Financial analysis report generation method and system based on large model thinking map

Abstract

The invention discloses a financial analysis report generation method and system based on a large model thinking graph, and the method comprises the steps: obtaining enterprise financial data uploaded by a user, and combining the enterprise financial data with external financial data provided by the system to serve as input data; the method comprises the following steps: disassembling a manufacturing process of a financial analysis report into a plurality of nodes, and constructing an instruction template Prompts; inputting the input data into an instruction template Prompts in a json format for processing, and outputting node content of each thinking map to obtain a node result; summarizing according to the node results to form a financial analysis report meeting template requirements, and terminating operation on nodes with obvious errors; for the node content which cannot be directly judged to be right or wrong, the deviation degree between the node content and the input data is judged, and preset node operation is executed through the controller; and combining the summary node of each paragraph into a final generation node, and forming and outputting an enterprise financial analysis report.

Description

Technical Field

[0001] The present invention relates to the field of large model technology, and more specifically, to a method and system for generating a financial analysis report based on a large model mind map. Background Art

[0002] With the continuous acceleration of the financial informatization and digital transformation of enterprises, the intelligent operation of various types of office in enterprises is increasingly used as a standard for judging the degree of digitalization of enterprises. Financial work is an important part of enterprise operations, and financial analysis reports are important supporting materials for enterprise business analysis, strategic decision-making, and transformation guidance. It is an indispensable part of the enterprise operation process. It provides enterprises with key information for in-depth understanding and evaluation of their own financial status, and can help financial managers and company owners quickly understand risks, performance, resource optimization, and form decision support. Although financial analysis reports are crucial to enterprise operations, they also face some difficulties in the preparation process. Financial analysis reports need to collect a large amount of financial data, including financial statements, accounting records, transaction data, etc. The collection and processing of these data requires a lot of time and effort, and it is necessary to ensure the accuracy and completeness of the data. At the same time, financial analysis reports must not only provide data and analysis results, but also explain and interpret these results. This requires the writer to have deep financial knowledge and rich experience in order to accurately convey information and answer possible questions. Financial analysis reports usually need to be prepared regularly, such as quarterly reports, annual reports, etc. It is a huge challenge for writers to complete high-quality report writing within a limited time. Most of the existing financial analysis reports use a combination of manual and expert systems to form preliminary comparison results for key information through rule comparison, and then form key conclusions through manual organization. This makes it difficult to quickly and efficiently form professional financial analysis reports.

[0003] With the rapid development of big language model technology, the role of using big model text generation capabilities to create various reports, analyses, and contracts has been deeply explored and applied. However, due to their own architecture and mechanism problems, the current big models have limited reasoning capabilities because they are trained on large amounts of network text data. These data may contain errors, biases, or outdated information, causing the model to produce inaccurate conclusions when reasoning, and the model may be insufficient when dealing with problems that require deep logical analysis, complex mathematical calculations, or professional domain knowledge. Summary of the invention

[0004] According to the present invention, a method and system for generating a financial analysis report based on a large model mind map are provided to solve the problem that the current large models have limitations in reasoning ability due to their own architecture and mechanism problems, because they are trained on a large amount of network text data. These data may contain errors, biases, or outdated information, causing the model to produce inaccurate conclusions when reasoning, and the model may show inadequacies when dealing with problems that require deep logical analysis, complex mathematical calculations, or professional domain knowledge. When dealing with problems that require deep logical analysis, complex financial mathematical calculations, or professional domain knowledge, the model will show insufficient technical problems.

[0005] According to a first aspect of the present invention, there is provided a method for generating a financial analysis report based on a large model mind map, comprising:

[0006] Obtain the enterprise financial data uploaded by the user and combine it with the external financial data provided by the system as input data. The enterprise financial data includes balance sheet, profit and loss statement and cash flow statement. The external financial data includes judicial data, industrial and commercial data, industry dynamics and policies and regulations.

[0007] Break down the process of making financial analysis reports into multiple nodes, including generation, repetition, calculation, summary, and reinforcement, and construct instruction template prompts;

[0008] Input the input data in json format into the instruction template Prompts for processing, output the node content of each mind map, and obtain the node result;

[0009] Summarize the node results to form a financial analysis report that meets the template requirements, score the content of each node, and terminate the operation for nodes with obvious errors;

[0010] For node contents that cannot be directly judged as right or wrong, the active learner is used to correct the nodes that need to be corrected and feedback is given to update the node contents. The controller is used to execute preset node operations to determine the total node of each paragraph; the summary nodes of each paragraph are merged into the final generated node to form and output the corporate financial analysis report.

[0011] Optionally, the process of making a financial analysis report is broken down into multiple nodes, including generation, repetition, calculation, summary, and reinforcement, and instruction template prompts are constructed, including:

[0012] Generate inference results on the business operation status of the enterprise based on the judicial data, industrial and commercial data, industry dynamics, and policies and regulations;

[0013] Use different calculation methods and different generation methods to repeat the previous node to ensure the accuracy of the generated node;

[0014] For calculations between multiple data, the calculation steps are broken down into simple calculation formulas, and the calculation results are summarized after the calculations are performed;

[0015] Summarize the text content to form an abstract and core ideas;

[0016] Score some of the generated results, and use reinforcement methods for outputs with higher scores to enhance the output content of the model.

[0017] Optionally, the node results are summarized to form a financial analysis report that meets the template requirements, each node content is scored, and the operation is terminated for nodes with obvious errors, including:

[0018] Score each node content to judge the accuracy of numerical calculations, language logic, policy references, and conclusions;

[0019] For nodes with obvious calculation errors, logical errors, inconsistent content, and reference conflicts, the termination operation will be performed and the node will no longer participate in the combination and calculation of the mind map.

