Complex document-oriented multi-mode internal control compliance intelligent auditing method and system

Abstract

The invention discloses a complex document-oriented multi-mode internal control compliance intelligent auditing method and system, and relates to the technical field of intelligent auditing. Comprising the following steps: obtaining a system text, an approval record, a business document, a contract attachment and an image scanning copy, performing structured analysis to obtain system semantic information, business behavior information and evidence information, constructing a standard approval path, comparing the standard approval path with the approval record, and obtaining an approval path comparison result; identifying a system version and a time label, constructing a system version tense semantic mapping model, matching the corresponding system version according to the occurrence time of the business behavior, and obtaining a system version tense semantic mapping result; analyzing the system rule, identifying an exception trigger condition, constructing a semantic rule map of system exception terms, judging whether the business behavior is suitable for the exception rule, and obtaining an exception rule judgment result; according to the method, the objectivity and the interpretability of rule reasoning are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent auditing technology, and in particular to a multimodal intelligent auditing method and system for complex documents. Background Technology

[0002] As enterprises deepen their digital transformation, the data environment faced by internal control and compliance audits has evolved from simple structured records to a complex scenario encompassing multimodal and heterogeneous documents from multiple sources, including policy documents, approval records, business documents, contract attachments, and scanned images. Traditional auditing methods rely on auditors manually reviewing, comparing, and judging data, which is not only inefficient but also struggles to address deeper logical issues hidden within massive amounts of documents, such as rule temporal changes, exception clause identification, evidence chain correlation, and rule conflict resolution. In recent years, some intelligent auditing systems have attempted to introduce rule engines or knowledge bases to transform policy rules into executable judgment logic, aiming to improve the level of audit automation.

[0003] However, existing technologies still have significant shortcomings in handling the correlation, dynamism, and uncertainty of multimodal documents. First, the rules and regulations upon which audits are based often have multiple versions running concurrently with overlapping effective dates. Existing systems struggle to automatically match the applicable rule version based on the timing of business actions, leading to temporal misalignments in rule application. Second, exception clauses in regulations are typically expressed in scattered natural language, lacking structured semantic modeling, making it difficult for the system to accurately determine whether a business action falls under an exception. More critically, there is a lack of unified correlation and credibility measurement mechanisms between evidence from different sources, such as documents, images, and approval records, resulting in broken evidence chains or an inability to assess evidence quality, thus affecting the reliability of rule-based judgments. When multiple rules produce conflicting conclusions regarding the same business action, existing systems also lack the ability to select the best rule based on evidence credibility and regulatory hierarchy. Summary of the Invention

[0004] This invention proposes an intelligent auditing method that enables dynamic adaptation of system versions, accurate identification of exception rules, unified modeling of multimodal evidence chains, and intelligent resolution of rule conflicts in complex document environments characterized by multi-source heterogeneity, temporal interleaving, and rule conflicts.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A multimodal intelligent auditing method for complex documents, comprising: The system obtains and structures the policy text, approval records, business documents, contract attachments and image scans to obtain policy semantic information, business behavior information and evidence information. After constructing a standard approval path, it is compared with the approval records to obtain the approval path comparison results. Identify the policy version and time tag, construct a policy version temporal semantic mapping model, and match the corresponding policy version according to the time of business behavior to obtain the policy version temporal semantic mapping result; Analyze the rules and regulations, identify the exception triggering conditions and construct a semantic rule graph of the exception clauses, determine whether the business behavior is subject to the exception rules, and obtain the exception rule determination results; The source identification, structural analysis and generation path tracing of business documents, contract attachments and image scans are performed to construct a multimodal evidence chain. The credibility quantification calculation of the evidence nodes in the multimodal evidence chain is performed to obtain the credibility quantification result of the evidence. Based on the information of institutional rules and the quantification results of evidence credibility, an evidence-rule coupling relationship graph is constructed. The rule applicability confidence of candidate institutional rules is calculated, and conflicting rules are resolved by combining the institutional hierarchy relationship to obtain an effective rule set. Based on the approval path comparison results, the temporal semantic mapping results of the system version, the exception rule determination results, and the set of effective rules, a system application status transition diagram is constructed to determine the application path of the target system. An audit reasoning path relationship diagram is constructed along the target system application path, and an audit interpretation path is generated. Based on the audit interpretation path and the results of evidence credibility quantification, a comprehensive analysis is conducted to output the compliance risk level and the corresponding audit interpretation results.

[0006] As a preferred technical solution of the present invention, the acquisition of the approval path comparison result includes: extracting each approval node and the sequential relationship between nodes according to the approval process rules in the system text, and constructing a standard approval path; extracting the actual approval nodes and their execution order from the approval records to form the actual approval path; performing node matching and sequence consistency comparison between the actual approval path corresponding to the approval record and the standard approval path, identifying approval nodes that exist in the standard approval path but do not appear in the actual approval path, and taking the identified approval nodes as implicit missing nodes to generate the approval path comparison result.

[0007] As a preferred technical solution of the present invention, the acquisition of the system version temporal semantic mapping result includes: extracting system identifier, version identifier, and time tag information from the system text, and classifying different versions of system text according to the system identifier; constructing a system version temporal semantic mapping model based on the version identifier and time tag information, wherein the system version temporal semantic mapping model is used to establish the correspondence between each system version and its corresponding system semantic information and applicable time interval; determining its applicable time interval according to the time tag corresponding to each system version; mapping the time of occurrence of business behavior to the corresponding applicable time interval, determining the system version corresponding to the business behavior information through the system version temporal semantic mapping model, and generating the system version temporal semantic mapping result.

[0008] As a preferred technical solution of the present invention, the acquisition of the exception rule determination result includes: performing rule semantic parsing on the system text to identify the condition content in the system rules and their corresponding exception triggering conditions; storing the system rules and their corresponding exception triggering conditions in association to construct a semantic rule graph of system exception clauses, wherein different system rules are associated with their corresponding exception triggering conditions; extracting feature information corresponding to the exception triggering conditions based on business behavior information, matching the feature information with the exception triggering conditions to determine whether the business behavior meets the exception triggering conditions, and determining that the corresponding system rule is subject to the exception rule when the exception triggering conditions are met, thereby generating an exception rule determination result.

[0009] As a preferred technical solution of the present invention, the acquisition of the evidence credibility quantification result includes: dividing the evidence information obtained from parsing business documents, contract attachments, and image scans into multiple evidence nodes; determining the temporal sequence relationship between evidence nodes based on the generation time information corresponding to each evidence node, determining the source association relationship between evidence nodes based on the source information of each evidence node, determining the content correspondence relationship between evidence nodes based on the content information of each evidence node, and establishing the association relationship between evidence nodes based on the temporal sequence relationship, source association relationship, and content correspondence relationship to construct a multimodal evidence chain; and performing credibility quantification calculation on each evidence node based on the generation time information, source information, and content consistency of each evidence node to obtain the evidence credibility quantification result.

[0010] As a preferred technical solution of the present invention, the construction of the evidence-rule coupling relationship diagram includes: extracting rule element information corresponding to each rule based on the rule information, and associating and matching the rule element information with evidence nodes to determine the set of evidence nodes on which each rule depends; establishing the association relationship between the rule and the evidence nodes based on the evidence node set and the corresponding evidence credibility quantification results, and constructing the evidence-rule coupling relationship diagram; performing weighted calculation based on the evidence credibility quantification results of each evidence node in the evidence node set corresponding to the rule, and correcting it in combination with the content consistency and time sequence relationship between each evidence node to obtain the rule applicability confidence.

[0011] As a preferred technical solution of the present invention, the acquisition of the effective rule set includes: comparing the consistency of the application results of candidate system rules; when different candidate system rules have inconsistent application results for the same business behavior, they are determined to be conflicting rules; ranking the candidate system rules according to the rule application confidence of each candidate system rule; calculating the evidence confidence distribution characteristics based on the evidence confidence quantification results of the evidence node set corresponding to each candidate system rule, and correcting the rule application confidence according to the evidence confidence distribution characteristics; and comprehensively comparing the corrected rule application confidence in combination with the system hierarchy, selecting candidate system rules whose rule application confidence meets the preset conditions as effective rules, and forming an effective rule set.

[0012] As a preferred technical solution of the present invention, the determination of the target system application path includes: combining the approval completion status, system version, and rule application results to generate multiple candidate system application states; arranging the candidate system application states chronologically according to the time of business behavior occurrence, and establishing state transition relationships between candidate system application states based on exception rule judgment results, approval path comparison results, effective rule set, and evidence credibility quantification results to obtain a system application state transition diagram; performing constraint consistency verification on each state transition relationship in the system application state transition diagram, and deleting state transition relationships that do not meet the constraint conditions; performing path traversal on the retained state transition relationships to obtain multiple candidate paths, and comparing the rule application confidence and evidence credibility quantification results corresponding to the candidate paths to determine the target system application path corresponding to the business behavior information.

