An Optimization Method and System for Audit Working Paper Verification Based on Spatiotemporal Awareness and Pseudo-Regulations Generation Strategy

Abstract

This invention discloses an audit working paper verification optimization method and system based on spatiotemporal awareness and pseudo-regulation generation strategy. First, an audit knowledge base containing multi-level regulations and systems is constructed. Then, the audit knowledge base is used to preprocess audit working paper samples to obtain preprocessed audit working papers. Next, timeline features are extracted from the preprocessed audit working papers. This invention achieves the function of accurately reproducing the legal environment under specific spatiotemporal conditions and introducing a pseudo-regulation generation strategy to bridge the semantic gap between business and legal principles. Furthermore, the multi-source burst retrieval mechanism enables full coverage of multi-level systems. Simultaneously, the composite mode of spatiotemporal awareness, reverse generation, and burst retrieval effectively eliminates the defects in existing audit qualitative analysis, such as invalid regulation citations, high retrieval noise, and unclear scope of system application. This not only makes audit work more compliant, intelligent, and efficient but also ensures the accuracy and reliability of audit conclusions, making it suitable for widespread promotion and use.

Description

Technical Field

[0001] This invention relates to the field of audit working paper verification and optimization technology, specifically to an audit working paper verification and optimization method and system based on spatiotemporal awareness and pseudo-regulation generation strategy. Background Technology

[0002] Financial auditing is one of the cornerstones of modern capital markets and business society. Through a rigorous and systematic methodology, independent professionals verify corporate financial statements and issue authoritative opinions, greatly enhancing the reliability of financial information, reducing information asymmetry in socio-economic activities, and playing an irreplaceable role in maintaining investor confidence and ensuring the fair and effective operation of the market.

[0003] Currently, with the digitalization of business operations, the data environment faced by auditing work has fundamentally changed. The massive amount of unstructured audit working papers and the ever-evolving regulatory framework constitute a seemingly irreconcilable contradiction. Existing keyword matching technologies generally lack sensitivity to the spatiotemporal dimension, making it impossible to determine the timeliness of regulations and match regional applicability. Furthermore, most existing tools cannot distinguish the lifecycle of regulations, often recommending new regulations that only take effect after the business transaction or old regulations that have already been repealed to auditors. This spatiotemporal misalignment in search results can easily lead to problems in the audit's qualitative assessment. Legal risks exist, and traditional retrieval technologies are generally inadequate in semantic understanding, leading to retrieval systems easily missing key compliance evidence and being filled with a large amount of noisy data. This forces auditors to revert to an inefficient manual screening mode. However, relying on manual screening is difficult to accurately locate applicable clauses in such a vast database of regulations. This is not only extremely inefficient, but also prone to omissions in the application of regulations or errors in hierarchical citation due to incomplete information, which seriously affects the authority and accuracy of audit conclusions. Therefore, it is necessary to design an audit working paper verification optimization method and system based on spatiotemporal awareness and pseudo-regulation generation strategies. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies. Existing keyword matching technologies generally lack sensitivity to spatiotemporal dimensions, leading to an inability to determine the validity of regulations and match regional applicability. Furthermore, most existing tools cannot distinguish the lifecycle of regulations, often recommending new regulations that only take effect after the business transaction or old regulations that have already been repealed to auditors. This spatiotemporal misalignment in search results can easily create legal risks in audit assessment. Traditional search technologies are also generally inadequate in semantic understanding, causing search systems to easily miss key compliance evidence and be filled with a large amount of noisy data. This forces auditors to revert to an inefficient manual screening mode. However, relying on manual methods to accurately locate applicable clauses in such a vast regulatory database is not only inefficient... The extremely low efficiency and the high risk of omissions in the application of regulations or errors in hierarchical citation due to incomplete information have seriously affected the authority and accuracy of audit conclusions. This paper proposes an image target detection enhancement method and system based on controllable synthesis and reprojection correction. It can accurately restore the legal environment under specific time and space and introduce a pseudo-regulation generation strategy to fill the gap between business and legal semantics. Moreover, through the set multi-source burst retrieval mechanism, it can fully cover multi-level regulations. At the same time, the composite mode of time and space awareness, reverse generation and burst retrieval can effectively eliminate the defects of invalid regulation citation, large retrieval noise and unclear scope of application of regulations in the existing audit qualitative analysis. This not only makes the audit work more compliant, intelligent and efficient, but also ensures the accuracy and reliability of audit conclusions.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] An optimization method for audit working paper verification based on spatiotemporal awareness and pseudo-regulation generation strategy includes the following steps:

[0007] Step A: Construct an audit knowledge base containing multi-level regulations and systems, and then use the audit knowledge base to preprocess the audit working paper samples to obtain preprocessed audit working papers;

[0008] Step B: Extract timeline features from the preprocessed audit working papers, then perform spatiotemporal awareness filtering on the timeline features to obtain the filtered audit working papers.