[0020] Optionally, for node contents that cannot be directly judged as right or wrong, the active learner is used to correct the nodes that need to be corrected and feedback is provided to update the node contents, and the controller is used to execute preset node operations to determine the total node of each paragraph, including:

[0021] For node contents that cannot be directly judged as right or wrong, the active learner is used to determine the degree of deviation between the node content and the input data;

[0022] If the correlation between the node output content and the input data is less than the threshold, it is necessary for financial accountants to judge and correct it, and update the node content based on the corrected result feedback. The controller will execute the preset node operation and decide how to merge or split the nodes at this layer.

[0023] Optionally, for node contents that cannot be directly judged as right or wrong, an active learner is used to judge the degree of deviation between the node contents and the input data, including:

[0024] The active learner is divided into two parts: training phase and discrimination phase;

[0025] During the training phase, the company’s financial data and financial analysis reports are split, and the paragraphs of the financial analysis report and the company’s financial data are used as input to train the classifier P(x1, x2). If the company’s financial data can infer the corresponding paragraphs of the company’s report, it is marked as y=1. If no relevant conclusion can be drawn, it is marked as 0. The degree of deviation between the node content and the input original text is determined based on the trained classifier.

[0026] Optionally, if the correlation between the node output content and the input data is less than a threshold, it is necessary to be judged and corrected by financial accountants, and the node content is updated based on the corrected result feedback, and the preset node operation is executed by the controller, including:

[0027] In the discrimination stage, the node results of the large model are input. The correlation between the node output content and the input data is determined by the trained discriminator, and the content with a correlation less than the threshold is selected as highly uncertain content and sent to the financial accounting personnel responsible for the corresponding profession. The financial accounting personnel responsible for the corresponding profession judge and modify the content. The modified content is loaded on the original node as the input content for the next round of model analysis.

[0028] According to another aspect of the present invention, there is also a financial analysis report generation system based on a large model mind map, comprising:

[0029] Determine an input data module, which is used to obtain the enterprise financial data uploaded by the user, and combine it with the external financial data provided by the system as input data. The enterprise financial data includes a balance sheet, a profit and loss statement, and a cash flow statement. The external financial data includes judicial data, industrial and commercial data, industry trends, and policies and regulations;

[0030] Build instruction template module, which is used to break down the production process of financial analysis report into multiple nodes, including generation, repetition, calculation, summary, reinforcement, and build instruction template prompts;

[0031] A node result obtaining module is used to input the input data in json format into the instruction template Prompts for processing, output the node content of each mind map, and obtain the node result;

[0032] A node scoring module is used to summarize the node results to form a financial analysis report that meets the template requirements, score the content of each node, and terminate the operation of nodes with obvious errors;

[0033] The node content correction and update module is used to correct the nodes that need to be corrected and feedback the updated node content through the active learner for the node content that cannot be directly judged as right or wrong, and execute the preset node operation through the controller to determine the total node of each paragraph;

[0034] The enterprise financial analysis report output module is used to merge the summary nodes of each paragraph into the final generation node to form and output the enterprise financial analysis report.

[0035] Optionally, construct a command template module, including:

[0036] A submodule for generating inference results is used to generate inference results on the operation of enterprises based on the judicial data, industrial and commercial data, industry dynamics, and policies and regulations;

[0037] Repeat the previous node submodule to use different calculation methods and different generation methods to repeat the previous node in order to ensure the accuracy of the generated node;

[0038] The calculation result summary submodule is used to break down the calculation steps into simple calculation formulas for calculations between multiple data, and summarize the calculation results after calculations;

[0039] The core submodule of the summary is used to summarize the text content and form a summary and core ideas;

[0040] The enhanced scoring result submodule is used to score some of the generated results, and to enhance the outputs with higher scoring results to enhance the output content of the model.

[0041] Optionally, a node scoring module includes:

[0042] The node content scoring submodule is used to score each node content and judge the accuracy of numerical calculations, language logic, policy references, and conclusion summaries;

[0043] The conflict node termination submodule is used to terminate nodes that have obvious calculation errors, logical errors, inconsistent content, and reference conflicts. The node will no longer participate in the combination and calculation of the mind map.

[0044] Optionally, modify and update the node content module, including:

[0045] The deviation degree judgment submodule is used to judge the deviation degree between the node content and the input data through the active learner for the node content that cannot be directly judged as right or wrong;

[0046] The execution node operation submodule is used to determine and correct the node output content if the correlation between the node output content and the input data is less than the threshold value. The node content is updated based on the feedback of the corrected result, and the preset node operation is executed through the controller to determine the merge and split operation method for the nodes at this layer.

[0047] Optionally, the deviation degree determination submodule includes:

[0048] The active learner is divided into two parts: training phase and discrimination phase;

[0049] The deviation degree judgment submodule is used to split the enterprise financial data and financial analysis report during the training stage. The paragraphs of the financial analysis report and the enterprise financial data are used as input to train the classifier P(x1, x2). If the enterprise financial data can infer the corresponding paragraphs of the enterprise report, it is marked as y=1. If no relevant conclusion can be drawn, it is marked as 0. The deviation degree between the node content and the input original text is judged based on the trained classifier.

[0050] Optionally, the execution node operation submodule includes:

[0051] The node content modification and update submodule is used to input the node results of the large model in the discrimination stage. The correlation between the node output content and the input data is determined according to the trained discriminator, and the content with a correlation less than the threshold is selected as highly uncertain content and sent to the financial accounting personnel responsible for the corresponding profession. The financial accounting personnel responsible for the corresponding profession judge and modify the content. The modified content is loaded on the original node as the input content for the next round of model analysis.

