Method, device and equipment for disclosure auditing, storage medium and program product

Abstract

The invention discloses a disclosure auditing method, device and equipment, a storage medium and a program product, and relates to the technical field of data processing, and the disclosure auditing method comprises the following steps: obtaining a to-be-audited billboard; constructing a knowledge base on the basis of the stock inviting book to obtain a title knowledge base; a preset large language model is finely adjusted through the title knowledge base, a finely-adjusted large language model is obtained, and the large language model is a pre-training model; and through the fine-tuned large language model, performing disclosure auditing on the stock inviting book to obtain a disclosure auditing result. According to the method and the device, the stock inviting book disclosure auditing based on the retrieval enhancement and fine tuning large language model is realized, the problem of low accuracy of stock inviting book disclosure auditing in the prior art is solved, and the accuracy of stock inviting book disclosure auditing is improved.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a disclosure review method, device, equipment, storage medium and program product. Background Art

[0002] Before a company goes public, regulatory authorities will review the information disclosed by the company in accordance with regulations. The reviewed prospectus is related to financial, business, risk and other content. Usually, the regulatory authorities will compare the keywords in the prospectus to determine whether the prospectus contains the necessary disclosure information.

[0003] Existing technology is based on machine learning text matching. It uses a trained model to perform text analysis on the new prospectus, identify key information points, and compare them with the disclosure rules. However, when faced with long text content, the processing speed will be limited by performance and affect the accuracy of text matching.

[0004] The above contents are only used to assist in understanding the technical solution of the present application and do not constitute an admission that the above contents are prior art. Summary of the invention

[0005] The main purpose of this application is to provide a disclosure review method, device, equipment, storage medium and program product, aiming to solve the technical problem of low accuracy of disclosure review in prospectus in the prior art.

[0006] To achieve the above objectives, this application proposes a disclosure review method, which includes:

[0007] Obtain the prospectus for review;

[0008] Constructing a knowledge base based on the prospectus to obtain a title knowledge base;

[0009] Fine-tune the preset large language model through the title knowledge base to obtain a fine-tuned large language model, where the large language model is a pre-trained model;

[0010] The fine-tuned large language model is used to conduct a disclosure review on the prospectus to obtain a disclosure review result.

[0011] In one embodiment, the prospectus includes several chapters, and the step of constructing a knowledge base based on the prospectus to obtain a title knowledge base includes:

[0012] Extract concepts from each chapter in the prospectus and obtain the first-level heading rules;

[0013] Abstract and summarize the subdivision rules of each chapter into secondary heading rules;

[0014] Determining whether the subdivision rules of each chapter contain further subdivision rules;

[0015] If the subdivision rules of each chapter include further subdivision rules, the further subdivision rules are abstracted and summarized into three-level heading rules;

[0016] A knowledge base is constructed based on the first-level title rule, the second-level title rule and the third-level title rule to obtain a title knowledge base.

[0017] In one embodiment, the step of fine-tuning a preset large language model through the title knowledge base to obtain a fine-tuned large language model includes:

[0018] According to the prospectus, searching is performed through the title knowledge base to obtain search results;

[0019] Obtaining historical disclosure audit records based on the search results;

[0020] The large language model is fine-tuned according to the retrieval results and the historical disclosure audit records to obtain a fine-tuned large language model.

[0021] In one embodiment, the title knowledge base includes a number of title rules, and the step of searching the title knowledge base according to the prospectus to obtain search results includes:

[0022] Constructing the disclosure rules based on the complete narrative content of the several title rules;

[0023] The titles corresponding to the disclosure rules in the prospectus and the titles in the title knowledge base are calculated for relevance by using a preset word embedding vector model, and the titles with the highest relevance in the title knowledge base are obtained as relevant titles;

[0024] The search results are obtained according to the relevant titles and the corresponding text contents in the prospectus.

[0025] In one embodiment, the step of fine-tuning the large language model according to the search results and the historical disclosure audit records to obtain a fine-tuned large language model includes:

[0026] Based on a preset prompt word template, generating prompt words according to the search results;

[0027] Generate a sample set based on the prompt words and historical disclosure audit records;

[0028] The sample set is input into the large language model for training and testing, so as to fine-tune the large language model and obtain a fine-tuned large language model.

[0029] In one embodiment, the disclosure rules include at least one or more of issuance overview, operating conditions and risk factors, the prompt words include at least one or more of relevant titles and text contents, and the disclosure review results include at least one or more of review conclusions, judgment basis, improvement suggestions and text location.

[0030] In addition, to achieve the above-mentioned purpose, the present application also proposes a disclosure review device, which includes:

[0031] The prospectus acquisition module is used to obtain the prospectus to be reviewed;

[0032] A knowledge base construction module, used to construct a knowledge base based on the prospectus to obtain a title knowledge base;

[0033] A model fine-tuning module, used to fine-tune a preset large language model through the title knowledge base to obtain a fine-tuned large language model, wherein the large language model is a pre-trained model;

[0034] The disclosure review module is used to perform a disclosure review on the prospectus through the fine-tuned large language model to obtain a disclosure review result.

[0035] In addition, to achieve the above-mentioned purpose, the present application also proposes a disclosure review device, which includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the disclosure review method described above.

[0036] In addition, to achieve the above-mentioned purpose, the present application also proposes a storage medium, which is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, the steps of the disclosure review method described above are implemented.

[0037] In addition, to achieve the above-mentioned purpose, the present application also provides a computer program product, which includes a computer program, and when the computer program is executed by a processor, it implements the steps of the disclosure review method described above.

