A system and method for verifying repeat offenders based on big data auditing and artificial intelligence.

Abstract

This invention relates to the field of computer data processing technology, specifically to a system and method for verifying repeated offenses based on big data auditing and artificial intelligence. The system includes: a data input module that acquires a list of historical and current audit issues for the same entity; a data processing module that performs compliance verification and structured transformation; a data calculation module that cleans and segments the audit qualitative, issue description, and legal basis texts, converts them into dense vectors using a text embedding model, calculates the similarity between current and historical issues across dimensions, weights the results according to preset weights to obtain a comprehensive similarity, and compares it with a threshold to filter potential clues; and an intelligent judgment module that triggers a retrieval enhancement generation mechanism for potential clues, retrieves matching context from a knowledge base, constructs a prompt word input large language model with the current issue information, determines whether a repeated offense relationship exists, and outputs the result. This solution can optimize computer system resource allocation during auditing, improve computational efficiency, and enhance the accuracy of identifying repeated offenses.

Description

Technical Field

[0001] This invention relates to the field of computer data processing technology, specifically to a system and method for verifying repeat offenders based on big data auditing and artificial intelligence. Background Technology

[0002] In corporate internal auditing practice, a long-standing pain point is the recurring issue of the same unit or similar problems appearing repeatedly in multiple audits (i.e., repeated offenses). This not only reduces the effectiveness of audit supervision but also reflects weaknesses in the audit rectification tracking and accountability processes. Current audit rectification work emphasizes establishing a ledger registration and closure system and a full-process tracking and inspection mechanism, and utilizing information platforms to achieve dynamic and diversified management of problems. In daily monitoring, big data auditing systems regularly push risk issues through data mining and require verification and rectification within a specified time limit. However, existing systems and procedures lack automated tools specifically designed for cross-project and cross-time intelligent comparison of repeatedly audited offenses, relying heavily on manual verification, which is inefficient and prone to oversight.

[0003] Specifically, existing audit information systems face the following technical challenges when conducting investigations into repeated offenses: The large volume of unstructured audit data from multiple sources is difficult to standardize, limiting the accuracy of cross-period audit data comparisons. In audit practice, the issue lists of historical audit projects and current audit projects for the same audit entity generally originate from different systems and formats, containing a large amount of unstructured information such as textual qualitative and descriptive information and the legal basis for these qualitative aspects. Current technology typically employs manual, item-by-item comparison for cross-period verification, which not only consumes significant human resources but also, due to the subjective nature of human judgment, is highly prone to omissions and misjudgments, making it difficult to guarantee the objectivity and consistency of the verification results.

[0004] Traditional data processing methods load all historical data at once, resulting in inefficient resource allocation within the computer system and low computational efficiency. Current technologies typically load all historical audit data into memory at once for processing when comparing historical and current data. When the historical data volume is large, server memory usage becomes excessive, and computing resources cannot be dynamically allocated based on data characteristics. This leads to a significant decrease in equipment efficiency, lengthy verification and processing cycles, and an inability to meet the timeliness requirements of digital audit rectification implementation.

[0005] Existing technologies lack standardized, multi-dimensional quantitative comparison methods and objective, intelligent judgment mechanisms, resulting in insufficient accuracy in identifying repeat offenders. While existing audit systems can perform preliminary data mining and risk alerts, the identification results rely on human experience when determining whether a repeat offender relationship exists between current and historical audit issues, making it difficult to guarantee accuracy. Summary of the Invention

[0006] The purpose of this invention is to propose a system and method for verifying repeated offenses based on big data auditing and artificial intelligence, which can optimize the allocation of computer system resources and improve computing efficiency during auditing, while also improving the accuracy of identifying repeated offenses.

[0007] To achieve the above objectives, in a first aspect, this invention proposes a repeat offender verification system based on big data auditing and artificial intelligence, comprising: The data input module is used to obtain historical audit project issue list data and current audit project issue list data for the same audit entity; The data processing module is used to perform compliance verification and structure transformation on the raw data output by the data input module, and output standardized structured data. The data calculation module is used to perform text cleaning and word segmentation on the structured data. The processed text of the audit problem qualitative, audit problem description and qualitative legal basis is input into the text embedding model to be converted into dense vectors. Based on the vectors of each dimension, the similarity between the current audit problem and the historical audit problem in each dimension is calculated, and the comprehensive similarity is calculated by weighting according to the preset weight ratio. The comprehensive similarity is compared with the preset threshold to screen out potential clues. The intelligent judgment module is used to trigger the retrieval enhancement generation mechanism for the potential clues, retrieve context content that matches the potential clues from the knowledge base, construct structured prompt words with the context content and the current audit issue information, input them into the large language model, and the large language model comprehensively judges whether there is a repeated audit and recidivism relationship between the current audit issue and the historical audit issues, and outputs the judgment result.

[0008] Beneficial effects of the basic solution: This solution uses a text embedding model, retrieval enhancement generation, and a large language model architecture to perform structured parsing, vector conversion, and multi-dimensional similarity weighted calculations on massive amounts of heterogeneous audit data. Through modular division of labor, the full historical and current audit data are split into various functional modules for step-by-step processing. This avoids the problems of excessive memory consumption, computing power congestion, and computational lag caused by loading massive amounts of unstructured audit data all at once on a single machine. Without upgrading the server hardware configuration, it shortens the processing time of large batches of audit data, improves the operation of computer system data storage, scheduling, and computation, achieves fine-grained allocation of hardware resources, and optimizes the overall machine performance.

