Financial disclosure document auditing method and device based on large model

Abstract

The application relates to the technical field of finance, and discloses a financial disclosure document auditing method and device based on a large model, which are applied to an intelligent auditing system. The method comprises the following steps: obtaining a financial document to be audited and auditing rule information of at least one disclosure problem; decomposing each problem and the auditing rule information to obtain a sub-problem of the problem and a rule thereof; based on a retriever, searching for a matched text segment in the financial document according to the rule of each sub-problem; when the decomposition mode is a target decomposition mode, inputting the rule of the sub-problem and the text segment into a large language model for auditing; when the decomposition mode is not the target decomposition mode, judging whether the sub-problem satisfies the condition of calling the large language model according to the rule of the sub-problem; auditing the text segment according to the judgment result and the rule; and integrating answers of each problem to obtain a target answer. It can be seen that the application can improve the auditing accuracy and compliance of whether the financial document has a disclosure problem.

Description

Technical Field

[0001] This invention relates to the field of financial technology, and in particular to a method and apparatus for reviewing financial information disclosure documents based on a large model. Background Technology

[0002] With the rapid development and widespread adoption of natural language processing and artificial intelligence technologies, researchers have begun to explore these technologies for compliance auditing of financial documents. In particular, the flourishing development of Large Language Models (LLMs) has spurred the creation of several large models specifically for the financial sector, leading to numerous applications of these models in financial operations and regulation.

[0003] However, practice has shown that large language models inevitably suffer from illusions when handling real-world business scenarios, leading to high uncertainty in model generation and impacting normal business processes. For example, using large language models to verify whether a financial document contains disclosure issues can result in low accuracy and compliance. Therefore, proposing a technical solution to improve the accuracy and compliance of verifying whether financial documents contain disclosure issues is particularly important. Summary of the Invention

[0004] This invention provides a method and apparatus for reviewing financial information disclosure documents based on a large model, which can improve the accuracy and compliance of reviewing whether financial documents have information disclosure problems.

[0005] To address the aforementioned technical problems, the first aspect of this invention discloses a method for reviewing financial information disclosure documents based on a large model. This method is applied to an intelligent review system and includes:

[0006] The intelligent review system obtains financial documents uploaded by users for review and review rule information for one or more preset information disclosure questions for the financial documents. Each information disclosure question corresponds to a question type.

[0007] The intelligent review system decomposes each type of information disclosure issue and the review rule information of that type of information disclosure issue according to the determined decomposition method, and obtains a set of sub-issues of that type of information disclosure issue and sub-review rule information of each sub-issue in the set of sub-issues. Each sub-issue has a corresponding retrieval device.

[0008] For each of the sub-problems, the intelligent review system, based on the retrieval device and according to the sub-review rule information of the sub-problem, retrieves the text fragment that matches the sub-problem from the financial document;

[0009] When the decomposition method is the target decomposition method, the intelligent review system inputs the sub-review rule information of the sub-problem and its matching text fragment into the corresponding large language model for review, and obtains the answer to the sub-problem. The target decomposition method is the method of decomposing using the large language model.

[0010] When the decomposition method is not the target decomposition method, the intelligent review system determines whether the sub-problem meets the target review condition of calling the corresponding large language model based on the sub-review rule information of the sub-problem, and obtains the judgment result. Each of the problem types has a corresponding large language model.

[0011] The intelligent review system reviews the text fragments that match the sub-question based on the judgment result and the sub-review rule information of the sub-question, and obtains the answer to the sub-question.

[0012] The intelligent review system integrates the answers to all sub-questions of each disclosure issue to obtain the target answer for each disclosure issue. The target answer includes whether the financial document has the corresponding disclosure issue or whether the financial document does not have the corresponding disclosure issue.

[0013] As an optional implementation, in the first aspect of the present invention, for each of the sub-problems, the intelligent review system, based on the retrieval device and according to the sub-review rule information of the sub-problem, retrieves a text fragment matching the sub-problem from the financial document, including:

[0014] The intelligent review system, based on the retrieval device corresponding to the sub-question, and according to the sub-review rule information of the sub-question, filters out at least one initial text fragment from the financial document that has a relevance to the sub-question greater than or equal to a preset relevance.

[0015] The intelligent review system analyzes the completeness of the coverage of each initial text fragment for the sub-question.

[0016] The intelligent review system determines a first weight coefficient corresponding to the relevance and a second weight coefficient corresponding to the coverage completion rate;

[0017] The intelligent review system calculates the matching degree between each initial text segment and the sub-question based on the relevance of each initial text segment, the first weight coefficient corresponding to the relevance, the coverage completeness of each initial text segment, and the second weight coefficient corresponding to the coverage completeness.

[0018] The intelligent review system determines the one with the highest matching degree among all the initial text fragments and the sub-question, based on the matching degree between all the initial text fragments and the sub-question, and uses it as the text fragment that matches the sub-question.

[0019] As an optional implementation, in the first aspect of the present invention, the intelligent review system determines whether the sub-problem meets the target review conditions for calling the corresponding large language model based on the sub-review rule information of the sub-problem, and obtains a judgment result, including:

[0020] The intelligent review system analyzes the complexity of the sub-problem based on the sub-review rule information of the sub-problem;

[0021] The intelligent review system determines whether the complexity of the sub-problem is greater than or equal to a preset complexity level;

[0022] When the complexity of a sub-problem is determined to be greater than or equal to the preset complexity, the intelligent review system determines that the sub-problem meets the target review conditions for calling the corresponding large language model;

[0023] When the complexity of a sub-problem is determined to be less than the preset complexity, the intelligent review system determines that the sub-problem does not meet the target review conditions for calling the corresponding large language model.

[0024] As an optional implementation, in the first aspect of the present invention, the intelligent review system reviews the text fragments matching the sub-question based on the judgment result and the sub-review rule information of the sub-question to obtain the answer to the sub-question, including:

[0025] When the judgment result indicates that the sub-question meets the target review conditions, the intelligent review system is triggered to perform the above-mentioned operation of inputting the sub-review rule information of the sub-question and its matching text fragment into the corresponding large language model for review, and obtaining the answer to the sub-question;

[0026] When the judgment result is that the sub-problem does not meet the target review conditions, the intelligent review system extracts the text fragment that matches the sub-problem according to the sub-review rule information of the sub-problem, and obtains the target text fragment that matches the sub-problem.

[0027] The intelligent review system extracts and infers the target text fragment that matches the sub-question based on the sub-review rule information of the sub-question, and obtains the answer to the sub-question.

[0028] As an optional implementation, in the first aspect of the invention, each of the problem types includes one of the following: related party transaction type, defective capital contribution type, share pledge or freeze type, and target illegal type;

[0029] Furthermore, the intelligent review system, based on the determined decomposition method, decomposes each type of information disclosure issue and its review rule information to obtain a set of sub-issues for that type of information disclosure issue and sub-review rule information for each sub-issue within the set of sub-issues, including:

[0030] For each remaining question type other than the aforementioned related-party transaction type, the intelligent review system decomposes the information disclosure question and its review rule information for that question type according to the large language model corresponding to that question type, obtaining a set of sub-questions for the information disclosure question of that question type and sub-review rule information for each sub-question within the set of sub-questions; or,

[0031] The intelligent review system obtains decomposition data obtained by the user performing a decomposition operation on the intelligent review system for the information disclosure question and its review rule information of this type of question, and determines the sub-question set of the information disclosure question of this type of question and the sub-review rule information of each sub-question in the sub-question set based on the decomposition data.

[0032] For related transaction issues of the aforementioned related transaction type, the intelligent review system decomposes the related transaction issue and its review rule information according to the large language model corresponding to the related transaction type, thereby obtaining a set of sub-issues of the related transaction issue and sub-review rule information for each sub-issue within the set of sub-issues.

[0033] As an optional implementation, in the first aspect of the invention, the large language model corresponding to each of the aforementioned problem types is trained in the following manner:

[0034] Collect positive and negative sample sets at different review stages. Each positive sample in the positive sample set is a reference financial document that has not had relevant information disclosure issues, and each negative sample in the negative sample set is a reference financial document that has had relevant information disclosure issues.

[0035] Collect the response files to the inquiry letters for each of the aforementioned reference financial documents;

[0036] For each of the aforementioned response documents, target information related to each of the aforementioned information disclosure issues is retrieved from the response document, the target information including inquiry information and response information; and all of the target information is analyzed to obtain reference questions.

[0037] Based on the ChatGPT model, a correlation analysis is performed between the reference question and each of the information disclosure questions to obtain the analysis results of the reference question; and based on the analysis results of the reference question, the reference question is decomposed to obtain one or more related sub-questions of the reference question, and each related sub-question corresponds to a question type;

[0038] Obtain the relevant paragraphs in the corresponding reference financial documents for each of the analyzed sub-problems; and formulate reference review rules for each sub-problem based on each sub-problem and the relevant paragraphs.

[0039] For each of the aforementioned question types, the relevant sub-questions of each question type and the reference review rules information of the relevant sub-questions of that question type are taken as input, and the response information corresponding to the relevant sub-questions of that question type is taken as output. At least one preset deep learning model is trained to obtain all the initial large language models corresponding to that question type.

[0040] Select one of the initial large language models corresponding to this problem type that meets the preset optimal conditions, and use it as the large language model corresponding to this problem type.

