Knowledge base intelligent deep auditing method and system based on ai large model

Abstract

The application provides an AI large model-based knowledge base intelligent deep auditing method and system, the method comprising: obtaining an unstructured file uploaded by a user, parsing the unstructured document, extracting a semantic auditing unit, and constructing a semantic auditing unit set; constructing a graph structure based on context dependence according to the semantic auditing unit set, labeling continuous, neighbor, reference, and reverse logic edge types, and updating node representation by fusing compliance attention mechanism to obtain a node semantic representation set; introducing a regulation knowledge embedding library, combining the graph structure and the node semantic representation set, and designing a double-layer risk discrimination function to identify candidate problem items; and performing causal chain backtracking on the candidate problem items to generate a structured auditing conclusion. The application significantly improves auditing accuracy and efficiency and reduces the burden of manual review.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of AI large models, and particularly relates to a knowledge base intelligent deep auditing method and system based on an AI large model. BACKGROUND

[0002] In high-complexity business scenarios such as bidding, budget review and government contracts, the knowledge base within an organization plays a core role in risk control and decision support. With the expansion of project scale and the frequent updating of regulations, the to-be-audited materials extend from structured fields to page-numbered fine PDF contract texts with frequent cross-chapter references, accompanied by scanned copies, charts and multiple versions of regulations, forming a highly heterogeneous information ecosystem. Manual auditing relies on personal experience and item-by-item comparison, which is not only time-consuming but also difficult to ensure consistency, especially when facing cross-chapter logical chains, reference chains and reverse condition expressions, it is easy to miss key risks; and traditional automated tools mostly stay at the level of keyword comparison or fixed template filtering, and cannot understand the causal structure between paragraphs and the applicable boundaries of regulations, nor can they provide reliable judgments in the case of multiple versions of regulations and non-uniform expression of clauses. As a result, the auditing period is long, there are many misjudgments and omissions, and the cost of manual review is high, which seriously restricts the efficiency of bidding and budget circulation, and also makes the organization face uncontrollable potential risks in terms of regulation compliance. SUMMARY

[0003] The purpose of the present application is to provide a knowledge base intelligent deep auditing method and system based on an AI large model to solve the above problems.

[0004] To achieve the above purpose, in the first aspect of the present application, a knowledge base intelligent deep auditing method based on an AI large model is provided, which comprises the following steps:

[0005] S1, obtaining a user-uploaded unstructured file, parsing the unstructured document and extracting semantic auditing units, and constructing a set of semantic auditing units;

[0006] S2, constructing a graph structure based on context dependence according to the set of semantic auditing units, labeling continuous, neighboring, reference and reverse logical edge types, and updating node representation by fusing compliance attention mechanism to obtain a set of node semantic representations;

[0007] S3, introducing a regulation knowledge embedding library, combining the graph structure and the set of node semantic representations to design a double-layer risk discrimination function to identify candidate problem items;

[0008] S4, performing causal chain backtracking on the candidate problem items to generate a structured auditing conclusion.

[0009] Further, the parsing of the unstructured document comprises:

[0010] The structured preprocessing is performed on each unstructured document by the integrated analysis tool PDFPlumber, including the following key content collection: text content collection, numbering identification, title format extraction, and structure position labeling.

[0011] Further, the unstructured document is parsed, and the following is performed:

[0012] The set of paragraphs formed by traversing the unstructured document is audited by a template matching function and a set of preset semantic templates to identify audit units, generating a corresponding set of semantic audit units; wherein the template matching function is used to perform keyword and regular rule identification on each paragraph to generate a hit result;

[0013] The semantic vector representation of the set of semantic audit units is calculated by a light vector encoder to generate the semantic vector representation of the corresponding semantic audit unit, representing the overall semantic expression of its content.

[0014] Further, the light vector encoder is a BERT-base model with the first four layers of frozen parameters, and only outputs the [CLS] vector as the representation.