[0052] Therefore, starting from the financial analysis report writing work in the financial and taxation scenarios, combining the big model capabilities and all the internal and external financial and taxation data of the enterprise, completing the big model logical reasoning, and using the structure of the mind map to enhance the output accuracy and professionalism of the model. By introducing active learning, the manual correction in the output process is enhanced, which can be better applied to scenarios with high requirements for financial reports, urgent work tasks, and difficulties in collecting external financial knowledge. At the same time, it can further help the company's financial accounting personnel to improve work efficiency, improve the professionalism and comprehensiveness of reports, and realize the intelligence of financial work and the digitalization of corporate finance. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] A more complete understanding of exemplary embodiments of the present invention may be obtained by referring to the following drawings:

[0054] Figure 1 A schematic diagram of the process of a method for generating a financial analysis report based on a large model mind map according to this embodiment;

[0055] Figure 2 A schematic diagram of a method for generating a financial analysis report based on a large model mind map according to this embodiment;

[0056] Figure 3 This is an example diagram of prompts for a financial analysis report based on a mind map as described in this embodiment;

[0057] Figure 4 This is a schematic diagram of a financial analysis report generation system based on a large model mind map described in this embodiment. DETAILED DESCRIPTION

[0058] Now, exemplary embodiments of the present invention are described with reference to the accompanying drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. These embodiments are provided to disclose the present invention in detail and completely and to fully convey the scope of the present invention to those skilled in the art. The terms used in the exemplary embodiments shown in the accompanying drawings are not intended to limit the present invention. In the accompanying drawings, the same units / elements are marked with the same reference numerals.

[0059] Unless otherwise specified, the terms (including technical terms) used herein have the commonly understood meanings to those skilled in the art. In addition, it is understood that the terms defined in commonly used dictionaries should be understood to have the same meanings as those in the context of the relevant fields, and should not be understood as idealized or overly formal meanings.

[0060] According to a first aspect of the present invention, a method 100 for generating a financial analysis report based on a large model mind map is provided. Figure 1 As shown, the method 100 includes:

[0061] S101: Acquire the enterprise financial data uploaded by the user and combine it with the external financial data provided by the system as input data. The enterprise financial data includes a balance sheet, a profit and loss statement, and a cash flow statement. The external financial data includes judicial data, industrial and commercial data, industry trends, and policies and regulations.

[0062] S102: Break down the production process of financial analysis report into multiple nodes, including generation, repetition, calculation, summary, and reinforcement, and construct instruction template prompts;

[0063] S103: input the input data in json format into the instruction template Prompts for processing, output the node content of each mind map, and obtain the node result;

[0064] S104: Summarize the node results to form a financial analysis report that meets the template requirements, score the content of each node, and terminate the operation for nodes with obvious errors;

[0065] S105: For node contents that cannot be directly judged as right or wrong, the active learner is used to correct the nodes that need to be corrected and feedback is given to update the node contents. The controller is used to execute preset node operations to determine the total node of each paragraph. S106: The summary nodes of each paragraph are merged into the final generated node to form and output the enterprise financial analysis report.

[0066] Specifically, refer to Figure 2As shown, first collect the company's financial data, including the balance sheet, income statement, cash flow statement, judicial data, industrial and commercial data, industry dynamics, policies and regulations of the last three periods (year / quarter / month), construct a prompt project based on the mind map, and divide the input into input data, prompts and expert feedback. After the user uploads the company's own relevant financial information, the system combines the company information, industry dynamics, judicial and industrial and commercial data as common input data. The large model extracts, combines, calculates, generates and scores information based on prompts, and feeds back the intermediate results to the user based on the active learning results. The user participates in the adjustment of the intermediate results and feeds back to the model the accuracy of the scoring. The model repeats multiple rounds of the process based on the logical structure of the directed acyclic graph, and finally outputs a complete financial analysis report.

[0067] The present invention is divided into three core stages, namely, construction of instruction template prompts, large model analysis and active learning control.

[0068] Instruction Template Prompts Building Module

[0069] The large model can output possible answers with probability based on user needs and the background knowledge of its own model training. However, since the essential structure of the model is a Transformer-Decoder structure, although the model saves a lot of background knowledge, the model cannot understand the user's intention very well, and therefore cannot output professional and accurate results. The present invention proposes to use the form of instruction templates, use input as a prompt, simulate the way humans think, and break down the process of making financial analysis reports into generation, repetition, calculation, summary, and refinement. The generation of any financial analysis report requires the combination and decomposition of these processes for the original information and intermediate results. Examples of Prompts based on mind maps are as follows: Figure 3 shown.

[0070] Each node represents the processing of a certain data, such as calculating sales gross margin, calculating return on net assets, summarizing cash flow, generating industry analysis reports, or calculating a certain result in multiple ways. Each edge represents a new operation derived from an old node, such as splitting a calculation node into multiple scattered simple calculation operations, or converging operations between multiple nodes, such as summarizing the calculation results and generated results. Specifically:

[0071] Generate: The big model generates inference results about the business operation status based on data, industry conditions, and judicial and industrial and commercial conditions.

[0072] Repeat: Repeat the previous node, using different calculation methods, different generation methods, etc., to ensure the accuracy of the generated node (similar to voting).

[0073] Compute: For calculations between multiple data, since it is difficult for large models to directly calculate complex financial analysis data, a mind map is used to decompose the calculation steps into simple calculation formulas, and then summarize the calculation results. For example, (a+b)*(c+d)=e is decomposed into (a+b)=a1, (c+d)=b1, and a1*b1=e.

[0074] Summary: A summary of the text content to form an abstract and core ideas.

[0075] Refine: Some generated results will be scored by the scoring model, and outputs with higher scores will be refined to enhance the output content of the model.

[0076] In addition, some nodes will be identified by themselves or by the scoring module and active learning control module and closed during the model execution due to calculation errors or conclusion errors.

[0077] Finally, the enterprise data and external data are combined and input into prompts in json format to form the final model input data. The following is the prompts for generating financial analysis reports based on mind maps constructed by the present invention:

[0078] Initial / system prompt

[0079] Hello, I would like to generate a financial analysis report based on the company's financial statements: {Input data}

[0080] (Note: The data is in JSON format. The balance sheet, income statement, and cash flow statement are historical data for three periods.)

