Document auditing method and device, equipment, storage medium and program product

Abstract

The invention provides a document auditing method and device, equipment, a storage medium and a program product, and relates to the technical field of financial science and technology. The method comprises the following steps: splitting an input document into a plurality of modules and marking content elements in each module; according to the document type and the field information, calling an associated auditing rule from a knowledge base and generating an auditing normal form containing a plurality of auditing entries; and scheduling a plurality of agents to execute the auditing items based on the auditing normal form and the content elements in the modules. According to the method, the document auditing efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of financial technology, and in particular to a document auditing method and device, equipment, storage medium and program product. BACKGROUND

[0002] In the process of bank business, document auditing is a core and complex task. Usually in the process of document submission, the document needs to be preliminarily audited to ensure that the document meets the requirements defined for format and content.

[0003] However, the traditional document auditing process highly depends on human experience, and needs to check clauses one by one, such as format specification, integrity, consistency, etc., which is time-consuming and laborious and is easy to miss or misjudge due to subjective judgment. In the case of errors in manual auditing, the submitter is notified to modify and resubmit, and the auditor is notified to re-audit, and even multiple rounds of modification and auditing. Therefore, the efficiency of document auditing in the related art is low. SUMMARY

[0004] The present application provides a document auditing method, device, equipment, storage medium and program product to provide a document auditing method to improve the accuracy of extracting video frames.

[0005] In a first aspect, the present application provides a document auditing method, comprising:

[0006] Splitting the input document into multiple modules and marking the content elements in each module;

[0007] According to the document type and field information, calling the associated auditing rules from the knowledge base and generating an auditing schema containing multiple auditing items;

[0008] Scheduling multiple agents to execute the auditing items based on the auditing schema and the content elements in each module.

[0009] Optionally, the knowledge base is a knowledge graph library, and the calling of the associated auditing rules from the knowledge base according to the document type and field information and the generation of the auditing schema containing multiple auditing items comprises:

[0010] Retrieving the auditing rules matching the document type and the field information from the knowledge graph library;

[0011] Generating an auditing schema containing multiple auditing items based on the auditing rules.

[0012] Optionally, the scheduling of multiple agents to execute the auditing items based on the auditing schema and the content elements in each module further comprises:

[0013] According to the dependency relationship between the auditing items, the execution order of the multiple agents is sorted.

[0014] The agent is invoked sequentially to execute the corresponding audit items.

[0015] Optionally, sorting the execution order of the multiple agents according to the dependencies between the audit items includes:

[0016] Retrieve the item dependency graph from the knowledge base;

[0017] The execution order of the agents is generated based on the item dependency graph.

[0018] Optionally, after scheduling multiple agents to execute review entries based on the review paradigm and the content elements in each module, the process further includes:

[0019] Perform causal chain analysis on the review results of each intelligent agent;

[0020] The causal chain analysis results are correlated with the terms in the knowledge base to generate an audit report.

[0021] Optionally, the method of scheduling multiple intelligent agents to execute review entries based on the review paradigm and content elements in each module further includes:

[0022] Real-time monitoring of resource utilization of intelligent agents;

[0023] The scheduling order of intelligent agents is dynamically adjusted based on monitoring results.

[0024] Optionally, after scheduling multiple agents to execute review entries based on the review paradigm and the content elements in each module, the process further includes:

[0025] The conflict resolution mechanism handles the conflict results of multiple agents on the same review item;

[0026] The conflict resolution mechanism specifically includes at least one of the following:

[0027] A voting mechanism based on rule weights;

[0028] The expert knowledge base is invoked for adjudication.

[0029] Optionally, after scheduling multiple agents to execute review entries based on the review paradigm and the content elements in each module, the process further includes:

[0030] The confidence level of the audit results of each agent is evaluated by ensemble learning methods.

[0031] When the confidence level is lower than the preset threshold, the expert knowledge base is invoked for secondary verification.

[0032] Secondly, this application provides a document review device, comprising:

[0033] Split modules are used to divide the input document into multiple modules and mark the content elements in each module;

[0034] The generation module is used to call the associated review rules from the knowledge base and generate a review paradigm containing multiple review items based on the document type and domain information.

[0035] The execution module is used to schedule multiple agents to execute audit entries based on the audit paradigm and the content elements in each module.

[0036] Thirdly, this application provides an electronic device, including: a processor, and a memory communicatively connected to the processor;

[0037] The memory stores computer-executed instructions;

[0038] The processor executes computer execution instructions stored in the memory to implement the method as described in any of the first aspects.

[0039] Fourthly, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any of the first aspects.

