Document auditing method, device and equipment and storage medium

Abstract

The invention discloses a document auditing method and device, equipment and a storage medium. When the method provided by the embodiment of the invention is executed, firstly, a candidate document, an auditing scene of the candidate document and an official rule file matched with the auditing scene can be obtained, and image information in the candidate document and image information in the official rule file are converted into editable text formats to obtain a target document and a standard rule file. Performing rule extraction on the standard rule file to obtain a rule base, and performing rule traversal on auditing rules in the rule base based on the auditing scene to obtain a plurality of rule applicable scenes and target auditing rules matched with the auditing scene. And checking the target document based on a plurality of target checking rules to obtain a real-time checking suggestion of the document to be checked. According to the method and the device, the auditing time is shortened, the efficiency is improved, misjudgment caused by fatigue or subjective deviation is reduced, and the consistency and the accuracy of the auditing result are ensured.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, specifically to a document review method, apparatus, device, and storage medium. Background Technology

[0002] In the daily operations of many industries, including government affairs, finance, and corporate offices, document review is a crucial step in ensuring business compliance and mitigating risks. This process heavily relies on official rules and regulations, such as policies, regulations, business standards, and contract templates. These documents serve as the core basis for the review process, clearly defining the boundaries and requirements for business operations. However, most industries currently still primarily employ a manual review model, where reviewers meticulously verify contracts, policy application materials, business forms, and other documents against complex official rules and regulations.

[0003] However, this manual review model has significant shortcomings: on the one hand, it is extremely inefficient. When faced with massive amounts of documents, reviewers need to spend a lot of time going through the rules and checking the details. The review of a single document of hundreds of pages may take several days, which is difficult to meet the needs of rapid business progress. On the other hand, the accuracy is poor. Manual review is easily affected by factors such as fatigue and subjective judgment bias. It may not only miss key clauses in the rule documents, but may also lead to misjudgment due to misunderstanding of the rules. Especially when the rules are complex and the applicable scenarios are diverse, it is even more difficult to ensure the consistency and accuracy of the review results, thus creating compliance risks for the industry's business operations.

[0004] Therefore, how to improve the efficiency and accuracy of document review is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of the above problems, this application provides a document review method, apparatus, device and storage medium, which can improve the efficiency and accuracy of document review.

[0006] The embodiments of this application disclose the following technical solutions:

[0007] A document review method, the method comprising:

[0008] Obtain the pending documents and the review scenarios for the pending documents, and obtain the official rule files that are compatible with the review scenarios; the content of the pending documents and the official rule files are in either image format or portable document format PDF format;

[0009] The image information in the candidate document and the official rule file are converted into editable text format to obtain the target document and the standard rule file, respectively.

[0010] The standard rule file is used to extract rules to obtain a rule base; the rule base includes multiple audit rules marked with quality scores; the audit rules include rule identifier ID, rule content, rule application scenario and rule judgment criteria;

[0011] Based on the audit scenario, the audit rules in the rule base are traversed to obtain multiple target audit rules; the applicable scenarios of the target audit rules match the audit scenario.

[0012] The target document is reviewed based on the multiple target review rules to obtain real-time review comments on the candidate document.

[0013] In one possible implementation, the step of extracting rules from the standard rule file to obtain a rule base includes:

[0014] Each text paragraph in the standard rule file is analyzed to obtain multiple usable rules;

[0015] Calculate the quality score of each available rule and mark the quality score of each available rule on the corresponding available rule; the quality score of the rule is calculated based on the rule executability score, the rule specificity score, and the rule scenario applicability score. The rule executability score is obtained by quantitatively evaluating the degree to which the rule meets the executability requirements. The rule specificity score is obtained by quantitatively evaluating the degree to which the rule meets the specificity requirements. The rule scenario applicability score is obtained by quantitatively evaluating the degree to which the rule meets the applicable scenario requirements.

[0016] All available rules labeled with quality scores are organized and stored to obtain the rule base.

[0017] In one possible implementation, paragraph analysis is performed on each text paragraph in the standard rule file to obtain multiple available rules, including:

[0018] Based on the first screening requirement, each text paragraph in the standard rule file is initially screened to obtain multiple screened paragraphs; the first screening requirement includes screening out text paragraphs that contain at least one target keyword and conform to at least one target sentence structure.

[0019] The fit score of each selected paragraph is calculated, and the selected paragraphs with a fit score greater than the fit threshold are retained to obtain candidate paragraphs; the fit score is obtained by quantitative evaluation based on the number of target keywords and the number of target sentence structures included in the rule.

