Audit report automatic generation method based on natural language processing

Abstract

The invention provides an audit report automatic generation method based on natural language processing, and relates to the technical field of natural languages, and the method comprises the steps: obtaining audit document data passing the examination and approval, the audit document data comprising an audit manuscript, an audit problem, an audit evidence obtaining sheet and an audit verification sheet; performing semantic analysis on the audit document data through a natural language processing model, and extracting key information of an audit subject, a risk point and a compliance conclusion; inputting the key information into a structured processing algorithm to generate structured data matched with a preset auditing template; and inputting the structured data into a pre-training text generation model for language reconstruction, generating an audit report text, automatically filling the audit report text into a corresponding field of an audit template, and outputting a complete audit report.

Description

Technical Field

[0001] This application belongs to the field of natural language technology, and specifically relates to a method for automatically generating audit reports based on natural language processing. Background Art

[0002] In recent years, the rapid development of natural language processing (NLP) technology and computer technology has provided new possibilities for the intelligent transformation of the audit field. In traditional audit work, the generation of audit reports highly depends on manual operations. Auditors need to manually sort, analyze, and write report content based on materials such as audit workpapers, audit issues, evidence collection forms, and verification forms. Although there are already some tools in the existing technology to support electronic document management (such as Word, Excel, etc.), its core process is still dominated by manual work, and there are the following significant defects:

[0003] Audit reports need to check data item by item and be manually written, which is time-consuming and laborious, and it is difficult to meet the needs of large-scale or high-frequency audits; manual operations are easily affected by subjective factors, such as differences in auditors' experience or negligence, which may lead to errors in report content or omission of key risk points; different auditors have different understandings and expression methods for the same audit project, resulting in lack of unified standards for report formats, logic, and conclusions, affecting the controllability of audit quality. Summary of the Invention

[0004] This application provides a method for automatically generating audit reports based on natural language processing to solve the technical problems of low efficiency, insufficient accuracy, and poor consistency in the existing audit report generation technology.

[0005] The technical solution adopted by this application is as follows:

[0006] An embodiment of this application provides a method for automatically generating audit reports based on natural language processing, including:

[0007] Obtain approved audit document data, where the audit document data includes audit workpapers, audit issues, audit evidence collection forms, and audit verification forms;

[0008] Perform semantic parsing on the audit document data through a natural language processing model to extract key information such as audit subjects, risk points, and compliance conclusions;

[0009] Input the key information into a structured processing algorithm to generate structured data that matches a preset audit template;

[0010] Input the structured data into a pre-trained text generation model for language reconstruction, generate an audit report text, and automatically fill it into the corresponding fields of the audit template to output a complete audit report.

[0011] According to an embodiment of the present application, the obtaining of the audited document data approved through review specifically includes: obtaining the audited document data by means of API interface call or batch file import, and performing a review status verification on the audited document data, and only retaining the document content with the review status being the final version.

[0012] According to an embodiment of the present application, after obtaining the audited document data approved through review, it further includes: performing compliance verification through a rule engine, and the rule engine includes:

[0013] Verifying the consistency between the document version and the review process log;

[0014] Verifying the integrity of the association relationship between the audit issues and the evidence collection forms;

[0015] Filtering out the documents that do not contain the signatures of the audit objects.

[0016] According to an embodiment of the present application, the natural language processing model adopts a sequence annotation model based on bidirectional LSTM to identify entities of risk points in the audit document, and extract the logical chain of compliance conclusions based on context dependence relationships.

[0017] According to an embodiment of the present application, the semantic parsing includes eliminating ambiguities in the fuzzy expressions in the audit document, and is specifically implemented by the following methods:

[0018] Constructing a synonym library in the audit field to map non-standard terms to preset standard vocabulary;

[0019] Based on the dependency syntax tree to parse the negative semantic scope and correct the polarity judgment of the compliance conclusion.

[0020] According to an embodiment of the present application, the structured processing algorithm includes a dynamic template adaptation mechanism, specifically:

[0021] According to the chapter structure of the audit template, reorganize the key information in the hierarchy of background description - risk analysis - compliance conclusion;

[0022] If it is detected that the fields corresponding to the key information are missing in the audit template, automatically create extended fields and generate filling guidelines.

[0023] According to an embodiment of the present application, the generation of the structured data further includes:

[0024] Through knowledge graph association technology, cross-verify the extracted audit entities with the associated entities in the historical audit records, and dynamically weight and correct the severity of the risk points.

[0025] According to an embodiment of the present application, the pre-trained text generation model is a domain adaptation model based on the GPT architecture, and the fine-tuning data of the pre-trained text generation model includes:

[0026] Standard expression paragraphs in the regulatory documents of the auditing industry;

[0027] Manually annotated high-frequency compliance description templates in historical audit reports;

[0028] Dataset of matching pairs of audit issues and recommended measures.

[0029] According to an embodiment of the present application, the corresponding fields automatically filled into the audit template include:

[0030] Identify the table area in the audit template through OCR, and inject the numerical results in the structured data according to the row and column coordinates;

[0031] Start an adaptive indentation strategy for text fields that are too long, including automatic pagination, appendix generation, or abbreviation replacement.

[0032] According to an embodiment of the present application, after inputting the structured data into the pre-trained text generation model for language reconstruction, generating the audit report text and automatically filling it into the corresponding fields of the audit template, and outputting the complete audit report, it further includes: automatically verifying the generated complete audit report through a preset audit logic rule library, and the verification includes:

[0033] Verify the logical consistency between the risk point description and the compliance conclusion;

[0034] Check whether the deviation threshold between the numerical data in the audit report and the original records in the audit working papers is within the allowable range;

[0035] If a conflict or overlimit is detected, automatically trigger a revision prompt and locate the error field.

[0036] Due to the adoption of the above technical solutions, the beneficial effects obtained by the present application are:

[0037] The present application automatically obtains approved audit document data (such as audit working papers, audit issues, audit evidence collection forms, and audit verification forms), and combines with a natural language processing model (such as spaCy) for semantic parsing. The system can quickly extract key information such as audit subjects, risk points, and compliance conclusions. Compared with the traditional manual item-by-item sorting and analysis, this method greatly shortens the data processing time, realizes end-to-end automation from data input to preliminary report generation, and the overall efficiency is improved by more than 70%.

[0038] This application uses a structured processing algorithm (such as rule-based text compression and rearrangement logic) to standardize the extracted key information, ensuring that the data format strictly matches the preset audit template. In addition, through a pre-trained text generation model (such as a GPT or BERT derivative model based on deep learning) for language reconstruction, the system can automatically avoid logical contradictions or expression deviations that may occur in manual writing, making the generated audit reports highly consistent in terms of content integrity and term standardization, and significantly reducing the human error rate.

[0039] The audit report text generated by this application is presented through a visual interface (such as an online editable Word document), allowing auditors to dynamically adjust the content according to actual needs. The system retains the relevance between the original structured data and the generation logic, supports secondary configuration of template fields, placeholders, and generation rules, ensuring that the report not only meets the standardization requirements but also can be flexibly adapted to different audit scenarios, enhancing the universality of practical applications.

[0040] The automated process of this application reduces the need for manual intervention, decreases the dependence on highly skilled auditors, and continuously optimizes the model performance through machine learning algorithms (such as decision trees and random forests), avoiding resource waste caused by repetitive work. In addition, the reports generated by the system can be directly stored as structured data (such as PDF or database records), facilitating version management and historical traceability, and reducing the additional costs caused by document loss or version confusion.

[0041] The system of this application has a built-in unified audit report management module, which uses technologies such as Apache POI or Aspose.Words to achieve document format conversion and storage, ensuring that all reports have complete metadata records (such as generation time, modification records, and approval status). This design not only strengthens the compliance of the audit process but also provides a reliable data basis for subsequent audit review and quality assessment. Description of the Drawings

[0042] The drawings described herein are used to provide a further understanding of this application and form a part of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an improper limitation of this application. In the drawings:

[0043] Figure 1 It is a schematic flowchart of a method for automatically generating an audit report based on natural language processing provided by an embodiment of this application. Detailed Embodiments

[0044] To more clearly illustrate the overall concept of this application, the following will be described in detail by way of examples in combination with the drawings of the specification.