[0015] Further, the S2 specifically includes:

[0016] The fields of the semantic audit unit include: original text content, position of the unstructured document, semantic classification label, and semantic vector representation of the corresponding semantic audit unit;

[0017] The semantic audit unit is mapped to a set of nodes, and the numbering continuous edge, semantic neighbor edge, text reference edge, and reverse logic edge are mapped to a set of edges to construct a graph structure.

[0018] The adjacent nodes are weighted, and the semantic vector representation of the corresponding semantic audit unit is combined to generate the context embedding representation of the aggregated corresponding semantic audit unit by adjacency aggregation, and is constructed into a set of node semantic representations.

[0019] Further, the numbering continuous edge represents that if the numbering difference of any two semantic audit units is and located in the same chapter, then is added.

[0020] The semantic neighbor edge represents that if the cosine similarity of any two semantic audit units is greater than 0.85, it is considered as content expression repetition or expansion.

[0021] The text reference edge represents that if the position of the unstructured document appears a reference phrase, it is hit and located to the corresponding semantic audit unit.

[0022] The reverse logic edge indicates that if the current semantic audit unit contains a reverse logic phrase and the content is opposite to the expression direction of another semantic audit unit, a reverse marking edge is established.

[0023] Further, the double-layer risk discrimination function is specifically:

[0024] The first layer is a regulation conflict possibility score function, which aims to minimize the conflict possibility score, and combines the context embedding representation and the embedding vector of the article in the regulation knowledge base to generate the minimum conflict possibility score between the current semantic audit unit and the regulation library.

[0025] The second layer is a structure inconsistency logic residual function, which is used to identify problems caused by reference logic errors or numerical inconsistencies. By combining the node set directly referenced by the current semantic audit unit in the graph structure, the edge weight, and the context embedding representation, a structure consistency residual value is generated.

[0026] Further, the combination of the graph structure and the node semantic representation set is designed to identify candidate problem items through a double-layer risk discrimination function, specifically:

[0027] If the minimum conflict possibility score is greater than a first preset threshold or the structure consistency residual value is greater than a second preset threshold, the system records the current semantic audit unit as a problem candidate item and marks it as a regulation conflict candidate item or a structure inconsistency candidate item, and outputs it to the problem candidate set. Each problem candidate item contains the original text, the page number, the structure number, the problem category, the reference path node set, and the article matching number.

[0028] Further, the S4 specifically includes:

[0029] For each problem candidate item, path backtracking is performed in the graph structure, with a maximum backtracking depth of , and all text reference edges and reverse logic edges are retained to form a causal support path for building a support chain.

[0030] Based on the causal support path, all nodes in the path are extracted to generate a structured conclusion object.

[0031] For each problem candidate item, a structured conclusion object is generated, combined with adjacent semantic audit units, to generate a structured display weight vector for highlighting the most influential nodes in the path.

[0032] The final output is a structured audit report conclusion set.

[0033] The structured conclusion object includes an audit conclusion label, original paragraph content, a page number, a node number involved in a cause-effect chain, a directed edge sequence in a cause-effect support path, a text segment from a node in the cause-effect support path, and a system recommended operation label.

[0034] In a second aspect of the application, an AI large model-based knowledge base intelligent deep audit system is provided, and the system comprises:

[0035] A file acquisition unit is configured to acquire a user-uploaded unstructured file, parse the unstructured document, and extract a semantic audit unit, and construct a semantic audit unit set;

[0036] A recognition unit is configured to construct a graph structure based on context dependence according to the semantic audit unit set, label continuous, near-neighbor, reference, and reverse logic edge types, and update node representation by fusing a compliance attention mechanism to obtain a node semantic representation set;

[0037] A file analysis unit is configured to introduce a regulation knowledge embedding library, combine the graph structure and the node semantic representation set, and design a double-layer risk discrimination function to identify candidate problem items;

[0038] A problem tracing and report generation unit is configured to perform cause-effect chain backtracking on the candidate problem items to generate a structured audit conclusion.