[0020] Based on preset rule extraction requirements, rule extraction is performed on each candidate paragraph to obtain multiple candidate rules; the rule extraction requirements are used to indicate the rule ID, rule content, rule application scenario, and rule judgment criteria in the extracted text;

[0021] Based on the second screening criteria, each candidate rule is subjected to quality screening to obtain the plurality of usable rules; the second screening criteria include screening out rules that simultaneously meet the executability requirement, the specificity requirement, and the applicable scenario requirement;

[0022] Specifically, if a rule has a clear judgment criterion, then the rule is determined to meet the enforceability requirement; if a rule has no vague expression, then the rule is determined to meet the specificity requirement; if a rule includes a clear applicable scenario, then the rule is determined to meet the applicable scenario requirement.

[0023] In one possible implementation, the method further includes:

[0024] Based on the aforementioned review scenario, the review rules in the rule base are retrieved to obtain at least one secondary review rule; the number of secondary review rules is less than the number of target review rules.

[0025] The target document is reviewed based on all secondary review rules to obtain supplementary review comments for the document under review;

[0026] Calculate the similarity between the supplementary review comments and the real-time review comments;

[0027] If the similarity is greater than or equal to the similarity threshold, then the supplementary review comments and the real-time review comments are deduplicated.

[0028] If the similarity is less than the similarity threshold, the accuracy of the supplementary review opinions and the real-time review opinions is evaluated based on the quality score and number of rules of each secondary review rule, as well as the quality score and number of rules of each target review rule. The review opinion with higher accuracy after the accuracy evaluation is taken as the final standardized review result.

[0029] Among them, the more rules there are and the higher the average quality score of the rules, the more accurate the review comments are.

[0030] In one possible implementation, the method further includes:

[0031] Obtain the metadata corresponding to the target document and the standard rule file respectively, and annotate them in the target document and the standard rule file; the metadata includes the name of the document or file, the upload time, and the business scenario.

[0032] In one possible implementation, the method further includes:

[0033] Data cleaning is performed on the target document and the standard rule file respectively;

[0034] The data cleaning process includes: removing meaningless characters and redundant information from the document, and correcting text garbled characters and sentence breaks.

[0035] A document review device, the device comprising:

[0036] The acquisition unit is used to acquire the candidate document and the review scenario of the candidate document, and to acquire the official rule file adapted to the review scenario; the content of the candidate document and the official rule file are in either image format or portable document format PDF format;

[0037] The conversion unit is used to convert the image information in the candidate document and the official rule file into an editable text format, respectively, to obtain the target document and the standard rule file;

[0038] The rule base construction unit is used to extract rules from the standard rule file to obtain a rule base; the rule base includes multiple audit rules marked with quality scores; the audit rules include rule ID, rule content, rule application scenario and rule judgment criteria;

[0039] The rule traversal unit is used to traverse the review rules in the rule base based on the review scenario to obtain multiple target review rules; the applicable scenario of the target review rules matches the review scenario.

[0040] The first review unit is used to review the target document based on the multiple target review rules and obtain real-time review opinions on the candidate document.

[0041] In one possible implementation, the rule base construction unit specifically includes:

[0042] The paragraph analysis unit is used to perform paragraph analysis on each text paragraph in the standard rule file to obtain multiple usable rules;

[0043] The scoring and labeling unit is used to calculate the quality score of each available rule and label the quality score of each available rule on the corresponding available rule. The quality score of a rule is calculated based on the rule executability score, the rule specificity score, and the rule scenario applicability score. The rule executability score is obtained by quantitatively evaluating the degree to which the rule meets the executability requirements. The rule specificity score is obtained by quantitatively evaluating the degree to which the rule meets the specificity requirements. The rule scenario applicability score is obtained by quantitatively evaluating the degree to which the rule meets the applicable scenario requirements.

[0044] The organization and storage unit is used to organize and store all available rules labeled with quality scores to obtain the rule base.

[0045] A document review device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the document review method described above.

[0046] A computer-readable storage medium storing instructions that, when executed on a terminal device, cause the terminal device to perform the document review method as described above.

[0047] Compared with the prior art, this application has the following beneficial effects:

[0048] This application provides a document review method, apparatus, device, and storage medium. Specifically, when executing the document review method provided in this application embodiment, the candidate document and its review scenario are first obtained, and a suitable official rule file is matched (the two types of files are scanned images or Portable Document Format (PDF)). This step ensures that the review basis is highly consistent with business needs through scenario-based rule matching, avoiding interference from irrelevant rules. Subsequently, the image information of the two types of files is converted into editable text to obtain the target document and standard rule file, breaking through the review barrier of non-editable format documents and solving the problem of information omission caused by complex formats in manual review. Next, rules are extracted from the standard rule file and a rule library containing quality scores, rule identifiers (IDs), applicable scenarios, and judgment criteria is constructed. The standardized rule element definitions not only save the time of manually comparing rules word by word, but also improve the reliability of the review basis by filtering high-quality rules through quality scores. Then, based on the review scenario, the rule library is traversed to obtain multiple target review rules that match the applicable scenarios of the rules and the review scenario, realizing accurate rule filtering and avoiding rule abuse or misuse. Finally, the target documents are reviewed according to the target review rules, and real-time review comments are output. The fully automated process greatly shortens the review cycle, meets the needs of rapid business progress, and ensures the consistency and accuracy of the review results.

