Enterprise financial risk analysis method, device and computer readable storage medium

Abstract

The application discloses an enterprise financial risk analysis method and device and a computer readable storage medium. The method comprises the following steps: obtaining an inquiry letter, a financial statement and a financial statement note; generating an inquiry letter risk analysis report and a note risk analysis report by using a large language model to analyze the inquiry letter and the financial statement through prompt words for the inquiry letter and the note, wherein the two reports both contain triplet data associated with an enterprise and financial risk description text; storing the triplet data in a graph database to form a knowledge graph, and analyzing the risk transmission relationship between the target enterprise and other enterprises based on the knowledge graph; performing vectorization on the financial risk description text and storing the vectorized data in a vector database, and performing similarity retrieval on risk characteristics based on the vector database to find other enterprises with a risk characteristic similar to that of the target enterprise and meeting a predetermined threshold. The above technical solution can realize more accurate, comprehensive and fast risk analysis and mining.

Description

Technical Field

[0001] This invention relates to the field of data risk analysis technology, and more specifically, to a method, apparatus, and computer-readable storage medium for analyzing corporate financial risks. Background Technology

[0002] With economic development, analyzing corporate financial risks is essential. However, some existing methods are not satisfactory in all aspects.

[0003] For example, patent application number CN202410620058.X, entitled "A Method, Device, and Medium for Verifying the Authenticity of Enterprise Financial Data," relies on indicator verification rules and a pre-set financial data model for its effectiveness. Such models, derived from expert rules, are highly subjective and have weak generalization capabilities, making them unable to identify sophisticated financial fraud. Furthermore, maintaining the indicator verification rules and financial data model requires manual operation, resulting in high costs.

[0004] For example, patent application number CN202410718888.6, entitled "A Method and Apparatus for Financial Health Analysis," only incorporates publicly available historical financial statements. Using unaudited quarterly or semi-annual reports results in low reliability; while using audited annual reports offers greater reliability, it lacks timeliness. Furthermore, the financial risk analysis lacks complex models and a comprehensive analysis of risk transmission, ultimately leading to room for improvement in the accuracy and comprehensiveness of the analysis results.

[0005] For example, patent application number CN202410675835.0, entitled "A Method and System for Financial Data Risk Analysis Based on Large Model Agents," only uses clustering and fitting estimation of current and historical financial data, without clustering or searching other companies, resulting in insufficient comprehensiveness of risk analysis. Furthermore, the setting of clustering thresholds involves subjective factors and lacks objectivity.

[0006] The three patent application schemes mentioned above share the following common drawbacks:

[0007] (1) The data from regulatory inquiry letters with strong timeliness and the complex notes in the financial statements were not used, and the risk information contained in these two types of data was not effectively mined.

[0008] (2) The model is not complex enough, and the ability to analyze risks depends on subjective experience.

[0009] (3) The analysis results of the model did not involve a more comprehensive risk mining process, nor did it include a risk transmission process. Summary of the Invention

[0010] In view of the above-mentioned problems in related technologies, the present invention proposes a method, device and computer-readable storage medium for enterprise financial risk analysis, which can achieve more accurate, comprehensive and fast risk analysis and mining.

[0011] According to one aspect of the present invention, a method for analyzing corporate financial risks is provided, comprising the following steps: obtaining an inquiry letter issued to a company, and obtaining financial statements and notes to the financial statements issued by the company; generating an inquiry letter risk analysis report by analyzing the inquiry letter and the financial statements using a large language model based on prompts for the inquiry letter, and generating a notes risk analysis report by analyzing the notes to the financial statements using the large language model based on prompts for the notes to the financial statements, wherein both the inquiry letter risk analysis report and the notes risk analysis report contain: triple data associated with the company, and a description text of the company's financial risks; storing the triple data in a graph database to form a knowledge graph, and analyzing the risk propagation relationship between the target company and other companies based on the knowledge graph; vectorizing the description text of financial risks and storing it in a vector database, and performing a similarity retrieval of risk features based on the vector database to find other companies whose risk features are similar to those of the target company to a predetermined threshold.

[0012] In some embodiments, the enterprise financial risk analysis method further includes: performing deduplication processing on the triplet data of the inquiry letter risk analysis report and the triplet data of the note risk analysis report, wherein the triplet data of the note risk analysis report is retained when data conflicts exist; and performing union processing on the financial risk description text of the inquiry letter risk analysis report and the financial risk description text of the note risk analysis report.

[0013] In some embodiments, analyzing the risk propagation relationship between the target enterprise and other enterprises based on the knowledge graph includes at least one of the following: performing N-degree association queries using a graph database query language to trace multi-level risk transmission paths between enterprises, where N is an integer between 2 and 5; constructing a risk propagation tree based on a breadth-first search algorithm to identify the chain risks of upstream and downstream related enterprises; using a community partitioning algorithm combined with density clustering to identify abnormal subgraph structures to identify abnormal enterprise association circles; and iteratively calculating the importance weights of enterprise nodes using the PageRank algorithm to identify core hub enterprises in the risk network.

[0014] In some embodiments, the enterprise financial risk analysis method further includes: vectorizing the financial risk description text after taking the union using embedding; and performing the similarity search using the Faiss search engine, wherein different search strategies are selected according to the application scenario, the application scenario including at least one of enterprise benchmarking, risk warning and document tracing.

[0015] In some embodiments, the inquiry letter risk analysis report and the note risk analysis report also include financial indicator data, which includes the name and value of the financial indicator. The enterprise financial risk analysis method further includes: storing the financial indicator data in a relational database; and applying preset judgment rules to score the enterprise's risk based on the financial indicator data in the relational database.

[0016] In some embodiments, the enterprise financial risk analysis method further includes: performing deduplication processing on the financial indicator data of the inquiry letter risk analysis report and the financial indicator data of the note risk analysis report, wherein when data conflicts exist, the financial indicator data of the note risk analysis report is retained; and storing the deduplicated financial indicator data into the relational database.

[0017] In some embodiments, the preset judgment rules include at least one of the following: abnormal accounts receivable turnover, excessive proportion of related-party transactions, insufficient inventory write-downs, cash flow disruption warning, and goodwill impairment risk. Each judgment rule defines the threshold conditions of financial indicators, the corresponding risk level, risk label, and verification suggestions. When the financial indicator data meets the threshold conditions of the corresponding rule, the corresponding rule is triggered and the risk level, risk label, and verification suggestions are output.

[0018] In some embodiments, the enterprise financial risk analysis method further includes: using a PDF text extraction tool to automatically identify and extract text information from PDF-formatted inquiry letters and financial statement notes for use in generating inquiry letter risk analysis reports and note risk analysis reports.

