A large model-based automatic audit report generation full-process method

Abstract

The application discloses an automatic audit report generation full-process method based on a large model. The application avoids the illusion and content omission problems that are easily caused by a single large language model when processing massive heterogeneous data and complex logic at one time by decomposing a complex report preparation task into multiple independent and professional subtasks, combining the dynamic coordination of the task scheduling matrix, introducing a prompt engineering and post-check mechanism, and strictly correcting the logic, smoothing the text, and standardizing the professional terms of the preliminary audit text. The application filters the possible rough expressions and logical jumps in the content generated by the large language model, ensures that the target audit report file produced has factual accuracy and logical consistency, and guarantees the uniqueness, tamper resistance, and full-process version traceability of the target audit report file issued through the automatic rendering of the document structure tree and the generation of digital fingerprints, so as to meet the requirements of audit archive compliance management.

Description

Technical Field

[0001] This invention relates to the field of audit report generation technology, specifically to a method for automated audit report generation based on a large model. Background Technology

[0002] With the exponential expansion of enterprise operations and the deepening of digital transformation, the data environment faced by auditing work has undergone fundamental changes. The massive amount of unstructured verification documents such as scanned contracts, on-site evidence photos, and interview records, coupled with increasingly stringent audit quality standards, constitutes an irreconcilable contradiction. In this complex operational scenario, the traditional audit report preparation model has gradually revealed its limitations. This model is essentially a "piecework" operation based on human experience. Auditors need to manually extract key information from scattered documents and then fill them into a fixed Word template word by word. This method is not only time-consuming and laborious, but also prone to data citation errors or logical inconsistencies due to human negligence.

[0003] While large language models possess powerful text generation capabilities, their inherent "illusion" problem is fatal in serious auditing scenarios. If large language models are directly used to write reports, there is a high possibility of fabricating facts, citing non-existent regulations, or making arbitrary interpretations.

[0004] Therefore, inventing a fully automated audit report generation method based on a large model, which leverages the powerful semantic capabilities of the large model while constraining its generation behavior through a rigorous engineering framework to ensure the factual accuracy and logical consistency of the generated audit report documents, is a core technical challenge that urgently needs to be solved in the field of intelligent auditing. Summary of the Invention

[0005] The purpose of this invention is to provide a fully automated audit report generation method based on a large model, which automates the entire process of audit report generation from draft extraction to final output, significantly improving the efficiency and standardization of report preparation, promoting the intelligent transformation of audit operations, and addressing the aforementioned shortcomings in the technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for generating automated audit reports based on a large model, comprising the following steps:

[0007] Step 1: Construct a multi-agent collaborative framework to decompose the complex report compilation task into sub-tasks that can be executed independently;

[0008] Step 2: Start the data reading workflow, perform multimodal analysis and feature extraction on heterogeneous audit working paper data, and construct a full high-dimensional feature vector set for the audit project;

[0009] Step 3: Execute the content matching workflow, perform semantic understanding on the features of the draft, calculate its matching degree with the preset report template slots, and generate the factual context of the draft;

[0010] Step 4: Run the report generation workflow, and guide the large language model to generate structured preliminary audit text through prompts and engineering.

[0011] Step 5: Trigger the language optimization workflow, segment the initial audit text, build a context consistency verification model to rewrite the segmented text in multiple rounds, and output the final text set of each chapter;

[0012] Step 6: Construct a document structure tree, map the optimized text set of each chapter to the document structure tree for format rendering, generate the target audit report file, and establish a version anti-counterfeiting and traceability mechanism for the entire lifecycle of the audit report.

[0013] The aforementioned method for automated audit report generation based on a large model, in step 1, constructs a multi-agent collaborative framework to decompose the complex report preparation task into independently executable sub-tasks. The specific steps are as follows:

[0014] 1.1 Determine the intelligent agent workflow set based on the audit business logic. The specific formula is as follows:

[0015] ;

[0016] in, For data reading workflow, Match content to workflow, To generate a workflow for the report, Optimize workflows for languages; the collection of agent workflows covers the entire lifecycle from data input to report output;

[0017] 1.2 Construct a task scheduling matrix to monitor the status of each workflow in real time. The specific formula is as follows:

[0018] ;

[0019] in, Indicates the current time step The system scheduling decision result, that is, the next best workflow selected for execution; It is a function that maximizes the parameters; For pre-trained task scheduling probability models; These are candidate agent workflow variables; Indicates the previous time step System execution status feedback; This indicates the global context status information of the current audit project, including the amount of data processed and the current report completion status.

[0020] The aforementioned automated audit report generation method based on a large model, in step 2, initiates a data reading workflow to perform multimodal analysis and feature extraction on heterogeneous audit working paper data, constructing a high-dimensional feature vector set for the entire audit project. The specific steps are as follows:

[0021] 2.1. By integrating optical character recognition and natural language processing technologies, the multi-source heterogeneous manuscripts are structured and analyzed to identify and separate tables, signatures, and text. Based on the separation results of each region, key entity information of the project is extracted.

[0022] Among them, multi-source heterogeneous drafts include tables, scanned copies, and documents;

[0023] 2.2 Data Reading Workflow Defined as a multimodal feature mapping process, it uses a feature encoding function to extract semantic and spatial features from key entity information of the project, and converts these into a high-dimensional feature vector set of the entire audit project that incorporates layout information. The specific formula is as follows:

[0024] ;

[0025] in, For feature encoding function, This is the processing function for the data reading workflow; This is a collection of heterogeneous audit working papers for input. This is the full set of high-dimensional feature vectors for the audit project; For aggregation or concatenation operations on sets; This refers to the total number of heterogeneous audit working papers entered into the audit project. This is the index number of the draft document, with a value ranging from 1 to... ; Indicates the first The table structure data identified in step 2.1 in the draft includes cell content and its corresponding row and column logical relationships; Indicates the first Visual features of signatures identified in the original draft; Indicates the first The unstructured main text in the draft.

