Financial statement processing method and device, storage medium and electronic equipment

Abstract

The invention discloses a financial statement processing method and device, a storage medium and electronic equipment, and relates to the field of artificial intelligence. The method comprises the following steps: receiving a target financial statement of a customer; determining multi-modal information in the target financial statement; target rules and target content are determined from a target knowledge base based on the multi-modal information, the target knowledge base is constructed according to historical financial statement analysis results and risk control rules, and the target rules are used for representing risk rules associated with target financial statements; the target content is the content needing to be analyzed when the target financial statement is analyzed; updating the cue word template according to the target rule and the target content; and inputting the multi-modal information and the updated cue word template into a target large model, and generating a report analysis result for the target financial report according to the target large model. The technical problem that in the prior art, when financial statement analysis is conducted through AI, the timeliness is poor is solved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, and more specifically, to a financial statement processing method, apparatus, storage medium, and electronic device. Background Technology

[0002] In the current credit business environment of financial institutions, financial statement analysis is a key step in assessing a company's financial condition and risk; however, this process generally faces serious timeliness challenges.

[0003] Traditional analytical methods rely on manual interpretation and evaluation of financial data, which is not only time-consuming but also inefficient and ill-suited to a rapidly changing business environment. With the increasing complexity of corporate financial statement data, including the emergence of multimodal formats such as combinations of charts, text, and tables, and the use of non-standard accounting items, the difficulty and time required for manual analysis have further increased. Furthermore, the construction and training of customized deep learning models require large amounts of corporate financial statement data and professional data annotation, resulting in long implementation cycles, high costs, and limited model generalization capabilities, making it difficult to cope with newly emerging financial statement formats and accounting items.

[0004] There is currently no effective solution to the above problems. Summary of the Invention

[0005] This application provides a financial statement processing method, apparatus, storage medium, and electronic device to at least solve the technical problem of poor timeliness in the prior art when using AI for financial statement analysis.

[0006] According to one aspect of this application, a financial statement processing method is provided, comprising: receiving a client's target financial statement; determining multimodal information in the target financial statement; determining target rules and target content from a target knowledge base based on the multimodal information, wherein the target knowledge base is constructed based on historical financial statement analysis results and risk control rules, the target rules are used to characterize risk rules associated with the target financial statement, and the target content is the content to be analyzed when performing the target financial statement analysis; updating a prompt word template according to the target rules and target content, wherein the prompt word template includes a set of instructions for guiding a target big model to perform financial statement analysis; inputting the multimodal information and the updated prompt word template into the target big model, and generating financial statement analysis results for the target financial statement based on the target big model.

[0007] Optionally, determining the multimodal information in the target financial statement includes: converting the target financial statement into a target format, and using the text information extracted from the target financial statement in the target format as the first text information; extracting text information from the images and tables in the target financial statement, and using the text information extracted from the images and tables in the target financial statement as the second text information; fusing the first text information and the second text information to obtain initial multimodal information; standardizing the initial multimodal information according to a preset standard to obtain the first multimodal information, wherein the standardization process is used to convert the initial multimodal information into the same format; filtering sensitive information in the first multimodal information, and using the filtered first multimodal information as the multimodal information.

[0008] Optionally, determining target rules and target content from the target knowledge base based on multimodal information includes: retrieving target rules and target content associated with multimodal information from the target knowledge base through a target approach.

[0009] Optionally, after determining the multimodal information in the target financial statement, the method further includes: detecting whether the multimodal information conforms to preset rules through a rule engine, wherein the rule engine is used to verify the completeness and consistency of the multimodal information, and the preset rules are used to constrain the structure and format of the multimodal information; when the multimodal information is detected to conform to the preset rules, detecting whether there is abnormal data in the multimodal information; when abnormal data is detected in the multimodal information, generating a prompt message.

[0010] Optionally, after inputting the multimodal information and the updated prompt word template into the target large model, and generating the report analysis results for the target financial statements based on the target large model, the method further includes: receiving the review results of the report analysis results from the target personnel; and dynamically adjusting the structure in the prompt word template and the weights of different prompt words in the prompt word template based on the review results.

[0011] Optionally, the report analysis results may include at least the earnings trend, earnings change rate, ratio analysis results, attribution analysis results, and confidence level of the report analysis results in the client's target financial statements.

[0012] Optionally, the prompt word template is obtained through the following steps: obtaining the analysis objective for the target financial statement and determining the target financial indicator set based on the analysis objective; determining the analysis process for the target financial statement; and constructing the prompt word template based on the target example content, the target financial indicator set, the analysis process, business knowledge, and business rules, wherein the target example content is a preset report analysis example.

[0013] According to another aspect of this application, a financial statement processing apparatus is also provided, comprising: a receiving unit for receiving a target financial statement from a client; a first determining unit for determining multimodal information in the target financial statement; a second determining unit for determining target rules and target content from a target knowledge base based on the multimodal information, wherein the target knowledge base is constructed based on historical financial statement analysis results and risk control rules, the target rules are used to characterize risk rules associated with the target financial statement, and the target content is the content to be analyzed when performing the target financial statement analysis; an updating unit for updating a prompt word template according to the target rules and target content, wherein the prompt word template includes a set of instructions for guiding a target big model to perform financial statement analysis; and a generating unit for inputting the multimodal information and the updated prompt word template into the target big model, and generating a statement analysis result for the target financial statement based on the target big model.