[0009] Step C: Based on the filtered audit working papers, a combination of pseudo-legal generation strategy and database retrieval is used to search and obtain legal clause search results;

[0010] Step D: Based on the filtered audit working papers, a multi-source knowledge base burst search strategy is used to conduct targeted searches and obtain the system attribution search results;

[0011] Step E involves using a semantic re-ranking model to optimize the applicability dimension of the legal clause search results and the system attribution search results, and obtaining the qualitative basis after verification and optimization.

[0012] Step F: Automatically generate and verify the optimized audit working papers based on the optimized qualitative criteria, and complete the audit working paper verification and optimization work.

[0013] The aforementioned audit working paper verification optimization method based on spatiotemporal awareness and pseudo-regulation generation strategy includes step A, which involves constructing an audit knowledge base containing multi-level regulations and systems, and then using the audit knowledge base to preprocess the audit working paper samples to obtain preprocessed audit working papers. The specific steps are as follows.

[0014] Step A1 involves constructing an audit knowledge base that includes multi-level regulations and systems. Specifically, this involves using national laws and regulations, industry regulatory provisions, internal group systems and implementation rules of subordinate units, and combining these with the geographical scope of application to determine a multi-source knowledge acquisition strategy, thereby obtaining the audit knowledge base, as shown in formula (1).

[0015] (1)

[0016] in, It is a collection of knowledge from multiple sources. This is a multi-source knowledge acquisition strategy function. For the specific content of Article j, For the first The corresponding publication or effective timestamp of each regulation. This is the geographical identifier corresponding to Article j. For the audit field, For a set of resource types, A collection of geographical regions;

[0017] Step A2 involves preprocessing the audit working paper samples using an audit knowledge base to obtain preprocessed audit working papers. Specifically, this involves using Optical Character Recognition (OCR) and document layout analysis technology to digitize scanned documents, images, and PDF documents based on the multimodal characteristics of the audit working paper samples, obtaining multimodal knowledge collection data. Then, metadata annotation is performed on the multimodal knowledge collection data, and a traceable knowledge data genealogy is established to obtain the preprocessed audit working papers, as shown in formula (2).

[0018] (2)

[0019] in, Metadata functions, For knowledge entries, This refers to the index number of the regulatory entries in the audit working paper sample. and These are the effective date and the repeal date of the regulations, respectively. The hierarchy of the effectiveness of regulations. For industries to which the regulations apply, The geographical area to which the regulations apply.

[0020] The aforementioned audit working paper verification optimization method based on spatiotemporal awareness and pseudo-regulation generation strategy, step B, involves extracting timeline features from the preprocessed audit working papers, then applying spatiotemporal awareness filtering to the timeline features to obtain the filtered audit working papers. The specific steps are as follows.

[0021] Step B1 involves extracting timeline features from the preprocessed audit working papers. Specifically, this involves entity identification and extracting key time points in which audit issues occurred from the preprocessed audit working papers. First, determine the accounting period to which the business belongs and the geographical location where the problem occurred. Then, traverse the regulatory database and retrieve each regulation. Timeline characteristics;

[0022] Step B2 involves performing spatiotemporal awareness filtering on the timeline features and obtaining the filtered audit working papers. Specifically, this involves constructing spatiotemporal awareness filtering logic to filter key time points. with regulations By comparing timeline features and matching geographical locations with applicable regulatory regions, an effective set of regulations can be constructed. Specifically, as shown in formula (3),

[0023] (3)

[0024] in, This is the set of valid regulations after screening. For the complete set of regulations, The key time points when audit issues occur. The geographical location where the audit issue occurred. To simultaneously satisfy the spatiotemporal conditions, For regulations The applicable geographical scope.

[0025] The aforementioned audit working paper verification optimization method based on spatiotemporal awareness and pseudo-legal generation strategy, step C, involves using a combination of pseudo-legal generation strategy and database retrieval to search for legal clauses based on the filtered audit working papers, and the specific steps are as follows.

[0026] Step C1: Based on the description of audit issues in the filtered audit working papers, reverse reasoning is used to deduce the idealized regulatory text that the behavior would violate and generate a pseudo-regulatory description, as shown in formula (4).

[0027] (4)

[0028] in, This is a false regulatory description. For generative large language models, Descriptions of audit issues in the filtered audit working papers. These are clues for reverse reasoning. These are model parameters;

[0029] Step C2, generate the pseudo-regulatory description Transform into a vector and match it with the vectors of real regulations in the audit knowledge base. The similarity between them is calculated to obtain the legal clause search results, as shown in formula (5):

[0030] (5)

[0031] in, This is the initial set of candidate regulations. To obtain the similarity score The function that takes the value of each result. The cosine similarity function is used. For vector embedding functions, This is a preset threshold for the number of candidate regulations.

[0032] The aforementioned audit working paper verification optimization method based on spatiotemporal awareness and pseudo-regulation generation strategy, step D, involves using a multi-source knowledge base burst retrieval strategy to perform targeted retrieval based on the filtered audit working papers and obtain the system attribution retrieval results. The specific steps are as follows.