[0019] According to another aspect of the present invention, a corporate financial risk analysis apparatus is provided, comprising: a document acquisition module, configured to acquire an inquiry letter issued to a company, and to acquire financial statements and notes to financial statements issued by the company; a large language model analysis module, configured to generate an inquiry letter risk analysis report by analyzing the inquiry letter and the financial statements using a large language model based on prompts for the inquiry letter, and to generate a notes risk analysis report by analyzing the notes to financial statements using the large language model based on prompts for the notes to financial statements, wherein both the inquiry letter risk analysis report and the notes risk analysis report include: triple data associated with the company, and a description text of the company's financial risks; a risk propagation analysis module, configured to store the triple data in a graph database to form a knowledge graph, and to analyze the risk propagation relationship between the target company and other companies based on the knowledge graph; and a similarity retrieval module, configured to vectorize the description text of financial risks and store it in a vector database, and to perform a similarity retrieval of risk features based on the vector database to find other companies whose risk features are similar to those of the target company to a predetermined threshold.

[0020] According to another aspect of the present invention, a computer-readable storage medium is provided, wherein at least one instruction is stored therein, the at least one instruction being executed by a processor in a computer device to implement the above-described enterprise financial risk analysis method.

[0021] The technical solution of this invention utilizes timely inquiry letters and financial statement notes, designs reasonable prompts for these documents, and employs a large language model for risk analysis and data extraction. This data forms the basis for subsequent analysis of risk propagation relationships and similarity retrieval, avoiding reliance on subjective experience. From a real-time perspective, this invention extracts data and performs risk analysis on daily inquiry letters issued by regulatory agencies, improving the timeliness of risk warnings. Using a large language model to analyze inquiry letters and financial statement notes effectively utilizes important information within these documents and leverages the excellent text parsing capabilities of the large language model. Furthermore, this invention analyzes risk propagation relationships using a knowledge graph to realize the risk transmission process and obtains more companies with similar financial characteristics through similarity retrieval, achieving more comprehensive risk mining and thus enabling more accurate, comprehensive, and rapid risk analysis and mining. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart of a corporate financial risk analysis method according to an embodiment of the present invention.

[0024] Figure 2 A flowchart of a corporate financial risk analysis method according to another embodiment of the present invention is shown.

[0025] Figure 3 A flowchart of a corporate financial risk analysis method according to another specific embodiment of the present invention is shown.

[0026] Figure 4 It shows Figure 3 Risk analysis pipeline for three services.

[0027] Figure 5 This is a block diagram of a corporate financial risk analysis device implemented according to this application. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0029] According to an embodiment of the present invention, a method for analyzing corporate financial risks is provided. Figure 1 This is a flowchart of a corporate financial risk analysis method 100 according to an embodiment of the present invention. It should be understood that although the flowchart is shown and described below as a series of actions or steps, the order of these actions or steps should not be construed as restrictive. For example, some actions may occur in a different order and / or simultaneously with other actions or steps not shown and / or described herein.

[0030] refer to Figure 1 As shown, the enterprise financial risk analysis method 100 includes steps S102-S108. In step S102, the inquiry letter issued to the enterprise is obtained, along with the enterprise's financial statements and notes to the financial statements. The enterprise can refer to a listed company. The financial statements include the enterprise's annual reports, semi-annual reports, quarterly reports, and monthly reports.

[0031] (1) Regarding the inquiry letter

[0032] Inquiry letters are typically issued by regulatory agencies (such as the Shanghai Stock Exchange and the Shenzhen Stock Exchange). Inquiry letter data is usually published on the official websites of the Shanghai Stock Exchange and the Shenzhen Stock Exchange, and the letters are usually PDF files. You can download the PDF files of the inquiry letters directly from these official websites, or you can obtain them through procurement.

[0033] Stock exchanges typically issue inquiry letters promptly after a company's relevant events (such as the release of its annual report) to alert investors to potential risks. These letters raise questions about discrepancies or issues requiring further clarification in the listed company's information disclosure, demanding specific responses from the company. Such inquiries usually occur when regulatory agencies review a company's financial statements and discover potentially inaccurate, incomplete, or clarifying information. By sending an inquiry letter, regulators request additional explanations, clarifications, or evidence from the company to confirm the truthfulness and accuracy of its financial statements.

[0034] From a risk control perspective, the types of inquiry letters that are more helpful for risk analysis are the periodic report information disclosure regulatory inquiry letters from the Shanghai Stock Exchange and the annual / semi-annual / quarterly report inquiry letters from the Shenzhen Stock Exchange. Their main content is that the stock exchange conducts a comprehensive review of the information disclosed in the financial statements of listed companies, raises questions about financial data, business models, related-party transactions, major events, etc., and requires the company to supplement disclosures or explanations.

[0035] In recent years, financial fraud and scandals involving listed companies have become increasingly frequent. According to statistics, taking May, a month with a high frequency of annual report releases by listed companies, as an example, the number of annual report inquiry functions issued by the Shenzhen Stock Exchange's main board over the past five years is shown in Table 1 below. The data comes from inquiry letters on the Shenzhen Stock Exchange's official website:

[0036] Table 1. Number of annual report inquiry functions on the Shenzhen Stock Exchange Main Board in May of the past 5 years

[0037] time May 2020 May 2021 May 2022 May 2023 May 2024 Quantity / month 112 151 175 142 111 Quantity / Working Days 5.9 7.9 8.8 6.8 5.3

[0038] As can be seen from the data in Table 1, during the peak season for annual reports in May each year, the Shenzhen Stock Exchange has issued an average of 5-9 annual report inquiry letters per working day over the past 5 years, and the vast majority of these are issued within one month of the companies releasing their annual reports, demonstrating the extremely high timeliness of the annual report inquiry letters.

[0039] Furthermore, as of September 13, 2024, the number of companies listed on the Shenzhen Stock Exchange Main Board was 1,488, data sourced from the Shenzhen Stock Exchange's official website. Annual report inquiry letters received by companies listed on the Shenzhen Main Board account for approximately 10% of all listed companies. Given that financial fraud is a low-probability event, the scope of these inquiry letters is extremely broad.

[0040] For bank lending departments, credit defaults can lead to direct economic losses (unrecoverable loan principal, interest losses, legal costs, etc.), reputational damage (damaged market image, customer trust crisis, etc.), and impacts on business strategies (increased risk management pressure, adjustments to credit policies, etc.). Inquiry letters, as a type of data that is highly timely, professional, and targeted, have extremely important reference value for bank lending departments in analyzing corporate financial risks. However, due to the complexity and highly specialized nature of inquiry letters, they have not been efficiently utilized in the field of corporate financial risk analysis.

[0041] (2) Notes to the financial statements

[0042] Financial statements and notes to financial statements of listed companies are typically published on information disclosure websites designated by the China Securities Regulatory Commission (such as CNINFO), the official websites of the Shanghai and Shenzhen Stock Exchanges, and the official websites of the listed companies. Financial statements and notes are usually PDF files. Similar to inquiry letters, you can directly download the PDF files of the financial statements and notes from the above websites, or you can purchase the relevant data.