[0041] As an optional implementation, in the first aspect of the present invention, for each of the aforementioned problem types, the method of selecting one of the preset optimal conditions from all initial large language models corresponding to that problem type as the large language model corresponding to that problem type specifically includes:

[0042] Obtain the output results obtained by performing the review operation for the information disclosure question of the question type through each of the initial language models; and determine the test results of the initial large language model based on each of the output results, wherein the test results include one or more combinations of precision, recall and runtime;

[0043] Based on the test results of each initial large language model, analyze the performance of each initial large language model;

[0044] Select the best-performing one from all the initial large language models corresponding to this problem type as the large language model corresponding to this problem type.

[0045] A second aspect of this invention discloses a financial information disclosure document review device based on a large model, the device being applied in an intelligent review system, the device comprising:

[0046] The acquisition module is used to acquire the financial documents uploaded by the user that are to be reviewed, as well as the review rule information for one or more preset information disclosure questions for the financial documents. Each information disclosure question corresponds to a question type.

[0047] The decomposition module is used to decompose each type of information disclosure issue and the review rule information of that type of information disclosure issue according to the determined decomposition method, to obtain a set of sub-issues of that type of information disclosure issue and sub-review rule information of each sub-issue in the set of sub-issues, and each sub-issue has a corresponding retrieval device;

[0048] The retrieval module is used to retrieve, based on the retrieval device and according to the sub-audit rule information of the sub-question, the text fragment matching the sub-question in the financial document for each sub-question;

[0049] The review module is used to input the sub-review rule information of the sub-problem and its matching text fragments into the corresponding large language model for review when the decomposition method is the target decomposition method, so as to obtain the answer to the sub-problem. The target decomposition method is the decomposition method using the large language model.

[0050] The judgment module is used to determine whether the sub-problem meets the target review conditions for calling the corresponding large language model when the decomposition method is not the target decomposition method, based on the sub-review rule information of the sub-problem, and to obtain the judgment result. Each of the problem types has a corresponding large language model.

[0051] The review module is also used to review the text fragment that matches the sub-question based on the judgment result and the sub-review rule information of the sub-question, so as to obtain the answer to the sub-question;

[0052] An integration module is used to integrate the answers to all sub-questions of each disclosure question to obtain a target answer for each disclosure question. The target answer includes whether the financial document has the corresponding disclosure question or whether the financial document does not have the corresponding disclosure question.

[0053] As an optional implementation, in the second aspect of the present invention, for each sub-question, the retrieval module, based on the retrieval instrument and according to the sub-review rule information of the sub-question, retrieves the text fragment matching the sub-question from the financial document in the following specific manner:

[0054] The intelligent review system, based on the retrieval device corresponding to the sub-question, and according to the sub-review rule information of the sub-question, filters out at least one initial text fragment from the financial document that has a relevance to the sub-question greater than or equal to a preset relevance.

[0055] The intelligent review system analyzes the completeness of the coverage of each initial text fragment for the sub-question.

[0056] The intelligent review system determines a first weight coefficient corresponding to the relevance and a second weight coefficient corresponding to the coverage completion rate;

[0057] The intelligent review system calculates the matching degree between each initial text segment and the sub-question based on the relevance of each initial text segment, the first weight coefficient corresponding to the relevance, the coverage completeness of each initial text segment, and the second weight coefficient corresponding to the coverage completeness.

[0058] The intelligent review system determines the one with the highest matching degree among all the initial text fragments and the sub-question, based on the matching degree between all the initial text fragments and the sub-question, and uses it as the text fragment that matches the sub-question.

[0059] As an optional implementation, in the second aspect of the present invention, the judgment module determines whether the sub-problem meets the target review conditions for calling the corresponding large language model based on the sub-review rule information of the sub-problem, and the specific method for obtaining the judgment result includes:

[0060] The intelligent review system analyzes the complexity of the sub-problem based on the sub-review rule information of the sub-problem;

[0061] The intelligent review system determines whether the complexity of the sub-problem is greater than or equal to a preset complexity level;

[0062] When the complexity of a sub-problem is determined to be greater than or equal to the preset complexity, the intelligent review system determines that the sub-problem meets the target review conditions for calling the corresponding large language model;

[0063] When the complexity of a sub-problem is determined to be less than the preset complexity, the intelligent review system determines that the sub-problem does not meet the target review conditions for calling the corresponding large language model.

[0064] As an optional implementation, in a second aspect of the present invention, the review module reviews the text fragments matching the sub-question based on the judgment result and the sub-review rule information of the sub-question, and obtains the answer to the sub-question in the following specific ways:

[0065] When the judgment result indicates that the sub-question meets the target review conditions, the intelligent review system is triggered to perform the above-mentioned operation of inputting the sub-review rule information of the sub-question and its matching text fragment into the corresponding large language model for review, and obtaining the answer to the sub-question;

[0066] When the judgment result is that the sub-problem does not meet the target review conditions, the intelligent review system extracts the text fragment that matches the sub-problem according to the sub-review rule information of the sub-problem, and obtains the target text fragment that matches the sub-problem.

[0067] The intelligent review system extracts and infers the target text fragment that matches the sub-question based on the sub-review rule information of the sub-question, and obtains the answer to the sub-question.

[0068] As an optional implementation, in the second aspect of the invention, each of the problem types includes one of the following: related party transaction type, defective capital contribution type, share pledge or freeze type, and target illegal type;

[0069] Furthermore, the decomposition module decomposes each type of information disclosure issue and its review rule information according to the determined decomposition method, obtaining a set of sub-issues for that type of information disclosure issue and sub-review rule information for each sub-issue within the set of sub-issues. Specifically, this includes:

[0070] For each remaining question type other than the aforementioned related-party transaction type, the intelligent review system decomposes the information disclosure question and its review rule information for that question type according to the large language model corresponding to that question type, obtaining a set of sub-questions for the information disclosure question of that question type and sub-review rule information for each sub-question within the set of sub-questions; or,

[0071] The intelligent review system obtains decomposition data obtained by the user performing a decomposition operation on the intelligent review system for the information disclosure question and its review rule information of this type of question, and determines the sub-question set of the information disclosure question of this type of question and the sub-review rule information of each sub-question in the sub-question set based on the decomposition data.

[0072] For related transaction issues of the aforementioned related transaction type, the intelligent review system decomposes the related transaction issue and its review rule information according to the large language model corresponding to the related transaction type, thereby obtaining a set of sub-issues of the related transaction issue and sub-review rule information for each sub-issue within the set of sub-issues.

[0073] As an optional implementation, in the second aspect of the invention, the large language model corresponding to each of the aforementioned problem types is trained in the following manner:

[0074] Collect positive and negative sample sets at different review stages. Each positive sample in the positive sample set is a reference financial document that has not had relevant information disclosure issues, and each negative sample in the negative sample set is a reference financial document that has had relevant information disclosure issues.

[0075] Collect the response files to the inquiry letters for each of the aforementioned reference financial documents;

[0076] For each of the aforementioned response documents, target information related to each of the aforementioned information disclosure issues is retrieved from the response document, the target information including inquiry information and response information; and all of the target information is analyzed to obtain reference questions.

[0077] Based on the ChatGPT model, a correlation analysis is performed between the reference question and each of the information disclosure questions to obtain the analysis results of the reference question; and based on the analysis results of the reference question, the reference question is decomposed to obtain one or more related sub-questions of the reference question, and each related sub-question corresponds to a question type;

[0078] Obtain the relevant paragraphs in the corresponding reference financial documents for each of the analyzed sub-problems; and formulate reference review rules for each sub-problem based on each sub-problem and the relevant paragraphs.

[0079] For each of the aforementioned question types, the relevant sub-questions of each question type and the reference review rules information of the relevant sub-questions of that question type are taken as input, and the response information corresponding to the relevant sub-questions of that question type is taken as output. At least one preset deep learning model is trained to obtain all the initial large language models corresponding to that question type.

[0080] Select one of the initial large language models corresponding to this problem type that meets the preset optimal conditions, and use it as the large language model corresponding to this problem type.

[0081] As an optional implementation, in the second aspect of the present invention, for each of the aforementioned problem types, the method of selecting one of the preset optimal conditions from all initial large language models corresponding to that problem type as the large language model corresponding to that problem type specifically includes:

[0082] Obtain the output results obtained by performing the review operation for the information disclosure question of the question type through each of the initial language models; and determine the test results of the initial large language model based on each of the output results, wherein the test results include one or more combinations of precision, recall and runtime;

[0083] Based on the test results of each initial large language model, analyze the performance of each initial large language model;

[0084] Select the best-performing one from all the initial large language models corresponding to this problem type as the large language model corresponding to this problem type.

[0085] A third aspect of this invention discloses another financial information disclosure document review device based on a large model, the device comprising:

[0086] Memory containing executable program code;

[0087] A processor coupled to the memory;

[0088] The processor calls the executable program code stored in the memory to execute the financial information disclosure document review method based on a large model disclosed in the first aspect of the present invention.

[0089] The fourth aspect of the present invention discloses a computer storage medium storing computer instructions, which, when invoked, are used to execute the financial information disclosure document review method based on a large model disclosed in the first aspect of the present invention.