[0009] Unlike traditional methods that involve meticulously reviewing audit files across years and projects for comparison, which can easily lead to missed, incorrect, or statistical oversights, this system uses algorithms to transform audit problem descriptions, qualitative entries, and legal bases into standardized, dense vectors. It then uses multi-dimensional weighted similarity quantification, combined with enhanced retrieval to generate a matching knowledge base. A large model comprehensively analyzes the relationships between recurring audits and repeat offenses, utilizing text feature mapping and vector space operation logic that follow natural laws to accurately uncover objective content correlations between problem data from different audit periods. This eliminates identification errors caused by subjective human judgment and significantly improves the accuracy of identifying recurring audit and repeat offense clues.

[0010] Traditional methods require significant manpower and time to cross-reference audit issues from previous years for the same entity. This system automates the entire process, from data import and organization to vectorization calculations, clue screening, and intelligent judgment. This reduces the need to repeatedly review paper documents and fragmented electronic files, shortening the investigation cycle for recurring violations. The system provides standardized, structured judgment conclusions that can be directly integrated with existing audit rectification information management systems. This eliminates the need for manual secondary processing and summarization of verification results, optimizing the user experience for auditors conducting rectification tracking and problem tracing through information systems, and facilitating digital improvements in audit supervision.

[0011] This solution uses compliance regulations and historical audit precedents from the knowledge base as the basis for the large model's reasoning. By enhancing the generation mechanism through retrieval, it binds objective legal provisions and historical case data to generate prompt words. This ensures that the model's judgment logic focuses on the objective characteristics of existing audit regulations and historical issues. It avoids the problem of inconsistent standards for identifying repeated violations due to differences in the professional experience and familiarity with regulations among different auditors. This makes the verification conclusions objective and standardized, and improves the operational effectiveness of the digital audit supervision system from a technical perspective.

[0012] As a feasible preferred embodiment, the data input module includes: The template upload input submodule is used to provide download and upload of standardized input templates. The templates are preset with fields for problem characterization, problem description, and legal basis. Users fill in the problem list of historical audit projects and the problem list of current audit projects according to the template specifications and then upload them. The system performs format verification on the uploaded template files. The project selection input submodule is used to obtain a list of available projects by connecting with the audit project management system. Users select the current audit project and one or more historical audit projects, and the system automatically pulls the audit issue list data associated with the selected project. Choose one of the two input methods. After input is complete, the system will automatically trigger the data verification process.

[0013] As a feasible preferred embodiment, the data processing module includes: The data validation submodule is used to judge the compliance of the input data in terms of encoding format, file type, and number of fields. It checks whether key fields are empty, whether the problem classification belongs to the system's preset classification system, and whether the legal basis field can match the valid entries in the system's built-in legal knowledge base. The structured parsing submodule is used to parse the validated raw data, map each column in the raw data to the system's internal standard field names, mark records with missing key fields as invalid records and filter them, and output a standardized structured data table.

[0014] As a feasible preferred embodiment, the data calculation module includes: The data merging submodule is used to merge the structured historical issue list with the current issue list into a unified dataset; The text cleaning and word segmentation submodule is used to clean the audit text using regular expressions, remove noisy data, and call the Chinese word segmentation engine to perform word segmentation on the cleaned text. The text vectorization submodule is used to feed the text of the three dimensions of audit problem qualitative analysis, audit problem description and qualitative legal basis after word segmentation into the text embedding model, and output a dense vector representation with fixed dimensions. The comprehensive judgment submodule is used to calculate the similarity of audit issues, the similarity of audit issue descriptions, and the similarity of audit basis based on cosine similarity, and to calculate the comprehensive similarity by weighting them according to a preset weight ratio. The comprehensive similarity is then compared with a set threshold, and data that is below the threshold is marked as irrelevant and filtered.

[0015] As a feasible preferred solution, the text embedding model is a Chinese semantic vector model that undergoes domain-adaptive fine-tuning based on audit domain corpus; the formula for calculating the comprehensive similarity is: Overall similarity = Audit issue similarity × Audit issue similarity weight + Audit issue description similarity × Audit issue description similarity weight + Audit basis similarity × Audit basis similarity weight.

[0016] As a feasible preferred embodiment, the intelligent determination module includes: The knowledge base retrieval submodule is used to transform the problem description text of potential clues into query vectors through a text embedding model, perform approximate nearest neighbor retrieval in the vector index of the knowledge base, recall multiple knowledge base records with the smallest cosine distance to the query vector, sort them by cosine similarity, filter irrelevant records, and then concatenate them into context text. The semantic and reasoning decision-making submodule is used to construct a structured prompt word template that includes role settings, judgment criteria, context injection, data to be judged, and output format constraints. After embedding the context text and current and historical audit issue information into the prompt words, it is sent to the large language model to obtain a structured output that includes the judgment result, confidence level, and judgment reason. The data entry and retention submodule is used to merge and organize audit data that is confirmed to be repeatedly audited and then save it to the knowledge base.

[0017] As a feasible preferred solution, the judgment criteria in the semantic and reasoning decision submodule include: the audit objects of the current problem and the historical problem must be the same unit or the same responsible entity; the problem characteristics of the current problem and the historical problem must belong to the same or highly related categories; the current problem must have occurred after the historical problem was pointed out, and the rectification status of the historical problem must be unrectified or inadequate; when the confidence level output by the large language model is lower than the preset confidence level threshold, the system marks the judgment result as pending manual review.

[0018] As a feasible and preferred option, it also includes: The threshold configuration module is used to customize similarity thresholds, which include high similarity thresholds, medium similarity thresholds, and low similarity thresholds. The threshold types include comprehensive similarity thresholds, qualitative similarity thresholds for questions, and similarity thresholds based on qualitative criteria. The association level selection module is used to select the classification level of audit issues for association comparison to adjust the comparison scope. It supports comparison by primary or secondary classification. When the issue list comes from multiple subordinate units under the superior unit, it automatically aggregates the data of all subordinate units for full verification. The results display and functional interaction module is used to output the comparison results in a standardized table format. The table includes issue details, similarity scores for each dimension, judgment results and judgment basis information, and supports result download. The comparison record viewing module is used to store historical comparison task records, and supports retrieval, viewing, and review. The knowledge base management module is used to manage the audit data in the knowledge base. It uses an approximate nearest neighbor index structure for vector indexing and writes newly confirmed repeat offender data into the knowledge base to achieve incremental updates after each comparison task is completed.