[0039] The application has at least the following beneficial technical effects:

[0040] The application proposes an end-to-end deep audit method to solve the above problems, which first divides an unstructured PDF contract into SRU blocks with independent audit significance by using a structure rule-driven semantic audit unit division technology, then explicitly labels the numbering continuity, semantic near-neighbor, reference, and reverse logic as multi-type edges by graph structure modeling to generate a computable context-dependent network; on this basis, the node representation is updated by using a compliance attention aggregation and a content avoidance factor, and the regulation matching deviation and the structural consistency residual are simultaneously measured by a double-layer risk discrimination function to accurately screen out potential high-risk candidate items; finally, the trigger path, support nodes, and articles of law of each problem clause are aggregated into a structured audit conclusion by a cause-effect chain backtracking algorithm, the key causes are highlighted by an interpretable weight vector, and a closed loop from problem discovery, cause tracing to evidence-based audit report generation is realized. This method can automatically complete deep semantic analysis and regulation adaptation in semi-structured, cross-chapter, and highly-referenced project documents without relying on large-scale labeling or black box models, significantly improves the audit accuracy and efficiency, reduces the artificial review burden, and provides the government and enterprises with reproducible, traceable, and compliant intelligent audit capabilities that meet the latest regulation requirements. BRIEF DESCRIPTION OF DRAWINGS

[0041] The application is further illustrated by the accompanying drawings, but the embodiments in the drawings do not constitute any limitation to the application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of the following drawings.

[0042] Figure 1 The flow chart of the AI large model-based knowledge base intelligent deep review method of the application. DETAILED DESCRIPTION

[0043] Embodiments of the application are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by reference to the drawings are exemplary and are only used to explain the application and cannot be understood as a limitation to the application.

[0044] As Figure 1 shown, the AI large model-based knowledge base intelligent deep review method provided by the embodiments of the application comprises:

[0045] S1, obtaining an unstructured file uploaded by a user, parsing the unstructured document and extracting a semantic review unit, and constructing a semantic review unit set.

[0046] Specifically, this step is aimed at "structured and formatted" project PDF files such as bidding procurement, project budget, government contract, etc. The goal is to extract semantic review units (SRU) with independent review value from them to support subsequent dependency modeling and review decisions. Although such documents have neat layouts, the content expression has instability, for example, the same "payment conditions" may appear in different chapters, and the wording may change due to different writing habits. This determines that content splitting cannot simply rely on paragraph segmentation or title extraction, and a balance between structure boundary and semantic recognition needs to be achieved.

[0047] The input of this step is a set of PDF files uploaded by users in the system . The system will perform structured preprocessing on each PDF file through the integrated parsing tool PDFPlumber, including the following key content collection: text content collection (extracting by page and paragraph); number recognition (recognizing structure numbers such as "3.2" and "4.1.5"); title format extraction (recognizing font, font size, bold state, etc.); structure position labeling (labeling the page number and line number of the paragraph).

[0048] For example, there is the following content in page 3 of a certain file:

[0049] 3.2 Bidder qualification requirements

[0050] The bidder shall be a registered independent legal person in the People's Republic of China, and there shall be no major illegal records in the past three years.

[0051] The system identifies the number "3.2" as a sub-chapter title and concatenates the paragraphs under it into a structured paragraph object . Then the system traverses the paragraph set formed by the document , and uses the template matching function and the preset semantic template set to identify the audit unit:

[0052] ;

[0053] Among them, : the set of semantic audit units extracted from the th PDF document; : the paragraph set after PDF structure analysis, each paragraph containing number, font, content and location; : semantic template set, including keyword group, regular expression, label name, matching threshold; : structure matching function, which performs keyword and regular rule identification on each paragraph to generate hit results.

[0054] For example, a paragraph is:

[0055] Payment terms: 30% of the total contract amount shall be paid within 10 working days after signing the contract, and the remaining 70% shall be paid after the project is completed and accepted.

[0056] The paragraph hits the keyword and sentence pattern of the "payment clause template" and is assigned the label "payment clause" to form the SRU unit , and records its original position and hit template.