[0049] This application overcomes the compatibility barriers between different document formats by accurately associating the review scenarios of pending documents with suitable official rule documents, and then using image information to convert into editable text preprocessing. This avoids information omissions or tedious verification issues caused by complex formats during manual review. Simultaneously, by systematically extracting rules and constructing a rule library containing quality scores, clear rule IDs, applicable scenarios, and judgment criteria, it saves the significant time spent manually comparing complex rules word by word, eliminating the need for several days to review a single hundred-page document, greatly improving review efficiency to meet the needs of rapid business advancement. Furthermore, standardized rule element definitions and quality screening eliminate the impact of subjective factors such as human misunderstanding and fatigue. In addition, accurately recalling target review rules based on review scenarios effectively avoids misjudgments caused by rule abuse and misuse, ensuring the consistency and accuracy of review results across different scenarios and documents, fundamentally reducing the compliance risks that may arise from manual review. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in this embodiment or the prior art, the drawings used in the description of the embodiment or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 A flowchart illustrating a document review method provided in this application embodiment;

[0052] Figure 2 A flowchart illustrating a rule base construction method provided in this application embodiment;

[0053] Figure 3 A flowchart illustrating a method for filtering available rules provided in an embodiment of this application;

[0054] Figure 4 A flowchart illustrating a document secondary review method provided in this application embodiment;

[0055] Figure 5 This is a schematic diagram of a document review device provided in an embodiment of this application. Detailed Implementation

[0056] To facilitate understanding of the technical solutions provided in the embodiments of this application, the background technology involved in the embodiments of this application will be described below.

[0057] Document review is a crucial step in various industries, such as government, finance, and corporate offices. These industries rely on official regulatory documents, such as policies, regulations, business standards, and contract templates, for manual document review. However, manual review methods have significant shortcomings. First, inefficiency is a major problem, as reviewers need to compare document content line by line with the clauses in the regulatory documents—a time-consuming and tedious process. Second, the accuracy of manual review is difficult to guarantee, as reviewers may overlook important details due to fatigue or human error, leading to inaccurate results. Therefore, although official regulatory documents provide clear guidance, traditional manual review methods still face significant challenges in improving review efficiency and accuracy.

[0058] To address this issue, this application provides a document review method, apparatus, device, and storage medium. First, it acquires the document awaiting review and its review scenario, and simultaneously retrieves official rule files (both types of files are scanned images or PDFs) compatible with the scenario. Then, using image-to-text technology, the document awaiting review and the official rule files are converted into editable target documents and standard rule files, respectively, resolving the review compatibility issue for non-editable documents. Next, rule extraction is performed on the standard rule files to construct a rule base containing multiple review rules. Each rule is labeled with a quality score and clearly includes key elements such as rule ID, rule content, applicable scenario, and judgment criteria. Then, based on the review scenario of the document awaiting review, all review rules in the rule base are fully traversed to select multiple target review rules that match the scenario's applicable scenario and review scenario. Finally, compliance verification is performed on the target document based on these target review rules, and real-time review opinions for the document awaiting review are output. This application overcomes format adaptation barriers and significantly improves review efficiency by using scenario-based rule matching, image-to-text preprocessing, standardized rule base construction, and precise rule traversal. It also eliminates human subjective bias and ensures the consistency and accuracy of review results.

[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0060] See Figure 1 The figure is a flowchart of a document review method provided in an embodiment of this application, as shown below. Figure 1 As shown, this document review method may include steps S101-S105:

[0061] S101: Obtain the pending document and the review scenario of the pending document, and obtain the official rule file that is compatible with the review scenario.

[0062] To ensure that document review is closely aligned with specific business scenarios and based on authoritative compliance evidence, three key pieces of information are required: the document awaiting review, the specific review scenario for that document, and the official rule document highly compatible with that scenario. The document awaiting review is the object of review; the review scenario determines the core direction and focus of the review; and the official rule document is the fundamental basis for determining document compliance. All three are indispensable. Furthermore, considering the actual document management habits of industries such as government, finance, and corporate offices, scanned image formats (such as scanned contracts and photocopies of policy documents) and PDF formats are the most common document storage formats. Therefore, in this application, the document awaiting review and the official rule document can adopt, but are not limited to, these two formats. This aligns with industry practices and facilitates subsequent unified image-to-text processing.