[0081] Please input the format as follows: {{companyName}} Financial Analysis Report

[0082] 1. Comprehensive financial evaluation

[0083]

Company Type

[0084]

Business Scope

[0085] {{summary}}: Example: In June 2024, the company's operating and growth capabilities are poor, but its cash flow is excellent, its debt repayment ability is good, and its asset quality and profitability are acceptable.

[0086] 2. Profitability

[0087] {{integrated}}

[0088] 2.1. Profitability

[0089] Key indicators of profitability

[0090] ●Trend of change

[0091] {{bdqs}}

[0092] Abnormal prompt

[0093] {{ycts}}

[0094] 2.2. Profitability indicators

[0095] 2.2.1. Gross profit margin

[0096] Gross profit margin refers to the ratio of the sales revenue of the enterprise in the current period after deducting the operating costs to the sales revenue, which can reflect the profitability of the commodity production and operation that constitutes the main business of the enterprise. Generally speaking, the higher the gross profit margin, the stronger the ability of the enterprise to offset various cost expenditures and the stronger its profitability, and vice versa.

[0097] ●Enterprise performance

[0098] {{xsmllqybx}}

[0099] ●Trend of change

[0100] {{qybxbdqs}}

[0101] 2.2.2.EBIT profit margin

[0102] EBIT profit margin refers to the ratio between the profit before deducting income tax and interest expenses and the operating income of the enterprise. This ratio can exclude the impact of differences in capital structure and income tax policies on profits. Generally speaking, the higher the ratio, the more profit the enterprise generates and the stronger its ability to provide cash flow for debt repayment and continuous operation, and vice versa.

[0103] ●Enterprise performance

[0104] {{ebttqybx}}

[0105] ●Trend of change

[0106] {{ebttbdqs}}

[0107] 2.2.3. Return on Equity

[0108] Return on equity is the ratio of a company's net profit to its average net assets. It is an important indicator for measuring a company's profitability, reflecting the level of compensation for owners' equity and the ability of its own capital to obtain net income. Generally speaking, the higher the index, the higher the return on investment, and vice versa, the weaker the profitability of the company's equity owners.

[0109] ●Enterprise performance

[0110] {{jzcsylqybx}}

[0111] ●Trend of change

[0112] {{jzcsylbdqs}}

[0113] 2.2.4. Cost-profit ratio

[0114] The cost-to-profit ratio is the ratio of the total profit of an enterprise in a certain period to the total cost, reflecting how much profit the enterprise can obtain for paying one unit of cost, that is, the ability of the enterprise to obtain profit. Generally speaking, the larger the ratio, the greater the return rate of the enterprise for paying unit cost, and the stronger the profitability of the enterprise's business operations.

[0115] ●Enterprise performance

[0116] {{cbfyqybx}}

[0117] ●Trend of change

[0118] {{cbfybdqs}}

[0119] 2.2.5. Earnings to Cash Ratio

[0120] The profit cash ratio refers to the ratio between the net cash flow generated by the company's operating activities and the net profit in a certain period, reflecting how much of the realized net profit is guaranteed by cash. Generally speaking, the larger the ratio, the better the company's profit quality.

[0121] ●Enterprise performance

[0122] {{ylxjqybx}}

[0123] ●Trend of change

[0124] {{ylxjbdqs}}

[0125] 3. Cash Flow

[0126] {{xjllgs}}

[0127] 3.1. Cash flow

[0128] Cash flow structure

[0129] ●Trend of change

[0130] {{xjllbdqs}}

[0131] Abnormal prompt

[0132] {{xjllycts}}

[0133] 3.2. Cash flow indicators

[0134] 3.2.1. Cash flow from operating activities

[0135] Cash flow from operating activities: reflects the cash inflow and outflow generated by the company's operating activities. A positive number indicates a net inflow, indicating that the company is in good operating condition; a negative number indicates a net outflow, and attention should be paid to the cash flow from operating activities. By accurately understanding the changing trend of cash flow, companies can adjust their operating strategies in a timely manner, provide a reliable basis for financing decisions, and effectively avoid capital chain breaks and reduce investment risks.

[0136] ●Enterprise performance

[0137] {{jyhdxjlqybx}}

[0138] ●Trend of change

[0139] {{jyhdxjlbdqs}}

[0140] 3.2.2. Cash flow from investing activities

[0141] Cash flow from investing activities reflects the cash inflow and outflow from a company's investing activities. A positive number indicates a good investment recovery or return on investment; a negative number indicates that the company is making expansionary investments.

[0142] ●Enterprise performance

[0143] {{tzhdxjlqybx}}

[0144] ●Trend of change

[0145] {{tzhdxjlbdqs}}

[0146] 3.2.3. Cash flow from financing activities

[0147] Cash flow from financing activities reflects the cash inflow and outflow obtained by the company through financing activities (such as issuing stocks and bonds, etc.). A positive number means that the company has obtained funds through financing; a negative number means that the company has repaid debts or paid dividends, etc.

[0148] ●Enterprise performance

[0149] {{czhdxjlqybx}}

[0150] ●Trend of change

[0151] {{czhdxjlbdqs}}

[0152] A prompt used by generate:

[0153] <instruction>

[0154] Input the company's EBIT information and return on net assets information, compare the company's market income, only reflect the change in income, and analyze the reasons for the change in income compared with previous periods. Express in text form.

[0155] < / instruction>

[0156] <example>

[0157] Input: The company's EBIT profit margin in the previous period was 0.0368, the profit margin in the previous period was 0.0366, and the profit margin in this period was 0.0287. The return on net assets in the previous period was 1.3216, the previous period was 1.5432, and the current period was 2.0735.

[0158] Output: Compared with the previous period, the company's EBIT profit margin and return on net assets have both declined. Among them, the EBIT profit margin has dropped from 0.0368 in the previous period to 0.0287 in the current period, and the return on net assets has dropped from 1.3261 in the previous period to 2.0735 in the current period. This may indicate that the company's profitability has weakened in the short term. However, the specific reasons and impacts need to be comprehensively analyzed in combination with the company's operating conditions, market environment and other factors.