[0057] In order to ensure the semantic integrity of the extracted content, the system further calculates the vector representation of each and evaluates the coherence between sentences to eliminate audit units with mixed structures or fragmented content. The embedding process is as follows:

[0058] ;

[0059] Among them, : the semantic vector of the audit unit , representing the overall semantic expression of its content; : light vector encoder, based on BERT-base frozen parameter model, only output [CLS] vector as representation; : the audit unit generated by hitting the structure template, whose text content is input into the encoder.

[0060] The system calculates the average cosine similarity of the vector and the internal sentences, and if it is lower than 0.6, it is considered that the semantic of the paragraph is not coherent, for example, there are mixed sentences such as “see the figure” and “please refer to the appendix”, which are removed or split.

[0061] Finally, the system outputs a set of review units with unified structure, clear semantics and consistent granularity , which serves as the basis input for constructing the logical dependency structure graph in the subsequent step.

[0062] S2, according to the semantic review unit set, a graph structure based on context dependency is constructed, the continuous, neighbor, reference and reverse logic edge types are labeled, and the node representation is updated by fusing compliance attention mechanism to obtain a set of node semantic representations.

[0063] Specifically, after the semantic review unit (SRU) extraction is completed, the context dependency relationship between each review unit is established, the document structure graph is constructed, and the context enhancement modeling is performed on each node to provide a semantic traceable structure basis for subsequent risk identification and compliance review. In the “tender procurement, project budget, government contract” type documents faced by the present application, although the numbering structure is clear, there are a large number of nonlinear relationships between clauses, such as cross-section reference, sequential condition dependency, and logical reverse expression. The traditional method of establishing logical connection based on numbering or adjacent paragraphs cannot meet the needs of “semantic consistency review” and “article application chain tracking”. Therefore, based on the structure graph construction, the present application innovatively introduces “compliance attention aggregation item” and “context avoidance factor”.

[0064] The input is the semantic review unit set output by the previous step , wherein each contains the following fields: original text content , position in PDF , semantic classification label , and semantic vector representation generated by the embedding model in step one . This set is the basis data for structure graph construction and context modeling in this step.

[0065] Further, in order to capture the implicit reference and structural relationship between the semantic review units, the present application takes as the node to construct the semantic structure graph . The types of edges in the graph not only include the common numbering continuous edge and semantic neighbor edge , but also innovatively introduce “reference trigger edge” and “chapter spanning reverse edge” , the latter is dedicated to mark the logical inversion relationship appearing in the "contract condition inverse logic expression" (e.g. "not pay if not complete acceptance") in the document. The edge set is defined as follows:

[0066] ;

[0067] wherein, : the constructed structure graph, consisting of node set and edge set ; : composed of all SRU units, i.e. ; : numbered structure edge, if and are numbered with a difference of and are in the same chapter, then add; : semantic similarity edge, if the cosine similarity , it is considered as content expression duplication or expansion; : text reference edge, if contains phrases such as "as mentioned above" "see the above article" etc., regular hit and locate to ; : inverse logic edge, if contains inverse logic phrases such as "if not... then..." sentence, and the content is opposite to another SRU expression direction, then an inverse marking edge is established. is a mapping function.

[0068] Further, after the structure graph is constructed, in order to capture the context dependence of each audit unit in the semantic structure graph, the system introduces a structure-aware node representation updating strategy. This strategy is based on the modification of the adjacency aggregation method of classical graph neural network, and innovatively adds "compliance attention term" and "content avoidance term" to improve the influence weight of important review nodes and weaken the influence of noise nodes irrelevant to the law. The specific formula is as follows:

[0069] ;

[0070] wherein, : the context embedding representation of the aggregated semantic audit unit ; : the base semantic vector (dimension 768) calculated in step one for node ; : the set of adjacent nodes connected to in graph ; : the structure attention weight from to , valued according to semantic relevance + edge type, for example: Edge weight is higher than ; : content evasion factor, if there is a non-examination content prompt (such as "see the figure" "as listed in the annex"), set , otherwise ; : for the normalization term, .