[0038] This application provides a disclosure review method, which obtains the prospectus to be reviewed to understand the content that needs to be disclosed and reviewed; uses the obtained prospectus to build a title knowledge base, so as to build a title knowledge base for the current prospectus to be reviewed in real time for retrieval enhancement; fine-tunes the pre-trained large language model through the constructed title knowledge base to obtain a fine-tuned large language model, so that the large language model can better adapt to the disclosure review context of the prospectus for processing; uses the fine-tuned large language model to conduct a disclosure review of the prospectus, and obtains the disclosure review results, so as to conduct a disclosure review of the prospectus by combining retrieval enhancement and model fine-tuning, thereby improving the accuracy of the disclosure review text matching. The disclosure review of the prospectus based on retrieval enhancement and model fine-tuning is realized, which solves the problem of low accuracy of the disclosure review of the prospectus in the prior art and improves the accuracy of the disclosure review of the prospectus. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0041] Figure 1 A flowchart of the first embodiment of the audit method disclosed in this application is provided;

[0042] Figure 2 A flowchart diagram of a second embodiment of the disclosure review method of this application;

[0043] Figure 3 A flowchart diagram of the detailed step S1 provided for the second embodiment of the disclosure review method of this application;

[0044] Figure 4 A flowchart diagram of the detailed step S3 provided for the second embodiment of the disclosure review method of this application;

[0045] Figure 5 A flowchart diagram of the detailed step S4 provided for the second embodiment of the disclosure review method of this application;

[0046] Figure 6 A schematic diagram of the module structure of the audit device disclosed in the embodiment of the present application;

[0047] Figure 7 A schematic diagram of the device structure of the hardware operating environment involved in the audit method disclosed in the embodiments of the present application.

[0048] The purpose, features and advantages of this application will be further described in conjunction with the embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0049] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of the present application and are not used to limit the present application.

[0050] In order to better understand the technical solution of the present application, a detailed description will be given below in conjunction with the accompanying drawings and specific implementation methods.

[0051] The main solution of the embodiment of the present application is: obtain the prospectus to be reviewed; construct a knowledge base based on the prospectus to obtain a title knowledge base; fine-tune the preset large language model through the title knowledge base to obtain a fine-tuned large language model, which is a pre-trained model; conduct a disclosure review of the prospectus through the fine-tuned large language model to obtain a disclosure review result.

[0052] In the existing technology, the disclosure review of the prospectus is usually based on two methods: keyword retrieval and text matching based on machine learning.

[0053] The keyword search method extracts keywords or phrases related to finance, business, risk, etc. from the prospectus, and then compares these keywords with the keyword library in the pre-set disclosure rules to determine whether the prospectus contains all the necessary disclosure information. This keyword search method may be affected by the subjectivity of keyword selection. Different auditors may choose different keywords due to differences in personal understanding, which may lead to the omission or misidentification of key information.

[0054] Machine learning-based text matching first uses a large amount of labeled prospectus data to train the machine learning model so that it can learn and understand the structure and content of the prospectus. Then, the trained model is used to perform text analysis on the new prospectus, identify key information points, and compare them with the disclosure rules. Machine learning-based text matching technology may encounter performance challenges when processing long texts, which may not only reduce processing speed but also affect matching accuracy.

[0055] In addition, prospectuses often contain complex multi-level headings and structures, which is a challenge for existing technologies because they may have difficulty in accurately parsing and understanding these complex text structures.

[0056] Technical terms involved in this application:

[0057] Disclosure review: refers to the process by which regulatory agencies or relevant institutions review and verify the information disclosed by enterprises in accordance with laws, regulations, industry standards and exchange rules. It ensures that the information disclosed by enterprises can truly, accurately and completely reflect their financial status, operating results and cash flow, provide decision-making basis for investors, creditors and other stakeholders, and maintain market order and fair competition.

[0058] Word Embedding Vector Model: It is an important technology in the field of natural language processing (NLP). It aims to convert words or characters into numerical vectors so that these numerical values ​​can represent the semantic relationship between words. This representation method makes words with similar meanings have similar representations in the vector space.

[0059] Large Language Models: refers to machine learning models with a large number of parameters and computing resources, especially deep neural network models. These models can provide higher performance and representation capabilities when processing complex tasks.

[0060] Supervised Fine-Tuning (SFT): refers to further training a pre-trained model (such as a large language model) by providing labeled data to make the model perform better in specific tasks or fields.

[0061] This application establishes a set of standardized disclosure rules and uses advanced large-scale model technology for auditing, which greatly reduces the reliance on the auditor's personal experience and knowledge level, effectively avoids the problem of inconsistent audit standards, and reduces the risk of missing key information. It not only significantly improves the accuracy of the audit work, but also enhances the reliability and consistency of the audit results. At the same time, it can automatically perform text recall and review according to disclosure requirements. This automated function greatly reduces the time and cost required for manual review.

[0062] It should be noted that the execution subject of this embodiment can be a computing service device with data processing, network communication and program running functions, such as a tablet computer, a personal computer, a mobile phone, etc., or an electronic device capable of realizing the above functions, a disclosure review device, etc. The following takes the disclosure review device as an example to illustrate this embodiment and the following embodiments.

[0063] Based on this, the present application embodiment provides a disclosure review method, referring to Figure 1 , Figure 1 This is a flowchart of the first embodiment of the audit method disclosed in this application.

[0064] In this embodiment, the disclosure review method includes steps S10 to S40:

[0065] Step S10, obtaining the prospectus to be reviewed;

[0066] It should be noted that the prospectus refers to the document disclosed to the public before a company publicly issues its shares. It can also be called a prospectus, which contains the company's financial status, business model, management information, etc.