[0013] As a preferred technical solution of the present invention, the comprehensive analysis includes: along the target system application path, associating and expanding each system version, rule application result, and corresponding evidence node in the path according to the state transition sequence, constructing an audit reasoning path relationship diagram, and generating an audit interpretation path; based on the evidence credibility quantification results of the evidence nodes corresponding to each rule application result in the audit interpretation path, performing a consistency assessment on each rule application result, and correcting or eliminating rule application results that do not meet the consistency conditions; based on the corrected rule application results, performing a compliance assessment on business behavior information, and outputting the compliance risk level and corresponding audit interpretation results.

[0014] A multimodal intelligent audit system for internal control and compliance, oriented towards complex documents, includes: Path comparison module: Acquires and parses policy texts, approval records, business documents, contract attachments and image scans in a structured manner to obtain policy semantic information, business behavior information and evidence information. After constructing a standard approval path, it is compared with the approval records to obtain the approval path comparison results. Version mapping module: Identifies policy versions and time tags, constructs a policy version temporal semantic mapping model, matches the corresponding policy version according to the time of business behavior occurrence, and obtains the policy version temporal semantic mapping results; Exception determination module: Parses the rules and regulations, identifies the exception triggering conditions and constructs a semantic rule graph of the exception clauses, determines whether the business behavior is subject to the exception rules, and obtains the exception rule determination results; Trust measurement module: performs source identification, structural analysis and generation path tracing on business documents, contract attachments and image scans, constructs a multimodal evidence chain, performs trust measurement calculation on evidence nodes in the multimodal evidence chain, and obtains the trust measurement results of the evidence. Effective rule module: Based on the information of institutional rules and the quantification results of evidence credibility, an evidence-rule coupling relationship graph is constructed, the rule applicability confidence of candidate institutional rules is calculated, and conflicting rules are resolved by combining institutional hierarchy relationships to obtain a set of effective rules; The system application module constructs a system application status transition diagram based on the approval path comparison results, system version temporal semantic mapping results, exception rule determination results, and effective rule set to determine the application path of the target system. Comprehensive Audit Module: Constructs an audit reasoning path relationship diagram along the target system application path and generates an audit interpretation path. Based on the audit interpretation path and the quantification results of evidence credibility, it performs a comprehensive analysis and outputs the compliance risk level and the corresponding audit interpretation results.

[0015] The present invention has the following advantages: This invention solves the technical problem of automatically discovering "not-occurring but should have occurred" in the compliance check of the approval process by automatically extracting the approval process rules in the institutional text into standard approval paths and comparing them with the actual approval records at the node level.

[0016] This invention constructs a semantic rule graph of institutional exception clauses, which structurally associates the exception triggering conditions scattered in institutional texts with the corresponding rules, and automatically determines the applicability of exception rules based on business feature information, thereby realizing the automated system identification of exception clauses.

[0017] This invention divides multi-source evidence into evidence nodes and constructs a multimodal evidence chain based on chronological order, source association, and content correspondence. It also performs credibility quantification calculation on each evidence node. By constructing an evidence-rule coupling relationship graph, it associates and matches the credibility quantification results of evidence nodes with institutional rule elements and calculates the rule applicability confidence level with weights, thereby improving the objectivity and interpretability of rule reasoning.

[0018] This invention effectively solves the problem of how to select the best rule to apply when multiple rules produce conflicting conclusions about the same business behavior by introducing the distribution characteristics of evidence credibility to modify the confidence of rule application when multiple rules conflict, and by combining the hierarchical relationship of the system for comprehensive comparison and screening.

[0019] This invention constructs a system application state transition diagram, performs temporal arrangement, constraint verification, and path traversal on candidate states, and determines the optimal system application path by combining rule application confidence and evidence credibility, thereby achieving dynamic adaptation and optimization of the audit process.

[0020] This invention constructs an audit reasoning path relationship diagram along the target system application path, and performs consistency assessment and correction on the rule application results based on the credibility of evidence, ultimately generating a traceable audit interpretation path. This solves the problem of audit results being "black boxed" and difficult to explain to non-technical users, and significantly improves the transparency and credibility of audit conclusions. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are only schematic diagrams of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort. Figure 1 This is a schematic diagram of the structure of a multimodal internal control compliance intelligent audit system for complex documents, as used in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0023] Example 1: A multimodal intelligent auditing method for complex documents, comprising the following steps: Step S1: Obtain the policy text, approval records, business documents, contract attachments and image scans and perform structured parsing to obtain policy semantic information, business behavior information and evidence information. After constructing a standard approval path, compare it with the approval records to obtain the approval path comparison results. The acquisition of the approval path comparison results includes: extracting each approval node and the sequential relationship between nodes according to the approval process rules in the policy document, and constructing a standard approval path; extracting the actual approval nodes and their execution order from the approval records to form the actual approval path; performing node matching and sequence consistency comparison between the actual approval path corresponding to the approval record and the standard approval path, identifying approval nodes that exist in the standard approval path but do not appear in the actual approval path, and using the identified approval nodes as implicit missing nodes to generate the approval path comparison results.

[0024] In this step, the policy text primarily originates from the policy management platform, document management directory, or business archive files. The policy text includes at least the policy name, policy number, version identifier, effective date, scope of application, process clauses, and node constraint clauses. When performing structured parsing of the policy text, the text is first segmented into its layout, identifying title levels, clause numbers, paragraph boundaries, and process description fragments. Then, rule element extraction is performed on the process description fragments to obtain the approval node name, node triggering conditions, node responsible entity, node sequence relationship, and node entry conditions. This content is then organized into policy semantic information. This policy semantic information is normative data used to characterize the structured rule content related to the approval process in the policy text.

[0025] The approval records primarily originate from office approval process data, business flow logs, or financial approval trace data. Each approval record includes at least the process number, business order number, node name, submission time, approval time, approver identifier, approval result, return identifier, transfer identifier, and process status. When performing structured parsing of the approval records, approval logs corresponding to the same business item are aggregated according to the process number and business order number. The occurrence sequence, execution status, and time position of each approval node are extracted to form the approval execution sequence in the business behavior information. This business behavior information belongs to business execution data and is used to characterize the node flow of a specific business item in the actual processing process.

[0026] The business documents, contract attachments, and scanned images primarily originate from archived business forms, archived contract data, and scanned uploaded files. Business documents must include at least the application number, business type, applying department, application amount, application time, and associated business identifier; contract attachments must include at least the contract number, contract signing time, payment terms, signature page, and supplementary agreement; scanned images must include at least scanned invoices, scanned paper approval forms, images of signed receipts, and images of supplementary explanation pages. When performing structured parsing on the business documents, contract attachments, and scanned images, the document identifier, associated business identifier, creation time, upload time, source type, amount field, signature field, and text content field are extracted to form evidentiary information. This evidentiary information constitutes factual supporting data, used to characterize the source, formation sequence, and content load of the materials corresponding to the business matter.

[0027] When extracting approval nodes and their sequential relationships based on the approval process rules in the policy document, the process clauses corresponding to the current business matter are located in the policy document. The node descriptions in these clauses are decomposed, identifying sequential connectors, conditional qualifiers, and node action words. Node names in different expressions are mapped to a unified node identifier, and node connections are established based on the dependencies between clauses to construct a standard approval path. This standard approval path is an ordered sequence of nodes generated according to the policy rules. When the policy document contains both basic path rules and additional node rules, the additional nodes are assembled based on the business type, amount range, and matter attributes in the business document to generate a standard approval path corresponding to the current business matter.

[0028] When extracting actual approval nodes and their execution sequence from approval records, the approval logs are sorted according to approval time, node status, and flow identifier. Returned, resubmitted, transferred, and supplementary approval records are then linked and merged to form the actual approval path. During merging, node records with the same business order number, the same process master identifier, or the same approval item identifier are grouped into the same business processing chain. For records with inconsistent node names but consistent meaning, the node names are standardized and mapped to a unified node identifier to eliminate interference caused by heteronymous nodes such as "Legal Review," "Legal Examination," and "Contract Law Review" in path identification.

[0029] When performing node matching and sequence consistency comparison between the actual approval path and the standard approval path corresponding to the approval record, a node correspondence relationship is first established between the node sequence in the standard approval path and the node sequence in the actual approval path. Then, sequence consistency verification is performed according to the node position index, predecessor node relationship, and successor node relationship. The node matching includes node identifier matching, node attribute matching, and node position matching; the sequence consistency comparison includes verification of the sequential relationship between adjacent nodes, verification of the existence of necessary nodes, and verification of the integrity of whether a condition node appears after being triggered.

[0030] When identifying approval nodes that exist in the standard approval path but do not appear in the actual approval path, the corresponding node identifier in the actual approval path is sequentially retrieved for each node in the standard approval path. If a target node does not exist in the actual approval path, but its preceding and following related nodes have already appeared in the actual approval path, or if the triggering condition corresponding to the node has been met in the business document but the node still does not appear in the approval record, the node is marked as an implicitly missing node. These implicitly missing nodes are not generated based on subjective judgment, but are determined based on the structured comparison results between policy rules, business attributes, and approval execution logs. For example, in a procurement payment scenario, if the business document states a procurement amount of 820,000 yuan, meeting the triggering condition of "adding a legal review node for large-amount procurements" in the policy clause, but no legal review node is found in the actual approval path formed by the approval record, then the legal review node is marked as an implicitly missing node.

