Audit risk point checking method, device, equipment, medium and product

Abstract

The invention discloses an audit risk point checking method and device, equipment, a medium and a product, and relates to the technical field of large language models.The method comprises the steps that a to-be-checked material file is received, and the to-be-checked material file is vectorized and then stored in a vector knowledge base; performing scene classification on the to-be-checked material file to obtain a classification result, and selecting a large language model based on the classification result; and performing risk point checking on the to-be-checked material file through the large language model to obtain a risk point checking result, and extracting text fragments in the vector knowledge base based on the risk point checking result. Therefore, the problems of low efficiency and error proneness caused by the fact that risk point induction depends on experts and mass materials need to be checked under an enterprise personalized system are solved, and the audit risk point checking efficiency is improved.

Description

Technical Field

[0001] This application relates to the field of large language model technology, and in particular to a method, apparatus, equipment, medium and product for audit risk point inspection. Background Technology

[0002] In the field of auditing, the existing audit content mainly covers financial auditing, internal control auditing, business auditing, asset auditing, auditing of major decisions, auditing of legal and regulatory compliance, auditing of social responsibility, and other special audits. In the existing technology, audit risk points need to be summarized and organized from the system documents (such as corporate internal systems, industry regulatory regulations, financial regulations, etc.) corresponding to the above audit content to identify specific inspection points. Then, the auditors will review the relevant business documents of the enterprise one by one based on these risk points to determine whether the business operations meet the audit compliance requirements.

[0003] However, the existing methods for identifying and checking audit risks have significant shortcomings: On the one hand, different companies have unique institutional systems, and risk identification requires experienced domain experts to conduct targeted analysis. Moreover, institutional documents are often lengthy and complex, making the risk identification process cumbersome and time-consuming, which seriously restricts the efficiency of audit preparation. On the other hand, risk identification requires manual verification of a large amount of company materials, which is cumbersome and has a huge workload. It is not only inefficient, but also prone to affecting the accuracy of audit results due to human fatigue or oversight.

[0004] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main purpose of this application is to provide a method, apparatus, equipment, medium, and product for auditing risk points, aiming to solve the technical problems of low efficiency and error-proneness caused by the reliance on experts to summarize risk points under the enterprise's personalized system and the need to verify a large amount of material.

[0006] To achieve the above objectives, this application proposes a method for checking audit risk points, which includes: Receive the material file to be inspected, vectorize the material file to be inspected, and store it in the vector knowledge base; The material files to be inspected are classified into scenarios to obtain classification results, and a large language model is selected based on the classification results. The large language model is used to perform risk point checks on the material document to be inspected, and the risk point check results are obtained. Based on the risk point check results, text fragments are extracted from the vector knowledge base.

[0007] In one embodiment, the large language model includes a first large language model and a second large language model. The step of classifying the material document to be inspected into scenarios, obtaining classification results, and selecting a large language model based on the classification results includes: The material file to be inspected is identified by a redundancy identification algorithm to obtain redundant information and unstructured data, and the redundant information and unstructured data are then removed from the material file to be inspected. The material file to be inspected is subjected to data noise reduction, and after the noise reduction is completed, the material file to be inspected is divided into several file sub-units based on the file logical structure and semantic content of the material file to be inspected by a text segmentation algorithm. Based on the aforementioned file sub-units, the material files to be inspected are classified into scenarios using a machine learning algorithm to obtain classification results; If the classification result indicates a professional scenario, then the first major language model is selected; If the classification result is a typical scenario, then the second largest language model is selected.

[0008] In one embodiment, before the step of classifying the material file to be inspected into scenarios, obtaining classification results, and selecting a large language model based on the classification results, the method further includes: Collect historical materials and documents; After preprocessing the historical material files, the data of the historical material files are labeled using an expert model to obtain training samples; Based on the training samples, the initial large language model is fine-tuned using a cue-based fine-tuning method to obtain the first large language model.

[0009] In one embodiment, the step of performing risk point checks on the material document to be inspected using the large language model to obtain the risk point check results includes: If the classification result is a professional scenario, then the dedicated resources for the professional scenario are loaded through the first large language model; The document slices of the material to be inspected are retrieved by using the dedicated resources; The retrieved document slices are then assembled into a context to obtain concatenated text. The concatenated text is then subjected to risk point checks using the first large language model, and the risk point check results are obtained.

[0010] In one embodiment, the step of performing risk point checks on the material document to be inspected using the large language model to obtain the risk point check results further includes: If the classification result is a normal scenario, then natural language understanding is performed on the material file to be inspected to obtain key information about the file and the context of the user's request; A semantic representation is constructed based on the key information of the file and the user request information, and key prompt words are obtained by calculating the context of the user request through a prompt word generation algorithm. The key prompts and semantic representations are input into the second large language model to obtain the risk point inspection results.

[0011] In one embodiment, after the steps of performing risk point checks on the material document to be inspected using the large language model to obtain risk point check results, and extracting text fragments from the vector knowledge base based on the risk point check results, the method further includes: A correlation analysis was performed on the risk point inspection results and text fragments to obtain a correlation map; The association graph is weighted and scored to obtain the scoring results; Based on the aforementioned correlation map and scoring results, a risk point view is obtained through visualization processing using a preset visual framework. Send the risk point view to the management terminal and receive the feedback results sent by the management terminal; Based on the feedback results, the large language model and vector knowledge base are optimized to obtain the optimized results.