[0043] Notes to the financial statements provide supplementary explanations and detailed descriptions of the data and information in the financial statements. They typically include a variety of content that is of great help in risk analysis, such as basic information about the company (place of registration, form of organization, nature of business, principal activities, parent company, etc.), the basis of preparation of the statements, a statement of accounting standards, significant accounting policies and accounting estimates, explanations of significant items, and information on capital structure and cost of capital.

[0044] From a risk analysis perspective, the impact of the notes to the financial statements includes: (A) Significant accounting policies and estimates directly affect the data in the financial statements. Understanding the accounting policies and estimates adopted by a company helps assess the reliability and potential risks of its financial statements. For example, different depreciation methods and impairment testing standards can have a significant impact on a company's profits and asset value; (B) Accounting policies and estimates are more likely to indicate changes in the company's operating environment or adjustments in management's strategies. These changes may have a significant impact on the company's financial condition and operating results; (C) Detailed disclosure of the composition and changes of significant items helps risk analysts understand the company's core assets, liabilities, and revenue structure, thereby assessing its potential risks. For example, an increase in accounts receivable and inventory may indicate that the company's sales are growing but the collection speed is slowing down or there is a risk of inventory backlog; (D) The contents of significant items such as related party transactions, commitments, and contingent liabilities may reveal specific risks or uncertainties faced by the company. For example, large related party transactions may indicate the risk of the company relying on a single customer or supplier.

[0045] It is evident that the notes to the financial statements published by companies contain a wealth of information that is of great help in risk analysis. A thorough reading of the notes to the financial statements can enable bank credit departments to gain a more comprehensive understanding of the company's financial condition and operating performance, and make informed decisions.

[0046] Then, in step S104, a Large Language Model (LLM) can be used to analyze the inquiry letter and the notes to the financial statements, respectively. A Large Language Model is a deep learning model used for natural language processing and natural language generation tasks. It is trained on large amounts of text data to learn the complexity and connections of language, thereby generating natural and fluent text, answering questions, translating languages, and more. Compared to traditional language models, Large Language Models have a higher parameter capacity and deeper neural networks, enabling them to handle more complex language structures and contexts.

[0047] In some embodiments, large language models can employ DeepSeek-R1, Qwen2.5-32B, GPT (Generative Pre-Training) models, etc. The GPT model is a pre-trained language model for natural language processing, proposed by OpenAI in 2018. The GPT model uses a unidirectional Transformer architecture, learning the syntax, semantics, and contextual information of natural language through unsupervised pre-training on large-scale text data. The architecture retains only the decoder part of the Transformer architecture, and the training steps are divided into two stages: 1) Unsupervised pre-training: using the decoder of a multi-layer Transformer model as the language model, performing unsupervised learning based on a large text corpus; 2) Supervised fine-tuning: when the downstream question has structured features, such as ordered sentence pairs or document, question, and answer triples, it is first converted into a specific sequence structure, and then fine-tuned based on a targeted dataset.

[0048] Specifically, in step S104, a risk analysis report for the inquiry letter can be generated by analyzing the inquiry letter and financial statements using a large language model based on the prompts in the inquiry letter. The risk analysis report includes: triple data associated with the enterprise, and a text describing the enterprise's financial risks. In some embodiments, the triples include at least one of "entity-relationship-entity" (e.g., ICBC-shareholder-Central Huijin) and "entity-relationship-attribute" (e.g., Kangmei Pharmaceutical-industry-pharmaceutical manufacturing).

[0049] Furthermore, in step S104, a risk analysis report for the notes to the financial statements is generated by analyzing the notes using a large language model based on the prompts in the notes. Similar to the risk analysis report for the inquiry letter, the risk analysis report for the notes also includes: triplet data associated with the company, and a text describing the company's financial risks.

[0050] In step S104, the prompt words may include multiple items from the following: Role setting, Background description, Profile, Goals, Skills, Goals, Constraints, Workflow, Examples, OutputFormat, and Initialization.

[0051] Due to the complexity and sheer volume of inquiry letters and financial statement notes—with approximately 5,000 companies publishing annual reports each year, along with semi-annual, quarterly, and monthly reports—and the sheer volume of financial statement notes (over 10,000 words) including various tables, traditional corporate financial analysis methods are ineffective in extracting deep, valuable information from these notes due to their high cost and poor cost-effectiveness. This invention utilizes a large language model to analyze financial statement notes, fully leveraging its ability to analyze long texts and extracting key risk management-related information from a large volume of financial statement notes.

[0052] In step S106, the triplet data from the inquiry letter risk analysis report and the supplementary risk analysis report are stored in a graph database to form a knowledge graph. The risk propagation relationship between the target company and other companies is then analyzed based on the knowledge graph. The target company's inquiry letter risk analysis report and supplementary risk analysis report can be obtained through steps S102 and S104, and the risk propagation relationship between the target company and other companies is analyzed based on the triplet data and the knowledge graph from these two reports.

[0053] Knowledge graphs are a structured and semantic representation of knowledge. They organize entities, relationships, and attributes into a vast knowledge network in a graphical form. Knowledge graphs use, for example, "entity-relationship-entity" triples as basic units to organize knowledge into a structured form. They can also imbue these entities and relationships with rich semantic information through attributes and other metadata. Finally, they are visualized through graph databases, presenting the knowledge in an intuitive way for easy user understanding and analysis.

[0054] A knowledge graph mainly consists of three parts: entities, relations, and attributes. 1) Entities are the most basic component of a knowledge graph. They can be concrete objects, abstract concepts, events, people, places, organizations, etc. Each entity has a unique identifier (ID) for unique identification and indexing within the knowledge graph. 2) Relationships are the interactions or connections between entities. They can be associations, dependencies, subordinate relationships, or other types of relationships between two entities. Each relationship also has a unique identifier (ID) for unique identification and indexing within the knowledge graph. 3) Attributes are the characteristics or descriptions of entities and relationships. They can include the entity's name, definition, type, classification, label, etc., and can also include the direction, weight, strength, type, etc., of the relationship. Each attribute also has a unique identifier (ID) for unique identification and indexing within the knowledge graph.

[0055] Step S108 involves vectorizing the financial risk description text in the inquiry letter risk analysis report and the appendix risk analysis report and storing it in a vector database. A similarity search of risk characteristics is then performed based on the vector database to find other companies whose risk characteristics have a similarity to the target company that meets a predetermined threshold. In some embodiments, a similarity greater than the predetermined threshold indicates a higher similarity; therefore, this step involves finding other companies with a similarity greater than the predetermined threshold. The inquiry letter risk analysis report and the appendix risk analysis report of the target company can be obtained through steps S102 and S104, and a similarity search is performed based on the financial risk description text in the two reports and the vector database.