[0090] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0091] In this embodiment of the invention, the intelligent review system obtains the financial documents uploaded by the user and the review rule information of one or more preset information disclosure questions for the financial documents. Each information disclosure question corresponds to a question type. The intelligent review system decomposes each type of information disclosure question and its review rule information according to a determined decomposition method, obtaining a set of sub-questions and sub-review rule information for each sub-question within the set. Each sub-question has a corresponding retrieval tool. For each sub-question, the intelligent review system, based on the retrieval tool and the sub-review rule information, retrieves the matching text fragment from the financial document. When the decomposition method is the target decomposition method, the intelligent review system retrieves the sub-review rule information of the sub-question and its matching text fragment. The text fragments are input into the corresponding large language model for review, yielding the answer to the sub-question. The target decomposition method is the decomposition using the large language model. When the decomposition method is not the target decomposition method, the intelligent review system determines whether the sub-question meets the target review conditions for calling the corresponding large language model based on the sub-review rule information of the sub-question, obtaining the judgment result. Each question type has a corresponding large language model. Based on the judgment result and the sub-review rule information of the sub-question, the intelligent review system reviews the text fragments matching the sub-question, obtaining the answer to the sub-question. The intelligent review system integrates the answers to all sub-questions of each information disclosure question to obtain the target answer for each information disclosure question. The target answer is used as the basis for judging whether the financial document contains the corresponding information disclosure question. Therefore, implementing this invention can provide an efficient, automated, and diversified intelligent review system for reviewing information disclosure questions, which is beneficial to improving the accuracy, standardization, and compliance of the review of whether the financial document contains the corresponding information disclosure question. Furthermore, the compliance and standardization review of the intelligent review system can alleviate the illusion problem faced by large language models in practical applications, thus contributing to the further development of the financial field. Attached Figure Description

[0092] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0093] Figure 1 This is a flowchart illustrating a financial information disclosure document review method based on a large model, as disclosed in an embodiment of the present invention.

[0094] Figure 2 This is a system architecture diagram of an intelligent auditing system disclosed in an embodiment of the present invention;

[0095] Figure 3 This is an overall framework diagram of a method for reviewing financial information disclosure documents through an intelligent review system, as disclosed in an embodiment of the present invention.

[0096] Figure 4 This is a schematic diagram of a functional module for decomposing various issues based on an intelligent auditing system, as disclosed in an embodiment of the present invention.

[0097] Figure 5 This is a flowchart illustrating a financial information disclosure document review method based on a large model, as disclosed in an embodiment of the present invention.

[0098] Figure 6 This is a schematic diagram of the structure of a financial information disclosure document review device based on a large model, as disclosed in an embodiment of the present invention.

[0099] Figure 7 This is a schematic diagram of another financial information disclosure document review device based on a large model disclosed in an embodiment of the present invention. Detailed Implementation

[0100] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0101] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product, or end that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or ends.

[0102] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0103] This invention discloses a method and apparatus for reviewing financial information disclosure documents based on a large-scale model. It can acquire the financial document to be reviewed and the review rule information for at least one information disclosure question, and decompose each question and its review rule information to obtain sub-questions and their rules, thereby improving the accuracy and efficiency of decomposing each type of question and its review rules. Subsequently, based on a retrieval tool, matching text fragments are retrieved from the financial document according to the rules of each sub-question, which helps improve the accuracy and efficiency of extracting information disclosure elements matching each type of question from the financial document. When the decomposition method is a target decomposition method, the rules and text fragments of the sub-question are directly input into the large-scale language model for review, which helps improve the review efficiency and accuracy of the sub-questions. When it is not a target decomposition method, the rules and text fragments of the sub-question are first retrieved according to the rules of each sub-question. The system determines whether a question meets the conditions for calling a large language model based on its rules. Then, it reviews the text fragments based on the judgment results and rules, which helps improve the diversity, flexibility, and accuracy of sub-question review. Finally, it integrates the answers to each type of question to obtain the final answer for each category of questions. For example, it determines whether a financial document contains a corresponding information disclosure issue, or whether a financial document does not contain a corresponding information disclosure issue. This solution provides an efficient, automated, and diversified intelligent review system for information disclosure issues. This helps improve the accuracy, standardization, and compliance of the review of whether financial documents contain corresponding information disclosure issues. Furthermore, the system's compliance and standardization review helps alleviate the illusion problem faced by large language models in practical applications, thus contributing to the further development of the financial sector. The following sections provide detailed explanations.

[0104] Example 1

[0105] Please see Figure 1 , Figure 1 This is a flowchart illustrating a financial information disclosure document review method based on a large model, as disclosed in an embodiment of the present invention. Figure 1 The described large-model-based financial information disclosure document review method can be applied to intelligent review systems or to large-model-based financial information disclosure document review devices. These devices can include review equipment or review servers, where the review server can include a cloud server or a local server; this embodiment of the invention does not impose limitations. Figure 1 As shown, this model-based method for reviewing financial information disclosure documents may include the following operations:

[0106] 101. The intelligent review system obtains the financial documents uploaded by users that are to be reviewed, as well as the review rules information for one or more preset information disclosure questions for the financial documents.

[0107] In this embodiment of the invention, the financial documents may include one or more combinations of prospectuses, investment analysis reports, and bond issuance prospectuses. The prospectus refers to a detailed document provided by a company to potential investors when publicly offering shares, containing important information such as the company's history, business, financial condition, management, market conditions, risk factors, and the terms and conditions of the share issuance. Each disclosure issue corresponds to a question type; optionally, each question type may include one of the following: related-party transaction type, defective capital contribution type, share pledge or freeze type, and target violation type.

[0108] Specifically, related-party transactions are a common issue in prospectuses. If the issuer has related-party transactions with related parties, the prospectus must explain the reasonableness and fairness of the pricing of these transactions. Potential problems with related-party transactions may include, but are not limited to, one or more combinations of procedural irregularities (failure to disclose normal decision-making processes, lack of approval from an independent board of directors or supervisory board, or a written agreement), excessively high proportions of related-party transactions in revenue, leading to inflated claims of independent operation and profitability, transaction prices below market value potentially resulting in profit transfers, and undisclosed information (such as non-transactional fund transfers with related parties). This embodiment of the invention does not limit these issues.

[0109] Problems that may arise from defective capital contributions may include verification of capital contributions for companies without securities and futures business qualifications, delayed capital contributions (where the issuer fails to pay the share price after a capital contribution is made on behalf of another party), transfer of equity to employees, employee capital increases at prices lower than the share price, and failure to promptly complete the transfer or registration of ownership of physical assets used for capital contributions. This embodiment of the invention does not limit these issues.

[0110] Potential issues arising from share pledges or freezes include related parties having assets frozen by the court (whether these have been repaid or whether they affect the listing). Target violations (such as major violations) mainly target the following two situations: investigations into directors, supervisors, and senior management (directors, supervisors, and senior management of listed companies) and major violations by the issuer's major shareholders.

[0111] For example, such as Figure 2 As shown, Figure 2 This is a system architecture diagram of an intelligent review system disclosed in an embodiment of the present invention, and Figure 2 The intelligent auditing system shown takes issues including related-party transactions, defective capital contributions, share pledges or freezes, and major violations as examples. Figure 2The system architecture of the intelligent review system shown includes a user layer, a rule layer, and a model layer. The user layer is primarily responsible for uploading PDF files, decrypting encrypted files, and receiving results from the rule layer. The rule layer focuses on handling four types of information leakage issues, with its processing including paragraph-level contextual retrieval and sub-problem decomposition. For complex tasks such as specific information retrieval or semantic understanding, the model layer is delegated for processing, and the results are then returned to the upper layer.

[0112] For example, such as Figure 3 As shown, Figure 3 This is an overall framework diagram of a method for reviewing financial information disclosure documents through an intelligent review system, as disclosed in an embodiment of the present invention. Figure 3 The overall framework of the audit method shown is based on the prospectus as an example, targeting... Figure 3 For further description of the overall framework, please refer to the specific description of the relevant content in Embodiment 1. The embodiments of the present invention will not be repeated here.

[0113] 102. The intelligent review system decomposes each type of information disclosure issue and its review rules according to the determined decomposition method, and obtains a set of sub-issues of that type of information disclosure issue and sub-review rules for each sub-issue within the set of sub-issues.

[0114] In this embodiment of the invention, each problem type corresponds to a decomposition method, and each decomposition method may include either calling the corresponding Large Language Model (LLM) for decomposition or manual decomposition. Each problem type has a corresponding Large Language Model. For example, for related transaction types, the Large Language Model corresponding to the related transaction type can be directly called for decomposition; for other problem types, one can choose to call the Large Language Model corresponding to the problem type for decomposition or choose manual decomposition. This embodiment of the invention does not impose any limitations.

[0115] For example, such as Figure 4 As shown, Figure 4 This is a schematic diagram of a functional module for decomposing various issues based on an intelligent review system, as disclosed in an embodiment of the present invention. Figure 4 The functional modules shown take issues such as related-party transactions, defective capital contributions, share pledges or freezes, and major violations as examples.