[0159] < / example>

[0160] A prompt used by Repeat:

[0161] <instruction>

[0162] Sort the input data by size and output the average value based on the size sorting results.

[0163] < / instruction>

[0164] <approach>

[0165] 1. Split the input data into two random parts and sort them by size.

[0166] 2. Merge two arrays, compare the numbers in the merged array with the merged array one by one and put them in the appropriate position according to the size, and repeat the above process

[0167] < / approach>

[0168] <example>

[0169] Input:1,3,5,7,9,2,4,6,8,10

[0170] Sorted:[1,2,3,7,9],[2,4,6,8,10]

[0171] Outputs:[1,2,3,4,5,6,7,8,9,10]

[0172] < / example>

[0173] A prompt used by compute:

[0174] <instruction>

[0175] You are an efficient numerical calculation assistant who can handle various mathematical operations and logical problems. In the following conversation, I will ask you a series of mathematical problems and expect you to give accurate answers and brief explanations to calculate the input data. You need to break down the calculation formula into basic addition, subtraction, multiplication and division formulas and calculate them step by step, and finally combine the calculation results.

[0176] < / instruction>

[0177] <example>

[0178] Input:(3.12+5.12)*(2.45+4.67)=

[0179] Output:3.12+5.12=8.24

[0180] 2.45+4.67=7.12

[0181] 8.24*7.12=58.6688

[0182] < / example>

[0183] A prompt used by summary:

[0184] <instruction>

[0185] Please read the following information and generate a concise summary for it, combining industry information and judicial and industrial information to form financial-related conclusions.

[0186] < / instruction>

[0187] <example>

[0188] Input:

[0189] In terms of policy: In 2024, the state issued an announcement by the Ministry of Industry and Information Technology of the People's Republic of China, the Ministry of Finance, and the State Administration of Taxation on adjusting the technical requirements for energy-saving new energy vehicles that enjoy vehicle and vessel tax preferences. New energy vehicles can enjoy tax benefits. Driven by favorable policies, preferential promotions by car companies, and the launch of many popular models, purchasing a new energy vehicle during the Dragon Boat Festival holiday has become a new consumption choice for many citizens.

[0190] In terms of business operations: the cash flow of the company was 761477601.84 in the previous period, 861477601.84 in the previous period, and 1861477601.84 in this period. The cash flow ratio from financing activities was 1.3216 in the previous period, 1.5432 in the previous period, and 2.0735 in this period.

[0191] Outpu: The net cash flow generated by the company's operating activities showed an upward trend, while the net cash flow generated by investment activities and financing activities showed a downward trend. Specifically, the net cash flow generated by operating activities in this period increased by 761,477,601.84 yuan compared with the previous period, the net cash flow generated by investment activities decreased by 1361,130,236.89 yuan, and the net cash flow generated by financing activities decreased by 223,697,6224.75 yuan. The main reason for this trend is that with the introduction of the "New Energy Vehicle Preferential Policy" this year, consumers are more willing to consume environmentally friendly products, and companies are adjusting their cash flow management strategies to reduce reliance on foreign investment and financing and increase sales from operating activities.

[0192] < / example>

[0193] A prompt used by refine:

[0194] <instruction>

[0195] Please recalculate the future trend of corporate income tax based on the following tax incentives. Tax rate calculation method:

[0196] (Output tax - input tax - retained tax refund amount from the previous period) * tax rate = tax amount.

[0197] < / instruction>

[0198] <example>

[0199] Input: For small-scale VAT taxpayers, the taxable sales income subject to a 3% tax rate will be subject to a reduced VAT rate of 1%; for prepaid VAT items subject to a 3% prepayment rate, the prepayment rate will be reduced to 1%. The above preferential policies will expire on December 31, 2027.

[0200] The company's monthly sales volume this period is 80,000 yuan, the monthly input tax is 55,000 yuan, the tax refund from the previous period is 4,000 yuan, and the tax amount this month is (80,000-55,000-4,000)*0.03=630 yuan.

[0201] Output: (80000-55000-4000)*0.03-(80000-55000-4000)*0.01=420. The company will pay 420 yuan less in tax every month.

[0202] < / example>

[0203] Large model analysis module

[0204] The large model training and analysis module is the core module for model content output, which consists of two parts: analysis model and scoring model. Since this patent adopts the mind map method, it is necessary to score and verify the output results of each layer of the map. By testing the large model parameter scale above 70B, the financial and tax background information can be better recorded, so as to better follow the prompts instructions and output more accurate and professional data and calculation results.

[0205] The main function of the analysis model is to follow the prompts to output the node content of each mind map, and to summarize the node results to form a financial analysis report that meets the template requirements. The scoring model is a large language model of the same scale. The scoring model is used to score the content of each node, judge the accuracy of numerical calculations, language logic, policy references, and conclusion summaries, and terminate the operation for nodes with obvious calculation errors, logical errors, inconsistent content, and reference conflicts. The node will no longer participate in the combination and calculation of the mind map.

[0206] Active Learning Control Module

[0207] The active learning control module is the core module of this patent, which consists of a Bayesian-based active learner and a controller.

[0208] The active learner is a traditional Bayesian statistical method of active learning framework learning strategy. It combines sample information, uses Bayesian theorem to find the posterior distribution, and then estimates the overall distribution. It uses probability to represent all forms of uncertainty, and uses probability rules to implement the learning and reasoning process. For node content that the scoring model cannot directly judge right or wrong, the active learner is used to judge the degree of deviation between the node content and the input data. If the deviation is too high, it is considered that the output content of the node has a high uncertainty or high information entropy. It needs to be judged and corrected by financial accountants, and the node content is updated based on the corrected results.