[0056] The technical solution of this invention utilizes timely inquiry letters and financial statement notes, designs reasonable prompts for these documents, and employs a large language model for risk analysis and data extraction. This data forms the basis for subsequent analysis of risk propagation relationships and similarity retrieval, avoiding reliance on subjective experience. From a real-time perspective, this invention extracts data and performs risk analysis on daily inquiry letters issued by regulatory agencies, not just monthly, quarterly, semi-annual, and annual reports. This enables daily-level real-time risk monitoring based on external data, improving the timeliness of risk warnings. Furthermore, regulatory inquiries to companies indicate potential financial risks. Combining the reasons for the inquiries with the financial risks analyzed by the large language model enhances risk assessment capabilities. Using a large language model to analyze inquiry letters and financial statement notes effectively utilizes important information within these documents and leverages the excellent text parsing capabilities of the large language model. Furthermore, this invention uses knowledge graphs to analyze risk propagation relationships and realizes the risk transmission process. It also uses similarity retrieval to obtain more companies with similar financial characteristics, thus achieving more comprehensive risk mining and enabling more accurate, comprehensive, and faster risk analysis and mining.

[0057] In other embodiments, risk scoring can also be performed on companies based on inquiry letters and notes to financial statements, and the scoring may include scores and ratings. Figure 2 A flowchart of a corporate financial risk analysis method 200 according to another embodiment of the present invention is shown. See also Figure 2 As shown, the enterprise financial risk analysis method 200 includes steps S202-210.

[0058] In step S202, similar to step S102, the inquiry letter issued to the company is obtained, as well as the financial statements and notes to the financial statements issued by the company.

[0059] In step S204, a risk analysis report for the inquiry letter can be generated by analyzing the inquiry letter and financial statements using a large language model based on the prompts in the inquiry letter. Similarly, a risk analysis report for the notes to the financial statements can be generated by analyzing the notes to the financial statements using a large language model based on the prompts in the notes to the financial statements.

[0060] In this embodiment, the inquiry letter risk analysis report and the appendix risk analysis report, in addition to containing triplet data and financial risk description text, also include financial indicator data, which includes the name and value of the financial indicator. Financial indicators can be, for example, the current ratio or quick ratio, and can be set according to actual needs.

[0061] At step S206, an operation similar to that in step S106 is performed. At step S208, an operation similar to that in step S108 is performed.

[0062] In this embodiment, the enterprise financial risk analysis method 200 further includes step S210, which involves storing financial indicator data into a relational database; and applying preset judgment rules to score the enterprise's risk based on the financial indicators in the relational database. In some embodiments, this model capable of applying preset judgment rules to score the enterprise's risk is referred to as an expert-experience-based financial risk model.

[0063] The preset judgment rules can include at least one of the following: abnormal accounts receivable turnover, excessive related-party transactions, insufficient inventory write-downs, cash flow disruption warnings, and goodwill impairment risks. Each judgment rule defines a threshold condition for a financial indicator, a corresponding risk level, a risk label, and verification suggestions. When the financial indicator data meets the threshold condition of the corresponding rule, the rule is triggered, and the risk level, risk label, and verification suggestions are output.

[0064] In this embodiment, the comprehensiveness of risk analysis and mining is further enhanced by using knowledge graphs, expert-based financial risk models, and similarity retrieval to empower existing enterprise risk analysis methods.

[0065] Figure 3 A flowchart of a corporate financial risk analysis method according to another specific embodiment of the present invention is shown. The following will refer to... Figure 3 To describe a specific embodiment of the present invention. See also Figure 3 As shown, in step S202, the inquiry letter issued by the stock exchange and the financial statements and notes to the financial statements issued by the listed company are obtained. After obtaining them, since the inquiry letter, financial statements and notes to the financial statements are usually in the form of PDF files, a PDF text extraction tool can be used to automatically identify and extract the text information in the PDF format inquiry letter and notes to the financial statements, so as to generate an inquiry letter risk analysis report and a note risk analysis report.

[0066] In some embodiments, it may be necessary to process a large number of PDF files in batches and extract text from them. Therefore, programming languages ​​(such as Python) and corresponding libraries (such as PyPDF2, PDFMiner, etc.) can be used to write scripts to automate the processing. These libraries provide rich APIs for processing PDF files, including extracting text, tables, and other content. It should be understood that those skilled in the art can implement batch text extraction from PDF documents based on any applicable existing technology, and this application does not limit this.

[0067] Then, in step S204, a prompt word can be designed first. Literally, it is a context given to the model along with the input. It tells and guides the model what task it should do next; it is a prompt. Or, to put it another way, as mentioned earlier, it can transform the downstream task into the form expected by the pre-trained model.

[0068] Prompt is an auxiliary technique for large language models, used to guide the model in generating responses. A prompt is typically a short text string that provides context and task-related information to help the model better understand the request and generate the correct output. For example, in question-answering tasks, a prompt can provide a relevant question to help the language model better understand it. Prompt techniques can effectively improve the performance of large language models, especially for tasks such as natural language generation and question answering. In recent years, prompt techniques have been widely used in the field of natural language processing and have become an important auxiliary technique for large language models.

[0069] In one embodiment, the prompts in the annual report inquiry letter are analyzed as follows.

[0070] - Role: Financial analyst and business risk assessment expert.

[0071] - Background: Users need to conduct in-depth analysis of the annual report inquiry letters of listed companies on the Shanghai Stock Exchange and Shenzhen Stock Exchange to extract key information, including company name, company structure, business operations, authenticity of financial indicators, changes in financial data, related-party transactions, industry and financial risks, etc., to help users analyze corporate risks.

[0072] - Profile: You are an experienced financial analyst and business risk assessment expert, skilled at extracting key information from complex financial documents and assessing business risks.

[0073] - Skills: You possess strong financial accounting knowledge, business risk assessment skills, and data analysis capabilities, enabling you to accurately interpret and analyze key information in annual report inquiry letters.

[0074] - Goals: Extract key data such as corporate entity information, reasons for inquiry, and financial risks from annual report inquiry letters to support users in analyzing corporate risks.

[0075] - Constraints: The extracted information must be accurate and error-free, and the analysis results must be objective and impartial, without containing any subjective assumptions or false information.

[0076] - OutputFormat: A structured inquiry letter risk analysis report, containing key information such as the enterprise entity, the reason for the inquiry, and financial risks, and providing risk assessment conclusions. The inquiry letter risk analysis report provides "entity-relationship-entity" or "entity-relationship-attribute" triples and financial indicator values.

[0077] - Workflow:

[0078] 1. Read and analyze the annual report inquiry letter to identify the main corporate information.