[0045] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application. However, the present application may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present application is not limited by the specific embodiments disclosed below. It should be noted that, without conflict, the embodiments of the present application and the features in each embodiment may be combined with each other.

[0046] In the present application, unless otherwise clearly defined and limited, the first feature being "on" or "under" the second feature may be that the first and second features are in direct contact, or the first and second features are in indirect contact through an intermediate medium. In the description of this specification, the description referring to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0047] Embodiment 1

[0048] As Figure 1 shown, a method for automatically generating an audit report based on natural language processing includes:

[0049] S100. Obtain the approved audit document data, where the audit document data includes audit work papers, audit issues, audit evidence collection forms, and audit verification forms.

[0050] Specifically, audit work papers: contain all the basic information collected during the audit, including but not limited to financial statements, accounting records, etc. It is the basic material for the audit work and provides a basic understanding of the financial status of the audited entity.

[0051] Audit issues: refer to the problems or questions found during the audit. These problems usually require further investigation and evidence collection to determine whether there are any violations or misstatements.

[0052] Audit evidence collection forms: When an auditor discovers certain specific problems, evidence will be collected to verify the authenticity of these problems. The audit evidence collection form is the document that records this evidence, and it is crucial for proving the correctness of the audit conclusion.

[0053] Audit verification forms: are used to record the process and results of the auditor's verification of the above problems, ensuring that all audit activities are well-documented and can support the final audit conclusion.

[0054] In actual operation, the content of these documents can be obtained through API interface calls or batch file imports, and the system will perform an approval status verification on these documents, only retaining the content of those documents with the approval status of the final version, that is, the final version that has been reviewed and confirmed to be correct. The purpose of doing this is to ensure the accuracy and reliability of the data input into the natural language processing model and to avoid the impact of un-reviewed information on the quality of the final audit report. In addition, after obtaining the approved audit document data, compliance verification can also be carried out through a rules engine, such as checking the consistency of document versions, the integrity of the association between audit issues and evidence collection forms, etc., so as to further improve the accuracy and consistency of the audit report.

[0055] For example, assume that a company is conducting an annual financial audit. The following are the specific steps to obtain the approved audit document data:

[0056] Collect audit working papers: The audit team will first collect all relevant financial statements, accounting records, etc. of the audited company. These materials form the basis of the audit working papers. For example, an Excel spreadsheet containing the company's monthly income, expenses, and asset-liability situation for the past year will be marked as an audit working paper and, after internal review and confirmation of its accuracy, marked as the final version.

[0057] Identify audit issues: During the preliminary review process, the auditor may find some doubts or anomalies, such as a significant increase in expenses in certain months. These doubts will be recorded as audit issues. For example: "Why is the marketing expense in August 30% higher than in other months?"

[0058] Prepare audit evidence collection forms: For each audit issue, the auditor needs to conduct further investigations and collect evidence. For example, for the above-mentioned marketing expense issue, the auditor may need to review original documents such as relevant contracts, invoices, and payment vouchers. These evidences will be organized into audit evidence collection forms to ensure that each issue is supported by sufficient evidence.

[0059] Complete audit verification forms: After collecting sufficient evidence, the auditor will verify each issue and record the results on the audit verification form. For example, after verifying the evidence related to the marketing expense in August, the auditor confirms that this expense is reasonable and indicates the reasons in the verification form.

[0060] Obtain approved data: All of the above documents (audit working papers, audit issues, audit evidence collection forms, and audit verification forms) need to go through a strict approval process. Only when the status of these documents is marked as "final version", that is, after final review and confirmation of their accuracy, will they be accepted by the system as the basic data for generating the audit report.

[0061] In actual operation, these approved documents can be imported into the system through API interface calls or batch file imports. The system will automatically check the approval status of the documents and only retain the content of those documents with a final approval status.

[0062] S200. Semantically parse the audit document data through a natural language processing model to extract key information such as the audit subject, risk points, and compliance conclusions.

[0063] Specifically, input of audit document data: The data sources include audit workpapers, audit issues, audit evidence collection forms, and audit verification forms, etc. After the approval status of these documents is verified, it is ensured that only the final and verified data is used for subsequent analysis.

[0064] Use NLP technology for text analysis: Lexical analysis: Decompose sentences into words and determine the part of speech and possible meanings of each word. Syntactic analysis: Identify the sentence structure and understand how words are combined to form meaningful phrases or sentences. Semantic analysis: Deeply understand the actual meaning of the text, which involves advanced functions such as entity recognition and relationship extraction.

[0065] Key information extraction

[0066] Audit subject, definition and extraction: The audit subject refers to the main object of the audit work, such as an enterprise or an organization. By constructing a synonym library in the audit field, map non-standard terms to preset standard vocabulary to accurately identify enterprise names, organizational structures, etc. in the document as the audit subject.

[0067] Risk points, definition and extraction: Risk points refer to potential problems found during the audit process or risk areas that need attention. Use a sequence annotation model based on bidirectional LSTM (Long Short-Term Memory Network) to identify entities of risk points in the audit document. This model can accurately find information segments related to risks according to context dependencies and mark them as risk points.

[0068] Compliance conclusions, definition and extraction: Compliance conclusions are the judgment results on whether the audited entity complies with laws, regulations, and relevant provisions. Use dependency syntactic tree parsing to eliminate ambiguities in ambiguous expressions and correct the polarity judgment of compliance conclusions. For example, the system can distinguish positive statements and negative statements, so as to correctly interpret the compliance situation in the document. If the document mentions "no major violations were found", the system should be able to recognize this as a positive compliance conclusion.

[0069] Selection of natural language processing model: Use a sequence labeling model based on bidirectional LSTM for entity recognition, especially for the recognition of risk points. This model can capture context information in the text well, which is crucial for improving the recognition accuracy. Disambiguation: To ensure the accuracy of the information extracted from the text, especially for those expressions that may have multiple interpretations, a method of constructing a synonym library in the audit field is adopted, and the negative semantic scope is parsed based on the dependency syntax tree to correct the judgment of compliance conclusions.

[0070] For example, assume that an annual financial audit is being conducted for an enterprise named "XYZ Company". The following types of documents are collected during the audit: audit workpapers, audit issue lists, audit evidence forms, and audit verification forms. These documents are confirmed as final versions after passing the approval status verification and are ready for subsequent analysis and report generation.

[0071] Semantic parsing and key information extraction

[0072] Recognition of audit entity

[0073] Content of the sample document: "The object of this audit is XYZ Company located in Beijing, which mainly engages in the sales of electronic products."

[0074] Parsing process: The NLP model identifies the entity "XYZ Company" in the sentence as the audit entity through lexical analysis and syntactic analysis. At the same time, based on the preset synonym library in the audit field, the model can correctly map non-standard terms (such as "sales of electronic products") to standard vocabulary, further enriching the information of the audit entity.

[0075] Result: The audit entity is "XYZ Company", and its main business is "sales of electronic products".

[0076] Extraction of risk points

[0077] Content of the sample document: "When reviewing the financial statements of XYZ Company for the third quarter of 2024, it was found that the marketing expenses increased abnormally by 50%, which requires further investigation."

[0078] Parsing process: Use a sequence labeling model based on bidirectional LSTM to analyze the above sentence and identify "the marketing expenses increased abnormally by 50%" as a potential risk point. The model understands the meaning of this expression through context dependency and marks it.

[0079] Result: The identified risk point is "abnormal increase in marketing expenses".

[0080] Extraction of compliance conclusions

[0081] Example document content: "Based on the existing evidence, no major violations were found in XYZ Company in 2024, but it is recommended to strengthen the management of marketing expenses."