[0063] For example, in the loan contract review scenario in the financial industry, the pending document is a PDF file of a loan contract submitted by a company, and the review scenario is clearly "compliance review of bank micro and small enterprise loan contracts". The corresponding official rule documents include the "Guidelines for Due Diligence in Bank Micro and Small Enterprise Credit Granting" (PDF format) and scanned copies of the "General Rules for Loans". Another example is the administrative license application review scenario in the government sector, where the pending document is a scanned copy of the license application materials submitted by the applicant, and the review scenario is "review of construction permit application for building projects". The corresponding official rule documents are the PDF file of the "Administrative Measures for Construction Permits for Building Projects" and scanned copies of the application specifications issued by the local housing and construction department.

[0064] S102: Convert the image information in the candidate document and the official rule file into editable text format respectively to obtain the target document and the standard rule file.

[0065] Since the obtained candidate documents and official rule documents are mostly scanned images or PDFs (non-editable text), direct rule matching and content verification are not possible. Therefore, it is necessary to perform format conversion processing on the image information of both types of documents: using text extraction technologies such as Optical Character Recognition (OCR), the image content in the candidate documents is converted into an editable and searchable text format to generate the target document. At the same time, the same technology is used to process the official rule documents from images to text to obtain the standard rule document.

[0066] This conversion process not only eliminates the limitations of non-editable formats on subsequent review work, enabling the system to accurately identify and compare document content with rule clauses, but also achieves unified standardization of the two types of file formats, providing an operable text foundation for subsequent rule extraction, intelligent matching and other processes. For example, after converting scanned loan contracts into editable text, the system can directly locate key clauses in the contract and compare them sentence by sentence with the compliance requirements in the standard rule documents, avoiding omissions or misreadings when manually identifying image content.

[0067] S103: Extract rules from the standard rule file to obtain a rule base.

[0068] The standard rule files for image-to-text conversion contain a large amount of unstructured rule descriptions, which cannot be directly used for intelligent review. Therefore, they need to be processed by structured rule extraction: Natural language processing technology is used to break down the core clauses in the standard rule files, extracting review rules with independent judgment value, and assigning each rule a unique rule ID for accurate identification. The rule content (specific compliance requirements), applicable scenarios (such as review of loan contracts for micro and small enterprises, and review of construction permit applications for building projects), and judgment criteria (specific conditions for compliance / non-compliance) are also clearly defined. Based on this, a quantitative score is generated by combining the rules' enforceability, specificity, and scenario adaptability, and a quality score is marked on the corresponding rule. Finally, all structured review rules are integrated and stored to form a standardized rule library.

[0069] For example, based on the standard rule document converted from the "Guidelines for Due Diligence in Banking Micro and Small Enterprise Credit Granting," an audit rule was extracted: Rule ID "JX001," rule content "Banks should implement a customer manager system for micro and small enterprise credit granting business, designating a specific person to be responsible for customer development, maintenance, and management," the applicable scenario of the rule is "bank micro and small enterprise loan contract auditing," and the rule judgment standard is "if the name and scope of responsibilities of the customer manager are clearly stipulated in the contract, it is compliant; if not stipulated or the stipulation is ambiguous, it is non-compliant." Furthermore, based on the rule's enforceability (clear judgment standards), specificity (unambiguous clause description), and scenario adaptability (fully matching the micro and small enterprise loan auditing scenario), it was assigned a quality score of 95 points, and this rule was ultimately included in the rule base. For example, a rule was extracted from the standard rule document of the "Administrative Measures for Construction Project Construction Permits": the rule ID is "JZ003", the rule content is "the construction unit applying for a construction permit shall have the condition that the construction funds have been secured", the rule application scenario is "application and review of construction project construction permits", the rule judgment standard is "if a bank-issued proof of funds or an investor's commitment to the availability of funds is provided, it is compliant; if no relevant supporting materials are provided, it is not compliant", the quality score is 92 points, and it is entered into the rule library simultaneously.

[0070] S104: Based on the audit scenario, perform rule traversal on the audit rules in the rule base to obtain multiple target audit rules.

[0071] To ensure that the review criteria are highly aligned with the current business scenario and to avoid irrelevant rules interfering with the review results, a full rule traversal of all review rules in the rule base is required based on the clearly defined review scenario of the documents to be reviewed. The system verifies the "applicable scenario" field of each rule one by one, filters out the rules that are completely matched with the current review scenario, and forms a set of target review rules specifically for this review. This ensures that subsequent reviews are conducted only based on the compliance requirements adapted to the scenario, thereby improving the accuracy of the review.