[0071] It can be understood that the core innovation of the formula is:

[0072] Instead of using fixed average, the adjacent nodes are given weight, and structure perception is introduced;

[0073] The SRU with "non-examination expression" is punished and the weight is reduced, and the structural noise is effectively excluded;

[0074] The weight calculation is combined with semantics and edge type, and the structural influence of "reference type reference" and "logical reversal" in compliance review is strengthened.

[0075] For example, in the "project acceptance" chapter, indicates "no payment without acceptance", which points to (payment condition definition), (acceptance process), then the system establishes a directional edge through + , and gives higher attention value , and if contains a statement like "see annex two", set to , so that aggregation is biased towards payment content rather than acceptance details.

[0076] Compared with the traditional "vector average" or "paragraph alignment" method, this step has obvious improvement in structural expression and audit logic support. It does not depend on supervised labeling and learnable parameters, but only through structural rules and attention control to express the potential dependency between clauses in complex regulatory documents.

[0077] S3, introduce a regulatory knowledge embedding library, combine the graph structure and node semantic representation set to design a double-layer risk discrimination function to identify candidate problem items.

[0078] Specifically, after the semantic review unit (SRU) has completed structure mapping and context representation aggregation, the goal is to identify content fragments that may have potential risks such as regulatory conflicts, logical contradictions, and structural inconsistencies, and mark them as candidate problem items for subsequent conclusion confirmation. Unlike traditional methods based on keywords or rule triggers, this step combines structural graph , node context representation and regulation knowledge embedding library , by constructing a set of computing functions with invention characteristics, systematically screening content units with "compliance doubts". This mechanism is particularly suitable for the application scenario of the invention: structured paragraphs such as project contracts, budget documents, procurement instructions, and complex text systems coexisting with cross-references and compliance requirements.

[0079] The input is the output of step two, including the structure diagram , the structure-aware representation of each semantic review unit , and the regulation knowledge base embedding representation . Among them The clauses from the actual regulation text (such as the "Government Procurement Law" and the "Bidding and Tenders Law") are generated using the same BERT-4 layer frozen model as the SRU. The embedding has been cleaned before the format and digital unit.

[0080] Among them, in practical applications, the invention finds that some problems frequently occur in project documents, but are very difficult to identify through traditional techniques, such as:

[0081] Surface compliance but semantic reversal ("no payment without acceptance" vs "payment without acceptance criteria");

[0082] Amount logic inconsistency ("pay 30%, 30%, 40%" vs two periods of detail items);

[0083] Clause missing and content reference ("see performance requirements" but actually no such clause);

[0084] Reference to outdated regulations or irrelevant legal texts.

[0085] To solve such problems, the invention introduces a candidate identification function with a double-layer discrimination logic, the first layer detects the possibility of regulation conflict, and the second layer assesses the inconsistency of structural reasoning, and finally fuses the problem risk score.

[0086] The first layer is the regulation conflict possibility scoring function, which combines the semantic bias main item + clause offset adjustment item + reference jump structure penalty item, and is constructed as follows:

[0087] ;

[0088] Among them, : represents the minimum conflict possibility score between and the regulation library, the larger the more suspicious; : the context-aware semantic representation of generated in step two, with a dimension of 768; ​: Embedding vector of the clause in the regulation knowledge base; Semantic similarity of the clause; : Regulation clause offset adjustment term, indicating whether the clause is a "weakly adapted clause", such as an old law or a non-general law, then , marked by manual review or historical adaptation frequency estimation; : Structural jump adjustment factor, if cross-chapter reference more than 2 layers, or more than 2 references in the reference edge, (inverse edge), then set to 1.5, otherwise to 1; Maximum structural number span of the referenced clause (for example, when referencing "4.3" and the current number is "1.2", the span is 3.1), indicating the logical span; Number of adjacent nodes of , used to normalize the reference density.

[0089] It can be understood that the innovation of the formula design is that not only the minimum distance matching of the semantic vector is considered, but also

[0090] Clause weight adjustment ( ), giving priority to new laws;

[0091] Reference path cost factor ( ), to suppress long-distance low-confidence references;

[0092] Number span explicit penalty ( ), to suppress the "pseudo-legality" of cross-chapter references.