[0067] It is understandable that step S10 is the basis or starting step of the audit process. In order to ensure that the audit work has clear objects and contents, provide necessary inputs for subsequent steps, and collect prospectuses that need to be disclosed for audit.

[0068] Step S20, constructing a knowledge base based on the prospectus to obtain a title knowledge base;

[0069] It should be noted that the process of building a knowledge base is usually to build the processed data into a knowledge base, which can be done by dividing the text into smaller fragments through natural language processing, converting these fragments into vectors using a text embedding model, and storing these vectors in a vector database.

[0070] It is understandable that providing standardized and customized training data for the large language model will enable it to better understand and process text data related to the prospectus, and build a structured title knowledge base by analyzing the information in the prospectus.

[0071] Step S30, fine-tuning the preset large language model through the title knowledge base to obtain a fine-tuned large language model, where the large language model is a pre-trained model;

[0072] It should be noted that a large language model refers to an artificial intelligence model that can understand and generate natural language text. In order to optimize the model's task processing in a specific field, it is necessary to use the general knowledge of the field to pre-train the large language model so that the model can be fine-tuned in combination with task requirements during task processing. Fine-tuning refers to the process of further training the pre-trained model on a dataset for a specific task to improve the model's adaptability and accuracy and adapt it to new tasks or fields.

[0073] It is understandable that improving the performance of the large language model on specific tasks enables it to more accurately understand and process the information in the prospectus, and using the constructed knowledge base to fine-tune the pre-trained large language model to adapt it to specific disclosure review tasks.

[0074] Step S40: using the fine-tuned large language model, conduct a disclosure review on the prospectus to obtain a disclosure review result.

[0075] It is understandable that in order to improve the efficiency and accuracy of the audit and reduce the workload of manual audit, the prospectus that needs to be audited can be input into the fine-tuned large language model, and the fine-tuned large language model can be used to perform automated disclosure audit on the prospectus to obtain the disclosure audit results output by the model to check the completeness, consistency and compliance of the information disclosed in the prospectus.

[0076] In a feasible implementation manner, the prospectus includes several chapters, and the step of constructing a knowledge base based on the prospectus to obtain a title knowledge base includes:

[0077] Step S201, extracting concepts from each chapter in the prospectus to obtain first-level title rules;

[0078] It should be noted that concept extraction can be the process of identifying and extracting key concepts or themes from text or data. These concepts are usually the core ideas or important information points in the text. They can be nouns, noun phrases, terms or other words that express key information. For example, key concepts such as "financial status", "market risk", "business model" and so on are extracted from various chapters of the prospectus. These concepts form the basis of the knowledge base. The first-level title rule refers to the main concepts or themes of each chapter of the prospectus, which serve as the basic framework of the knowledge base.

[0079] In addition, it should be noted that the prospectus contains multiple chapters, each chapter corresponds to a first-level title, each chapter has subdivision rules, and the subdivision rules under the chapters will have further subdivision rules. Therefore, it is necessary to generate a structured title knowledge base and perform the same structured processing on the internal structure of the prospectus.

[0080] It can be understood that in order to determine the main content and purpose of each chapter and provide a basis for subsequent rule construction, by analyzing the various chapters of the prospectus, identifying and refining the core concepts and themes of each chapter, and using the first-level heading rules as the basis of the knowledge base, the model can understand the high-level structure and main parts of the prospectus.

[0081] Step S202, abstracting and summarizing the subdivision rules of each chapter into secondary title rules;

[0082] It should be noted that abstraction and generalization is the process of extracting general rules or patterns from specific examples or data, which involves summarizing specific information or data points into broader categories or concepts, thereby simplifying complex information and making it easier to understand and process. For example, on the basis of extracting key concepts from various chapters of the prospectus, further abstraction and generalization of broader rules or categories, such as abstracting specific financial indicators such as "revenue growth" and "cost control" into the broader concept of "financial health". Secondary heading rules refer to subheadings or subdivisions within each chapter, which are used to increase the depth of the knowledge base.

[0083] It is understandable that in order to refine the content of the knowledge base so that it can cover more details within the chapter, on the basis of the first-level heading rules, the subdivided content within each chapter is further analyzed and more specific rules are abstracted. Among them, the second-level heading rules increase the depth of the knowledge base, enabling the model to understand and process the specific content in the prospectus more accurately.

[0084] Step S203, determining whether the subdivision rules of each chapter include further subdivision rules;

[0085] It is understandable that the key to determine whether the knowledge base needs to be further subdivided is to ensure that the structure of the knowledge base matches the actual structure of the prospectus, check whether there are more detailed classifications or rules under the secondary heading rules that need to be identified and abstracted, and ensure the comprehensiveness and meticulousness of the knowledge base, covering all relevant information in the prospectus.

[0086] Step S204, if the subdivision rules of each chapter include further subdivision rules, the further subdivision rules are abstracted and summarized into three-level heading rules;

[0087] It should be noted that the third-level heading rules refer to specific rules that are further subdivided under the second-level headings to increase the detail of the knowledge base.

[0088] It is understandable that more detailed levels are added to the knowledge base to cover the most detailed information in the prospectus. On the basis of the second-level heading rules, the subdivision rules of each chapter can be further subdivided, and the further subdivided rules can be abstracted and summarized into more specific third-level heading rules, so that the knowledge base can cover the structure of the prospectus more carefully and comprehensively, which will help the subsequent large language model to deeply understand the complex structure of the prospectus.

[0089] Step S205, constructing a knowledge base based on the first-level title rule, the second-level title rule and the third-level title rule to obtain a title knowledge base.