[0040] Fifthly, this application provides a computer program product including a computer program that, when executed by a processor, is described as described in any of the first aspects.

[0041] The document review method, apparatus, device, storage medium, and program products provided in this application involve splitting the input document into multiple modules and labeling the content elements in each module; based on the document type and domain information, retrieving associated review rules from a knowledge base and generating a review paradigm containing multiple review items; and scheduling multiple intelligent agents to execute the review items based on the review paradigm and the content elements in each module. The method of this application decomposes the document into structured modules, dynamically generates a review paradigm adapted to the document type and domain information, and achieves automated review through the collaborative execution of review items by multiple intelligent agents, thereby improving the efficiency of document review. Attached Figure Description

[0042] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0043] Figure 1 Flowchart of the document review method provided in this application embodiment Figure 1 ;

[0044] Figure 2Flowchart of the document review method provided in this application embodiment Figure 2 ;

[0045] Figure 3 Flowchart of the document review method provided in this application embodiment Figure 3 ;

[0046] Figure 4 A schematic diagram of the structure of a document review device provided in an embodiment of this application;

[0047] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0048] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0049] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0050] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of the relevant data all comply with the relevant laws, regulations, and standards of the relevant countries and regions, have taken necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation access points for users to choose to authorize or refuse.

[0051] Furthermore, the technical solution involved in this application, which involves big data analysis of user information (including but not limited to personal biometrics, identity data, consumption data, asset data, electronic terminal operation data, etc.) and the use of artificial intelligence technology for automated decision-making, and makes decisions that have a significant impact on personal rights based on the results of automated decision-making, provides users with corresponding operation entry points for users to choose to agree to or reject the results of automated decision-making; if the user chooses to reject, the process will proceed to the expert decision-making process.

[0052] It should be noted that the document review method, apparatus, equipment, storage medium and program products provided in this application can be used in the field of fintech, or in any field other than fintech. The application field of the document review method, apparatus, equipment, storage medium and program products in this application is not limited.

[0053] In related technologies, the traditional document review process relies heavily on human experience, requiring clause-by-clause verification. If errors are found during manual review, the submitter is notified to make corrections and resubmit, followed by further review by the reviewers, sometimes involving multiple rounds of revisions and reviews. Therefore, the document review efficiency in related technologies is relatively low.

[0054] In view of this, this application proposes a document review method that decomposes documents into structured modules, dynamically generates review paradigms adapted to document types and domain information, and achieves automated review by having multiple intelligent agents collaboratively execute review items. This method improves the efficiency of document review by achieving automated document review through the division of labor among intelligent agents.

[0055] This application applies to document processing devices, apparatuses, or systems with document review functions. The document review function is implemented on the document processing device, apparatus, or system. For example, a document processing system reviews documents uploaded to the document processing device. Another example is a third-party application calling a document review method through an Application Programming Interface (API) to review documents and output the review results. The following explanation uses a document processing system as the executing entity.

[0056] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0057] Figure 1 Flowchart of the document review method provided in this application embodiment Figure 1 .like Figure 1 As shown, document review methods may include:

[0058] S101. Divide the input document into multiple modules and mark the content elements in each module.

[0059] A module refers to multiple logical modules obtained after splitting the content of the input document.

[0060] Content elements refer to the types of content within each module and are the smallest auditable unit in a document. Examples of content elements include paragraph numbering, figure references, figure numbers, and tables.

[0061] Optionally, the document processing system can split the input document into multiple logical modules and mark the content elements in each module using preset structured rules. For example, the input document can be split using page breaks, headings, paragraphs, etc. The preset structured rules can be determined based on the document type of the input document.

[0062] For example, the document processing system pre-defines various document types and their mapping relationships with structured rules. In this approach, when uploading an input document, the user selects the corresponding document type. The document processing system then searches for the corresponding structured rule from the mapping relationship between document types and structured rules, and uses this structured rule to break down the input document into multiple logical modules. Further, based on the pre-defined structured rules, the system parses the format of the content elements in each module, and marks each content element, such as titles, paragraphs, and charts, within each module.

[0063] For example, a document processing system can determine the document type based on the method of uploading the input document, or it can determine the document type by parsing the content and / or format of the input document, and find the corresponding structured rule from the mapping relationship between document type and structured rule. The input document is then split into multiple logical modules and the content elements in each module are marked by the structured rule.

[0064] For example, the document processing system can target a document type of financial product introduction, and the corresponding structured rule is to split the input document according to page breaks and / or header content, and determine the modules of product overview, market background, and detailed plan based on the content. Based on the rules for the "detailed plan" module in the structured rules, the system can identify the paragraph number, figure references, tables and other content elements of each paragraph.