[0031] When generating the approval path comparison results, the standard approval path identifier, the actual approval path identifier, the node matching result, the sequence consistency verification result, the implicit missing node identifier, the node abnormal position, and the corresponding business item identifier are associated and summarized to form the approval path comparison result. The approval path comparison result includes at least the set of matched nodes, the set of unmatched nodes, the set of implicitly missing nodes, and the set of nodes with abnormal sequence.

[0032] Step S2: Identify the policy version and time tag, construct a policy version temporal semantic mapping model, and match the corresponding policy version according to the time of business behavior occurrence to obtain the policy version temporal semantic mapping result; The acquisition of the temporal semantic mapping result of the system version includes: extracting system identifier, version identifier, and time tag information from the system text, and classifying different versions of system text according to the system identifier; constructing a system version temporal semantic mapping model based on the version identifier and time tag information, which is used to establish the correspondence between each system version and its corresponding system semantic information and applicable time interval; determining its applicable time interval according to the time tag corresponding to each system version; mapping the time of occurrence of business behavior to the corresponding applicable time interval, determining the system version corresponding to the business behavior information through the system version temporal semantic mapping model, and generating the system version temporal semantic mapping result.

[0033] In this step, the system identifiers in the policy text mainly originate from the identifier fields in the policy's homepage, title bar, policy number field, document information field, or main body, and at least include the policy name, policy number, issuing department, and policy category. Version identifiers mainly originate from the revision notes, version number, document number, revision batch, replacement notes, or version description fields in the main body, and at least include the version number, revision order, release batch, and the correspondence between old and new versions. Time stamp information mainly originates from the policy's homepage, effective clauses, repealed clauses, trial implementation instructions, revision notes, and document signature, and at least includes the release date, effective date, implementation date, repeal date, replacement date, and transition period. When performing structured parsing of the policy text, the policy title area, document area, clause area, and footnote area are divided into regions. System identifiers, version identifiers, and time stamp information are extracted from the corresponding regions to form the basic data for version identification.

[0034] When categorizing different versions of policy documents according to policy identifiers, the policy name, policy number, and issuing department combination are used as the primary policy identifier. Multiple policy documents belonging to the same primary identifier are grouped into the same policy set. For policy documents with abbreviations, older names, or differences in supplementary descriptions, the policy number, issuing department, and descriptions of applicable matters in the main text are used as auxiliary matching fields for merging, avoiding the splitting of multiple versions of the same policy into different policy sets due to naming differences. After categorization, a version sequence is established within the same policy set. Each version document in the version sequence retains its corresponding version identifier, time stamp information, and policy semantic information.

[0035] When constructing the temporal semantic mapping model of system versions, the main identifier of the system is used as the index unit, the version identifier and time tag information are used as the time positioning unit, and the semantic information of the system is used as the rule content unit to establish the correspondence between system versions, semantic information of the system, and applicable time intervals.

[0036] When determining the applicable time interval based on the time tags corresponding to each version of the system, the extracted release date, effective date, repeal date, replacement date, and transition period time range are standardized, and time information in different forms of expression is uniformly converted into calculable time interval data. For systems that explicitly state "effective from a certain date," that date is used as the starting point of application; for systems that explicitly state "simultaneously repealing a certain original system," the end point of application of the replaced system is truncated to the time before the replacement system takes effect; for systems that state "applicable during the trial period," the start and end dates of the trial period are determined as independent applicable time intervals; for systems with transitional clauses, the transition period is marked as a separate time period, and the boundary relationship between the transition period and the formal application period is preserved in the system version temporal semantic mapping model.

[0037] When mapping the occurrence time of business activities to the corresponding applicable time intervals, time fields related to the determination of the applicability of the system are extracted from the business activity information. The occurrence time of these business activities mainly comes from approval records, business documents, contract attachments, and related trace data, and includes at least the business application time, approval time, contract signing time, acceptance time, payment application time, and payment execution time. After uniformly formatting the above time fields, the target time used for version matching is determined according to the preset time anchor point corresponding to the business type.

[0038] When determining the system version corresponding to business behavior information using the system version temporal semantic mapping model, the applicable time interval into which the target time falls is searched to locate the system version corresponding to that time interval. When multiple time intervals overlap within the same system set, trial and official versions coexist, or supplementary notices and main systems are applied in parallel, the target system version is determined based on the substitution relationship in the version identifier, the priority application instructions in the time tag, and the system hierarchy relationship. For cases where "main system + supplementary notice" are applied in parallel, the association between the main system version and the supplementary notice version is preserved in the system version temporal semantic mapping model.

[0039] When generating the temporal semantic mapping result of the system version, the business item identifier, the time of occurrence of the business behavior, the main identifier of the system, the matched system version identifier, the corresponding applicable time interval, and the associated system semantic information are associated and output to form the system version temporal semantic mapping result. The system version temporal semantic mapping result includes at least the system set identifier corresponding to the business item, the matched system version identifier, the start and end time of the version's application, the time field of the version matching basis, and the corresponding system semantic information identifier.

[0040] Step S3: Parse the rules and regulations, identify the exception triggering conditions and construct a semantic rule graph of the exception clauses, determine whether the business behavior is subject to the exception rules, and obtain the exception rule determination result; The acquisition of the exception rule determination result includes: performing rule semantic parsing on the policy text to identify the conditional content in the policy rules and their corresponding exception triggering conditions; associating and storing the policy rules and their corresponding exception triggering conditions to construct a semantic rule graph of policy exception clauses, wherein different policy rules are associated with their corresponding exception triggering conditions; extracting feature information corresponding to the exception triggering conditions based on business behavior information, matching the feature information with the exception triggering conditions to determine whether the business behavior meets the exception triggering conditions, and determining that the corresponding policy rule is subject to the exception rule when the exception triggering conditions are met, thereby generating the exception rule determination result.

[0041] In this step, the rules and regulations are primarily derived from the main clauses, supplementary clauses, proviso clauses, appendix clauses, temporary notice clauses, and special provisions clauses in the policy text. The conditions refer to the structured constraints in the rules and regulations used to define the scope of application, implementation methods, and approval requirements, including at least business type conditions, monetary conditions, time conditions, subject conditions, material conditions, and status conditions. The exception triggering conditions refer to the special applicable conditions set for general rules in the policy text, including at least emergency procurement conditions, emergency repair and supply conditions, sole source conditions, supplementary approval conditions, supplementary contract signing conditions, temporary authorization conditions, and special exemption conditions. When performing rule semantic parsing on the policy text, the conditional phrases, restrictive phrases, exception descriptive phrases, and applicable object descriptive phrases in the clauses are segmented and identified, and the conditional items, constraint items, and exception items are extracted to form the policy rules and their corresponding exception triggering condition data.

[0042] When identifying the conditions in the rules and regulations and their corresponding exception triggering conditions, the rule statements in the rule text are syntactically segmented and semantically labeled. Expressions such as "under - circumstances", "except -", "but belong to -", "after - approval", and "due to - reason" are identified as exception triggering structures, and the related rule content before and after them is bound to the exception situation content.

[0043] When storing rules and their corresponding exception triggering conditions in association, the system's main identifier, version identifier, and rule identifier are used as index fields. Rule entries and exception entries under the same system version are then associated and encoded to construct a semantic rule graph of system exception clauses. This semantic rule graph of system exception clauses is a structured representation of system rules, exception triggering conditions, and their interrelationships. Nodes include at least system rule nodes, condition nodes, exception triggering condition nodes, and exception execution result nodes. Edge relationships include at least "applicable condition association," "exception triggering relationship," "exception substitution relationship," and "supplementary constraint relationship." For exception clauses with different expressions but similar semantics in different system versions, they are distinguished and stored using system version identifiers and rule identifiers to avoid mixing exception conditions from different versions.

[0044] When extracting feature information corresponding to exception triggering conditions based on business behavior information, the business attributes, time status, approval status, and explanatory content related to the determination of system exceptions are extracted from the business behavior information. The feature information includes at least the business type, purchase amount, application reason, urgency level indicator, whether production is suspended, whether emergency repairs are needed, supplier source type, approval supplementary entry status, contract signing status, and special explanation status. The above feature information mainly comes from the application reason field in business documents, the opinion field in approval records, the explanatory page in contract attachments, supplementary explanation content in image scans, and the business status field. When extracting the above data, structured fields are directly mapped to feature items, and the reason description, emergency item description, and supplementary explanation content in unstructured text content are parsed into matchable feature tags. For example, "Sudden production line shutdown, requiring urgent replacement of core components" is parsed into three business feature tags: "Sudden Failure," "Production Stoppage Risk," and "Emergency Procurement."

[0045] When matching feature information with exception triggering conditions, the feature items in the business behavior information are compared item by item according to the condition item structure recorded in the semantic rule graph of the exception clauses of the system to determine whether the business behavior meets the exception triggering conditions. The matching is not based on a single field comparison, but on the combined verification results of the set of condition items, including necessary condition verification, additional condition verification, and exclusion condition verification. Necessary condition verification is used to determine whether the core facts required for the exception to be established are complete, additional condition verification is used to determine whether supplementary materials or supplementary approvals required by the system have been submitted, and exclusion condition verification is used to exclude exceptions that are explicitly prohibited by the system.