[0090] It is understandable that in order to create a comprehensive and structured knowledge base to guide and optimize the disclosure review process of the prospectus, all levels of title rules of the prospectus (i.e., first-level title rules, second-level title rules, and third-level title rules) are integrated to build a multi-level and structured title knowledge base, which provides rich and structured training data for the subsequent fine-tuning of the large language model, so as to more accurately understand and process the disclosure information in the prospectus.

[0091] In this embodiment, a highly structured, detailed and standardized title knowledge base is achieved through a multi-level and comprehensive knowledge base construction method, so that the large language model can more accurately understand and process the complex information in the prospectus. It is suitable for processing prospectuses containing complex multi-level titles and structures, and can provide a comprehensive and multi-level perspective to understand and parse the complex text structure in the prospectus, improve the accuracy and efficiency of the prospectus disclosure review, and reduce the risk of omission or misidentification of key information.

[0092] In a feasible implementation manner, the step of fine-tuning the preset large language model through the title knowledge base to obtain the fine-tuned large language model includes:

[0093] Step S301, searching through the title knowledge base according to the prospectus to obtain search results;

[0094] It should be noted that the search results refer to the results obtained after searching the prospectus through the title knowledge base, which contain information points that match the standards and rules defined in the title knowledge base, for example, the standardized titles in the title knowledge base and the corresponding titles in the prospectus.

[0095] It is understandable that the structured information defined in the title knowledge base (such as primary, secondary, and tertiary title rules) is used to match and retrieve the contents of the prospectus, identify and locate the key information in the prospectus, and provide precise data support for subsequent model fine-tuning and disclosure review, ensuring that the disclosure review process can focus on the key parts of the prospectus, improving the efficiency of text matching and the accuracy of disclosure review.

[0096] Step S302, obtaining historical disclosure audit records according to the search results;

[0097] It should be noted that the historical disclosure audit record pointer is a retrieval result, which is a record of the same key information points in the past during the disclosure audit process. It can include audit conclusions, judgment basis, improvement suggestions and related analysis, which are used for training and fine-tuning the model.

[0098] It is understandable that by utilizing the key information points in the search results, relevant disclosure audit records can be found from the historical database, and past audit cases and decision-making logic related to the key information points of the current prospectus can be collected to provide actual case support for model fine-tuning, and the decision-making ability of the model can be enhanced through historical data, so that it can learn and optimize based on actual cases.

[0099] Step S303: fine-tune the large language model according to the search results and the historical disclosure audit records to obtain a fine-tuned large language model.

[0100] It is understandable that the pre-trained large language model is further trained through machine learning technology using retrieval results and historical audit records as training data to adjust the model parameters of the model and optimize the performance of the model, so that the model can understand the complex structure in the prospectus and process the complex text information in the prospectus, thereby improving its accuracy and reliability in actual disclosure audits.

[0101] In this implementation, by combining the retrieval enhancement technology of the knowledge base, it is possible to ensure that key information matching the title knowledge base is accurately extracted from the prospectus, and obtain relevant historical disclosure audit records, providing rich training data for the model. Finally, the large language model is fine-tuned through supervised learning and reinforcement learning methods to optimize the performance of the model so that it can better adapt to the prospectus disclosure audit task, which not only improves the efficiency and accuracy of the audit, but also can adapt to the complexity and diversity of the prospectus structure, thereby significantly improving the quality of the disclosure audit and effectively supporting the complex prospectus document audit work.

[0102] In a feasible implementation manner, the title knowledge base includes a number of title rules, and the step of searching through the title knowledge base according to the prospectus to obtain search results includes:

[0103] Step S3011, constructing disclosure rules based on the complete description content of the several title rules;

[0104] It should be noted that the title rules refer to the rules used in the title knowledge base to describe the key information and structure that each part of the prospectus should contain. Disclosure rules refer to a set of rules that are constructed based on the complete narrative content of different title rules and are used to guide how to retrieve and identify key information from the prospectus.

[0105] In addition, it should be noted that the complete narrative content is for each title or chapter in the prospectus, and it is necessary to ensure that all relevant information and details contained therein are fully described and presented, so as to ensure the integrity and transparency of information disclosure and meet regulatory requirements and investors' information needs.

[0106] It is understandable that the key information points that should be included in different title rules in the title knowledge base should be analyzed, and corresponding disclosure rules should be defined to ensure that these information points are fully described in the prospectus to provide guidance on how to retrieve and identify key information from the prospectus and ensure the integrity of the information, thereby ensuring that all information under each title can be accurately identified and evaluated during the disclosure review process, thereby improving the integrity of information disclosure.

[0107] Step S3012, calculating the relevance of the titles corresponding to the disclosure rules in the prospectus and the titles in the title knowledge base using a preset word embedding vector model, and obtaining the title with the highest relevance in the title knowledge base as the relevant title;

[0108] It should be noted that the word embedding vector model refers to a model that converts text into a numerical vector, which enables computers to process and analyze text data and is used to convert titles into computable vector form. Relevance calculation evaluates the semantic relevance of two vectors by calculating the similarity between them, and is used to determine the degree of match between the prospectus title and the title in the title knowledge base. Relevant titles refer to the titles in the title knowledge base that best match the disclosure rules in the prospectus, and are used to guide subsequent text content retrieval.

[0109] It can be understood that the titles corresponding to the disclosure rules in the prospectus are converted into vectors using a pre-trained word embedding vector model, and the correlation between the titles in the prospectus and the titles in the title knowledge base is determined by calculating the similarity between these vectors and the title vectors in the title knowledge base (for example, using cosine similarity), so as to more accurately retrieve and extract relevant information, improve the accuracy of retrieval, and ensure that the information processed in subsequent steps is highly relevant to the disclosure requirements.

[0110] Step S3013, obtaining search results based on the relevant titles and the corresponding text content in the prospectus.

