An AI-based audit data processing and pre-audit analysis report automatic generation method and system

Abstract

The application discloses an AI-based audit data processing and pre-audit analysis report automatic generation method and system, relates to the technical field of natural language data processing, and comprises the following steps: receiving input and performing semantic standardization preprocessing, and performing semantic vectorization and intention analysis; dynamically searching an audit knowledge base to obtain a structured data table, and fusing the structured data table with business data to form collaborative data; calling an intelligent analysis engine to analyze the collaborative data and generate audit analysis results; based on a report engine combining template driving and semantic generation, dynamically adjusting a report template, and converting the audit analysis results into a report text to generate an audit report based on the report template; and based on a closed-loop linkage mechanism, triggering recalculation and report synchronous updating when source data is updated or feedback is received. Through the technical scheme, unified expression of requirements is realized, manual writing workload is significantly reduced, and timeliness and whole-process consistency of results are improved.

Description

Technical Field

[0001] This invention relates to the field of natural language data processing technology, and in particular to an AI-based method for automatically generating audit data processing and pre-audit analysis reports, as well as an AI-based system for automatically generating audit data processing and pre-audit analysis reports. Background Technology

[0002] Existing audit data analysis and report generation typically employ a hierarchical process: structured data is extracted from systems such as ERP, financial control, and marketing operations, cleaned and formatted, and then subjected to statistical analysis, clustering, or classification mining. Finally, auditors manually write and fill in charts and conclusions based on fixed templates, resulting in limited automation.

[0003] The existing process suffers from four common pain points: First, the cleaning and transformation rules are mostly preset manually, which is insufficient for recognizing massive and complex anomaly scenarios, and manual verification is costly and prone to errors. Second, the analysis relies on manually set dimensions, making it difficult to capture non-linear or dynamic risk patterns across periods and businesses, and the consistency of conclusions is affected by personnel experience. Third, report generation relies on handwritten text and fixed templates, making it difficult to dynamically restructure the structure and generate standardized risk statements based on data results. Fourth, the collection, analysis, and report generation stages are relatively independent, and the entire process needs to be manually retried after data updates, making it impossible to automatically link recalculation and iteration in real time, which is difficult to meet the needs of real-time performance and continuous optimization. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides an AI-based method and system for automatically generating audit data processing and pre-audit analysis reports. By standardizing and vectorizing semantics and parsing intent, it opens up natural language entry points and eliminates ambiguity, achieving a unified expression of requirements. Through an intelligent analysis engine configured according to semantic vectors, it enables adaptive analysis based on scenarios. Through template-driven and semantic generation linkage, it automatically transcribes quantitative conclusions into audit standard expressions, achieving dynamic adjustment and one-click generation of report structures, significantly reducing the workload of manual writing. Through a closed-loop linkage mechanism, it can automatically trigger recalculation and report synchronization updates, improving the timeliness of results and consistency throughout the entire process.

[0005] To achieve the above objectives, this invention provides an AI-based method for automatically generating audit data processing and pre-audit analysis reports, comprising: Step S1: Receive the natural language input from the auditor and perform semantic standardization preprocessing. Then, perform semantic vectorization and intent parsing on the preprocessed natural language input to obtain the target intent and semantic vector. Step S2: Based on the target intent and the semantic vector, dynamically retrieve the audit knowledge base to obtain a structured data table, and merge it with the business data read from the business system accordingly to form collaborative data; Step S3: Call the intelligent analysis engine to analyze the collaborative data and generate audit analysis results based on the semantic vector; Step S4: Based on the report engine that combines template-driven and semantic generation, dynamically adjust the report template corresponding to the audit scenario, transcribe the audit analysis results into report text, and generate an audit report based on the adjusted report template; Step S5: Based on the closed-loop linkage mechanism of data collection-analysis-reporting-feedback, the recalculation and report synchronization update in steps S2-S4 are triggered when the source data is updated or when feedback is received from auditors.