[0019] Secondly, the present invention also proposes a method for verifying repeated offenses based on big data auditing and artificial intelligence, applied to the system as described in any one of claims 1 to 8, characterized by comprising the following steps: Step S1: Obtain the historical audit project issue list data and the current audit project issue list data for the same audit entity; Step S2: Input the historical audit project data and the current audit project data by uploading a template or selecting a project; Step S3: Perform format validation and field mapping on the input data, and output a standardized structured data table; Step S4: Set the similarity threshold and select the comparison classification level and range; Step S5: The structured data is cleaned, segmented and vectorized. The cosine similarity of the three dimensions of audit problem characterization, audit problem description and characterization legal basis is calculated respectively. The comprehensive similarity is calculated by weighting according to the preset weight ratio and compared with the threshold to output potential clues. Step S6: Search the knowledge base for the potential clues to obtain matching context content, construct structured prompt words with the current audit issue information, input them into the large language model for reasoning and judgment, and output the final judgment result of whether it constitutes a repeated offense. Step S7: The final judgment results are presented in a standardized table format and can be downloaded. At the same time, the data confirmed as repeated offenders are stored in the knowledge base. Step S8: Archive the comparison task to support subsequent retrieval and review.

[0020] As a feasible preferred solution, in step S6, the knowledge base retrieval includes: converting the problem description text of potential clues into a query vector, performing an approximate nearest neighbor search in the vector index of the knowledge base, recalling multiple knowledge base records with the smallest cosine distance, filtering irrelevant records, and concatenating them into context text. The reasoning and judgment process includes: constructing a structured prompt word template containing role settings, judgment criteria, context injection, data to be judged, and output format constraints, and feeding it into a large language model to obtain a structured output containing the judgment result, confidence level, and judgment reason; the judgment criteria include judgment conditions for the identity of the audit object, the qualitative relevance of the problem, and the rectification status of historical problems. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the architecture of a repeat offender verification system based on big data auditing and artificial intelligence.

[0022] Figure 2 This is a logical diagram of a method for verifying repeated offenses based on big data auditing and artificial intelligence. Detailed Implementation

[0023] To make the technical solution and advantages of this application clearer, the technical solution of the present invention will be further described in detail below with reference to the accompanying drawings. It is understood that the specific embodiments described herein are only some embodiments of the present invention, and are only used to explain this application, not to limit it. It should be noted that the technical features or combinations of technical features described in the following embodiments should not be considered isolated; they can be combined with each other to achieve better technical effects. The same reference numerals appearing in the accompanying drawings of the following embodiments represent the same features or components, and can be applied to different embodiments.

[0024] Furthermore, unless otherwise defined, the technical or scientific terms used in this invention description shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains.

[0025] The present invention will now be described in further detail with reference to the accompanying drawings.

[0026] Reference Figure 1 This disclosure provides a system for verifying repeated offenses based on big data auditing and artificial intelligence, including a data input module, a data processing module, a data calculation module, an intelligent judgment module, a threshold configuration module, an association level selection module, a result display and function interaction module, a comparison record viewing module, and a knowledge base management module.

[0027] The data input module serves as the data entry point for the entire system, providing two flexible data input methods that balance standardization and operational efficiency, ensuring that the input data meets the requirements for audit issue comparison. It includes a template upload input submodule and a project selection input submodule.

[0028] The template upload input submodule provides a download portal for standardized input templates. The templates include pre-defined core fields such as problem type, problem description, and legal basis, as well as filling guidelines. This includes the following: The system includes a template download button on the front-end interface. Clicking this button generates and downloads a preset-format Excel or CSV template file. The template includes the following required fields: Audit Project Number, Audit Entity Name, Audit Project Name, Issue Number, Issue Classification (Primary / Secondary Category), Issue Description (Detailed Text Description), Legal Basis for Classification (Referenced Legal Articles or Policy Clauses Numbers), Issue Amount (if any), and Rectification Status.

[0029] Users fill out two compliant audit issue lists for the same entity according to the template specifications: one for historical audit projects and one for the current audit project.

[0030] Users can upload two completed template files through the front-end upload interface.

[0031] The system automatically validates the uploaded template for formatting. Validation rules include, but are not limited to: whether the file format is a preset format (e.g., xlsx or csv), whether required fields are complete, and whether field data types conform to specifications (e.g., whether the problem classification belongs to a preset category enumeration value, and whether the legal basis is a valid article number format). Only after successful validation can the next step be taken; otherwise, a specific error message is returned, requiring the user to correct and re-upload.

[0032] The project selection input submodule is used to select a target audit project from the audit project management list. The system automatically reads the audit issue list associated with the selected project as the input data source, eliminating the need for manual uploading. This includes the following: The system interfaces with the existing audit project management system via API to obtain a list of available projects. Project status restrictions: Only audit projects with statuses of "Closed," "Archived," or "Reporting Stage" can be selected.

[0033] In the project selection interface, you can select a current audit project (a new project to be audited) and multiple historical audit projects, supporting a one-to-many comparison mode of "current project - multiple historical projects".

[0034] After confirming the selection, the system automatically retrieves the audit issue list data associated with the selected project from the audit project management system, including fields such as issue type, issue description, and legal basis for the type.

[0035] The system performs compliance verification on the retrieved data, and automatically transfers the data to the data verification and structuring processing module after the verification is passed.

