enterprise risks to industrial chains, regions, and even the entire economic system. S5. Risk statistics module provides multi-dimensional aggregate analysis and visual presentation of enterprise risks; S6, scenario recognition and trend prediction module dynamically monitors enterprise risks and predicts future trends; S7. The user interaction module makes the risk assessment results understandable and verifiable.

2. The enterprise risk comprehensive assessment method according to claim 1, characterized in that: In step S1, a hierarchical update strategy is adopted, with industrial and commercial information updated in full daily, financial data synchronized quarterly, and news and public opinion data streamed in real time. At the same time, missing values, outliers, and duplicate data are automatically cleaned and alarmed to monitor data quality.

3. The enterprise risk comprehensive assessment method according to claim 2, characterized in that: In step S2, during the data cleaning, a rule engine is used to perform logical verification on structured data; for unstructured text data, word segmentation and stop word removal are first performed, and then named entity recognition is performed through the BERT-CRF model.

4. The enterprise risk comprehensive assessment method according to claim 3, characterized in that: In step S3, a hybrid architecture of rule engine + machine learning model is used to build an expert knowledge base. For news and public opinion data, a fine-tuned Llama-3 model is used to perform multi-dimensional analysis. First, the news subject is identified, then the risk type is determined, and finally the severity of the risk is evaluated. The risk label is generated using a dynamic weighting algorithm.

5. The enterprise risk comprehensive assessment method according to claim 4, characterized in that: In step S4, based on the graph neural network (GNN), first, a dynamic enterprise association graph is constructed, and the graph data is updated in real time; Develop an incremental graph learning algorithm to automatically identify three types of transmission paths when new risk events occur: direct transmission, indirect transmission, and industry resonance.

6. A comprehensive enterprise risk assessment method according to claim 5, characterized in that: In step S5, including risk dimension statistics, risk classification and label management, The risk dimension statistics include 10 first-level risk dimensions and 38 second-level sub-dimensions, and adopt a four-level risk classification system; The four-level risk classification system includes: (1) Serious warning: three or more high-risk indicators are triggered simultaneously; (2) Warning: Combination of two medium- to high-risk indicators; (3) Concern: Continuous deterioration of a single risk indicator; (4) Tips: potential risk signals; In label management, a dynamic threshold method is adopted, and the thresholds of various industries are adjusted automatically.

7. The enterprise risk comprehensive assessment method according to claim 6, characterized in that: In step S6, a time series anomaly detection model is constructed to process the time series data of the enterprise's multidimensional risk indicators. The nonlinear relationship between indicators is captured through the self-attention mechanism, and an early warning is automatically triggered when abnormal fluctuations are detected.

8. The enterprise risk comprehensive assessment method according to claim 6, characterized in that: In step S7, including intelligent question-answering system, natural language query and traceability query; The intelligent question-answering system generates enterprise risk reports with one click based on a large language model in the financial field with 70B parameters; The traceability query adopts the RAG retrieval enhanced generation architecture, and the answer is based on the latest data and internal risk control strategies.

9. A comprehensive enterprise risk assessment device, characterized in that: include: at least one memory and at least one processor; The at least one memory is configured to store a machine-readable program; The at least one processor is configured to call the machine-readable program to execute the method according to any one of claims 1 to 8.

Images