[0006] In the above technical solution, preferably, the specific process of the semantic standardization preprocessing includes: Perform spell correction, semantic completion, and semantic disambiguation on the input natural language; By using the content in the expert knowledge base, the input natural language is rewritten to obtain a normalized representation; The natural language question is rewritten based on the conversation context to obtain a statement consistent with the context.

[0007] In the above technical solution, preferably, the specific process of semantic vectorization and intent parsing includes: The input natural language text is fed into the semantic representation model, and the model captures the contextual relationships within the text through its multi-layer transformation structure, mapping the entire text into a fixed-length feature vector containing high-dimensional semantic features. Based on the contextual representation of high-dimensional semantic vectors, a prompt word template is constructed. The prompt word template includes an intent classification task description, a preset list of intent categories in the audit domain, and the user question text to be classified. The prompt word template is input into a large model, and its autoregressive generation capability is used to generate corresponding target intent labels based on the high-dimensional semantic vectors, thus completing intent parsing and output.

[0008] In the above technical solution, preferably, the specific process of step S2 includes: The target intent obtained in step S1 is compared with the semantic vector input to a pre-built vector index library for similarity retrieval; the vector index library stores audit rules, semantic vectors of business data models and associated metadata; the distance between the current semantic vector and each vector in the library is calculated, and the candidate with the highest similarity is recalled; Extract key elements from the input natural language to determine the required audit rules, business data models, and field mapping relationships; Extract field mapping tables and definition dictionaries from the defined business data model; the field mapping table defines the correspondence between natural language concepts and physical tables and fields in the business system, and the definition dictionary describes the business meaning and calculation logic of the fields; combine the slot information parsed from the natural language input, use the field mapping table to align the slots to specific business table fields, and generate query conditions; Based on the query conditions and the data source connection information contained in the business data model, a structured query statement is dynamically generated. A query request is sent to the corresponding business system database through the database connection interface, and data is read according to the target time to obtain structured data. For the acquired structured data, the data is aligned according to a unified business semantics based on the field mapping table and the caliber dictionary to resolve the issues of homonyms or homonyms. Secondly, the selected fields are calculated according to the indicator calculation logic defined in the audit rules to generate derived indicators. Finally, the data is concatenated into a wide table according to dimensions such as time and subject to form a collaborative dataset with a clear structure, consistent semantics, and risk clues, which directly supports the subsequent generation and analysis of audit reports.

[0009] In the above technical solution, preferably, the specific process of step S3 includes: The collaborative data is subjected to feature processing operations to generate analytical input features; The intelligent analysis engine is a hybrid model that includes a deep learning model and statistical analysis methods. The rule sub-engine and the model sub-engine run in parallel within the intelligent analysis engine. The rule sub-engine generates rule hit records based on the audit rule set, and the model sub-engine selects models from the model set and selects analysis dimensions and model configurations according to the configuration instructions of the semantic vector. The rule hit records and model outputs are weighted or sorted and synthesized by the fusion strategy layer to form audit analysis results.

[0010] In the above technical solution, preferably, the specific process of step S4 includes: The report templates corresponding to the audit scenarios include at least one of the following: comprehensive analysis report, single model data analysis report, and pre-audit analysis report for prefecture-level cities; The report engine selects the chapter structure and template variants based on the semantic vector; Fill the template slots with indicator tables, graphical results, and questionable items; Perform normalization mapping on terminology and indicator definitions; Generate the main body of the report and the appendices.

[0011] In the above technical solution, preferably, the specific process of step S5 includes: Based on the closed-loop linkage mechanism of data collection, analysis, reporting and feedback, when a source data change event is detected, steps S2-S4 are re-executed, and the new version report is associated with the old version.

[0012] In the above technical solution, preferably, the specific process of step S5 further includes: Based on the closed-loop linkage mechanism of data collection-analysis-reporting-feedback, when multiple rounds of dialogue are received from auditors, the target intent and semantic vector of the new round of dialogue are determined based on step S1, and the contextual association is maintained based on the task tracking of intent recognition. Steps S2-S4 are re-executed, and the new version report is associated with the old version.