[0116] The module on defective capital contributions is divided into the following sub-modules: disclosure of information on shareholders holding 5% or more of the shares (whether the information on shareholders holding 5% or more of the shares of the issuer is fully disclosed: including missing items in the shareholder list or incomplete disclosure of partners in partnerships), disclosure of information on shareholders with equity changes (the positioning of capital increase / transfer events each year and the judgment on whether there is corresponding price fairness and background of capital increase / transfer in the document), verification of capital contributions (whether the issuer's shareholders have issued a corresponding "Capital Verification Report" after declaring capital contributions in the prospectus; if not, it proves that there may be problems with this capital contribution or whether the issuer's shareholders have disclosed the corresponding type of capital contribution after declaring capital contributions in the prospectus), price fluctuations (determining whether there are fluctuations in the capital contribution / transfer price in the short term and whether the fluctuation range of the two capital contribution / transfer prices exceeds the threshold α), and disclosure of information on newly added shareholders in the past year (whether the list of newly added shareholders of the issuer in the past year matches the equity change situation and whether there is sufficient disclosure).

[0117] The share pledge and freeze issue module is divided into the following sub-issue modules: issue statement (whether the prospectus states that there are no share pledges or freezes by the issuer's major shareholders or directors, supervisors and senior management), shareholder pledge situation explanation (whether there are share pledges or freezes by the issuer's major shareholders or directors, supervisors and senior management but the specific circumstances are not explained), and release situation explanation (whether the issuer's share pledges or freezes have been released or whether the issuer has the ability to repay).

[0118] The major illegal issues module is divided into the following sub-modules: illegal events (determining whether the company and its holding subsidiaries, including its affiliates, have been fined in the past), criminal acts (determining whether the issuer's major shareholders, directors, supervisors, and senior management have committed or fully disclosed criminal acts and explained their impact on this listing), personal income tax (whether the issuer has paid personal income tax in full), and financial acts (whether the existence of major illegal acts was disclosed before the subsidiary or affiliated company was deregistered or transferred to an external party).

[0119] 103. For each sub-problem, the intelligent review system uses a retrieval tool to search for the matching text fragment in the financial documents based on the sub-review rule information of that sub-problem.

[0120] In this embodiment of the invention, each sub-problem has a corresponding retrieval tool. Optionally, each retrieval tool can be based on Retrieval-Augmented Generation (RAG) technology or on Chain of Thought (CoT) technology; this embodiment of the invention does not impose any limitations. The text fragments retrieved from the financial documents by the retrieval tool that match each sub-problem can be the most relevant and comprehensive text fragments in the financial documents related to that sub-problem.

[0121] In this embodiment of the invention, when the decomposition method in step 103 is the target decomposition method, step 104 is triggered; when the decomposition method in step 103 is not the target decomposition method, step 105 is triggered.

[0122] 104. The intelligent review system inputs the sub-review rule information of the sub-question and its matching text fragment into the corresponding large language model for review, and obtains the answer to the sub-question.

[0123] In this embodiment of the invention, the target decomposition method is to use the large language model for decomposition.

[0124] 105. The intelligent review system determines whether the sub-problem meets the target review conditions for calling the corresponding large language model based on the sub-review rule information of the sub-problem, and obtains the judgment result.

[0125] In this embodiment of the invention, the judgment result may include the result that the corresponding sub-problem meets the target review condition for calling the corresponding large language model, or the result that the corresponding sub-problem does not meet the target review condition for calling the corresponding large language model.

[0126] 106. Based on the judgment result and the sub-review rule information of the sub-question, the intelligent review system reviews the text fragments that match the sub-question and obtains the answer to the sub-question.

[0127] 107. The intelligent review system integrates the answers to all sub-questions of each information disclosure question to obtain the target answer for each information disclosure question.

[0128] In this embodiment of the invention, the target answer may include whether the financial document has a corresponding information disclosure problem, or whether the financial document does not have a corresponding information disclosure problem.

[0129] It is evident that implementation Figure 1The described large-scale model-based financial information disclosure document review method can acquire the financial document to be reviewed and the review rule information of at least one information disclosure question, and decompose each question and its review rule information to obtain the sub-questions and their rules, which can improve the accuracy and efficiency of the decomposition of each type of question and its review rules. Subsequently, based on the retrieval tool, matching text fragments are retrieved in the financial document according to the rules of each sub-question, which helps to improve the accuracy and efficiency of extracting information disclosure elements that match each type of question in the financial document. When the decomposition method is a target decomposition method, the rules and text fragments of the sub-question are directly input into the large language model for review, which helps to improve the review efficiency and accuracy of the sub-question. When it is not a target decomposition method, the rules of the sub-question are first used to review the sub-question. The system determines whether the conditions for calling the large language model are met; then, based on the judgment results and rules, the text fragments are reviewed, which helps improve the diversity, flexibility, and accuracy of the review of sub-questions; finally, the answers to each type of question are integrated to obtain the final answer for each type of question, such as: whether the financial document has a corresponding information disclosure issue, or whether the financial document does not have a corresponding information disclosure issue. This solution provides an efficient, automated, and diversified intelligent review system for information disclosure issues, which helps improve the accuracy, standardization, and compliance of the review of whether corresponding information disclosure issues exist in financial documents. It also helps alleviate the illusion problem faced by the large language model in practical applications through the compliance and standardization review of the intelligent review system, thereby contributing to the further development of the financial field.

[0130] In an optional embodiment, the large language model corresponding to each question type is trained in the following way:

[0131] Collect positive and negative sample sets at different review stages. Each positive sample in the positive sample set is a reference financial document that has not had relevant information disclosure issues, and each negative sample in the negative sample set is a reference financial document that has had relevant information disclosure issues.

[0132] Collect the response files to the inquiry letters for each reference financial document;

[0133] For each response file, retrieve the target information related to each disclosure question from that response file; and analyze all the target information to obtain reference questions.

[0134] Based on the ChatGPT model, a correlation analysis is performed between the reference question and each information disclosure question to obtain the analysis results of the reference question; and based on the analysis results of the reference question, the reference question is decomposed to obtain one or more related sub-questions of the reference question, and each related sub-question corresponds to a question type;

[0135] Obtain the relevant paragraphs in the corresponding reference financial documents for each relevant sub-issue identified in the analysis; and formulate reference review rules for each relevant sub-issue based on each relevant sub-issue and its relevant paragraphs.

[0136] For each question type, the relevant sub-questions of each question type and the reference review rules information of the relevant sub-questions of that question type are taken as input, and the response information corresponding to the relevant sub-questions of that question type is taken as output. At least one preset deep learning model is trained to obtain all the initial large language models corresponding to that question type.

[0137] Select one of the pre-defined optimal conditions from all the initial large language models corresponding to this problem type, and use it as the large language model corresponding to this problem type.

[0138] In this embodiment of the invention, the review stage may include one or more combinations of the application stage (the stage of reviewing the application draft), the meeting stage (the stage of reviewing the meeting draft), and the registration stage (the stage of reviewing the registration draft). Optionally, the target information may include inquiry information and response information, or information obtained by clustering inquiry information and response information. Optionally, the relevant paragraphs for each relevant sub-question may be retrieved by a retrieval tool or obtained through expert experience analysis. Optionally, the deep learning model may be the open-source large model Shusheng·Puyu 2.5-20B-chat, the large financial model Alphabox, or other pre-defined large models; this embodiment of the invention does not limit the scope of the model.

[0139] It should be noted that the training operation of the large language model corresponding to each question type can be performed by the intelligent review system or by the model training system that interacts with the intelligent review system (after the model training system finishes its operation, it sends the large language model corresponding to the question type to the intelligent review system for the intelligent review system to download the model). This embodiment of the invention does not limit the scope of the operation.

[0140] As can be seen, this optional embodiment can collect positive and negative samples and their response documents to inquiry letters at different review stages; retrieve target information related to each disclosure issue, such as inquiry information and response information, from each response document; analyze all target information to obtain reference questions, which helps improve the accuracy of reference question analysis; subsequently, based on the ChatGPT model, a correlation analysis is performed on the reference questions and each disclosure issue to obtain the analysis results of the reference questions, which helps improve the accuracy of the correlation analysis between the reference questions corresponding to the samples and the disclosure issues of the financial documents to be reviewed; and decompose the reference questions according to the analysis results and obtain the relevant sub-questions obtained from the decomposition in the corresponding reference financial documents. By integrating relevant paragraphs in the document, the review rules for related sub-questions are accurately formulated. Then, each type of related sub-question and its review rules are used as input, and the response information of that type of related sub-question is used as output to train at least one preset deep learning model, obtaining all initial large language models for the corresponding question type. This helps improve the training accuracy and reliability of the large language model for each type of question. Then, one of the preset optimal conditions is selected from all the initial large language models corresponding to the question type as the final large language model used. This helps improve the selection accuracy of the large language model, thereby helping to improve the accuracy and efficiency of subsequent document processing based on the accurately selected large language model.

[0141] In this optional embodiment, as an optional implementation method, for each problem type, one of the preset optimal conditions is selected from all initial large language models corresponding to that problem type as the large language model. Specifically, this selection process may include:

[0142] Obtain the output results of the review operation for the information disclosure question of the question type performed by each initial language model; and determine the test results of the initial large language model based on each output result. The test results include one or more combinations of precision, recall and runtime.

[0143] Based on the test results of each initial large language model, analyze the performance of each initial large language model;

[0144] Select the best-performing one from all the initial large language models corresponding to this problem type as the large language model for this problem type.

[0145] For example, the test results of the Shusheng·Puyu 2.5-20B-chat model are shown in Table 1, and the test results of the Alphabox model are shown in Table 2.