[0014] According to another aspect of this application, a computer-readable storage medium is also provided, which includes a stored executable program, wherein, when the executable program is executed, it controls the device where the computer-readable storage medium is located to perform the above-described financial statement processing method.

[0015] According to another aspect of this application, an electronic device is also provided, including one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to perform the above-described financial statement processing method.

[0016] According to another aspect of the embodiments of this application, a computer program product is also provided, including computer instructions that, when executed by a processor, implement the steps of the above-described financial statement processing method.

[0017] In this application, the client's target financial statements are first received. Then, the multimodal information in the target financial statements is determined. Based on the multimodal information, target rules and target content are determined from a target knowledge base. The target knowledge base is constructed based on historical financial statement analysis results and risk control rules. The target rules are used to characterize the risk rules associated with the target financial statements. The target content is the content that needs to be analyzed when performing the target financial statement analysis. Then, the prompt word template is updated according to the target rules and target content. The prompt word template includes a set of instructions used to guide the target big model in performing financial statement analysis. Finally, the multimodal information and the updated prompt word template are input into the target big model, and the report analysis results for the target financial statements are generated based on the target big model. By employing multimodal intelligent analysis and an adaptive prompt word template adjustment mechanism, the efficiency of AI financial statement analysis has been significantly improved. Specifically, this approach leverages the characteristics of a general-purpose large model to achieve rapid and accurate analysis of financial statements without relying on extensive fine-tuning of financial data. It eliminates the need for time-consuming data annotation by personnel, lowering the technical application threshold and shortening the analysis cycle. This achieves the technical effects of rapid response, accurate identification, and intelligent analysis based on a general-purpose large model, thereby solving the technical problem of poor timeliness in existing AI-based financial statement analysis. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0019] Figure 1 This is a flowchart of an optional financial statement processing method according to an embodiment of this application;

[0020] Figure 2 This is a schematic diagram of an optional financial statement processing apparatus according to an embodiment of this application. Detailed Implementation

[0021] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0022] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0023] It should be noted that the information collected in this application (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for display, data used for analysis, etc.) are information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of this data all comply with relevant laws, regulations, and standards, necessary confidentiality measures have been taken, and they do not violate public order and good morals. Corresponding access points are provided for users to choose to authorize or refuse. For example, interfaces are set up between this system and relevant users or organizations, providing users with corresponding access points to choose to agree to or refuse automated decision-making results; if the user chooses to refuse, the process proceeds to the expert decision-making stage.

[0024] According to an embodiment of this application, a method embodiment for financial statement processing is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0025] It should be noted that an intelligent processing system can serve as the execution entity for the financial statement processing method in this application embodiment. It is understood that the financial statement processing method provided in this application embodiment can also be executed by other systems or devices, and this application embodiment does not specifically limit this.

[0026] Figure 1 This is a flowchart of an optional financial statement processing method according to an embodiment of this application, such as... Figure 1 As shown, the method includes the following steps:

[0027] Step S101: Receive the customer's target financial statement.

[0028] Optionally, the target financial statements refer to the financial reports of a specific company that a financial institution needs to analyze in its corporate lending business, including the balance sheet, income statement, and cash flow statement.

[0029] Optionally, the intelligent processing system receives financial statements submitted by corporate clients of financial institutions, typically in PDF, Excel, or other electronic document formats, in preparation for subsequent multimodal data analysis and intelligent parsing.

[0030] Step S102: Determine the multimodal information in the target financial statement.

[0031] Optionally, multimodal information refers to various types of data information, such as text, images, and tables, contained in the target financial statements.

[0032] Optionally, leveraging the capabilities of a multimodal general-purpose large model, the intelligent processing system will analyze all available information in the financial statements, including but not limited to textual descriptions, various charts (such as bar charts and pie charts), and data tables, to gain a more comprehensive understanding of the company's financial situation.

[0033] Optionally, identifying multimodal information can help supplement the shortcomings of single-source data (usually text) and improve the accuracy and depth of the analysis results.

[0034] Step S103: Determine target rules and target content from the target knowledge base based on multimodal information.

[0035] In step S103, the target knowledge base is constructed based on the analysis results of historical financial statements and risk control rules. The target rules are used to characterize the risk rules associated with the target financial statements, and the target content is the content that needs to be analyzed when conducting the analysis of the target financial statements.

[0036] Optionally, the target knowledge base consists of historical financial statement analysis results and risk control rules accumulated by financial institutions, including financial analysis models, industry standards, regulatory policies, etc.

[0037] Optionally, target rules refer to risk assessment rules related to target financial statements, used to verify the reasonableness of financial data and identify potential financial risk points.

[0038] Optionally, the target content refers to the key financial indicators and analytical directions that the intelligent processing system should focus on when conducting financial statement analysis.