[0082] Analysis process: Using the dependency syntactic tree analysis technology, the model first identifies the key words "no major violations were found..." within the scope of negative semantics, and then corrects the polarity judgment of the compliance conclusion. In addition, for the subsequent recommendation part, the model can also identify and record it as supplementary information.

[0083] Result: The compliance conclusion is "no major violations were found", with the attached recommendation of "strengthen the management of marketing expenses".

[0084] In this example, it shows how to use a natural language processing model to perform semantic analysis on audit document data, so as to accurately extract the audit entity (XYZ Company), risk points (abnormal growth of marketing expenses), and compliance conclusions (no major violations were found). In this way, not only the efficiency of the audit work is improved, but also the accuracy and consistency of the audit report are enhanced. This process emphasizes the importance of extracting key information from specific texts without involving other technical features or subsequent steps.

[0085] Furthermore, in order to improve accuracy, a thesaurus or glossary specific to the audit field can be constructed to map non-standard expressions to preset standard vocabulary. For example, different expressions such as "XYZ Group" and "XYZ Co., Ltd." can be correctly identified as the same audit entity.

[0086] Further using the dependency syntactic analysis technology can help the model better understand the logical relationships in complex sentence structures, such as causal relationships and turning relationships, so as to more accurately locate risk points. For example, "Due to the intensification of market competition, the marketing expenses in the third quarter of 2024 increased by 50%", by analyzing the internal relationships of the sentence, the specific reasons for the risk can be more clearly identified.

[0087] In order to handle ambiguous expressions or complex language structures, a negative scope detection mechanism can be introduced, that is, not only identifying a single negative word, but considering all relevant information within the entire scope of negative semantics to ensure the accuracy of the compliance conclusion.

[0088] S300. Input the key information into a structured processing algorithm to generate structured data that matches a preset audit template.

[0089] Specifically, the structured processing algorithm

[0090] Structuring of Key Information, Definition: Structuring processing refers to the transformation of key information extracted from unstructured text (such as audit work papers, audit issue descriptions, etc.) into a data form with clear fields and values. Purpose: To enable the information to be automatically recognized and processed by the system, facilitating subsequent filling into the audit report template.

[0091] Algorithm Workflow

[0092] Data Mapping: Based on the extracted key information (such as the name of the audit entity, identified risk points, compliance conclusions, etc.), map it to the corresponding fields in the preset audit template. For example, "XYZ Company" as the audit entity will be mapped to the "Company Name" field, and "The marketing expense has an abnormal increase of 50%" as a risk point will be mapped to the "Main Risk Point" field.

[0093] Dynamic Template Adaptation Mechanism: If it is detected that a field corresponding to the extracted key information is missing in the audit template, automatically generate an extended field and provide filling guidelines. For example, if there is no dedicated field for "marketing expenses" in the template, the system can automatically create a new field to accommodate this situation.

[0094] Hierarchical Reorganization: According to the chapter structure of the audit template, such as the hierarchical order of background description - risk analysis - compliance conclusion, reorganize the extracted key information. The purpose of doing this is to ensure the logical coherence and structure of the information in the report.

[0095] Knowledge Graph Association Technology: Utilize knowledge graph association technology to cross-verify the extracted audit entity with associated entities in historical audit records, and dynamically weight and correct the severity of risk points. This step helps to improve the accuracy of risk assessment and adjust the risk rating in the current report based on historical data.

[0096] Generate Structured Data

[0097] Result Output: After being processed through the above steps, the originally scattered and difficult-to-directly-use text information is transformed into a data format with clear structure, easy for computers to read and process. For example, data represented in JSON or XML format, which contains clear identifiers (such as field names) and corresponding values.

[0098] Automated Filling: Finally, this structured data can be directly used to fill into the pre-designed audit report template, thereby automatically generating a preliminary draft of the audit report.

[0099] The core of this process lies in how to effectively convert the information obtained from natural language processing into a structured data format, ensuring that the information can be accurately reflected in the final generated audit report while maintaining the logicality and integrity of the report content. The work at this stage is the cornerstone for building an efficient and accurate audit report generation system.

[0100] For example, it is necessary to map the extracted key information to the corresponding fields in the preset audit template. For instance:

[0101] Audit entity ("Audit entity name" field): XYZ Company

[0102] Risk point ("Main risk point" field): The marketing expense has increased abnormally by 50%

[0103] Compliance conclusion ("Compliance conclusion" field): No major violations were found, but it is recommended to strengthen the management of marketing expenses

[0104] If there are no directly corresponding fields in the preset audit template to accommodate certain specific information (such as the specific details of "marketing expenses"), the system will automatically create new fields or adjust existing fields to adapt to this information. For example, a new section named "Special attention matters" can be added to the template to record remarks like "The marketing expenses need further review".

[0105] According to the chapter structure of the audit template, such as the hierarchical order of background description - risk analysis - compliance conclusion, reorganize the extracted key information. For example:

[0106] Background description:

[0107] Company name: XYZ Company

[0108] Main business: Electronic product sales

[0109] Risk analysis:

[0110] Main risk point: The marketing expense has increased abnormally by 50%

[0111] Special attention matters: It is necessary to conduct an in-depth investigation into the reasons for the increase in marketing expenses

[0112] Compliance conclusion:

[0113] Preliminary conclusion: No major violations were found

[0114] Suggested measures: Strengthen the management and supervision of marketing expenses

[0115] Using knowledge graph association technology, cross-validate the extracted audit entities with the associated entities in historical audit records to dynamically adjust the severity of risk points. For example, if historical data shows that XYZ Company has had similar issues with marketing expenses in the past few years, the weight of the current risk point can be increased and this can be emphasized in the report.

[0116] Example of structured data

[0117] The final generated structured data may be as follows (taking JSON format as an example):

[0118]

[0119]

[0120] Automatically fill into the audit template

[0121] The above structured data can be directly used to fill into a pre-designed audit report template to automatically generate a preliminary draft of the audit report. For example, positions for fields such as "Company Name", "Main Risk Points", and "Preliminary Conclusion" are reserved in the template, and the system can automatically identify and fill in the corresponding structured data to form a complete audit report.

[0122] Furthermore, it is also possible to determine which extracted information should be mapped to which field in the template. For example, "XYZ Company" as the audit entity should be mapped to the "Company Name" field in the template. The identified key information is filled as the value into the corresponding field. For instance, "The marketing expense has increased abnormally by 50%" as a risk point should be filled into the "Main Risk Points" field. If it is detected that there is no field in the audit template corresponding to the extracted key information, the system will automatically create an extended field and provide corresponding filling guidelines. For example, if there is no specific field for "marketing expense" in the template, a new field will be automatically generated to record relevant information. According to the chapter structure of the audit template (such as the background description - risk analysis - compliance conclusion hierarchy), the extracted key information is rearranged and combined logically. If a certain piece of information does not fit the existing chapter classification, new chapters can be created or the order of existing chapters can be adjusted. Based on the structure of the audit template, the extracted information is assigned to the appropriate chapters. For example, the basic information about the audit entity is placed in the background description section; the discovered risk points are placed in the risk analysis section, etc. Inside each chapter, the content is sorted according to the importance or logical relationship of the information. For example, within the risk analysis chapter, the most serious risk points are listed first, followed by the less serious ones. The extracted audit entity information is compared with relevant entities in the knowledge graph to check if there are similar historical cases or patterns. This helps to evaluate the authenticity and severity of the currently discovered risk points. Based on the results of cross - verification, the severity of the risk points is dynamically adjusted. For example, if historical data shows that a certain type of risk has occurred frequently in the past, the current evaluation weight can be appropriately increased.

[0123] S400. Input the structured data into a pre - trained text generation model for language reconstruction, generate the audit report text and automatically fill it into the corresponding fields of the audit template, and output the complete audit report.