[0026] The aforementioned automated audit report generation method based on a large model, in step 3, executes a content matching workflow, performs semantic understanding on the characteristics of the audit report, calculates its matching degree with the preset report template slots, and generates the factual context of the audit report. The specific steps are as follows:

[0027] 3.1 Parse the preset standardized audit report template, extract the natural language description text of the current text slot to be filled, and call the feature encoding function to map it into a semantic feature vector of the template slot. The specific formula is as follows:

[0028] ;

[0029] in, This refers to the natural language description text of the current text slot to be filled after parsing the preset standardized audit report template; the working paper features refer to the extracted set of key semantics, content attributes and core elements inherent in the business working paper itself, which are used to perform semantic matching and align template slots for the large model;

[0030] 3.2 Using template slot semantic feature vectors As a query condition, in the full high-dimensional feature vector set of the audited projects A global traversal search is performed, and the semantic cosine similarity between the candidate underlying feature vectors in the database and the query conditions is calculated one by one to extract the manuscript fact feature vector corresponding to the maximum similarity score. The specific formula is as follows:

[0031] ;

[0032] in, It is a function that maximizes the parameters; For traversal Candidate low-level feature vectors at that time Used to serve as a retrieval database; express and The product of the magnitudes of two vectors;

[0033] This is a vector dot product operation;

[0034] 3.3 Using template slot semantic feature vectors Based on the query criteria, the full high-dimensional feature vector set of the audited projects is used. Feature retrieval is performed to extract matching factual feature vectors from the original manuscript. Then, the factual feature vector of the draft was calculated. semantic feature vector of template slot The semantic cosine similarity is obtained by calculating the cosine similarity, and the specific formula is as follows:

[0035] ;

[0036] in, For semantic cosine similarity; express and The product of the magnitudes of two vectors;

[0037] Among them, records The corresponding maximum semantic cosine similarity And the numerical range is controlled within between;

[0038] 3.4 Match content to workflow Defined as a similarity-based piecewise decision function, based on semantic cosine similarity. Extract the matching content, obtain and output the factual context of the draft. The specific formula is as follows:

[0039] ;

[0040] in, This is a dynamic confidence threshold. Decision execution function for content matching workflow; This means that when the semantic cosine similarity is greater than the dynamic confidence threshold, the system automatically extracts the original text content corresponding to the factual feature vector of the draft as the factual context of the draft. This means that when the semantic cosine similarity is lower than or equal to the dynamic confidence threshold, the system marks the item as pending manual review. After review and confirmation, it will be directly converted into the factual context of the draft.

[0041] The aforementioned automated audit report generation method based on a large model, in step 4, runs the report generation workflow, guiding the large language model to generate structured preliminary audit text through prompts. The specific steps are as follows:

[0042] 4.1 Based on the factual context of the draft manuscript The project builds a dedicated audit template that includes role settings and task descriptions. The specific splicing formula is as follows:

[0043] ;

[0044] in, As a basis for audit regulations; This indicates a text sequence concatenation and integration operation; Preset prompts for the audit expert role; audit-specific template To construct a complete sequence of audit prompt text;

[0045] 4.2 Use the audit template The input is fed into a large language model for inference and generation, and the output is a structured preliminary audit text. The specific formula is as follows:

[0046] ;

[0047] in, The execution function that generates the workflow for the report; For large language model inference algorithms; Includes an overview of audit issues and audit recommendations.

[0048] The aforementioned automated audit report generation method based on a large model, in step 5, triggers a language optimization workflow to segment the initial audit text, constructs a context consistency verification model to perform multiple rounds of iterative rewriting on the segmented text, and outputs the final text set of each chapter. The specific steps are as follows:

[0049] 5.1. Submit the preliminary audit documents. Divided according to the report's chapter structure A sequence of paragraphs, denoted as The sequence of each paragraph is input into a pre-built quality assessment model;

[0050] in, This represents the total number of specific chapters and modules that need to be filled in for the current report;

[0051] 5.2 The quality assessment model calculates the text perplexity and logical consistency scores for each input paragraph sequence, establishing a paragraph-level optimization objective function. The specific formula is as follows:

[0052] ;

[0053] in, Assess the text perplexity of the current paragraph. Score the logical consistency of the currently evaluated paragraph; Weighting coefficients to balance smoothness; Weighting coefficients to balance logical consistency;

[0054] 5.3 Based on the optimization objective function The paragraph sequence is rewritten in multiple independent iterations to generate text. Calculate the corresponding objective function value. Determine whether it converges to the preset range;

[0055] The specific criteria for determining the preset range are as follows:

[0056] or ;

[0057] in, The paragraph number is a number ranging from 1 to 1. ; Regarding the first The first paragraph sequence The text generated by rewriting in the next iteration; for The corresponding objective function value; The upper limit of the objective function is an absolute threshold used to measure whether the text quality has reached the exemption standard. The relative threshold for tolerance of changes between two iterations is used to measure whether the optimization margin of a large language model has reached a state of diminishing marginal returns.

[0058] 5.4, ​​Let When the first The sequence of segments satisfies the set range criteria, or the current iteration number. Reaching the set maximum number of iterations At that time, language optimization workflow Terminate the current iteration of the paragraph sequence and directly transform the current text state into the final optimized text. The specific conversion and assignment formula is as follows:

[0059] ;

[0060] in, This indicates that each segmented initial paragraph is considered as the initial state for the corresponding iterative rewriting; Indicates the first The final optimized text is output after the sequence of paragraphs has reached convergence. This represents the last text generated in the paragraph sequence when the termination condition is met.

[0061] 5.5 The system has completed traversing and processing all... After a sequence of paragraphs, the final output is a collection of texts for each chapter, after logical error correction and standardization of technical terms. This is used for the input of entity data in the downstream layout and rendering module.