[0033] Step D1 involves classifying and linking the filtered audit working papers to entities and parsing the attribute characteristics of the auditee. The attribute characteristics of the audited entity mentioned above This includes the organizational level to which the audited entity belongs and the geographical location attribute of the audited entity;

[0034] Step D2: Dynamically activate the retrieval domain and construct a multi-source retrieval routing matrix based on the unit attributes, as shown in formula (6):

[0035] (6)

[0036] in, For the retrieval domain, For the national legal database, As a provincial-level legal database, The geographical location attribute of the audited entity. For the group's unified system database, This is a filtering condition function based on the attributes of the audited entity. A personalized policy database for organizations. It is an empty set;

[0037] Step D3 involves generating and executing search requests for different levels of audit knowledge bases, then aggregating the search results from each source to obtain the system attribution search results.

[0038] The aforementioned audit working paper verification optimization method based on spatiotemporal awareness and pseudo-regulation generation strategy, step E, involves using a semantic re-ranking model to optimize the applicability dimension of the legal clause retrieval results and the system attribution retrieval results, and obtaining the qualitative basis after verification optimization. The specific steps are as follows.

[0039] Step E1 involves aggregating the legal clause search results and the system attribution search results into a search candidate set. Then, the audit facts are concatenated with each regulation in the search candidate set to form a text pair for calculating the semantic relevance score, as shown in formula (7).

[0040] (7)

[0041] in, Score the semantic relevance. for Activation function For cross encoder models, For audit facts, Concatenate text sequences. These are the weights for the reordering model;

[0042] Step E2, based on the score Rank the regulations in descending order and apply a cutoff threshold. Then, weakly relevant clauses with scores below the threshold are removed, and the top-N clauses are selected as the qualitative basis for verification and optimization.

[0043] The aforementioned audit working paper verification and optimization method based on spatiotemporal awareness and pseudo-regulation generation strategy, step F, involves automatically generating the verified and optimized audit working papers based on the optimized qualitative criteria and completing the audit working paper verification and optimization operation. The specific steps are as follows.

[0044] Step F1: Construct the final audit working paper template, which includes four preset slots: problem characterization, fact description, regulatory reference, and audit recommendations.

[0045] Step F2: Based on the final audit working paper template, fill in the reordered and optimized qualitative basis and audit facts into the final audit working paper template to obtain the optimized audit text, as shown in formula (8).

[0046] (8)

[0047] in, To verify the optimized audit text, Generate a dedicated large model for the draft. To verify the qualitative basis after optimization, This is an audit working paper template. These are the fine-tuned model parameters;

[0048] Step F3 involves performing format verification on the optimized audit text and ensuring that the cited legal provisions are completely consistent with the original text, thereby obtaining the optimized audit working paper.

[0049] An audit working paper verification and optimization system based on spatiotemporal awareness and pseudo-regulation generation strategy includes an audit working paper preprocessing module, a spatiotemporal awareness filtering module, a legal clause retrieval module, a system attribution retrieval module, a semantic reordering module, and an audit working paper generation module. The audit working paper preprocessing module constructs an audit knowledge base containing multi-level regulations and systems, and then uses the audit knowledge base to preprocess audit working paper samples to obtain preprocessed audit working papers. The spatiotemporal awareness filtering module extracts timeline features from the preprocessed audit working papers, and then applies spatiotemporal awareness filtering to the timeline features to obtain filtered audit working papers. The legal clause verification and generation module... The search module is used to retrieve legal clause search results based on the filtered audit working papers using a combination of pseudo-regulation generation strategy and database retrieval; the system attribution search module is used to perform targeted retrieval based on the filtered audit working papers using a multi-source knowledge base burst search strategy and obtain system attribution search results; the semantic re-ranking module is used to optimize the applicability dimension of the legal clause search results and system attribution search results using a semantic re-ranking model and obtain the qualitative basis after verification and optimization; the audit working paper generation module is used to automatically generate the verified and optimized audit working papers based on the optimized qualitative basis and complete the audit working paper verification and optimization operation.