[0036] Users can choose any input method, but two methods cannot be used simultaneously. After input, the system will automatically trigger the data verification process. Only after the verification is passed can the subsequent threshold configuration and comparison process begin.

[0037] The data processing module receives the raw data output from the data input module, performs compliance checks and structure transformations on it, and provides standardized input for subsequent vectorization processing. It includes a data validation submodule and a structure parsing submodule.

[0038] The data validation submodule is used to perform compliance and format checks on uploaded audit items, verifying whether the data meets the system's processing requirements. It includes the following: Determine if the encoding format (UTF-8), file type, and number of fields of the uploaded file are consistent with the template.

[0039] Check whether key fields are empty, whether the problem classification belongs to the system's preset classification system (such as primary classification including: financial management, engineering management, procurement management, etc.; secondary classification is a subdivision under primary classification), and whether the legal basis field can match valid entries in the system's built-in legal knowledge base.

[0040] If the verification fails, the system will return a message to the user containing the specific error fields and the reason for the error, requiring the user to correct the error and resubmit.

[0041] The structured parsing submodule is used to parse the incoming raw data using the Pandas library, transforming it into structured data that the system can process. It includes the following: The system calls Python's Pandas data processing library to read the uploaded data file and generate a structured data table.

[0042] Map each column in the original data to a standard field name within the system. For example, map "Problem Name" to "Problem Description", "Problem Category" to "Problem Characterization", and "Legal Basis" to "Qualitative Legal Basis".

[0043] For missing values ​​in non-critical fields, the system fills them with default values ​​or marks them as empty strings; for missing values ​​in critical fields (problem type, problem description, legal basis), the system marks the record as invalid and filters it.

[0044] Output a standardized structured data table and pass it to the data calculation module.

[0045] The data computation module is used for multiple deep processing steps, including a data merging submodule, a text cleaning and word segmentation submodule, a text vectorization submodule, and a comprehensive judgment submodule.

[0046] The data merging submodule is used to merge historical and current audit data using Pandas. The system merges the structured historical issue list output by the previous module with the current issue list into a unified dataset, providing a data foundation for subsequent item-by-item comparisons.

[0047] The text cleaning and word segmentation submodule is used to clean the data and perform text segmentation. It combines Pandas and regular expressions to clean the data, removing irrelevant symbols, spaces, etc. The Jieba Chinese word segmentation component is used to perform Chinese text segmentation on audit texts (problem descriptions, qualitative analysis, etc.).

[0048] Specifically, the system performs regular cleaning on the "Problem Description" field and the "Qualitative Regulatory Basis" field of each audit problem respectively, removing noise data such as special characters, extra spaces, line breaks, etc.; then calls the Jieba word segmentation engine to perform accurate mode word segmentation on the cleaned Chinese text, cutting the continuous text string into an independent sequence of words as the input for subsequent vectorization.

[0049] The text vectorization sub-module is used to input the cleaned and segmented text (audit problem qualification, audit problem description, audit basis) into a vector model and convert it into a dense vector that can be computed by a computer.

[0050] Specifically, the system sends the three core text fields of each audit problem - namely, "Audit Problem" (qualitative description), "Audit Problem Description" (detailed text), and "Audit Basis" (regulatory provisions) - into the text embedding model respectively.

[0051] In this embodiment, the text embedding model specifically uses the Qwen-embedding Chinese semantic vector model, which is domain-adaptively fine-tuned based on the Tongyi Qianwen base for audit domain corpus. The fine-tuning training set contains 50,000 pairs of manually annotated audit problem similarities and 200,000 unlabeled audit regulatory provisions. During the training process, the InfoNCE contrastive learning loss function is used, the batch size is set to 32, the learning rate is 2e-5, and a total of 10 epochs are trained. The maximum Token length limit for the model input text is 8192. For text exceeding the length, a tail truncation strategy is adopted, and the insufficient part is filled with zero vectors. The model finally outputs a dense vector representation with a fixed dimension of 1024. This vector can effectively represent the distribution characteristics of audit problems in the semantic space, ensuring that the cosine distance between audit problems with similar semantics in the vector space is significantly smaller than that between problems with unrelated semantics, and at the same time giving full play to the semantic understanding advantage of the Qwen-embedding model for long text audit records. [[ID=】]

[0052] Specifically, when sending the above three core text fields into the text embedding model for vectorized encoding respectively, the system adopts the following specific encoding and pooling strategies: Input text formatting and instruction prefix injection: To enable the model to better distinguish the semantic roles of different fields, the system dynamically concatenates a specific instruction prefix (InstructPrefix) for each field before inputting the text into the Qwen-embedding model. For the "Audit Problem Qualitative" field, the prefix "Instruct: Identify audit problem category\nQuery:" is concatenated; for the "Audit Problem Description" field, the prefix "Instruct: Retrieve similar audit problem descriptions\nQuery:" is concatenated; and for the "Qualitative Legal Basis" field, the prefix "Instruct: Match audit legal provisions\nQuery:" is concatenated. The concatenated complete string serves as the final input sequence for the model.

[0053] Attention Mask and Padding Mechanism: To address the issue of varying input sequence lengths, the system transforms the input text into a Token ID sequence using a tokenizer. For sequences shorter than the maximum Token length limit (8192), a specific padding marker ([PAD] Token) is added to the end of the sequence instead of directly padding with zero vectors. Simultaneously, the system generates a binary attention mask vector of the same length as the input sequence, where the value at the position corresponding to the real Token is 1, and the value at the position corresponding to the [PAD] Token is 0. During the multi-head self-attention calculation process, this attention mask is used to reset the attention weight at the padding position to negative infinity, thereby ensuring that the padding portion does not participate in the semantic feature calculation and avoiding noise interference.