[0124] Specifically, a pre - trained text generation model refers to those models that have been pre - trained on a large amount of text data. These models can understand complex language structures and context information and are suitable for text generation tasks in specific domains.

[0125] In this scenario, a domain - adaptive model based on the GPT architecture is adopted. This model has been fine - tuned and uses standard expression paragraphs in audit industry regulatory documents, high - frequency compliance description templates manually marked in historical audit reports, and a dataset of matching pairs of audit problems and recommended measures as training data.

[0126] Data Input and Language Reconstruction

[0127] Input Preparation: First, the data obtained from the structured processing algorithm needs to be prepared to ensure that its format and content can be correctly understood and processed by the pre-trained text generation model. For example, if the structured data is in JSON format, it needs to be converted into a form suitable for model input.

[0128] Language Reconstruction: Input the structured data into the pre-trained text generation model and utilize the powerful capabilities of the model to reconstruct the language of this data. This means that the model will generate coherent and professional audit report text based on the input data. For example, for the content in the "Main Risk Points" field (such as "The marketing expenses have increased abnormally by 50%"), the model will generate a detailed description of this risk point, including possible cause analysis, impact assessment, etc.

[0129] Automatically Fill into the Audit Template

[0130] Filling Process

[0131] Field Matching: The system automatically fills the generated audit report text into the corresponding fields of the audit report template according to preset rules or mapping relationships. For example, the "Company Name" field will be automatically filled with "XYZ Company", and the part of "Main Risk Points" will be filled with the specific description of the abnormal increase in marketing expenses.

[0132] Table Area Processing: For the part containing numerical results, the table area in the audit template can be identified through OCR technology, and then the numerical results in the structured data can be accurately injected into the corresponding cells according to row and column coordinates. This ensures the accuracy of the data and the correctness of the position.

[0133] Management of Overlong Fields: When encountering the situation of overly long text, an adaptive indentation strategy can be initiated, including methods such as automatic pagination, appendix generation, or abbreviation replacement to handle overlong fields, ensuring that the overall layout of the report is reasonable and easy to read.

[0134] Output the Complete Audit Report

[0135] Final Check: Before outputting the final audit report, the generated report is usually automatically verified through a preset audit logic rule library to ensure that all key information is correctly included and there are no logical errors.

[0136] Report Generation: Once all the above steps are completed and passed the verification, the system will output a complete audit report. This report not only contains the necessary audit conclusions and identified risk points but also maintains a high degree of professionalism and standardization, and can be directly used for formal submission or further review.

[0137] For example, assume that key information has been extracted from an audit document through a natural language processing (NLP) model and transformed into the following structured data through a structured processing algorithm:

[0138]

[0139] Application Example of Pretrained Text Generation Model

[0140] Data Input and Language Reconstruction

[0141] Input Preparation:

[0142] Provide the above structured data in JSON format as input to a domain adaptation model based on the GPT architecture. This model has been fine-tuned using, as training data, standard statement paragraphs from audit industry regulatory documents, manually annotated high-frequency compliance description templates from historical audit reports, and a dataset of matching pairs of audit issues and recommended measures.

[0143] Language Reconstruction:

[0144] The model generates coherent and professional audit report text based on the input data. For example, for the content of the "Main Risk Points" field (such as "The marketing expenses have increased abnormally by 50%"), the model may generate text similar to the following:

[0145] During this audit, we noticed a significant increase in marketing expenses, specifically, the marketing expenses in the third quarter of 2024 increased by 50% compared to the previous quarter. This growth rate exceeds the expected range and may imply potential risks or management loopholes.

[0146] Similarly, for the "Compliance Conclusion", the model will generate the corresponding paragraph:

[0147] According to our review results, no major violations were found in this company during the current year. However, considering the growth of marketing expenses, it is recommended to strengthen the management and supervision of this part of the expenses to ensure their reasonableness and transparency.

[0148] Automatically Fill into Audit Template

[0149] Field Matching:

[0150] Audit report templates usually contain multiple chapters, such as "Background Description", "Risk Analysis", "Compliance Conclusion", etc. The system will automatically fill the generated audit report text into the corresponding chapters according to preset rules or mapping relationships.

[0151] For example, "Company Name: XYZ Company" will be filled into the "Company Name" field in the "Background Description" section; the relevant description of "The marketing expense has increased abnormally by 50%" will be filled into the "Main Risk Points" field in the "Risk Analysis" section.

[0152] Table Area Processing:

[0153] If the audit report contains data in tabular form, such as the summary of financial statements, the table area in the audit template can be identified through OCR technology, and then the numerical results in the structured data can be accurately injected into the corresponding cells according to the row and column coordinates. For example, the specific value of the marketing expense can be correctly placed in the relevant position of the financial summary table.

[0154] Management of Ultra-Long Fields:

[0155] For the case of ultra-long text, an adaptive indentation strategy is launched, including automatic pagination, appendix generation, or abbreviation replacement, etc. to handle ultra-long fields. For example, if the description of the reason for the increase in marketing expenses is too long, the system can automatically create an appendix section, place the detailed analysis in the appendix, and provide a brief overview and a link to the appendix in the main report.

[0156] Output the Complete Audit Report

[0157] Finally, the audit report processed through the above steps will include the following content:

[0158] Background Description: Introduces the basic situation of the audited entity, such as the company name, main business, etc.

[0159] Risk Analysis: Details the main risk points found and their impacts, such as the analysis of the reasons for the abnormal increase in marketing expenses.

[0160] Compliance Conclusion: Summarizes the results of the audit and puts forward specific improvement suggestions, such as strengthening the management of marketing expenses.

[0161] Furthermore, a domain adaptation model based on the GPT architecture can also be selected. This model not only has powerful natural language processing capabilities, but also can better adapt to the language style and term usage in a specific domain through fine-tuning. For audit report generation, this means that the model can understand and correctly use the professional terms in the audit industry.

[0162] Fine-tuning dataset: Standard statement paragraphs: Include standard statements from audit industry regulatory documents to help the model learn how to accurately express audit findings in formal reports. High-frequency compliance description templates: These templates usually contain manually annotated high-frequency compliance descriptions, enabling the model to identify and generate texts that meet actual needs. Audit issue and recommended measure matching pair dataset: Helps the model learn how to propose reasonable recommended measures based on specific audit issues.

[0163] Data format conversion: Ensure that the data obtained from the structured processing algorithm (such as in JSON format) is correctly converted into a form suitable for model input. This may involve adjusting field names to match the expected input format of the model, or appropriately formatting numerical data.

[0164] Context information construction: To enable the model to generate more coherent and relevant texts, a context containing background information needs to be constructed. For example, before inputting information about "the marketing expense has increased abnormally by 50%", information about the company's business scope and recent market strategies is provided first.

[0165] Natural and fluent text generation: Utilize the powerful capabilities of the pre-trained text generation model to convert structured data into natural, fluent, and logically consistent audit report texts. For example, for the input of "major risk points", the model may generate a detailed description that not only explains the existence of the risk but also analyzes its potential causes and impacts.

[0166] Logical consistency check: During the text generation process, the model considers the logical relationships between the input data to ensure that the generated content is logically correct and consistent. For example, if the compliance conclusion is "no major violations were found", but there are significant risk points, the model will reasonably explain the relationship between the two.

[0167] Exact matching: The system fills the generated audit report text precisely into the corresponding fields of the audit report template according to preset rules or mapping relationships. For example, the "company name" field will be filled with "XYZ Company", and the part about "major risk points" will be filled with the specific description of the abnormal increase in marketing expenses.

[0168] Intelligent adjustment: When encountering situations where certain fields are not suitable for direct filling, the system can appropriately adjust the content as needed. For example, if the space of a certain field is limited but the generated text is long, automatic summarization can be performed to retain the most critical information.

[0169] Table area processing

[0170] Numeric result injection: For parts containing numeric results, the table areas in the audit template are identified through OCR technology, and then the numeric results in the structured data are accurately injected into the corresponding cells according to row and column coordinates. This ensures the accuracy of the data and its correct position in the report.

