A knowledge graph-based intelligent auditing method

Abstract

This application discloses an intelligent auditing method based on knowledge graph constraints, relating to the field of natural language processing technology. The method includes: constructing an audit rule knowledge graph; training an initial large language model under constraints to obtain an audit fact generation model; receiving a work document to be audited and inputting it into the pre-trained audit fact generation model to generate candidate fact triple paths; calling a pre-constructed audit hook knowledge graph to calculate hook matching degrees; determining target audit events and target audit hooks; based on associated target questioning keywords, performing deep information verification through the audit fact generation model to generate confirmatory fact triples; and determining risk types and risk levels through the audit rule knowledge graph to generate an audit conclusion report. By specifically training the large language model using the audit rule knowledge graph, the model's internal logic and output results are guided by structured domain rules when generating audit facts.

Description

Technical Field

[0001] This application relates to the field of natural language processing technology, specifically to an intelligent auditing method based on knowledge graph constraints. Background Technology

[0002] In areas such as compliance and oversight of large enterprises, a large number of work reports and project minutes are generated that await auditing. Automating the auditing of these documents through artificial intelligence technologies, such as audit fact generation models trained on large language models, has become a key technological trend for improving audit efficiency and reducing compliance risks.

[0003] However, current automated auditing solutions based on large language models typically use general large language models directly to extract audit facts. But general large language models lack a deep understanding of the professional knowledge and rule system in the auditing field, making it difficult to align with legal provisions and internal regulations. This makes it easy for the model to generate biased or incomplete factual descriptions when dealing with complex and professional audit texts, resulting in insufficient accuracy and completeness in fact extraction. Summary of the Invention

[0004] To address the aforementioned issues, this application proposes an intelligent auditing method based on knowledge graph constraints, comprising:

[0005] An audit rule knowledge graph is constructed, and an initial large language model is trained under constraints based on the audit rule knowledge graph to obtain an audit fact generation model.

[0006] The constrained training is achieved by minimizing the fusion loss function, which is defined as follows: ,in, This is the initial large language model. These are all the trainable model parameters of the initial large language model. For pre-built training samples, for from The set of fact triples extracted from it. for The set of rule triples mapped onto the audit rule knowledge graph. for and The divergence value between them For a pre-labeled set of standard rule triples, For calculation based on the maximum conjugate string matching algorithm and Matching reward value, For hyperparameters;

[0007] Receive the work document to be audited, input it into the audit fact generation model, generate candidate fact triple paths corresponding to each audit event in the work document to be audited, and obtain a set of candidate fact triples;

[0008] The pre-built audit hook knowledge graph is invoked, and the candidate fact triple path is matched with each audit hook in the audit hook knowledge graph. The hook matching degree corresponding to each audit event is calculated. Based on the hook matching degree and the triggering condition of the corresponding audit hook, the target audit event and the target audit hook that trigger the audit are determined.

[0009] Based on the target questioning keywords associated with the target audit hook, the audit fact generation model performs in-depth information verification on the target audit event within the limited scope of the audited work document, generating a confirmatory fact triplet corresponding to the target audit event.

[0010] Using the audit rule knowledge graph and based on the confirmatory fact triples, the risk type and risk level corresponding to the target audit event are determined to generate an audit conclusion report.

[0011] The intelligent auditing method based on knowledge graph constraints proposed in this application can bring the following beneficial effects:

[0012] By constructing an audit rule knowledge graph and designing a loss function that integrates knowledge constraint loss and supervision matching reward, the large language model is specifically trained. This ensures that the model's internal logic and output results are structured and guided by domain rules when generating audit facts. This training mechanism effectively corrects potential biases or ambiguities that may arise from the model's free generation, thereby improving the standardization of fact triples extracted from complex documents in terms of business meaning, logical consistency, and compliance with actual legal provisions, laying a reliable factual foundation for subsequent risk assessment.

[0013] Furthermore, compared to traditional passive matching rules, the pre-built audit hook knowledge graph can intelligently scan and match the factual chains initially generated by the model, automatically locating high-risk suspicious points and triggering targeted in-depth verification. This proactive questioning and multi-round verification mechanism enables the audit fact generation model to penetrate the surface text, enhancing its ability to detect implicit contradictions, logical loopholes, or deliberately evasive statements in the text, thereby more effectively uncovering systemic and interconnected chains of violations. Attached Figure Description

[0014] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0015] Figure 1 This is a flowchart illustrating an intelligent auditing method based on knowledge graph constraints in an embodiment of this application.

[0016] Figure 2 This is a flowchart illustrating a traditional automated auditing method in the embodiments of this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0019] like Figure 1 As shown in the figure, this application provides an intelligent auditing method based on knowledge graph constraints, including:

[0020] S101. Construct an audit rule knowledge graph, and perform constraint training on the initial large language model based on the audit rule knowledge graph to obtain an audit fact generation model.

[0021] By accessing the rule data source interface related to the target audit scope within the audit business area, audit rule documents are obtained. These documents are then cleaned and preprocessed, including removing irrelevant content, format conversion, and text segmentation. The audit rule documents include standardized rule texts, such as departmental regulations, industry standards, and other normative documents, as well as audit standards, operational guidelines, and typical cases issued by audit institutions.