[0054] Hidden State Extraction and Mean Pooling Strategy: After the text sequence is processed by the Transformer encoder of the Qwen-embedding model, the output is a token-level hidden state matrix with shape [sequence length, 1024]. To transform it into a fixed-dimensional sentence-level dense vector, the system adopts a mean pooling strategy. The specific calculation process is as follows: First, the hidden states corresponding to the [PAD] token are filtered out using the attention mask mentioned above, retaining only the hidden state vectors of the real tokens; then, the arithmetic mean of the 1024-dimensional hidden state vectors of all real tokens is calculated along the sequence dimension to obtain an initial 1024-dimensional sentence vector; finally, L2 normalization is performed on the initial sentence vector to make its magnitude 1. The normalized 1024-dimensional vector is the final dense vector representation of this field.

[0055] Alignment of long and short text vector spaces: Given that "qualitative audit questions" are usually short texts (such as "financial management - fund management"), while "description of audit questions" are long texts, in order to alleviate the distribution bias of long and short texts in the vector space, when fine-tuning the Qwen-embedding model, the system deliberately constructs positive sample pairs of "short text qualitative - long text description" in the training set, and uses the temperature coefficient (Temperature Parameter, set to 0.05 in this embodiment) in the InfoNCE loss function for scaling, forcing the model to ignore the difference in text length during encoding and focus on the extraction of core semantic features, thereby ensuring that the vectors of the three dimensions have high comparability in the same semantic space.

[0056] The comprehensive judgment submodule calculates the final comprehensive similarity based on a certain weight ratio. The formula is as follows: Overall Similarity = Audit Issue Similarity × Audit Issue Similarity Weight + Audit Issue Description Similarity × Audit Issue Description Similarity Weight + Audit Basis Similarity × Audit Basis Similarity Weight In this embodiment, the similarity of audit issues accounts for 40%, the similarity of audit issue descriptions accounts for 40%, and the similarity of audit basis accounts for 20%; the similarity of each dimension is calculated using cosine similarity. The formula for calculating cosine similarity is:

[0057] In the formula, A is the vector representation of a certain dimension of the current audit problem, B is the vector representation of the corresponding dimension of the historical audit problem, A·B is the dot product of the two vectors, and ||A|| and ||B|| are the Euclidean norms (modulus lengths) of the two vectors, respectively. The cosine similarity ranges from zero to one. The closer the value is to one, the more similar the two texts are semantically; the closer the value is to zero, the greater the semantic difference.

[0058] The overall similarity score is compared to a set threshold. If the score does not meet the threshold (below the threshold), the data is marked as irrelevant and filtered out.

[0059] The intelligent judgment module is used to intelligently identify potential clues that have passed the basic similarity threshold screening. It includes a knowledge base retrieval submodule, a semantic and reasoning decision-making submodule, and a data entry and retention submodule.

[0060] The knowledge base retrieval submodule is used to trigger the RAG (Retrieval Enhancement Generation) mechanism, retrieve contextual content that highly matches the clue from the internal knowledge base, and input it into the Large Language Model (LLM).

[0061] Specifically, the system reuses the Qwen-embedding text embedding model in the data calculation module to transform the problem description text of potential repeat offenders into a 1024-dimensional query vector. This query vector is then used to perform an Approximate Nearest Neighbor (ANN) retrieval in the FastGPT knowledge base's vector index. During the retrieval, the Top-K recall is set to 5, meaning the 5 knowledge base records with the smallest cosine distance to the query vector are retrieved.

[0062] Each recalled knowledge base record includes a description of the historical audit case, the nature of the problem, the legal basis for the nature of the problem, the rectification status, and information about the auditing unit. The system sorts the five recalled records from highest to lowest cosine similarity and filters out irrelevant records with a similarity below 0.5, concatenating the remaining records into contextual text. If the number of remaining records after filtering is zero, the RAG enhancement step is skipped, and the original clue data is directly fed into the large language model for judgment.

[0063] The Semantic and Reasoning Decision Submodule utilizes the semantic understanding and reasoning capabilities of LLM to gain a deep understanding of the context and the current problem, and comprehensively judge whether there is a genuine case of repeated offenses.

[0064] Specifically, in the semantics and reasoning decision-making submodule, the system constructs prompt words and obtains judgment results in the following manner: The system predefines a structured prompt word template, which includes the following components: Role setting section: The big language model is set as a senior audit expert, and its task is to determine whether there is a repeated audit relationship between the current audit problem and the historical audit problem; Judgment Criteria: The judgment rules for repeated offenses are embedded in the prompts. The judgment rules include: the audit objects of the current problem and the historical problem must be the same unit or the same responsible entity; the current problem and the historical problem must be classified as the same or highly related categories; the current problem must have occurred after the historical problem was pointed out, and the rectification status of the historical problem must be unrectified or inadequate. Context injection section: The historical audit cases, problem characterization descriptions, and rectification tracking records recalled by the knowledge base retrieval submodule are used as reference context, and prompt words are embedded after being wrapped with reference material tags; The data to be judged section includes the current audit issue's qualitative nature, description, legal basis for the qualitative nature, and corresponding information of historical audit issues, wrapped in labels for current and historical issues and then embedded with prompt words. Output format constraints: The large language model is required to output the judgment results in JSON format. The JSON structure contains three fields: the judgment field takes the value "yes" or "no", indicating whether it constitutes repeated offenses; the confidence field takes the value of a floating-point number between 0 and 1, representing the confidence level of the judgment; and the reason field is a string type, representing a natural language description of the reason for the judgment.

[0065] The specific judgment logic has two possible outcomes: A "No" result indicates that although the issues are similar, they are not truly repeat offenses, and the data has been filtered out. For example, a historical issue might be "Company A failed to conduct bidding for a contract as required," while the current issue is "Company B failed to conduct bidding for a contract as required." Although the texts are highly similar, they involve different companies and do not constitute repeat offenses.