[0079] 2. Based on the content of the inquiry letter, summarize the reasons for the inquiry.

[0080] 3. Analyze financial statements to identify potential financial risks.

[0081] 4. Conduct a business risk assessment based on the reasons for the inquiry and the financial risks.

[0082] 5. Prepare a structured inquiry letter risk analysis report, presenting the analysis results and risk assessment conclusions. The inquiry letter risk analysis report should provide the "entity-relationship-entity" or "entity-relationship-attribute" triples and financial indicator values.

[0083] Examples:

[0084] - Example 1: The business entity "XX Technology Co., Ltd." was questioned about "the timing of revenue recognition". The financial risks included "excessive accounts receivable" and "tight cash flow".

[0085] Example 2: The main entity of the enterprise, "YY Manufacturing Co., Ltd.", was questioned for "insufficient provision for inventory depreciation". The financial risks included "decline in inventory turnover" and "risk of inventory backlog".

[0086] - Example 3: The business entity "ZZ E-commerce Co., Ltd." was questioned about "goodwill impairment testing issues". The financial risks included "high proportion of goodwill" and "increased impairment risk".

[0087] - Initialization: In the first conversation, please directly output the following: Hello, I am your financial analyst and corporate risk assessment expert. I will help you analyze the annual report inquiry letters of companies listed on the Shanghai Stock Exchange and Shenzhen Stock Exchange, extract key information, and assess corporate risks. Please provide the annual report inquiry letter documents you need to analyze.

[0088] In one embodiment, the prompts used for analyzing the notes to the financial statements are as follows.

[0089] - Role: Financial statement analyst and risk assessment expert.

[0090] - Background: Users need to conduct in-depth analysis of the notes to the financial statements of various financial statements (such as annual reports, semi-annual reports, quarterly reports, and monthly reports) issued by listed companies on the Shanghai Stock Exchange and Shenzhen Stock Exchange in order to obtain key information such as the company's basic situation, changes in accounting policies, changes in accounting estimates, financial indicator risks, explanations of important items, and capital structure, and then assess the company's financial risks.

[0091] - Profile: You are a senior financial statement analyst and risk assessment expert with the professional ability to interpret and analyze the notes to financial statements. You are able to extract key information from complex data and assess the financial health and potential risks of a company.

[0092] - Skills: You possess solid financial accounting knowledge, financial analysis skills, risk assessment capabilities, and a deep understanding of accounting standards. You are able to accurately interpret the information in the notes to the financial statements and conduct comprehensive analysis.

[0093] - Goals: Extract key information from the notes to the financial statements, analyze the company's financial condition, identify potential financial risks, and provide professional conclusions.

[0094] - Constraints: Analysis must be based on objective data in the notes to the financial statements, and conclusions must be supported by sufficient data to maintain objectivity and professionalism.

[0095] - OutputFormat: A detailed risk analysis report with notes, including basic information about the company, changes in accounting policies, changes in accounting estimates, risk assessment of financial indicators, explanations of important projects and capital structure analysis, as well as the final financial risk assessment conclusion. The risk analysis report with notes provides triples of "entity-relationship-entity" or "entity-relationship-attribute" and financial indicator values.

[0096] - Workflow:

[0097] 1. Review the notes to the financial statements and extract basic information about the company.

[0098] 2. Analyze changes in accounting policies and accounting estimates.

[0099] 3. Assess financial indicators and identify potential financial risks.

[0100] 4. Provide detailed descriptions of important items, including key items from the balance sheet, income statement, and cash flow statement.

[0101] 5. Analyze the capital structure and assess the efficiency and risk of the company's capital operations.

[0102] 6. Based on the above analysis, provide a conclusion on the financial risk assessment of the enterprise, and provide the triplet of "entity-relationship-entity" or "entity-relationship-attribute" and the financial indicator values.

[0103] Examples:

[0104] Example 1: Analysis of basic enterprise information, including enterprise size, industry position, main business, etc.

[0105] Example 2: Analysis of changes in accounting policies, such as the impact of changes in revenue recognition policies on financial statements.

[0106] Example 3: Analysis of changes in accounting estimates, such as the impact of adjustments to the provision for bad debts ratio on profits.

[0107] Example 4: Risk assessment of financial indicators, such as analysis of indicators like current ratio and quick ratio.

[0108] Example 5: Explanation of important items, such as the impact of an increase in accounts receivable on cash flow.

[0109] Example 6: Capital structure analysis, such as the analysis of debt ratio and equity structure.

[0110] - Initialization: In the first conversation, please directly output the following: Hello, I am your financial statement analyst and risk assessment expert. I will help you conduct an in-depth analysis of the financial statement notes of listed companies, extract key information, and assess financial risks. Please provide the financial statement notes you need to analyze.

[0111] By using the above prompts and uploading the inquiry letter and financial statement notes, you can extract key information related to risk management.

[0112] In some embodiments, after the large language model generates an inquiry letter risk analysis report and a note risk analysis report based on the inquiry letter and financial statement notes, it can be transformed, including data merging and cleaning (duplicate removal), to provide a data foundation for subsequent knowledge graph queries, financial risk models, and similarity retrieval.

[0113] At this point, deduplication can be performed on the triplet data from the inquiry letter risk analysis report and the note risk analysis report. The same <subject, relation, object> criteria can be used to extract the triples from both reports, deduplicate them, and then proceed to step S206. When data conflicts exist, the triplet data from the note risk analysis report is prioritized and retained because its data source is audited. The extracted triples can be used as data sources for subsequent knowledge graph queries and graph algorithms.

[0114] The union of the financial risk description texts in the two reports can also be performed to provide a data basis for the similarity search in the subsequent step S208.

[0115] In addition, financial indicator data (including the names and values ​​of financial indicators) can be extracted from the two reports, deduplicated, and stored in a relational database. When data conflicts exist, the financial indicator data from the risk analysis report in the notes is prioritized and retained, as its data source is audited. The extracted financial indicator data can be used as the data source for the financial risk model in subsequent step S210.

[0116] After data merging and cleaning, the graph query and graph algorithm in step S206, the similarity retrieval in step S208, and the financial risk model scoring in step S210 are deployed.

[0117] In step S206, the deduplicated triple data is stored in a graph database to form a knowledge graph. In some embodiments, the applicable graph database is Neo4j Community Edition, which is open-source and highly mature. NetworkX is used as the graph algorithm engine, which is open-source and commercially available.

[0118] Neo4j is an open-source graph database management system that uses a graph structure composed of nodes, relations, and attributes to store data. Its core features include the following aspects: (1) Native graph storage: Data is organized in the form of "entity-relationship-entity" triples, which supports efficient storage and querying of complex relationships; (2) Cypher query language: A declarative language designed specifically for graph data, which supports multi-hop queries (such as N-degree association analysis), path finding, etc.; (3) Real-time analysis capability: Data associations are quickly mined through graph traversal algorithms, which is suitable for scenarios such as risk transmission analysis and social network modeling; (4) In this patent, Neo4j is used to build enterprise association graphs, supporting multi-degree queries (such as guarantee chain risk mining) and graph algorithm analysis (such as community discovery).