[0146] Problem Type Average accuracy Average recall Related party transactions 0.602 0.6505 Capital contribution defects 0.7845 0.8225 Shares pledged or frozen 0.912 0.94 serious violations 0.925 0.9333

[0147] Table 1. Test results of the Shusheng Puyu 2.5-20B-chat model.

[0148] Problem Type Average accuracy Average recall Related party transactions 0.65 0.6859 Capital contribution defects 0.7705 0.8225 Shares pledged or frozen 0.8975 0.905 serious violations 0.84 0.857

[0149] Table 2 Test results of the Alphabox model

[0150] The results in Table 1 show that the average precision and average recall of the Shusheng·Puyu 2.5-20B-chat model used by the intelligent auditing system for the four types of questions are approximately 0.8, indicating good performance. Comparing with the results in Table 2, except for related-party transaction questions, Shusheng·Puyu 2.5-20B-chat outperforms Alphabox in the other three question types. Regarding the runtime of each large language model, Shusheng·Puyu 2.5-20B-chat takes approximately 10 minutes to process each type of question in each financial document, while Alphabox takes approximately 15 minutes.

[0151] As can be seen, this optional implementation can obtain the output results of the review operation for the information disclosure question of the question type performed by each initial language model; and accurately determine the test results of the initial large language model, such as precision, recall and runtime, based on each output result. This is beneficial to improving the accuracy and diversity of the test results obtained for each large language model. Subsequently, based on the test results of each initial large language model, the performance of each initial large language model is analyzed, and then the one with the best performance is selected from all the initial large language models corresponding to the question type as the large language model corresponding to the question type. This is beneficial to improving the accuracy and reliability of the final selection of the large language model used.

[0152] Example 2

[0153] Please see Figure 5 , Figure 5 This is a flowchart illustrating a financial information disclosure document review method based on a large model, as disclosed in an embodiment of the present invention. Figure 5 The described large-model-based financial information disclosure document review method can be applied to intelligent review systems or to large-model-based financial information disclosure document review devices. These devices can include review equipment or review servers, where the review server can include a cloud server or a local server; this embodiment of the invention does not impose limitations. Figure 2 As shown, this model-based method for reviewing financial information disclosure documents may include the following operations:

[0154] 201. The intelligent review system obtains the financial documents uploaded by users that are to be reviewed, as well as the review rules information for one or more preset information disclosure questions for the financial documents.

[0155] 202. The intelligent review system decomposes each type of information disclosure issue and its review rules based on the determined decomposition method, thereby obtaining a set of sub-issues for that type of information disclosure issue and sub-review rules for each sub-issue within the set of sub-issues.

[0156] 203. For each sub-problem, the intelligent review system uses a retrieval tool to search for the matching text fragment in the financial documents based on the sub-review rule information of that sub-problem.

[0157] In this embodiment of the invention, when the decomposition method in step 203 is the target decomposition method, step 204 is triggered; when the decomposition method in step 203 is not the target decomposition method, step 205 is triggered.

[0158] 204. The intelligent review system inputs the sub-review rule information of the sub-question and its matching text fragment into the corresponding large language model for review, and obtains the answer to the sub-question.

[0159] 205. The intelligent review system determines whether the sub-problem meets the target review conditions for calling the corresponding large language model based on the sub-review rule information of the sub-problem, and obtains the judgment result.

[0160] In this embodiment of the invention, when the judgment result in step 205 is that the sub-question meets the target review conditions, that is, when the judgment result in step 205 is yes, the intelligent review system is triggered to execute the operation in step 204 above, which inputs the sub-review rule information of the sub-question and its matching text fragment into the corresponding large language model for review, and obtains the answer to the sub-question; when the judgment result is that the sub-question does not meet the target review conditions, that is, when the judgment result in step 205 is no, step 206 is triggered.

[0161] 206. The intelligent review system extracts the text fragments that match the sub-question based on the sub-review rule information of the sub-question, and obtains the target text fragments that match the sub-question.

[0162] 207. The intelligent review system extracts and infers the target text fragment that matches the sub-question based on the sub-review rule information of the sub-question, and obtains the answer to the sub-question.

[0163] 208. The intelligent review system integrates the answers to all sub-questions of each information disclosure question to obtain the target answer for each information disclosure question.

[0164] Specifically, for a particular problem type Q, the problem type is first broken down into n sub-problems q using a large language model or manually. i For each subproblem q iUsing the appropriate search engine (i.e., the corresponding search rules and strategies), financial documents (such as prospectuses) are retrieved to obtain the most relevant and comprehensive text fragments related to the question. Paragraph-level searching can be used during the retrieval process. For sub-questions that do not meet the target review criteria, the answer 'a' is obtained directly through further extraction and reasoning using the proposed question review rules. i For sub-questions that meet the target review criteria, the retrieved text fragments and sub-question q will be... i The inputs are combined with a large language model for reasoning and generation to obtain the answer 'a' to the corresponding sub-question. i Finally, the answers to each sub-question are combined to obtain the final answer A for this question type.

[0165] It is evident that implementation Figure 5 The described large-scale model-based financial information disclosure document review method can acquire the financial document to be reviewed and the review rule information of at least one information disclosure question, and decompose each question and its review rule information to obtain the sub-questions and their rules, which can improve the accuracy and efficiency of the decomposition of each type of question and its review rules. Subsequently, based on the retrieval tool, matching text fragments are retrieved in the financial document according to the rules of each sub-question, which helps to improve the accuracy and efficiency of extracting information disclosure elements that match each type of question in the financial document. When the decomposition method is a target decomposition method, the rules and text fragments of the sub-question are directly input into the large language model for review, which helps to improve the review efficiency and accuracy of the sub-question. When it is not a target decomposition method, the rules of the sub-question are first used to review the sub-question. The system determines whether a text fragment meets the conditions for invoking a large language model. Based on the determination results and rules, the text fragment is then reviewed, improving the diversity, flexibility, and accuracy of sub-question review. Finally, the answers to each type of question are integrated to obtain the final answer for each category of questions. For example, if a financial document contains a corresponding information disclosure issue, or if a financial document does not contain a corresponding information disclosure issue, this solution provides an efficient, automated, and diversified intelligent review system for information disclosure issues. This improves the accuracy, standardization, and compliance of the review of whether financial documents contain corresponding information disclosure issues. Furthermore, the system's compliance and standardization review helps alleviate the illusion problem faced by large language models in practical applications, thus contributing to the further development of the financial sector. In addition, when the determination result indicates that the corresponding sub-question meets the target review conditions, the intelligent review system is triggered to perform a review using a large language model, improving the accuracy and reliability of such reviews. When the determination result indicates that the corresponding sub-question does not meet the target review conditions, the intelligent review system directly extracts and infers the answer to this type of question through the proposed review rules, improving the efficiency and speed of such reviews.

[0166] In an optional embodiment, step 203 above, where the intelligent review system retrieves a matching text fragment from the financial document based on the sub-review rule information for each sub-question using a retrieval tool, may include:

[0167] Based on the retrieval system corresponding to the sub-question, the intelligent review system filters out at least one initial text fragment from the financial documents that has a relevance to the sub-question that is greater than or equal to a preset relevance, according to the sub-review rule information of the sub-question.

[0168] The intelligent review system analyzes the completeness of the coverage of each initial text fragment for the sub-question.

[0169] The intelligent review system determines the first weight coefficient corresponding to the degree of relevance and the second weight coefficient corresponding to the degree of coverage completion.

[0170] The intelligent review system calculates the matching degree between each initial text fragment and the sub-question based on the relevance of each initial text fragment, the first weight coefficient corresponding to the relevance, the completeness of the coverage of each initial text fragment, and the second weight coefficient corresponding to the completeness of the coverage.

[0171] The intelligent review system determines the text fragment with the highest matching degree among all the initial text fragments and uses it as the text fragment that matches the sub-question.

[0172] For example, for a certain subproblem, suppose there are initial text fragments A, B, and C that are related to it. The relevance of initial text fragment A is 90% and the coverage completeness is 5%, the relevance of initial text fragment B is 80% and the coverage completeness is 75%, and the relevance of initial text fragment C is 10% and the coverage completeness is 60%. After calculating the weight coefficients of each parameter (for example, the first weight coefficient and the second weight coefficient are both 0.5), we can find that the initial text fragment B has the highest matching degree. At this time, the initial text fragment B is determined as the text fragment that matches the subproblem.

[0173] As can be seen, this optional embodiment can use the intelligent review system to filter at least one initial text fragment from financial documents based on the retrieval tool corresponding to the sub-question and the sub-review rule information of the sub-question. This helps to improve the efficiency and speed of the initial screening of text fragments. It also automatically analyzes the coverage completeness of each initial text fragment for the sub-question and determines the first weight coefficient corresponding to the relevance and the second weight coefficient corresponding to the coverage completeness. Based on the relevance of each initial text fragment, the first weight coefficient corresponding to the relevance, the coverage completeness of each initial text fragment, and the second weight coefficient corresponding to the coverage completeness, it calculates the matching degree between each initial text fragment and the sub-question. This helps to improve the accuracy and reliability of the matching degree calculation between each text fragment and the corresponding sub-question. Then, based on the matching degree between all initial text fragments and the sub-question, it determines the one with the highest matching degree from all initial text fragments as the text fragment that matches the sub-question. This helps to improve the matching accuracy and reliability between each sub-question and the retrieved text fragment.