[0022] The audit rule documents are parsed using a natural language processing engine to extract structured rule information. In practice, the deployed NLP engine parses the collected audit rule documents to identify key entities within the audit rules, including rule subjects and rule objects. The rule subject refers to the main entity bound or targeted by the rule, typically the party issuing the action or the responsible party. The rule object refers to the entity or target to which the rule relationship refers, typically the party receiving the action or a related object. Furthermore, the rule relationship between the rule subject and rule object is identified, i.e., the specific association or constraint between them, such as "basis," "prohibition," "penalty," or "attribution." The system also identifies the restrictions imposed on the rule subject, rule relationship, and rule object, such as monetary thresholds, timeframes, and procedural requirements.

[0023] Furthermore, while extracting structured rule information, the system assigns explicit risk semantic labels and quantitative guidance to this information. Specifically, rule relationships are mapped to a predefined risk type system, transforming abstract rule relationships into risk types. For example, the relationship type "prohibition" is usually directly associated with compliance risk types, "penalty" with legal liability risk types, and "basis" may be associated with financial risk types if it involves funding standards. Rule constraints are standardized and parameterized as the basis for risk level assessment. Rule constraints, along with their corresponding rule relationships and entities, are stored together to form the risk level determination conditions for logical reasoning within the knowledge graph. For example, "amount threshold" is parsed into specific numerical ranges, with different ranges corresponding to different risk severity levels (e.g., low, medium, high); the "time range" requirement is transformed into a judgment logic for whether a period is overdue, with overdue itself constituting a condition for an increase in risk level.

[0024] It should be noted that the judgment logic for risk level determination is as follows: based on the structured rule constraints bound to the associated edges of each rule in the audit rule knowledge graph, the judgment logic is executed in two scenarios. If the rule constraint is a quantitative judgment condition, the specific numerical attributes corresponding to the confirmatory fact triple are automatically extracted, and compared with the preset thresholds and operators within the condition to directly map to the corresponding risk level. If the rule constraint is a qualitative compliance clause, the procedural, time-sensitive, or descriptive textual features of the fact triple in the relevant context of the audited work document are extracted, compared with the clause requirements, and the corresponding risk level is finally determined based on the severity of feature loss or violation.

[0025] Furthermore, using rule subjects and rule objects as rule entity nodes and rule relationships as rule association edges, a preliminary rule knowledge graph is constructed. Attribute information is added to the rule entity nodes and rule association edges to obtain the completed rule knowledge graph. The attribute information includes the rule source, effective time, and scope of application. The completed rule knowledge graph is stored in a graph database, and a corresponding index structure is established to obtain the audit rule knowledge graph.

[0026] It should be noted that the audit rules knowledge graph stores the laws, regulations, policy provisions and business standards in the audit field in a structured form, and each rule node is associated with a clear definition of risk type and risk level.

[0027] Furthermore, the initial large language model is trained under constraints based on the audit rule knowledge graph to obtain an audit fact generation model. Constraint training is achieved by minimizing the fusion loss function, which is defined as: ,in, For the initial large language model, For all trainable model parameters of the initial large language model, For pre-built training samples, for from The set of fact triples extracted from it. for The set of rule triples mapped onto the audit rule knowledge graph. for and The divergence value between them A pre-labeled set of standard rule triples, For calculation based on the maximum conjugate string matching algorithm and Matching reward value, This is a hyperparameter.

[0028] In practice, a large amount of audit-related data is collected, such as audit reports, working papers, and policy documents. This data is then cleaned, and training samples q are constructed based on the cleaned audit-related data. For some training samples, audit experts annotate the set of audit fact triples they contain, which serve as the standard rule triple set. .

[0029] Input the training sample q into the initial large language model In the middle, through the initial large language model Two forward inferences are performed. Specifically, the first forward inference is: the initial large language model... The training sample q is analyzed, and the set of fact triples interpreted is output. It also records the first probability distribution of each token in the output layer for each fact triple during the generation process. The second forward inference involves inputting the training sample q along with the audit rule knowledge graph into the initial large language model. Guide its output to the set of fact triples A set of rule triples that are semantically corresponding but conform to the audit rule knowledge graph specification. It also records the second probability distribution of each token in the output layer for each rule triple during the generation process. .

[0030] In the first forward inference process, the initial large language model identifies sample audit entities in the training samples, including sample responsible parties, sample audit objects, and sample key elements. It also analyzes the semantic relationships between sample audit entities and extracts the entity relationships to describe their behavior and state. During extraction, the model simultaneously analyzes and records the temporal order of each sample audit entity's appearance in the text context, as well as the logical connectors linking different factual statements.

[0031] The sample audit entities and their relationships are automatically combined and encapsulated according to a pre-defined structured triplet format of subject, relationship, and object to construct initial fact triplets for the samples. All initial fact triplets extracted from the training samples constitute the set of fact triplets corresponding to the training samples. When the initial large model generates the text sequence of each sample fact triplet word by word, its output layer calculates a probability value for each candidate word in the vocabulary, completely recording the token probability distribution at all positions during the model's generation of the entire set of fact triplets, forming the first probability distribution.

[0032] It should be noted that the processing of training samples and audit rule knowledge graph by the initial large language model in the second forward inference process is consistent with that in the first forward inference process.

[0033] Calculate the constraint loss of the audit rule knowledge graph on the initial large language model. Specifically, calculate the first probability distribution. With the second probability distribution The Kullback-Leibler (KL) divergence between the fact triples and the rule triples is used to obtain the divergence value of the fact triples set and the rule triples set in the same generated content space. The calculation formula is as follows: Divergence value The larger the value, the greater the loss, indicating that the model deviates more from the rule constraints when generating facts, and stronger penalties need to be applied during training.