[0171] Automated typesetting: Considering the design differences of different audit templates, the system needs to have a certain degree of automated typesetting ability to ensure that all information can be correctly displayed even in complex layouts.

[0172] Logical verification: Before outputting the final audit report, the system will automatically verify the generated report to ensure that all key information is correctly included and there are no logical errors. For example, verify whether there is a logical conflict between the risk point description and the compliance conclusion.

[0173] Format correction: Check whether the overall format of the report meets the standard requirements, including details such as font size, line spacing, and page margins, to ensure the professionalism and readability of the report.

[0174] Version control: Establish a unified audit report system for storing, querying, and managing all audit reports. This can not only ensure the version control and traceability of audit reports but also facilitate collaboration and communication among auditors. Document conversion and storage: Use libraries such as Apache POI, Aspose Words, or Spire.Doc in Java to convert Word documents into PDF format and store them in the database for subsequent reference and archiving.

[0175] According to an embodiment of the present application, the obtaining of the approved audit document data is specifically: obtaining the audit document data through API interface call or file batch import, and performing an approval status verification on the audit document data, and only retaining the document content with the approval status of the final version.

[0176] Specifically, obtain the approved audit document data

[0177] Data acquisition method

[0178] API interface call: One method is to automatically obtain the required audit document data by calling the application programming interface (API). This method allows the system to directly interact with the data source storing the audit documents and obtain the latest and approved data in real time.

[0179] File batch import: Another method is to upload multiple audit documents through batch import. This method is applicable to document collections that already exist in local or external storage, and users can choose to upload these files to the system for processing at one time.

[0180] Approval status verification

[0181] Perform approval status verification: Whether the data is obtained through API interface calls or batch file import, it needs to be verified before it is officially used. This means that the system will check whether each document has completed all review processes and is marked as "final".

[0182] Final document content: Only documents whose approval status is clearly marked as "final" will be accepted by the system and used for subsequent natural language processing and audit report generation. The purpose of this is to ensure that all data entered into the system is accurate and to avoid information that has not been finally reviewed affecting the quality of the audit report.

[0183] Assume that there is an audit project that needs to generate an audit report. The relevant documents of the project include audit working papers, audit question list, audit evidence collection form and audit verification form. The following are the specific steps:

[0184] Preparation phase: Confirm that all necessary audit documents have completed the internal review process and are marked as "final". Data acquisition: If you choose the API interface call method, the system will pull relevant audit document data from the specified data source (such as the company's internal database or cloud service) according to the predetermined interface specifications.

[0185] If you use the batch file import method, you must first organize all relevant documents, and then import them all at once using the upload function provided by the system.

[0186] Approval status verification: After obtaining the documents, the system will automatically check the approval status of each document. This usually involves reading the document metadata, which contains information about the document's approval status. For those documents that have not yet completed the final review or whose approval status is not "final", the system will not include them in the subsequent processing flow.

[0187] Screening final version documents: Only when the approval status of a document is confirmed as "final version" will it be retained as the basic material for further analysis and report generation.

[0188] According to an embodiment of the present application, after obtaining the approved audit document data, the method further includes: performing compliance verification through a rule engine, wherein the rule engine includes:

[0189] Verify the consistency of the document version and the approval process log;

[0190] Verify the integrity of the relationship between audit questions and evidence collection forms;

[0191] Filter documents that do not contain the audit subject's signature.

[0192] Specifically, the rules engine performs compliance verification

[0193] Verify the consistency between the document version and the approval process log to ensure that there are no inconsistencies between the document version and its approval process records. This is to prevent the document from being tampered with or the wrong version from being misused.

[0194] Version control check: The system automatically compares whether the current document version number matches the version number recorded in the approval process. For example, if a document has undergone multiple revisions during the approval process, the final version used must be the last approved version.

[0195] Log review: Review the approval history log of the document to ensure that each version has a corresponding approval record and that these records match the actual document content.

[0196] Verify the integrity of the association between audit issues and evidence collection forms to ensure that each audit issue has corresponding evidence support and that this evidence can accurately reflect the nature of the issue. This is crucial for improving the credibility of the audit report.

[0197] Relevance check: The system compares each audit issue with its related evidence collection form to confirm the integrity of the association between the two. For example, for an audit issue such as "abnormal increase in marketing expenses", corresponding contracts, invoices, etc. are required as evidence support.

[0198] Integrity verification: Not only check whether there is an association, but also ensure that the associated information is complete, that is, all audit issues can find corresponding evidence collection forms, and vice versa.

[0199] Filter out documents that do not contain the signature of the audit object to ensure that all documents used to generate the audit report have undergone a formal signing process, enhancing the legal effect and authority of the report.

[0200] Signature detection: The system scans each document to check for the existence of a legal and valid signature. Here, the signature usually refers to the confirmation document signed by the authorized representative of the audit object (such as the audited enterprise).

[0201] Filtering mechanism: Any document that fails to contain a valid signature will be automatically excluded and not included in the subsequent processing flow. This can avoid using unauthorized or unknown source materials, thus ensuring the solemnity and accuracy of the audit work.

[0202] According to an embodiment of the present application, the natural language processing model uses a sequence annotation model based on bidirectional LSTM to perform entity recognition on risk points in the audit document and extract the logical chain of compliance conclusions based on context dependencies.

[0203] Specifically, the sequence labeling model of bidirectional LSTM

[0204] Bidirectional LSTM: The Long Short-Term Memory (LSTM) is a special type of Recurrent Neural Network (RNN) that can learn long-term dependencies. The bidirectional LSTM processes sequence data both forward and backward simultaneously, enabling the model to utilize information before and after a given position, thereby better capturing the context information in the text.

[0205] Sequence labeling model: In natural language processing, the sequence labeling task refers to assigning a label or category to each element in an input sequence. For example, in the Named Entity Recognition (NER) task, the model assigns an entity type label (such as person name, place name, etc.) to each word. For audit documents, the sequence labeling task here is to identify and classify the risk points in the document.

[0206] Perform entity recognition on the risk points in audit documents

[0207] Entity recognition: In this context, an entity refers to a risk point or a key issue mentioned in the audit document. For example, "The marketing expense has increased abnormally by 50%" can be regarded as a risk point entity.

[0208] Training data preparation: First, a dataset containing a large number of labeled samples needs to be constructed, where each sample contains a piece of text and the corresponding risk point annotation. These annotations can help the model learn how to identify new unseen risk points.

[0209] Feature extraction: Using the bidirectional LSTM model, the system can extract rich feature representations from the input text, especially those features that can reflect context dependencies.

[0210] Prediction and annotation: Based on the trained model, the system can identify and annotate the risk points in the newly input audit document. The model will output the probability distribution of whether each word or phrase belongs to a certain risk point and make a final annotation decision accordingly.

[0211] Extract the logical chain of compliance conclusions based on context dependencies

[0212] Context dependencies: The advantage of the bidirectional LSTM model is that it can effectively utilize the information in the context, which is very important for understanding complex sentence structures and implicit meanings. For example, when analyzing the sentence "Although the marketing expense has increased, no major violations have been found", the model needs to understand the logical relationship of "although...but" to correctly interpret the compliance conclusion.

[0213] Dependency syntactic analysis: To gain a deeper understanding of the text meaning, in addition to simple sequence labeling, dependency syntactic analysis techniques can also be combined. This helps to identify the relationships within a sentence, such as causal relationships, adversative relationships, etc., so as to more accurately extract the logical chain behind the compliance conclusion.

[0214] Logical reasoning: Based on the extracted context dependencies, the system can infer the specific logic of the compliance conclusion. For example, if the document mentions that "although some expenses exceed the budget, all expenditures have legal bases", then the system should be able to identify this as a positive compliance conclusion and record its logical chain.