[0209] The active learner is divided into two parts: the training stage and the discrimination stage. In the training stage, the historical financial data and financial analysis report of the enterprise are split, and the paragraphs of the financial analysis report and the financial data of the enterprise are used as input to train the classifier P(x1, x2). If the enterprise financial data can infer the corresponding paragraphs of the enterprise report, it is marked as y=1, and if it is weak and cannot draw relevant conclusions, it is marked as 0. The trained classifier can better judge the degree of deviation between the node content and the input original text. In the discrimination stage, the node output of the large model is input. In order to save time, this patent only inputs the summary and generate nodes. According to the trained discriminator, the correlation between the output content of the Summary and Generate nodes and the input data is judged, and the correlation is less than the threshold (the threshold of this patent is 0.2) as highly uncertain content is sent to the financial accounting personnel responsible for the corresponding profession. The accounting personnel responsible for the corresponding profession judge and modify the content. The modified content loads the original node as the input content for the next round of model analysis.

[0210] The controller is the core control part of each operation of the financial analysis report. The controller executes the preset node operation, thereby determining the operation mode of merging and splitting the nodes of this layer. The more node levels there are, the longer the total model analysis time is. After experiments, this patent uses a 7-layer analysis method to complete the generation of financial analysis reports. The first layer consists of compute and generate. The number of compute is the number of calculation formulas in the template, and generate is the number of each title. The 2nd to 5th layers are composed of compute, generate, repeat, refine, and summarize. The final quantitative method adopted by this patent is that the number of compute is the number of upper-level formulas after splitting. If it is not split, no new number is generated. The number of generate is the number of paragraphs in the document. The number of summary is the number of generated nodes and compute nodes in the previous level in the document. The number of repeat is the formula calculation repeat = 1, the text class ratio is set to 20% of the number of nodes, and the number of refine is the number of high scores in the evaluation model. The sixth layer consists of 5 summaries, which generate the comprehensive evaluation, profitability, profitability indicators, cash flow conditions, and cash flow indicators in the financial analysis report. The seventh layer is the final report output layer, which consists of 1 aggregate.

[0211] Report output module

[0212] After being controlled by the multi-layer thinking tree of the active learning control module, the summary of each paragraph is finally merged into the final aggregate node to form a corporate financial analysis report.

[0213] Therefore, starting from the financial analysis report writing work in the financial and taxation scenarios, combining the big model capabilities and all the internal and external financial and taxation data of the enterprise, completing the big model logical reasoning, and using the structure of the mind map to enhance the output accuracy and professionalism of the model. By introducing active learning, the manual correction in the output process is enhanced, which can be better applied to scenarios with high requirements for financial reports, urgent work tasks, and difficulties in collecting external financial knowledge. At the same time, it can further help the company's financial accounting personnel to improve work efficiency, improve the professionalism and comprehensiveness of reports, and realize the intelligence of financial work and the digitalization of corporate finance.

[0214] Optionally, the process of making a financial analysis report is broken down into multiple nodes, including generation, repetition, calculation, summary, and reinforcement, and instruction template prompts are constructed, including:

[0215] Generate inference results on the business operation status of the enterprise based on the judicial data, industrial and commercial data, industry dynamics, and policies and regulations;

[0216] Use different calculation methods and different generation methods to repeat the previous node to ensure the accuracy of the generated node;

[0217] For calculations between multiple data, the calculation steps are broken down into simple calculation formulas, and the calculation results are summarized after the calculations are performed;

[0218] Summarize the text content to form an abstract and core ideas;

[0219] Score some of the generated results, and use reinforcement methods for outputs with higher scores to enhance the output content of the model.

[0220] Optionally, the node results are summarized to form a financial analysis report that meets the template requirements, each node content is scored, and the operation is terminated for nodes with obvious errors, including:

[0221] Score each node content to judge the accuracy of numerical calculations, language logic, policy references, and conclusions;

[0222] For nodes with obvious calculation errors, logical errors, inconsistent content, and reference conflicts, the termination operation will be performed and the node will no longer participate in the combination and calculation of the mind map.

[0223] Optionally, for node contents that cannot be directly judged as right or wrong, the active learner is used to correct the nodes that need to be corrected and feedback is provided to update the node contents, and the controller is used to execute preset node operations to determine the total node of each paragraph, including:

[0224] For node contents that cannot be directly judged as right or wrong, the active learner is used to determine the degree of deviation between the node content and the input data;

[0225] If the correlation between the node output content and the input data is less than the threshold, it is necessary for financial accountants to judge and correct it, and update the node content based on the corrected result feedback. The controller will execute the preset node operation and decide how to merge or split the nodes at this layer.

[0226] Optionally, for node contents that cannot be directly judged as right or wrong, an active learner is used to judge the degree of deviation between the node contents and the input data, including:

[0227] The active learner is divided into two parts: training phase and discrimination phase;

[0228] During the training phase, the company’s financial data and financial analysis reports are split, and the paragraphs of the financial analysis report and the company’s financial data are used as input to train the classifier P(x1, x2). If the company’s financial data can infer the corresponding paragraphs of the company’s report, it is marked as y=1. If no relevant conclusion can be drawn, it is marked as 0. The degree of deviation between the node content and the input original text is determined based on the trained classifier.

[0229] Optionally, if the correlation between the node output content and the input data is less than a threshold, it is necessary to be judged and corrected by financial accountants, and the node content is updated based on the corrected result feedback, and the preset node operation is executed by the controller, including:

[0230] In the discrimination stage, the node results of the large model are input. The correlation between the node output content and the input data is determined by the trained discriminator, and the content with a correlation less than the threshold is selected as highly uncertain content and sent to the financial accounting personnel responsible for the corresponding profession. The financial accounting personnel responsible for the corresponding profession judge and modify the content. The modified content is loaded on the original node as the input content for the next round of model analysis.