[0174] In another optional embodiment, the intelligent review system in step 205 above determines whether the sub-question meets the target review conditions for calling the corresponding large language model based on the sub-review rule information of the sub-question, and obtains the judgment result, which may include:

[0175] The intelligent review system analyzes the complexity of the sub-problem based on the sub-review rule information of the sub-problem;

[0176] The intelligent review system determines whether the complexity of the sub-problem is greater than or equal to the preset complexity level.

[0177] When the complexity of a sub-problem is determined to be greater than or equal to the preset complexity, the intelligent review system determines that the sub-problem meets the target review conditions for calling the corresponding large language model.

[0178] When the complexity of a sub-problem is determined to be less than the preset complexity, the intelligent review system determines that the sub-problem does not meet the target review conditions for calling the corresponding large language model.

[0179] As can be seen, this optional embodiment can accurately analyze the complexity of each sub-problem according to the review rules of each sub-problem through the intelligent review system, and determine whether the complexity of the sub-problem is greater than or equal to the preset complexity. If so, it is determined that the sub-problem meets the target review conditions for calling the corresponding large language model, which improves the accuracy of determining sub-problems that meet the target review conditions for calling the corresponding large language model. If not, it is determined that the sub-problem does not meet the target review conditions for calling the corresponding large language model, which improves the accuracy of determining sub-problems that do not meet the target review conditions for calling the corresponding large language model.

[0180] In another optional embodiment, the intelligent review system in step 202 above decomposes each type of information disclosure issue and its review rule information according to the determined decomposition method, obtaining a set of sub-issues of that type of information disclosure issue and sub-review rule information for each sub-issue within the set of sub-issues, which may include:

[0181] For each remaining question type other than related-party transaction types, the intelligent review system decomposes the information disclosure questions and their review rules for that question type based on the corresponding large language model, resulting in a set of sub-questions for that question type and sub-review rules for each sub-question within that set; or,

[0182] The intelligent review system obtains the decomposition data obtained by the user's decomposition operation on the intelligent review system for the information disclosure question and its review rule information of this type of question, and determines the sub-question set of the information disclosure question of this type of question and the sub-review rule information of each sub-question in the sub-question set based on the decomposition data;

[0183] For related-party transaction issues, the intelligent auditing system decomposes the related-party transaction issues and their auditing rules based on the large language model corresponding to the related-party transaction type, resulting in a set of sub-issues of the related-party transaction issues and sub-auditing rules for each sub-issue within the set of sub-issues.

[0184] For example, the decomposition method for related-party transaction issues is as follows: The Chain of Thought (CoT) method, employing a zero-shot strategy, decomposes the problem using a large model. Optionally, constraints can be added to the prompt instruction to reduce the material required for the decomposition steps, thereby improving decomposition efficiency and speed. Subsequently, RAG is used to extract relevant content from sections such as "Issuer Basic Information," "Business and Technology," and "Corporate Governance and Independence" within the financial document. Each decomposed problem and the extracted content are then concatenated and input into the large model to obtain the generated model results.

[0185] As can be seen, this optional embodiment can directly decompose related-party transaction issues and their review rules into multiple sub-issues and review rules for each sub-issue through the intelligent review system based on the corresponding large language model, which is conducive to improving the accuracy and efficiency of decomposing related-party transaction issues. For other issues, decomposition can be performed through the large language model or manually by the user, which is conducive to improving the diversity, flexibility and accuracy of decomposition of other issues.

[0186] Example 3

[0187] Please see Figure 6 , Figure 6 This is a schematic diagram of a financial information disclosure document review device based on a large model, as disclosed in an embodiment of the present invention. Figure 6 The described large-scale model-based financial information disclosure document review device can be applied to intelligent review systems. This device can include review equipment or a review server, where the review server can be a cloud server or a local server; this embodiment of the invention does not impose limitations. Figure 6 As shown, the financial information disclosure document review device based on a large model may include:

[0188] The acquisition module 301 is used to acquire the financial documents uploaded by the user that are to be reviewed, as well as the review rule information of one or more preset information disclosure questions for the financial documents. Each information disclosure question corresponds to a question type.

[0189] The decomposition module 302 is used to decompose each type of information disclosure issue and the review rule information of that type of information disclosure issue according to the determined decomposition method, to obtain a set of sub-issues of that type of information disclosure issue and sub-review rule information of each sub-issue within the set of sub-issues, and each sub-issue has a corresponding retrieval device;

[0190] The retrieval module 303 is used to retrieve, based on the retrieval tool and according to the sub-review rule information of the sub-question, the text fragment that matches the sub-question in the financial document for each sub-question;

[0191] The review module 304 is used to input the sub-review rule information of the sub-problem and its matching text fragments into the corresponding large language model for review when the decomposition method is the target decomposition method, so as to obtain the answer of the sub-problem. The target decomposition method is the decomposition method using the large language model.

[0192] The judgment module 305 is used to determine whether the sub-problem meets the target review conditions for calling the corresponding large language model when the decomposition method is not the target decomposition method, based on the sub-review rule information of the sub-problem, and obtain the judgment result. Each problem type has a corresponding large language model.

[0193] The review module 304 is also used to review the text fragments that match the sub-question based on the judgment result and the sub-review rule information of the sub-question, and obtain the answer to the sub-question;

[0194] The integration module 306 is used to integrate the answers to all sub-questions of each information disclosure question to obtain the target answer for each information disclosure question. The target answer includes whether the financial document contains the corresponding information disclosure question or whether the financial document does not contain the corresponding information disclosure question.

[0195] It is evident that implementation Figure 6 The described large-scale model-based financial information disclosure document review device can acquire the financial document to be reviewed and the review rule information of at least one information disclosure question, and decompose each question and its review rule information to obtain sub-questions and their rules, which can improve the accuracy and efficiency of decomposing each type of question and its review rules. Subsequently, based on the retrieval device, matching text fragments are retrieved in the financial document according to the rules of each sub-question, which helps to improve the accuracy and efficiency of extracting information disclosure elements that match each type of question in the financial document. When the decomposition method is the target decomposition method, the rules and text fragments of the sub-question are directly input into the large language model for review, which helps to improve the review efficiency and accuracy of the sub-question. When it is not the target decomposition method, the rules of the sub-question are first used to review the sub-question. The system determines whether the conditions for calling the large language model are met; then, based on the judgment results and rules, the text fragments are reviewed, which helps improve the diversity, flexibility, and accuracy of the review of sub-questions; finally, the answers to each type of question are integrated to obtain the final answer for each type of question, such as: whether the financial document has a corresponding information disclosure issue, or whether the financial document does not have a corresponding information disclosure issue. This solution provides an efficient, automated, and diversified intelligent review system for information disclosure issues, which helps improve the accuracy, standardization, and compliance of the review of whether corresponding information disclosure issues exist in financial documents. It also helps alleviate the illusion problem faced by the large language model in practical applications through the compliance and standardization review of the intelligent review system, thereby contributing to the further development of the financial field.

[0196] In an optional embodiment, for each sub-question, the retrieval module 303 retrieves the matching text fragment in the financial document based on the sub-review rule information of the sub-question using the retrieval tool. Specifically, this may include:

[0197] Based on the retrieval system corresponding to the sub-question, the intelligent review system filters out at least one initial text fragment from the financial documents that has a relevance to the sub-question that is greater than or equal to a preset relevance, according to the sub-review rule information of the sub-question.

[0198] The intelligent review system analyzes the completeness of the coverage of each initial text fragment for the sub-question.

[0199] The intelligent review system determines the first weight coefficient corresponding to the degree of relevance and the second weight coefficient corresponding to the degree of coverage completion.

[0200] The intelligent review system calculates the matching degree between each initial text fragment and the sub-question based on the relevance of each initial text fragment, the first weight coefficient corresponding to the relevance, the completeness of the coverage of each initial text fragment, and the second weight coefficient corresponding to the completeness of the coverage.

[0201] The intelligent review system determines the text fragment with the highest matching degree among all the initial text fragments and uses it as the text fragment that matches the sub-question.

[0202] As can be seen, this optional embodiment can use the intelligent review system to filter at least one initial text fragment from financial documents based on the retrieval tool corresponding to the sub-question and the sub-review rule information of the sub-question. This helps to improve the efficiency and speed of the initial screening of text fragments. It also automatically analyzes the coverage completeness of each initial text fragment for the sub-question and determines the first weight coefficient corresponding to the relevance and the second weight coefficient corresponding to the coverage completeness. Based on the relevance of each initial text fragment, the first weight coefficient corresponding to the relevance, the coverage completeness of each initial text fragment, and the second weight coefficient corresponding to the coverage completeness, it calculates the matching degree between each initial text fragment and the sub-question. This helps to improve the accuracy and reliability of the matching degree calculation between each text fragment and the corresponding sub-question. Then, based on the matching degree between all initial text fragments and the sub-question, it determines the one with the highest matching degree from all initial text fragments as the text fragment that matches the sub-question. This helps to improve the matching accuracy and reliability between each sub-question and the retrieved text fragment.

[0203] In another optional embodiment, the judgment module 305 determines whether the sub-question meets the target review conditions for calling the corresponding large language model based on the sub-review rule information of the sub-question. The specific method for obtaining the judgment result may include:

[0204] The intelligent review system analyzes the complexity of the sub-problem based on the sub-review rule information of the sub-problem;

[0205] The intelligent review system determines whether the complexity of the sub-problem is greater than or equal to the preset complexity level.