[0072] For example, assuming the current review scenario is "bank micro and small enterprise loan contract review", the system will traverse all rules in the rule base and filter out all rules whose "applicable scenario" is marked as "bank micro and small enterprise loan contract review" (such as the aforementioned rule with rule ID JX001). These rules together constitute the target review rules for this review. However, rules in the rule base whose applicable scenario is "construction engineering construction permit application review" (such as the rule with rule ID JZ003) will be excluded because the scenario does not match and will not be included in the target review rules.

[0073] S105: Review the target document based on the multiple target review rules to obtain real-time review opinions on the candidate document.

[0074] After determining the target review rules suitable for the current review scenario, the system will use these rules as the basis for judgment and conduct a line-by-line verification of the converted target document (the editable text version of the document awaiting review). Through text comparison, keyword extraction, logical matching and other technologies, the system will check whether the content of the target document meets the judgment criteria of each target review rule, record the results of compliance or non-compliance and the specific discrepancies, and finally integrate all the verification conclusions to form a real-time review opinion for the document awaiting review. The opinion will clearly mark the compliance items, non-compliance items and the corresponding rule basis of the document to ensure that the review results are traceable and explainable.

[0075] For example, assuming a scenario of "bank micro and small enterprise loan contract review," the target review rule includes JX001 (requiring the contract to clearly state the name and responsibilities of the account manager). When the system verifies the target document (editable text of the loan contract), if it finds that the contract only mentions "designated person in charge" but does not specify the name and scope of responsibilities of the account manager, it determines that the clause does not meet the judgment criteria of rule JX001. Simultaneously, combining the verification results of other target review rules, the system ultimately generates a real-time review opinion: "1. The contract does not clearly state the name and specific responsibilities of the account manager for micro and small enterprise loan business, which does not comply with the relevant rules of the 'Guidelines for Due Diligence in Bank Micro and Small Enterprise Credit Granting' (Rule ID: JX001); 2. The clauses in the contract regarding the purpose of loan funds and repayment methods all comply with the corresponding review rule requirements; In summary, this contract has one violation and needs to be supplemented and improved before resubmitting for review." If the target document fully complies with all target review rules, the real-time review opinion is: "After review, all clauses in this document comply with the corresponding compliance rules requirements, and the review is passed."

[0076] Based on the content of S101-S105, the process first involves acquiring the candidate document and its review scenario, and then matching it with the official rule file for that scenario (both types of files are scanned images or PDFs). Next, image-to-text technology is used to convert both into editable target documents and standard rule files, overcoming the limitations of non-editable formats. Then, rules are extracted from the standard rule file to build a rule base. Each review rule in the base is labeled with a quality score and includes core information such as rule ID, content, applicable scenario, and judgment criteria. Following this, a full rule traversal is performed on the rule base based on the review scenario to select suitable target review rules. Finally, the target document is reviewed according to these target review rules, and real-time review opinions are output for the candidate document, achieving intelligent processing from document preprocessing to review conclusion output. This application not only shortens review time and improves efficiency but also reduces misjudgments caused by fatigue or subjective bias, ensuring the consistency and accuracy of review results.

[0077] In one possible implementation, this application also provides a method for constructing a rule base, such as... Figure 2 As shown, correspondingly, step S103 extracts rules from the standard rule file to obtain a rule base, which can be specifically implemented through steps S201-S203:

[0078] S201: Perform paragraph analysis on each text paragraph in the standard rule file to obtain multiple usable rules.

[0079] To effectively extract rules suitable for document review, it's essential to first analyze each text paragraph in the standard rule file. This analysis process identifies and extracts multiple usable rules. This step is crucial because it ensures that the rules used in the review process are comprehensive and accurate, thus covering a wide range of possible scenarios and needs.

[0080] S202: Calculate the quality score of each available rule and mark the quality score of each available rule on the corresponding available rule.

[0081] To select high-quality rules that can be effectively used for auditing and to ensure the reliability and usability of the rule base, each available rule needs to be quantitatively scored, and the score needs to be linked to the corresponding rule. This quality score is not determined by a single dimension, but is calculated by comprehensively considering the scoring results of three core dimensions: rule executability, specificity, and scenario applicability.

[0082] The three dimensions are: whether the executability scoring rules have clear judgment criteria (whether they can clearly determine whether the document meets the requirements), whether the specificity scoring rules are accurately and unambiguously described (whether there are vague or general expressions), and whether the scenario applicability scoring rules match the preset business review scenario (whether they are applicable to the compliance judgment of the target scenario). After quantifying and scoring each of the three dimensions, the final quality score of the rule is obtained through weighted calculation and other methods.