[0231] According to another aspect of the present invention, a financial analysis report generation system 400 based on a large model mind map is also included. Figure 4 As shown, the system 400 includes:

[0232] Determine input data module 410, used to obtain enterprise financial data uploaded by users, combined with external financial data provided by the system as input data, the enterprise financial data includes balance sheet, profit and loss statement and cash flow statement, the external financial data includes judicial data, industrial and commercial data, industry dynamics and policies and regulations;

[0233] The instruction template construction module 420 is used to decompose the production process of the financial analysis report into multiple nodes, including generation, repetition, calculation, summary, and reinforcement, and to construct instruction template prompts;

[0234] The node result obtaining module 430 is used to input the input data in the json format into the instruction template Prompts for processing, output the node content of each mind map, and obtain the node result;

[0235] The node scoring module 440 is used to summarize the node results to form a financial analysis report that meets the template requirements, score the content of each node, and terminate the operation for nodes with obvious errors;

[0236] The node content correction and update module 450 is used to correct the node content that cannot be directly judged as right or wrong, and to feedback and update the node content through the active learner, and to execute the preset node operation through the controller to determine the total node of each paragraph;

[0237] The enterprise financial analysis report output module 460 is used to merge the summary nodes of each paragraph into a final generation node to form an enterprise financial analysis report and output it.

[0238] Optionally, construct a command template module, including:

[0239] A submodule for generating inference results is used to generate inference results on the operation of enterprises based on the judicial data, industrial and commercial data, industry dynamics, and policies and regulations;

[0240] Repeat the previous node submodule to use different calculation methods and different generation methods to repeat the previous node in order to ensure the accuracy of the generated node;

[0241] The calculation result summary submodule is used to break down the calculation steps into simple calculation formulas for calculations between multiple data, and summarize the calculation results after calculations;

[0242] The core submodule of the summary is used to summarize the text content and form a summary and core ideas;

[0243] The enhanced scoring result submodule is used to score some of the generated results, and to enhance the outputs with higher scoring results to enhance the output content of the model.

[0244] Optionally, a node scoring module includes:

[0245] The node content scoring submodule is used to score each node content and judge the accuracy of numerical calculations, language logic, policy references, and conclusion summaries;

[0246] The conflict node termination submodule is used to terminate nodes that have obvious calculation errors, logical errors, inconsistent content, and reference conflicts. The node will no longer participate in the combination and calculation of the mind map.

[0247] Optionally, modify and update the node content module, including:

[0248] The deviation degree judgment submodule is used to judge the deviation degree between the node content and the input data through the active learner for the node content that cannot be directly judged as right or wrong;

[0249] The execution node operation submodule is used to determine and correct the node output content if the correlation between the node output content and the input data is less than the threshold value. The node content is updated based on the feedback of the corrected result, and the preset node operation is executed through the controller to determine the merge and split operation method for the nodes at this layer.

[0250] Optionally, the deviation degree determination submodule includes: the active learner is divided into two parts: a training phase and a determination phase;

[0251] The deviation degree judgment submodule is used to split the enterprise financial data and financial analysis report during the training stage. The paragraphs of the financial analysis report and the enterprise financial data are used as input to train the classifier P(x1, x2). If the enterprise financial data can infer the corresponding paragraphs of the enterprise report, it is marked as y=1. If no relevant conclusion can be drawn, it is marked as 0. The deviation degree between the node content and the input original text is judged based on the trained classifier.

[0252] Optionally, modify and update the node content module, including:

[0253] The node content modification and update submodule is used to input the node results of the large model in the discrimination stage. The correlation between the node output content and the input data is determined according to the trained discriminator, and the content with a correlation less than the threshold is selected as highly uncertain content and sent to the financial accounting personnel responsible for the corresponding profession. The financial accounting personnel responsible for the corresponding profession judge and modify the content. The modified content is loaded on the original node as the input content for the next round of model analysis.

[0254] A financial analysis report generation system 400 based on a large model mind map according to an embodiment of the present invention corresponds to a financial analysis report generation method 100 based on a large model mind map according to another embodiment of the present invention, and will not be described in detail here.

[0255] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of complete hardware embodiments, complete software embodiments, or embodiments in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The scheme in the embodiments of the present application can be implemented in various computer languages, for example, object-oriented programming language Java and literal scripting language JavaScript, etc.

[0256] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0257] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0258] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0259] Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0260] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A method for generating a financial analysis report based on a large model mind map, characterized in that: include: Obtain the enterprise financial data uploaded by the user and combine it with the external financial data provided by the system as input data. The enterprise financial data includes balance sheet, profit and loss statement and cash flow statement. The external financial data includes judicial data, industrial and commercial data, industry dynamics and policies and regulations. Break down the process of making financial analysis reports into multiple nodes, including generation, repetition, calculation, summary, and reinforcement, and construct instruction template prompts; Input the input data in json format into the instruction template Prompts for processing, output the node content of each mind map, and obtain the node result; Summarize the node results to form a financial analysis report that meets the template requirements, score the content of each node, and terminate the operation for nodes with obvious errors; For node contents that cannot be directly judged as right or wrong, the active learner is used to correct the nodes that need to be corrected and feedback is given to update the node contents. The controller is used to execute the preset node operations to determine the total node of each paragraph. Merge the summary nodes of each paragraph into the final generated node to form and output the enterprise financial analysis report.

2. The method according to claim 1, characterized in that The production process of financial analysis reports is broken down into multiple nodes, including generation, repetition, calculation, summary, and reinforcement, and instruction template prompts are constructed, including: Generate inference results on the business operation status of the enterprise based on the judicial data, industrial and commercial data, industry dynamics, and policies and regulations; Use different calculation methods and different generation methods to repeat the previous node to ensure the accuracy of the generated node; For calculations between multiple data, the calculation steps are broken down into simple calculation formulas, and the calculation results are summarized after the calculations are performed; Summarize the text content to form an abstract and core ideas; Score some of the generated results, and use reinforcement methods for outputs with higher scores to enhance the output content of the model.