[0111] It is understandable that relevant titles are identified and text content directly related to these titles is extracted from the prospectus, so as to obtain specific text information directly related to the disclosure rules, provide a basis for subsequent review and analysis, ensure the practicality and relevance of the search results, and support efficient long text matching and disclosure review.

[0112] In this implementation, detailed disclosure rules are first constructed based on the complete description of the title rules in the title knowledge base to ensure that the retrieval process covers all important disclosure requirements, ensuring the accuracy and completeness of the disclosure rules while also improving the comprehensiveness of the retrieval. Next, the word embedding vector model and relevance calculation are used to accurately identify the titles most relevant to the disclosure rules to ensure the high relevance and accuracy of subsequent retrieval work. Then, specific text content is extracted from the prospectus based on these relevant titles, providing highly relevant information for the sample set acquisition of the model fine-tuning, thereby ensuring a high correlation between the sample set in the model fine-tuning and the prospectus that needs to be disclosed and reviewed this time, so that the model can better adapt to the disclosure review task that requires this prospectus, thereby significantly improving the quality and reliability of the prospectus disclosure review.

[0113] In a feasible implementation manner, the step of fine-tuning the large language model according to the search results and the historical disclosure audit records to obtain the fine-tuned large language model includes:

[0114] Step S3031, generating prompt words according to the search results based on a preset prompt word template;

[0115] It should be noted that the prompt word template is a predefined template, which defines how to convert the retrieval results into the prompt word form required for model training. It is used to combine the specific information filled in with the disclosure review-related prompt word framework defined in the template to generate prompt words and guide the model to focus on specific information.

[0116] It can be understood that by using the preset prompt word template, the search results are combined with the predefined prompt word template content to generate prompt words that can guide the large language model to pay attention to and learn key information related to the disclosure review, improve the targeted nature of the model training, and ensure that the model can focus on the information most relevant to the disclosure review.

[0117] Step S3032, generating a sample set according to the prompt words and historical disclosure audit records;

[0118] It should be noted that the sample set contains a collection of input texts and expected outputs, which are used to train and test models. The sample set may include the prompt word text of the input model (relevant titles and corresponding text content in the prospectus, etc.) and the expected output (the audit results of the audit history records disclosed related to the prompt word text, which may include audit results, reasons, etc.), and can be divided into training sets and test sets according to a certain ratio.

[0119] It can be understood that combining prompt words and historical disclosure audit records to create sample sets for training and testing models, creating sample sets containing actual audit cases for training and testing large language models, enables the model to learn and verify in actual disclosure audit scenarios, and improves its performance in real audit tasks.

[0120] Step S3033: input the sample set into the large language model for training and testing, so as to fine-tune the large language model and obtain a fine-tuned large language model.

[0121] It is understandable that the large language model is further trained (fine-tuned) using the sample set to obtain a fine-tuned large language model in order to optimize the performance of the large language model so that it can accurately perform disclosure review tasks, ensure that it can deeply understand complex language results and subtle semantic differences when processing the disclosure review of the prospectus, effectively process the long text content in the prospectus, provide high-quality output disclosure review results, and improve the accuracy and reliability of the disclosure review.

[0122] In this implementation, prompt words that can guide the model to focus on specific information are first generated based on a preset prompt word template. These prompt words are combined with historical disclosure audit records to generate sample sets. These sample sets contain actual audit cases and decisions, providing rich learning materials for the model. Finally, the sample set is input into a large language model for training and testing to fine-tune the model so that it can better adapt to specific disclosure audit tasks. This not only improves the pertinence and accuracy of the model, but also ensures its effectiveness and reliability in practical applications.

[0123] In a feasible implementation, the disclosure rules include at least one or more of the issuance overview, operating conditions and risk factors, the prompt words include at least one or more of the relevant titles and text contents, and the disclosure review results include at least one or more of the review conclusions, judgment basis, improvement suggestions and text location.

[0124] It should be noted that the disclosure rules are the basic guidelines for guiding the disclosure of information in the prospectus, which may include disclosure requirements for the issuance overview, operating conditions and risk factors. For example, the issuance overview may include: 1. The type, quantity and price of the issued shares; 2. The purpose and expected effect of the raised funds; 3. Changes in the equity structure before and after the issuance; 4. Information on financial advisors, lawyers and other intermediaries related to the issuance. The operating conditions may include: 1. The company's main business, products or services; 2. Market position, business model, major customers and suppliers; 3. Recent operating performance and future development plans; 4. Management's discussion and analysis of the company's future development. Risk factors may include: 1. The main risks faced, including market risks, financial risks, legal risks, etc.; 2. The potential impact of risks and the company's response measures; 3. Any factors that may have a significant adverse impact on the company's operations, financial conditions or stock prices.

[0125] Prompt words are keywords or phrases used to guide the large language model to focus on specific information, including at least relevant titles and text content. Among them, the relevant titles can be the corresponding chapter titles in the prospectus, such as "Issuance Overview", "Operating Conditions", "Risk Factors", etc., or any sub-titles that can indicate the key information part. The text content can be specific descriptions or data related to the disclosure rules, or it can be a summary or overview of key information, such as a summary of financial data, a description of major risk points, etc.

[0126] The disclosure audit results are the output of the audit process, which at least include the audit conclusions, judgment basis, improvement suggestions and text location. The audit conclusion may include an overall evaluation of the completeness and accuracy of the information disclosed in the prospectus, as well as a conclusion on whether it complies with relevant laws, regulations and disclosure rules. The judgment basis may include specific evidence or reasons supporting the audit conclusion, as well as references to relevant laws, regulations or standards. Improvement suggestions may include suggestions for improvements to problems or deficiencies found, as well as suggestions on how to improve information disclosure to increase transparency and credibility. The text location may include indicating the specific location of the text in the prospectus, including page numbers, chapters or paragraphs, which helps to quickly locate and reference relevant information.