[0034] The supervised matching reward of the set of standard rule triples to the initial large language model is calculated. Specifically, the longest common subsequence (LCS) algorithm is used to calculate the reward of the set of rule triples. With the set of standard rule triples Considering two sequences of triplets, calculate the length of the longest common subsequence between them. The length of the standard rule triple set. Length of the longest common subsequence Normalize to obtain the reward value The calculation formula is: Among them, the reward value The larger the value, the smaller the loss, indicating that the model output is closer to the standard answer, and a greater reward should be given during training.

[0035] Constructing the fusion loss function The initial large language model is comprehensively optimized. Specifically, the fusion loss function is defined as: .in, The negative log-likelihood loss of the initial large language model can ensure the basic generative ability of the model; For knowledge enhancement loss terms, It is added as a penalty term to drive the model output to conform to the rule; Subtracting the reward from the output of the incentive model, the model's output is closer to the standard answer. Hyperparameters used to balance the overall loss.

[0036] Minimize using gradient descent algorithms (such as Adam's algorithm) Iteratively update all trainable model parameters of the initial large language model. Once the losses converge, we obtain a trained audit fact generation model that internalizes audit rule knowledge. .

[0037] It should be noted that, as Figure 2 The diagram shows a flowchart of a traditional automated auditing method. First, a large language model is used to parse regulations and generate static audit rules. When rule conflicts occur, a post-hoc selection is made based on historical accuracy and human feedback. Then, a knowledge graph is constructed based on the selected rule chain, and audit conclusions are output. This approach results in static audit rules, making timely updates difficult. Conflict resolution relies on external human intervention, leading to inefficiency and a passive approach. Furthermore, it makes it difficult to perform reinforcement learning and online optimization on the audit fact generation model.

[0038] S102. Receive the work document to be audited, input it into the audit fact generation model, generate candidate fact triple paths corresponding to each audit event in the work document to be audited, and obtain a set of candidate fact triples.

[0039] Specifically, the audit process receives the work document to be audited and inputs it into the audit fact generation model. The model parses the work document to extract audit event information, which includes audit entities and their corresponding relationships, the contextual chronological order of audit entities, and logical connectors between audit entities. In practice, the audit fact generation model performs deep semantic analysis on the work document to be audited, identifying multiple independent audit events and defining boundaries for each event. For each identified audit event, all audit entities in the work document are simultaneously identified and labeled. An audit entity refers to the basic unit constituting a fact, including but not limited to: responsible party, audit object, time point, location, amount, and key actions. The model identifies the relationships between audit entities, such as approval, payment, affiliation, and violation. Simultaneously, it analyzes the context in which audit entities appear, records the contextual chronological order of each audit entity, and extracts logical connectors connecting different facts or entities, such as "therefore," "however," and "and." The audit event information is then composed of the audit entities and their corresponding relationships, the contextual chronological order of audit entities, and the logical connectors between audit entities.

[0040] Furthermore, by treating audit entities as subjects and objects, and entity relationships as relations, initial fact triples are constructed based on a pre-defined structured triple format of subject, relation, and object, resulting in a set of triples corresponding to each audit event. In specific implementation, each initial fact triple follows the standard subject-relationship-object structure. Each audit entity is combined with its corresponding relation to generate initial fact triples related to the audit entity. All initial fact triples belonging to the same audit event constitute the triple set for that audit event, and the triple sets for different events are independent of each other. For example, initial fact triples such as (Department A, Approved, Project B) or (Special Funds, Used for, Equipment Procurement) are generated.

[0041] Based on the contextual chronological order of the audit entities and the logical connectors between them, the temporal and logical relationships between the initial fact triples in the triplet set are determined. In practice, the chronological order of the facts described by each initial fact triplet in the triplet set is determined according to the chronological order of the audit entities within the audit event, establishing a temporal relationship; simultaneously, based on the logical connectors within the audit event, the causal, adversative, parallel, or progressive logical relationships between these facts are determined.

[0042] Based on temporal and logical relationships, the initial fact triples are linked and sorted to generate candidate fact triple paths corresponding to audit events. In practice, for each audit event, multiple related initial fact triples within the audit event are sequentially linked according to temporal and logical relationships as the rules for linking and sorting. For example, triples that occur sequentially and have a causal relationship are linked into a chain. By traversing the set of event triples, the system generates a candidate fact triple path representing the complete audit event. Finally, the system outputs the candidate paths corresponding to all audit events, forming a candidate path set covering the entire document.

[0043] For example, the initial triples (Department A, Approval, Project B), (Project B, Application, Special Funds), and (Special Funds, Misappropriation, Company C) are linked and ordered into a path after their logic and timing are determined based on the context: [(Department A, Approval, Project B) -> (Project B, Application, Special Funds) -> (Special Funds, Misappropriation, Company C)], to clearly illustrate a potential chain of illicit fund flows.

[0044] Before generating corresponding candidate fact triple paths from the auditable work documents, the process also includes constructing an audit hook knowledge graph. It should be noted that audit hooks refer to technical means used during the audit process to discover potential problems. Specifically, they extract key information through data analysis, logical reasoning, and other methods to reveal risks and issues during the audit process.

[0045] Specifically, historical audit case data within the target audit scope of the audit business area is obtained. The historical audit case data is then parsed using a natural language processing engine to extract structured audit information, which includes the type of violation, the responsible parties involved, and the relevant audit standard clauses.

[0046] In practice, the target audit scope is defined based on the specific audit needs corresponding to the audit business area. Historical audit case data within the target audit scope is obtained. This historical audit case data includes the specific audit objects, factual details discovered during the audit process such as violations, involved entities, fund flows, the audit standards and regulations on which they were based, and the final audit conclusions and handling results.