[0093] When (system threshold, usually 0.35), the system determines that there is a risk of regulation adaptation conflict.

[0094] The second layer is the structural inconsistency logical residual function, which is used to identify problems caused by logical errors or numerical inconsistencies in references, such as inconsistencies between summary numbers and sub-items, references to undefined content, etc. This function is based on the "structural consistency residual" between the structure path and the embedding, defined as follows:

[0095] ;

[0096] Where, : Structural consistency residual value, used to judge the degree of inconsistency between the semantics of and its referenced content; : Set of nodes directly referenced in the structure diagram; : Node ​​contextual semantic representation of the clause; : edge weight, derived from the structural attention coefficient in step two ; : represents the squared Euclidean distance between vectors.

[0097] Intuitively, if a clause references multiple sub-clauses, but the content vector of the clause is far from the aggregated representation of the sub-clauses, there is a semantic aggregation conflict in the clause, which may be a "description contradiction", "condition omission" or "inappropriate merger" problem.

[0098] The final candidate identification logic is: if or , the system records the as , and marks it as "regulation conflict candidate" or "structural inconsistency candidate", and outputs it to the problem candidate set , each candidate contains the original text, page number, structure number, problem category, reference path node set, regulation matching number, etc.

[0099] This step combines structural graph aggregation, regulation embedding comparison, and reference path analysis for the first time, identifies "potential logical problem points" that cannot be handled by traditional methods, not only improves the discovery rate, but also generates candidate problems with "structural causal chain", which can provide traceability basis for subsequent decision modules.

[0100] S4, perform causal chain backtracking on the candidate problem item, and generate structured review conclusion.

[0101] Specifically, this step has completed the identification of candidate problem items in the previous stage, that is, to generate a candidate problem set , each contains a source semantic review unit , trigger reason (such as , ), location in the structural graph , and its contextual embedding . The goal of this step is to backtrack the structural causal chain based on the identified problem items without introducing any new model or structural judgment, generate clear conclusion labels, organize structured display fields, and achieve visual, readable, and traceable review results output.

[0102] The project document handled by the present application generally has the characteristics of stable paragraph number, dense structure reference, and frequent semantic reference. Therefore, the conclusion generation cannot be simply "red" or "tagged", but must include: question paragraph body, question type, corresponding risk indicator, causal reference path chain, support content node set, reference article fragment (if any), and finally "conclusion suggestion label" (for example: pass, risk item, missing item, manual review, etc.).

[0103] Firstly, the present application performs path backtracking on each of the candidate questions In the structure diagram , the maximum backtracking depth is limited to , and all and type edges (i.e. reference edges and reverse logical edges) are retained to form causal support paths for building support chains. The path acquisition function is as follows:

[0104] ;

[0105] Wherein, : the causal support path set of the candidate question ; : the semantic audit unit that triggers the candidate item; : reference edge, representing intra-paragraph "as mentioned above" "see the previous article" and other reference relationships; : reverse expression edge, used to connect reverse type expressions such as "if not accepted, no payment"; the path depth does not exceed 3 layers to avoid information drift.

[0106] The system uses the path set to extract all nodes in the path for display of support content, reference structure and context explanation. Then a structured conclusion object is generated, which includes the following fields:

[0107] : audit conclusion label, from question type label;

[0108] : original paragraph content, ;

[0109] : page number in PDF, from ;

[0110] : node numbers involved in the causal chain, such as ["3.2", "3.4"];

[0111] : Path structure of directed edges in the path;

[0112] : Text snippet from node in path ;

[0113] : System recommended operation label, e.g.: manual review required / suggested modification / automatically repairable.

[0114] To improve the interpretability of the conclusion, the application generates a structured display weight vector for each conclusion, which is used to highlight which path nodes have the greatest impact in the display. The formula is as follows:

[0115] ;

[0116] wherein, : Candidate supporting node in the conclusion display weight; : Structural edge weight, derived from the edge weight matrix in the structure graph ; : Risk score of node from step three or , normalized; the weight vector is used to display the contribution degree and highlight the key factors in the front-end report.