[0039] Optionally, based on the collected multimodal information, the intelligent processing system will query the target knowledge base to find analytical rules and financial content that needs to be evaluated, such as key indicators like current ratio and return on assets, that match the report type and industry.

[0040] Optionally, defining the target rules and content helps to standardize and guide the analysis process of large models, making them more in line with the financial analysis standards and risk preferences of financial institutions.

[0041] Step S104: Update the prompt word template according to the target rules and target content.

[0042] In step S104, the prompt word template includes a set of instructions for guiding the target large model in financial statement analysis.

[0043] Optionally, a prompt template contains a series of guiding statements and questions to stimulate the thinking process of large models in financial statement analysis, ensuring the comprehensiveness and depth of the analysis results.

[0044] Optionally, based on the target rules and target content determined in the previous step, the system will customize and update the prompt word template, adding specific analysis instructions and questions, such as requiring the model to assess the company's solvency and market prospects.

[0045] Optionally, the updated prompt word template is designed as a set of instructions to guide large models to focus on specific financial indicators and risk points during the analysis process, while making more refined trend predictions and causal inferences.

[0046] Step S105: Input the multimodal information and the updated prompt word template into the target large model, and generate report analysis results for the target financial statements based on the target large model.

[0047] Optionally, the target large model refers to a general intelligent model that has been pre-trained and has powerful multimodal data processing and generalization capabilities.

[0048] Optionally, by leveraging the multimodal general-purpose large model's ability to understand different modalities of data such as text, images, and tables, the large model can make full use of different modalities and heterogeneous data to gain a comprehensive and accurate understanding of the company's financial situation.

[0049] Optionally, the system inputs the multimodal information and customized prompt word templates obtained in the preceding steps into the target large model. After receiving this information, the large model performs in-depth analysis based on the prompt word templates, which can not only quickly generate conclusions about the company's financial health but also predict future financial trends.

[0050] Optionally, the analysis results will include the direction and magnitude of changes in corporate financial returns and their attribution analysis, as well as the confidence level assessment of this analysis, to provide a scientific and objective basis for financial institutions' credit decisions.

[0051] Optionally, this step is the core of the entire analysis process. The results obtained through the calculation of the large model can effectively solve the problems of insufficient timeliness and accuracy in traditional methods, while lowering the technical threshold and improving the user experience.

[0052] As can be seen from steps S101 to S105, in this application, the target financial statements of the client are first received. Then, the multimodal information in the target financial statements is determined. Then, the target rules and target content are determined from the target knowledge base based on the multimodal information. The target knowledge base is constructed based on the analysis results of historical financial statements and risk control rules. The target rules are used to characterize the risk rules associated with the target financial statements. The target content is the content that needs to be analyzed when performing the analysis of the target financial statements. Then, the prompt word template is updated according to the target rules and target content. The prompt word template includes a set of instructions used to guide the target big model to perform financial statement analysis. Finally, the multimodal information and the updated prompt word template are input into the target big model, and the report analysis results for the target financial statements are generated based on the target big model. By employing multimodal intelligent analysis and an adaptive prompt word template adjustment mechanism, the efficiency of AI financial statement analysis has been significantly improved. Specifically, this approach leverages the characteristics of a general-purpose large model to achieve rapid and accurate analysis of financial statements without relying on extensive fine-tuning of financial data. It eliminates the need for time-consuming data annotation by personnel, lowering the technical application threshold and shortening the analysis cycle. This achieves the technical effects of rapid response, accurate identification, and intelligent analysis based on a general-purpose large model, thereby solving the technical problem of poor timeliness in existing AI-based financial statement analysis.

[0053] In one optional embodiment, the intelligent processing system converts the target financial statement into a target format, and uses the text information extracted from the target financial statement in the target format as first text information. It also extracts text information from the images and tables in the target financial statement and uses the text information extracted from the images and tables in the target financial statement as second text information. Then, it merges the first text information and the second text information to obtain initial multimodal information. After that, it performs standardization processing on the initial multimodal information according to a preset standard to obtain first multimodal information. The standardization processing is used to convert the initial multimodal information into the same format. Finally, it filters sensitive information in the first multimodal information and uses the filtered first multimodal information as multimodal information.

[0054] Optionally, the intelligent processing system first converts the target financial statements from the client into a unified PDF format (target format). This conversion ensures the consistency of the data source and lays the foundation for subsequent multimodal information extraction. Regardless of whether the original report is submitted in Excel, Word, or other electronic document format, after conversion to PDF format, the text, images, tables, and other elements in the report will be fully preserved, facilitating data parsing and extraction in subsequent steps.

[0055] Optionally, the intelligent processing system uses a PDF text extraction component to perform in-depth analysis of the converted financial statements, transforming all text content into structured data to form the first textual information. This information covers the basic structure and core data points of the company's financial reports, forming the foundation for financial analysis. Next, the intelligent processing system employs advanced OCR (Optical Character Recognition) technology and table recognition algorithms. The system delves into the PDF-formatted financial statements, extracting all possible textual information containing financial data from charts, images, and tables to form the second textual information. This step enhances the comprehensiveness of the information, ensuring that the analysis model can access all important data in the statements, not just the obvious textual information.