[0013] In the above technical solution, preferably, before step S4, the report generation process further includes: The audit analysis results generated by the intelligent analysis engine are used as candidate SQL statements. The candidate SQL statements are ranked by confidence level, and evaluated and ranked based on syntax compliance and business fit. The sorted SQL is validated and post-processed, including semantic compliance checks and SQL performance optimization, and the SQL is intelligently rewritten. The validated SQL is used to generate the data results for the audit report.

[0014] This invention also proposes an AI-based system for automatically generating audit data processing and pre-audit analysis reports, applying the AI-based method for automatically generating audit data processing and pre-audit analysis reports disclosed in any of the above technical solutions, including: The semantic interaction module is used to receive natural language input and perform semantic standardization preprocessing. The intent parsing module is used to perform semantic vectorization and intent parsing on the preprocessed natural language input to obtain the target intent and semantic vector. The data retrieval module is used to dynamically retrieve the audit knowledge base based on the target intent and the semantic vector to obtain a structured data table, and then merge it with the business data read from the business system accordingly to form collaborative data; The intelligent analysis module is used to call the intelligent analysis engine to analyze the collaborative data and generate audit analysis results based on the semantic vector; The report generation module is used to dynamically adjust the report template corresponding to the audit scenario based on the report engine that combines template-driven and semantic generation, and transcribe the audit analysis results into report text, and generate an audit report based on the adjusted report template; The closed-loop linkage module is used to trigger data recalculation and synchronously update the audit report when the source data is updated or feedback is received.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) By adopting the technical approach of semantic standardization → vectorization → intent parsing, we can unify the entry point and standard of natural language and reduce the cost of demand communication, rule matching and field mapping.

[0016] (2) By driving knowledge base retrieval and data table mapping through intent and semantic vectors, we can connect rule base, business data model and historical cases to build a reusable data and knowledge collaboration foundation.

[0017] (3) Through data cleaning and standardization, structured collaborative data is formed, which improves the comparability and stability of the analysis.

[0018] (4) Through the rule and model hybrid architecture of the intelligent analysis engine, the rule subset and model combination are selected according to the audit scenario, covering statistical analysis, anomaly identification and multidimensional correlation, which enhances the depth and breadth of risk identification.

[0019] (5) Through the template-driven and semantically generated report engine, the chapter structure is dynamically adjusted according to the audit scenario, and the quantitative results are automatically converted into standardized expressions, so that the report can be calculated and generated immediately.

[0020] (6) By sorting SQL confidence and performing SQL verification / post-processing, the correctness and executability of data retrieval statements are improved before generating reports, reducing manual rework.

[0021] (7) End-to-end automatic iteration is achieved through closed-loop linkage to ensure the timeliness and consistency of results.

[0022] (8) Through multi-round dialogue and contextual memory, it supports interactive analysis and incremental reporting, thereby improving the efficiency of audit operations and the quality of human-machine collaboration.

[0023] (9) Realize the visual analysis of audit graphs. Based on business data such as audit models, combined with big language models and natural language interaction technology, the auditors’ questions are automatically converted into SQL and rendered into charts. Multi-dimensional analysis of audit data is carried out according to dimensions such as time, region, and electricity consumption category. Visual analysis display boards are generated, which can intuitively present the audit data trends, abnormal distributions and correlations, and provide intuitive data basis for management decision-making. Attached Figure Description