[0050] The beneficial effects of this invention are as follows: This invention provides an audit working paper verification and optimization method and system based on spatiotemporal awareness and pseudo-regulation generation strategies. First, an audit knowledge base containing multi-level regulations and systems is constructed. Then, the audit knowledge base is used to preprocess audit working paper samples to obtain preprocessed audit working papers. Next, timeline features are extracted from the preprocessed audit working papers, and spatiotemporal awareness filtering is applied to these timeline features to obtain filtered audit working papers. Subsequently, based on the filtered audit working papers, a combination of pseudo-regulation generation strategies and database retrieval is used to perform a search to obtain legal clause search results. Then, based on the filtered audit working papers, a multi-source knowledge base burst search strategy is used to perform a targeted search to obtain system attribution search results. Finally, the legal clause search results and system attribution search results are optimized for applicability using a semantic re-ranking model to obtain verification results. The system verifies and optimizes the qualitative basis, then automatically generates and verifies the optimized audit working papers based on the optimized qualitative basis, and completes the audit working paper verification and optimization operation. This effectively realizes that the audit working paper verification and optimization method and system can accurately restore the legal environment under specific time and space and introduce pseudo-regulation generation strategies to fill the semantic gap between business and legal theory. Moreover, through the set multi-source burst retrieval mechanism, it can fully cover multi-level systems. At the same time, the composite mode of time and space awareness, reverse generation and burst retrieval can fully eliminate the defects of existing audit qualitative analysis, such as invalid regulatory references, large retrieval noise and unclear scope of application of systems. This not only makes the audit work more compliant, intelligent and efficient, but also ensures the accuracy and reliability of audit conclusions, avoids the problems of regulatory timeliness mismatch and insufficient retrieval accuracy in audit qualitative analysis, and improves the automation and intelligence of audit operations. Attached Figure Description

[0051] Figure 1 This is an overall flowchart of an audit working paper verification optimization method based on spatiotemporal perception and pseudo-regulation generation strategy of the present invention;

[0052] Figure 2 This is a schematic diagram of the spatiotemporal sensing filtering principle of the present invention;

[0053] Figure 3 This is a schematic diagram of the pseudo-regulation generation strategy of the present invention;

[0054] Figure 4 This is a schematic diagram of the burst retrieval strategy of the present invention. Detailed Implementation

[0055] The present invention will now be further described with reference to the accompanying drawings.

[0056] like Figure 1 As shown, the present invention provides an audit working paper verification optimization method based on spatiotemporal awareness and pseudo-regulation generation strategy, comprising the following steps:

[0057] Step A involves constructing an audit knowledge base containing multi-level regulations and systems, then using this knowledge base to preprocess the audit working paper samples and obtain preprocessed audit working papers. The specific steps are as follows:

[0058] Step A1 involves constructing an audit knowledge base that includes multi-level regulations and systems. Specifically, this involves using national laws and regulations, industry regulatory provisions, internal group systems and implementation rules of subordinate units, and combining these with the geographical scope of application to determine a multi-source knowledge acquisition strategy, thereby obtaining the audit knowledge base, as shown in formula (1).

[0059] (1)

[0060] in, It is a collection of knowledge from multiple sources. This is a multi-source knowledge acquisition strategy function. For the specific content of Article j, For the first The corresponding publication or effective timestamp of each regulation. This is the geographical identifier corresponding to Article j. For the audit field, For a set of resource types, A collection of geographical regions;

[0061] Step A2 involves preprocessing the audit working paper samples using an audit knowledge base to obtain preprocessed audit working papers. Specifically, this involves using Optical Character Recognition (OCR) and document layout analysis technology to digitize scanned documents, images, and PDF documents based on the multimodal characteristics of the audit working paper samples, obtaining multimodal knowledge collection data. Then, metadata annotation is performed on the multimodal knowledge collection data, and a traceable knowledge data genealogy is established to obtain the preprocessed audit working papers, as shown in formula (2).

[0062] (2)

[0063] in, Metadata functions, For knowledge entries, This refers to the index number of the regulatory entries in the audit working paper sample. and These are the effective date and the repeal date of the regulations, respectively. The hierarchy of the effectiveness of regulations. For industries to which the regulations apply, The geographical area to which the regulations apply.

[0064] like Figure 2As shown, step B involves extracting timeline features from the preprocessed audit working papers, then performing spatiotemporal-aware filtering on these features to obtain the filtered audit working papers. The specific steps are as follows.

[0065] Step B1 involves extracting timeline features from the preprocessed audit working papers. Specifically, this involves entity identification and extracting key time points in which audit issues occurred from the preprocessed audit working papers. First, determine the accounting period to which the business belongs and the geographical location where the problem occurred. Then, traverse the regulatory database and retrieve each regulation. Timeline characteristics;

[0066] Step B2 involves performing spatiotemporal awareness filtering on the timeline features and obtaining the filtered audit working papers. Specifically, this involves constructing spatiotemporal awareness filtering logic to filter key time points. with regulations By comparing timeline features and matching geographical locations with applicable regulatory regions, an effective set of regulations can be constructed. Specifically, as shown in formula (3),

[0067] (3)

[0068] in, This is the set of valid regulations after screening. For the complete set of regulations, The key time points when audit issues occur. The geographical location where the audit issue occurred. To simultaneously satisfy the spatiotemporal conditions, For regulations The applicable geographical scope.

[0069] like Figure 3 As shown, step C involves using a combination of pseudo-legal document generation strategy and database retrieval based on the filtered audit working papers to obtain legal clause search results. The specific steps are as follows:

[0070] Step C1: Based on the description of audit issues in the filtered audit working papers, reverse reasoning is used to deduce the idealized regulatory text that the behavior would violate and generate a pseudo-regulatory description, as shown in formula (4).