[0056] Optionally, the intelligent processing system integrates the first textual information (information extracted directly from PDF text) and the second textual information (information extracted from images and tables) to form an initial multimodal information set containing various types of data. This fusion process ensures that the system can utilize all available data when performing financial analysis, improving the comprehensiveness and depth of the analysis. Then, referring to the standardized format of Compustat (a widely used financial database) financial datasets, the system standardizes the fused initial multimodal information, including unifying data formats, standardizing numerical units, and standardizing the description of financial indicators. The standardized first multimodal information not only eliminates the differences caused by different report formats and data representations but also improves data readability and the processing efficiency of the analysis model.

[0057] Optionally, to ensure the security and compliance of financial analysis, the intelligent processing system will perform sensitive information filtering on the first multimodal information obtained after standardization, that is, remove all sensitive data associated with the enterprise identity and retain pure financial data and indicators. After this desensitization step, the filtered information is used as the final multimodal information for subsequent analysis.

[0058] As described above, through the specific implementation steps, the intelligent processing system can efficiently convert financial statements of any format into PDF format and comprehensively extract text, images, and tabular data. Subsequently, through standardization and desensitization of this multimodal information, the system prepares a unified and clean data foundation for subsequent financial analysis, ensuring the accuracy and compliance of the analysis. This process not only significantly improves the efficiency of financial statement analysis but also enhances the depth and breadth of the analysis through standardization.

[0059] In one alternative embodiment, the intelligent processing system retrieves target rules and target content associated with multimodal information from a target knowledge base in a target manner.

[0060] Optionally, the target approach (RAG paradigm): RAG, short for Retrieval-Augmented Generation, is an AI technology paradigm that combines retrieval and generation. In this embodiment, the intelligent processing system adopts the RAG paradigm to enhance the financial statement analysis generation process by retrieving professional rules and content related to multimodal information from the target knowledge base.

[0061] Optionally, upon receiving standardized and desensitized multimodal information, the intelligent processing system initiates the RAG paradigm, retrieving professional rules and analytical content closely related to the financial statement data from the target knowledge base. First, the system uses natural language processing technology to extract keywords from the multimodal information, such as "current ratio," "return on assets," and "market prospects." These keywords form the basis for retrieving target rules and content. Based on these keywords, the system searches and matches all relevant financial analysis rules and industry standards in the target knowledge base. For example, for the extracted "current ratio," the system will retrieve the corresponding calculation formula, industry average, and risk assessment standards. The system further analyzes the context within the multimodal information to ensure that the retrieved target rules and content match the specific circumstances of the current financial statements. This process may involve judging the applicability of different rules and content to ensure the relevance of the analysis results. Finally, the system integrates the retrieved target rules and content into the financial statement analysis process, providing professional guidance for subsequent prompt word templates and ensuring the depth and accuracy of the analysis.

[0062] Optionally, leveraging the RAG paradigm, the target knowledge base is cleverly used as an "external brain," effectively compensating for the general-purpose model's shortcomings in specialized domain knowledge. In specific implementation, the multimodal general-purpose model first parses and extracts multi-dimensional data from financial information, then retrieves relevant risk control rules and key points of financial analysis from the knowledge base, integrating these professional guidelines into the inference prompt word templates generated by the model, thereby significantly improving the accuracy and depth of financial statement analysis.

[0063] As described above, the intelligent processing system enhances financial statement analysis by employing the RAG paradigm to retrieve target rules and content from the target knowledge base. This paradigm not only enriches the professional knowledge base of the analysis process but also ensures the compliance and accuracy of the analysis results. The application of the RAG paradigm enables the system to automatically retrieve the most relevant analysis rules and content based on the specific information in the financial statements, avoiding the subjectivity and time-consuming nature of manual selection. This technology significantly improves the automation level of financial statement analysis, reduces reliance on professional personnel, and simultaneously ensures the professionalism and depth of the analysis. It provides strong technical support for financial institutions' credit operations, accelerates the decision-making process, and improves the overall efficiency and quality of business operations.

[0064] In one optional embodiment, the intelligent processing system uses a rule engine to detect whether the multimodal information conforms to preset rules. The rule engine is used to verify the integrity and consistency of the multimodal information, and the preset rules are used to constrain the structure and format of the multimodal information. When the multimodal information is detected to conform to the preset rules, the system detects whether there is abnormal data in the multimodal information. When abnormal data is detected in the multimodal information, a prompt message is generated.

[0065] Optionally, the rules engine is specifically designed to process and execute predefined business rules, automatically verifying and analyzing the integrity and consistency of data to ensure that the data conforms to specific structural and format requirements.

[0066] Optionally, preset rules: a set of rules predefined by financial institutions based on financial analysis needs and industry standards, including data format, numerical range, consistency checks, etc., used to constrain the structure and format of multimodal information and ensure data quality.