[0047] The deployed natural language processing engine automates the parsing of acquired audit standards documents and historical audit case data. Specifically, the engine identifies key entities in the text, such as responsible parties, audit targets, and behaviors, and normalizes entities that, while differing in description, refer to the same thing. Simultaneously, it identifies relationships between entities, linking the violations described in the cases to applicable clauses in the standards, outputting a structured set of audit information. Each piece of information clearly includes the violation, the responsible party involved, and the applicable clause violated.

[0048] Furthermore, the structured audit information is clustered and summarized using a clustering and inductive analysis engine to obtain a set of violation risk representations. This set of structured audit information is then input into the clustering and inductive analysis engine, which employs an unsupervised learning algorithm to automatically analyze the large amount of scattered audit information.

[0049] Structured audit information is vectorized to obtain a comprehensive feature vector. Specifically, the type of violation in each audit message is converted into a numerical vector through one-hot encoding or embedding technology based on a pre-trained language model. The responsible parties involved are encoded according to their attributes, and the audit criteria clauses are mapped into standardized clause identifier vectors.

[0050] Unsupervised clustering algorithms are used to group similar features in the comprehensive feature vectors into the same audit information cluster. Specifically, a pre-defined unsupervised clustering algorithm, such as the density-based DBSCAN algorithm or hierarchical clustering algorithm, is used to process the comprehensive feature vector set corresponding to all audit information. By calculating the distance metric between vectors, vectors that are close in distance and have similar features are automatically grouped into the same group, and each group is called an audit information cluster.

[0051] For each audit information cluster, the most frequently occurring violation types are extracted and summarized into violation patterns corresponding to the cluster. Common audit responsibility entities and audit standard clauses within the cluster are also extracted and summarized into corresponding risk scenarios. In practice, for each audit information cluster, violation types appearing in all audit information within that cluster are extracted, and the frequency of each violation type is calculated. One or more of the most frequently occurring violation types are selected and abstracted and summarized into a representative violation pattern. For example, if "fictitious expenditures" and "fake invoices" frequently occur in a cluster, the violation pattern "fictitious costs and expenses through fake invoices" might be summarized. All clusters are traversed, and each summarized violation pattern is added to a violation pattern set, which encompasses various common violation logics extracted from historical data.

[0052] Simultaneously, for the same audit information cluster, the common elements involved in the audit information within the cluster are analyzed, and the common roles or types of the responsible entities involved are extracted and statistically analyzed, such as "grassroots budget units" and "project leaders," as well as the common areas targeted by audit standard clauses. Based on these common elements, combined with the business background description of the case, a typical risk scenario is summarized.

[0053] By combining violation patterns and risk scenarios within the same audit information cluster, a violation risk representation corresponding to that cluster is obtained. The violation risk representations of each audit information cluster are then combined to construct a violation risk representation set.

[0054] Based on the set of violation patterns and risk scenarios, an audit hook set is constructed. Specifically, the structured descriptions corresponding to each violation risk representation in the set of violation risk representations are input into a pre-trained audit hook generation engine to generate triggering conditions and questioning keywords for the corresponding violation risk representation. The triggering conditions include triggering keywords, triggering entity types, and entity relationship logic. In specific implementation, the natural language descriptions and characteristics of each violation pattern and risk scenario are input into the pre-trained audit hook generation engine. The audit hook generation engine is an audit fact generation model that has been fine-tuned or optimized through prompting engineering. Based on its understanding of the audit domain, the engine automatically generates two core components for each risk pattern, including triggering conditions and questioning keywords. Triggering conditions include triggering keywords, triggering entity types, entity relationship logic, etc. Questioning keywords provide core vocabulary for subsequently generating specific questions, such as "procedural compliance," "competitiveness," and "approval process."

[0055] The generated trigger conditions are associated with the questioning keywords, and the corresponding audit hooks representing the violation risks are encapsulated. Based on the inherent logic of the audit business domain, all audit hooks are connected through semantic association to form an interconnected audit hook knowledge graph rich in semantic relationships.

[0056] S103. Call the pre-built audit hook knowledge graph, match the candidate fact triple path with each audit hook in the audit hook knowledge graph, calculate the hook matching degree corresponding to each audit event, and determine the target audit event and the target audit hook to trigger the audit based on the hook matching degree and the triggering condition of the corresponding audit hook.

[0057] Specifically, a pre-built audit hook knowledge graph is invoked to calculate the semantic similarity between the audit entities in the candidate fact triple path set and the trigger keywords and trigger entity types of each audit hook in the audit hook knowledge graph, thus obtaining the entity matching degree. The entity relationships in the candidate fact triple paths are then matched logically with the entity relationships of the audit hooks to obtain relationship matching results. Based on these results, the relationship matching degree is calculated. Finally, the entity matching degree and relationship matching degree are weighted and fused to obtain the comprehensive matching degree for each audit hook.

[0058] In practice, each audit hook in the audit hook knowledge graph is traversed. For each hook, based on its triggering conditions, a two-dimensional matching calculation is performed with the input data, including entity matching degree and relation matching degree. Entity matching degree is used to evaluate the degree of consistency between the document content and the hook at the entity level, while relation matching degree is used to evaluate the degree of consistency between the factual relationships described in the document and the risk logic pattern defined by the hook.

[0059] For entity matching, the semantic similarity between the audit entities in the candidate fact triple path set and the trigger keywords and trigger entity types of the current audit hook is calculated. Specifically, for each trigger keyword defined in the hook, a search and semantic similarity calculation are performed in the audit entity set of the document.