[0071] (4)

[0072] in, This is a false regulatory description. For generative large language models, Descriptions of audit issues in the filtered audit working papers. These are clues for reverse reasoning. These are model parameters;

[0073] Step C2, generate the pseudo-regulatory description Transform into a vector and match it with the vectors of real regulations in the audit knowledge base. The similarity between them is calculated to obtain the legal clause search results, as shown in formula (5):

[0074] (5)

[0075] in, This is the initial set of candidate regulations. To obtain the similarity score The function that takes the value of each result. The cosine similarity function is used. For vector embedding functions, This is a preset threshold for the number of candidate regulations.

[0076] like Figure 4 As shown, step D involves using a multi-source knowledge base burst retrieval strategy to perform targeted searches based on the filtered audit working papers and obtain the system attribution search results. The specific steps are as follows:

[0077] Step D1 involves classifying and linking the filtered audit working papers to entities and parsing the attribute characteristics of the auditee. The attribute characteristics of the audited entity mentioned above This includes the organizational level to which the audited entity belongs and the geographical location attribute of the audited entity;

[0078] Step D2: Dynamically activate the retrieval domain and construct a multi-source retrieval routing matrix based on the unit attributes, as shown in formula (6):

[0079] (6)

[0080] in, For the retrieval domain, For the national legal database, As a provincial-level legal database, The geographical location attribute of the audited entity. For the group's unified system database, This is a filtering condition function based on the attributes of the audited entity. A personalized policy database for organizations. It is an empty set;

[0081] Step D3 involves generating and executing search requests for different levels of audit knowledge bases, then aggregating the search results from each source to obtain the system attribution search results.

[0082] Step E involves using a semantic re-ranking model to optimize the applicability dimension of the legal clause search results and the system attribution search results, and obtaining the qualitative basis for verification and optimization. The specific steps are as follows:

[0083] Step E1 involves aggregating the legal clause search results and the system attribution search results into a search candidate set. Then, the audit facts are concatenated with each regulation in the search candidate set to form a text pair for calculating the semantic relevance score, as shown in formula (7).

[0084] (7)

[0085] in, Score the semantic relevance. for Activation function For cross encoder models, For audit facts, Concatenate text sequences. These are the weights for the reordering model;

[0086] Step E2, based on the score Rank the regulations in descending order and apply a cutoff threshold. Then, weakly relevant clauses with scores below the threshold are removed, and the top-N clauses are selected as the qualitative basis for verification and optimization.

[0087] Step F involves automatically generating and verifying the optimized audit working papers based on the optimized qualitative criteria, and completing the audit working paper verification and optimization task. The specific steps are as follows:

[0088] Step F1: Construct the final audit working paper template, which includes four preset slots: problem characterization, fact description, regulatory reference, and audit recommendations.

[0089] Step F2: Based on the final audit working paper template, fill in the reordered and optimized qualitative basis and audit facts into the final audit working paper template to obtain the optimized audit text, as shown in formula (8).

[0090] (8)

[0091] in, To verify the optimized audit text, Generate a dedicated large model for the draft. To verify the qualitative basis after optimization, This is an audit working paper template. These are the fine-tuned model parameters;

[0092] Step F3 involves performing format verification on the optimized audit text and ensuring that the cited legal provisions are completely consistent with the original text, thereby obtaining the optimized audit working paper.

[0093] An audit working paper verification and optimization system based on spatiotemporal awareness and pseudo-regulation generation strategy includes an audit working paper preprocessing module, a spatiotemporal awareness filtering module, a legal clause retrieval module, a system attribution retrieval module, a semantic reordering module, and an audit working paper generation module. The audit working paper preprocessing module constructs an audit knowledge base containing multi-level regulations and systems, and then uses the audit knowledge base to preprocess audit working paper samples to obtain preprocessed audit working papers. The spatiotemporal awareness filtering module extracts timeline features from the preprocessed audit working papers, and then applies spatiotemporal awareness filtering to the timeline features to obtain filtered audit working papers. The legal clause verification and generation module... The search module is used to retrieve legal clause search results based on the filtered audit working papers using a combination of pseudo-regulation generation strategy and database retrieval; the system attribution search module is used to perform targeted retrieval based on the filtered audit working papers using a multi-source knowledge base burst search strategy and obtain system attribution search results; the semantic re-ranking module is used to optimize the applicability dimension of the legal clause search results and system attribution search results using a semantic re-ranking model and obtain the qualitative basis after verification and optimization; the audit working paper generation module is used to automatically generate the verified and optimized audit working papers based on the optimized qualitative basis and complete the audit working paper verification and optimization operation.

[0094] To better illustrate the effects of this invention, a specific embodiment of the audit working paper verification optimization method of this invention is described below.