[0065] S102. Based on the document type and domain information, retrieve the associated audit rules from the knowledge base and generate an audit paradigm containing multiple audit items.

[0066] Document type is a classification based on the purpose of the document. For example, resumes, contracts, financial statements, audit reports, etc.

[0067] Domain information refers to the classification of the domains covered by the document content. For example, the domain information of an input document of type contract could be a loan agreement, a service agreement, etc. Exemplarily, domain information can be obtained from user input or determined based on content parsing.

[0068] A knowledge base refers to a structured database of review items. A knowledge base can also store structured domain review rules, clauses, and cases, as well as domain knowledge such as review rules for legal clauses and technical specifications.

[0069] The knowledge base supports retrieving related rules based on document type and domain information. For example, based on the document type "introduction to financial products" and the domain information "fixed income products", the relevant review rule "support test of return calculation data for product expected rate of return" can be retrieved.

[0070] Audit rules are predefined audit standards for specific domains and / or document types. They are audit standards abstracted from business logic, laws and regulations, industry standards, or internal specifications. Examples include audits of symmetrical liability for breach of contract, audits of benchmark interest rate requirements, and audits of the completeness of algorithm flowcharts.

[0071] Audit items are the executable checks performed on audit rules within the input document. Each audit item is a quantifiable judgment task, and its function is to transform abstract rules into actionable audit steps. For example, "Check whether the citations in the main text are included in the bibliography."

[0072] An audit paradigm refers to a predefined template that includes multiple audit items, rule base reference paths, and result aggregation logic.

[0073] For example, a document processing system can retrieve corresponding review rules from a knowledge base based on pre-stored mapping relationships, i.e., associated review rules, according to document type and domain information. Taking a document type of "financial product introduction" and domain information of "fixed-income products" as an example, the document processing system, based on the input document's document type of "financial product introduction" and domain information of "fixed-income products," retrieves associated review rules from the knowledge base and generates a review paradigm containing multiple review items. A review item could be, for example, "Check whether the detailed plan contains content that does not conform to industry standards."

[0074] S103. Schedule multiple agents to execute audit entries based on the audit paradigm and content elements in each module.

[0075] An intelligent agent is a software module with autonomous decision-making capabilities that executes specific auditing tasks through preset rules or algorithms.

[0076] For example, the document processing system can assign review tasks to corresponding agents based on the item dependencies in the review paradigm template and the agent's functionality. For instance, the format agent verifies number continuity, the logic agent detects reference relationships, and the quality agent compares internal database flag conflicts. The processing results of each agent are integrated into a review report through the result aggregation module, completing the document review.

[0077] For example, each agent executes audit entries based on the audit paradigm and the content elements in each module. The content elements in each module can be audited according to the audit paradigm. For instance, if the format agent verifies the continuity of table numbers in the module, it identifies the content elements marked with table numbers for each module and performs a continuity audit on them.

[0078] The document review method of this application splits the input document into multiple modules and marks the content elements in each module; based on the document type and domain information, it calls the associated review rules from the knowledge base and generates a review paradigm containing multiple review items; and it schedules multiple intelligent agents to execute the review items based on the review paradigm and the content elements in each module. This method decomposes the document into structured modules, dynamically generates a review paradigm adapted to the document type and domain information, and achieves automated review by having multiple intelligent agents collaboratively execute the review items, thus improving the efficiency of document review.

[0079] The following describes how the document review method of this application, based on document type and domain information, calls associated review rules from the knowledge base and generates multiple review items.

[0080] Optionally, the knowledge base is a knowledge graph library. The document processing system can retrieve review rules that match the document type and domain information from the knowledge graph library, and generate a review paradigm containing multiple review items based on the review rules.

[0081] A knowledge graph database is a graph-structured database that stores domain knowledge and supports semantic-based retrieval.

[0082] Optionally, when generating the review paradigm, the document processing system first retrieves review rules from the knowledge graph database that match the input document type and domain information. For example, when the document type is "introduction to financial products" and the domain information is "funds," the document processing system uses semantic retrieval to search for review rules related to "introduction to financial products" and "funds" in the knowledge graph database.

[0083] Multiple review items are generated based on review rules. For example, "Check if the detailed plan contains content that does not conform to industry standards," "Are the Chinese and English annotations consistent?" and "Test the support of the revenue calculation data for the product's expected rate of return." A semantic retrieval mechanism ensures that the review rules are adapted to the needs of different fields. For example, in the fixed-income product field, the document processing system retrieves the review rule "Test the support of the revenue calculation data for the product," and extracts the following review items from the knowledge graph: "Confirm whether the reinvestment rate, default rate, and prepayment rate are clearly disclosed in the materials," and "Check whether the calculation data includes the source of the basic data, the calculation time, the argumentation process, and a comparison of the calculation results under different market scenarios." The document processing system combines multiple review items into a review paradigm based on a preset construction method for constructing review paradigms.