[0012] Furthermore, to achieve the above objectives, this application also proposes an audit risk point detection device, which includes: The receiving module is used to receive the material file to be inspected, vectorize the material file to be inspected, and store it in the vector knowledge base. The classification module is used to classify the material files to be inspected into scenarios, obtain classification results, and select a large language model based on the classification results; The inspection module is used to perform risk point inspection on the material document to be inspected through the large language model, obtain the risk point inspection results, and extract text fragments from the vector knowledge base based on the risk point inspection results.

[0013] In addition, to achieve the above objectives, this application also proposes an audit risk point inspection device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the audit risk point inspection method as described above.

[0014] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the audit risk point inspection method described above.

[0015] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the audit risk point checking method described above.

[0016] One or more technical solutions proposed in this application have at least the following technical effects: This application proposes an audit risk point inspection method, apparatus, equipment, medium, and product. It receives a document to be inspected, vectorizes it, and stores it in a vector knowledge base. The document is then categorized by scenario to obtain a classification result, and a large language model is selected based on the classification result. The large language model is used to inspect the document for risk points, resulting in risk point inspection results. Text fragments are then extracted from the vector knowledge base based on these results. Therefore, after receiving the document to be inspected, vectorizing it and storing it in a vector knowledge base, then classifying it by scenario to select a suitable large language model, and finally using the large language model to inspect the document for risk points, obtaining risk point inspection results, and extracting text fragments from the vector knowledge base based on these results, this method solves the problem of low efficiency and error-proneness caused by relying on experts to summarize risk points under personalized enterprise systems and requiring the verification of massive amounts of materials. This improves the efficiency of audit risk point inspection. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the first embodiment of the audit risk point inspection method for this application. Figure 2 This is a flowchart illustrating the second embodiment of the audit risk point inspection method for this application. Figure 3A simplified flowchart illustrating the audit risk point detection method provided in Embodiment 2 of this application; Figure 4 This is a schematic diagram of the module structure of the audit risk point inspection device according to an embodiment of this application; Figure 5 This is a schematic diagram of the device structure of the hardware operating environment involved in the audit risk point inspection method in this application embodiment.

[0020] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0021] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0022] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0023] The main solution of this application embodiment is as follows: The material file to be inspected is identified using a redundancy identification algorithm to obtain redundant information and unstructured data, and the redundant information and unstructured data are removed from the material file to be inspected; the material file to be inspected is denoised, and after denoising, based on the file's logical structure and semantic content, the material file to be inspected is divided into several file sub-units using a text segmentation algorithm; according to the several file sub-units, the material file to be inspected is classified into scenarios using a machine learning algorithm to obtain classification results; if the classification result is a professional scenario, a first large language model is selected; if the classification result is a general scenario, a second large language model is selected. Historical material files are collected; after preprocessing the historical material files, data annotation is performed on the historical material files using an expert model to obtain training samples; based on the training samples, the initial large language model is fine-tuned using a prompting fine-tuning method to obtain a first large language model. If the classification result is a professional scenario, then the dedicated resources for that professional scenario are loaded through the first large language model; document slice retrieval is performed on the document file to be inspected using the dedicated resources to obtain retrieved document slices; context assembly is performed on the retrieved document slices to obtain concatenated text; risk point checks are performed on the concatenated text using the first large language model to obtain risk point check results. If the classification result is a general scenario, then natural language understanding is performed on the document file to be inspected to obtain key document information and the context of the user request; a semantic representation is constructed based on the key document information and the user request information, and key prompt words are calculated based on the context of the user request using a prompt word generation algorithm; the key prompt words and semantic representation are input into the second large language model to obtain risk point check results. The process involves correlation analysis of the risk point inspection results and text fragments to obtain a correlation graph; weighted scoring of the correlation graph to obtain a scoring result; visualization processing of the correlation graph and scoring result using a pre-defined visual framework to obtain a risk point view; sending the risk point view to the management terminal and receiving feedback from the management terminal; and optimizing the large language model and vector knowledge base based on the feedback result to obtain an optimized result. This solves the problem of low efficiency and error-proneness caused by the reliance on experts for risk point summarization under enterprise personalized systems and the need to verify massive amounts of material, thus realizing the inspection of audit risk points and improving the efficiency of audit risk point inspection.Based on the present invention, considering the unique differences in corporate systems across different enterprises, the need for experienced domain experts to conduct targeted analysis for risk point identification, and the often lengthy and complex nature of policy documents leading to a cumbersome and time-consuming risk point summarization process that severely restricts the efficiency of pre-audit preparation, a method for checking audit risk points was designed. The effectiveness of this method was verified during the checking of audit risk points, and the efficiency of audit risk point checking was significantly improved by the present invention.

[0024] In this embodiment, for ease of description, the audit risk point inspection device will be used as the execution subject in the following description.

[0025] Due to the high reliance on human experience and limitations of traditional processes in current audit risk point identification and inspection methods, audit efficiency and accuracy need improvement. One issue is the cumbersome process of risk point summarization. Different companies have unique systems, requiring targeted analysis by experts. Furthermore, the large size and complexity of these documents make the summarization process time-consuming, severely hindering audit preparation efficiency. Another issue is the inefficiency of manual verification. Risk point inspection requires manual review of massive amounts of business documents against key points, a cumbersome and labor-intensive process that is not only inefficient but also prone to affecting the accuracy of audit results due to human fatigue or oversight. Finally, there is insufficient domain adaptability. Audit content covers multiple areas, including finance and internal control, and different companies have significantly different systemic structures. Current methods have not established a standardized cross-company risk point handling mechanism, which also affects audit adaptability and efficiency. Therefore, current audit work faces challenges in scaling up and standardizing. Because different companies have different system frameworks, business models, and compliance requirements, if audit risk point identification and inspection methods are not adapted to individual company needs and process optimization is not implemented, efficiency and accuracy will decline.