3. The method according to claim 1, characterized in that The node results are summarized to form a financial analysis report that meets the template requirements, and each node content is scored. Operations are terminated for nodes with obvious errors, including: Score each node content to judge the accuracy of numerical calculations, language logic, policy references, and conclusions; For nodes with obvious calculation errors, logical errors, inconsistent content, and reference conflicts, the termination operation will be performed and the node will no longer participate in the combination and calculation of the mind map.

4. The method according to claim 1, characterized in that For node contents that cannot be directly judged as right or wrong, the active learner is used to correct the nodes that need to be corrected and feedback is given to update the node contents. The controller is used to execute the preset node operations to determine the total node of each paragraph, including: For node contents that cannot be directly judged as right or wrong, the active learner is used to determine the degree of deviation between the node content and the input data; If the correlation between the node output content and the input data is less than the threshold, it is necessary for financial accountants to judge and correct it, and update the node content based on the corrected result feedback. The controller will execute the preset node operation and decide how to merge or split the nodes at this layer.

5. The method according to claim 4, characterized in that For node contents that cannot be directly judged as right or wrong, the active learner is used to determine the degree of deviation between the node content and the input data, including: The active learner is divided into two parts: training phase and discrimination phase; During the training phase, the company’s financial data and financial analysis reports are split, and the paragraphs of the financial analysis report and the company’s financial data are used as input to train the classifier P(x1, x2). If the company’s financial data can infer the corresponding paragraphs of the company’s report, it is marked as y=1. If no relevant conclusion can be drawn, it is marked as 0. The degree of deviation between the node content and the input original text is determined based on the trained classifier.

6. The method according to claim 5, characterized in that If the correlation between the node output content and the input data is less than the threshold, it is necessary for the financial accountant to make a judgment and correction, update the node content based on the corrected result feedback, and execute the preset node operation through the controller, including: In the discrimination stage, the node results of the large model are input. The correlation between the node output content and the input data is determined by the trained discriminator, and the content with a correlation less than the threshold is selected as highly uncertain content and sent to the financial accounting personnel responsible for the corresponding profession. The financial accounting personnel responsible for the corresponding profession judge and modify the content. The modified content is loaded on the original node as the input content for the next round of model analysis.

7. A financial analysis report generation system based on a large model mind map, characterized in that: include: Determine an input data module, which is used to obtain the enterprise financial data uploaded by the user, and combine it with the external financial data provided by the system as input data. The enterprise financial data includes a balance sheet, a profit and loss statement, and a cash flow statement. The external financial data includes judicial data, industrial and commercial data, industry trends, and policies and regulations; Build instruction template module, which is used to break down the production process of financial analysis report into multiple nodes, including generation, repetition, calculation, summary, reinforcement, and build instruction template prompts; A node result obtaining module is used to input the input data in json format into the instruction template Prompts for processing, output the node content of each mind map, and obtain the node result; A node scoring module is used to summarize the node results to form a financial analysis report that meets the template requirements, score the content of each node, and terminate the operation of nodes with obvious errors; The node content correction and update module is used to correct the node content that cannot be directly judged as right or wrong through the active learner, and feedback the updated node content, and execute the preset node operation through the controller to determine the total node of each paragraph; The enterprise financial analysis report output module is used to merge the summary nodes of each paragraph into the final generation node to form and output the enterprise financial analysis report.

8. The system according to claim 7, characterized in that Build the instruction template module, including: A submodule for generating inference results is used to generate inference results on the operation of enterprises based on the judicial data, industrial and commercial data, industry dynamics, and policies and regulations; Repeat the previous node submodule to use different calculation methods and different generation methods to repeat the previous node in order to ensure the accuracy of the generated node; The calculation result summary submodule is used to break down the calculation steps into simple calculation formulas for calculations between multiple data, and summarize the calculation results after calculations; The core submodule of the summary is used to summarize the text content and form a summary and core ideas; The enhanced scoring result submodule is used to score some of the generated results, and to enhance the outputs with higher scoring results to enhance the output content of the model.

9. The system according to claim 7, characterized in that The node scoring module includes: The node content scoring submodule is used to score each node content and judge the accuracy of numerical calculations, language logic, policy references, and conclusion summaries; The conflict node termination submodule is used to terminate nodes that have obvious calculation errors, logical errors, inconsistent content, and reference conflicts. The node will no longer participate in the combination and calculation of the mind map.

10. The system according to claim 7, characterized in that Modify and update the node content module, including: The deviation degree judgment submodule is used to judge the deviation degree between the node content and the input data through the active learner for the node content that cannot be directly judged as right or wrong; The execution node operation submodule is used to determine and correct the node output content if the correlation between the node output content and the input data is less than the threshold value. The node content is updated based on the feedback of the corrected result, and the preset node operation is executed through the controller to determine the merge and split operation method for the nodes at this layer.

11. The system according to claim 10, characterized in that The submodule for judging the degree of deviation includes: The active learner is divided into two parts: training phase and discrimination phase; The deviation degree judgment submodule is used to split the enterprise financial data and financial analysis report during the training stage. The paragraphs of the financial analysis report and the enterprise financial data are used as input to train the classifier P(x1, x2). If the enterprise financial data can infer the corresponding paragraphs of the enterprise report, it is marked as y=1. If no relevant conclusion can be drawn, it is marked as 0. The deviation degree between the node content and the input original text is judged based on the trained classifier.

12. The system according to claim 11, characterized in that Execute node operation submodule, including: The node content modification and update submodule is used to input the node results of the large model in the discrimination stage. The correlation between the node output content and the input data is determined according to the trained discriminator, and the content with a correlation less than the threshold is selected as highly uncertain content and sent to the financial accounting personnel responsible for the corresponding profession. The financial accounting personnel responsible for the corresponding profession judge and modify the content. The modified content is loaded on the original node as the input content for the next round of model analysis.

Images