[0095] This example uses a compliance audit of procurement operations at a group's subsidiary as a scenario. The example receives a set of unstructured working papers submitted by the auditors, including: one scanned PDF file, one natively editable PDF document, and two JPG photos of on-site records. This example performs data preprocessing, employing appropriate parsing strategies for different file types.

[0096] (1) For scanned PDFs, this embodiment calls the optical character recognition (OCR) engine to perform text recognition on each page image, and combines it with the document layout analysis algorithm to automatically distinguish between headers, footers, tables, figure captions and body text areas, extract only the main text content related to the audit matters, and reconstruct the paragraph logical structure to avoid semantic breaks caused by layout errors.

[0097] For native PDFs, this embodiment directly extracts their embedded text layer and skips the image recognition process, while applying layout parsing technology to correct formatting anomalies.

[0098] For JPG photos, this embodiment first performs image preprocessing, then sends the image to the OCR module to recognize the text content, and uses context rules to automatically correct common recognition errors.

[0099] (2) In this embodiment, after the text extraction is completed, a unified post-processing process is executed. This embodiment uses regular expression matching and entity recognition rules to detect and desensitize sensitive information such as ID card number, bank account number, and mobile phone number; the whole text is standardized and cleaned, including unifying the date format and numerical unit, and removing redundant blank lines and irrelevant symbols. Finally, a standardized text with clear structure and standardized content is output as the input source for subsequent analysis.

[0100] Meanwhile, this embodiment has already built an audit knowledge base covering a multi-level system of regulations in the background. The knowledge base connects to the national legal database, provincial government information disclosure platform, group internal control system, and audited entity's OA system through interfaces, collecting normative documents including national, provincial, group, and unit-level documents. Each regulation is marked with structured metadata, including effective date, repeal date, level of validity, applicable industry, and geographical scope, forming a knowledge resource pool that supports conditional filtering and targeted retrieval.

[0101] (3) Entering the spatiotemporal filtering stage, this embodiment extracts entities from the preprocessed text to accurately identify the time interval of the problem as April 1, 2020 to June 30, 2020, and the geographical location as Nanjing City, Jiangsu Province. The system traverses all regulations and filters out regulations that meet two conditions simultaneously: first, the effective period covers the extracted time interval, that is, the effective date is no later than June 30, 2020, and it has not been repealed before that date or the repeal date is no earlier than April 1, 2020; the applicable geographical area includes the whole country and Jiangsu Province. After this filtering, a subset of regulations that are binding only under this specific spatiotemporal condition is formed.

[0102] (4) In this embodiment, for non-standard expressions such as the absence of public bidding or selection information in the draft, the existence of only internal departmental negotiation records, and the failure to follow procurement procedures under the pretext of emergency repair, this embodiment adopts a reverse derivation method to automatically generate an idealized compliance requirement text that conforms to the style of legal language based on the facts of the problem. After vectorization, the idealized compliance requirement text is used to calculate the semantic similarity with each real clause in the effective legal subset and to sort them according to the degree of matching to initially recall several highly relevant clauses.

[0103] (5) This embodiment further analyzes the organizational information in the working papers, confirming that the audited entity is Subsidiary A, whose registered address and business location are both in Nanjing, and which belongs to a centrally administered group. Accordingly, this embodiment simultaneously launches four search channels: the National Legal Database, the Jiangsu Provincial Local Regulations Database, the Group Regulations Database, and Subsidiary A's Local Regulations Database. Each channel independently performs searches in its corresponding sub-database that has passed spatiotemporal filtering, and the results are aggregated into a unified set of candidate regulations.

[0104] (6) In this embodiment, the original audit facts are combined with each regulation in the candidate set into a text pair, which is then input into a semantic reordering model for relevance scoring. The model focuses on evaluating the core elements of the clauses, including large-value service procurement, emergency situation identification conditions, price comparison requirements, and record-keeping requirements. In this embodiment, the clauses are sorted from high to low scores, and a threshold is set to eliminate weakly relevant items, ultimately retaining several regulations that best fit the facts as qualitative basis.

[0105] (7) This embodiment calls a preset audit working paper template, which includes four fixed parts: problem characterization, factual description, compliance basis, and audit recommendations. In this embodiment, the selected legal provisions and original facts are filled into the corresponding positions to generate a structured draft of the working paper. The content generated in this embodiment is compared word by word with the authoritative original text in the knowledge base by the verification module to ensure that the citation is accurate. After the verification is passed, this embodiment outputs the final standardized audit working paper file, completing the fully automated processing from original unstructured materials to compliant, accurate, and archiveable audit conclusions.