[0024] Figure 1 This is a logical schematic diagram of an AI-based audit data processing and pre-audit analysis report automatic generation method disclosed in one embodiment of the present invention; Figure 2 This is a flowchart illustrating an AI-based method for automatically generating audit data and pre-audit analysis reports, as disclosed in one embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] The present invention will now be described in further detail with reference to the accompanying drawings: like Figure 1 and Figure 2 As shown, the present invention provides an AI-based method for automatically generating audit data processing and pre-audit analysis reports, comprising: Step S1: Receive the natural language input from the auditor and perform semantic standardization preprocessing. Then, perform semantic vectorization and intent parsing on the preprocessed natural language input to obtain the target intent and semantic vector. Step S2: Based on the target intent and semantic vector, dynamically retrieve the audit knowledge base to obtain a structured data table, and merge it with the business data read from the business system accordingly to form collaborative data; Step S3: Call the intelligent analysis engine to analyze the collaborative data and generate audit analysis results based on semantic vectors; Step S4: Based on the report engine that combines template-driven and semantic generation, dynamically adjust the report template corresponding to the audit scenario, transcribe the audit analysis results into report text, and generate an audit report based on the adjusted report template; Step S5: Based on the closed-loop linkage mechanism of data collection-analysis-reporting-feedback, the recalculation and report synchronization update in steps S2-S4 are triggered when the source data is updated or when feedback is received from auditors.

[0027] In this implementation, semantic standardization and vectorization, along with intent parsing, open up the natural language entry point and eliminate ambiguity, achieving a unified expression of requirements. The intelligent analysis engine enables scenario-based adaptive analysis by configuring semantic vectors. Through template-driven and semantic generation linkage, quantitative conclusions are automatically transcribed into audit standard expressions, enabling dynamic adjustment and one-click generation of report structures, significantly reducing the workload of manual writing. Through a closed-loop linkage mechanism, recalculation and report synchronization updates can be automatically triggered, improving the timeliness of results and consistency throughout the entire process.

[0028] During implementation, auditors input natural language questions in the dialog box. First, the semantic interaction layer performs semantic standardization preprocessing, followed by semantic vectorization and intent parsing to obtain the target intent and semantic vector. Based on this, the audit rule base, business data model, and historical cases are dynamically retrieved, structured data tables are pulled, and collaborative data is formed by combining business knowledge. Then, the task chain drives the large model and analysis engine to generate analysis output. Multi-turn dialogue is supported until an audit report is generated, forming a service loop.

[0029] Within this process, the data side is implemented in a layered design of "data collection → preprocessing → analysis → report generation". Specifically, this includes: collecting structured data from ERP / finance / marketing systems, and uniformly converting and standardizing it; the analysis side supports both statistical methods and data mining methods; and the report side is carried out by templated modules.

[0030] In the above embodiments, preferably, the specific process of semantic standardization preprocessing includes: Perform spell correction, semantic completion, and semantic disambiguation on the input natural language; By using the content in the expert knowledge base, the input natural language is rewritten to obtain a normalized representation; By combining the conversational context, the natural language question is rewritten to obtain a statement consistent with the context.

[0031] During implementation, semantic standardization preprocessing includes three types of operations: a) Spelling correction based on error correction dictionaries and rules, semantic completion of incomplete expressions, and semantic disambiguation such as homonymy / referentiality resolution; b) Use the audit knowledge base to paraphrase the original questions to obtain a standardized expression with consistent wording, which is used to align "rules / models / standards". For example, "accounts receivable turnover rate decreased by 15%" is automatically rewritten as "accounts receivable collection efficiency has decreased, and there is a risk of capital occupation". c) Rewrite the context by asking multiple rounds of follow-up questions based on the dialogue context, while maintaining consistency of reference and continuity of intent.

[0032] In the above embodiments, preferably, the specific process of semantic vectorization and intent parsing includes: The input natural language text is fed into a semantic representation model (such as the BERT language model). The model captures the contextual relationships within the text through its multi-layer transformation structure, mapping the entire text into a fixed-length feature vector containing high-dimensional semantic features. Based on the contextual representation of high-dimensional semantic vectors, a prompt word template is constructed. The prompt word template includes a description of the intent classification task, a pre-defined list of intent categories in the audit domain, and the user question text to be classified. The prompt word template is input into a large model, and its autoregressive generation capability is used to generate corresponding target intent labels based on high-dimensional semantic vectors, thus completing intent parsing and output.