[0046] When determining whether a business behavior meets the exception triggering conditions, the matching results of each exception triggering condition are summarized, and the applicable version range of the exception rule is limited by combining the temporal semantic mapping results of the policy version. For exception clauses that exist only in a specific policy version, they only participate in the exception determination after the corresponding version is matched; for cases where the main policy and supplementary notices together constitute an exception rule, the main policy rule node and the exception triggering condition node in the supplementary notice are linked for matching. When a business behavior meets the exception triggering conditions, the corresponding policy rule is determined to be subject to the exception rule; when a business behavior does not meet the exception triggering conditions, the general policy rule remains in effect.

[0047] When generating exception rule determination results, the business item identifier, system version identifier, candidate system rule identifier, exception trigger condition identifier, business feature matching result, exception application status, and determination basis are associated and output to form the exception rule determination result. The exception rule determination result includes at least the rule set applicable to general rules, the rule set applicable to exception rules, the rule set that does not meet the exception trigger condition, and the condition matching details corresponding to each rule.

[0048] Step S4: Identify the source, analyze the structure, and trace the generation path of business documents, contract attachments, and image scans to construct a multimodal evidence chain. Perform credibility quantification calculation on the evidence nodes in the multimodal evidence chain to obtain the credibility quantification results of the evidence. The acquisition of the evidence credibility quantification result includes: dividing the evidence information obtained from parsing business documents, contract attachments, and image scans into multiple evidence nodes; determining the temporal relationship between evidence nodes based on the generation time information corresponding to each evidence node, determining the source association relationship between evidence nodes based on the source information of each evidence node, determining the content correspondence relationship between evidence nodes based on the content information of each evidence node, and establishing the association relationship between evidence nodes based on the temporal relationship, source association relationship, and content correspondence relationship to construct a multimodal evidence chain; and performing credibility quantification calculation on each evidence node based on the generation time information, source information, and content consistency of each evidence node to obtain the evidence credibility quantification result.

[0049] In this step, business documents primarily originate from application forms, approval forms, acceptance forms, and payment forms generated within the business system, including at least purchase requisitions, budget application forms, acceptance forms, payment application forms, and related business registration forms. Contract attachments primarily originate from archived contract materials and contract approval archived materials, including at least the contract text, supplementary agreements, technical agreements, payment terms, quotation lists, signature pages, and authorization documents. Scanned images primarily originate from scanned and uploaded paper materials, archived image files, and supplementary business record files, including at least scanned invoices, scanned paper approval forms, images of signed receipts, acceptance photos, and scanned supplementary instructions. When identifying the source of the above materials, the file identifier, upload subject identifier, archive path identifier, business association identifier, formation time field, upload time field, and file type field are extracted to form the source data and time data in the evidence information.

[0050] When dividing evidence information obtained from parsing business documents, contract attachments, and image scans into multiple evidence nodes, the node division is performed according to preset segmentation rules. These preset segmentation rules include document-level segmentation rules, page-level segmentation rules, field-level segmentation rules, and clause fragment segmentation rules. For business documents containing fixed field labels, fixed column positions, fixed document number fields, and fixed date fields, evidence nodes are segmented by field group; for contract attachments containing clause numbers, clause text, signature information, and attachment page identifiers, evidence nodes are segmented by clause paragraphs, signature pages, or attachment pages; for scanned image files, evidence nodes are segmented by scanned pages, image areas, and identified key field fragments. Each evidence node is assigned a node identifier and records the node type, source document identifier, business matter identifier, generation time information, source information, and content summary.

[0051] When performing structural analysis on business documents, contract attachments, and scanned images, the corresponding parsing rules are applied according to the document type. For text documents, page partitioning, field location, clause segmentation, and key field extraction are performed; for scanned documents, page segmentation, character recognition, seal area location, signature area location, and document field extraction are performed. The fields obtained after structural analysis include at least the amount field, date field, main body field, contract number field, document number field, signature field, clause field, and explanatory field. For different areas within the same document, their semantic roles are marked based on the presence of field labels, clause numbers, signature graphic features, or explanatory text to distinguish between the main text clause area, the notes and explanations area, the signature area, the document field area, and the header and footer area.

[0052] When tracing the generation path of business documents, contract attachments, and scanned images, trace information related to the evidence formation chain is extracted along the document creation, uploading, circulation, archiving, and replacement processes. This information includes at least the document creation time, file upload time, last modification time, uploading account identifier, archiving location, version number, and replacement record identifier. Generation path tracing is used to determine the formation, upload, and archiving locations of the same evidence node within the business processing chain.

[0053] When determining the temporal sequence of evidence nodes based on their generation time information, the generation time, upload time, and archiving time of each evidence node are uniformly converted into a standard time format. Evidence nodes belonging to the same business processing chain are then sorted according to their business transaction identifiers. The criteria for determining the temporal sequence include: the contract signing node should be earlier than the payment application node; the acceptance form node should be earlier than the final payment node; and the invoice node should meet the invoicing and payment time constraints stipulated in the regulations. For node pairs that do not meet the preset order constraints, the abnormal temporal relationship is recorded and added to the evidence node relationship set.

[0054] When determining the source relationships between evidence nodes based on their source information, the evidence nodes are merged and their source chains are traced according to the uploader's identifier, archive directory, file module, business association number, and version replacement record. The criteria for determining source relationships include: the business order number, contract number, and archive directory corresponding to the two evidence nodes being identical, or the file transfer record pointing to the same business matter. If a node lacks a business association number, lacks an archive path record, or there is no corresponding relationship between the uploader's identifier and the business handling entity, a source information missing marker or a source chain break marker is recorded in that node.

[0055] When determining the content correspondence between evidence nodes based on the content information of each evidence node, the amount, date, subject, matter name, contract number, goods name, approval item description, and clause constraints in the node are standardized and mapped. The basis for determining the content correspondence includes: the field values ​​are the same, the field values ​​meet the preset conversion relationship, the field values ​​are consistent with the clause conditions, or the standardized text expression points to the same business fact.

[0056] When establishing relationships between evidence nodes based on chronological order, source association, and content correspondence, the evidence nodes are treated as chain nodes, and chronological, source, and content associations are treated as chain edge relationships, constructing a multimodal evidence chain. Each node in the chain retains its node type, content summary, formation time, source identifier, and associated edge information; the edge relationships in the chain include at least chronological order edges, source attribution edges, content corroboration edges, and abnormal conflict edges. For a set of nodes with complete business association numbers, satisfying chronological order constraints, and corresponding content fields, their continuous association relationships are preserved in the multimodal evidence chain; for nodes lacking business association numbers, exhibiting chronological anomalies, or having field conflicts, anomaly markers are added to the multimodal evidence chain.

[0057] Based on the generation time information, source information, and content consistency of each evidence node, the credibility quantification calculation for each evidence node is performed separately according to preset evaluation items. The preset evaluation items include at least a time evaluation item, a source evaluation item, and a content consistency evaluation item. The time evaluation item is used to determine whether the node's formation time, upload time, and archiving time are consistent with the business process sequence; the source evaluation item is used to determine whether the node has a clear source file identifier, upload subject identifier, archiving path identifier, and version replacement record; the content consistency evaluation item is used to determine whether the node is consistent with other related nodes in terms of amount, date, subject, matter, and clause content. Each evaluation item outputs a corresponding evaluation result according to preset rules. The time evaluation result includes time sequence consistency, time sequence conflict, and missing time information; the source evaluation result includes a complete source chain, a broken source chain, and missing source identifier; and the content consistency evaluation result includes content consistency, content conflict, and missing key information.

[0058] When performing credibility quantification calculations on each evidence node, the time evaluation results, source evaluation results, and content consistency evaluation results are converted into corresponding quantitative markers, and node credibility quantification results are generated. The quantification method can be implemented by assigning values ​​to each item and then summarizing them. The time evaluation item, source evaluation item, and content consistency evaluation item each correspond to a preset score range. When a node has time sequence conflicts, broken source chains, or content conflicts, the corresponding evaluation item is recorded as an outlier. When a node has missing time information, missing source identifiers, or missing key information, the corresponding evaluation item is recorded as a missing value. When a node satisfies time sequence consistency, a complete source chain, and content consistency, the corresponding evaluation item is recorded as a valid value. After summarizing the results of each evaluation item, the evidence credibility quantification result for that evidence node is obtained.

[0059] Upon obtaining the evidence credibility quantification result, the evidence node identifier, node type, associated business matter identifier, time evaluation result, source evaluation result, content consistency evaluation result, and comprehensive quantification result are correlated and output to form the evidence credibility quantification result. The evidence credibility quantification result includes at least: the time status marker, source status marker, content status marker, comprehensive quantification value, and anomaly cause marker of the evidence node.

[0060] Step S5: Construct an evidence-rule coupling relationship graph based on institutional rule information and evidence credibility quantification results, calculate the rule applicability confidence of candidate institutional rules, and resolve conflicting rules by combining institutional hierarchy relationships to obtain an effective rule set; The construction of the evidence-rule coupling graph includes: extracting rule element information corresponding to each rule based on the rule information, and matching the rule element information with evidence nodes to determine the set of evidence nodes on which each rule depends; establishing the association between the rule and the evidence nodes based on the evidence node set and the corresponding evidence credibility quantification results, and constructing the evidence-rule coupling graph; performing weighted calculations based on the evidence credibility quantification results of each evidence node in the evidence node set corresponding to the rule, and correcting them by combining the content consistency and time sequence relationship between each evidence node, to obtain the rule applicability confidence.