[0106] In summary, the audit working paper verification and optimization method and system based on spatiotemporal awareness and pseudo-regulation generation strategy of the present invention first constructs an audit knowledge base containing multi-level regulations and systems. Then, the audit knowledge base is used to preprocess audit working paper samples to obtain preprocessed audit working papers. Next, timeline features are extracted from the preprocessed audit working papers, and spatiotemporal awareness filtering is applied to these timeline features to obtain filtered audit working papers. Subsequently, based on the filtered audit working papers, a search is performed using a combination of pseudo-regulation generation strategy and database retrieval to obtain legal clause search results. Then, based on the filtered audit working papers, a targeted search is performed using a multi-source knowledge base burst search strategy to obtain system attribution search results. Finally, the legal clause search results and system attribution search results are optimized for applicability dimension using a semantic re-ranking model to obtain verification optimization. The system uses the post-qualification basis to automatically generate and verify the optimized audit working papers, completing the audit working paper verification and optimization process. This effectively enables the audit working paper verification and optimization method and system to accurately recreate the legal environment under specific time and space conditions and introduce pseudo-regulation generation strategies to bridge the semantic gap between business and legal principles. Furthermore, the multi-source burst retrieval mechanism provides full coverage of multi-level regulations. The combined mode of time-space awareness, reverse generation, and burst retrieval effectively eliminates the defects in existing audit qualitative analysis, such as invalid regulatory citations, high retrieval noise, and unclear scope of application of regulations. This not only makes the audit work more compliant, intelligent, and efficient but also ensures the accuracy and reliability of audit conclusions, avoiding problems such as mismatch of regulatory timeliness and insufficient retrieval accuracy in audit qualitative analysis, and improving the automation and intelligence of audit operations.

[0107] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An optimization method for audit working paper verification based on spatiotemporal awareness and pseudo-regulation generation strategy, characterized in that: Includes the following steps, Step A: Construct an audit knowledge base containing multi-level regulations and systems, and then use the audit knowledge base to preprocess the audit working paper samples to obtain preprocessed audit working papers; Step B involves extracting timeline features from the preprocessed audit working papers, then performing spatiotemporal awareness filtering on these features to obtain the filtered audit working papers. The specific steps are as follows. Step B1 involves extracting timeline features from the preprocessed audit working papers. Specifically, this involves entity identification and extracting key time points in which audit issues occurred from the preprocessed audit working papers. First, determine the accounting period to which the business belongs and the geographical location where the problem occurred. Then, traverse the regulatory database and retrieve each regulation. Timeline characteristics; Step B2 involves performing spatiotemporal awareness filtering on the timeline features and obtaining the filtered audit working papers. Specifically, this involves constructing spatiotemporal awareness filtering logic to filter key time points. with regulations By comparing timeline features and matching geographical locations with applicable regulatory regions, an effective set of regulations can be constructed. Specifically, as shown in formula (3), （3） in, This is the set of valid regulations after screening. For the complete set of regulations, The key time points when audit issues occur. The geographical location where the audit issue occurred. To simultaneously satisfy the spatiotemporal conditions, For regulations The applicable geographical scope; Step C involves using a combination of pseudo-legal regulation generation and database retrieval based on the filtered audit working papers to obtain legal clause search results. The specific steps are as follows: Step C1: Based on the audit issue description in the filtered audit working papers, reverse reasoning is used to derive the idealized regulatory text that the audit issue description would violate and generate a pseudo-regulatory description, as shown in formula (4). （4） in, This is a false regulatory description. For generative large language models, Descriptions of audit issues in the filtered audit working papers. These are clues for reverse reasoning. These are model parameters; Step C2, generate the pseudo-regulatory description Transform into a vector and match it with the vectors of real regulations in the audit knowledge base. The similarity between them is calculated to obtain the legal clause search results, as shown in formula (5): （5） in, This is the initial set of candidate regulations. To obtain the similarity score The function that takes the value of each result. The cosine similarity function is used. For vector embedding functions, The preset threshold for the number of candidate regulations; Step D involves using a multi-source knowledge base burst retrieval strategy to perform targeted searches based on the filtered audit working papers and obtain the system attribution search results. The specific steps are as follows: Step D1 involves classifying and linking the filtered audit working papers to entities and parsing the attribute characteristics of the auditee. The attribute characteristics of the audited entity mentioned above This includes the organizational level to which the audited entity belongs and the geographical location attribute of the audited entity; Step D2: Dynamically activate the retrieval domain and construct a multi-source retrieval routing matrix based on the unit attributes, as shown in formula (6): （6） in, For the retrieval domain, For the national legal database, As a provincial-level legal database, The geographical location attribute of the audited entity. For the group's unified system database, This is a filtering condition function based on the attributes of the audited entity. A personalized policy database for organizations. It is an empty set; Step D3: Execute search requests for different levels of audit knowledge base, then aggregate the search results from each source to obtain the system attribution search results; Step E involves using a semantic re-ranking model to optimize the applicability dimension of the legal clause search results and the system attribution search results, and obtaining the qualitative basis after verification and optimization. Step F: Automatically generate and verify the optimized audit working papers based on the optimized qualitative criteria, and complete the audit working paper verification and optimization work.