[0119] NetworkX is an open-source graph algorithm library written in Python, focusing on complex network analysis. Its core functions include the following: (1) Algorithm support: providing classic graph algorithms (such as shortest path, PageRank, community detection, minimum spanning tree, etc.); (2) Dynamic modeling: supporting dynamic graph construction and visualization, and flexibly handling node / edge attribute updates; (3) Strong extensibility: compatible with scientific computing libraries such as NumPy and SciPy, suitable for large-scale network simulation.

[0120] In this patent, NetworkX serves as a graph algorithm engine, working in conjunction with Neo4j to analyze risk propagation paths (such as simulating risk contagion in a guarantee network). Therefore, based on Neo4j and NetworkX, multi-degree query and graph algorithm services based on knowledge graphs can be built.

[0121] It should be understood that multi-degree query refers to the technique of data retrieval in a knowledge graph through multi-level relationship paths (i.e., "hop count" or "degree") between nodes. It can uncover indirect connections between entities, revealing implicit information in complex relationship networks, and is one of the core capabilities in risk analysis scenarios. Each entity (such as a company, person, or event) is connected by edges (relationships), forming a "node-edge-attribute" network structure. Multi-degree query searches for nodes that are indirectly related to the target entity by traversing multiple levels of relationships (such as A→B→C→D). A first-degree query is a direct connection (A→B), a second-degree query is a connection through an intermediate node (A→B→C), and an N-degree query is a connection through N-1 intermediate nodes (A→B→C→…→N).

[0122] Specifically, as shown in Table 2 below, based on Neo4j and NetworkX, a knowledge graph-based multi-degree query and graph algorithm service is built to realize multi-level enterprise risk propagation analysis. The query and algorithm service includes at least one of the following: (1) Using graph database query language to perform N-degree association queries to track multi-level risk transmission paths between enterprises, where N is an integer between 2 and 5; (2) Constructing a risk propagation tree based on breadth-first search algorithm to identify the chain risks of upstream and downstream related enterprises; (3) Using community partitioning algorithm combined with density clustering to identify abnormal subgraph structures to identify abnormal enterprise association circles; (4) Iteratively calculating the importance weight of enterprise nodes through PageRank algorithm to identify core hub enterprises in the risk network, where the iteration is 10 times and the damping coefficient is 0.85.

[0123] Table 2 Knowledge Graph Query and Algorithm Services

[0124] Functional modules Technical Implementation Application scenario examples Multi-hop relationship mining Implement N-degree joins (N=2~5) using Neo4j query statements. Tracing the risk transmission path from listed company A → related party B → guarantor C → debt defaulting company Risk Contagion Analysis Constructing a risk propagation tree based on the breadth-first search (BFS) algorithm After a real estate company went bankrupt, identify the chain of risks involving its upstream and downstream suppliers, partners, and related shareholder companies. Anomaly Subgraph Detection The Louvain algorithm is used for community partitioning, combined with density clustering to identify anomalous subgraphs. A hidden guarantee network consisting of eight subsidiaries of a certain group was discovered, with the total amount of guarantees exceeding three times the group's net assets. PageRank Assess the importance weight of enterprise nodes (10 iterations, damping coefficient 0.85) Identifying the core hub enterprises within a guarantee network, whose risks could trigger a systemic crisis.

[0125] In step S208, embedding is used to vectorize the financial risk description text after the union is taken, and the vector is stored in a vector database for similarity retrieval. In some embodiments, the vectorization technique that can be used is the m3e-base model, which performs well with Chinese characters and has low deployment costs.

[0126] Vectorized retrieval using embedding is a widely used technique in Natural Language Processing (NLP). It converts text data into vector representations, enabling efficient retrieval and analysis within a vector space. Embedding refers to the process of transforming text (such as words, sentences, numbers, etc.) into one or more numerical vectors. These vectors typically contain multiple dimensions, each representing an abstract feature or attribute of the text data. Methods for obtaining embeddings include bag-of-words models (such as TF-IDF) and deep learning-based methods (such as Word2Vec, GloVe, BERT, etc.). Among these, deep learning methods are currently the most popular and effective. By training neural network models, text data is mapped to a high-dimensional vector space, making the texts corresponding to vectors that are close in distance in the vector space more semantically similar.

[0127] Using embeddings for retrieval typically involves the following steps: (1) Building an index: To improve retrieval efficiency, it is usually necessary to build an index on the embeddings in the text library. There are various methods for building an index, including tree-based indexes (such as KD-trees, ball trees), hash-based indexes (such as LSH), and vector-based approximate nearest neighbor search algorithms (such as Faiss, Annoy, Elasticsearch, etc.). (2) Similarity calculation: When the text to be queried is input, it is first converted into an embedding, and then the index is searched for the embedding that is most similar to the vector. There are various methods for calculating similarity, among which cosine similarity is a commonly used method. It measures the similarity between two vectors by calculating the cosine value of the angle between them. In multidimensional space, the more similar two vectors are, the smaller their angle is, and the closer the cosine value is to 1. To make the results more obvious, the inverse cosine is calculated for cosine similarity, that is, the angle between the test vector and the feature vector is obtained. The smaller the angle, the more similar they are. (3) Result sorting and display: Based on the similarity calculation results, the retrieved texts are sorted, and the top-ranked texts are displayed to the user. These texts are semantically most similar to the user's query text, and therefore are most likely to meet the user's needs. Vectorized retrieval of text embedding is an efficient, accurate, and scalable text retrieval method with broad application prospects in the search field.

[0128] In some embodiments, the similarity retrieval service used is Faiss, an open-source and commercially available retrieval engine with mature technology. Faiss is a high-efficiency similarity retrieval library open-sourced by Facebook, designed specifically for high-dimensional vector search. Its core advantages are as follows: (1) Fast index building: Supports multiple index structures (such as IVF-PQ, HNSW), accelerating the retrieval of large-scale vector data; (2) GPU acceleration: Utilizes CUDA to achieve parallel computing, significantly improving the nearest neighbor search efficiency of a vector library with hundreds of millions of entries; (3) Flexible integration: Compatible with mainstream deep learning frameworks (such as PyTorch, TensorFlow), suitable for text and image retrieval scenarios. In this patent, Faiss is used to store text vectors of financial statement notes, achieving millisecond-level matching of similar risk cases.