[0061] The acquisition of the effective rule set includes: comparing the consistency of the application results of candidate system rules; when different candidate system rules have inconsistent application results for the same business behavior, they are determined to be conflicting rules; ranking the candidate system rules according to their rule application confidence; calculating the evidence confidence distribution characteristics based on the evidence confidence quantification results of the evidence node set corresponding to each candidate system rule, and correcting the rule application confidence based on the evidence confidence distribution characteristics; and comprehensively comparing the corrected rule application confidence in combination with the system hierarchy, selecting candidate system rules whose rule application confidence meets the preset conditions as effective rules, and forming an effective rule set.

[0062] In this step, the rule information refers to the semantic information of the rules and the results of their application, including at least the rule identifier, the system identifier to which the rule belongs, the version identifier of the rule, the text content of the rule, the conditions for application of the rule, the content of the constraints of the rule, the content of the result of the rule, the exception application mark, and the system level identifier. The rule element information refers to the rule components that can establish a corresponding relationship with the evidence nodes, including at least the subject element, the amount element, the time element, the matter element, the process element, the material element, and the status element.

[0063] When extracting rule element information corresponding to each rule based on the system and rules information, rule splitting is performed on the rule text, dividing it into condition sections, constraint sections, and result sections, and identifying and classifying key fields in each section. Subject fields are mapped to subject elements, amount fields to amount elements, date and effective time fields to time elements, business type and item name fields to item elements, approval node fields to process elements, contract, invoice, and acceptance form material fields to material elements, and business status fields such as approval completion, acceptance completion, and payment execution to status elements. After extraction, a corresponding set of rule elements is generated for each rule, and the position and semantic role of each rule element in the system and rules are recorded.

[0064] When matching rule element information with evidence nodes, the rule element type is used as the matching entry point. The corresponding evidence fields, clause fragments, or status records are retrieved from the evidence node set. The amount element is matched with the application amount, contract amount, payment amount, and invoice amount fields in the evidence node; the time element is matched with fields such as contract signing date, acceptance date, payment application date, and payment execution date; the subject element is matched with the application department, supplier name, and approval subject fields; the process element is matched with the approval node record and approval path comparison results; the material element is matched with contract attachments, scanned invoices, acceptance form images, and explanatory material nodes; and the status element is matched with the approval status, contract status, acceptance status, and payment status fields. During matching, the rule element identifier, the matched evidence node identifier, the matched field name, the matching result status, and the reason for any mismatch are recorded.

[0065] When determining the set of evidence nodes upon which each rule depends, evidence nodes that match the same set of rule elements are merged to form the set of evidence nodes corresponding to that rule. This set of evidence nodes is not arbitrarily aggregated, but determined based on the coverage relationship of rule elements; that is, the elements required by the rule, such as the subject, amount, time, matter, process, materials, and status, should all have corresponding supporting nodes in the set of evidence nodes. Rule elements that do not match are marked as missing.

[0066] Based on the set of evidence nodes and the corresponding quantification results of evidence credibility, when establishing the association between institutional rules and evidence nodes, the institutional rules are treated as rule nodes, and each evidence node in the evidence node set is treated as an evidence node in the relationship graph. Element-matching edges, field-supporting edges, status-supporting edges, and anomaly-marking edges are established between them to construct an evidence-rule coupling relationship graph. In this graph, rule nodes record the rule identifier, institutional identifier, version identifier, and rule application conditions; evidence nodes record the node identifier, node type, and quantification results of evidence credibility; and edge relationships record the matching element type, matching result status, missing marker, conflict marker, and time correction marker. This graph structure explicitly associates the basis for the application of institutional rules with the set of evidence nodes supporting those rules.

[0067] When performing weighted calculations based on the credible quantification results of each evidence node in the evidence node set corresponding to the system rules, corresponding weights are assigned to the evidence nodes according to the rule element type. These weights are pre-configured based on the degree of influence of each rule element on the rule's validity. Core elements that directly determine whether a rule is valid correspond to core weights, while elements that provide supplementary evidence correspond to auxiliary weights. For example, in the amount threshold rule, the amount node and process node are core evidence nodes, while the explanatory material node is an auxiliary evidence node; in the final payment rule, the acceptance completion status node and acceptance form node are core evidence nodes, while the supplementary explanatory node is an auxiliary evidence node. The credible quantification results of each evidence node and their corresponding weights are aggregated to obtain the basic quantified value of the rule.

[0068] When making corrections based on the consistency of content and chronological order among various evidence nodes, a consistency check is performed on the node relationships within the evidence node set corresponding to the rule. If there is a conversion relationship between the amount fields and the conversion result satisfies the clause constraints, a content consistency correction mark is recorded; if there are conflicts between the amount fields, inconsistencies in the main fields, or key clause pages not corresponding to the application content, a content conflict correction mark is recorded. If the contract signing time, acceptance time, payment application time, and payment execution time satisfy the sequential relationship constrained by the system, a chronological order consistency correction mark is recorded; if there are situations such as the payment application preceding the contract signing, or the final payment preceding the acceptance completion, a chronological order conflict correction mark is recorded. Based on the above correction marks, the basic quantitative value of the rule is corrected to obtain the rule applicability confidence level. The rule applicability confidence level is the quantitative result of the system rule being supported by the evidence node set in the current business matter.

[0069] When comparing the consistency of the application results of candidate rules, all candidate rules under the same business matter are categorized according to the business object, processing stage, and rule objective, and the results of the categorized rules are compared. When different candidate rules give mutually exclusive conclusions for the same business behavior, or when inconsistent results such as "payment allowed" and "payment prohibited," or "approval node required" and "no approval node required" are simultaneously derived under the same business state, it is determined that there are conflicting rules. Conflicting rules include conflicts between general rules and exception rules, conflicts between rules of different system versions, and conflicts between rules of different system levels.

[0070] When ranking candidate rules based on their applicability confidence, the ranking is based on the rule applicability confidence, with additional comparison fields including the rule's institutional level, version date, exception application flag, and evidence missing flag. The ranking results are used for subsequent conflict rule resolution and do not directly replace institutional hierarchical relationships. For candidate rules with the same applicability confidence or whose differences fall within a preset threshold range, further screening is performed in the next step, combining evidence credibility distribution characteristics and institutional hierarchical relationships.

[0071] When calculating the evidence credibility distribution characteristics based on the evidence credibility quantification results of the evidence node sets corresponding to each candidate rule, the distribution of valid values, missing values, and outliers in the evidence node sets corresponding to the rule is statistically analyzed, as well as the distribution of various quantification results in core evidence nodes and auxiliary evidence nodes. The evidence credibility distribution characteristics include at least the proportion of valid values ​​in core nodes, the number of outliers in core nodes, the number of missing values ​​in auxiliary nodes, and the number of content conflicts between nodes. When correcting the rule applicability confidence based on the evidence credibility distribution characteristics, if there are outliers or missing values ​​in core evidence nodes, the rule applicability confidence is lowered; if there are missing values ​​in auxiliary evidence nodes but the core evidence nodes are complete, the original rule applicability confidence is retained and a missing material marker is recorded; if there are content conflicts or chronological order conflicts between core evidence nodes, a conflict correction marker is added to the rule applicability confidence.

[0072] When comprehensively comparing the application confidence of the revised rules based on the hierarchical relationship of the systems, priority is determined by invoking the system hierarchy identifier, system version relationship, and exception rule relationship from the semantic information of the systems. The hierarchical relationship of the systems includes at least the following: higher-level systems take precedence over lower-level systems; formal systems take precedence over trial notices; special systems take precedence over general systems within their corresponding scope; and newer versions of systems take precedence over older versions within their applicable time frame. The exception rule relationship is used to limit the exception rule to covering the general rule only when the exception triggering condition is met and the evidence is complete; when the exception triggering condition is not met or the evidence supporting the exception is missing a core element, it still participates in the comprehensive comparison according to the general rule. After the comprehensive comparison, candidate system rules that meet the preset conditions for application confidence and the system hierarchy priority meet the application requirements are selected as valid rules and form a set of valid rules.

[0073] When generating a valid rule set, the business matter identifier, candidate rule set, conflict rule identifier, conflict resolution result, retained rule identifier, removed rule identifier, rule application confidence level and evidence support summary corresponding to each rule are correlated and output to form a valid rule set. The valid rule set includes at least the set of rule nodes that actually participate in the subsequent system application status transition in each processing stage of the current business matter.

[0074] Step S6: Based on the approval path comparison results, the temporal semantic mapping results of the system version, the exception rule determination results, and the set of valid rules, construct a system application status transition diagram to determine the application path of the target system; The determination of the target system application path includes: combining the approval completion status, system version, and rule application results to generate multiple candidate system application states; arranging the candidate system application states chronologically according to the time of business behavior occurrence, and establishing state transition relationships between candidate system application states based on exception rule judgment results, approval path comparison results, effective rule set, and evidence credibility quantification results to obtain a system application state transition diagram; performing constraint consistency verification on each state transition relationship in the system application state transition diagram, and deleting state transition relationships that do not meet the constraint conditions; performing path traversal on the retained state transition relationships to obtain multiple candidate paths, and comparing the rule application confidence and evidence credibility quantification results corresponding to the candidate paths to determine the target system application path corresponding to the business behavior information.