[0067] Optionally, before starting the analysis, the intelligent processing system first activates the rules engine and loads preset rules. These rules include, but are not limited to, format checks of financial data, verification of the reasonableness of numerical ranges, and consistency checks between data. The system submits the standardized multimodal information to the rules engine, which then performs a comprehensive check on the data, verifying its completeness, correct format, and whether the values ​​are within a reasonable range. The preset rules constrain the structure and format of the multimodal information, ensuring that the data is presented in a unified standard; for example, all currency values ​​must be expressed in the same unit, and all ratio values ​​must be between 0 and 1.

[0068] Optionally, after verification by the rules engine, the system will further check for abnormal data in the multimodal information, such as isolated extremely large or small values, or data changes that do not conform to industry trends. If data that does not conform to preset rules or is abnormal is found in the multimodal information during the detection process, the system will automatically generate corresponding prompts to remind analysts or automated processes to pay attention to these potential problems.

[0069] As demonstrated above, the intelligent processing system, by introducing a rule engine, ensures the high quality and compliance of multimodal information. The comprehensive verification by the rule engine not only guarantees data integrity and consistency but also ensures that the data structure and format conform to industry standards and internal regulations of financial institutions through the constraints of preset rules. The detection of abnormal data and the generation of alerts further enhance the system's robustness and the credibility of the analysis results. The implementation of this process significantly reduces the risk of data errors in financial statement analysis, improves the accuracy and efficiency of analysis, and provides a high-quality data foundation for subsequent financial analysis. Overall, the implementation of this step significantly improves the intelligence level of financial statement analysis, provides strong technical support for risk assessment and decision-making in financial institutions' credit business, and effectively avoids analytical biases and decision-making errors caused by data quality issues.

[0070] In one alternative embodiment, the intelligent processing system receives the target personnel's review of the report analysis results, and then dynamically adjusts the structure of the prompt word template and the weight of different prompt words in the prompt word template based on the review results.

[0071] Optionally, the target personnel refers to reviewers with professional knowledge and experience, typically including senior financial analysts, credit approval team members, or external financial advisors from financial institutions.

[0072] Optionally, the CoT (Chain-of-Thought) prompt template is a prompt template used to guide AI models in in-depth thinking and reasoning. In this embodiment, the CoT prompt template is designed to guide a multimodal general-purpose model in accurate analysis of financial statements, including trend analysis, ratio analysis, and attribution analysis.

[0073] Optionally, the intelligent processing system first delivers the generated financial statement analysis results to the target personnel for review. Reviewers will use their professional knowledge and experience to judge the accuracy, comprehensiveness, and logic of the analysis results, as well as their practicality for business decisions. The system receives feedback from the target personnel, including affirmations of the analysis results, suggestions for correction, and additional analytical perspectives or focuses. This feedback is recorded and serves as an important basis for optimizing the CoT prompt template.

[0074] Optionally, based on the review results, the intelligent processing system will analyze which analytical steps received positive feedback and which steps need improvement. The system may adjust the structure of the CoT prompt template, such as adding new analytical dimensions or adjusting the order of the analytical process, to better adapt to the characteristics of financial statements. The intelligent processing system will dynamically adjust the weight of different prompts in the CoT prompt template based on the review feedback from the target personnel. If a prompt plays a key role in the analysis and receives high praise, the system may increase its weight; conversely, it may decrease its weight or even exclude it from future analyses.

[0075] Optionally, this adjustment process is not static but continuous. The intelligent processing system treats each review by the target personnel as a learning opportunity, and gradually optimizes the CoT prompt template through continuous structural adjustments and weight updates, making it more aligned with the actual needs of financial institutions' corporate client financial statement analysis.

[0076] As described above, the intelligent processing system can dynamically optimize the CoT prompt template based on the review results of the target personnel, achieving continuous improvement in analysis effectiveness. This mechanism ensures the flexibility and adaptability of the AI ​​analysis model, enabling it to continuously learn and improve its analytical skills to cope with various complex situations in financial statements. Over time, the CoT prompt template will become increasingly accurate, more effectively guiding large models to conduct in-depth financial analysis and improving the quality and reliability of the analysis results. Furthermore, this dynamic adjustment mechanism promotes the deep integration of AI models and human expertise, allowing financial institutions to leverage the advantages of AI technology while maintaining the rigor and professionalism of their analysis, providing stronger data support for credit decisions.

[0077] In one optional embodiment, the report analysis results include at least the revenue trend, revenue change magnitude, ratio analysis results, attribution analysis results, and confidence level of the report analysis results in the client's target financial statements.

[0078] Optionally, earnings trend refers to the direction of earnings change over time in the client's target financial statements, and is an important indicator for assessing a company's financial health and future potential.

[0079] Optionally, the magnitude of revenue change: quantifies the magnitude of revenue change over time, helping to understand the stability or volatility of the company's operations.

[0080] Optionally, ratio analysis results: assess a company’s financial condition and operational efficiency by calculating and analyzing financial ratios (such as current ratio, return on assets, etc.).

[0081] Optionally, attribution analysis results: explain the reasons behind the analysis results, especially the reasons for changes in earnings, to help understand the specific factors affecting the company's financial performance.