[0026] This application provides a solution that receives a document to be inspected, vectorizes it, and stores it in a vector knowledge base; classifies the document into scenarios to obtain classification results, selects a large language model based on the classification results; performs risk point checks on the document using the large language model to obtain risk point check results, and extracts text fragments from the vector knowledge base based on the risk point check results. Thus, after receiving the document to be inspected, vectorizing it and storing it in a vector knowledge base, classifying it into scenarios to select a suitable large language model, and finally using the large language model to perform risk point checks on the document to obtain risk point check results, and extracting text fragments from the vector knowledge base based on the risk point check results, this solution solves the problem of low efficiency and error-proneness caused by relying on experts for risk point summarization under enterprise personalized systems and requiring the verification of massive amounts of materials, thereby improving the efficiency of audit risk point checks.

[0027] Based on this, the embodiments of this application provide a method for checking audit risk points, referring to Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the audit risk point inspection method of this application.

[0028] In this embodiment, the audit risk point inspection method includes steps S01 to S03: Step S01: Receive the material file to be inspected, vectorize the material file to be inspected, and store it in the vector knowledge base; Before the implementation of this embodiment, it should be clear that the existing content in the audit field covers multiple types of audits, such as financial and internal control. Risk points need to be summarized from the system documents and the audit personnel manually check the business documents to judge compliance. However, the existing method has defects. First, the enterprise system varies greatly and the documents are complicated. It is tedious and time-consuming for experts to summarize risk points. Second, manual verification of massive materials is inefficient and easily affected by omissions.

[0029] To address the aforementioned issues, this embodiment, upon receiving a material file to be inspected, first vectorizes the file and stores it in a pre-built vector knowledge base. This vector knowledge base enables unified vector representation of material files of different types and formats. Through optimized storage structure and indexing mechanisms, it achieves efficient retrieval and correlation analysis of massive amounts of text data, providing underlying data support for quickly locating relevant information during subsequent risk point inspections.

[0030] Step S02: Classify the material file to be inspected into scenarios to obtain classification results, and select a large language model based on the classification results; Subsequently, the materials to be inspected are categorized by scenario to obtain the scenario corresponding to the current material to be inspected (in this embodiment, this includes general scenarios and professional scenarios). Based on this, a large language model is selected to obtain a large language model suitable for the current scenario. In general scenarios, a basic version of the large language model is selected. This model has extensive text understanding capabilities and can handle common materials such as financial statements and basic contract terms in conventional formats. In professional scenarios, a professional version of the large language model is matched with data fine-tuned for specific fields. For example, for the internal control audit scenario, the model has learned a large amount of professional knowledge such as internal control process specifications and typical risk cases, and can accurately identify deep-seated professional risk points such as abnormal approval authority settings and failure to separate incompatible positions.

[0031] Through scenario classification and model adaptation, intelligent and professional processing of different types of materials to be inspected has been achieved, laying a model foundation for the accurate extraction of subsequent risk points.

[0032] Step S03: Perform risk point checks on the material document to be checked using the large language model to obtain risk point check results, and extract text fragments from the vector knowledge base based on the risk point check results.

[0033] Finally, the selected large language model is used to check the risk points of the materials to be inspected, and the risk point check results are obtained. Based on the risk point check results, the corresponding text fragments are extracted from the vector knowledge base. The extracted text fragments are the original evidence that is highly related to the risk points, which can provide specific contextual support for subsequent risk analysis.

[0034] Specifically, the aforementioned large language model includes a first large language model and a second large language model. Step S02 involves classifying the material document to be inspected into scenarios to obtain classification results. The step of selecting a large language model based on the classification results includes: Step S021: The material file to be inspected is identified by a redundancy identification algorithm to obtain redundant information and unstructured data, and the redundant information and unstructured data are removed from the material file to be inspected. Step S022: Denoise the data of the material file to be inspected, and after the denoising is completed, divide the material file to be inspected into several file sub-units based on the file logical structure and semantic content of the material file to be inspected using a text segmentation algorithm. Step S023: Based on the aforementioned file sub-units, the material file to be inspected is classified into scenarios using a machine learning algorithm to obtain a classification result; Step S024: If the classification result is a professional scenario, then select the first major language model; Step S025: If the classification result is a normal scenario, then select the second largest language model.

[0035] First, deep data cleaning is achieved through a categorized redundancy identification algorithm. Specifically, this includes constructing a two-dimensional identification system for semantic redundancy and format redundancy. The semantic redundancy side uses an algorithm combining TF-IDF and cosine similarity to identify repetitive paragraphs and meaningless filler words (such as high-frequency, non-information words like "as stated in this notice" and "see attachment for details"). The format redundancy side develops a tag parser to automatically remove non-content information such as HTML tags, PDF format tags, and headers and footers. In this embodiment, for unstructured data (such as garbled text in scanned documents and table-like images), the convertibility of the data is determined through an OCR quality assessment model. Low-quality unstructured data that cannot be converted is directly removed, while convertible data is temporarily stored for subsequent structured processing. This achieves accurate filtering of redundant and invalid data and improves preprocessing efficiency.