[0062] The aforementioned automated audit report generation method based on a large model, in step 6, constructs a document structure tree, maps the optimized text sets of each chapter to the document structure tree for format rendering, generates the target audit report file, and establishes a version anti-counterfeiting and traceability mechanism for the entire lifecycle of the audit report. The specific steps are as follows:

[0063] 6.1 Based on relevant auditing standards and the current audit project type, invoke the structure tree construction function to automatically generate a standardized audit report document structure tree. The specific construction formula is as follows:

[0064] ;

[0065] in, A function for automating the generation of structure trees, used to dynamically assemble report templates based on different business scenarios; The relevant audit criteria and formatting constraint rule set are pre-imported into the system; Labels for the current audit project type; This represents the set of normalized hierarchical directory nodes defined in the document's structure tree; The total number of hierarchical directory modules dynamically generated based on the current project type, and ;

[0066] Standardized hierarchical directory nodes The basic information, audit findings, and audit recommendations all end with text slots to be filled, which provide a layout framework for subsequent optimization of the text content;

[0067] The standardized hierarchical directory includes basic information, audit findings, and audit recommendations;

[0068] 6.2. Compile the final text of each chapter. Corresponding to the document structure tree The target audit report file is obtained by calling the rendering engine to perform format rendering in the leaf node slot. The specific formula is as follows:

[0069] ;

[0070] in, For document typesetting and formatting rendering engine; target audit report files For structured audit report files containing the complete format;

[0071] 6.3. Submit the target audit report document Export as a standard format document;

[0072] Standard format documents are set to PDF format or other non-editable formats;

[0073] 6.4 Establish a version anti-counterfeiting and traceability mechanism for the entire lifecycle of audit reports.

[0074] The aforementioned automated audit report generation method based on a large model establishes a version anti-counterfeiting and traceability mechanism for the entire lifecycle of audit reports in step 6.4. The specific steps are as follows:

[0075] 6.4.1 Based on a secure hash algorithm, perform hash mapping calculations on the binary data stream of the standard format document, the system's current generation timestamp, and the corresponding auditor's identity identifier to generate a unique digital fingerprint for the standard format document. ;

[0076] 6.4.2. Digital fingerprint The watermark or document metadata is embedded into the exported standard format document and simultaneously stored in the system's audit log database for immutable and fixed filing.

[0077] Compared with the prior art, the beneficial effects of the present invention are:

[0078] 1. This invention constructs an intelligent agent framework that operates collaboratively through multiple sub-workflows. First, a data reading workflow extracts multi-source heterogeneous working paper data from the audit database and performs deep semantic analysis using a large model. Subsequently, a content matching workflow achieves accurate mapping and automatic filling of audit elements and report templates. Then, a report generation workflow automatically generates problem characterization, cause analysis, and audit recommendations based on working paper features and audit criteria, utilizing the reasoning capabilities of the large model. Finally, a language optimization workflow performs logical coherence review and standardization of professional terminology on the text, achieving full automation of the audit report process from working paper extraction to final output. This significantly improves the efficiency and standardization of report preparation and promotes the intelligent transformation of audit operations.

[0079] 2. This invention decomposes the complex report preparation task into multiple independent and professional sub-tasks, and combines this with the dynamic coordination of the task scheduling matrix. This avoids the illusions and content omissions that can easily occur when a single large language model processes massive amounts of heterogeneous data and complex logic at once. At the same time, it introduces prompting engineering and post-verification mechanisms, which can perform strict logical error correction, fluency polishing, and standardization and alignment of professional terminology on the initial audit text. This filters out any awkward expressions and logical jumps that may remain in the content generated by the large language model, ensuring that the produced target audit report document has factual accuracy and logical consistency. Through the automated rendering of the document structure tree and the generation of digital fingerprints, the uniqueness, tamper-proofness, and full-process version traceability of the issued target audit report document are guaranteed, meeting the strict requirements for audit file compliance management. Attached Figure Description

[0080] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0081] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0082] This invention provides the following: Figure 1 The method for automated audit report generation based on a large model, as shown, includes the following steps:

[0083] Step 1: Construct a multi-agent collaborative framework to decompose the complex report compilation task into sub-tasks that can be executed independently. The specific steps are as follows:

[0084] 1.1 Determine the intelligent agent workflow set based on the audit business logic. The specific formula is as follows:

[0085] ;

[0086] in, For data reading workflow, Match content to workflow, To generate a workflow for the report, Optimize workflows for languages; the collection of agent workflows covers the entire lifecycle from data input to report output;

[0087] 1.2 Construct a task scheduling matrix to monitor the status of each workflow in real time. The specific formula is as follows:

[0088] ;

[0089] in, Indicates the current time step The system scheduling decision result, that is, the next best workflow selected for execution; For parameter maximization function, used to extract from Select the workflow that maximizes the probability value from the selection process; It is a pre-trained task scheduling probability model used to predict the execution necessity of each workflow under a specific state; These are candidate agent workflow variables; Indicates the previous time step System execution status feedback; This indicates the global context status information of the current audit project, including the amount of data processed and the current report completion status;

[0090] In step 1, the complex task of report preparation is broken down into several independent and specialized sub-tasks, namely... , , , Furthermore, by combining dynamic coordination with the task scheduling matrix, it effectively avoids the illusion and content omission problems that are prone to occur when a single large language model processes massive heterogeneous data and complex logic at once. At the same time, the probabilistic scheduling mechanism based on global context state gives the system a high degree of flexibility, enabling it to adaptively cope with abnormal situations such as data parsing failure or insufficient matching degree, which greatly improves the robustness of the automated audit process.

[0091] Step 2: Start the data reading workflow Multimodal analysis and feature extraction are performed on heterogeneous audit working paper data to construct a high-dimensional feature vector set for the entire audit project. The specific steps are as follows:

[0092] 2.1. By integrating optical character recognition and natural language processing technologies, the multi-source heterogeneous manuscripts are structured and analyzed to identify and separate tables, signatures, and text. Based on the separation results of each region, key entity information of the project is extracted.