[0082] Optionally, the confidence level of the report analysis results: This represents the reliability of the report analysis results and is calculated based on factors such as the prediction accuracy of the model and the completeness of the data.

[0083] Optionally, the intelligent processing system first performs a revenue trend analysis on the client's target financial statements to determine whether revenue is increasing, decreasing, or remaining stable. This analysis is typically based on a comparison of financial data from multiple periods, using predictive models to identify patterns of revenue change over time. After determining the revenue trend, the system further calculates the magnitude of revenue changes, quantifying the size of the company's financial fluctuations and providing specific numerical support for the company's risk assessment.

[0084] Optionally, the intelligent processing system performs ratio analysis, calculating and interpreting a range of financial ratios, such as the current ratio and return on assets, to assess a company's financial health and operational efficiency. This analysis is based on standardized financial statements, ensuring comparability and objectivity.

[0085] Optionally, the system can perform attribution analysis on the trend and magnitude of changes in revenue, identify and explain key drivers affecting corporate financial performance, such as market environment, cost control, and investment decisions, providing in-depth insights.

[0086] Optionally, while generating the above analysis results, the system also calculates the confidence level of the report analysis results based on indicators such as the model's prediction accuracy, data integrity, and consistency, to ensure that users can understand the reliability of the analysis results.

[0087] As described above, the implementation method ensures that the report analysis results generated by the intelligent processing system cover key aspects of the enterprise's financial situation, including earnings trends, magnitude of changes, ratio analysis results, attribution analysis results, and confidence level assessments of these results. This comprehensive analysis not only provides an in-depth interpretation of financial data but also quantifies the reliability and uncertainty of the analysis results, enhancing decision-makers' understanding and trust in the results. Through this implementation method, credit approval personnel in financial institutions can obtain more comprehensive, accurate, and reliable risk assessment information, helping to make more informed credit decisions and avoiding the risks of decisions based on partial information or subjective judgment. Simultaneously, the system provides a measure of result quality by calculating confidence levels, making the analysis process more transparent and enhancing the scientific and professional nature of financial analysis.

[0088] In one optional embodiment, the prompt word template is obtained through the following steps: First, the intelligent processing system obtains the analysis target for the target financial statement and determines the target financial indicator set based on the analysis target. Then, it determines the analysis process for the target financial statement. Finally, based on the target example content, the prompt word template is constructed according to the target financial indicator set, analysis process, business knowledge, and business rules. The target example content is a preset report analysis example.

[0089] Optionally, the analysis objectives are defined based on the needs of the financial institution's credit business, aiming to achieve certain goals through financial statement analysis, such as assessing a company's solvency, market competitiveness, or profitability; the target financial indicator set is a series of key financial indicators selected based on the analysis objectives to assess the company's financial condition, such as the current ratio, debt-to-equity ratio, and net profit growth rate; the analysis process refers to a set of analytical steps and strategies followed in financial statement analysis according to specific business logic or professional requirements, ensuring the systematic nature and accuracy of the analysis; business knowledge includes professional knowledge and experience in the field of financial institution credit business, including but not limited to industry characteristics, risk management principles, and credit policies, which are important bases for the design of the analysis process; business rules are internal rules and standards set by the financial institution regarding financial analysis, credit approval, etc., used to constrain analytical behavior and ensure the compliance and consistency of decision-making; and the target example content includes preset report analysis examples used to guide and calibrate the design of the CoT prompt word template, ensuring that the analysis results generated by the model meet the expected format and depth.

[0090] Optionally, the intelligent processing system first receives and understands the analytical objectives of the financial institution's credit business, clarifying the focus and purpose of the analysis, such as assessing the company's financial soundness. Based on the analytical objectives, the system selects a series of relevant financial indicators from a financial indicator library to form a target financial indicator set. These indicators include, but are not limited to, current ratio, debt-to-equity ratio, net profit growth rate, and return on assets, which are directly related to the analytical objectives. Next, based on business needs and professional knowledge, the system designs a reasonable analytical process to ensure the structure and logic of the analysis. The process includes the calculation, comparison, and interpretation of various financial indicators, as well as an indicator-based assessment of financial health.

[0091] Optionally, the system references preset report analysis examples, which may be created based on past successful cases and contain the results of a complete financial statement analysis. These examples guide the construction of the CoT prompt template. Finally, combining the analysis objectives, target financial indicator set, analysis process, business knowledge and rules, and target example content, the intelligent processing system constructs the CoT prompt template. The template contains prompts that guide large models in complex reasoning and trend prediction, ensuring the depth and accuracy of the analysis results.

[0092] Optionally, the intelligent processing system will apply the constructed CoT prompt template to actual financial statement analysis tasks, verify the effectiveness of the template through expert review and actual business results, and make necessary adjustments based on feedback to continuously improve and optimize the template.

[0093] Optionally, the intelligent processing system, through specially designed CoT prompt templates, can also request the large model to analyze and predict the direction of financial changes in the future (t+1, t+2) based on the company's current financial statements (t), in order to simulate the thought process of human analysts. The CoT prompts from this stage are then further stacked into the CoT prompt template and finally sent to the large model for reasoning and generating financial statement analysis results.