[0075] In this step, the approval path comparison results mainly provide the node matching status of business items in the approval process, implicit missing node markers, and sequence anomaly markers; the system version temporal semantic mapping results mainly provide the system version identifiers and applicable time intervals corresponding to different processing stages of business items; the exception rule determination results mainly provide the general rule application status, the exception rule application status, and unclosed exception markers; the valid rule set mainly provides the system rule nodes retained after evidence support and conflict resolution. These results serve as the basic inputs for state generation and state transition calculations.

[0076] When combining approval completion status, system version, and rule application results to generate multiple candidate system application states, the approval status, version status, and rule status of the business matter at each processing stage are used as the state constituent units for combination. The approval completion status includes at least the node executed status, node missing status, node sequence abnormal status, and supplementary review completed status; the system version includes at least the system version identifier and its applicable time interval identifier that are matched in the current stage of the business; the rule application result includes at least the general rule application status, the exceptional rule application status, the rule conflict resolution result, and the rule not satisfied flag. After combining the above state constituent units, a status identifier is generated for each combination result, forming a candidate system application state. Each candidate system application state records the business matter identifier, processing stage identifier, approval status set, system version identifier, rule result set, and its corresponding evidence summary identifier.

[0077] When chronologically arranging the application status of candidate systems according to the occurrence time of business activities, time anchors related to state transitions are extracted from the business activity information. These time anchors include at least the business application time, approval initiation time, contract signing time, acceptance time, payment application time, and payment execution time. The application statuses of candidate systems are then sorted according to their corresponding time anchors, and states within the same time period are further sorted according to the business process sequence to form a state time series. For cases where the same business matter corresponds to different system versions at different stages, the corresponding version status is retained in the state time series, ensuring that the system application status transition diagram reflects the actual application process of the system version as the business stage changes.

[0078] Based on the exception rule determination results, approval path comparison results, valid rule set, and evidence credibility quantification results, when establishing the state transition relationship between the applicable states of candidate systems, the transition conditions from one state to the next are structurally represented as approval conditions, version conditions, rule conditions, and evidence conditions. Approval conditions are used to limit the reachability relationship between approval states before and after the transition. For example, an approval node that should have been completed in the previous state has been executed in the next state, or a missing node in the previous state has been converted to supplementary review and completed in the next state. Version conditions are used to limit whether the system version switch during the state transition conforms to the applicable time interval. Rule conditions are used to limit whether the rule results between the previous and next states are consistent with the valid rule set. Evidence conditions are used to limit whether the evidence nodes supporting the state transition have time anomalies, source anomalies, or content conflicts. When the above conditions are met, a state transition edge is established between the corresponding applicable states of the candidate system.

[0079] When obtaining the system application status transition diagram, candidate system application statuses are used as graph nodes, and status transition relationships are used as graph edges, forming the system application status transition diagram corresponding to the business item. Each node in the diagram records the status identifier, stage identifier, system version identifier, approval status, rule application status, and evidence summary; each edge in the diagram records the transition start point, transition end point, transition condition set, constraint verification result, and exception marker. For status transitions triggered by exception rules, exception trigger markers and exception closure markers are added to the graph edges; for status corrections caused by approval supplementation or supplementary review, supplementation correction markers are added to the graph edges.

[0080] When performing constraint consistency verification on each state transition relationship in the system application state transition diagram, preset constraint rules are invoked to verify each edge of the diagram. These preset constraint rules include at least time constraints, version constraints, approval constraints, rule constraints, and evidence constraints. Time constraints are used to verify whether the time sequence of state transitions is consistent with the business process; version constraints are used to verify whether the system version changes before and after the transition fall within the corresponding applicable time interval; approval constraints are used to verify whether the transition to the subsequent payment or execution state is allowed when a mandatory approval node is missing; rule constraints are used to verify whether a state transition contrary to a prohibitive rule still occurs when one exists; and evidence constraints are used to verify whether there are missing or outlier values ​​in the core evidence nodes supporting the transition. For state transition relationships that do not meet the constraints, the corresponding constraint conflict marker is recorded and the relationship is deleted from the system application state transition diagram.

[0081] After deleting state transition relationships that do not meet the constraints, path traversal is performed on the remaining state transition relationships to obtain multiple candidate paths. During path traversal, the initial state of the business matter is used as the starting point, and the current processing result state or final execution state of the business matter is used as the ending point. The traversal is performed on the reachable paths in the graph, and the state sequence, rule sequence, system version sequence, and anomaly marker sequence on each path are recorded. Paths containing deleted transition edges are not included in the generation of the candidate path set; for paths with unclosed exception states, unresolved approval missing states, or conflicting core evidence states, their anomaly markers are retained in the candidate paths.

[0082] When comparing the rule applicability confidence and evidence credibility quantification results corresponding to candidate paths, a summary calculation is performed on the rule nodes and evidence nodes on each candidate path. The comparison includes at least the summary of rule applicability confidence for valid rule nodes within the path, the summary of credibility quantification results for core evidence nodes within the path, the statistics of the number of anomaly markers within the path, and the path integrity verification results. Path integrity is used to determine whether the candidate path covers the necessary processing stages of a business matter from its initial state to its final state; the number of anomaly markers is used to record the number of approval gaps, time conflicts, source gaps, content conflicts, and unclosed exceptions present in the path. After comprehensively comparing the above results, the target system applicability path corresponding to the business behavior information is determined.

[0083] The target system application path is not simply a business trajectory selected in chronological order, but rather a state transition result that meets the conditions for system version application, approval constraints, rule constraints, and evidence support. After the target system application path is determined, the system version, rule results, evidence summary, and anomaly markers corresponding to each state node and each transition edge in the path are retained.

[0084] Step S7: Construct an audit reasoning path relationship diagram along the target system application path and generate an audit interpretation path. Based on the audit interpretation path and the results of evidence credibility quantification, conduct a comprehensive analysis and output the compliance risk level and the corresponding audit interpretation results.

[0085] The comprehensive analysis includes: along the target system application path, associating and expanding the various system versions, rule application results, and corresponding evidence nodes in the path according to the state transition sequence, constructing an audit reasoning path relationship diagram, and generating an audit interpretation path; based on the evidence credibility quantification results of the evidence nodes corresponding to the rule application results in the audit interpretation path, performing a consistency assessment on the rule application results, and correcting or eliminating rule application results that do not meet the consistency conditions; based on the corrected rule application results, performing a compliance assessment on business behavior information, and outputting the compliance risk level and corresponding audit interpretation results.

[0086] In this step, along the target system application path, when associating and expanding the various system versions, rule application results, and corresponding evidence nodes in the path according to the state transition sequence, each state node in the path is used as the starting point for expansion. The system version identifier, the valid rule identifier, the rule application result identifier, and the set of evidence nodes that the corresponding state node are retrieved. Association edges are established for system version nodes, rule nodes, and evidence nodes under the same state node. Specifically, a version application edge is established between system version nodes and rule nodes, an evidence support edge is established between rule nodes and evidence nodes, and a state determination edge is established between rule nodes and state nodes. State transition edges are retained between adjacent state nodes. This results in a reasoning relationship chain arranged in chronological order.

[0087] When constructing the audit reasoning path diagram, the nodes in the diagram should include at least the following: system version node, rule node, evidence node, status node, and anomaly marker node. System version nodes record the system identifier, version identifier, and applicable time interval; rule nodes record the rule identifier, rule content, applicable results, and rule application confidence level; evidence nodes record the node identifier, node type, source status marker, time status marker, content status marker, and comprehensive quantitative value; status nodes record the processing stage identifier, approval status, and migration results; anomaly marker nodes record anomalies such as missing approvals, time conflicts, broken source chains, content conflicts, and unclosed exceptions. The edges in the diagram should include at least the version application edge, rule support edge, evidence corroboration edge, status determination edge, and anomaly correction edge, thus unifying the rule application basis, evidence support relationship, and status change process within the same reasoning structure.

[0088] When generating an audit interpretation path, a reachable reasoning chain consisting of system version nodes, rule nodes, evidence nodes, and status nodes is extracted from the audit reasoning path relationship graph, and an interpretation sequence is generated according to the state transition order. The audit interpretation path includes at least the system version applicable to each processing stage of the business matter, the rules hit at each stage, the evidence support for the rules, the status change results derived from the rule application results, and anomaly markers in the path.

[0089] Based on the quantified results of the evidence nodes corresponding to the applicable results of each rule in the audit interpretation path, when conducting a consistency assessment of the applicable results of each rule, an evidence support verification is performed on the set of evidence nodes upon which each rule node depends. The consistency assessment includes at least a check of the consistency between the rule results and the evidence nodes in the amount, time, subject, matter, and status fields, as well as a check of the sufficiency of support between the rule results and the comprehensive quantified values ​​of the evidence nodes. If the core evidence nodes upon which a rule node depends have outliers, missing values, or content conflict markers, an evidence support inconsistency marker is recorded in that rule node; if all core evidence nodes upon which a rule node depends meet the requirements of chronological order, source chain, and content correspondence, an evidence support consistency marker is recorded.