[0036] Then, data denoising and structured slicing are integrated into a coordinated process. In the data denoising stage, a terminology mapping library and context-aware error correction technology are used. On the one hand, a terminology library specific to the audit field is constructed (covering standard terms for eight types of audit content, including finance and internal control), which intelligently corrects inconsistent terminology usage in documents (such as the consistency between "accounts payable" and "accounts payable"). On the other hand, the BERT semantic model is used to analyze the context and intelligently correct grammatical errors and semantic contradictions (such as the logical conflict of "pay overdue fees within 30 days"). After denoising, a hierarchical text segmentation algorithm is used to achieve structured slicing. First, coarse segmentation is performed based on the document's logical structure (chapter and title levels). Then, a semantic boundary detection model (based on sentence vector similarity calculation using Sentence-BERT) is used to optimize the coarse segmentation results, ensuring the semantic integrity of each document subunit. At the same time, metadata such as chapter tags, core themes, and keywords are added to the subunits to provide accurate support for subsequent classification and analysis.

[0037] Subsequently, an enhanced machine learning classification algorithm is used to achieve accurate scenario classification. First, a multi-dimensional feature set is constructed, including document metadata (file type, creation time, industry tags), content features (sub-unit keywords, terminology density, topic distribution), and historical interaction features (past classification results of similar documents, high-frequency audit concerns). Then, a classification model using LightGBM and an attention mechanism is employed. LightGBM handles high-dimensional features for initial classification, while the attention mechanism assigns higher weight to core audit terminology features (such as professional terms like "internal control" and "compliance"). A dynamic threshold adjustment mechanism is introduced to optimize the classification decision boundary in real time based on the classification accuracy feedback from different industry audit scenarios, ensuring that the materials to be inspected are accurately classified into professional scenarios (such as audits of financial industry regulatory laws and regulations, and financial audits of listed companies) or routine scenarios (such as audits of general corporate attendance systems). Finally, based on the classification results, the model is accurately matched. If the classification result is a professional scenario (such as financial audit involving complex financial instruments or audit of specific industry regulatory regulations), the first language model is selected. This model is a customized LLM that has been fine-tuned specifically for the audit field. It integrates industry regulatory libraries (such as the "Accounting Standards for Business Enterprises" and the "Financial Regulatory Regulations") and typical audit cases through LoRA technology. It has the ability to deeply understand professional terminology and reason about complex compliance logic. If the classification result is a routine scenario (such as audit of general corporate internal management systems), the second language model is selected. This model is a lightweight general-purpose LLM that optimizes inference speed and resource consumption. It can efficiently handle standardized and low-complexity audit risk point inspection tasks, improving processing efficiency while ensuring basic accuracy, and achieving the dual goals of high accuracy in professional scenarios and high efficiency in routine scenarios.

[0038] More specifically, before step S02 above, which involves classifying the material document to be inspected into scenarios to obtain classification results, and selecting a large language model based on the classification results, the method further includes: Step S0201: Collect historical materials and documents; Step S0202: After preprocessing the historical material files, the historical material files are annotated with data using an expert model to obtain training samples; Step S0203: Based on the training samples, the initial large language model is fine-tuned and trained using a prompting fine-tuning method to obtain the first large language model.

[0039] First, a comprehensive audit data collection system is built to achieve efficient collection of historical materials. On the one hand, the dimensions of data sources are expanded to include not only company policies, regulatory laws and regulations, and financial regulations, but also historical audit reports, rectification tracking records, industry typical case libraries (such as audit risk cases of listed companies and compliance cases in the financial industry), and cross-departmental business data (such as financial vouchers and internal control process records), forming a multi-source data pool. On the other hand, a data quality stratification mechanism is established. Through compliance verification and value density assessment, data is divided into a core layer (such as authoritative regulations and verified audit cases), a support layer (such as corporate standards and regulations and routine business documents), and a reference layer (such as industry research reports and expert interpretations). Core layer data is prioritized for model training, while support layer and reference layer data serve as supplements to ensure the accuracy and comprehensiveness of training data.

[0040] Then, a closed loop is formed by integrating refined preprocessing and high-quality annotation. In the preprocessing stage, data quality assurance requirements are followed, and audit semantic alignment processing is added. By constructing an audit domain knowledge graph, synonymous terms (such as "compliance inspection" and "compliance verification") and cross-domain mapping relationships (such as the correspondence between financial terms and business terms) are unified. At the same time, redundant fragments and format noise in the data (such as garbled characters in scanned documents and duplicate attachment descriptions) are removed. After preprocessing, a human-machine collaborative annotation mode guided by expert models and reviewed by humans is adopted. First, the expert model (a classifier trained based on historical annotation experience) pre-annotates the data and outputs the risk point type, inspection basis, and annotation confidence level. Then, domain experts manually review and correct the annotation results with a confidence level of less than 90%, and the correction results are fed back to the expert model for iterative optimization. The annotated training samples are divided into training set, validation set, and test set in an "8:1:1" ratio, and metadata such as "scenario label (such as financial audit / internal control audit) and difficulty level" are added to improve the efficiency of sample utilization.

[0041] Finally, based on the prompting fine-tuning method (P-turning), efficient fine-tuning of the initial large language model is achieved. First, a prompting template system specific to the audit domain is designed. The template includes "system role (audit risk analyst), task instructions (risk point identification / compliance judgment), domain rules (such as the provisions of the "Enterprise Accounting Standards"), and example samples (3-5 high-confidence labeled samples)". The template is then transformed into learnable continuous prompting vectors. Then, a prompting encoder and parameter freezing strategy is adopted to freeze the main parameters of the initial large language model. The prompting vectors are optimized only through training (embedded into the model input layer and intermediate attention layer). During the training process, an audit semantic loss function is introduced to enhance the model's learning of professional terminology and compliance logic. After fine-tuning, an incremental fine-tuning mechanism is established. For example, newly labeled samples (from new audit cases and regulatory updates) are input into the model every quarter for secondary fine-tuning. A low learning rate (1e-5) is used for 1-2 rounds of training. At the same time, the model performance is monitored in real time through the validation set. When the accuracy of risk point identification drops by more than 5%, a full fine-tuning is triggered to ensure that the first large language model continues to adapt to new scenarios and requirements in the audit domain, achieving the dual goals of low resource consumption and high-precision output.