[0093] Among them, multi-source heterogeneous drafts include tables, scanned copies, and documents;

[0094] 2.2 Data Reading Workflow Defined as a multimodal feature mapping process, it uses a feature encoding function to extract semantic and spatial features from key entity information of the project, and converts these into a high-dimensional feature vector set of the entire audit project that incorporates layout information. The specific formula is as follows:

[0095] ;

[0096] in, For feature encoding function, This is the processing function for the data reading workflow; This is a collection of heterogeneous audit working papers for input. This is the full set of high-dimensional feature vectors for the audit project; This is an aggregation or splicing operation for a collection, used to integrate the features of a single draft into a project-level feature library; This refers to the total number of heterogeneous audit working papers entered into the audit project. This is the index number of the draft document, with a value ranging from 1 to... ; Indicates the first The table structure data identified in step 2.1 in the draft includes cell content and its corresponding row and column logical relationships; Indicates the first Visual features of signatures identified in the original document are used to record the legal validity status of the document; Indicates the first Unstructured text in the draft;

[0097] Among them, the feature encoding function A multimodal pre-trained model based on the Transformer architecture is used as the basic network structure. The features of heterogeneous regions are tensor-concatenated and dimension-unified, and finally mapped into a high-dimensional feature vector containing all the information of the manuscript.

[0098] To adapt to specific audit business scenarios, feature encoding function Before practical application, a domain fine-tuning training process is included, which involves using anonymized historical audit working papers and standard audit terminology as training corpus to construct semantically related and unrelated positive and negative sample pairs. A contrastive learning algorithm is used to update the weights of the basic network structure, enabling the multimodal pre-trained model to accurately identify the deep semantics of specific audit facts such as "abnormal contract amount" and "missing price comparison sheet".

[0099] After feature encoding function After mapping, the final generated is a dense floating-point vector of fixed dimensions (such as 768-dimensional or 1024-dimensional). This vector integrates the semantic and spatial features of the original draft, providing standardized data support for high-precision similarity retrieval in the subsequent vector space.

[0100] In step 2, through multimodal parsing and feature vectorization of multi-source heterogeneous data, the originally messy unstructured drafts (such as charts, scanned documents, and documents) were successfully transformed into a unified, machine-understandable high-dimensional vector representation. This completely eliminated the risk of omissions and errors that are easily generated when manually extracting and entering data in the traditional way, and laid a standardized, high-quality data foundation for the accurate semantic understanding and efficient retrieval of subsequent large models (large language models, reordering models).

[0101] Step 3: Execute the content matching workflow The document performs semantic understanding on its features, calculates its matching degree with the preset report template slots, and generates factual context for the document. The specific steps are as follows:

[0102] 3.1 Parse the preset standardized audit report template, extract the natural language description text of the current text slot to be filled, and call the feature encoding function to map it into a semantic feature vector of the template slot. The specific formula is as follows:

[0103] ;

[0104] in, This refers to the natural language description text of the current text slot to be filled after parsing the preset standardized audit report template; the working paper features refer to the extracted set of key semantics, content attributes and core elements inherent in the business working paper itself, which are used to perform semantic matching and align template slots for the large model; The use of template slot semantic feature vectors can ensure The features of the draft are projected into the same high-dimensional vector space;

[0105] 3.2 Using template slot semantic feature vectors As a query condition, in the full high-dimensional feature vector set of the audited projects A global traversal search is performed, and the semantic cosine similarity between the candidate underlying feature vectors in the database and the query conditions is calculated one by one to extract the manuscript fact feature vector corresponding to the maximum similarity score. The specific formula is as follows:

[0106] ;

[0107] in, This is a parameter maximization function used to locate specific candidate vectors that maximize the similarity function value. For traversal The candidate low-level feature vectors represent the individual feature terms in the low-level database. Used to serve as a retrieval database; express and The product of the magnitudes of the two vectors is used to represent the normalization process;

[0108] This is a vector dot product operation;

[0109] 3.3 Using template slot semantic feature vectors Based on the query criteria, the full high-dimensional feature vector set of the audited projects is used. Feature retrieval is performed to extract matching factual feature vectors from the original manuscript. Then, the factual feature vector of the draft was calculated. semantic feature vector of template slot The semantic cosine similarity is obtained by calculating the cosine similarity, and the specific formula is as follows:

[0110] ;

[0111] in, Semantic cosine similarity is used to measure the degree of fit between the content of the draft and the chapters of the report; express and The product of the magnitudes of the two vectors is used to represent the normalization process;

[0112] In this process, normalization eliminates the impact of text length differences on matching results, ensuring that lengthy descriptions and concise titles can be correctly aligned in the semantic space. The system records... The corresponding maximum semantic cosine similarity And the numerical range is controlled within between;

[0113] 3.4 Match content to workflow Defined as a similarity-based piecewise decision function, based on semantic cosine similarity. Extract the matching content, obtain and output the factual context of the draft. The specific formula is as follows:

[0114] ;

[0115] in, This is a dynamic confidence threshold. Decision execution function for content matching workflow; This means that when the semantic cosine similarity is greater than the dynamic confidence threshold, the system automatically extracts the original text content corresponding to the factual feature vector of the draft as the factual context of the draft. This means that when the semantic cosine similarity is lower than or equal to the dynamic confidence threshold, the system marks the item as pending manual review. After review and confirmation, it will be directly converted into the factual context of the draft.

[0116] In step 3, by calculating semantic cosine similarity, accurate semantic alignment of the factual feature vector of the draft and the semantic feature vector of the template slot is achieved, avoiding the high-cost operation of traditional manual comparison and splicing. While improving the efficiency of automated filling, low-confidence matching results are intercepted by setting a dynamic confidence threshold and transferred to manual review, thereby effectively ensuring the accuracy and rigor of audit data.