[0084] Figure 2 Flowchart of the document review method provided in this application embodiment Figure 2 .like Figure 2 As shown, document review methods may include:

[0085] S201. Divide the input document into multiple modules and mark the content elements in each module.

[0086] For example, the document processing system can find the corresponding structured rule from the mapping relationship between document types and structured rules based on the document type selected by the user, and then split the input document into multiple logical modules using the structured rule. Further, it matches the content in each module according to the preset structured rules, parses the format of the content elements in each module, and marks the content elements such as titles, paragraphs, figure numbers, and table numbers in each module one by one.

[0087] S202. Retrieve review rules from the knowledge graph database that match the document type and domain information.

[0088] For example, a document processing system can retrieve corresponding review rules from a knowledge graph database based on document type and domain information.

[0089] S203. Generate an audit paradigm containing multiple audit items based on audit rules.

[0090] Generating multiple review items based on review rules can be achieved in two ways. One approach is to extract the review items corresponding to the review rules from a knowledge base. Another approach is to generate them using a pre-trained model. In this method, the model is pre-trained to learn a review paradigm for outputting multiple review items based on the review rules. Optionally, the model can output multiple review items based on the review rules and construct a review paradigm based on these multiple review items.

[0091] For example, a review paradigm can be constructed based on multiple review items. This can be done using a pre-defined unified construction method or using a pre-trained model. This model can learn in advance the ability to output a review paradigm based on the input document type, domain information, and review items.

[0092] S204. Schedule multiple agents to execute audit entries based on the audit paradigm and content elements in each module.

[0093] For example, the document processing system can assign review tasks to corresponding agents based on the item dependencies in the review paradigm template and the agent's functionality. Each agent executes the review items based on the review paradigm and the content elements in each module.

[0094] It should be noted that, in Figure 2 The various processing steps (S201-S204) shown in the embodiments can be implemented with reference to the same or similar steps in the above embodiments. Figure 2 The processing steps shown in the embodiments do not constitute a specific limitation on the document review process. In other embodiments of this application, the document review process may include more than Figure 2 The embodiments may involve more or fewer steps. For example, a document review process may include... Figure 2 Some steps in the embodiments, or, Figure 2 Some steps in the embodiments can be replaced by steps with the same function, or Figure 3 Some steps in the embodiments can be broken down into multiple steps, etc.

[0095] The method in this application embodiment, through retrieval of a knowledge graph database, leverages the graph structure data of the knowledge graph database to support semantic-based association rule retrieval, enabling the generated review entries to accurately cover the review requirements demanded by document type and domain information. This method of dynamically determining review rules based on document type and domain information improves the flexibility and accuracy of the review process.

[0096] Optionally, the document processing system can schedule multiple agents to execute review items based on the review paradigm and the content elements in each module. It can also sort the execution order of multiple agents according to the dependencies between review items; and call agents in sequence to execute the corresponding review items.

[0097] Dependencies are the logical order of audit items; for example, format validation must precede logical validation.

[0098] The execution order is the priority order in which agents perform tasks. For example, the format agent is executed first, followed by the logic agent.

[0099] For example, the document processing system can analyze the dependencies between audit items and generate an execution order. For instance, if "verify number continuity" needs to be executed before "detect reference relationship", then the agent executing the "verify number continuity" audit item will be executed first, followed by the agent executing the "detect reference relationship" audit item, and the agents will be called in order to execute the corresponding audit items.

[0100] For example, taking a formatting agent and a logic agent as examples, the formatting agent verifies the document structure, and the logic agent detects citation relationships. For instance, the document processing system first calls the formatting agent to verify the continuity of figure numbers, and then calls the logic agent to detect whether the figure numbers exist in the text.

[0101] The method in this application embodiment ensures the logical rigor of the review process by dynamically sorting the execution order of intelligent agents, avoiding review omissions caused by incorrect task order. It optimizes the review process by prioritizing tasks, thereby improving collaborative efficiency.

[0102] Optionally, the document processing system can schedule multiple agents to execute audit items based on the audit paradigm and the content elements in each module, and can also call the item dependency graph from the knowledge base; and generate the execution order of the agents based on the item dependency graph.