[0206] When the complexity of a sub-problem is determined to be greater than or equal to the preset complexity, the intelligent review system determines that the sub-problem meets the target review conditions for calling the corresponding large language model.

[0207] When the complexity of a sub-problem is determined to be less than the preset complexity, the intelligent review system determines that the sub-problem does not meet the target review conditions for calling the corresponding large language model.

[0208] As can be seen, this optional embodiment can accurately analyze the complexity of each sub-problem according to the review rules of each sub-problem through the intelligent review system, and determine whether the complexity of the sub-problem is greater than or equal to the preset complexity. If so, it is determined that the sub-problem meets the target review conditions for calling the corresponding large language model, which improves the accuracy of determining sub-problems that meet the target review conditions for calling the corresponding large language model. If not, it is determined that the sub-problem does not meet the target review conditions for calling the corresponding large language model, which improves the accuracy of determining sub-problems that do not meet the target review conditions for calling the corresponding large language model.

[0209] In another optional embodiment, the review module 304 reviews the text fragments matching the sub-question based on the judgment result and the sub-review rule information of the sub-question, and the specific method for obtaining the answer to the sub-question may include:

[0210] When the judgment result is that the sub-question meets the target review conditions, the intelligent review system is triggered to perform the above-mentioned operation of inputting the sub-review rule information of the sub-question and its matching text fragment into the corresponding large language model for review, and obtaining the answer to the sub-question;

[0211] When the judgment result is that the sub-question does not meet the target review conditions, the intelligent review system extracts the text fragment that matches the sub-question based on the sub-review rule information of the sub-question, and obtains the target text fragment that matches the sub-question.

[0212] The intelligent review system extracts and infers the target text fragment that matches the sub-question based on the sub-review rule information of the sub-question, and obtains the answer to the sub-question.

[0213] As can be seen, this optional embodiment can trigger the intelligent review system to perform the review operation through the large language model when the judgment result is that the corresponding sub-problem meets the target review conditions. This is beneficial to improving the accuracy and reliability of the review of such problems. When the judgment result is that the corresponding sub-problem does not meet the target review conditions, the intelligent review system can directly extract and reason to obtain the answer to such problems through the proposed problem review rules. This is beneficial to improving the review efficiency and speed of such problems.

[0214] In another optional embodiment, each issue type includes one of the following: related-party transaction type, defective capital contribution type, share pledge or freeze type, and target violation type. Furthermore, the decomposition module 302 decomposes each type of information disclosure issue and its review rule information according to the determined decomposition method, obtaining a set of sub-issues and sub-review rule information for each sub-issue within the set of sub-issues. Specifically, this decomposition may include:

[0215] For each remaining question type other than related-party transaction types, the intelligent review system decomposes the information disclosure questions and their review rules for that question type based on the corresponding large language model, resulting in a set of sub-questions for that question type and sub-review rules for each sub-question within that set; or,

[0216] The intelligent review system obtains the decomposition data obtained by the user's decomposition operation on the intelligent review system for the information disclosure question and its review rule information of this type of question, and determines the sub-question set of the information disclosure question of this type of question and the sub-review rule information of each sub-question in the sub-question set based on the decomposition data;

[0217] For related-party transaction issues, the intelligent auditing system decomposes the related-party transaction issues and their auditing rules based on the large language model corresponding to the related-party transaction type, resulting in a set of sub-issues of the related-party transaction issues and sub-auditing rules for each sub-issue within the set of sub-issues.

[0218] As can be seen, this optional embodiment can directly decompose related-party transaction issues and their review rules into multiple sub-issues and review rules for each sub-issue through the intelligent review system based on the corresponding large language model, which is conducive to improving the accuracy and efficiency of decomposing related-party transaction issues. For other issues, decomposition can be performed through the large language model or manually by the user, which is conducive to improving the diversity, flexibility and accuracy of decomposition of other issues.

[0219] In yet another optional embodiment, the large language model corresponding to each question type is trained in the following manner:

[0220] Collect positive and negative sample sets at different review stages. Each positive sample in the positive sample set is a reference financial document that has not had relevant information disclosure issues, and each negative sample in the negative sample set is a reference financial document that has had relevant information disclosure issues.

[0221] Collect the response files to the inquiry letters for each reference financial document;

[0222] For each response file, target information related to each disclosure question is retrieved from that response file. The target information includes inquiry information and response information. All target information is analyzed to obtain reference questions.

[0223] Based on the ChatGPT model, a correlation analysis is performed between the reference question and each information disclosure question to obtain the analysis results of the reference question; and based on the analysis results of the reference question, the reference question is decomposed to obtain one or more related sub-questions of the reference question, and each related sub-question corresponds to a question type;

[0224] Obtain the relevant paragraphs in the corresponding reference financial documents for each relevant sub-issue identified in the analysis; and formulate reference review rules for each relevant sub-issue based on each relevant sub-issue and its relevant paragraphs.

[0225] For each question type, the relevant sub-questions of each question type and the reference review rules information of the relevant sub-questions of that question type are taken as input, and the response information corresponding to the relevant sub-questions of that question type is taken as output. At least one preset deep learning model is trained to obtain all the initial large language models corresponding to that question type.

[0226] Select one of the initial large language models corresponding to this problem type that meets the preset optimal conditions, and use it as the large language model corresponding to this problem type.

[0227] As can be seen, this optional embodiment can collect positive and negative samples and their response documents to inquiry letters at different review stages; retrieve target information related to each disclosure issue, such as inquiry information and response information, from each response document; analyze all target information to obtain reference questions, which helps improve the accuracy of reference question analysis; subsequently, based on the ChatGPT model, a correlation analysis is performed on the reference questions and each disclosure issue to obtain the analysis results of the reference questions, which helps improve the accuracy of the correlation analysis between the reference questions corresponding to the samples and the disclosure issues of the financial documents to be reviewed; and decompose the reference questions according to the analysis results and obtain the relevant sub-questions obtained from the decomposition in the corresponding reference financial documents. By integrating relevant paragraphs in the document, the review rules for related sub-questions are accurately formulated. Then, each type of related sub-question and its review rules are used as input, and the response information of that type of related sub-question is used as output to train at least one preset deep learning model, obtaining all initial large language models for the corresponding question type. This helps improve the training accuracy and reliability of the large language model for each type of question. Then, one of the preset optimal conditions is selected from all the initial large language models corresponding to the question type as the final large language model used. This helps improve the selection accuracy of the large language model, thereby helping to improve the accuracy and efficiency of subsequent document processing based on the accurately selected large language model.

[0228] In this optional embodiment, as an optional implementation method, for each problem type, one of the preset optimal conditions is selected from all initial large language models corresponding to that problem type as the large language model. Specifically, this includes:

[0229] Obtain the output results of the review operation for the information disclosure question of the question type performed by each initial language model; and determine the test results of the initial large language model based on each output result. The test results include one or more combinations of precision, recall and runtime.

[0230] Based on the test results of each initial large language model, analyze the performance of each initial large language model;

[0231] Select the best-performing one from all the initial large language models corresponding to this problem type as the large language model for this problem type.

[0232] As can be seen, this optional implementation can obtain the output results of the review operation for the information disclosure question of the question type performed by each initial language model; and accurately determine the test results of the initial large language model, such as precision, recall and runtime, based on each output result. This is beneficial to improving the accuracy and diversity of the test results obtained for each large language model. Subsequently, based on the test results of each initial large language model, the performance of each initial large language model is analyzed, and then the one with the best performance is selected from all the initial large language models corresponding to the question type as the large language model corresponding to the question type. This is beneficial to improving the accuracy and reliability of the final selection of the large language model used.

[0233] Example 4

[0234] Please see Figure 7 , Figure 7 This is a schematic diagram of another financial information disclosure document review device based on a large model disclosed in an embodiment of the present invention. Figure 7 As shown, the financial information disclosure document review device based on a large model may include:

[0235] Memory 401 storing executable program code;

[0236] Processor 402 coupled to memory 401;

[0237] The processor 402 calls the executable program code stored in the memory 401 to execute the steps in the financial information disclosure document review method based on a large model as described in Embodiment 1 or Embodiment 2 of the present invention.

[0238] Example 5

[0239] This invention discloses a computer storage medium storing computer instructions. When these computer instructions are invoked, they are used to execute the steps in the financial information disclosure document review method based on a large model as described in Embodiment 1 or Embodiment 2 of this invention.

[0240] Example 6

[0241] This invention discloses a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program, and the computer program is operable to cause a computer to perform the steps in the large-model-based financial information disclosure document review method described in Embodiment 1 or Embodiment 2.