[0042] This embodiment, through the above-described scheme, specifically involves receiving the material file to be inspected, vectorizing the material file, and storing it in a vector knowledge base; classifying the material file into scenarios to obtain classification results, and selecting a large language model based on the classification results; using the large language model to perform risk point checks on the material file to obtain risk point check results, and extracting text fragments from the vector knowledge base based on the risk point check results. Therefore, after receiving the material file to be inspected, vectorizing it and storing it in a vector knowledge base, then classifying the material file into scenarios to select a suitable large language model, and finally using the large language model to perform risk point checks on the material file to obtain risk point check results, and extracting text fragments from the vector knowledge base based on the risk point check results, this approach solves the problem of low efficiency and error-proneness caused by the reliance on experts for risk point summarization under enterprise personalized systems and the need to verify massive amounts of materials, thus improving the efficiency of audit risk point checks.

[0043] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to the first embodiment described above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 In step S03, the audit risk point inspection method further includes steps S031-S034, whereby the risk point inspection results are obtained by performing risk point inspection on the document to be inspected using the large language model. Step S031: If the classification result is a professional scenario, then load the dedicated resources for the professional scenario through the first large language model; Step S032: Use the dedicated resource to perform document slice retrieval on the material file to be inspected to obtain the retrieved document slices; Step S033: Perform context assembly on the retrieved document slices to obtain concatenated text; Step S034: Perform risk point checks on the concatenated text using the first large language model to obtain the risk point check results.

[0044] First, the large language model adopts a two-layer resource loading mechanism of core and extension layers to accurately activate professional capabilities. The core layer loads basic resources specific to the audit field, including a professional terminology library aligned with terminology (covering high-frequency terms and easily confused concepts in 8 types of audit content such as finance, internal control, and regulations) and a structured authoritative rule library (such as breaking down legal provisions such as the "Accounting Standards for Business Enterprises" and "Auditing Standards" into "rule ID, applicable scenario, and judgment logic" triples). The extension layer dynamically loads scenario-based resources according to specific professional scenarios. For example, financial industry audits load financial instrument valuation rules and regulatory penalty case libraries, while medical industry audits load medical compliance indicator libraries and medical insurance policy interpretation libraries. Resource loading adopts a combination of pre-caching and real-time calling mode. Core resources are pre-loaded into the model memory to improve response speed, while extension resources are called on demand through vector library association retrieval, ensuring that the model can accurately understand professional terms and adapt to scenario-based audit needs.

[0045] Then, the integration of vector library optimization technology enables efficient and accurate document slice retrieval. First, based on the semantic embedding model, a dual vector representation of basic semantic vectors and scene feature vectors is generated for document slices of special resources and materials to be inspected. The basic semantic vectors capture the core meaning of the text, while the scene feature vectors incorporate professional scene tags (such as "financial statement audit, revenue recognition"). Then, a hybrid retrieval strategy of semantic retrieval and metadata filtering is adopted. First, the top 20 slices with semantic similarity to the materials to be inspected are matched from the vector library through near nearest neighbor (ANN) search (based on HNSW index). Then, irrelevant slices are filtered out by combining scene tags, chapter levels and other metadata. Finally, a retrieval result re-ranking mechanism is introduced. The risk relevance score between the slice and the audit task is calculated through the first major language model (including factors such as comprehensive term density and rule matching degree). The top 10 retrieval slices are output in descending order of score to ensure that the retrieval results cover the core content and focus on high-risk related segments.

[0046] Next, the concatenated text is constructed with semantic coherence as the core. First, the retrieved document slices undergo deduplication and completion preprocessing to remove duplicate or highly similar slices. For semantically incomplete slices (such as those containing only half of the clause content), context fragments are supplemented by searching a vector library. Then, an assembly strategy using logical connectors and hierarchical tags is adopted to sort the slices according to chapter logic and risk relevance. Connective connectors are added between slices (such as "According to the above clauses, the subsequent business operation requirements are:") to enhance semantic coherence. At the same time, each slice is labeled with scenario tags and rule association markers (such as "[Financial Audit - Revenue Recognition - Standard No. 14]"). Finally, the length is dynamically controlled. If the concatenated text exceeds the context window of the first large language model, the core semantic summarization technology is used to generate a semantic summary of 30% length for non-critical slices to ensure that the concatenated text retains the complete risk relevance logic while meeting the model input length requirements.

[0047] Finally, leveraging the professional capabilities of a customized LLM model, in-depth risk point output is achieved. The first language model employs a multi-round inference and cross-validation mechanism for risk point checking. The first round, based on concatenated text and loaded dedicated resources, outputs preliminary risk points (including risk description, professional basis, and scenario matching degree). The second round uses rule-based reverse verification, calling associated rules in the rule base to logically verify the preliminary risk points, eliminating misjudged risks that do not conform to the rule judgment logic. Finally, structured risk point check results are output, including six elements: risk ID, risk type, risk level (high / medium / low), specific clause reference, professional rule basis, and risk impact analysis. At the same time, the source page number and vector similarity score of each risk point are associated with the search slice, making the risk points traceable and verifiable. In addition, risk warning suggestions are output for high-risk points (such as "This risk point involves revenue recognition violations, and the corresponding business vouchers need to be checked carefully"), improving the practicality of the results.