[0103] An item dependency graph is a graph structure that describes the logical dependencies between review items. For example, an item dependency graph might include a structure like "format validation → logical validation → quality validation," representing the execution order in which review items for format validation, logical validation, and quality validation must be satisfied. Another example is a graph structure like "compound structure description completeness validation → experimental data repeatability validation → guidance clause comparison," representing the sequential execution of the agent performing compound structure description completeness validation, experimental data repeatability validation, and guidance clause comparison.

[0104] For example, the generated audit paradigm, which retrieves associated audit rules from the knowledge base based on document type and domain information, includes audit entries for format validation and audit entries for quality validation. If the entry dependency graph includes "format validation - logic validation - quality validation", then the execution order of the generated execution bodies indicates that the agent performing the audit entries for format validation executes first, and the execution order of the execution bodies is that the agent performing the audit entries for format validation executes last.

[0105] For example, the knowledge base calls a predefined entry dependency graph that includes "verifying number continuity → detecting reference relationships → comparing conflict schemes." Based on this entry dependency graph, an agent execution order is generated. The agents are then called to execute the audit entries based on the audit paradigm and the content elements in each module. For example, if the audit entry in the audit paradigm is "checking whether the figure numbers referenced in the text are correct," then the module in the text is determined based on the content elements in multiple modules, and the audit is performed based on the content elements labeled with figure numbers.

[0106] The method in this application generates a corresponding intelligent agent scheduling order through an item dependency graph, avoiding omissions in the review process due to incorrect task order and improving the accuracy of document review.

[0107] Optionally, after the document processing system schedules multiple agents to execute audit items based on the audit paradigm and the content elements in each module, it can also perform causal chain analysis on the audit results of each agent; and associate the causal chain analysis results with the terms in the knowledge base to generate an audit report.

[0108] Causal chain analysis refers to analyzing the causal relationship path between audit conclusions and document elements.

[0109] The terms and conditions are the original text of the specific basis for the review, such as legal terms and institutional terms.

[0110] For example, a document processing system can generate a causal chain path for the review conclusion by inputting the dependencies between various modules of a document, and then associate it with relevant legal provisions related to content retrieval or semantic retrieval in related technologies to generate a review report.

[0111] Figure 3 Flowchart of the document review method provided in this application embodiment Figure 3 .like Figure 3 As shown, document review methods may include:

[0112] S301. Divide the input document into multiple modules and mark the content elements in each module.

[0113] S302. Retrieve review rules from the knowledge graph database that match the document type and domain information.

[0114] S303. Generate an audit paradigm containing multiple audit items based on audit rules.

[0115] S304. Sort the execution order of multiple agents according to the dependencies between the audit items.

[0116] S305. Call the intelligent agent in sequence to execute the corresponding audit items.

[0117] S306. Perform causal chain analysis on the audit results of each intelligent agent.

[0118] For example, taking a contract as the input document, the document processing system can, based on the correspondence between the service terms module and the performance details module of the input document, associate the audit result "The 7×24-hour customer service response promised in the service terms is not reflected in the specific response process in the performance details" with the audit result "The performance details do not stipulate the responsibilities and measures for late response," thus establishing a causal relationship between "The 7×24-hour customer service response promised in the service terms is not reflected in the specific response process in the performance details" and "The performance details do not stipulate the responsibilities and measures for late response." The above audit results can be obtained from the same intelligent agent or different intelligent agents.

[0119] For example, causal relationships can be indicated by symbols such as "→". For instance, performing a causal chain analysis on the two audit results above can be represented as: "The 7×24-hour customer service response promised in the terms of service is not reflected in the specific response process in the performance details" → "The performance details do not stipulate the responsibilities and measures for late response", which indicates that there is a causal relationship between the two.

[0120] S307. Associate the causal chain analysis results with the clauses in the knowledge base to generate an audit report.

[0121] For example, based on keyword matching or semantic matching methods in related technologies, matching clauses are retrieved from the knowledge base, associated with the clauses in the knowledge base, and an audit report is generated. The associated clause in the knowledge base could be, for example, Article m of the "Implementation Rules of X".

[0122] It should be noted that, in Figure 3 The various processing steps (S301-S307) shown in the embodiments can be implemented with reference to the specific implementation of the same or similar steps in the above embodiments. Figure 3 The processing steps shown in the embodiments do not constitute a specific limitation on the document review process. In other embodiments of this application, the document review process may include more than Figure 3 The embodiments may involve more or fewer steps. For example, a document review process may include... Figure 3 Some steps in the embodiments, or, Figure 4 Some steps in the embodiments can be replaced by steps with the same function, or Figure 4 Some steps in the embodiments can be broken down into multiple steps, etc.