[0033] During implementation, the preprocessed question text is fed into a vector model and mapped into a high-dimensional semantic vector. Intent recognition is performed under the constraints of audit domain knowledge, and target intent labels are output. The semantic vectors are subsequently used for: ① nearest neighbor retrieval in the vector index library, ② providing the analysis engine with selection signals for "analysis dimension / model configuration," and ③ selecting template variants and chapter structures in the report engine.

[0034] In the above embodiments, preferably, the specific process of step S2 includes: The target intent obtained in step S1 is compared with the semantic vector input to a pre-built vector index library for similarity retrieval; the vector index library stores audit rules, semantic vectors of business data models and associated metadata; the distance between the current semantic vector and each vector in the library is calculated, and the candidate with the highest similarity is recalled; Extract key elements from the input natural language to determine the required audit rules, business data models, and field mapping relationships; Extract field mapping tables and definition dictionaries from the defined business data model; the field mapping table defines the correspondence between natural language concepts and physical tables and fields in the business system, and the definition dictionary describes the business meaning and calculation logic of the fields; combine the slot information (such as time range and unit subject) parsed from the natural language input, use the field mapping table to align the slots to specific business table fields, and generate query conditions; Based on the query conditions and the data source connection information contained in the business data model, a structured query statement is dynamically generated. A query request is sent to the corresponding business system database through the database connection interface, and data is read according to the target time to obtain structured data. For the acquired structured data, the data is aligned according to a unified business semantics based on the field mapping table and the caliber dictionary to resolve the issues of homonyms or homonyms. Secondly, the selected fields are calculated according to the indicator calculation logic defined in the audit rules to generate derived indicators. Finally, the data is concatenated into a wide table according to dimensions such as time and subject to form a collaborative dataset with a clear structure, consistent semantics, and risk clues, which directly supports the subsequent generation and analysis of audit reports.

[0035] During implementation, based on the target intent and semantic vector, similarity retrieval is performed in the knowledge base to recall audit rules, business data models, and historical cases. Combined with the auditee and the audit period, structured data tables are read from the enterprise's ERP / financial control / marketing system through interfaces. Subsequently, fields and definitions are aligned according to the business data model definition and integrated into collaborative data for use by the subsequent engine.

[0036] In the above embodiments, preferably, the specific process of step S3 includes: Perform feature processing on collaborative data to generate analytical input features; The intelligent analysis engine is a hybrid model that includes deep learning models and statistical analysis methods. The rule sub-engine and model sub-engine run in parallel within the intelligent analysis engine. The rule sub-engine generates rule hit records based on the audit rule set, and the model sub-engine selects models from the model set and selects analysis dimensions and model configurations according to the configuration instructions of semantic vectors. By combining the rule hit records and model outputs through a fusion strategy layer, a weighted or sorted synthesis is formed to generate audit analysis results.

[0037] The intelligent analysis engine is a hybrid of "rules + models": The rule sub-engine executes the audit rule set and generates rule hit records; The model sub-engine selects from the statistical analysis and data mining algorithm pool, such as descriptive statistics, trend analysis, association rules, clustering, classification, etc., and selects the analysis dimensions and model parameters according to the configuration instructions carried by the semantic vector; The fusion strategy layer performs weighted / ranked synthesis of "rule hit + model output" to form audit analysis results.

[0038] In the above embodiment, preferably, the specific process of step S4 includes: The report templates corresponding to the audit scenarios include at least one of the following: comprehensive analysis report, single model data analysis report, and pre-audit analysis report for prefecture-level cities; The report engine selects chapter structure and template variants based on semantic vectors; Fill the template slots with indicator tables, graphical results, and questionable items; Perform normalization mapping on terminology and indicator definitions; Generate the main body of the report and the appendices.