[0117] Step 4: Workflow for generating the run report By guiding the large language model through prompting engineering, structured preliminary audit text is generated. The specific steps are as follows:

[0118] 4.1 Based on the factual context of the draft manuscript The project builds a dedicated audit template that includes role settings and task descriptions. The specific splicing formula is as follows:

[0119] ;

[0120] in, As a basis for audit regulations, the system retrieves and accesses audit regulations applicable to the current business scenario. This indicates a text sequence concatenation and integration operation; These are preset prompts for the role of an audit expert, used to limit the output context and professional tone of the large language model;

[0121] Among them, audit-specific templates To construct a complete sequence of audit prompt text;

[0122] 4.2 Use the audit template The input is fed into a large language model for inference and generation, and the output is a structured preliminary audit text. The specific formula is as follows:

[0123] ;

[0124] in, The execution function that generates the workflow for the report; For large language model inference algorithms;

[0125] in, Includes an overview of audit issues and audit recommendations;

[0126] In step 4, a dedicated audit template is constructed that integrates audit expert role prompts, factual context from working papers, and legal basis for auditing. This effectively constrains and guides the reasoning boundaries of the large language model, enabling previously fragmented facts to be automatically and quickly transformed into logically rigorous, objectively expressed problem summaries and recommendations that conform to the "legal language" of auditing. This significantly reduces the text writing burden on auditors and avoids qualitative biases caused by subjective interpretation.

[0127] Step 5: Trigger the language optimization workflow The preliminary audit text The text is segmented, and a context consistency verification model is built to perform multiple rounds of iterative rewriting of the segmented text, outputting the final text set of each chapter. The specific steps are as follows:

[0128] 5.1. Submit the preliminary audit documents. Divided according to the report's chapter structure A sequence of paragraphs, denoted as The sequence of each paragraph is input into a pre-built quality assessment model;

[0129] in, This represents the total number of specific chapters and modules that need to be filled in for the current report;

[0130] The quality assessment model adopts a dual-stream parallel assessment architecture, which includes: a causal language model for assessing the fluency of text expression, and a natural language inference model for testing semantic conflict prevention between paragraphs.

[0131] 5.2 The quality assessment model calculates the text perplexity and logical consistency scores for each input paragraph sequence, establishing a paragraph-level optimization objective function. The specific formula is as follows:

[0132] ;

[0133] in, Assess the text perplexity of the current paragraph. Score the logical consistency of the currently evaluated paragraph; Weighting coefficients to balance smoothness; Weighting coefficients to balance logical consistency;

[0134] 5.3 Based on the optimization objective function The paragraph sequence is rewritten in multiple independent iterations to generate text. Calculate the corresponding objective function value. Determine whether it converges to the preset range;

[0135] The specific criteria for determining the preset range are as follows:

[0136] or ;

[0137] in, The paragraph number is a number ranging from 1 to 1. ; Regarding the first The first paragraph sequence The text generated by rewriting in the next iteration; for The corresponding objective function value; The upper limit of the objective function is an absolute threshold used to measure whether the text quality has reached the exemption standard. The relative threshold for tolerance of changes between two iterations is used to measure whether the optimization margin of a large language model has reached a state of diminishing marginal returns.

[0138] 5.4, ​​Let When the first The sequence of segments satisfies the set range criteria, or the current iteration number. Reaching the set maximum number of iterations At that time, language optimization workflow Terminate the current iteration of the paragraph sequence and directly transform the current text state into the final optimized text. The specific conversion and assignment formula is as follows:

[0139] ;

[0140] in, This indicates that each segmented initial paragraph is considered as the initial state for the corresponding iterative rewriting; Indicates the first The final optimized text is output after the sequence of paragraphs has reached convergence. This represents the last text generated in the paragraph sequence when the termination condition is met.

[0141] Among them, a maximum number of iterations is set. This is to prevent the system from falling into an infinite rewrite loop;

[0142] 5.5 The system has completed traversing and processing all... After a sequence of paragraphs, the final output is a collection of texts for each chapter, after logical error correction and standardization of technical terms. This is used for the input of entity data in the downstream layout and rendering module.

[0143] In step 5, a paragraph-level optimization objective function is constructed, with text perplexity and logical consistency scores as the core. Iterative rewriting was performed to build an automated quality assessment and post-verification mechanism for the initial draft report, capable of verifying the initial audit text. By performing rigorous logical error correction, fluency enhancement, and standardization and alignment of professional terminology, the awkward expressions and logical jumps that may remain in the content generated by the large language model are further filtered out, greatly reducing the cost of manual secondary review.

[0144] Step 6: Construct a document structure tree, which will contain the final text collection of each optimized chapter. The format is rendered by mapping to the document structure tree, the target audit report file is generated, and a version anti-counterfeiting and traceability mechanism is established throughout the lifecycle of the audit report. The specific steps are as follows:

[0145] 6.1 Based on relevant auditing standards and the current audit project type, invoke the structure tree construction function to automatically generate a standardized audit report document structure tree. The specific construction formula is as follows:

[0146] ;

[0147] in, A function for automating the generation of structure trees, used to dynamically assemble report templates based on different business scenarios; The relevant audit criteria and formatting constraint rule set are pre-imported into the system; Label the current audit project type, such as departure audit or economic responsibility audit; This represents the set of normalized hierarchical directory nodes defined in the document's structure tree; The total number of hierarchical directory modules dynamically generated based on the current project type, and ;

[0148] Standardized hierarchical directory nodes The basic information, audit findings, and audit recommendations all end with text slots to be filled, which provide a layout framework for subsequent optimization of the text content;

[0149] The standardized hierarchical directory includes basic information, audit findings, and audit recommendations;

[0150] 6.2. Compile the final text of each chapter. Corresponding to the document structure tree The target audit report file is obtained by calling the rendering engine to perform format rendering in the leaf node slot. The specific formula is as follows:

[0151] ;

[0152] in, For document typesetting and formatting rendering engine; target audit report files For structured audit report files containing the complete format;

[0153] 6.3. Submit the target audit report document Export as a standard format document;

[0154] Standard format documents are set to PDF format or other non-editable formats;

[0155] 6.4 Establish a version anti-counterfeiting and traceability mechanism for the entire lifecycle of audit reports. The specific steps are as follows:

[0156] 6.4.1 Based on a secure hash algorithm, perform hash mapping calculations on the binary data stream of the standard format document, the system's current generation timestamp, and the corresponding auditor's identity identifier to generate a unique digital fingerprint for the standard format document. ;

[0157] 6.4.2. Digital fingerprint Embedded into the exported standard format document in the form of implicit watermark or document metadata, and simultaneously stored in the system's audit log database for immutable and fixed filing;

[0158] Among them, the secure hash algorithm is a type of one-way irreversible, fixed-length output, collision-resistant encrypted digest algorithm, which is specifically used for file fingerprinting, integrity verification, anti-tampering, and identity tracing, and can be applied to the digital fingerprint generation scenario of this application;

[0159] Furthermore, the secure hash algorithm is set to SHA-256, a standard hash algorithm commonly used in national standards, government and enterprise, and audit evidence storage scenarios;

[0160] In step 6, the automatic rendering of the document structure tree and the generation of digital fingerprints enable the standardized generation of complex layout formats and multi-chapter content within seconds. This also ensures the uniqueness, tamper-proof nature, and full-process version traceability of the target audit report documents at the mechanism level, meeting the stringent requirements for audit file compliance management.