[0048] Specifically, step S03 above, which involves performing risk point checks on the document to be inspected using the large language model to obtain the risk point check results, further includes: Step S035: If the classification result is a normal scenario, then perform natural language understanding on the material file to be inspected to obtain key information about the file and the context of the user request. Step S036: Construct a semantic representation based on the key information of the file and the user request information, and calculate the context of the user request using a prompt word generation algorithm to obtain key prompt words; Step S037: Input the key prompt words and semantic representations into the second large language model to obtain the risk point inspection results.

[0049] First, use natural language processing methods to extract core information from the documents to be inspected, such as document name, effective date, responsible party, deadline, and amount. At the same time, sort out the user requests, clarify the key points to be checked (such as the compliance of travel expense reimbursement), remove useless information, and keep only the key content.

[0050] The extracted file information and user requests are then organized into semantic features that the model can understand, including the meaning of the text, the file structure, and the specific task the user wants to search. Then, prompt words are generated using a basic template and a focus-highlighting approach. The template is fixed as "As a routine audit assistant, output risk results based on the file information [{extracted content}] and the user request [{search what}]", and the key points that the user is concerned about (such as "reimbursement amount") are highlighted in the prompt words.

[0051] Finally, the semantic features and prompt words are input into the second language model, and a fast reasoning mode is used (to ensure results within 10 seconds). After reasoning, the results are checked against a standard risk rule library (such as general regulations like "reimbursements exceeding 500 yuan require approval"). Obvious errors are removed, and a simple structured result is output, including risk description, basis, and two levels of risk (high and low). It can also provide simple compliance suggestions for low-risk points.

[0052] More specifically, after step S03 above, which involves performing risk point checks on the document to be inspected using the large language model to obtain risk point check results, and extracting text fragments from the vector knowledge base based on the risk point check results, the method further includes: Step S04: Perform correlation analysis on the risk point inspection results and text fragments to obtain a correlation map; Step S05: Perform a weighted score on the association graph to obtain the score result; Step S06: Based on the correlation map and the scoring results, perform visualization processing through a preset visual framework to obtain a risk point view; Step S07: Send the risk point view to the management terminal and receive the feedback result sent by the management terminal; Step S08: Optimize the large language model and vector knowledge base based on the feedback results to obtain the optimization results.

[0053] First, a three-layer association graph is constructed using risk, text, and rules. The first layer is semantic association, which uses Sentence-BERT to calculate the semantic similarity between risk point inspection results and text fragments, establishing association edges from risk points to similar text fragments. The second layer is positional association, which constructs source-tracing edges from risk points to their original text locations based on the page number and chapter information of the text fragments in the material to be inspected, achieving precise source tracing. The third layer is rule association, which associates the professional / general rule base entries corresponding to the risk points and adds edges to the basis of the judgment rules. The graph nodes in this embodiment are designed in three categories: risk point nodes (labeled with risk level and type), text fragment nodes (labeled with source document and page number), and rule nodes (labeled with rule ID and core content). The edge attributes label the association strength value (semantic similarity / positional relevance), forming a structured association network.

[0054] Then, a two-dimensional weighted scoring model of basic attributes and related features is constructed. The weights of basic attributes include risk level (0.6 / 0.3 / 0.1 for high / medium / low respectively) and user attention popularity (an additional 0.2 weight is given for high-frequency risk types clicked in the management terminal). The weights of related features include semantic association strength (0.3 weight if similarity ≥ 0.8) and rule matching completeness (0.2 weight based on the completeness of rule entries). The scoring calculation in this embodiment adopts the weighted summation formula: total score = Σ(basic attribute weight × attribute value) + Σ(related feature weight × feature value), with a score range of 0-10 points. The scores are divided into "core risk (≥ 8 points), important risk (6-7.9 points), and general risk (< 6 points)", providing a basis for subsequent visualization priority display.

[0055] Subsequently, based on the ECharts / Neo4j Bloom visualization framework, a three-layer interactive view was designed, comprising an overview, details, and source tracing. The overview view uses a pie chart to display the risk percentage of different rating levels, overlaid with a risk distribution heatmap to mark high-frequency risk areas. The details view presents a force-directed graph of relationships, and clicking on a risk point node expands to related text fragments and rule nodes. Text fragment nodes support one-click jump to the corresponding page number in the original text (achieving precise source tracing). The source tracing view provides hierarchical drill-down functionality from risk points to text fragments to the original document, and also supports filtering by risk level, relationship strength, and other conditions. In addition, the view in this embodiment also includes a risk description floating window, where hovering the mouse over a node displays key information such as risk description and rating basis.

[0056] Next, multiple types of feedback entry points are embedded in the risk point view. Buttons such as misjudgment markers and omission supplements are set for risk point nodes. Clicking them will pop up a feedback form, which supports uploading supporting materials (such as correct rule basis, text fragments of omissions). Interactive suggestion input boxes are set for view functions to collect management's optimization opinions on visualization effects. In addition, the system receives feedback results from the management in real time via WebSocket, automatically labels each feedback with tags such as "risk correction" and "function optimization", and associates it with the corresponding risk point / view module ID to form a feedback ledger, providing precise guidance for subsequent optimization.