[0127] This embodiment provides a disclosure review method. First, by obtaining the prospectus to be reviewed, the smooth progress of the review work is ensured. Then, natural language processing technology is used to automatically extract key information from the prospectus to build a title knowledge base to ensure the comprehensiveness and accuracy of the title knowledge base. Then, the model is fine-tuned according to the characteristics of the review task to improve the adaptability and accuracy of the model. Finally, combined with automated disclosure review, the accuracy and compliance of the review results are ensured. It can not only improve the review efficiency and reduce manual workload, but also improve the accuracy and reliability of the review through artificial intelligence technology, especially when dealing with complex multi-level titles and structures, it can better understand and parse these complex text structures, thereby providing higher quality review results. The prospectus disclosure review based on retrieval enhancement and model fine-tuning is realized, which solves the problem of low accuracy of prospectus disclosure review in the prior art and improves the accuracy of prospectus disclosure review.

[0128] Based on the first embodiment of the present application, in the second embodiment of the present application, the same or similar contents as those of the above-mentioned first embodiment can be referred to the above introduction and will not be repeated later.

[0129] For example, in order to help understand the implementation process of the disclosure review method obtained by combining this embodiment with the above-mentioned first embodiment, as shown in FIG. Figure 2 As shown, Figure 2 The present application discloses a flow chart of a second embodiment of the audit method, which may include the following steps:

[0130] S1. Build a standardized title knowledge base for title retrieval in prospectuses;

[0131] S2, construct a mapping between disclosure rules and titles in the title knowledge base;

[0132] S3. Based on the constructed standardized title knowledge base, use the word embedding vector model to retrieve the relevant titles and text content of the prospectus;

[0133] S4. Fine-tune the large language model based on the collected training data and review the prospectus in accordance with the disclosure requirements.

[0134] Among them, Figure 3 As shown, Figure 3 The flowchart of step S1 provided in the second embodiment of the disclosure review method of this application is detailed. Step S1 may also include the following steps:

[0135] S11. Extract concepts from the chapter titles in the prospectus document and construct a first-level title rule that covers the entire document;

[0136] S12. Based on the first-level title rules, further refine them and abstractly summarize the subdivision rules under each chapter into second-level title rules;

[0137] S13. If there are more detailed requirements under the second-level heading, further summarize to form more precise third-level heading rules to achieve comprehensive coverage of information granularity; if there are no more detailed requirements under the second-level heading, do not perform this step, that is, do not set up corresponding third-level heading rules;

[0138] S14. In view of the diversity of prospectus title naming, in order to improve the search accuracy and applicability, the present invention designs a series of title aliases for each level of title rules. These aliases are organized into a title set.

[0139] Wherein, step S2 may further include the following steps:

[0140] Step S21: Based on the complete description of the three-level title rules, the relevant description content is accurately disassembled to form a set of specific disclosure rules. Each three-level title rule corresponds to a set of disclosure rules, where each set of disclosure rules contains all disclosure requirements related to a specific three-level title. If there is no corresponding three-level title under the chapter, the second-level title is used to construct the disclosure rules. This design ensures that the applicability and effectiveness of the disclosure rules can be ensured even if the information level is not detailed enough.

[0141] Among them, Figure 4 As shown, Figure 4 The flowchart of step S3 provided in the second embodiment of the disclosure review method of this application is detailed. Step S3 may also include the following steps:

[0142] S31. Use Aspose document parsing tool to parse the Word format prospectus uploaded by the user. This step involves analyzing the title structure and text structure of the prospectus to ensure that its information is accurately captured.

[0143] S32, constructing the titles at all levels into a title knowledge base in the manner described in S1;

[0144] S33, for each disclosure rule obtained in S2, determine whether there is a corresponding third-level title in the title knowledge base. If yes, execute step S331. If not, execute step S332;

[0145] S331, using a word embedding vector model to respectively calculate the relevance between the third-level title and the third-level title alias and each title in the title knowledge base, and selecting the title with the highest relevance as the search result;

[0146] S332, using a word embedding vector model to respectively calculate the relevance of the secondary title and the secondary title alias with each title in the title knowledge base, and selecting the title with the highest relevance as the search result;

[0147] S34. Use the retrieved title to match and locate its corresponding text content in the prospectus.

[0148] Among them, Figure 5 As shown, Figure 5 The flowchart of step S4 provided in the second embodiment of the disclosure review method of this application is detailed. Step S4 may also include the following steps:

[0149] S41. For each disclosure rule in S2, use the relevant text content retrieved in S3, match the customized prompt template, collect relevant records in the historical audit records as a sample set, and divide the sample set into a training set and a test set in a ratio of 9:1;

[0150] S42, fine-tune the open source large model Qwen1.5-14B-Chat using the training set;

[0151] S43. Use the fine-tuned large language model to conduct disclosure review on the input disclosure requirements and prospectus. The review results include the review conclusion (review passed, review failed), judgment basis, and corresponding position information in the prospectus, which is convenient for manual review at a later stage.

[0152] In this embodiment, automation is used to reduce the time required for manual review, improve the accuracy and reliability of the review work, and speed up the review process of the prospectus. The use of large models for automated evaluation effectively avoids the problems of inconsistent review standards and missing key information due to different personal experience and knowledge levels. At the same time, thanks to the long text understanding and processing capabilities of the large model, the present invention can deeply understand complex language structures and subtle semantic differences. In addition, it can automatically perform text retrieval and review tasks according to disclosure requirements, reducing manual intervention.