[0129] In some embodiments, different search strategies are selected based on the application scenario, which includes at least one of enterprise benchmarking, risk warning, and document tracing. In some embodiments, a search service is built based on Faiss to find enterprises with similar risk performance, thereby increasing the coverage of risk analysis. Examples of search strategies and applications are shown in Table 3. Vector similarity matching technology is used to search for other enterprises similar to the target enterprise in terms of risk characteristics, and corresponding thresholds are set to control the accuracy of the search results.

[0130] Table 3 Search Strategies Based on Enterprise Risk

[0131] Scene Search methods Similarity threshold Typical applications Enterprise benchmarking FlatL2 brute-force search 0.75 Find comparable companies with similar revenue size Risk warning IVF-PQ Quantitative Search 0.82 Matching characteristics of historical financial fraud cases Document source tracing HNSW Dynamic Index 0.88 The appendix to the inquiry letter regarding the location of the problem

[0132] In step S210, the deduplicated financial indicator data is stored in a relational database to perform risk scoring on the enterprise according to preset judgment rules. In some embodiments, the relational database that can be used is a MySQL database, which is open source, commercially available, and technologically mature and reliable.

[0133] In some embodiments, a financial risk model based on expert experience is used to score the enterprise's risk, ensuring the accuracy of the risk analysis. Examples of the application of such a financial risk model are shown in Table 4.

[0134] Table 4 Financial Risk Model

[0135] Rule Name Applicable Scenarios Judgment Logic Output Application Examples Accounts Receivable Turnover Abnormal Rules Credit risk assessment If the current accounts receivable turnover days are greater than 1.3 times the industry average and the revenue growth rate is less than 10%, then it is considered abnormal. Risk Level: Medium-High Risk; Tag: Decreased Collection Ability; Recommendation: Verify Customer Creditworthiness. A tech company's annual report showed that its accounts receivable turnover days increased from 45 days to 68 days, while revenue growth was only 3%, triggering this rule. Related-party transaction ratio exceeding the limit rules Related Party Transaction Risk Identification If (related party sales / total revenue) is greater than 0.5 and the gross profit margin is greater than the industry average plus 0.15, then there is a risk of inflating profits by relying on related party transactions. Risk Level: High Risk; Tag: Questionable Revenue Authenticity; Recommendation: Review Transaction Pricing Fairness. A manufacturing company's related-party transactions accounted for 62% of its total transactions, and its gross profit margin exceeded the industry average by 20%, triggering this rule. Insufficient Provision for Inventory Depreciation Inventory risk monitoring If the year-on-year growth rate of inventory is greater than 0.3 and the ratio of impairment provision to total inventory is less than 0.05, there may be a risk of inventory impairment. Risk Level: Medium Risk; Tag: Asset Overstatement; Recommendation: Analyze Net Realizable Value of Inventory A retail company's inventory increased by 45% year-on-year, but its provision for inventory obsolescence was only 2%, triggering this rule. Cash flow disruption early warning rules Liquidity Risk Analysis If net operating cash flow is less than 0 for two consecutive years and the year-on-year growth rate of short-term borrowings is greater than 0.2%, then the company will face debt repayment pressure. Risk Level: High Risk; Tag: Liquidity Crisis; Recommendation: Assess the stability of financing channels. A real estate company has had negative operating cash flow for two consecutive years, and its short-term borrowings have increased by 35%, triggering this rule. Goodwill Impairment Risk Rules Post-merger risk monitoring If the goodwill / net assets ratio is greater than 0.3 and the actual net profit of the acquired party is less than 80% of the committed value, there is a risk of significant impairment. Risk Level: Medium Risk; Tag: Deterioration in Asset Quality; Recommendation: Calculate the impact of potential impairment on profits. A listed company's goodwill accounts for 40% of its assets, and the acquired target only achieved 50% of its performance commitment, triggering this rule.

[0136] To delve deeper and broader into the risks of target and related companies, a collaborative working mechanism was constructed using graph queries and algorithms, expert-based financial risk model scoring, and similarity retrieval services, forming a risk analysis pipeline. (See [link to relevant documentation]). Figure 4 .

[0137] In some embodiments, after steps S206-S210, the risk propagation relationships between enterprises, risk scores, and enterprises with similar risk characteristics to the target enterprise can be fused into a large language model to output a comprehensive evaluation result. The comprehensive evaluation result may include information such as enterprise risk warnings, risk transmission (risk propagation relationships), risk measurement (scores), and risk labels.

[0138] In some embodiments, if no other companies with risk propagation relationships with the target company are found when analyzing risk propagation relationships based on knowledge graphs, text representing this result can also be output in the comprehensive evaluation results. In some embodiments, if no other companies with risk characteristics similar to the target company are found when performing similarity searches, text representing this result can also be output in the comprehensive evaluation results.

[0139] In some embodiments, in addition to inquiry letters and notes to financial statements, public opinion information about the company can be crawled in real time. In some such embodiments, large language models can be used to analyze public opinion information and generate corresponding public opinion risk analysis reports. These reports may include triple data associated with the company, as well as financial risk description text, and may also include financial indicator data for risk propagation relationship analysis, similarity retrieval, and risk scoring. By supplementing public opinion information as the risk analysis data source of this invention, the data sources for risk analysis are expanded.

[0140] In summary, Figure 3 The technical solution presented includes three main services: acquisition and text extraction of inquiry letters and financial statement data; analysis of text and extraction of enterprise information using a large language model; data merging, cleaning, and transformation; and deployment of the transformed data. These services comprise a knowledge graph, an expert-based financial risk model scoring system, and a similarity retrieval system. For inquiry letters and financial statement notes, appropriate prompts are designed, and a large language model is used for risk analysis and data extraction. The results of risk analysis and data extraction using the large language model are transformed into triples, financial indicator data, and embedding vectors, serving as the data foundation for the knowledge graph, expert-based financial risk model, and similarity retrieval services. The knowledge graph is used to mine risks associated with the enterprise, the expert-based financial risk model is used to analyze the enterprise's own financial risks, and a search engine is used to find enterprises with similar financial risks. This expands the scope of risk analysis from three aspects, ensuring no conflicting results. Finally, the results are summarized using the large language model to generate a comprehensive risk assessment report. This invention provides a solution that uses a large language model to analyze inquiry letters from regulatory agencies and notes to financial statements published by companies to obtain risk assessment results. These results are then combined with knowledge graphs, similarity retrieval, and traditional expert-based financial analysis methods to identify more companies with similar financial risks. Ultimately, this solution helps banks and credit departments to comprehensively and effectively analyze corporate financial risks. This invention improves the accuracy, comprehensiveness, and timeliness of financial risk identification, achieving more accurate, comprehensive, and rapid risk analysis and discovery.

[0141] Embodiments of this application also provide an enterprise financial risk analysis device 400. Figure 5 This is a block diagram of the enterprise financial risk analysis device 400 implemented according to this application. (Reference) Figure 5 As shown, the enterprise financial risk analysis device 400 may include a file acquisition module 410 and a large language model analysis module 420 that are connected to each other.