[0039] In this implementation, the report engine adopts an architecture that combines template-driven and semantic generation: three types of templates are pre-set: comprehensive analysis report, single model data analysis report, and pre-audit analysis report for prefecture-level cities; the rule engine can dynamically adjust the chapter structure and level of detail, and fill the template slots with the indicator tables, graphical visualization results and suspicious items obtained from the analysis; at the same time, the large language model is used to transcribe the structured analysis conclusions into expressions that conform to audit standards, and automatically generate the report text and attachments.

[0040] Specifically, the data analysis report of a single model is mainly based on a specified audit model. Users interact with the system through natural language to automatically generate a well-structured and detailed risk analysis report. The report includes an overview, classification analysis, analysis conclusions, and rectification suggestions, which can quickly locate abnormal issues and assist auditors in accurately verifying high-risk clues.

[0041] The pre-audit analysis report for specific cities and prefectures integrates data from multiple audit models to generate a pre-audit risk report, clarifying the key areas and high-risk zones for on-site audits. This report, focusing on the core risks of electricity pricing in the audited city, is applied in the pre-audit phase and, through collaboration with the audit project, generates a list of key audit recommendations to guide pre-audit communication.

[0042] The comprehensive analysis report primarily aggregates audit model results to generate a holistic risk insight analysis report, supporting strategic decision-making. This report systematically presents a complete picture of audit risks, including risk distribution, trend analysis, and unit analysis. It is applied to annual audit planning and management decision-making scenarios. The audit department can periodically push risk evolution trend analysis reports to business management departments, providing data support for electricity price policy adjustments and audit resource allocation.

[0043] In the above embodiments, preferably, the specific process of step S5 includes: Based on the closed-loop linkage mechanism of data collection, analysis, reporting and feedback, when a source data change event is detected, steps S2-S4 are re-executed, and the new version report is associated with the old version.

[0044] In addition, it can simultaneously collect feedback from auditors to optimize NLP extraction and model parameters, enabling continuous iteration.

[0045] In the above embodiments, preferably, the specific process of step S5 further includes: Based on the closed-loop linkage mechanism of data collection-analysis-reporting-feedback, when multiple rounds of dialogue are received from auditors, the target intent and semantic vector of the new round of dialogue are determined based on step S1, and the contextual association is maintained based on the task tracking of intent recognition. Steps S2-S4 are re-executed, and the new version report is associated with the old version.

[0046] During implementation, in a continuous session, S1 is repeated for a new round of natural language input to maintain task tracking and contextual relevance, drive S2-S4 to recalculate and synchronously report versions.

[0047] In the above embodiment, preferably, before step S4, the report generation process further includes: The audit analysis results generated by the intelligent analysis engine are used as candidate SQL. The candidate SQL is ranked by confidence and evaluated and ranked based on syntax compliance and business fit. Perform validation and post-processing on the sorted SQL, including semantic compliance checks and SQL performance optimization, and intelligently rewrite the SQL. The validated SQL is used to generate the data results for the audit report.

[0048] Specifically, this process serves as a pre-processing check for report generation, performing confidence ranking and post-verification processing on candidate queries generated by the intelligent analysis engine (including structured data expressions produced by large models). The ranking dimensions cover syntactic compliance and business fit; verification includes semantic compliance checks and performance optimization, and intelligent rewriting when necessary. The results that pass the verification are then used for report generation and visualization in the database.

[0049] This invention also proposes an AI-based system for automatically generating audit data processing and pre-audit analysis reports, applying the AI-based method for automatically generating audit data processing and pre-audit analysis reports disclosed in any of the above embodiments, including: The semantic interaction module is used to receive natural language input and perform semantic standardization preprocessing. The intent parsing module is used to perform semantic vectorization and intent parsing on the preprocessed natural language input to obtain the target intent and semantic vector. The data retrieval module is used to dynamically retrieve the audit knowledge base based on target intent and semantic vectors to obtain structured data tables, and then integrate them with business data read from the business system to form collaborative data. The intelligent analysis module is used to call the intelligent analysis engine to analyze collaborative data and generate audit analysis results based on semantic vectors; The report generation module is used to dynamically adjust the report template corresponding to the audit scenario based on the report engine that combines template-driven and semantic generation, and transcribe the audit analysis results into report text, and generate audit reports based on the adjusted report template; The closed-loop linkage module is used to trigger data recalculation and synchronously update the audit report when the source data is updated or feedback is received.