[0117] The final output is a structured audit report conclusion set , which can be displayed on the page through the front-end system, including main problem content, supporting quote snippet, quote path graph, and suggested operation item, etc.

[0118] For example, a conclusion item structure is as follows:

[0119] Main clause: Payment method is 30%+30%, and the remaining 40% is paid after the project is completed;

[0120] Problem type: Structural inconsistency;

[0121] Quote chain: 3.2→3.3 (project acceptance), 3.5 (contract termination clause);

[0122] Supporting node content: Page 3 "acceptance criteria have not been defined", page 5 "payment is based on completion of acceptance";

[0123] Suggestion: Manual review, there may be a problem that the payment clause is not closed.

[0124] Through the mechanism of backtracking and structured presentation of the cause-effect path, the system not only realizes problem positioning, but also can "explain how the problem comes", and breaks through the complete chain of "triggering-> attribution-> presentation".

[0125] The embodiment of the application also provides an AI large model-based knowledge base intelligent deep auditing system, which comprises:

[0126] A file acquisition unit is configured to acquire an unstructured file uploaded by a user, parse the unstructured document, extract a semantic auditing unit, and construct a semantic auditing unit set.

[0127] A recognition unit is configured to construct a graph structure based on context dependence according to the semantic auditing unit set, label continuous, neighbor, reference and reverse logic edge types, and update node representation by fusing a compliance attention mechanism to obtain a node semantic representation set.

[0128] A file analysis unit is configured to introduce a regulation knowledge embedding library, combine the graph structure and the node semantic representation set, and design a double-layer risk discrimination function to identify a candidate problem item.

[0129] A problem backtracking and report generation unit is configured to perform cause-effect chain backtracking on the candidate problem item to generate a structured auditing conclusion.

[0130] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system, device and unit described above can refer to the corresponding process in the foregoing method embodiments, which will not be repeated here.

[0131] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of the units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interface, device or unit, and can be electrical, mechanical or other forms.

[0132] If the functions are implemented in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the parts that contribute to the prior art or parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0133] Although the embodiments of the present application have been shown and described, those skilled in the art can understand that various changes, modifications, replacements and deformations can be made to the embodiments without departing from the principles and purposes of the present application, and the scope of the present application is defined by the claims and their equivalents.

Claims

1. A knowledge base intelligent deep review method based on AI large model, characterized in that, The method includes the following steps: S1. Obtain unstructured files uploaded by users, parse the unstructured documents and extract semantic review units, and construct a set of semantic review units; S2. Construct a context-dependent graph structure based on the set of semantic review units, label the types of continuous, nearest, reference and reverse logic edges, and update the node representation by integrating the compliance attention mechanism to obtain the set of node semantic representations; S3. Introduce a regulatory knowledge embedding library, and design a two-layer risk discrimination function to identify candidate problem items by combining the graph structure and node semantic representation set; S4. Perform causal chain backtracking on the candidate issue items to generate a structured audit conclusion; The two-layer risk discrimination function is specifically as follows: The first layer is a regulatory conflict probability scoring function. With the goal of minimizing the probability of conflict, it combines the context embedding representation and the embedding vector of legal provisions in the regulatory knowledge base to generate the minimum probability of conflict between the current semantic review unit and the regulatory base. The second layer is a structural inconsistency logic residual function, which is used to identify problems caused by incorrect reference logic or numerical discrepancies. It generates structural consistency residual values ​​by combining the set of nodes directly referenced by the current semantic review unit in the graph structure, edge weights, and context embedding representation. The design, which combines the graph structure and node semantic representation set, identifies candidate problem items through a two-layer risk discriminant function, specifically as follows: If the minimum conflict probability score is greater than the first preset threshold or the structural consistency residual value is greater than the second preset threshold, the system records the current semantic review unit as a candidate issue and marks it as a candidate for legal conflict or a candidate for structural inconsistency, and outputs it to the candidate issue set. Each candidate issue includes the original text, the page number, the structure number, the issue category, the set of reference path nodes, and the legal provision matching number. S4 specifically includes: For each problem candidate, perform path backtracking in the graph structure, limiting the maximum backtracking depth to [value missing]. And retain all text reference edges and reverse logic edges to form a causal support path for building a support chain; Based on the causal support path, nodes in all paths are extracted to generate a structured conclusion object; For each question candidate's structured conclusion object, combined with the adjacency semantic review unit, a structured display weight vector is generated to display the node with the greatest influence in the highlighted path. The final output is a set of structured audit report conclusions; The structured conclusion object includes: review conclusion tags, original paragraph content, page numbers, node numbers involved in the causal chain, directed edge sequences in the causal support path, text fragments from nodes in the causal support path, and system recommended operation tags.