[0094] As described above, the implementation method ensures that the intelligent processing system, when constructing the CoT prompt template, fully considers the analysis objectives, financial indicators, analysis processes, business knowledge and rules of financial institutions' credit business, as well as past report analysis examples. This systematic construction process not only improves the relevance and depth of model analysis but also enhances the accuracy and reliability of analysis results, reducing subjectivity and bias. Furthermore, continuous validation and adjustment of the template ensures its effectiveness, enabling it to adapt to constantly changing business needs and market environments, providing flexible, efficient, and professional financial statement analysis technical support for financial institutions' credit business.

[0095] In one optional embodiment, building upon the strong feature fitting and generalization emergence capabilities of a general-purpose large model, this embodiment constructs a financial statement processing approach that combines a rule engine with a pre-trained large model. The rule engine handles and verifies basic financial rules and anomaly detection, while the large model focuses on complex reasoning and trend prediction. This collaborative architecture between the large model and engineering technology significantly improves the robustness and accuracy of financial statement analysis. Furthermore, leveraging the general-purpose multimodal large model's ability to understand different modalities of data such as text, images, and tables, it fully utilizes rich textual information in financial reports, including charts, fonts, and complex reports. Finally, through an automated feedback mechanism, the model can automatically adjust the structure and weights of CoT prompts based on expert review and actual business results, forming a closed loop to continuously optimize its analytical capabilities, thereby constantly improving the model's financial analysis skills and accuracy. Through this design, this embodiment achieves financial statement analysis directly using a pre-trained general-purpose large language model without fine-tuning the large model based on financial data. This significantly lowers the technical threshold for applying this method, eliminating the need for time-consuming data annotation and the need to prepare substantial GPU resources for fine-tuning large models, thus greatly reducing the implementation cost of this method in real-world business scenarios.

[0096] This application also provides a financial statement processing apparatus. It should be noted that the financial statement processing apparatus of this application can be used to execute the financial statement processing method provided in this application. The following describes the financial statement processing apparatus provided in this application.

[0097] According to an embodiment of this application, an apparatus for implementing the above-described financial statement processing method is also provided. Figure 2 This is a schematic diagram of an optional financial statement processing apparatus according to an embodiment of this application, such as... Figure 2 As shown, the device includes: a receiving unit 201 for receiving a client's target financial statement; a first determining unit 202 for determining multimodal information in the target financial statement; a second determining unit 203 for determining target rules and target content from a target knowledge base based on the multimodal information, wherein the target knowledge base is constructed based on historical financial statement analysis results and risk control rules, the target rules are used to characterize risk rules associated with the target financial statement, and the target content is the content to be analyzed when performing target financial statement analysis; an updating unit 204 for updating a prompt word template according to the target rules and target content, wherein the prompt word template includes a set of instructions for guiding the target big model to perform financial statement analysis; and a generating unit 205 for inputting the multimodal information and the updated prompt word template into the target big model, and generating report analysis results for the target financial statement based on the target big model.

[0098] Optionally, the first determining unit 202 includes: a first extraction subunit, a second extraction subunit, a first fusion subunit, a first processing subunit, and a second processing subunit. The first extraction subunit is used to convert the target financial statement into a target format and use the text information extracted from the target financial statement in the target format as first text information; the second extraction subunit is used to extract text information from images and tables in the target financial statement and use the text information extracted from images and tables in the target financial statement as second text information; the first fusion subunit is used to fuse the first text information and the second text information to obtain initial multimodal information; the first processing subunit is used to standardize the initial multimodal information according to a preset standard to obtain first multimodal information, wherein the standardization process is used to convert the initial multimodal information into the same format; the second processing subunit is used to filter sensitive information in the first multimodal information and use the filtered first multimodal information as multimodal information.

[0099] Optionally, the second determining unit 203 includes: a first retrieval subunit, used to retrieve target rules and target content associated with multimodal information from the target knowledge base in a target manner.

[0100] Optionally, the financial statement processing device further includes: a first detection unit, a second detection unit, and a first generation unit. The first detection unit is used to detect whether the multimodal information conforms to preset rules using a rule engine. The rule engine is used to verify the completeness and consistency of the multimodal information, and the preset rules are used to constrain the structure and format of the multimodal information. The second detection unit is used to detect whether there is abnormal data in the multimodal information when it is detected that the multimodal information conforms to the preset rules. The first generation unit is used to generate a prompt message when abnormal data is detected in the multimodal information.

[0101] Optionally, the financial statement processing device further includes a first receiving unit and a first adjusting unit. The first receiving unit is used to receive the review results of the target personnel on the report analysis results; the first adjusting unit is used to dynamically adjust the structure of the prompt word template and the weights of different prompt words in the prompt word template based on the review results.

[0102] Optionally, the report analysis results may include at least the earnings trend, earnings change rate, ratio analysis results, attribution analysis results, and confidence level of the report analysis results in the client's target financial statements.