[0050] The AI-based audit data processing and pre-audit analysis report automatic generation system disclosed in the above embodiments has modules whose functions correspond to the steps of the AI-based audit data processing and pre-audit analysis report automatic generation method disclosed in the above embodiments. During implementation, the above embodiments are referred to for operation, and will not be repeated here.

[0051] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for automatically generating audit data processing and pre-audit analysis reports based on AI, characterized in that, include: Step S1: Receive the natural language input from the auditor and perform semantic standardization preprocessing. Then, perform semantic vectorization and intent parsing on the preprocessed natural language input to obtain the target intent and semantic vector. Step S2: Based on the target intent and the semantic vector, dynamically retrieve the audit knowledge base to obtain a structured data table, and merge it with the business data read from the business system accordingly to form collaborative data; Step S3: Call the intelligent analysis engine to analyze the collaborative data and generate audit analysis results based on the semantic vector; Step S4: Based on the report engine that combines template-driven and semantic generation, dynamically adjust the report template corresponding to the audit scenario, transcribe the audit analysis results into report text, and generate an audit report based on the adjusted report template; Step S5: Based on the closed-loop linkage mechanism of data collection-analysis-reporting-feedback, the recalculation and report synchronization update in steps S2-S4 are triggered when the source data is updated or when feedback is received from auditors.

2. The AI-based audit data processing and pre-audit analysis report automatic generation method according to claim 1, characterized in that, The specific process of the semantic standardization preprocessing includes: Perform spell correction, semantic completion, and semantic disambiguation on the input natural language; By using the content in the expert knowledge base, the input natural language is rewritten to obtain a normalized representation; The natural language question is rewritten based on the conversation context to obtain a statement consistent with the context.

3. The AI-based audit data processing and pre-audit analysis report automatic generation method according to claim 1, characterized in that, The specific process of semantic vectorization and intent parsing includes: The input natural language text is fed into the semantic representation model, and the model captures the contextual relationships within the text through its multi-layer transformation structure, mapping the entire text into a fixed-length feature vector containing high-dimensional semantic features. Based on the contextual representation of high-dimensional semantic vectors, a prompt word template is constructed. The prompt word template includes an intent classification task description, a preset list of intent categories in the audit domain, and the user question text to be classified. The prompt word template is input into a large model, and its autoregressive generation capability is used to generate corresponding target intent labels based on the high-dimensional semantic vectors, thus completing intent parsing and output.

4. The AI-based audit data processing and pre-audit analysis report automatic generation method according to claim 1, characterized in that, The specific process of step S2 includes: The target intent obtained in step S1 is compared with the semantic vector input to a pre-built vector index library for similarity retrieval; the vector index library stores audit rules, semantic vectors of business data models and associated metadata; the distance between the current semantic vector and each vector in the library is calculated, and the candidate with the highest similarity is recalled; Extract key elements from the input natural language to determine the required audit rules, business data models, and field mapping relationships; Extract field mapping tables and definition dictionaries from the defined business data model; the field mapping table defines the correspondence between natural language concepts and physical tables and fields in the business system, and the definition dictionary describes the business meaning and calculation logic of the fields; combine the slot information parsed from the natural language input, use the field mapping table to align the slots to specific business table fields, and generate query conditions; Based on the query conditions and the data source connection information contained in the business data model, a structured query statement is dynamically generated. A query request is sent to the corresponding business system database through the database connection interface, and data is read according to the target time to obtain structured data. For the acquired structured data, the data is aligned according to a unified business semantics based on the field mapping table and the caliber dictionary to resolve the issues of homonyms or homonyms. Secondly, the selected fields are calculated according to the indicator calculation logic defined in the audit rules to generate derived indicators. Finally, the data is concatenated into a wide table according to dimensions such as time and subject to form a collaborative dataset with a clear structure, consistent semantics, and risk clues, which directly supports the subsequent generation and analysis of audit reports.