[0057] Finally, a closed-loop feedback-optimization chain is established. For feedback on "risk correction," if it is a model misjudgment, the corrected risk points and supporting materials are converted into labeled samples, and incremental fine-tuning is performed on the large language model (LoRA is used for professional scenarios, and lightweight fine-tuning is used for general scenarios). If it is a deviation in vector library retrieval, the semantic vectors of the corresponding text fragments are regenerated and the vector library index is updated. For feedback on "functional optimization," if it involves view interaction issues, the parameters of the view framework are adjusted (such as drill-down level and filtering conditions). If it involves scoring logic issues, the weight coefficients of the weighted scoring model are optimized. After optimization, an optimization report is generated to record the adjusted content and effect verification data (such as the percentage increase in model accuracy and changes in vector retrieval recall), and synchronized to the management end to achieve continuous iterative upgrades of the large language model and vector knowledge base.

[0058] This embodiment, through the above-described scheme, specifically involves loading specialized resources for a professional scenario using the first large language model if the classification result indicates such a scenario; performing document slice retrieval on the document to be inspected using the specialized resources to obtain retrieved document slices; assembling the retrieved document slices into context to obtain concatenated text; and performing risk point checks on the concatenated text using the first large language model to obtain risk point check results. Thus, after receiving the document to be inspected, the document is vectorized and stored in a vector knowledge base. Subsequently, the document to be inspected is classified into scenarios to select an appropriate large language model. Finally, the large language model is used to perform risk point checks on the document to be inspected, obtaining risk point check results. Based on the risk point check results, text fragments are extracted from the vector knowledge base. This solves the problem of low efficiency and error-proneness caused by the reliance on experts for risk point summarization under enterprise personalized systems and the need to verify massive amounts of materials, thereby improving the efficiency of audit risk point checks.

[0059] For example, to help understand the implementation process of the audit risk point detection method obtained by combining this embodiment with the above embodiment one, please refer to... Figure 3 , Figure 3 A simplified flowchart of an audit risk point inspection method is provided, specifically: Risk point inspection model preparation process: (1) Historical data input: Input historical data of risk points and inspection materials.

[0060] (2) Output labeled samples: Output the risk point results and corresponding inspection materials, and have experts label (mark the results), and then use the processed text combined with the labels to construct samples.

[0061] (3) Model training and deployment: P-turning training is performed based on the LLM model, and the customized model is deployed in the production environment after completion.

[0062] Risk point inspection main process: (1) Material input and preprocessing: Input production data risk points and check material documents, and perform risk point and check material preprocessing.

[0063] (2) Scenario classification judgment: make a preliminary judgment on risk points and materials to determine whether they are classified scenarios.

[0064] (3) Non-classification scenario path: Use a custom model to judge complex situations, and execute the following steps in sequence: read content, obtain context related to user request, fill template with request and context content to obtain prompt words, input prompt words into large language model, and finally output results from large model.

[0065] (4) Classification scenario path: Traditional machine learning algorithms are used (images use OCR, machine learning algorithms, etc.), and the materials are saved to the vector knowledge base. Then, the large model and traditional algorithms access the original materials, and the results of the large model check are combined to trace the original text fragments corresponding to the vector library.

[0066] (5) Output the final result: Integrate the output of all paths to obtain and output the final result of the risk point inspection.

[0067] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the audit risk point inspection method of this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0068] This application also provides an audit risk point detection device, please refer to... Figure 4 The audit risk point inspection device includes: The receiving module 10 is used to receive the material file to be inspected, and to vectorize the material file to be inspected and store it in the vector knowledge base. Classification module 20 is used to classify the material file to be inspected into scenarios, obtain classification results, and select a large language model based on the classification results; The inspection module 30 is used to perform risk point inspection on the material document to be inspected through the large language model, obtain the risk point inspection results, and extract text fragments from the vector knowledge base based on the risk point inspection results.

[0069] The audit risk point checking device provided in this application, employing the audit risk point checking method in the above embodiments, can solve the technical problem of low efficiency and error-proneness caused by the reliance on experts for risk point summarization under enterprise personalized systems and the need to verify massive amounts of material. Compared with the prior art, the beneficial effects of the audit risk point checking device provided in this application are the same as those of the audit risk point checking method provided in the above embodiments, and other technical features in the audit risk point checking device are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0070] This application provides an audit risk point checking device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the audit risk point checking method in the above embodiment 1.

[0071] The following is for reference. Figure 5 The diagram illustrates a structural schematic suitable for implementing an audit risk point detection device according to embodiments of this application. The audit risk point detection device in embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 5 The audit risk point inspection device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0072] like Figure 5As shown, the audit risk point inspection device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory 1002 or a program loaded from a storage device 1003 into a random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the audit risk point inspection device. The processing unit 1001, the read-only memory 1002, and the random access memory 1004 are interconnected via a bus 1005. An input / output interface 1006 is also connected to the bus. Typically, the following systems can be connected to the input / output interface 1006: input devices 1007 including, for example, a touch screen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 1003 including, for example, magnetic tape, hard disk, etc.; and communication devices 1009. Communication device 1009 allows the audit risk point inspection device to communicate wirelessly or wiredly with other devices to exchange data. While the figure shows audit risk point inspection devices with various systems, it should be understood that implementation or possession of all the systems shown is not required. More or fewer systems may be implemented alternatively.

[0073] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0074] The audit risk point checking device provided in this application, employing the audit risk point checking method described in the above embodiments, can solve the technical problem of low efficiency and error-proneness caused by the reliance on experts for risk point summarization under enterprise personalized systems and the need to verify massive amounts of material. Compared with the prior art, the beneficial effects of the audit risk point checking device provided in this application are the same as those of the audit risk point checking method provided in the above embodiments, and other technical features of this audit risk point checking device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0075] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0076] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0077] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to perform the audit risk point checking method in the above embodiments.