[0060] Simultaneously, it verifies whether the audit entity type in the document matches the trigger entity type defined by the hook. Combining the matching results of all keywords and entity types, a weighted average or other aggregation method is used to obtain an entity match score between 0 and 1. The higher the score, the more relevant the entity appearing in the document is to the risk entity of interest to the hook.

[0061] For relation matching, the compatibility between the entity relations in the candidate fact triple path set and the entity relation logic of the current audit hook is calculated. Entity relation logic is a set of logical rules defined in the hook that describes typical patterns of violations, such as: Entity A and Entity B are related AND Entity A transferred funds to Entity B.

[0062] The entity relationships and initial fact triples extracted from the audited work document are matched against logical rules. This matching process checks not only the consistency of relationship types but also whether the entity types conform to the constraints in the rules. A relationship matching degree is calculated based on the extent to which the factual relationships in the audited work document satisfy the hook's logical rules. A higher value indicates that the chain of facts described in the document more closely matches the potential violation pattern defined by the hook.

[0063] For each audit hook that is scanned, the entity matching degree and relationship matching degree calculated above are weighted and fused to obtain the comprehensive matching degree corresponding to the audit hook.

[0064] Traverse each audit hook in the audit hook knowledge graph, compare the comprehensive matching degree of each audit hook with the predefined triggering conditions of the audit hook one by one. If the comprehensive matching degree of an audit hook meets its own triggering conditions, then mark the audit hook as a candidate triggering hook.

[0065] Generate a candidate trigger hook list, which contains all audit hooks that meet the trigger conditions and their corresponding overall match scores. Sort the list in descending order of overall match score, and based on a preset global maximum number of triggers, select the top N candidate trigger hooks from the list.

[0066] Furthermore, the candidate trigger hooks after screening are finally confirmed, and the candidate trigger hooks that have been screened and sorted are officially determined as target audit hooks. The complete information of each target audit hook is recorded, including its triggering conditions, associated questioning keywords and corresponding comprehensive matching degree, and a set of target audit hooks is generated.

[0067] It should be noted that the triggering conditions adopt a dynamic threshold mechanism, which can be adaptively adjusted according to the risk level associated with the audit hook. High-risk audit hooks can be set with lower trigger thresholds to improve detection sensitivity.

[0068] S104. Based on the target challenge keywords associated with the target audit hook, the audit fact generation model is used to perform deep information verification on the target audit event within the scope of the audited work document, and generate a confirmatory fact triplet corresponding to the target audit event.

[0069] Specifically, the content of the audit documents is vectorized to generate semantic representation vectors, which are then stored in a pre-defined vector database to build an audit document index. In practice, the document is segmented into semantic paragraphs, and each paragraph is converted into a vector representation using an embedding model to generate semantic representation vectors. All semantic representation vectors are then stored in a pre-defined vector database to establish a complete audit document index.

[0070] Obtain the target challenge keywords associated with the target audit hooks, and generate multiple natural language challenge questions based on the target challenge keywords to obtain a target challenge question sequence.

[0071] In practice, the target audit hooks are identified by acquiring the target challenge keywords and inputting them into the question generation module. This module dynamically combines the keywords to generate multiple natural language challenge questions that fit the audit context, forming a sequence of target challenge questions.

[0072] Furthermore, the questions raised are vectorized to generate query vectors. In practice, each question in the target question sequence is vectorized by converting its natural language text into a numerical query vector using a pre-defined embedding model. Subsequently, the query vectors are semantically retrieved from a pre-defined vector database. By calculating the cosine similarity between the query vector and each semantic representation vector in the database, the K most similar semantic representation vectors are identified as the most similar semantic representation vectors corresponding to the question raised.

[0073] The most similar semantic representation vector is used to set the most relevant text fragment in the original text of the work document to be audited as the most relevant text fragment corresponding to the question, thus obtaining the most relevant text fragment sequence corresponding to the target question sequence, which serves as the limiting context. In specific implementation, based on the association relationship between semantic representation vectors and original text fragments stored in a preset vector database, the retrieved most similar semantic representation vectors are mapped one by one to obtain their corresponding original text fragments in the work document to be audited; these fragments are combined in order to set them as the most relevant text fragments corresponding to the question; all question questions in the target question sequence are traversed, and the obtained most relevant text fragments are arranged in question order to form the most relevant text fragment sequence corresponding to the target question sequence, which serves as the limiting context required for subsequent in-depth information verification.

[0074] The target question sequence and the limited context are input into the audit fact generation model. The audit fact generation model performs reasoning and verification for each question within the limited context, and generates the corresponding intermediate verification triples based on the preset structured triple format.

[0075] Specifically, semantic role labeling is performed on the most relevant text fragments within the defined context to identify core actions, executing subjects, executing objects, and subject-object logical relationships. Based on a preset structured triplet format, local logical triples for the most relevant text fragments are constructed. In practice, the target question sequence and its corresponding defined context are input into the audit fact generation model. For each question, semantic role labeling is performed on the text fragments within the defined context to identify core actions, executing subjects, and executing objects. Logical relationship connectors between executing subjects and executing objects are extracted from the text fragments to obtain the subject-object logical relationship, and local logical triples in the format of executing subject, executing object, and subject-object logical relationship are constructed.