[0161] This invention proposes an automated audit report generation system based on a large model, aiming to fully realize the entire process of automated audit report generation based on a large model from the system architecture level;

[0162] The system adopts a modular design concept and consists of five modules: agent scheduling module, data parsing module, semantic matching module, report generation module, and language optimization module.

[0163] The five modules work together through standardized interfaces to complete the entire process from inputting original audit working papers to outputting the final structured report. The functions of each module are as follows:

[0164] The agent scheduling module, as the control center of the system, includes a workflow definition unit and a task scheduling engine. Its main responsibility is to initialize a multi-agent collaborative workflow and monitor the context state of each subtask in real time during operation. It uses a probabilistic model to dynamically coordinate the task allocation and timing synchronization of data reading, content matching, report generation and language optimization, thereby ensuring the consistency of the entire system logic and execution efficiency.

[0165] The data parsing module serves as the system's perception entry point, integrating a multimodal document processing unit and a feature encoding unit. This module combines optical character recognition and natural language processing technologies to perform in-depth layout analysis and key entity extraction on audit working papers from diverse sources and in various formats. It also transforms unstructured content into high-dimensional structured feature vectors that machines can understand, providing a unified data foundation for subsequent processing.

[0166] The semantic matching module is responsible for establishing a precise mapping relationship between the draft content and the report template. It includes a template slot parsing unit and a similarity calculation unit. This module first parses the field structure in the audit report template, then calculates the cosine similarity between the draft feature vector and each template slot vector, and automatically determines the optimal matching scheme to achieve intelligent content filling.

[0167] The report generation module relies on the powerful text generation capabilities of the large model and includes a prompt word engineering management unit and a large model reasoning unit. This module constructs a Prompt template that integrates audit role settings and compliance constraints, and calls the generative large model to automatically generate a clear and accurate summary of issues and audit opinions by combining the factual basis of the working papers and relevant legal provisions.

[0168] The language optimization module, serving as the final quality checkpoint before output, includes a quality assessment unit and an iterative polishing unit. This module calculates the perplexity and logical consistency scores of the generated text and uses an objective function-driven optimization mechanism to perform multiple rounds of iterative polishing on the initial draft, ultimately improving the fluency of the language and ensuring the standardization and consistency of the use of professional terminology.

[0169] Verification Experiment

[0170] The following examples illustrate the effectiveness of this invention using the scenario of "compliance audit of procurement business of a group's subsidiaries".

[0171] The auditors uploaded a scanned copy of the verification worksheet. The text description included information such as "From April to June 2020, Subsidiary A signed a contract with Company B for 'XX equipment maintenance services.' The contract amount was substantial, and no public bidding documents were found during the procurement process. Only departmental negotiation records and payment approvals were provided." The system performed OCR recognition on the scanned document, separating the table (payment approval form), signature (company seal), and text. Key entities (time, subject, event) were extracted, and this information was converted into a high-dimensional feature vector set. The system parses the natural language description of the "Procurement Issue Description" slot in the preset report template and converts it into a query vector. Using this query vector, it searches the feature vector set of the working paper and calculates the cosine similarity. If the similarity score of the working paper's main text exceeds a preset threshold, the system automatically extracts the original text of the working paper as its factual context. Based on this factual context, the system constructs a dedicated audit template that includes the roles of audit experts and the legal basis for auditing. This template is then input into the large language model to generate a structured preliminary audit text. Based on the aforementioned factual context of the working papers, the system concatenates audit-specific prompts: "You are an audit expert. Based on the following facts: From April to June 2020, Subsidiary A had contracts with substantial amounts but no public bidding was observed. According to the 'XX Group Procurement Management Measures' (2020 Edition), which stipulates that 'amounts exceeding XX million yuan must be subject to public bidding,' generate the problem characterization and audit recommendations." Inputting these prompts into the large language model generates preliminary audit text: "Problem Characterization: Inadequate execution of procurement procedures. Fact Description: From April to June 2020, Subsidiary A… Audit Recommendation: Supplement bidding or price comparison materials."

[0172] The system divides the initial text into three paragraphs: "Problem Characterization," "Fact Description," and "Audit Recommendations," and inputs them into the quality assessment model. Each paragraph is scored for perplexity and logical consistency. If the objective function value of the "Fact Description" paragraph does not reach the convergence standard, it is rewritten. After multiple iterations, the objective function value converges to the preset range, the iteration is terminated, and the optimized text of the three paragraphs is output. Based on auditing standards and procurement compliance audit project types, the system generates a document structure tree containing nodes for basic information, audit findings, and audit recommendations. The optimized three paragraphs are mapped to corresponding node slots, and a PDF report is generated. A hash calculation is performed based on the PDF file, the current timestamp, and the auditor's ID to generate a digital fingerprint, which is then embedded as an implicit watermark and stored in the audit log database.

[0173] As can be seen from the above experiments, the work that originally required auditors to spend several hours organizing, writing and proofreading can be completed in a few minutes through the design of this invention, while effectively avoiding the risks of errors in manual transcription and non-standard expression.

[0174] In summary, this invention decomposes complex report compilation tasks into multiple independent and specialized sub-tasks, and combines this with the dynamic coordination of a task scheduling matrix, thus avoiding the illusion and content omission problems that are prone to occur when a single large language model processes massive heterogeneous data and complex logic at once.