[0123] The method in this application embodiment improves the interpretability of the audit report by performing causal chain analysis on the audit results of each intelligent agent, associating the causal chain analysis results with the terms in the knowledge base.

[0124] Optionally, the document processing system can schedule multiple agents to execute review items based on the review paradigm and content elements in each module, and can also monitor the resource utilization of agents in real time; and dynamically adjust the scheduling order of agents based on the monitoring results.

[0125] Optionally, the document processing system can use a reinforcement learning framework to monitor the resource utilization of the agent in real time. The reinforcement learning framework optimizes the task allocation strategy through Markov decision processes.

[0126] For example, a document processing system can use a reinforcement learning framework to model the agent scheduling problem as a Markov decision process, monitor resource utilization in real time, and dynamically adjust scheduling strategies. For instance, when a logical agent experiences delays due to processing complex reference relationships, the document processing system can allocate some tasks to idle quality agents.

[0127] The method in this application embodiment monitors the resource utilization of intelligent agents in real time and dynamically adjusts the scheduling order of intelligent agents based on the monitoring results, thereby balancing the load on intelligent agents. In high-concurrency scenarios, it avoids review delays caused by the overload of a single intelligent agent. Through real-time resource monitoring and strategy adjustment, the stability and efficiency of the review process are improved.

[0128] Optionally, after the document processing system schedules multiple agents to execute audit items based on the audit paradigm and the content elements in each module, it can also handle the conflict results of multiple agents on the same audit item through a conflict resolution mechanism.

[0129] The conflict resolution mechanism specifically includes at least one of the following: a voting mechanism based on rule weights; and invoking an expert knowledge base for adjudication.

[0130] A conflict resolution mechanism is a mechanism used to resolve conflicts arising from multiple agents' opinions on the same review item.

[0131] An expert knowledge base refers to a database that stores the experience of experts in a particular field and is used to adjudicate conflicts.

[0132] For example, if the formatting agent, quality agent, and logic agent disagree on the compliance judgment of the same paragraph, the document processing system can make a decision based on a voting mechanism using rule weights, resulting in a conflicting outcome. For instance, if the formatting agent's review result is non-compliant, then the non-compliant review result is weighted according to its weight; if the quality agent's review result is compliant, then the compliant review result is weighted according to its weight; if the logic agent's review result is compliant, then the compliant review result is weighted according to its weight. The weighted sums of the two review results are compared, and the higher review result is considered a conflicting outcome.

[0133] For example, if the formatting agent and the logic agent disagree on the compliance judgment of the same paragraph, the document processing system can invoke the adjudication rules in the expert knowledge base to make a decision. By comparing keywords in the expert knowledge base, the system determines the most matching adjudication rule and obtains the conflict result based on the matching adjudication rule.

[0134] The method in this application embodiment, through a conflict resolution mechanism, ensures that the conflict results of multiple intelligent agents are reasonably adjudicated, avoids review delays or erroneous conclusions caused by conflicts, and improves the accuracy of review.

[0135] Optionally, after scheduling multiple agents to execute review items based on the review paradigm and content elements in each module, the document processing system can also evaluate the confidence level of the review results of each agent through an ensemble learning method; when the confidence level is lower than a preset threshold, the expert knowledge base is invoked for secondary verification.

[0136] Ensemble learning methods refer to calculating reliability scores by comparing the audit results of multiple agents with a historical case library.

[0137] For example, the review results of multiple agents are compared with a historical case library, and a confidence score is calculated using an ensemble learning method. If the score is below a threshold, secondary verification is triggered. For instance, if the confidence score of a logical agent in detecting a reference relationship is <70%, a secondary verification is performed by calling an expert knowledge base.

[0138] The method in this application identifies low-confidence results through dynamic confidence assessment, avoiding erroneous conclusions caused by model uncertainty and improving the accuracy of audit results.

[0139] Optionally, the document processing system can adapt the knowledge graph rules of existing domain information to the domain based on the transfer learning framework; and dynamically generate review items adapted to the new domain by aligning the semantic spaces of different domains through comparative learning.

[0140] For example, based on domain adaptation technology and utilizing a multi-task learning framework, the existing knowledge graph audit entry "Compliance and Consistency Check of Fund Type Terminology" for the domain "Fund" is migrated to the audit entry "Compliance and Consistency Check of Trust Type Terminology" for the new domain "Trust". By aligning the semantic spaces of different domains through contrastive learning, the system ensures that the generated audit entries accurately match domain requirements. For instance, when the input document type is "Trust", the system extracts the audit entry "Compliance and Consistency Check of Trust Type Terminology".