[0076] The process involves matching the challenged issue with local logical triples. Based on the matching results, the attitude of the most relevant text fragment towards the challenged issue is determined. This attitude includes support, rejection, and insufficient proof. Intermediate validation triples are generated based on these attitudes. Conflict detection and information fusion are performed on all intermediate validation triples to generate confirmatory fact triples. In practice, the conclusion fields of different intermediate validation triples are compared to identify semantically contradictory conclusions. For contradictory conclusions, a weighted vote is performed based on their evidence source coverage and confidence level. Based on the weighted voting results, the conclusion with the highest confidence level is adopted. The final conclusion, along with relevant entities and relationships, is reconstructed into a structured confirmatory fact triple.

[0077] Furthermore, conflict detection and information fusion are performed on all intermediate validation triples to generate confirmatory fact triples corresponding to the target audit event. In practice, the claims and core evidence in different intermediate validation triples are compared to identify logically contradictory or semantically inconsistent conclusions. Based on pre-defined rules, such as prioritizing high-confidence results and weighting based on evidence coverage, conflicting conclusions are arbitrated and reconciled, and core factual elements consistently agreed upon or established after arbitration are extracted from all results. These factual elements are combined to generate confirmatory fact triples in a pre-defined structured triple format representing the final verification conclusion of the target audit event. It should be noted that the intermediate validation triples ensure that the final output triples are rigorous products resulting from multi-faceted questioning, cross-validation of evidence, and conflict resolution, thereby significantly improving the reliability and accuracy of the audit conclusions.

[0078] S105. Using the audit rule knowledge graph and based on the confirmatory fact triples, determine the risk type and risk level corresponding to the target audit event to generate an audit conclusion report.

[0079] The process involves obtaining the verification execution subject, verification execution object, and the logical relationship between the verification subject and object from the confirmatory fact triples. The semantic similarity between this logical relationship and the rule relationships in the audit rule knowledge graph is then calculated, identifying candidate rule association edges whose semantic similarity exceeds a preset similarity threshold. Specifically, after fusing intermediate verification triples, confirmatory fact triples are obtained, and the verification execution subject, verification execution object, and logical relationship between the verification subject and object are obtained from these triples. A pre-built audit rule knowledge graph is then invoked to automatically match the confirmatory fact triples corresponding to the target audit event with the audit rule knowledge graph. Specifically, a pre-trained semantic similarity calculation model is used to convert the text of the rule relationship carried by each rule association edge in the audit rule knowledge graph into vector representations, and the semantic similarity between the two is calculated. Rule association edges with similarity exceeding a preset similarity threshold are selected as candidate rule association edges for preliminary matching.

[0080] The process involves obtaining the candidate rule subject and object associated with each candidate rule's associated edge, and determining whether the verification execution subject and object belong to the same entity type as the candidate rule subject and object. In practice, for each candidate rule associated edge, the connected candidate rule subject and object nodes are obtained, and their corresponding node type attributes are read. Simultaneously, the entity type identifiers of the verification execution subject and object are obtained. Entity type-level matching and verification are completed by comparing the entity types of the verification execution subject and candidate rule subject, and the verification execution object and candidate rule object.

[0081] If the semantic similarity of a candidate edge meets the standard and the entity type verification passes, it is determined that the confirmatory fact triplet matches the rule-related edge successfully. The pre-set risk level judgment conditions and the mapped risk type field value of the rule-related edge are read, the mapped risk type is used as the risk type of the target audit event, and the audit rule clause index bound to the rule-related edge is determined.

[0082] Compliance assessments are conducted based on verification attribute values ​​and risk level determination criteria. Based on the compliance assessment results, the risk level corresponding to the target audit event is determined. In practice, information extraction is performed based on the location and context of the verification execution entity and object within the audited work document. Methods such as named entity recognition and regular expression matching are used to extract relevant numerical quantitative parameters and textual state parameters, collectively referred to as verification attribute values.

[0083] If the risk level determination condition is a quantitative rule, then the numerical quantitative parameters are substituted into the inequalities or threshold expressions in the condition for calculation; if the risk level determination condition is a qualitative rule, then the textual status parameters are matched with the compliance requirement text in the condition using pattern matching or keyword conformity analysis. Based on the judgment result and a preset mapping table, the corresponding risk level, such as "high," "medium," or "low," is output, thereby completing the risk level determination of the target audit event.

[0084] The process involves obtaining the audit rule clause index corresponding to the candidate rule association edges, and generating an audit conclusion report based on confirmatory fact triples, the audit rule clause index, risk type, and risk level using a pre-set audit report template. In practice, the process involves obtaining the audit rule clause index corresponding to the candidate rule association edges in the audit rule knowledge graph, and structurally integrating the identified confirmatory fact triples, matching rule texts, risk types, and risk levels to generate standardized audit findings items. These audit findings items are automatically associated with their positions in the original text of the audited document, and quantitative evidence of semantic matching and compliance judgments is embedded to form a traceable chain of evidence. Using a pre-set audit report template, the audit findings items are sorted and categorized by risk level to generate an audit conclusion report, thereby achieving automation, structuring, and traceability of audit conclusions.

[0085] This application embodiment also includes optimizing the model parameters of the audit fact generation model through a paradigmatic audit knowledge graph after the audit conclusion report is generated.

[0086] Specifically, the matching degree between confirmatory fact triples and pre-stored standard answers is calculated to obtain an accuracy score. The comprehensive semantic relevance between confirmatory fact triples and the target question sequence is calculated to obtain a relevance score. Based on the number of occurrences of confirmatory fact triples in the candidate fact triple path set, an information addition score is calculated. The accuracy score, relevance score, and information addition score are weighted and summed to obtain the questioning reward value. Based on the questioning reward value, a reinforcement loss function term for the audit fact generation model is defined, and the model parameters of the audit fact generation model are adjusted using the gradient descent algorithm to minimize the loss function, thereby optimizing the audit fact generation model.