[0242] The device embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0243] Through the detailed description of the above embodiments, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-Erasable Programmable Read-Only Memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0244] Finally, it should be noted that the financial information disclosure document review method and apparatus based on a large model disclosed in the embodiments of the present invention are merely preferred embodiments of the present invention, and are only used to illustrate the technical solutions of the present invention, not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for reviewing financial information disclosure documents based on a large model, characterized in that, The method is applied to an intelligent auditing system, and the method includes: The intelligent review system obtains financial documents uploaded by users for review and review rule information for one or more preset information disclosure questions for the financial documents. Each information disclosure question corresponds to a question type. The intelligent review system decomposes each type of information disclosure issue and the review rule information of that type of information disclosure issue according to the determined decomposition method, and obtains a set of sub-issues of that type of information disclosure issue and sub-review rule information of each sub-issue in the set of sub-issues. Each sub-issue has a corresponding retrieval device. For each of the sub-problems, the intelligent review system, based on the retrieval device and according to the sub-review rule information of the sub-problem, retrieves the text fragment that matches the sub-problem from the financial document; When the decomposition method is the target decomposition method, the intelligent review system inputs the sub-review rule information of the sub-problem and its matching text fragment into the corresponding large language model for review, and obtains the answer to the sub-problem. The target decomposition method is the method of decomposing using the large language model. When the decomposition method is not the target decomposition method, the intelligent review system determines whether the sub-problem meets the target review condition of calling the corresponding large language model based on the sub-review rule information of the sub-problem, and obtains the judgment result. Each of the problem types has a corresponding large language model. The intelligent review system reviews the text fragments that match the sub-question based on the judgment result and the sub-review rule information of the sub-question, and obtains the answer to the sub-question. The intelligent review system integrates the answers to all sub-questions of each disclosure issue to obtain the target answer for each disclosure issue. The target answer includes whether the financial document has the corresponding disclosure issue or whether the financial document does not have the corresponding disclosure issue.

2. The financial information disclosure document review method based on a large model according to claim 1, characterized in that, For each of the aforementioned sub-questions, the intelligent review system, based on the retrieval tool and according to the sub-review rule information for that sub-question, retrieves the text fragment matching that sub-question from the financial document, including: The intelligent review system, based on the retrieval device corresponding to the sub-question, and according to the sub-review rule information of the sub-question, filters out at least one initial text fragment from the financial document that has a relevance to the sub-question greater than or equal to a preset relevance. The intelligent review system analyzes the completeness of the coverage of each initial text fragment for the sub-question. The intelligent review system determines a first weight coefficient corresponding to the relevance and a second weight coefficient corresponding to the coverage completion rate; The intelligent review system calculates the matching degree between each initial text segment and the sub-question based on the relevance of each initial text segment, the first weight coefficient corresponding to the relevance, the coverage completeness of each initial text segment, and the second weight coefficient corresponding to the coverage completeness. The intelligent review system determines the one with the highest matching degree among all the initial text fragments and the sub-question, based on the matching degree between all the initial text fragments and the sub-question, and uses it as the text fragment that matches the sub-question.

3. The financial information disclosure document review method based on a large model according to claim 1 or 2, characterized in that, The intelligent review system determines whether the sub-problem meets the target review conditions for calling the corresponding large language model based on the sub-review rule information of the sub-problem, and obtains the judgment result, including: The intelligent review system analyzes the complexity of the sub-problem based on the sub-review rule information of the sub-problem; The intelligent review system determines whether the complexity of the sub-problem is greater than or equal to a preset complexity level; When the complexity of a sub-problem is determined to be greater than or equal to the preset complexity, the intelligent review system determines that the sub-problem meets the target review conditions for calling the corresponding large language model; When the complexity of a sub-problem is determined to be less than the preset complexity, the intelligent review system determines that the sub-problem does not meet the target review conditions for calling the corresponding large language model.

4. The financial information disclosure document review method based on a large model according to claim 3, characterized in that, The intelligent review system reviews the text fragments matching the sub-question based on the judgment result and the sub-review rule information of the sub-question, and obtains the answer to the sub-question, including: When the judgment result indicates that the sub-question meets the target review conditions, the intelligent review system is triggered to perform the above-mentioned operation of inputting the sub-review rule information of the sub-question and its matching text fragment into the corresponding large language model for review, and obtaining the answer to the sub-question; When the judgment result is that the sub-problem does not meet the target review conditions, the intelligent review system extracts the text fragment that matches the sub-problem according to the sub-review rule information of the sub-problem, and obtains the target text fragment that matches the sub-problem. The intelligent review system extracts and infers the target text fragment that matches the sub-question based on the sub-review rule information of the sub-question, and obtains the answer to the sub-question.

5. The financial information disclosure document review method based on a large model according to claim 4, characterized in that, Each of the aforementioned problem types includes one of the following: related party transaction type, defective capital contribution type, share pledge or freeze type, and target violation type; Furthermore, the intelligent review system, based on the determined decomposition method, decomposes each type of information disclosure issue and its review rule information to obtain a set of sub-issues for that type of information disclosure issue and sub-review rule information for each sub-issue within the set of sub-issues, including: For each remaining question type other than the aforementioned related-party transaction type, the intelligent review system decomposes the information disclosure question and its review rule information for that question type according to the large language model corresponding to that question type, obtaining a set of sub-questions for the information disclosure question of that question type and sub-review rule information for each sub-question within the set of sub-questions; or, The intelligent review system obtains decomposition data obtained by the user performing a decomposition operation on the intelligent review system for the information disclosure question and its review rule information of this type of question, and determines the sub-question set of the information disclosure question of this type of question and the sub-review rule information of each sub-question in the sub-question set based on the decomposition data. For related transaction issues of the aforementioned related transaction type, the intelligent review system decomposes the related transaction issue and its review rule information according to the large language model corresponding to the related transaction type, thereby obtaining a set of sub-issues of the related transaction issue and sub-review rule information for each sub-issue within the set of sub-issues.

6. The financial information disclosure document review method based on a large model according to any one of claims 1, 2, 4, and 5, characterized in that, The large language model corresponding to each of the aforementioned problem types is trained in the following way: Collect positive and negative sample sets at different review stages. Each positive sample in the positive sample set is a reference financial document that has not had relevant information disclosure issues, and each negative sample in the negative sample set is a reference financial document that has had relevant information disclosure issues. Collect the response files to the inquiry letters for each of the aforementioned reference financial documents; For each of the aforementioned response files, target information related to each of the aforementioned information disclosure issues is retrieved from the response file, the target information including inquiry information and response information; And by analyzing all the aforementioned target information, a reference problem is derived; Based on the ChatGPT model, a correlation analysis is performed between the reference question and each of the information disclosure questions to obtain the analysis results of the reference question. Based on the analysis results of the reference problem, the reference problem is decomposed to obtain one or more related sub-problems of the reference problem, and each related sub-problem corresponds to a problem type; Obtain the relevant paragraphs in the corresponding reference financial documents for each of the analyzed sub-problems; and formulate reference review rules for each sub-problem based on each sub-problem and the relevant paragraphs. For each of the aforementioned question types, the relevant sub-questions of each question type and the reference review rules information of the relevant sub-questions of that question type are taken as input, and the response information corresponding to the relevant sub-questions of that question type is taken as output. At least one preset deep learning model is trained to obtain all the initial large language models corresponding to that question type. Select one of the initial large language models corresponding to this problem type that meets the preset optimal conditions, and use it as the large language model corresponding to this problem type.

7. The financial information disclosure document review method based on a large model according to claim 6, characterized in that, For each of the aforementioned problem types, the method of selecting one of the preset optimal conditions from all initial large language models corresponding to that problem type as the large language model for that problem type specifically includes: Obtain the output results obtained by performing the review operation for the information disclosure question of the question type through each of the initial language models; and determine the test results of the initial large language model based on each of the output results, wherein the test results include one or more combinations of precision, recall and runtime; Based on the test results of each initial large language model, analyze the performance of each initial large language model; Select the best-performing one from all the initial large language models corresponding to this problem type as the large language model corresponding to this problem type.

8. A financial information disclosure document review device based on a large model, characterized in that, The device is used in an intelligent auditing system, and the device includes: The acquisition module is used to acquire the financial documents uploaded by the user that are to be reviewed, as well as the review rule information for one or more preset information disclosure questions for the financial documents. Each information disclosure question corresponds to a question type. The decomposition module is used to decompose each type of information disclosure issue and the review rule information of that type of information disclosure issue according to the determined decomposition method, to obtain a set of sub-issues of that type of information disclosure issue and sub-review rule information of each sub-issue in the set of sub-issues, and each sub-issue has a corresponding retrieval device; The retrieval module is used to retrieve, based on the retrieval device and according to the sub-audit rule information of the sub-question, the text fragment matching the sub-question in the financial document for each sub-question; The review module is used to input the sub-review rule information of the sub-problem and its matching text fragments into the corresponding large language model for review when the decomposition method is the target decomposition method, so as to obtain the answer to the sub-problem. The target decomposition method is the decomposition method using the large language model. The judgment module is used to determine whether the sub-problem meets the target review conditions for calling the corresponding large language model when the decomposition method is not the target decomposition method, based on the sub-review rule information of the sub-problem, and to obtain the judgment result. Each of the problem types has a corresponding large language model. The review module is also used to review the text fragment that matches the sub-question based on the judgment result and the sub-review rule information of the sub-question, so as to obtain the answer to the sub-question; An integration module is used to integrate the answers to all sub-questions of each disclosure question to obtain a target answer for each disclosure question. The target answer includes whether the financial document has the corresponding disclosure question or whether the financial document does not have the corresponding disclosure question.

9. A financial information disclosure document review device based on a large model, characterized in that, The device includes: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the financial information disclosure document review method based on a large model as described in any one of claims 1-7.

10. A computer storage medium, characterized in that, The computer storage medium stores computer instructions, which, when invoked, are used to execute the financial information disclosure document review method based on a large model as described in any one of claims 1-7.

Images