[0142] The document acquisition module 410 is used to acquire inquiry letters issued to enterprises, as well as the financial statements and notes to the financial statements issued by the enterprises. The large language model analysis module 420 can be used to generate an inquiry letter risk analysis report by analyzing the inquiry letter and financial statements using a large language model based on the prompts for the inquiry letters, and to generate a note-to-note risk analysis report by analyzing the notes to the financial statements using a large language model based on the prompts for the notes to the financial statements. Both the inquiry letter risk analysis report and the note-to-note risk analysis report include: triplet data associated with the enterprise, and a text describing the enterprise's financial risks.

[0143] The enterprise financial risk analysis device 400 may further include a risk propagation analysis module 430 and a similarity retrieval module 440, which are communicatively connected to the large language model analysis module 420. The risk propagation analysis module 430 can be used to store triplet data into a graph database to form a knowledge graph, and analyze the risk propagation relationship between the target enterprise and other enterprises based on the knowledge graph. The similarity retrieval module 440 can be used to vectorize the financial risk description text and store it in a vector database, and perform similarity retrieval of risk features based on the vector database to find other enterprises whose risk features are similar to the target enterprise to a predetermined threshold. The enterprise financial risk analysis device 400 may possess the advantages described above regarding enterprise financial risk analysis methods.

[0144] In some embodiments, the enterprise financial risk analysis device 400 may further include a risk scoring module 450 communicatively connected to the large language model analysis module 420. In this embodiment, both the inquiry letter risk analysis report and the supplementary risk analysis report may also include financial indicator data. The risk scoring module 450 can be used to: store the financial indicator data in a relational database; and apply preset judgment rules to score the enterprise's risk based on the financial indicator data in the relational database.

[0145] Embodiments of this application also provide a computer-readable storage medium storing at least one instruction, which is executed by a processor in a computer device to implement the above-described enterprise financial risk analysis method, for example, referring to... Figures 1 to 4 The methods for analyzing corporate financial risks are described.

[0146] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for analyzing corporate financial risks, characterized in that, Includes the following steps: Obtain the inquiry letters issued to the company, and obtain the financial statements and notes to the financial statements issued by the company; By analyzing the prompts in the inquiry letter and the financial statements using a large language model, an inquiry letter risk analysis report is generated. Similarly, by analyzing the prompts in the notes to the financial statements using the same large language model, a notes risk analysis report is generated. Both the inquiry letter risk analysis report and the notes risk analysis report include: triplet data associated with the company and a description of the company's financial risks. The triplet data is stored in a graph database to form a knowledge graph, and the risk propagation relationship between the target enterprise and other enterprises is analyzed based on the knowledge graph. The financial risk description text is vectorized and stored in a vector database. Based on the vector database, a similarity retrieval of risk features is performed to find other companies whose risk features are similar to those of the target company and meet a predetermined threshold.

2. The method as described in claim 1, characterized in that, Also includes: The triplet data in the inquiry letter risk analysis report and the triplet data in the note risk analysis report are deduplicated, wherein the triplet data in the note risk analysis report is retained when data conflicts exist. Perform a union operation on the financial risk description text of the inquiry letter risk analysis report and the financial risk description text of the footnote risk analysis report.

3. The method as described in claim 2, characterized in that, The risk transmission relationship between the target enterprise and other enterprises is analyzed based on the knowledge graph, including at least one of the following: Use graph database query language to perform N-degree relational queries to trace multi-level risk transmission paths between enterprises, where N is an integer between 2 and 5; A risk propagation tree is constructed based on a breadth-first search algorithm to identify the chain risks of upstream and downstream related enterprises; A community partitioning algorithm combined with density clustering is used to identify abnormal subgraph structures in order to identify abnormal enterprise association circles; The PageRank algorithm is used to iteratively calculate the importance weights of enterprise nodes in order to identify core hub enterprises in risk networks.

4. The method as described in claim 2, characterized in that, Also includes: The financial risk description text after taking the union is vectorized using embedding; The similarity search was performed using the Faiss search engine. Different retrieval strategies are selected based on the application scenario, which includes at least one of enterprise benchmarking, risk warning, and document tracing.

5. The method as claimed in claim 1, characterized in that, Both the inquiry letter risk analysis report and the notes risk analysis report include financial indicator data, which includes the name and value of the financial indicator. The methodology also includes: Store the financial indicator data in a relational database; Risk scoring is performed on enterprises based on the financial indicator data in the relational database and the application of preset judgment rules.

6. The method as described in claim 5, characterized in that, Also includes: The financial indicator data in the risk analysis report of the inquiry letter and the financial indicator data in the risk analysis report in the notes are deduplicated. When there is a data conflict, the financial indicator data in the risk analysis report in the notes is retained. The deduplicated financial indicator data is stored in the relational database.

7. The method as claimed in claim 5, characterized in that, The preset judgment rules include at least one of the following: abnormal accounts receivable turnover, excessive related-party transactions, insufficient inventory write-downs, cash flow disruption warnings, and goodwill impairment risks. Each judgment rule defines the threshold conditions for financial indicators, the corresponding risk level, risk label, and verification suggestions. When the financial indicator data meets the threshold conditions of the corresponding rule, the corresponding rule is triggered and the risk level, risk label, and verification suggestions are output.

8. The method as claimed in claim 1, characterized in that, Also includes: Using a PDF text extraction tool, text information in PDF format inquiry letters and financial statement notes can be automatically identified and extracted to generate inquiry letter risk analysis reports and note risk analysis reports.

9. A device for analyzing corporate financial risks, characterized in that, include: The document acquisition module is used to acquire inquiry letters issued to enterprises, as well as financial statements and notes to financial statements issued by the enterprises. The large language model analysis module is used to generate an inquiry letter risk analysis report by analyzing the inquiry letter and the financial statements using a large language model based on the prompt words in the inquiry letter. It is also used to generate a financial statement notes risk analysis report by analyzing the financial statement notes using the large language model based on the prompt words in the financial statement notes. Both the inquiry letter risk analysis report and the financial statement notes risk analysis report include: triplet data associated with the enterprise, and a description of the enterprise's financial risks. The risk propagation analysis module stores the triplet data into a graph database to form a knowledge graph, and analyzes the risk propagation relationship between the target enterprise and other enterprises based on the knowledge graph. The similarity retrieval module vectorizes the financial risk description text and stores it in a vector database. Based on the vector database, it performs a similarity retrieval of risk features to find other companies whose risk features are similar to those of the target company and meet a predetermined threshold.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one instruction, which is executed by a processor in a computer device to implement the enterprise financial risk analysis method as described in any one of claims 1 to 8.

Images