[0215] According to an embodiment of the present application, the semantic parsing includes eliminating the ambiguity of the fuzzy expressions in the audit document, which is specifically achieved by the following methods:

[0216] Construct a synonym library for the audit field to map non-standard terms to preset standard vocabulary;

[0217] Based on the dependency syntactic tree, parse the negative semantic scope to correct the polarity judgment of the compliance conclusion.

[0218] Specifically, the elimination of the ambiguity of fuzzy expressions

[0219] Construct a synonym library for the audit field to solve the problem of inconsistent understanding caused by using different terms or expressions, and ensure that all extracted information uses unified standard terms.

[0220] Synonym library construction: Create a synonym library specifically for the audit field, which contains different expressions of common terms and their corresponding standardized terms. For example, "sales expenses", "marketing expenditures", etc. can be mapped to the standard term "marketing expenses".

[0221] Term mapping: During the semantic parsing process, the system will automatically identify non-standard terms in the text and convert them into preset standard vocabulary. This not only improves the consistency of information but also facilitates subsequent data analysis and report generation.

[0222] Based on the dependency syntactic tree, parse the negative semantic scope to accurately understand sentences containing negative meanings, especially those parts related to the compliance conclusion, and avoid making wrong judgments due to misunderstandings.

[0223] Dependency syntactic analysis: Use the Dependency Parsing Tree to parse the sentence structure and determine the dependency relationships between various components. This is particularly important for understanding complex sentences, such as sentences containing multiple negative words or negative scopes.

[0224] Negative semantic scope recognition: By analyzing the dependency syntactic tree, the system can identify the scope of action of negative words (such as "not", "none", "without", etc.) and adjust the polarity of the compliance conclusion accordingly. For example, in the sentence "No major violations were found", the negative word "no" acts on the entire phrase "found major violations", indicating a positive compliance conclusion.

[0225] Suppose there is the following statement in an audit document: "Although the marketing expenses of some projects have increased, no violations of company policies have been found."

[0226] Term mapping: The system first scans the text to identify the term "marketing expenses" and determines whether it needs to be mapped to other forms of expression in the thesaurus (such as "sales expenses", "marketing expenditures"). In this example, "marketing expenses" is already a standard term, so no further conversion is required.

[0227] Negative semantic scope analysis: Next, the system uses the dependency syntactic tree to analyze the sentence structure, especially focusing on the scope of action of the negative word "no". Through analysis, "no" acts on "found any violations of company policies" here, which means this part is a positive compliance conclusion.

[0228] Based on this, the system can correctly extract the compliance conclusion - that is, no major violations were found - and accurately reflect it in the final audit report.

[0229] According to an embodiment of the present application, the structured processing algorithm includes a dynamic template adaptation mechanism, specifically:

[0230] Reorganize the key information according to the chapter structure of the audit template at the level of background description - risk analysis - compliance conclusion;

[0231] If it is detected that a field corresponding to the key information is missing in the audit template, automatically create an extended field and generate a filling guide.

[0232] Specifically, the dynamic template adaptation mechanism

[0233] Reorganize key information according to the chapter structure of the audit template

[0234] Chapter structure: An audit report usually contains multiple chapters or sections, such as background description, risk analysis, and compliance conclusion. Each section has its specific content and format requirements.

[0235] Background description: Provide basic information about the audited entity, such as company name, main business, etc.

[0236] Risk analysis: List the main risk points found and their impact assessments.

[0237] Compliance conclusion: Summarize the audit results and put forward improvement suggestions.

[0238] Key information for reorganization:

[0239] Hierarchical reorganization: After obtaining the key information extracted from the natural language processing model, the system reorganizes this information according to the chapter structure of the audit template (such as background description - risk analysis - compliance conclusion). For example, if the extracted information includes "XYZ Company" as the audit subject, "Abnormal 50% increase in marketing expenses" as the main risk point, and "No major violations found" as the preliminary conclusion, the system will classify this information into the corresponding chapters respectively.

[0240] Logical sorting: Inside each chapter, the system also sorts the information according to the importance or logical relationship of the information. For example, in the risk analysis section, the most serious risk points are listed first, followed by the secondary risk points.

[0241] Automatically create extended fields and generate filling guidelines

[0242] Detect missing fields: During the reorganization process, the system checks whether the current audit template contains all the key information fields that need to be displayed. If it is found that some key information does not have a corresponding field position (that is, the relevant field is missing in the template), the system will take measures to supplement it.

[0243] Automatically create extended fields:

[0244] Identify requirements: When it is detected that the template is missing the corresponding field for a certain key information, the system will automatically generate a new field to accommodate this information. For example, assume that there is no dedicated field for "marketing expenses" in the audit template, and the system can automatically create a new field named "Details of marketing expenses".

[0245] Generate filling guidelines:

[0246] Guide user operation: To ensure the correct use of the newly added fields, the system will generate detailed filling guidelines. These guidelines may include descriptions of field usage, examples of recommended filling content, etc. For example, for the newly created "Details of marketing expenses" field, the system may prompt the user to enter specific expense amounts, time ranges, and relevant contract numbers, etc.

[0247] Flexibility and scalability: This mechanism not only improves the flexibility of the system, enabling it to adapt to various different audit requirements, but also ensures that the finally generated audit report has a high degree of integrity and professionalism.

[0248] Suppose in an audit project, the extracted key information is as follows:

[0249] Audit entity: XYZ Company

[0250] Main risk point: The marketing promotion expenses have increased abnormally by 50%

[0251] Compliance conclusion: No major violations were found, but it is recommended to strengthen the management of marketing promotion expenses

[0252] Key information on restructuring:

[0253] The system will classify "XYZ Company" into the "Background description" section.

[0254] Classify "The marketing promotion expenses have increased abnormally by 50%" into the "Risk analysis" section and sort it according to importance.

[0255] Classify "No major violations were found, but it is recommended to strengthen the management of marketing promotion expenses" into the "Compliance conclusion" section.

[0256] Automatically create extended fields and generate filling guidelines:

[0257] If it is found that there is no dedicated field in the existing template for recording the reasons for the increase in marketing promotion expenses, the system will automatically create a new field, such as "Analysis of the reasons for the increase in marketing promotion expenses".

[0258] For this new field, the system will generate filling guidelines, such as: "Please elaborate on the specific reasons leading to the increase in marketing promotion expenses, including but not limited to factors such as market strategy adjustments and changes in the competitive environment."

[0259] According to an embodiment of the present application, the generation of the structured data further includes:

[0260] Through knowledge graph association technology, cross-validate the extracted audit entity with the associated entities in the historical audit records, and dynamically weight and correct the severity of the risk points.

[0261] Specifically, a knowledge graph is a graphical database used to represent entities and their interrelationships. In this context, cross-validation refers to comparing and validating the audit entity (such as company name, business scope, etc.) extracted in the current audit with the relevant entities in the historical audit records.

[0262] Data integration: First, the system needs to integrate data from different sources, including but not limited to data from the current audit project, historical audit records, and industry standards or regulatory requirements, etc.

[0263] Entity matching: Identify and match the current audit entity with relevant entities in historical records through algorithms. For example, if the current audit is of "XYZ Company", the system will search for all past audit records related to "XYZ Company".

[0264] Dynamic weighted correction: Based on historical data and other relevant information, adjust the severity score of the risk points found in the current audit. This helps to more accurately reflect the actual risk level and provides a basis for subsequent decision-making.

[0265] Weight calculation: Based on similar cases in historical records, the system can calculate the impact weights of different factors on the severity of risk points. For example, if a company has had abnormal growth in marketing expenses similar to the current situation multiple times in the past, the weight of this type of risk point may be increased.

[0266] Real-time update: As new data is continuously added, the system can dynamically adjust these weight values to ensure that the risk assessment results are always up-to-date and accurate.