[0083] For example, suppose we score the following dimensions based on the applicable rule extracted from the "Guidelines for Due Diligence in Banking Small and Micro Enterprise Credit Granting": "Banks should implement a customer manager system for small and micro enterprise credit granting business, and must designate a specific person to be responsible for customer development, maintenance and management. If the name and specific responsibilities of the customer manager are not clearly stated in the contract, the loan contract will not be approved."

[0084] Feasibility score: The rules clearly define the core compliance requirements, as well as the judgment conditions and corresponding consequences, and have an implementable review standard (if the contract clearly states the account manager information, it is compliant; if not, it is not compliant). Therefore, the feasibility score is 90 points (out of 100).

[0085] Specificity score: The scope of responsibilities for "account manager system", "account development, maintenance and management" and the consequences of "approval failure" in the rules are all accurate and unambiguous, without any vague expressions. Therefore, the specificity score is 95 points.

[0086] Scenario Applicability Score: This rule perfectly matches the "bank micro and small enterprise loan contract review" scenario, therefore the scenario applicability score is 98 points;

[0087] If the weight of each of the three dimensions is 1 / 3, then the final quality score of the rule is (90+95+98)÷3≈94.3 points, and this score will be directly marked on the rule.

[0088] For example, a usable rule states that "banks must strengthen credit risk management for micro and small enterprises. If they fail to establish and regularly update risk control ledgers as required, they will be prohibited from issuing credit loans to micro and small enterprises." However, because the core requirement of "strengthening credit risk management for micro and small enterprises" lacks specific standards (it is judged only by the single condition of "establishing risk ledgers," without defining specific risk management measures), it cannot be fully implemented, and its enforceability score is only 40 points. The description of "strengthening risk management" is too general and lacks detailed definition, resulting in a specificity score of 50 points. Although it matches the target scenario of "bank micro and small enterprise loan contract review" (scenario applicability score of 90 points), the final quality score is (40+50+90)÷3≈60 points, which is far lower than the previous rule, reflecting its quality shortcomings.

[0089] S203: Organize and store all available rules labeled with quality scores to obtain the rule base.

[0090] To establish a standardized rule system that is clearly structured, easy to retrieve and call, and reusable over the long term, all available rules with quality scores need to be uniformly organized and persistently stored, ultimately building a rule library. The organization phase mainly focuses on the classification, identification, and association of rules, including assigning a unique rule ID to each rule, categorizing them according to applicable scenarios (such as "Commercial Bank SME Loan Review Rules" and "Construction Project Construction Permit Review Rules"), and associating the original policy basis corresponding to the rules. The storage phase involves entering the organized rule information (including rule ID, content, applicable scenarios, judgment criteria, quality scores, etc.) into the database, ensuring that the system can quickly retrieve and match rules, providing stable support for subsequent review work.

[0091] By implementing the above steps, the construction of the rule base not only improves the usability and effectiveness of the rules, but also helps to optimize and refine the overall document review process.

[0092] In one possible implementation, this application also provides a filtering method using available rules, such as... Figure 3 As shown, correspondingly, step S201 performs paragraph analysis on each text paragraph in the standard rule file to obtain multiple usable rules, which can be specifically implemented through steps S301-S304:

[0093] S301: Based on the first screening requirements, each text paragraph in the standard rule file is initially screened to obtain multiple screened paragraphs.

[0094] In order to quickly identify the core content with compliance determination value from a large number of text paragraphs in the standard rule documents and filter out redundant paragraphs such as background descriptions and supplementary explanations that have no determination significance, the system needs to perform preliminary screening on each text paragraph based on the first screening requirement, and finally obtain multiple filtered paragraphs.

[0095] It should be noted that the first screening requirement sets two core conditions: the paragraph must contain at least one target keyword (such as words with mandatory constraints such as "should", "must", "must", "prohibited") and conform to at least one target sentence structure (such as "condition + requirement + consequence", "if...then...otherwise...", etc., which have a clear logical judgment relationship), so as to ensure that the selected paragraphs are all valid content that can be converted into review rules.

[0096] For example, suppose that in the text paragraph of the "Guidelines for Due Diligence in Banking Credit Granting to Micro and Small Enterprises", paragraph 1 is "Banks should implement a customer manager system for credit granting to micro and small enterprises. If the name and specific responsibilities of the customer manager are not clearly stated in the contract, the loan contract will not be approved." This paragraph contains the target keyword "should" and meets the target sentence structure of "requirement + condition + consequence". Therefore, it passes the initial screening and becomes the screening paragraph.

[0097] Paragraph 2, which reads "Credit business for micro and small enterprises is an important type of business for banks to support the real economy, and the relevant approval process must follow internal management regulations," is excluded because it contains neither target keywords such as "should" nor "must," nor does it use a specific target sentence structure; it is merely a description of the business background.