[0141] Optionally, the document processing system can employ a multi-task learning framework, training the model using data with labeled domain information, while simultaneously introducing a small number of manually labeled samples from new domain information for fine-tuning. The model aligns the semantic spaces of different domains through comparative learning, dynamically generating review items adapted to the new domain.

[0142] The method in this application embodiment adapts existing domain information knowledge graph rules to a domain using a transfer learning framework, quickly constructs new domain review rules, maintains the new review rule base through a domain adaptive model, and can quickly generate an adapted review paradigm.

[0143] Optionally, during the knowledge base maintenance phase, natural language processing and knowledge extraction technologies can be used to automatically parse newly added regulatory texts, extract key clauses, and update the knowledge graph. This ensures the timeliness and accuracy of the review rules, adapting to dynamic changes in regulations and business needs. Through an incremental update mechanism, the document processing system can respond to new regulations in real time.

[0144] The document review method of this application splits the input document into multiple modules and marks the content elements in each module; based on the document type and domain information, it calls the associated review rules from the knowledge base and generates a review paradigm containing multiple review items; and it schedules multiple intelligent agents to execute the review items based on the review paradigm and the content elements in each module. This method decomposes the document into structured modules, dynamically generates a review paradigm adapted to the document type and domain information, and achieves automated review by having multiple intelligent agents collaboratively execute the review items, thus improving the efficiency of document review.

[0145] Figure 5 This is a schematic diagram of a document review device provided in an embodiment of this application. Figure 5 As shown, the document review device 400 may include, for example, a mutation frame detection module 401, a generation module 402, and an execution module 403. Optionally, it may also include an analysis module.

[0146] Splitting module 401 is used to split the input document into multiple modules and mark the content elements in each module;

[0147] The generation module 402 is used to call the associated review rules from the knowledge base and generate a review paradigm containing multiple review items based on the document type and domain information.

[0148] The execution module 403 is used to schedule multiple agents to execute audit entries based on the audit paradigm and the content elements in each module.

[0149] One possible implementation is that the knowledge base is a knowledge graph library, and the generation module 402 is specifically used for:

[0150] Retrieve review rules from the knowledge graph database that match document type and domain information;

[0151] Generate an audit paradigm containing multiple audit items based on audit rules.

[0152] One possible implementation is that execution module 403 is specifically used for:

[0153] The execution order of multiple agents is sorted according to the dependencies between the audit items;

[0154] The agent is invoked sequentially to execute the corresponding audit items.

[0155] One possible implementation is that execution module 403 is specifically used for:

[0156] Retrieve the item dependency graph from the knowledge base;

[0157] The execution order of agents is generated based on the item dependency graph.

[0158] One possible implementation involves scheduling multiple agents to execute review items based on the review paradigm and content elements in each module. The analysis module is then specifically used for:

[0159] Perform causal chain analysis on the review results of each intelligent agent;

[0160] The causal chain analysis results are correlated with the terms in the knowledge base to generate an audit report.

[0161] One possible implementation involves scheduling multiple agents to execute review entries based on a review paradigm and content elements in each module, and also includes:

[0162] Real-time monitoring of resource utilization of intelligent agents;

[0163] The scheduling order of intelligent agents is dynamically adjusted based on monitoring results.

[0164] One possible implementation involves scheduling multiple agents to execute review items based on the review paradigm and content elements in each module. The analysis module is then specifically used for:

[0165] The conflict resolution mechanism handles the conflict results of multiple agents on the same review item;

[0166] Conflict resolution mechanisms specifically include at least one of the following:

[0167] A voting mechanism based on rule weights;

[0168] The expert knowledge base is invoked for adjudication.

[0169] One possible implementation involves scheduling multiple agents to execute review items based on the review paradigm and content elements in each module. The analysis module is then specifically used for:

[0170] The confidence level of the audit results of each agent is evaluated by ensemble learning methods.

[0171] When the confidence level is lower than the preset threshold, the expert knowledge base is invoked for secondary verification.

[0172] It should be understood that the apparatus of this application embodiment can implement any of the above embodiments, and this application will not elaborate on them.

[0173] ​ This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. ​ As shown, the electronic device may include at least one processor 501 and a memory 502.

[0174] The memory 502 is used to store programs. Specifically, the program may include program code, which includes computer operation instructions.

[0175] Memory 502 may include high-speed RAM memory, and may also include non-volatile memory.

[0176] The processor 501 is used to execute computer execution instructions stored in the memory 502 to implement the method of the foregoing method embodiments. The processor 501 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application.