[0175] By using multimodal analysis and feature vectorization of multi-source heterogeneous data, the originally messy unstructured drafts were successfully transformed into a unified, machine-understandable high-dimensional vector representation, eliminating the risk of omissions and errors that are easily generated when manually extracting and entering data in the traditional way.

[0176] By calculating semantic cosine similarity, accurate semantic alignment between the factual feature vector of the draft and the semantic feature vector of the template slot is achieved, which improves the efficiency of automated filling. Furthermore, by setting a dynamic confidence threshold to intercept low-confidence matching results, the accuracy and rigor of audit data can be effectively guaranteed.

[0177] By constructing a dedicated audit template that integrates audit expert role prompts, factual context of working papers, and audit regulatory basis, the text writing burden of auditors has been significantly reduced and qualitative deviations caused by subjective interpretation have been avoided.

[0178] Furthermore, by constructing an automated quality assessment and post-verification mechanism for the initial draft of the report, it is possible to strictly correct logical errors, improve fluency, and standardize and align professional terms in the initial audit text. This further filters out any awkward expressions and logical jumps that may remain in the content generated by the large language model, greatly reducing the cost of manual secondary review.

[0179] By automatically rendering the document structure tree and generating digital fingerprints, the system achieves standardized generation of complex layout formats and multi-chapter content within seconds. It also ensures the uniqueness, tamper-proofness, and full-process version traceability of the target audit report documents at the mechanism level, meeting the strict requirements for audit file compliance management.

[0180] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used above are only some embodiments described in this invention. Obviously, those skilled in the art can obtain other drawings based on these drawings.

[0181] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A method for automated audit report generation based on a large model, characterized by the following steps: Step 1: Construct a multi-agent collaborative framework to decompose the complex report compilation task into sub-tasks that can be executed independently. The specific steps are as follows: 1.1 Determine the intelligent agent workflow set based on the audit business logic. The specific formula is as follows: ； in, For data reading workflow, Match content to workflow, To generate a workflow for the report, Optimize workflows for languages; the collection of agent workflows covers the entire lifecycle from data input to report output; 1.2 Construct a task scheduling matrix to monitor the status of each workflow in real time. The specific formula is as follows: ； in, Indicates the current time step The system scheduling decision result, that is, the next best workflow selected for execution; It is a function that maximizes the parameters; For pre-trained task scheduling probability models; These are candidate agent workflow variables; Indicates the previous time step System execution status feedback; This indicates the global context status information of the current audit project, including the amount of data processed and the current report completion status; Step 2: Initiate the data reading workflow to perform multimodal analysis and feature extraction on heterogeneous audit working paper data, and construct a high-dimensional feature vector set for the entire audit project. The specific steps are as follows: 2.

1. By integrating optical character recognition and natural language processing technologies, the multi-source heterogeneous manuscripts are structured and analyzed to identify and separate tables, signatures, and text. Based on the separation results of each region, key entity information of the project is extracted. Among them, multi-source heterogeneous drafts include tables, scanned copies, and documents; 2.2 Data Reading Workflow Defined as a multimodal feature mapping process, it uses a feature encoding function to extract semantic and spatial features from key entity information of the project, and converts them into a high-dimensional feature vector set of the entire audit project that incorporates layout information. The specific formula is as follows: ； in, For feature encoding function, For processing functions in the data reading workflow; This is a collection of heterogeneous audit working papers for input. This is the full set of high-dimensional feature vectors for the audit project; For aggregation or concatenation operations on sets; This refers to the total number of heterogeneous audit working papers entered into the audit project. This is the index number of the draft document, with a value ranging from 1 to... ; Indicates the first The table structure data identified in step 2.1 in the draft includes cell content and its corresponding row and column logical relationships; Indicates the first Visual features of signatures identified in the original draft; Indicates the first Unstructured text in the draft; Step 3: Execute the content matching workflow, perform semantic understanding on the features of the draft, calculate its matching degree with the preset report template slots, and generate the factual context of the draft; Step 4: Run the report generation workflow, and guide the large language model to generate structured preliminary audit text through prompts and engineering. Step 5: Trigger the language optimization workflow, segment the initial audit text, build a context consistency verification model to rewrite the segmented text in multiple rounds, and output the final text set of each chapter; Step 6: Construct a document structure tree, map the optimized text set of each chapter to the document structure tree for format rendering, generate the target audit report file, and establish a version anti-counterfeiting and traceability mechanism for the entire lifecycle of the audit report.

2. The method for generating automated audit reports based on a large model according to claim 1, characterized in that: In step 3, the content matching workflow is executed to perform semantic understanding on the draft features, calculate their matching degree with the preset report template slots, and generate the draft factual context. The specific steps are as follows: 3.1 Parse the preset standardized audit report template, extract the natural language description text of the current text slot to be filled, and call the feature encoding function to map it into a semantic feature vector of the template slot. The specific formula is as follows: ； in, This refers to the natural language description text of the current text slot to be filled after parsing the preset standardized audit report template; the working paper features refer to the extracted set of key semantics, content attributes and core elements inherent in the business working paper itself, which are used to perform semantic matching and align template slots for the large model; 3.2 Using template slot semantic feature vectors As a query condition, in the full high-dimensional feature vector set of the audited projects A global traversal search is performed, and the semantic cosine similarity between the candidate underlying feature vectors in the database and the query conditions is calculated one by one to extract the manuscript fact feature vector corresponding to the maximum similarity score. The specific formula is as follows: ； in, It is a function that maximizes the parameters; For traversal Candidate low-level feature vectors at that time Used to serve as a retrieval database; express and The product of the magnitudes of two vectors; This is a vector dot product operation; 3.3 Using template slot semantic feature vectors Based on the query criteria, the full high-dimensional feature vector set of the audited projects is used. Feature retrieval is performed to extract matching factual feature vectors from the original manuscript. Then, the factual feature vector of the draft was calculated. semantic feature vector of template slot The semantic cosine similarity is obtained by calculating the cosine similarity, and the specific formula is as follows: ； in, For semantic cosine similarity; express and The product of the magnitudes of two vectors; Among them, records The corresponding maximum semantic cosine similarity And the numerical range is controlled within between; 3.4 Match content to workflow Defined as a similarity-based piecewise decision function, based on semantic cosine similarity. Extract the matching content, obtain and output the factual context of the draft. The specific formula is as follows: ； in, This is a dynamic confidence threshold. Decision execution function for content matching workflow; This means that when the semantic cosine similarity is greater than the dynamic confidence threshold, the system automatically extracts the original text content corresponding to the factual feature vector of the draft as the factual context of the draft. This means that when the semantic cosine similarity is lower than or equal to the dynamic confidence threshold, the system marks the item as pending manual review. After review and confirmation, it will be directly converted into the factual context of the draft.