[0087] In practice, confirmatory fact triples are matched with standard facts in a pre-built audit rule knowledge graph. A combination of exact matching and semantic matching is used to calculate the degree of agreement between the confirmatory fact triples and the standard facts. Based on the matching results, an accuracy score between 0 and 1 is generated, with a higher score indicating higher fact accuracy.

[0088] The semantic similarity calculation model is used to evaluate the comprehensive semantic relevance between the confirmatory fact triple and the target question sequence. It also considers the degree of direct response and indirect support of the verification results to the question, and outputs a relevance score between 0 and 1. The higher the score, the closer the association between the verification results and the question.

[0089] The confirmatory fact triples are compared with the initially generated candidate fact triple path set. The number and importance weight of new facts in the confirmatory fact triples that did not appear in the initial set are counted. The information addition score is calculated based on the total value of the added information. The higher the score, the more in-depth information the model has uncovered.

[0090] The three ratings are weighted and summed to generate the final questioning reward value.

[0091] Based on the challenge reward value, the model parameters of the audit fact generation model are iteratively optimized using a near-end policy optimization algorithm. Specifically, the current performance of the model in the audit task is used as the policy, and the challenge reward value is used as the reward signal for this policy. The gradient of the challenge reward value with respect to the model parameters is calculated to determine the direction and magnitude of parameter updates. While ensuring the magnitude of policy updates, the model parameters are updated along the gradient direction. This process is repeated so that the model gradually learns how to generate more accurate and insightful audit facts.

[0092] This application constructs an audit rule knowledge graph and designs a loss function that integrates knowledge constraint loss and supervision matching reward to specifically train a large language model. This ensures that when the model generates audit facts, its internal logic and output are guided by the structured rules of the domain. This training mechanism can effectively correct any biases or ambiguities that may arise from the model's free generation, thereby improving the standardization of the fact triples extracted from complex documents in terms of business meaning, logical consistency, and compliance with actual legal provisions, laying a reliable factual foundation for subsequent risk assessment.

[0093] Furthermore, compared to traditional passive matching methods, this application, through a pre-constructed audit hook knowledge graph, can intelligently scan and match the factual chains initially generated by the model, automatically locating high-risk suspicious points and triggering targeted in-depth verification. This proactive questioning and multi-round verification mechanism enables the audit fact generation model to penetrate the surface text, enhancing its ability to detect implicit contradictions, logical loopholes, or deliberately evasive statements in the text, thereby more effectively uncovering systemic and interconnected chains of violations.

[0094] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device and medium embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the description of the method embodiments.

[0095] The devices and media provided in this application are one-to-one with the methods. Therefore, the devices and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be repeated here.

[0096] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0097] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0098] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0099] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0100] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0101] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0102] Computer-readable media include both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0103] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0104] The above description is merely an embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of this application should be included within the scope of the claims of this application.