[0153] It should be noted that the above examples are only used to understand the present application and do not constitute a limitation on the review method disclosed in the present application. More simple transformations based on this technical concept are all within the scope of protection of the present application.

[0154] This application also provides a disclosure review device, please refer to Figure 6 , the disclosure review device comprises:

[0155] The prospectus acquisition module 10 is used to acquire the prospectus to be reviewed;

[0156] A knowledge base construction module 20 is used to construct a knowledge base based on the prospectus to obtain a title knowledge base;

[0157] A model fine-tuning module 30, used to fine-tune a preset large language model through the title knowledge base to obtain a fine-tuned large language model, wherein the large language model is a pre-trained model;

[0158] The disclosure review module 40 is used to perform a disclosure review on the prospectus through the fine-tuned large language model to obtain a disclosure review result.

[0159] Optionally, the prospectus includes several chapters, and the knowledge base construction module 20 is further used to:

[0160] Extract concepts from each chapter in the prospectus and obtain the first-level heading rules;

[0161] Abstract and summarize the subdivision rules of each chapter into secondary heading rules;

[0162] Determining whether the subdivision rules of each chapter contain further subdivision rules;

[0163] If the subdivision rules of each chapter include further subdivision rules, the further subdivision rules are abstracted and summarized into three-level heading rules;

[0164] A knowledge base is constructed based on the first-level title rule, the second-level title rule and the third-level title rule to obtain a title knowledge base.

[0165] Optionally, the model fine-tuning module 30 is further used for:

[0166] According to the prospectus, searching is performed through the title knowledge base to obtain search results;

[0167] Obtaining historical disclosure audit records based on the search results;

[0168] The large language model is fine-tuned according to the retrieval results and the historical disclosure audit records to obtain a fine-tuned large language model.

[0169] Optionally, the title knowledge base includes a number of title rules, and the model fine-tuning module 30 is further used to:

[0170] Constructing the disclosure rules based on the complete narrative content of the several title rules;

[0171] The titles corresponding to the disclosure rules in the prospectus and the titles in the title knowledge base are calculated for relevance by using a preset word embedding vector model, and the titles with the highest relevance in the title knowledge base are obtained as relevant titles;

[0172] The search results are obtained according to the relevant titles and the corresponding text contents in the prospectus.

[0173] Optionally, the model fine-tuning module 30 is further used for:

[0174] Based on a preset prompt word template, generating prompt words according to the search results;

[0175] Generate a sample set based on the prompt words and historical disclosure audit records;

[0176] The sample set is input into the large language model for training and testing, so as to fine-tune the large language model and obtain a fine-tuned large language model.

[0177] Optionally, the disclosure rules include at least one or more of issuance overview, operating conditions and risk factors, the prompt words include at least one or more of relevant titles and text contents, and the disclosure review results include at least one or more of review conclusions, judgment basis, improvement suggestions and text location.

[0178] The disclosure review device provided by the present application adopts the disclosure review method in the above embodiment, which can solve the technical problem of low accuracy of the disclosure review of the prospectus in the prior art. Compared with the prior art, the beneficial effects of the disclosure review device provided by the present application are the same as the beneficial effects of the disclosure review method provided by the above embodiment, and the other technical features in the disclosure review device are the same as the features disclosed in the above embodiment method, which will not be repeated here.

[0179] The present application provides a disclosure review device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the disclosure review method in the above-mentioned first embodiment.

[0180] Reference below Figure 7 , which shows a schematic diagram of the structure of a disclosure review device suitable for implementing the embodiment of the present application. The disclosure review device in the embodiment of the present application may include but is not limited to mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 7 The disclosure review device shown is merely an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.

[0181] like Figure 7 As shown, the disclosure audit device may include a processing device 1001 (e.g., a central processing unit, a graphics processor, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM: Read Only Memory) 1002 or a program loaded from a storage device 1003 to a random access memory (RAM: Random Access Memory) 1004. In RAM1004, various programs and data required for the operation of the disclosure audit device are also stored. The processing device 1001, ROM1002, and RAM1004 are connected to each other through a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Generally, the following systems can be connected to the I / O interface 1006: an input device 1007 including, for example, a touch screen, a touch pad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, etc.; an output device 1008 including, for example, a liquid crystal display (LCD: Liquid Crystal Display), a speaker, a vibrator, etc.; a storage device 1003 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 1009. The communication device 1009 can allow the disclosure audit device to communicate with other devices wirelessly or wired to exchange data. Although the disclosure audit device with various systems is shown in the figure, it should be understood that it is not required to implement or have all the systems shown. More or fewer systems can be implemented or have alternatively.

[0182] In particular, according to the embodiments disclosed in the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present application include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through a communication device, or installed from a storage device 1003, or installed from a ROM 1002. When the computer program is executed by the processing device 1001, the above-mentioned functions defined in the method of the embodiment disclosed in the present application are executed.

[0183] The disclosure review device provided by this application adopts the disclosure review method in the above embodiment, which can solve the technical problem of low accuracy of disclosure review in the prior art prospectus. Compared with the prior art, the beneficial effects of the disclosure review device provided by this application are the same as the beneficial effects of the disclosure review method provided by the above embodiment, and the other technical features in the disclosure review device are the same as the features disclosed in the method of the previous embodiment, which will not be repeated here.

[0184] It should be understood that the various parts disclosed in this application can be implemented by hardware, software, firmware or a combination thereof. In the description of the above embodiments, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0185] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0186] The present application provides a computer-readable storage medium having computer-readable program instructions (ie, computer programs) stored thereon, and the computer-readable program instructions are used to execute the disclosure review method in the above-mentioned embodiment.