[0090] When correcting or removing rule application results that do not meet consistency conditions, processing is performed according to the inconsistency type corresponding to the rule node. For inconsistencies caused by missing supporting evidence nodes, a missing material marker is added while retaining the rule node; for inconsistencies caused by missing core evidence nodes, core field conflicts, or key time sequence conflicts, the application result of the corresponding rule node is corrected; for rule nodes that still do not meet the evidence support conditions after correction, they are removed from the audit interpretation path, and the reason for removal is recorded. The correction content includes at least rule application status adjustment, rule application confidence update, and anomaly marker addition. Through this process, subsequent compliance assessments are performed only based on rule results that have completed consistency verification.

[0091] When conducting compliance assessments of business activity information based on the revised rule application results, the rule nodes retained in the audit interpretation path are categorized into prohibitive rules, mandatory rules, and conditional rules, and verified item by item in conjunction with the approval status, execution status, and exception status in the business activity information. Prohibitive rules are used to determine whether a business activity has resulted in a process explicitly prohibited by regulations; mandatory rules are used to determine whether necessary steps, required materials, and necessary statuses are complete; and conditional rules are used to determine whether the amount, proportion, time interval, and subject scope meet the regulations' limitations. After summarizing the verification results, a compliance assessment result corresponding to the business item is formed.

[0092] When outputting compliance risk levels, business matters are classified into risk levels based on the revised rule application results, the number of anomaly markers in the path, and the violation of core rules. The compliance risk levels include at least four states: Pass, Alert, Warning, and Blocked. Pass indicates that the business matter does not violate any rules and the key evidence nodes are complete; Alert indicates that the business matter does not violate any prohibitive rules, but there are missing supporting materials, pending supplementary information, or general anomaly markers; Warning indicates that the business matter has unmet mandatory rules, unclosed exception rules, or conflicting core evidence nodes; Blocked indicates that the business matter has violated prohibitive rules, missing and unresolved mandatory approval nodes, conflicting key time sequences, or direct conflict between payment execution and institutional rules. The judgment criteria for each level are retained in the output results.

[0093] When outputting the corresponding audit interpretation results, the business matter identifier, the target system application path identifier, the audit interpretation path identifier, the set of revised rule results, the compliance risk level, and the risk triggering reason will be output in a correlated manner. The audit interpretation results shall include at least: the system version that was hit, the effective rules that were hit, the rules that were revised or removed and the reasons therefor, the set of evidence nodes supporting the conclusion, the approval conditions that were not met, the exception conditions that were not closed, the evidence nodes that conflicted, and the basis for the final risk level determination.

[0094] Example 2: A multimodal intelligent audit system for complex documents, comprising the following modules: Path comparison module: Acquires and parses policy texts, approval records, business documents, contract attachments and image scans in a structured manner to obtain policy semantic information, business behavior information and evidence information. After constructing a standard approval path, it is compared with the approval records to obtain the approval path comparison results. Version mapping module: Identifies policy versions and time tags, constructs a policy version temporal semantic mapping model, matches the corresponding policy version according to the time of business behavior occurrence, and obtains the policy version temporal semantic mapping results; Exception determination module: Parses the rules and regulations, identifies the exception triggering conditions and constructs a semantic rule graph of the exception clauses, determines whether the business behavior is subject to the exception rules, and obtains the exception rule determination results; Trust measurement module: performs source identification, structural analysis and generation path tracing on business documents, contract attachments and image scans, constructs a multimodal evidence chain, performs trust measurement calculation on evidence nodes in the multimodal evidence chain, and obtains the trust measurement results of the evidence. Effective rule module: Based on the information of institutional rules and the quantification results of evidence credibility, an evidence-rule coupling relationship graph is constructed, the rule applicability confidence of candidate institutional rules is calculated, and conflicting rules are resolved by combining institutional hierarchy relationships to obtain a set of effective rules; The system application module constructs a system application status transition diagram based on the approval path comparison results, system version temporal semantic mapping results, exception rule determination results, and effective rule set to determine the application path of the target system. Comprehensive Audit Module: Constructs an audit reasoning path relationship diagram along the target system application path and generates an audit interpretation path. Based on the audit interpretation path and the quantification results of evidence credibility, it performs a comprehensive analysis and outputs the compliance risk level and the corresponding audit interpretation results.

[0095] based on Figure 1 The structural diagram shown illustrates the specific implementation of the present invention as follows: each module works collaboratively according to a preset data flow direction to achieve a complete process of multimodal internal control compliance intelligent auditing.

[0096] The path comparison module receives input document data, including policy texts, approval records, business documents, contract attachments, and scanned images. Internally, the module first performs structured parsing of these documents, extracting semantic information from the policy texts, business behavior information from the approval records, and evidentiary information from the remaining documents. Then, the module constructs a standard approval path based on the approval process rules in the policy texts and extracts the actual approval path from the approval records. The module performs node-level comparison between the actual approval path and the standard approval path, identifying approval nodes present in the standard path but missing in the actual path as implicit missing nodes. Finally, the module generates and outputs the approval path comparison result.

[0097] The version mapping module receives policy texts and business behavior information. It identifies policy version identifiers and corresponding timestamps from the policy texts and categorizes different versions of the policy texts according to the identifiers. The module constructs a policy version temporal semantic mapping model, which establishes the correspondence between each policy version, its semantic information, and its applicable time interval. Based on the occurrence time of the business behavior, the module maps it to the corresponding applicable time interval, determines the policy version that matches the current business behavior information through the policy version temporal semantic mapping model, generates the policy version temporal semantic mapping result, and outputs it.

[0098] The exception determination module receives policy text and business behavior information. The module performs rule semantic parsing on the policy text, identifying the conditions within the policy rules and their corresponding exception triggering conditions. The module associates and stores the policy rules with their corresponding exception triggering conditions, constructing a semantic rule graph of policy exception clauses, where different policy rules are associated with their respective exception triggering conditions. The module extracts feature information corresponding to the exception triggering conditions based on the business behavior information, matches this feature information with the exception triggering conditions, determines whether the business behavior meets the exception triggering conditions, and if the conditions are met, determines that the corresponding policy rule applies the exception rule, generating and outputting the exception rule determination result.

[0099] The credibility measurement module receives business documents, contract attachments, and scanned images. After parsing these documents, the module divides them into multiple evidence nodes. Based on the generation time information of each evidence node, the module determines the temporal relationship between nodes; based on the source information, it determines the source association; and based on the content information, it determines the content correspondence. Based on these three relationships, the module establishes the association between evidence nodes, constructing a multimodal evidence chain. Based on the generation time information, source information, and content consistency of each evidence node, the module performs credibility measurement calculations on each evidence node, obtaining and outputting the credibility measurement results.

[0100] The effective rules module receives institutional rule information and the evidence credibility quantification results output by the credibility quantification module. Based on the institutional rule information, the module extracts rule element information corresponding to each institutional rule and matches this information with evidence nodes to determine the set of evidence nodes upon which each institutional rule depends. Based on the evidence node set and the corresponding evidence credibility quantification results, the module establishes the association between institutional rules and evidence nodes, constructing an evidence-rule coupling graph. The module performs weighted calculations on the credibility quantification results of each node in the evidence node set corresponding to the institutional rule, and corrects these calculations based on content consistency and temporal order among the evidence nodes to obtain the rule applicability confidence level. The module compares the applicability results of different candidate institutional rules for consistency. When conflicting rules exist, the module ranks the candidate institutional rules according to their rule applicability confidence level, corrects the confidence level based on the evidence credibility distribution characteristics, and selects candidate institutional rules that meet preset conditions as effective rules, forming a set of effective rules and outputting it.

[0101] The system application module receives the approval path comparison results from the path comparison module, the system version temporal semantic mapping results from the version mapping module, the exception rule determination results from the exception determination module, and the set of valid rules from the valid rules module. The module combines the approval completion status, system version, and rule application results to generate multiple candidate system application states. The module arranges the candidate system application states chronologically according to the time of business behavior occurrence and establishes state transition relationships between these states based on the input results, resulting in a system application state transition graph. The module performs constraint consistency checks on each state transition relationship in the state transition graph, deletes state transition relationships that do not meet the constraints, and performs path traversal on the remaining state transition relationships to obtain multiple candidate paths. The module compares the rule application confidence and evidence credibility quantification results corresponding to the candidate paths to determine and output the target system application path corresponding to the business behavior information.

[0102] The comprehensive audit module receives the target system application path output by the system application module and the evidence credibility measurement results output by the credibility measurement module. Following the target system application path, the module correlates and expands the various system versions, rule application results, and corresponding evidence nodes in the path according to their state transition order, constructing an audit reasoning path relationship graph and generating an audit interpretation path. Based on the evidence credibility measurement results of the evidence nodes corresponding to the rule application results in the audit interpretation path, the module performs a consistency assessment on the rule application results, correcting or eliminating rule application results that do not meet the consistency conditions. Based on the corrected rule application results, the module performs a compliance assessment on business behavior information, ultimately outputting the compliance risk level and the corresponding audit interpretation results. These modules execute sequentially, with the output of the previous module serving as the input for the subsequent module, forming a complete audit processing chain and achieving automated intelligent auditing of multimodal internal control compliance in complex document environments.