Claims

1. An intelligent auditing method based on knowledge graph constraint, characterized in that, The method comprises the following steps: constructing an audit rule knowledge graph, and performing constraint training on an initial large language model based on the audit rule knowledge graph to obtain an audit fact generation model; The constrained training is achieved by minimizing the fusion loss function, which is defined as follows: ,in, This is the initial large language model. These are all the trainable model parameters of the initial large language model. For pre-built training samples, for from The set of fact triples extracted from it. for The set of rule triples mapped onto the audit rule knowledge graph. for and The divergence value between them For a pre-labeled set of standard rule triples, For calculation based on the maximum conjugate string matching algorithm and Matching reward value, For hyperparameters; receiving a work document to be audited, inputting the work document to be audited into the audit fact generation model, generating candidate fact triple paths corresponding to each audit event in the work document to be audited, and obtaining a candidate fact triple set; calling a pre-constructed audit hook knowledge graph, matching the candidate fact triple paths with each audit hook in the audit hook knowledge graph, calculating the hook matching degree corresponding to each audit event, and determining the target audit event triggering the audit and the target audit hook triggering the audit according to the hook matching degree and the trigger condition of the corresponding audit hook; based on the target audit hook associated target audit keyword, performing deep information verification on the target audit event in the limited range of the work document to be audited through the audit fact generation model, and generating a verification fact triple corresponding to the target audit event; determining the risk type and risk level corresponding to the target audit event based on the verification fact triple through the audit rule knowledge graph to generate an audit conclusion report. 2.The knowledge graph constraint-based intelligent auditing method of claim 1, wherein, The constraint training of the initial large language model based on the audit rule knowledge graph to obtain the audit fact generation model comprises the following steps: inputting pre-constructed training samples into an initial large language model, outputting a fact triple set of the training samples and a first probability distribution corresponding to each fact triple; inputting the audit rule knowledge graph and the fact triple set into the initial large language model, outputting a rule triple set and a second probability distribution corresponding to each rule triple; based on the first probability distribution and the second probability distribution, calculating the divergence value of the fact triple set and the rule triple set; The matching degree between the rule triple set calculated by the maximum conjugate string matching algorithm and the pre-labeled standard rule triple set is calculated, and the matching degree is normalized to obtain a matching degree reward value; wherein the matching degree reward value calculation formula is: , represents the length of the longest common subsequence between and obtained by the maximum conjugate string matching algorithm, represents the length of . based on the divergence value and the matching degree reward value, constructing a fusion loss function, and iteratively optimizing the initial large language model until the model converges to obtain an audit fact generation model. 3.The knowledge graph constraint-based intelligent auditing method of claim 1, wherein, The method comprises the following steps: obtaining an audit rule file of a target audit range in an audit business field, analyzing the audit rule file through a natural language processing engine, and extracting structured rule information; the structured rule information comprises a rule subject, a rule relationship, and a rule object; taking the rule subject and the rule object as rule entity nodes and taking the rule relationship as a rule association edge to construct a preliminary rule knowledge graph; adding attribute information to the rule entity nodes and the rule association edges to obtain a constructed rule knowledge graph; the attribute information comprises a rule constraint, a rule source, an effective time, an applicable range, and a mapped risk type; the rule constraint comprises a risk level determination condition; storing the constructed rule knowledge graph into a graph database and establishing a corresponding index structure to obtain an audit rule knowledge graph. 4.The knowledge graph constraint-based intelligent auditing method of claim 3, wherein, Before the method receives the work document to be audited, the method further comprises the following steps: Obtaining historical audit case data of the target audit scope, extracting structured audit information of the historical audit case data through the natural language processing engine; the structured audit information includes violation behavior type, involved responsible subject and involved audit standard clause; Vectorizing the structured audit information to obtain a comprehensive feature vector, and dividing similar features in the comprehensive feature vector into the same audit information cluster through an unsupervised clustering algorithm; For each audit information cluster, extracting the most frequently occurring violation behavior type and inducing it as a violation mode corresponding to the audit information cluster, and extracting the common involved audit responsible subject and audit standard clause in the audit information cluster and inducing them as a corresponding risk scenario; Combining the violation mode and the risk scenario to obtain a violation risk representation corresponding to the audit information cluster, so as to construct a violation risk representation set; Based on the violation risk representation set, an audit hook set is constructed, and each audit hook in the audit hook set is semantically associated according to the audit business of the audit business field, so as to construct an audit hook knowledge graph. 5.The knowledge graph constraint-based intelligent auditing method of claim 4, wherein, The construction of the audit hook set based on the violation risk representation set specifically includes: Inputting the structured description of each violation risk representation in the violation risk representation set into a pre-trained audit hook generation engine to generate a trigger condition and a questioning keyword corresponding to the violation risk representation; the trigger condition includes a trigger keyword, a trigger entity type and an entity relationship logic; The trigger condition and the questioning keyword are associated and encapsulated to generate an audit hook corresponding to the violation risk representation, so as to construct an audit hook set. 6.The knowledge graph constraint-based intelligent auditing method of claim 1, wherein, The receiving of the to-be-audited work document and the inputting of the audit fact generation model to generate a candidate fact triple path corresponding to each audit event in the to-be-audited work document specifically include: Receiving a to-be-audited work document and inputting the audit fact generation model to extract audit event information of the to-be-audited work document; the audit event information includes audit entities and corresponding entity relationships, context time sequence of the audit entities, and logical connection words between the audit entities; Taking the audit entities as the subject and the object and taking the entity relationship as the relationship, combining based on a preset structured triple format to construct an initial fact triple to obtain a triple set corresponding to each audit event; the preset structured triple format is a structured format of subject, relationship and object; According to the context time sequence of the audit entities and the logical connection words between the audit entities, the time sequence relationship and the logical relationship between each initial fact triple in the triple set are determined; According to the time sequence relationship and the logical relationship, the initial fact triples are linked and sorted to generate a candidate fact triple path corresponding to the audit event. 7.The knowledge graph constraint-based intelligent auditing method of claim 6, wherein, The target audit hook is associated with the target audit event, and the target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with the target audit hook. The target audit event is associated with ​ ​ ​ ​ ​ ​ 8.The knowledge graph constraint-based intelligent auditing method of claim 7, wherein, ​ ​ ​ ​ 9.The knowledge graph constraint-based intelligent auditing method of claim 8, wherein, ​ ​ ​ obtaining a candidate rule subject and a candidate rule object associated with the candidate rule association edge, and determining whether the verification execution subject and the verification execution object belong to the same entity type as the candidate rule subject and the candidate rule object; if yes, determining that the verification fact triple matches the candidate rule association edge successfully, obtaining a risk level determination condition and a mapped risk type from attribute information of the candidate rule association edge, and obtaining a risk type of the target audit event; extracting verification attribute values associated with the verification execution subject and the verification execution object from the to-be-audited work document; the verification attribute values include numerical quantitative parameters and text state parameters; performing compliance judgment based on the verification attribute values and the risk level determination condition, and determining a risk level corresponding to the target audit event according to a compliance judgment result; obtaining an audit rule clause index corresponding to the candidate rule association edge, and generating an audit conclusion report based on the verification fact triple, the audit rule clause index, the risk type, and the risk level through a preset audit report template. 10.The knowledge graph constraint-based intelligent auditing method of claim 7, wherein, After determining the risk type and the risk level corresponding to the target audit event based on the verification fact triple to generate the audit conclusion report, the method further includes: calculating a matching degree of the verification fact triple and a pre-stored standard answer to obtain an accuracy score; calculating a comprehensive semantic correlation degree of the verification fact triple and the target question sequence to obtain a relevance score; calculating an information addition score based on a number of occurrences of the verification fact triple in the candidate fact triple path set; performing weighted summation on the accuracy score, the relevance score, and the information addition score to obtain a question reward value; defining a reinforcement loss function item of the audit fact generation model based on the question reward value, and adjusting model parameters of the audit fact generation model through a gradient descent algorithm to minimize the loss function, so as to perform online optimization on the audit fact generation model.

Images