[0098] Paragraph 3, which reads "Banks are prohibited from issuing credit loans to micro and small enterprises without a real business scenario. If the business scenario is found to be false during the verification, the credit process will be terminated immediately," contains the target keyword "prohibition" and conforms to the target sentence structure of "requirement + condition + consequence." Therefore, it has also passed the screening and become a selected paragraph.

[0099] S302: Calculate the fit score of each filtered paragraph, and retain the filtered paragraphs with a fit score greater than the fit threshold to obtain candidate paragraphs.

[0100] To further improve the accuracy of rule extraction and eliminate screening paragraphs with weak adaptability, it is necessary to quantify and score the adaptability of each screening paragraph, and then retain paragraphs with scores exceeding the preset adaptability threshold as candidate paragraphs.

[0101] It should be noted that the fit score is calculated around two main dimensions: the number of target keywords and the number of target sentence structures. Both are positive scoring indicators: the more target keywords (such as "should" and "must") included and the richer the target sentence structures (such as "condition + requirement + consequence"), the higher the paragraph fit score. This allows for the precise selection of paragraphs with high rule conversion value, laying the foundation for extracting high-quality rules in the future.

[0102] For example, suppose the fit scoring criteria are set as follows: "20 points for a single target keyword, 30 points for a single target sentence structure (each type of target sentence structure is worth a maximum of 30 points; if only one type is met, 30 points are awarded; if two or more types are met, 30 points are still awarded), with a maximum score of 100 points," and the fit threshold is set at 60 points.

[0103] Paragraph 1: "Banks should implement a customer manager system for credit business to micro and small enterprises. If no specific person is designated to be responsible for customer development, maintenance and management, the loan contract review will not be approved." It contains one target keyword, "should" (20 points), and meets one target sentence structure, "requirement + condition + consequence" (30 points). The total fit score is 50 points, which is below the threshold of 60 points, so it is excluded.

[0104] Paragraph 2: "Banks must establish a credit risk ledger for micro and small enterprises and must not conceal risk data. If the ledger is not updated as required, the relevant business approval will be suspended." It contains two target keywords, "must" and "must not" (40 points), and meets one target sentence structure, "requirement + condition + consequence" (30 points). The total fit score is 70 points, which is higher than the threshold and is retained as a candidate paragraph.

[0105] Paragraph 3: "Banks are prohibited from issuing loans to micro and small enterprises without a real business scenario. If the business scenario is found to be false, the credit line will be terminated immediately and the enterprise will not be allowed to apply for credit again within 3 years." It contains two target keywords, "prohibited" and "must not" (40 points), and meets one target sentence structure, "requirement + condition + consequence" (30 points). The total fit score is 70 points, which is far above the threshold and is retained as a candidate paragraph.

[0106] Paragraph 4: "Banks should standardize the credit approval process for micro and small enterprises, and must clearly define the approval time limit. If the approval is not completed within the time limit, a written explanation must be given to the applicant, and the approval process must not be delayed without reason." It contains three target keywords: "should," "must," and "must not" (60 points), and meets one target sentence structure: "requirement + condition + consequence" (30 points). The total fit score is 90 points, which is higher than the threshold, and it is retained as a candidate paragraph.

[0107] S303: Based on the preset rule extraction requirements, rule extraction is performed on each candidate paragraph to obtain multiple candidate rules.

[0108] To transform candidate paragraphs in natural text form into structurally unified and element-complete rule prototypes for subsequent quality verification and rule base management, the system needs to precisely decompose and extract information from each candidate paragraph according to preset rule extraction requirements, ultimately generating multiple candidate rules. These rule extraction requirements clearly define four core extraction elements: rule ID (assigning a unique identifier to each rule to ensure the uniqueness and traceability of rule management), rule content (condensing the core compliance constraints in the paragraph), rule application scenario (identifying the specific business review scenario corresponding to the rule), and rule judgment criteria (transforming the logical relationships in the paragraph into clear compliance judgment criteria). This ensures that all candidate rules have a standardized structural framework, avoiding missing key information or inconsistent formatting.

[0109] For example, suppose we perform structured extraction for the candidate paragraph "Banks must establish a credit risk ledger for micro and small enterprises and must not conceal risk data. If the ledger is not updated as required, the relevant business approval will be suspended." The extraction should be performed according to the rules.

[0110] Rule ID: Set to "BANK-MICRO-RULE-001" (consisting of bank identifier, business domain, rule type, and sequence number to ensure global uniqueness);

[0111] The rules stipulate that banks must establish risk ledgers in their lending business to micro and small enterprises, and must not conceal risk data. They must also update the ledgers regularly as required.