[0103] Optionally, the financial statement processing apparatus further includes: a first acquisition unit, a third determination unit, and a first construction unit. The first acquisition unit is used to acquire the analysis objective for the target financial statement and determine the target set of financial indicators based on the analysis objective; the third determination unit is used to determine the analysis process for the target financial statement; and the first construction unit is used to construct a prompt template based on the target example content, the target set of financial indicators, the analysis process, business knowledge, and business rules, wherein the target example content is a preset report analysis example.

[0104] According to another aspect of this application, a computer-readable storage medium is also provided, which includes a stored executable program, wherein, when the executable program is executed, it controls the device where the computer-readable storage medium is located to perform the above-described financial statement processing method.

[0105] According to another aspect of this application, an electronic device is also provided, including one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to perform the above-described financial statement processing method.

[0106] According to another aspect of the embodiments of this application, a computer program product is also provided, including computer instructions that, when executed by a processor, implement the steps of the above-described financial statement processing method.

[0107] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0108] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0109] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0110] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0111] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0112] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0113] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for processing financial statements, characterized in that, include: Receive the client's target financial statements; Identify the multimodal information in the target financial statements; Based on the multimodal information, target rules and target content are determined from the target knowledge base, wherein the target knowledge base is constructed based on the analysis results of historical financial statements and risk control rules, the target rules are used to characterize the risk rules associated with the target financial statements, and the target content is the content that needs to be analyzed when analyzing the target financial statements; The prompt word template is updated according to the target rules and the target content, wherein the prompt word template includes a set of instructions for guiding the target large model to perform financial statement analysis; The multimodal information and the updated prompt word template are input into the target big model, and the report analysis results for the target financial statement are generated based on the target big model.

2. The financial statement processing method according to claim 1, characterized in that, Determining the multimodal information in the target financial statements includes: The target financial statement is converted into a target format, and the text information extracted from the target financial statement in the target format is used as the first text information; Text information is extracted from the images and tables in the target financial statement, and the text information extracted from the images and tables in the target financial statement is used as second text information; The first text information and the second text information are fused to obtain the initial multimodal information; The initial multimodal information is standardized according to a preset standard to obtain first multimodal information, wherein the standardization process is used to convert the initial multimodal information into the same format; Sensitive information in the first multimodal information is filtered out, and the filtered first multimodal information is used as the multimodal information.

3. The financial statement processing method according to claim 1, characterized in that, Based on the aforementioned multimodal information, target rules and target content are determined from the target knowledge base, including: The target rules and target content associated with the multimodal information are retrieved from the target knowledge base using a target-oriented approach.

4. The financial statement processing method according to claim 1, characterized in that, After determining the multimodal information in the target financial statements, the method further includes: The rule engine is used to detect whether the multimodal information conforms to preset rules. The rule engine is used to verify the integrity and consistency of the multimodal information, and the preset rules are used to constrain the structure and format of the multimodal information. When the multimodal information is detected to conform to the preset rules, it is detected whether there is abnormal data in the multimodal information; When abnormal data is detected in the multimodal information, a prompt message is generated.

5. The financial statement processing method according to claim 1, characterized in that, After inputting the multimodal information and the updated prompt word template into the target large model, and generating report analysis results for the target financial statement based on the target large model, the method further includes: Receive the review results of the target personnel on the report analysis results; The structure of the prompt word template and the weight of different prompt words in the prompt word template are dynamically adjusted based on the review results.

6. The financial statement processing method according to claim 1, characterized in that, The report analysis results include at least the revenue trend, revenue change range, ratio analysis results, attribution analysis results, and confidence level of the report analysis results in the client's target financial statements.

7. The financial statement processing method according to claim 1, characterized in that, The prompt word template is obtained through the following steps: Obtain the analysis objectives for the target financial statements, and determine the set of target financial indicators based on the analysis objectives; Determine the analysis process for the target financial statements; Based on the target example content, the prompt word template is constructed according to the target set of financial indicators, the analysis process, business knowledge, and business rules, wherein the target example content is a preset report analysis example.

8. A financial statement processing device, characterized in that, include: The receiving unit is used to receive the customer's target financial statements. The first determining unit is used to determine the multimodal information in the target financial statement; The second determining unit is used to determine target rules and target content from the target knowledge base based on the multimodal information. The target knowledge base is constructed based on the analysis results of historical financial statements and risk control rules. The target rules are used to characterize the risk rules associated with the target financial statements. The target content is the content that needs to be analyzed when analyzing the target financial statements. An update unit is used to update the prompt word template according to the target rules and the target content, wherein the prompt word template includes a set of instructions for guiding the target large model to perform financial statement analysis; The generation unit is used to input the multimodal information and the updated prompt word template into the target large model, and generate report analysis results for the target financial statement based on the target large model.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein when the computer program is executed, the device in which the computer-readable storage medium is located performs the financial statement processing method according to any one of claims 1 to 7.

10. An electronic device, characterized in that, It includes one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to perform the financial statement processing method according to any one of claims 1 to 7.

11. A computer program product comprising computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the steps of the financial statement processing method according to any one of claims 1 to 7.

Images