[0267] Suppose in an audit project, the key information extracted is as follows:

[0268] Audit entity: XYZ Company

[0269] Main risk point: The marketing expenses have increased abnormally by 50%

[0270] Cross-verification:

[0271] The system will first search the knowledge graph for all historical audit records related to "XYZ Company". Suppose it is found that the company has had two similar situations in the past three years - the marketing expenses increased by 30% and 40% respectively, and financial irregularities were indeed discovered after one of the events.

[0272] Dynamic weighted correction: Based on the above historical data, the system will re-evaluate the severity of the current risk point of "the marketing expenses have increased abnormally by 50%". Considering the two similar situations in the past and the fact that one of them led to irregularities, the system may increase the weight of this risk point.

[0273] Specifically, the system may introduce a dynamic weighting factor. For example, if the original risk point score was set at 5 points (out of 10), but due to historical records showing a high risk of irregularities in similar situations, the system automatically raises the score to 7 points.

[0274] This dynamic adjustment not only considers the quantitative growth percentage but also combines the qualitative historical performance and potential consequences, making the final risk assessment more comprehensive and accurate.

[0275] According to an embodiment of the present application, the pre-trained text generation model is a domain adaptation model based on the GPT architecture, and the fine-tuning data of the pre-trained text generation model includes:

[0276] Standard statement paragraphs in the regulatory documents of the audit industry;

[0277] Manually annotated high-frequency compliance description templates in historical audit reports;

[0278] A dataset of matching pairs of audit issues and recommended measures.

[0279] Specifically, the pre-trained text generation model

[0280] A domain adaptation model based on the GPT architecture

[0281] GPT architecture: GPT (Generative Pre-trained Transformer) is an advanced natural language processing model architecture that learns general knowledge of language through large-scale pre-training and can be fine-tuned to adapt to specific domain tasks.

[0282] Domain adaptation: It refers to further training (fine-tuning) a model (such as GPT) that has been pre-trained with a large amount of general data using specific domain data to enable it to better understand and generate text within that domain. In this example, the model is specifically adjusted to adapt to the language and norms of the audit industry.

[0283] Standard statement paragraphs in the regulatory documents of the audit industry to ensure that the generated audit reports comply with industry standards and regulatory requirements, enhancing the professionalism and compliance of the reports.

[0284] Data collection: Extract paragraphs containing standard statements from various regulatory documents in the audit industry, which may include laws, regulations, industry guidelines, etc.

[0285] Application: Use these standard statement paragraphs to fine-tune the model so that the model can imitate and generate a language style that complies with official standards. For example, for how to describe the review results or compliance statements of financial statements, the model can refer to these standard statements to ensure accuracy and consistency.

[0286] Manually annotated high-frequency compliance description templates in historical audit reports to improve the relevance and practicality of the text generated by the model, especially for common audit findings and conclusions.

[0287] Data preparation: Select those compliance description templates that are manually annotated as high-frequency occurrences from past real audit reports. These templates usually cover different types of risk points and corresponding compliance evaluations.

[0288] Fine-tuning process: By providing these high-frequency templates to the model as learning materials, the model is enabled to learn to recognize and generate similar compliance descriptions. For example, when encountering a risk point such as "abnormal growth of marketing expenses", the model can put forward reasonable compliance conclusions and suggestions based on the templates learned previously.

[0289] The matching pair dataset of audit issues and recommended measures helps the model understand how to put forward effective improvement suggestions according to specific audit issues, thus enhancing the practical value of the report.

[0290] Data construction: Create a dataset containing audit issues and their corresponding recommended measures. Each entry consists of a question and at least one related recommendation.

[0291] Train the model: Use this dataset to train the model so that it can automatically recommend appropriate improvement measures when facing new audit issues. For example, if it is detected that a certain company has problems with insufficient internal control, the model can recommend specific measures to strengthen the internal audit process based on the knowledge it has learned.

[0292] According to an embodiment of the present application, the corresponding fields automatically filled into the audit template include:

[0293] Identify the table area in the audit template through OCR, and inject the numerical results in the structured data according to the row and column coordinates;

[0294] Start an adaptive indentation strategy for text overlong fields, including automatic pagination, appendix generation, or abbreviation replacement.

[0295] Specifically, automatically fill into the audit template

[0296] Identify the table area in the audit template through OCR to ensure that the numerical results can be accurately placed in the correct position of the audit report template, especially when these data need to be filled into a table.

[0297] Table area identification: Use optical character recognition (OCR) technology to identify the table area in the audit template. OCR technology can analyze image files (such as PDF format templates) and identify the text and table layout therein.

[0298] Row and column coordinate mapping: Once the table is identified, the system will determine the specific position (i.e., row and column coordinates) of each cell. Then, according to the predefined rules or mapping relationships, inject the numerical results (such as specific financial figures) in the structured data into the corresponding cells according to the correct row and column coordinates.

[0299] Inject the numerical results in the structured data according to the row and column coordinates

[0300] Data Preparation: Extract all the numerical results that need to be filled into the table from the data obtained by the structured processing algorithm, and record their corresponding field names or identifiers.

[0301] Precise Filling: Using the row and column coordinate information provided by OCR technology, the system can accurately fill these numerical results into the specified positions in the audit template. For example, if a certain audit template requires filling in the specific amount of "Marketing Expenses in the First Quarter of 2024", the system will find the corresponding table area and fill in its value in the appropriate cell.

[0302] Activate the adaptive indentation strategy for text fields that are too long, to solve the typesetting problems caused by some text fields being too long, and ensure that the final generated audit report is both beautiful and easy to read.

[0303] Automatic Pagination: For those text fields that exceed the page limit in length, the system can automatically insert page breaks, so that some content is moved to the next page for display. This helps to keep the page layout clean.

[0304] Appendix Generation: If a certain piece of text is very long and not suitable to be directly placed in the main report, the system can choose to move it to the appendix section. The appendix is usually located at the end of the report and is used to store supplementary materials or detailed explanations.

[0305] Abbreviation Replacement: Another method is to appropriately abbreviate long text. For example, replace long professional terms with short abbreviations, or summarize key points, while providing links or references to the location of the complete content.

[0306] Suppose in an audit project, the key information extracted is as follows:

[0307] Financial Data: The marketing expenses in the first quarter of 2024 are 500,000 yuan.

[0308] Risk Analysis: The analysis of the reasons for the abnormal increase in marketing expenses and the impact assessment are very detailed and exceed the single-page limit.

[0309] Identify the table area in the audit template through OCR:

[0310] The system first uses OCR technology to scan the audit template and identify the table area containing financial data and its row and column coordinates.

[0311] Then, fill the corresponding value of 500,000 yuan for "Marketing Expenses in the First Quarter of 2024" accurately into the table according to the identified row and column coordinates.

[0312] Activate the adaptive indentation strategy for text fields that are too long:

[0313] For the content of "Analysis of the Reasons for the Abnormal Increase in Marketing Expenses and Impact Assessment", since its length exceeds the single-page limit, the system will take one of the following measures:

[0314] If automatic pagination is selected, page breaks will be inserted at appropriate positions so that some content continues to be displayed on the next page.

[0315] If generating an appendix is selected, a separate section will be created, perhaps titled "Analysis of the Reasons for the Abnormal Increase in Marketing Expenses", and the detailed content will be placed in the appendix, with only a brief summary and a link to the appendix retained in the main report.

[0316] If abbreviation replacement is adopted, key findings may be summarized, such as "The abnormal increase in marketing expenses is mainly caused by the expansion of new markets", and more detailed explanations will be provided in footnotes or annotations.

[0317] According to an embodiment of the present application, after inputting the structured data into the pre-trained text generation model for language reconstruction, generating the audit report text and automatically filling it into the corresponding fields of the audit template, and outputting the complete audit report, it further includes: automatically verifying the generated complete audit report through a preset audit logic rule library, and the verification includes:

[0318] Verifying the logical consistency between the risk point description and the compliance conclusion;

[0319] Checking whether the deviation threshold between the numerical data in the audit report and the original records in the audit working papers is within the allowable range;

[0320] If a conflict or overlimit is detected, automatically trigger a revision prompt and locate the error field.