5. The AI-based audit data processing and pre-audit analysis report automatic generation method according to claim 1, characterized in that, The specific process of step S3 includes: The collaborative data is subjected to feature processing operations to generate analytical input features; The intelligent analysis engine is a hybrid model that includes a deep learning model and statistical analysis methods. The rule sub-engine and the model sub-engine run in parallel within the intelligent analysis engine. The rule sub-engine generates rule hit records based on the audit rule set, and the model sub-engine selects models from the model set and selects analysis dimensions and model configurations according to the configuration instructions of the semantic vector. The rule hit records and model outputs are weighted or sorted and synthesized by the fusion strategy layer to form audit analysis results.

6. The AI-based audit data processing and pre-audit analysis report automatic generation method according to claim 1, characterized in that, The specific process of step S4 includes: The report templates corresponding to the audit scenarios include at least one of the following: comprehensive analysis report, single model data analysis report, and pre-audit analysis report for prefecture-level cities; The report engine selects the chapter structure and template variants based on the semantic vector; Fill the template slots with indicator tables, graphical results, and questionable items; Perform normalization mapping on terminology and indicator definitions; Generate the main body of the report and the appendices.

7. The AI-based audit data processing and pre-audit analysis report automatic generation method according to claim 1, characterized in that, The specific process of step S5 includes: Based on the closed-loop linkage mechanism of data collection, analysis, reporting and feedback, when a source data change event is detected, steps S2-S4 are re-executed, and the new version report is associated with the old version.

8. The AI-based audit data processing and pre-audit analysis report automatic generation method according to claim 7, characterized in that, The specific process of step S5 also includes: Based on the closed-loop linkage mechanism of data collection-analysis-reporting-feedback, when multiple rounds of dialogue are received from auditors, the target intent and semantic vector of the new round of dialogue are determined based on step S1, and the contextual association is maintained based on the task tracking of intent recognition. Steps S2-S4 are re-executed, and the new version report is associated with the old version.

9. The AI-based audit data processing and pre-audit analysis report automatic generation method according to claim 1, characterized in that, Prior to step S4, the report generation process further includes: The audit analysis results generated by the intelligent analysis engine are used as candidate SQL statements. The candidate SQL statements are ranked by confidence level, and evaluated and ranked based on syntax compliance and business fit. The sorted SQL is validated and post-processed, including semantic compliance checks and SQL performance optimization, and the SQL is intelligently rewritten. The validated SQL is used to generate the data results for the audit report.

10. An AI-based system for automatically generating audit data processing and pre-audit analysis reports, characterized in that: The method for automatically generating audit data processing and pre-audit analysis reports based on AI, as described in any one of claims 1 to 9, includes: The semantic interaction module is used to receive natural language input and perform semantic standardization preprocessing. The intent parsing module is used to perform semantic vectorization and intent parsing on the preprocessed natural language input to obtain the target intent and semantic vector. The data retrieval module is used to dynamically retrieve the audit knowledge base based on the target intent and the semantic vector to obtain a structured data table, and then merge it with the business data read from the business system to form collaborative data. The intelligent analysis module is used to call the intelligent analysis engine to analyze the collaborative data and generate audit analysis results based on the semantic vector; The report generation module is used to dynamically adjust the report template corresponding to the audit scenario based on the report engine that combines template-driven and semantic generation, and transcribe the audit analysis results into report text, and generate an audit report based on the adjusted report template; The closed-loop linkage module is used to trigger data recalculation and synchronously update the audit report when the source data is updated or feedback is received.

Images