[0103] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multimodal intelligent auditing method for complex documents, characterized in that, include: The system obtains and structures the policy text, approval records, business documents, contract attachments and image scans to obtain policy semantic information, business behavior information and evidence information. After constructing a standard approval path, it is compared with the approval records to obtain the approval path comparison results. Identify the policy version and time tag, construct a policy version temporal semantic mapping model, and match the corresponding policy version according to the time of business behavior to obtain the policy version temporal semantic mapping result; Analyze the rules and regulations, identify the exception triggering conditions and construct a semantic rule graph of the exception clauses, determine whether the business behavior is subject to the exception rules, and obtain the exception rule determination results; The source identification, structural analysis and generation path tracing of business documents, contract attachments and image scans are performed to construct a multimodal evidence chain. The credibility quantification calculation of the evidence nodes in the multimodal evidence chain is performed to obtain the credibility quantification result of the evidence. Based on the information of institutional rules and the quantification results of evidence credibility, an evidence-rule coupling relationship graph is constructed. The rule applicability confidence of candidate institutional rules is calculated, and conflicting rules are resolved by combining the institutional hierarchy relationship to obtain an effective rule set. Based on the approval path comparison results, the temporal semantic mapping results of the system version, the exception rule determination results, and the set of effective rules, a system application status transition diagram is constructed to determine the application path of the target system. An audit reasoning path relationship diagram is constructed along the target system application path, and an audit interpretation path is generated. Based on the audit interpretation path and the results of evidence credibility quantification, a comprehensive analysis is conducted to output the compliance risk level and the corresponding audit interpretation results.

2. The multimodal internal control compliance intelligent auditing method for complex documents according to claim 1, characterized in that, The acquisition of the approval path comparison results includes: extracting each approval node and the sequential relationship between nodes according to the approval process rules in the policy document, and constructing a standard approval path; extracting the actual approval nodes and their execution order from the approval records to form the actual approval path; performing node matching and sequence consistency comparison between the actual approval path corresponding to the approval record and the standard approval path, identifying approval nodes that exist in the standard approval path but do not appear in the actual approval path, and using the identified approval nodes as implicit missing nodes to generate the approval path comparison results.

3. The multimodal internal control compliance intelligent auditing method for complex documents according to claim 1, characterized in that, The acquisition of the temporal semantic mapping result of the system version includes: extracting system identifier, version identifier, and time tag information from the system text, and classifying different versions of system text according to the system identifier; constructing a system version temporal semantic mapping model based on the version identifier and time tag information, which is used to establish the correspondence between each system version and its corresponding system semantic information and applicable time interval; determining its applicable time interval according to the time tag corresponding to each system version; mapping the time of occurrence of business behavior to the corresponding applicable time interval, determining the system version corresponding to the business behavior information through the system version temporal semantic mapping model, and generating the system version temporal semantic mapping result.

4. The multimodal internal control compliance intelligent auditing method for complex documents according to claim 1, characterized in that, The acquisition of the exception rule determination result includes: performing rule semantic parsing on the policy text to identify the conditional content in the policy rules and their corresponding exception triggering conditions; associating and storing the policy rules and their corresponding exception triggering conditions to construct a semantic rule graph of policy exception clauses, wherein different policy rules are associated with their corresponding exception triggering conditions; extracting feature information corresponding to the exception triggering conditions based on business behavior information, matching the feature information with the exception triggering conditions to determine whether the business behavior meets the exception triggering conditions, and determining that the corresponding policy rule is subject to the exception rule when the exception triggering conditions are met, thereby generating the exception rule determination result.

5. The multimodal internal control compliance intelligent auditing method for complex documents according to claim 1, characterized in that, The acquisition of the evidence credibility quantification result includes: dividing the evidence information obtained from parsing business documents, contract attachments, and image scans into multiple evidence nodes; determining the temporal relationship between evidence nodes based on the generation time information corresponding to each evidence node, determining the source association relationship between evidence nodes based on the source information of each evidence node, determining the content correspondence relationship between evidence nodes based on the content information of each evidence node, and establishing the association relationship between evidence nodes based on the temporal relationship, source association relationship, and content correspondence relationship to construct a multimodal evidence chain; and performing credibility quantification calculation on each evidence node based on the generation time information, source information, and content consistency of each evidence node to obtain the evidence credibility quantification result.

6. The multimodal internal control compliance intelligent auditing method for complex documents according to claim 1, characterized in that, The construction of the evidence-rule coupling graph includes: extracting rule element information corresponding to each rule based on the rule information, and matching the rule element information with evidence nodes to determine the set of evidence nodes on which each rule depends; establishing the association between the rule and the evidence nodes based on the evidence node set and the corresponding evidence credibility quantification results, and constructing the evidence-rule coupling graph; performing weighted calculations based on the evidence credibility quantification results of each evidence node in the evidence node set corresponding to the rule, and correcting them by combining the content consistency and time sequence relationship between each evidence node, to obtain the rule applicability confidence.

7. The multimodal internal control compliance intelligent auditing method for complex documents according to claim 1, characterized in that, The acquisition of the effective rule set includes: comparing the consistency of the application results of candidate system rules; when different candidate system rules have inconsistent application results for the same business behavior, they are determined to be conflicting rules; ranking the candidate system rules according to their rule application confidence; calculating the evidence confidence distribution characteristics based on the evidence confidence quantification results of the evidence node set corresponding to each candidate system rule, and correcting the rule application confidence based on the evidence confidence distribution characteristics; and comprehensively comparing the corrected rule application confidence in combination with the system hierarchy, selecting candidate system rules whose rule application confidence meets the preset conditions as effective rules, and forming an effective rule set.

8. The multimodal internal control compliance intelligent auditing method for complex documents according to claim 1, characterized in that, The determination of the target system application path includes: combining the approval completion status, system version, and rule application results to generate multiple candidate system application states; arranging the candidate system application states chronologically according to the time of business behavior occurrence, and establishing state transition relationships between candidate system application states based on exception rule judgment results, approval path comparison results, effective rule set, and evidence credibility quantification results to obtain a system application state transition diagram; performing constraint consistency verification on each state transition relationship in the system application state transition diagram, and deleting state transition relationships that do not meet the constraint conditions; performing path traversal on the retained state transition relationships to obtain multiple candidate paths, and comparing the rule application confidence and evidence credibility quantification results corresponding to the candidate paths to determine the target system application path corresponding to the business behavior information.

9. The multimodal intelligent auditing method for complex documents according to claim 1, characterized in that, The comprehensive analysis includes: along the target system application path, associating and expanding the various system versions, rule application results, and corresponding evidence nodes in the path according to the state transition sequence, constructing an audit reasoning path relationship diagram, and generating an audit interpretation path; based on the evidence credibility quantification results of the evidence nodes corresponding to the rule application results in the audit interpretation path, performing a consistency assessment on the rule application results, and correcting or eliminating rule application results that do not meet the consistency conditions; based on the corrected rule application results, performing a compliance assessment on business behavior information, and outputting the compliance risk level and corresponding audit interpretation results.

10. A multimodal intelligent audit system for internal control and compliance in response to complex documents, characterized in that: A method for implementing a multimodal internal control compliance intelligent auditing method for complex documents as described in any one of claims 1 to 9 includes: Path comparison module: Acquires and parses policy texts, approval records, business documents, contract attachments and image scans in a structured manner to obtain policy semantic information, business behavior information and evidence information. After constructing a standard approval path, it is compared with the approval records to obtain the approval path comparison results. Version mapping module: Identifies policy versions and time tags, constructs a policy version temporal semantic mapping model, matches the corresponding policy version according to the time of business behavior occurrence, and obtains the policy version temporal semantic mapping results; Exception determination module: Parses the rules and regulations, identifies the exception triggering conditions and constructs a semantic rule graph of the exception clauses, determines whether the business behavior is subject to the exception rules, and obtains the exception rule determination results; Trust measurement module: performs source identification, structural analysis and generation path tracing on business documents, contract attachments and image scans, constructs a multimodal evidence chain, performs trust measurement calculation on evidence nodes in the multimodal evidence chain, and obtains the trust measurement results of the evidence. Effective rule module: Based on the information of institutional rules and the quantification results of evidence credibility, an evidence-rule coupling relationship graph is constructed, the rule applicability confidence of candidate institutional rules is calculated, and conflicting rules are resolved by combining institutional hierarchy relationships to obtain a set of effective rules; The system application module constructs a system application status transition diagram based on the approval path comparison results, system version temporal semantic mapping results, exception rule determination results, and effective rule set to determine the application path of the target system. Comprehensive Audit Module: Constructs an audit reasoning path relationship diagram along the target system application path and generates an audit interpretation path. Based on the audit interpretation path and the quantification results of evidence credibility, it performs a comprehensive analysis and outputs the compliance risk level and the corresponding audit interpretation results.

Images