[0177] Optionally, the electronic device may also include a communication interface 503. In specific implementations, if the communication interface 503, the memory 502, and the processor 501 are implemented independently, the communication interface 503, the memory 502, and the processor 501 can be interconnected via a bus to complete communication between them.

[0178] Optionally, in a specific implementation, if the communication interface 503, memory 502, and processor 501 are integrated on a single chip, then the communication interface 503, memory 502, and processor 501 can communicate through an internal interface.

[0179] This application also provides a computer-readable storage medium, which may include various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), and a random access memory (RAM). Specifically, the computer-readable storage medium stores program instructions, which are used to implement the actions of the above-described method implementation.

[0180] This application also provides a computer program product, which includes a computer program that, when executed by a processor, performs the actions of the above-described method implementation.

[0181] Furthermore, unless otherwise specified, the functional units / modules in the various embodiments of this application can be integrated into one unit / module, or each unit / module can exist physically separately, or two or more units / modules can be integrated together. The integrated units / modules described above can be implemented in hardware or as software program modules.

[0182] When an integrated unit / module is implemented in hardware, the hardware can be digital circuits, analog circuits, etc. The physical implementation of the hardware structure includes, but is not limited to, transistors. Unless otherwise specified, the processor can be any suitable hardware processor, such as a CPU, GPU, FPGA, DSP, and ASIC, etc. Unless otherwise specified, the storage unit can be any suitable storage medium, such as Resistive Random Access Memory (RRAM), Dynamic Random Access Memory (DRAM), Static Random Access Memory (SRAM), Enhanced Dynamic Random Access Memory (EDRAM), High-Bandwidth Memory (HBM), Hybrid Memory Cube (HMC), etc.

[0183] If the integrated unit / module is implemented as a software program module and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), and portable hard drives.

[0184] In the above embodiments, the descriptions of each embodiment have their own emphasis. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0185] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM and RAM.

Claims

1. A document review method, characterized in that, include: The input document is split into multiple modules and the content elements in each module are marked; Based on document type and domain information, retrieve the associated review rules from the knowledge base and generate a review paradigm containing multiple review items; Multiple agents are scheduled to execute audit entries based on the audit paradigm and the content elements in each module.

2. The method according to claim 1, characterized in that, The knowledge base is a knowledge graph library. The step of retrieving associated review rules from the knowledge base and generating a review paradigm containing multiple review items based on document type and domain information includes: Retrieve review rules from the knowledge graph database that match the document type and the domain information; An audit paradigm containing multiple audit items is generated based on the audit rules.

3. The method according to claim 1, characterized in that, The scheduling of multiple intelligent agents to execute audit entries based on the audit paradigm and content elements in each module includes: The execution order of the multiple agents is sorted according to the dependencies between the audit items; The agent is invoked sequentially to execute the corresponding audit items.

4. The method according to claim 3, characterized in that, The step of sorting the execution order of the multiple agents according to the dependencies between the audit items includes: Retrieve the item dependency graph from the knowledge base; The execution order of the agents is generated based on the item dependency graph.

5. The method according to any one of claims 1-4, characterized in that, After scheduling multiple agents to execute audit entries based on the audit paradigm and content elements in each module, the process further includes: Perform causal chain analysis on the review results of each intelligent agent; The causal chain analysis results are correlated with the terms in the knowledge base to generate an audit report.

6. The method according to any one of claims 1-4, characterized in that, The method further includes: Real-time monitoring of resource utilization of intelligent agents; Adjust the scheduling order of intelligent agents based on monitoring results.

7. The method according to any one of claims 1-4, characterized in that, After scheduling multiple agents to execute audit entries based on the audit paradigm and content elements in each module, the process further includes: The conflict resolution mechanism handles the conflict results of multiple agents on the same review item; The conflict resolution mechanism specifically includes at least one of the following: A voting mechanism based on rule weights; The expert knowledge base is invoked for adjudication.

8. The method according to any one of claims 1-4, characterized in that, After scheduling multiple agents to execute audit entries based on the audit paradigm and content elements in each module, the process further includes: The confidence level of the audit results of each agent is evaluated by ensemble learning methods. When the confidence level is lower than the preset threshold, the expert knowledge base is invoked for secondary verification.

9. A document review device, characterized in that, include: Split modules are used to divide the input document into multiple modules and mark the content elements in each module; The generation module is used to call the associated review rules from the knowledge base and generate a review paradigm containing multiple review items based on the document type and domain information. The execution module is used to schedule multiple agents to execute audit entries based on the audit paradigm and the content elements in each module.

10. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method as described in any one of claims 1 to 8.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1 to 8.

12. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method of any one of claims 1 to 8.

Images