2. The intelligent deep review method for knowledge base based on AI large model according to claim 1, characterized in that, The parsing of unstructured documents includes: Each unstructured file undergoes structured preprocessing using the integrated parsing tool PDFPlumber, including the following key content captures: text content capture, number recognition, title format extraction, and structural location annotation.

3. The intelligent deep review method for knowledge base based on AI large model according to claim 2, characterized in that, The parsing of unstructured documents involves the following steps: The paragraph set formed by the unstructured document is traversed, and the review unit is identified by using the template matching function and the preset semantic template set to generate the corresponding semantic review unit set; The template matching function is used to perform keyword and regular expression rule recognition on each paragraph and generate a matching result; The semantic vector representation of the set of semantic review units is calculated by a lightweight vector encoder, and the semantic vector representation of the corresponding semantic review unit is generated to represent the overall semantic expression of its content.

4. The intelligent deep review method for knowledge base based on AI large model according to claim 3, characterized in that, The lightweight vector encoder is based on the BERT-base first 4-layer frozen parameter model and only outputs the [CLS] vector as a representation.

5. The intelligent deep review method for knowledge base based on AI large model according to claim 3, characterized in that, S2 specifically includes: Obtain the fields of the semantic review unit, including: the original text content, the location of the unstructured document, the semantic classification tag, and the semantic vector representation of the corresponding semantic review unit; The semantic review unit is mapped to a node set, and the consecutive numbered edges, semantic nearest neighbor edges, text reference edges, and reverse logic edges are used as edge sets to construct a graph structure. By assigning weights to adjacent nodes and combining them with the semantic vector representation of the corresponding semantic review unit, the context embedding representation of the corresponding semantic review unit is generated through adjacency aggregation, and a set of node semantic representations is constructed.

6. The intelligent deep review method for knowledge base based on AI large model according to claim 5, characterized in that, The term "sequential edge" indicates that if the number difference between any two semantic review units is 1 and they are located in the same chapter, then an edge is added. The semantic nearest neighbor edge representation is as follows: if the cosine similarity between any two semantic review units is greater than 0.85, it is considered as content repetition or expansion. The text reference edge indicates that if a referential phrase appears in the location of the unstructured document, the corresponding semantic review unit is hit and located. The reverse logic edge means that if the current semantic review unit contains a reverse logic phrase, and the content is opposite to the expression direction of another semantic review unit, then a reverse annotation edge is established.

7. A system for implementing the AI-based large-scale model-based intelligent deep auditing method for knowledge bases as described in claim 1, characterized in that, The system includes: The file acquisition unit is used to acquire unstructured files uploaded by users, parse unstructured documents and extract semantic review units, and construct a set of semantic review units; The identification unit is used to construct a context-dependent graph structure based on the set of semantic review units, label the types of continuous, nearest, reference and reverse logic edges, and update the node representation by integrating the compliance attention mechanism to obtain a set of node semantic representations; The document analysis unit is used to introduce a regulatory knowledge embedding library and, in combination with the graph structure and node semantic representation set, design a two-layer risk discrimination function to identify candidate problem items. The issue tracing and report generation unit is used to perform causal chain backtracking on the candidate issue items and generate structured audit conclusions.

Images