[0066] A judgment of "yes" confirms the issue as a "repeated offense." For example, if historical and current issues not only have similar texts but also involve the same unit, the same type of problem, and failure to implement rectification measures, LLM will determine, based on the overall context, that it is a genuine case of repeated offenses.

[0067] Regarding the method of obtaining LLM confidence: After the system sends the structured prompts mentioned above into the large language model, the large language model outputs a confidence score in the confidence field when generating a JSON response, based on its self-assessment of the degree of matching between the judgment criteria and the data to be judged. The system parses the JSON response returned by the large language model, extracts the value of the confidence field as the LLM confidence, the value of the reason field as the judgment reason, and the value of the judgment field as the final judgment result. When the confidence is lower than the preset confidence threshold (preferably set to 0.7 in this embodiment), the system marks the judgment result as pending manual review and prompts the user to confirm manually with a prominent icon in the result display module.

[0068] The data entry and retention submodule is used to merge and organize all audit data identified as "repeatedly offending" using Pandas. Structured data that meets the criteria is then saved to the FastGPT knowledge base for subsequent audit report generation or risk warnings.

[0069] The threshold configuration module is used to customize similarity thresholds to meet the refined requirements of audit issue comparison and avoid misjudgment and omissions. It includes a similarity threshold submodule, a threshold definition submodule, and a threshold type submodule.

[0070] The similarity threshold submodule is used to set the threshold according to the comparison requirements, and it also has a default basic configuration stored in advance.

[0071] The threshold definition submodule is used to define the threshold for similarity. In this embodiment, it includes a high similarity threshold, a medium similarity threshold, and a low similarity threshold.

[0072] Specifically, high similarity is defined as complete semantic consistency, with only paraphrasing, word order adjustments, or minor detail additions, and no difference in core information. Moderate similarity is defined as high thematic relevance, with no difference in core information. Figure 1 Similarity is defined as high overlap in key information, differing only in details or wording. Low similarity is defined as sharing only a few non-core keywords, with different themes or core intentions, and superficial vocabulary overlap.

[0073] The threshold type submodule is used to define the similarity threshold for each comparison type, including the overall similarity threshold, the qualitative similarity threshold for the question, and the qualitative basis similarity threshold.

[0074] Specifically, overall similarity refers to the degree of overlap in the overall assessment of the issues. Qualitative similarity refers to the degree of overlap in the types of issues, which can be compared through the core qualitative descriptions of the issues. Qualitative basis similarity refers to whether the basis for the qualitative assessment of the issues (legal provisions, institutional norms, etc.) is the same, avoiding misjudgments caused by different basis.

[0075] The association level selection module is used to implement hierarchical association control for audit issue comparison. It can flexibly adjust the comparison scope according to the verification needs of different levels of units, ensuring accurate issue comparison and strong adaptability. It includes an advanced association settings submodule, a default comparison rules submodule, a subordinate unit data aggregation submodule, and a filtering function submodule.

[0076] The advanced correlation settings submodule is used to select the primary or secondary category of audit issues for correlation comparison, meeting the needs of refined comparison. Based on the selection, the system only performs comparison calculations on issues under the selected category; issues from other categories are not included in the current comparison. In this embodiment, the system provides a drop-down selection box, allowing users to choose to compare by primary category (e.g., "Financial Management") or secondary category (e.g., "Financial Management - Funds Management").

[0077] The default comparison rules submodule is used to treat uploaded files or selected items as the same verification subject and perform a full set match between the historical issue list and the current issue list.

[0078] The subordinate unit data aggregation submodule is used to automatically aggregate the data of all subordinate third-level units when the problem list comes from multiple third-level units under the second-level unit, and perform a full verification, and attach the third-level unit information in the output results.

[0079] The filtering function submodule is used to filter the aggregated three-level unit data and flexibly adjust the comparison range.

[0080] The results display and functional interaction module is used to present the comparison results in an intuitive and standardized form, support verification and review, and support rectification tracking. It includes a results display submodule and a results download submodule.

[0081] The results display submodule outputs results in a standardized tabular format, clearly listing the problem details (problem type, problem description, and legal basis for the type), the relevant fields, similarity scores for each dimension (total similarity, qualitative similarity, and legal basis similarity), the judgment result (whether it is a repeat offense), and the judgment basis, among other core information.

[0082] The table columns include: serial number, current project number, current issue number, current issue characterization, current issue description, current legal basis, relevant field, historical project number, historical issue number, historical issue characterization, historical issue description, historical legal basis, overall similarity score, characterization similarity score, legal basis similarity score, LLM judgment result, LLM confidence level, and judgment reason.

[0083] The results download submodule is used to download and save the comparison results in a standardized table format for easy offline verification and archiving without redirecting to another page. In this embodiment, a download results button is built into the results display area; clicking it directly downloads and saves the comparison results.

[0084] The comparison record viewing module is used to store the historical records of all past repeated offense identification tasks. It supports retrieval, viewing, and review, and enables traceability of comparison records. It includes a retrieval area submodule, a record list area submodule, and an operation area submodule.

[0085] The search area submodule is used to navigate to historical tasks based on historical sessions. In this embodiment, clicking the history entry on the current single page will bring up a search and record list pop-up window; after entering search criteria, the pop-up window will filter the history in real time to quickly locate the target task.

[0086] The record list section submodule is used to display all historical comparison tasks in a list format. Each record displays information such as task number, comparison time, involved projects, and comparison result summary.

[0087] The operation area submodule is used for interactive operations on historical records. In this embodiment, each historical record corresponds to a "View Details" button. Clicking the "View Details" button on a historical record in the pop-up window switches the pop-up window to the comparison result details for that task, allowing you to view the complete recognition details; closing the pop-up window returns you to the current single page.