3. The method for generating automated audit reports based on a large model according to claim 1, characterized in that, in step 4, the report generation workflow is run, and the large language model is guided by prompts to generate structured preliminary audit text. The specific steps are as follows: 4.1 Based on the factual context of the draft manuscript The project builds a dedicated audit template that includes role settings and task descriptions. The specific splicing formula is as follows: ； in, As a basis for audit regulations; This indicates a text sequence concatenation and integration operation; Preset prompts for the audit expert role; audit-specific template To construct a complete sequence of audit prompt text; 4.2 Use the audit template The input is fed into a large language model for inference and generation, and the output is a structured preliminary audit text. The specific formula is as follows: ； in, The execution function that generates the workflow for the report; For large language model inference algorithms; Includes an overview of audit issues and audit recommendations.

4. The method for generating automated audit reports based on a large model according to claim 1, characterized in that: In step 5, the language optimization workflow is triggered, the initial audit text is segmented, a context consistency verification model is built to rewrite the segmented text through multiple rounds of iterations, and the final text set of each chapter is output. The specific steps are as follows: 5.

1. Submit the preliminary audit documents. Divided according to the report's chapter structure A sequence of paragraphs, denoted as The sequence of each paragraph is input into a pre-built quality assessment model; in, This represents the total number of specific chapters and modules that need to be filled in for the current report; 5.2 The quality assessment model calculates the text perplexity and logical consistency scores for each input paragraph sequence, establishing a paragraph-level optimization objective function. The specific formula is as follows: ； in, Assess the text perplexity of the current paragraph. Score the logical consistency of the currently evaluated paragraph; Weighting coefficients to balance smoothness; Weighting coefficients to balance logical consistency; 5.3 Based on the optimization objective function The paragraph sequence is rewritten in multiple independent iterations to generate text. Calculate the corresponding objective function value. Determine whether it converges to the preset range; The specific criteria for determining the preset range are as follows: or ; in, The paragraph number is a number ranging from 1 to 1. ; Regarding the first The first paragraph sequence The text generated by rewriting in the next iteration; for The corresponding objective function value; The upper limit of the objective function is an absolute threshold used to measure whether the text quality has reached the exemption standard. The relative threshold for tolerance of changes between two iterations is used to measure whether the optimization margin of a large language model has reached a state of diminishing marginal returns. 5.4, ​​Let When the first The sequence of segments satisfies the set range criteria, or the current iteration number. Reaching the set maximum number of iterations At that time, language optimization workflow Terminate the current iteration of the paragraph sequence and directly transform the current text state into the final optimized text. The specific conversion and assignment formula is as follows: ； in, This indicates that each segmented initial paragraph is considered as the initial state for the corresponding iterative rewriting; Indicates the first The final optimized text is output after the sequence of paragraphs has reached convergence. This represents the last text generated in the paragraph sequence when the termination condition is met. 5.5 The system has completed traversing and processing all... After a sequence of paragraphs, the final output is a collection of texts for each chapter, after logical error correction and standardization of technical terms. This is used for the input of entity data in the downstream layout and rendering module.

5. The method for generating automated audit reports based on a large model according to claim 1, characterized in that: In step 6, a document structure tree is constructed, mapping the optimized text sets of each chapter to the document structure tree for format rendering, generating the target audit report file, and establishing a version anti-counterfeiting and traceability mechanism for the entire lifecycle of the audit report. The specific steps are as follows: 6.1 Based on auditing standards and the current audit project type, invoke the structure tree construction function to automatically generate a standardized audit report document structure tree. The specific construction formula is as follows: ； in, A function for automating the generation of structure trees, used to dynamically assemble report templates based on different business scenarios; The set of audit criteria and formatting constraints pre-imported into the system; Labels for the current audit project type; This represents the set of normalized hierarchical directory nodes defined in the document's structure tree; The total number of hierarchical directory modules dynamically generated based on the current project type, and ; Standardized hierarchical directory nodes The basic information, audit findings, and audit recommendations all end with text slots to be filled, which provide a layout framework for subsequent optimization of the text content; The standardized hierarchical directory includes basic information, audit findings, and audit recommendations; 6.

2. Compile the final text of each chapter. Corresponding to the document structure tree The target audit report file is obtained by calling the rendering engine to perform format rendering in the leaf node slot. The specific formula is as follows: ； in, For document typesetting and formatting rendering engine; target audit report files For structured audit report files containing the complete format; 6.

3. Submit the target audit report document Export as a standard format document; Standard format documents are set to PDF format or other non-editable formats; 6.4 Establish a version anti-counterfeiting and traceability mechanism for the entire lifecycle of audit reports.

6. The method for generating automated audit reports based on a large model according to claim 5, characterized in that: In step 6.4, a version anti-counterfeiting and traceability mechanism is established throughout the entire lifecycle of the audit report. The specific steps are as follows: 6.4.1 Based on a secure hash algorithm, perform hash mapping calculations on the binary data stream of the standard format document, the system's current generation timestamp, and the corresponding auditor's identity identifier to generate a unique digital fingerprint for the standard format document. ; 6.4.

2. Digital fingerprint The watermark or document metadata is embedded into the exported standard format document and simultaneously stored in the system's audit log database for immutable and fixed filing.

Images