[0187] The computer-readable storage medium provided in the present application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared or semiconductor systems, systems or devices, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, system or device. The program code contained on the computer-readable storage medium may be transmitted using any appropriate medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination of the above.

[0188] The computer-readable storage medium may be included in the disclosure review device; or may exist independently without being assembled into the disclosure review device.

[0189] The above-mentioned computer-readable storage medium carries one or more programs. When the above-mentioned one or more programs are executed by the disclosure review device, the disclosure review device: obtains the prospectus to be reviewed; constructs a knowledge base based on the prospectus to obtain a title knowledge base; fine-tunes the preset large language model through the title knowledge base to obtain a fine-tuned large language model, and the large language model is a pre-trained model; and performs a disclosure review on the prospectus through the fine-tuned large language model to obtain a disclosure review result.

[0190] Computer program code for performing the operations of the present application may be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0191] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present application. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a part of a code, and the module, a program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0192] The modules involved in the embodiments of the present application may be implemented by software or hardware, wherein the name of the module does not limit the unit itself in some cases.

[0193] The readable storage medium provided in this application is a computer-readable storage medium, which stores computer-readable program instructions (i.e., computer programs) for executing the above-mentioned disclosure review method, and can solve the technical problem of low accuracy of disclosure review in the prior art prospectus. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as the beneficial effects of the disclosure review method provided in the above-mentioned embodiment, and will not be elaborated here.

[0194] The present application also provides a computer program product, including a computer program, which implements the steps of the disclosure review method as described above when executed by a processor.

[0195] The computer program product provided by this application can solve the technical problem of low accuracy of disclosure review of prospectus in the prior art. Compared with the prior art, the beneficial effects of the computer program product provided by this application are the same as the beneficial effects of the disclosure review method provided by the above embodiment, which will not be repeated here.

[0196] The above descriptions are only some embodiments of the present application, and are not intended to limit the patent scope of the present application. All equivalent structural changes made using the contents of the present application specification and drawings under the technical concept of the present application, or direct / indirect applications in other related technical fields are included in the patent protection scope of the present application.

Claims

1. A disclosure review method, characterized in that: The method comprises: Obtain the prospectus for review; Constructing a knowledge base based on the prospectus to obtain a title knowledge base; Fine-tune the preset large language model through the title knowledge base to obtain a fine-tuned large language model, where the large language model is a pre-trained model; The fine-tuned large language model is used to conduct a disclosure review on the prospectus to obtain a disclosure review result.

2. The method according to claim 1, characterized in that: The prospectus includes several chapters. The step of constructing a knowledge base based on the prospectus to obtain a title knowledge base includes: Extract concepts from each chapter in the prospectus and obtain the first-level heading rules; Abstract and summarize the subdivision rules of each chapter into secondary heading rules; Determining whether the subdivision rules of each chapter contain further subdivision rules; If the subdivision rules of each chapter include further subdivision rules, the further subdivision rules are abstracted and summarized into three-level heading rules; A knowledge base is constructed based on the first-level title rule, the second-level title rule and the third-level title rule to obtain a title knowledge base.

3. The method according to claim 1, characterized in that: The step of fine-tuning the preset large language model through the title knowledge base to obtain the fine-tuned large language model includes: According to the prospectus, searching is performed through the title knowledge base to obtain search results; Obtaining historical disclosure audit records based on the search results; The large language model is fine-tuned according to the retrieval results and the historical disclosure audit records to obtain a fine-tuned large language model.

4. The method according to claim 3, characterized in that: The title knowledge base includes a number of title rules. The step of searching the title knowledge base according to the prospectus to obtain search results includes: Constructing the disclosure rules based on the complete narrative content of the several title rules; The titles corresponding to the disclosure rules in the prospectus and the titles in the title knowledge base are calculated for relevance by using a preset word embedding vector model, and the titles with the highest relevance in the title knowledge base are obtained as relevant titles; The search results are obtained according to the relevant titles and the corresponding text contents in the prospectus.

5. The method according to claim 3, characterized in that: The step of fine-tuning the large language model according to the search results and the historical disclosure audit records to obtain a fine-tuned large language model includes: Based on a preset prompt word template, generating prompt words according to the search results; Generate a sample set based on the prompt words and historical disclosure audit records; The sample set is input into the large language model for training and testing, so as to fine-tune the large language model and obtain a fine-tuned large language model.

6. The method according to any one of claims 4 to 5, characterized in that: The disclosure rules include at least one or more of the issuance overview, operating conditions and risk factors; the prompt words include at least one or more of the relevant titles and text contents; and the disclosure review results include at least one or more of the review conclusions, judgment basis, improvement suggestions and text location.

7. A disclosure review device, characterized in that: The device comprises: The prospectus acquisition module is used to obtain the prospectus to be reviewed; A knowledge base construction module, used to construct a knowledge base based on the prospectus to obtain a title knowledge base; A model fine-tuning module, used to fine-tune a preset large language model through the title knowledge base to obtain a fine-tuned large language model, wherein the large language model is a pre-trained model; The disclosure review module is used to perform a disclosure review on the prospectus through the fine-tuned large language model to obtain a disclosure review result.

8. A disclosure review device, characterized in that: The device comprises: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the disclosure review method according to any one of claims 1 to 6.

9. A storage medium, characterized in that: The storage medium is a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the disclosure review method according to any one of claims 1 to 6 are implemented.

10. A computer program product, characterized in that The computer program product comprises a computer program, and when the computer program is executed by a processor, the steps of the disclosure review method according to any one of claims 1 to 6 are implemented.

Images