[0088] The knowledge base management module manages the audit data in the FastGPT knowledge base, providing data support for RAG retrieval, and persistently stores the final results output by the intelligent judgment module. It includes the following: The knowledge base is initialized by pre-setting basic knowledge such as a legal provisions library, a historical audit case library, and a problem classification library in the audit field, which serve as the underlying data for RAG retrieval.

[0089] Knowledge base update: After each comparison task is completed, newly confirmed data on repeat offenders is automatically written into the knowledge base, realizing the continuous accumulation and updating of the knowledge base.

[0090] Knowledge base retrieval optimization: The system adopts vector indexing technology to ensure that it can quickly and accurately recall the most relevant contextual content to the current clue from massive audit data, thus ensuring the input quality of LLM reasoning.

[0091] Specifically, the implementation methods of vector indexing technology and knowledge base update mechanism in the knowledge base management module are as follows: The knowledge base's vector index uses an HNSW (Hierarchical Navigable Small World) approximate nearest neighbor index structure. The index parameters are set as follows: maximum number of connections M=16, search range during construction efConstruction=200, and search range during query efSearch=100. This index structure ensures retrieval accuracy while supporting millisecond-level retrieval response times even with millions of audit data points.

[0092] The data structure of each record in the knowledge base includes the following fields: unique record identifier ID, audit unit name, audit project number, problem type, problem description, legal basis for the type, rectification status, entry timestamp, and a 1024-dimensional dense vector generated by the Qwen-embedding model.

[0093] The knowledge base update process is as follows: When the intelligent judgment module confirms a record of repeated offenses, the data entry and retention submodule concatenates the problem description, problem classification, and relevant legal basis text for that record. Then, it uses the Qwen-embedding model to generate a corresponding 1024-dimensional dense vector. Subsequently, all fields of the record and its vector are written into the HNSW index of the FastGPT knowledge base, achieving incremental updates. After the knowledge base is updated, the newly written data can be immediately retrieved by subsequent RAG searches without needing to rebuild the full index.

[0094] Reference Figure 2 This disclosure provides a method for verifying repeated offenses based on big data auditing and artificial intelligence, including the following:

[0095] Step S1: Obtain the problem list data of the historical and current audit projects of the same unit through the data interface of the big data audit system.

[0096] Step S2: Input historical audit project data and current audit project data through template upload or project selection in the data input module.

[0097] Step S3: The data processing module performs format validation and field mapping on the input data and outputs a standardized data table.

[0098] Step S4: Set the similarity threshold through the threshold configuration module, and select the comparison classification level and range through the association level selection module.

[0099] Step S5: The vectorization and weighted similarity calculation module performs text cleaning, Jieba word segmentation, and vectorization on the structured data. It calculates the cosine similarity of the audit question, audit question description, and audit basis in three dimensions, and calculates the comprehensive similarity with a weight of 4:4:2. After comparing with the threshold, it outputs potential clues.

[0100] In step S6, the intelligent judgment module performs knowledge base retrieval and LLM reasoning on potential clues, and outputs the final judgment result of repeated offenders.

[0101] In step S7, the results display and functional interaction module presents the final results in a standardized table and supports downloading; at the same time, the data is stored in the FastGPT knowledge base.

[0102] Step S8: Historical record management. The comparison record viewing module archives the current comparison task, supporting subsequent retrieval and review.

[0103] The above content is merely an embodiment of the present invention. Commonly known structures and characteristics of the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can improve and implement this solution based on the guidance provided in this application and their own capabilities. Some typical known structures or operating methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A system for verifying repeat offenders based on big data auditing and artificial intelligence, characterized in that: include: The data input module is used to obtain historical audit project issue list data and current audit project issue list data for the same audit entity; The data processing module is used to perform compliance verification and structure transformation on the raw data output by the data input module, and output standardized structured data. The data calculation module is used to perform text cleaning and word segmentation on the structured data. The processed text of the audit problem qualitative, audit problem description and qualitative legal basis is input into the text embedding model to be converted into dense vectors. Based on the vectors of each dimension, the similarity between the current audit problem and the historical audit problem in each dimension is calculated, and the comprehensive similarity is calculated by weighting according to the preset weight ratio. The comprehensive similarity is compared with the preset threshold to screen out potential clues. The intelligent judgment module is used to trigger the retrieval enhancement generation mechanism for the potential clues, retrieve context content that matches the potential clues from the knowledge base, construct structured prompt words with the context content and the current audit issue information, input them into the large language model, and the large language model comprehensively judges whether there is a repeated audit and recidivism relationship between the current audit issue and the historical audit issues, and outputs the judgment result.

2. The repeat offender verification system based on big data auditing and artificial intelligence as described in claim 1, characterized in that, The data input module includes: The template upload input submodule is used to provide download and upload of standardized input templates. The templates are preset with fields for problem characterization, problem description, and legal basis. Users fill in the problem list of historical audit projects and the problem list of current audit projects according to the template specifications and then upload them. The system performs format verification on the uploaded template files. The project selection input submodule is used to obtain a list of available projects by connecting with the audit project management system. Users select the current audit project and one or more historical audit projects, and the system automatically pulls the audit issue list data associated with the selected project. Choose one of the two input methods. After input is complete, the system will automatically trigger the data verification process.

3. The repeat offender verification system based on big data auditing and artificial intelligence as described in claim 1, characterized in that, The data processing module includes: The data validation submodule is used to judge the compliance of the input data in terms of encoding format, file type, and number of fields. It checks whether key fields are empty, whether the problem classification belongs to the system's preset classification system, and whether the legal basis field can match the valid entries in the system's built-in legal knowledge base. The structured parsing submodule is used to parse the validated raw data, map each column in the raw data to the system's internal standard field names, mark records with missing key fields as invalid records and filter them, and output a standardized structured data table.