[0078] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0079] The aforementioned computer-readable storage medium may be included in the audit risk point inspection device; or it may exist independently and not be assembled into the audit risk point inspection device.

[0080] The aforementioned computer-readable storage medium carries one or more programs. When the aforementioned one or more programs are executed by the audit risk point inspection device, the audit risk point inspection device: receives a material file to be inspected, vectorizes the material file to be inspected and stores it in a vector knowledge base; performs scenario classification on the material file to be inspected, obtains a classification result, selects a large language model based on the classification result; performs risk point inspection on the material file to be inspected through the large language model, obtains a risk point inspection result, and extracts text fragments from the vector knowledge base based on the risk point inspection result.

[0081] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0082] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0083] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0084] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., computer programs) for executing the above-described audit risk point checking method. This addresses the technical problem of low efficiency and error-proneness caused by the reliance on experts for risk point summarization under enterprise-specific systems, requiring the verification of massive amounts of material. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the audit risk point checking method provided in the above embodiments, and will not be elaborated upon here.

[0085] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the audit risk point checking method described above.

[0086] The computer program product provided in this application can solve the technical problem that relying on experts to summarize risk points under enterprise personalized systems and requiring the verification of massive amounts of material leads to low efficiency and susceptibility to errors. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the audit risk point inspection method provided in the above embodiments, and will not be repeated here.

[0087] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for checking audit risk points, characterized in that, The methods for checking audit risk points include: Receive the material file to be inspected, vectorize the material file to be inspected, and store it in the vector knowledge base; The material files to be inspected are classified into scenarios to obtain classification results, and a large language model is selected based on the classification results. The large language model is used to perform risk point checks on the material document to be inspected, and the risk point check results are obtained. Based on the risk point check results, text fragments are extracted from the vector knowledge base.

2. The audit risk point inspection method as described in claim 1, characterized in that, The large language model includes a first large language model and a second large language model. The step of classifying the material document to be inspected into scenarios and obtaining classification results, and selecting a large language model based on the classification results, includes: The material file to be inspected is identified by a redundancy identification algorithm to obtain redundant information and unstructured data, and the redundant information and unstructured data are removed from the material file to be inspected. The material file to be inspected is subjected to data noise reduction, and after the noise reduction is completed, the material file to be inspected is divided into several file sub-units based on the file logical structure and semantic content of the material file to be inspected by a text segmentation algorithm. Based on the aforementioned file sub-units, the material files to be inspected are classified into scenarios using a machine learning algorithm to obtain classification results; If the classification result indicates a professional scenario, then the first major language model is selected; If the classification result is a typical scenario, then the second largest language model is selected.

3. The audit risk point inspection method as described in claim 2, characterized in that, Before the step of classifying the material file to be inspected into scenarios, obtaining classification results, and selecting a large language model based on the classification results, the method further includes: Collect historical materials and documents; After preprocessing the historical material files, the data of the historical material files are labeled using an expert model to obtain training samples; Based on the training samples, the initial large language model is fine-tuned using a cue-based fine-tuning method to obtain the first large language model.

4. The audit risk point inspection method as described in claim 3, characterized in that, The step of performing risk point checks on the document to be inspected using the large language model to obtain the risk point check results includes: If the classification result is a professional scenario, then the dedicated resources for the professional scenario are loaded through the first large language model; The document slices of the material to be inspected are retrieved by using the dedicated resources; The retrieved document slices are then assembled into a context to obtain concatenated text. The concatenated text is then subjected to risk point checks using the first large language model, and the risk point check results are obtained.

5. The audit risk point inspection method as described in claim 2, characterized in that, The step of performing risk point checks on the document to be inspected using the large language model to obtain the risk point check results further includes: If the classification result is a normal scenario, then natural language understanding is performed on the material file to be inspected to obtain key information about the file and the context of the user's request; A semantic representation is constructed based on the key information of the file and the user request information, and key prompt words are obtained by calculating the context of the user request through a prompt word generation algorithm. The key prompts and semantic representations are input into the second large language model to obtain the risk point inspection results.

6. The audit risk point inspection method as described in claim 2, characterized in that, After the steps of performing risk point checks on the document to be inspected using the large language model, obtaining risk point check results, and extracting text fragments from the vector knowledge base based on the risk point check results, the method further includes: A correlation analysis was performed on the risk point inspection results and text fragments to obtain a correlation map; The association graph is weighted and scored to obtain the scoring results; Based on the aforementioned correlation map and scoring results, a risk point view is obtained through visualization processing using a preset visual framework. Send the risk point view to the management terminal and receive the feedback results sent by the management terminal; Based on the feedback results, the large language model and vector knowledge base are optimized to obtain the optimized results.

7. An audit risk point detection device, characterized in that, The audit risk point detection device includes: The receiving module is used to receive the material file to be inspected, vectorize the material file to be inspected, and store it in the vector knowledge base. The classification module is used to classify the material files to be inspected into scenarios, obtain classification results, and select a large language model based on the classification results; The inspection module is used to perform risk point inspection on the material document to be inspected through the large language model, obtain the risk point inspection results, and extract text fragments from the vector knowledge base based on the risk point inspection results.

8. An audit risk point detection device, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the audit risk point detection method as described in any one of claims 1 to 6.

9. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the audit risk point inspection method as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the steps of the audit risk point inspection method as described in any one of claims 1 to 6.

Images