2. The audit working paper verification optimization method based on spatiotemporal awareness and pseudo-regulation generation strategy as described in claim 1, characterized in that: Step A involves constructing an audit knowledge base containing multi-level regulations and systems, then using this knowledge base to preprocess the audit working paper samples and obtain preprocessed audit working papers. The specific steps are as follows: Step A1 involves constructing an audit knowledge base that includes multi-level regulations and systems. Specifically, this involves using national laws and regulations, industry regulatory provisions, internal group systems and implementation rules of subordinate units, and combining these with the geographical scope of application to determine a multi-source knowledge acquisition strategy, thereby obtaining the audit knowledge base, as shown in formula (1). （1） in, It is a collection of knowledge from multiple sources. This is a multi-source knowledge acquisition strategy function. For the specific content of Article j, For the first The corresponding publication or effective timestamp of each regulation. This is the geographical identifier corresponding to Article j. For the auditing field, For a set of resource types, A collection of geographical regions; Step A2 involves preprocessing the audit working paper samples using an audit knowledge base to obtain preprocessed audit working papers. Specifically, this involves using Optical Character Recognition (OCR) and document layout analysis technology to digitize scanned documents, images, and PDF documents based on the multimodal characteristics of the audit working paper samples, obtaining multimodal knowledge collection data. Then, metadata annotation is performed on the multimodal knowledge collection data, and a traceable knowledge data genealogy is established to obtain the preprocessed audit working papers, as shown in formula (2). （2） in, Metadata functions, For knowledge entries, This refers to the index number of the regulatory entries in the audit working paper sample. and These are the effective date and the repeal date of the regulations, respectively. The hierarchy of the effectiveness of regulations. For industries to which the regulations apply, The geographical area to which the regulations apply.

3. The audit working paper verification optimization method based on spatiotemporal awareness and pseudo-regulation generation strategy as described in claim 2, characterized in that: Step E involves using a semantic re-ranking model to optimize the applicability dimension of the legal clause search results and the system attribution search results, and obtaining the qualitative basis for verification and optimization. The specific steps are as follows: Step E1 involves aggregating the legal clause search results and the system attribution search results into a search candidate set. Then, the audit facts are concatenated with each regulation in the search candidate set to form a text pair for calculating the semantic relevance score, as shown in formula (7). （7） in, Score the semantic relevance. for Activation function For cross encoder models, For audit facts, Concatenate text sequences. These are the weights for the reordering model; Step E2, based on the score Rank the regulations in descending order and apply a cutoff threshold. Then, weakly relevant clauses with scores below the threshold are removed, and the top-N clauses are selected as the qualitative basis for verification and optimization.

4. The audit working paper verification optimization method based on spatiotemporal awareness and pseudo-regulation generation strategy as described in claim 3, characterized in that: Step F involves automatically generating and verifying the optimized audit working papers based on the optimized qualitative criteria, and completing the audit working paper verification and optimization task. The specific steps are as follows. Step F1: Construct the final audit working paper template, which includes four preset slots: problem characterization, fact description, regulatory reference, and audit recommendations. Step F2: Based on the final audit working paper template, fill in the reordered and optimized qualitative basis and audit facts into the final audit working paper template to obtain the optimized audit text, as shown in formula (8). （8） in, To verify the optimized audit text, Generate a dedicated large model for the draft. To verify the qualitative basis after optimization, This is an audit working paper template. These are the fine-tuned model parameters; Step F3 involves performing format verification on the optimized audit text and ensuring that the cited legal provisions are completely consistent with the original text, thereby obtaining the optimized audit working paper.

5. An audit working paper verification and optimization system based on spatiotemporal awareness and pseudo-regulation generation strategy, wherein the specific verification and optimization process of the audit working paper verification and optimization system is based on the audit working paper verification and optimization method according to any one of claims 1-4, characterized in that: It includes an audit working paper preprocessing module, a spatiotemporal awareness filtering module, a legal clause retrieval module, a system attribution retrieval module, a semantic reordering module, and an audit working paper generation module. The audit working paper preprocessing module is used to construct an audit knowledge base containing multi-level laws and regulations, and then use the audit knowledge base to preprocess the audit working paper samples to obtain the preprocessed audit working papers. The spatiotemporal awareness filtering module is used to extract timeline features from the preprocessed audit working papers, then perform spatiotemporal awareness filtering on the timeline features to obtain the filtered audit working papers. The legal clause retrieval module is used to retrieve legal clauses based on the filtered audit working papers by combining a pseudo-regulation generation strategy with database retrieval. The system attribution retrieval module is used to perform targeted retrieval based on the filtered audit working papers using a burst retrieval strategy from a multi-source knowledge base and obtain system attribution retrieval results. The semantic re-ranking module is used to optimize the applicability dimension of the legal clause retrieval results and the system attribution retrieval results using a semantic re-ranking model and obtain the qualitative basis after verification and optimization. The audit working paper generation module is used to automatically generate and verify the optimized audit working papers based on the optimized qualitative criteria and to complete the audit working paper verification and optimization work.

Images