4. The repeat offender verification system based on big data auditing and artificial intelligence as described in claim 1, characterized in that, The data calculation module includes: The data merging submodule is used to merge the structured historical issue list with the current issue list into a unified dataset; The text cleaning and word segmentation submodule is used to clean the audit text using regular expressions, remove noisy data, and call the Chinese word segmentation engine to perform word segmentation on the cleaned text. The text vectorization submodule is used to feed the text of the three dimensions of audit problem qualitative analysis, audit problem description and qualitative legal basis after word segmentation into the text embedding model, and output a dense vector representation with fixed dimensions. The comprehensive judgment submodule is used to calculate the similarity of audit issues, the similarity of audit issue descriptions, and the similarity of audit basis based on cosine similarity, and to calculate the comprehensive similarity by weighting them according to a preset weight ratio. The comprehensive similarity is then compared with a set threshold, and data that is below the threshold is marked as irrelevant and filtered.

5. A repeat offender verification system based on big data auditing and artificial intelligence as described in claim 4, characterized in that, The text embedding model is a Chinese semantic vector model that undergoes domain-adaptive fine-tuning based on audit domain corpus; the formula for calculating the comprehensive similarity is: Overall similarity = Audit issue similarity × Audit issue similarity weight + Audit issue description similarity × Audit issue description similarity weight + Audit basis similarity × Audit basis similarity weight.

6. The repeat offender verification system based on big data auditing and artificial intelligence according to claim 1, characterized in that, The intelligent determination module includes: The knowledge base retrieval submodule is used to transform the problem description text of potential clues into query vectors through a text embedding model, perform approximate nearest neighbor retrieval in the vector index of the knowledge base, recall multiple knowledge base records with the smallest cosine distance to the query vector, sort them by cosine similarity, filter irrelevant records, and then concatenate them into context text. The semantic and reasoning decision-making submodule is used to construct a structured prompt word template that includes role settings, judgment criteria, context injection, data to be judged, and output format constraints. After embedding the context text and current and historical audit issue information into the prompt words, it is sent to the large language model to obtain a structured output that includes the judgment result, confidence level, and judgment reason. The data entry and retention submodule is used to merge and organize audit data that is confirmed to be repeatedly audited and then save it to the knowledge base.

7. A repeat offender verification system based on big data auditing and artificial intelligence as described in claim 6, characterized in that, The judgment criteria in the semantic and reasoning decision-making submodule include: the audit objects of the current problem and the historical problem must be the same unit or the same responsible entity; the problem characteristics of the current problem and the historical problem must belong to the same or highly related categories; the current problem must have occurred after the historical problem was pointed out, and the rectification status of the historical problem must be unrectified or inadequate; when the confidence level output by the large language model is lower than the preset confidence level threshold, the system marks the judgment result as pending manual review.

8. The repeat offender verification system based on big data auditing and artificial intelligence according to claim 1, characterized in that, Also includes: The threshold configuration module is used to customize similarity thresholds, which include high similarity thresholds, medium similarity thresholds, and low similarity thresholds. The threshold types include comprehensive similarity thresholds, qualitative similarity thresholds for questions, and similarity thresholds based on qualitative criteria. The association level selection module is used to select the classification level of audit issues for association comparison to adjust the comparison scope. It supports comparison by primary or secondary classification. When the issue list comes from multiple subordinate units under the superior unit, it automatically aggregates the data of all subordinate units for full verification. The results display and functional interaction module is used to output the comparison results in a standardized table format. The table includes issue details, similarity scores for each dimension, judgment results and judgment basis information, and supports result download. The comparison record viewing module is used to store historical comparison task records, and supports retrieval, viewing, and review. The knowledge base management module is used to manage the audit data in the knowledge base. It uses an approximate nearest neighbor index structure for vector indexing and writes newly confirmed repeat offender data into the knowledge base to achieve incremental updates after each comparison task is completed.

9. A method for verifying repeated offenses based on big data auditing and artificial intelligence, applied to the system as described in any one of claims 1 to 8, characterized in that, Includes the following steps: Step S1: Obtain the historical audit project issue list data and the current audit project issue list data for the same audit entity; Step S2: Input the historical audit project data and the current audit project data by uploading a template or selecting a project; Step S3: Perform format validation and field mapping on the input data, and output a standardized structured data table; Step S4: Set the similarity threshold and select the comparison classification level and range; Step S5: The structured data is cleaned, segmented and vectorized. The cosine similarity of the three dimensions of audit problem characterization, audit problem description and characterization legal basis is calculated respectively. The comprehensive similarity is calculated by weighting according to the preset weight ratio and compared with the threshold to output potential clues. Step S6: Search the knowledge base for the potential clues to obtain matching context content, construct structured prompt words with the current audit issue information, input them into the large language model for reasoning and judgment, and output the final judgment result of whether it constitutes a repeated offense. Step S7: The final judgment results are presented in a standardized table format and can be downloaded. At the same time, the data confirmed as repeated offenders are stored in the knowledge base. Step S8: Archive the comparison task to support subsequent retrieval and review.

10. A method for verifying repeated offenses based on big data auditing and artificial intelligence according to claim 9, characterized in that, In step S6, the knowledge base retrieval includes: converting the problem description text of potential clues into a query vector, performing an approximate nearest neighbor retrieval in the vector index of the knowledge base, recalling multiple knowledge base records with the smallest cosine distance, filtering irrelevant records, and concatenating them into context text. The reasoning and judgment process includes: constructing a structured prompt word template containing role settings, judgment criteria, context injection, data to be judged, and output format constraints, and feeding it into a large language model to obtain a structured output containing the judgment result, confidence level, and judgment reason; the judgment criteria include judgment conditions for the identity of the audit object, the qualitative relevance of the problem, and the rectification status of historical problems.

Images