[0321] Specifically, the automatic verification mechanism

[0322] The preset audit logic rule library, which is a database or rule set containing a series of predefined rules for checking whether various elements in the audit report conform to logical consistency, numerical accuracy, and other quality standards. Through a systematic inspection process, ensure that the generated audit report is not only in the correct format but also accurate in content.

[0323] Verifying the logical consistency between the risk point description and the compliance conclusion to ensure that there is no logical conflict between the risk points mentioned in the audit report and their corresponding compliance conclusions.

[0324] Logical relationship analysis: The system analyzes the description of each risk point and its corresponding compliance conclusion to determine whether the logical relationship between the two is reasonable. For example, if a risk point states that "the marketing expenses have increased abnormally by 50%", but the compliance conclusion is "no violations were found", in such a case, further explanation or adjustment of the conclusion may be required to maintain logical consistency.

[0325] Conflict detection: Once an inconsistent situation is detected, the system marks it and prompts for revision. This helps prevent incorrect conclusions caused by misunderstandings or information omissions.

[0326] Check whether the deviation thresholds between the numerical data in the audit report and the original records in the audit working papers are within the allowable range to ensure that all numerical data in the audit report (such as financial data, statistical data, etc.) match the records in the original audit working papers and the deviation is within the acceptable range.

[0327] Data comparison: The system automatically compares all numerical data appearing in the audit report with the corresponding data in the original audit working papers. This includes but is not limited to key financial indicators such as revenue, expenditure, and asset value.

[0328] Deviation calculation: For each pair of compared data, the system calculates its relative deviation. If the deviation exceeds a pre-set threshold (e.g., ±5%), it is considered to be out of limit.

[0329] Threshold setting: The threshold can be set according to specific business requirements and industry standards to ensure that it is neither too lenient nor too strict.

[0330] If a conflict or out-of-limit situation is detected, automatically trigger a revision prompt and locate the error field, providing immediate feedback to help auditors quickly identify and correct problems in the report.

[0331] Revision prompt: When the system detects a logical conflict or a numerical deviation exceeding the allowable range, it automatically generates a revision prompt. These prompts can be presented to the user in the form of pop-up windows, message notifications, or other forms.

[0332] Error location: In addition to the prompt, the system also precisely locates the specific error field or paragraph, enabling the user to directly jump to the problematic part for viewing and modification.

[0333] Revision suggestion: In some cases, the system can also provide preliminary revision suggestions to help users solve problems more quickly. For example, for numerical deviations, the system may suggest re-checking the relevant data sources or adjusting the numerical values in the report.

[0334] Suppose in an audit project, the following content appears in the generated audit report:

[0335] Risk point description: "The marketing expenses have increased abnormally by 50%."

[0336] Compliance conclusion: "Although the marketing expenses have increased, no major violations have been found."

[0337] At the same time, a certain financial data listed in the report is "The net profit in the first quarter of 2024 is 800,000 yuan", while the original record in the audit working papers shows that this value should be "760,000 yuan".

[0338] Verify logical consistency: The system will analyze the relationship between the risk point description and the compliance conclusion. Although there seems to be a contradiction on the surface (a 50% increase but no major violations), if it is further explained that it is due to a reasonable market expansion strategy rather than improper behavior, it can be regarded as logically consistent. Otherwise, the system will prompt that additional explanations or adjustments to the conclusion are needed.

[0339] Check the numerical deviation: The system calculates the deviation between the net profit in the report and the record in the working papers as

[0340] (800,000 - 760,000) / 760,000×100% = 5.26%. If the preset deviation threshold is ±5%, then this deviation is considered to exceed the limit.

[0341] The system will automatically generate a revision prompt and precisely point to the field of "Net profit in the first quarter of 2024" in the report, reminding the user to check and correct the data.

[0342] Automatically trigger the revision prompt: For the above two examples, the system will trigger the revision prompt and provide detailed error location. The user can directly jump to the problem location by clicking on the prompt to make necessary modifications or confirmations.

[0343] What is not described in this application can be achieved by adopting or referring to the existing technology.

[0344] Each embodiment in this specification is described in a progressive manner. For the same or similar parts between each embodiment, reference can be made to each other. Each embodiment focuses on the differences from other embodiments.

[0345] The above are only the embodiments of this application and are not used to limit this application. For those skilled in the art, this application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of this application shall be included within the scope of the claims of this application.

Claims

1. A method for automatically generating an audit report based on natural language processing, characterized in that: include: Acquire approved audit document data, including audit working papers, audit questions, audit evidence collection forms and audit verification forms; Perform semantic analysis on the audit document data through a natural language processing model to extract key information on the audit subject, risk points and compliance conclusions; Inputting the key information into a structured processing algorithm to generate structured data matching a preset audit template; The structured data is input into a pre-trained text generation model for language reconstruction, the audit report text is generated and automatically filled into the corresponding fields of the audit template, and a complete audit report is output.

2. The method according to claim 1, characterized in that The obtaining of the audit document data that has been approved specifically includes: obtaining the audit document data through an API interface call or a batch file import method, and performing an approval status check on the audit document data, and only retaining the document content with a final approval status.

3. The method according to claim 2, characterized in that After obtaining the audit document data that has been approved, the method further includes: performing compliance verification through a rule engine, wherein the rule engine includes: Verify the consistency of the document version and the approval process log; Verify the integrity of the relationship between audit questions and evidence collection forms; Filter documents that do not contain the audit subject's signature.

4. The method according to claim 1, characterized in that The natural language processing model adopts a bidirectional LSTM-based sequence labeling model to perform entity recognition on risk points in audit documents and extract the logical chain of compliance conclusions based on contextual dependencies.

5. The method according to claim 1, characterized in that The semantic analysis includes disambiguating the ambiguous expressions in the audit document, which is specifically achieved in the following ways: Build a synonym database in the audit field to map non-standard terms to preset standardized vocabulary; Based on the dependency syntactic tree, the semantic scope of the negative is parsed and the polarity judgment of the compliance conclusion is corrected.

6. The method according to claim 1, characterized in that The structured processing algorithm includes a dynamic template adaptation mechanism, specifically: According to the chapter structure of the audit template, the key information is reorganized into the hierarchy of background description-risk analysis-compliance conclusion; If it is detected that the fields corresponding to the key information are missing in the audit template, extended fields are automatically created and filling instructions are generated.

7. The method according to claim 1, characterized in that The generation of the structured data further includes: Through knowledge graph association technology, the extracted audit subjects are cross-validated with related entities in historical audit records, and the severity of risk points is dynamically weighted and corrected.

8. The method according to claim 1, characterized in that The pre-trained text generation model is a domain adaptive model based on the GPT architecture. The fine-tuning data of the pre-trained text generation model includes: Standardized paragraphs in audit industry regulatory documents; Templates for manually annotated high-frequency compliance descriptions in historical audit reports; A dataset of matched pairs of audit questions and recommended actions.

9. The method according to claim 1, characterized in that: The corresponding fields automatically filled into the audit template include: Use OCR to identify the table area in the audit template and inject the numerical results in the structured data according to row and column coordinates; Enable adaptive indentation strategies for extra-long text fields, including automatic page breaks, appendix generation, or abbreviation replacement.

10. The method according to claim 1, characterized in that After inputting the structured data into the pre-trained text generation model for language reconstruction, generating the audit report text and automatically filling it into the corresponding field of the audit template, and outputting the complete audit report, the method further includes: automatically verifying the generated complete audit report through a preset audit logic rule library, wherein the verification includes: Verify the logical consistency between risk point description and compliance conclusion; Check whether the deviation threshold between the numerical data in the audit report and the original records in the audit working papers is within the allowable range; If a conflict or limit violation is detected, a revision prompt is automatically triggered and the error field is located.

Images