[0112] Applicable scenarios for the rules: Risk assessment of bank credit business for micro and small enterprises;

[0113] The criteria for determining compliance with the rules are as follows: if a risk ledger has been established, no risk data has been concealed, and the data has been updated as required, then the rules are met; if a ledger has not been established, data has been concealed, or the data has not been updated on time, then the rules are not met.

[0114] S304: Based on the second screening requirements, perform quality screening on each candidate rule to obtain the multiple available rules.

[0115] To ensure that the final generated usable rules have the core attributes of being implementable, unambiguous, and scenario-matched, a quality screening of candidate rules is required based on the second screening criteria. Only rules that simultaneously meet the requirements of executability, specificity, and applicable scenarios will be retained as usable rules.

[0116] Among them, the enforceability requirement focuses on whether the rule has clear judgment criteria (whether it can clearly determine whether the document is compliant), the specificity requirement focuses on whether the rule description is free of ambiguity (avoiding vague and ambiguous content), and the applicable scenario requirement focuses on whether the rule is clearly marked with a business scenario (ensuring that it matches the review requirements). Through three layers of verification, unqualified candidate rules are eliminated.

[0117] Specifically, firstly, regarding the enforceability requirement, if a rule contains clear criteria for judgment, then the rule is considered enforceable. For example, the rule "must be stamped with an official seal" provides a clear criterion, indicating under what circumstances the rule can be enforced, thus satisfying the enforceability requirement. Conversely, expressions like "must comply with regulations" lack clear criteria for judgment, making it impossible to determine whether it is enforceable, and therefore do not meet the enforceability requirement.

[0118] Secondly, regarding the requirement of specificity, the wording of the rules must be unambiguous. For example, the phrases "within 30 days" and "double wages" are specific and their meanings can be clearly understood; while "relevant materials" and "reasonable period" are vague expressions because they do not provide sufficient information, making the implementation and understanding of the rules difficult. Therefore, only those rules without vague wording can be considered to meet the requirement of specificity.

[0119] Finally, the applicable scenario requirement is also crucial. The rule must include a clear applicable scenario so that users can clearly understand in what situations the rule applies. For example, "applies to labor contract review" provides a clear scope of application and therefore meets the applicable scenario requirement; while "applies to related documents" is too broad and not specific enough, and therefore does not meet this requirement.

[0120] By following the steps above, high-quality, usable rules can be systematically extracted from standard rule files, enhancing the practicality and reliability of the rule base and providing a solid foundation for subsequent applications.

[0121] In one possible implementation, this application also provides a secondary review method for documents, such as... Figure 4 As shown, this can be achieved through the following steps:

[0122] S401: Based on the audit scenario, retrieve the audit rules in the rule base to obtain at least one secondary audit rule.

[0123] To further enhance the reliability of audit results and focus on core compliance points, the system will selectively recall rules from the rule base based on the current audit scenario, forming at least one secondary audit rule.

[0124] It should be noted that the secondary review rules are not a simple repetition of the target review rules, but are constructed by selecting core rules from the rule base that have higher quality scores and more accurate matching with the scenario. Therefore, their number is strictly less than the target review rules. The aim is to complete the secondary verification of documents with concise and high-priority rules, and avoid redundant rules interfering with the review conclusion.

[0125] S402: Review the target document based on all secondary review rules to obtain supplementary review comments on the candidate document.

[0126] After recalling the secondary review rules, the system uses these streamlined and high-priority secondary review rules as a unified judgment basis to conduct a comprehensive review and verification of the target document (the editable text version of the document awaiting review). It checks whether the document content meets the judgment criteria of each secondary review rule, and finally integrates all secondary verification results to form a supplementary review opinion for the document awaiting review. This opinion focuses on the verification conclusion of the secondary review rules, is independent of the previous real-time review opinion, and is used to cross-verify with the real-time review opinion to further ensure the accuracy of the review results.

[0127] S403: Calculate the similarity between the supplementary review comments and the real-time review comments.

[0128] To determine whether the core conclusions and violation judgment criteria of the supplementary review comments and the real-time review comments are consistent, the system needs to perform similarity calculations on the two review comments. Specifically, text similarity algorithms (such as cosine similarity, Jaccard similarity, etc.) can be used to quantitatively analyze the degree of overlap between the two comments in key dimensions such as core semantics, rule references, and judgment results. This measures the level of consistency between the two and provides data support for subsequent integration or selection of the best final review result.

[0129] S4041: If the similarity is greater than or equal to the similarity threshold, then the supplementary review comments and the real-time review comments are deduplicated.

[0130] If the similarity between the supplementary review comments and the real-time review comments reaches or exceeds the preset similarity threshold, it means that the core judgment conclusions, violation basis, and rule citations of the two comments are highly overlapping, with only differences in the details of the wording. In this case, the two comments need to be deduplicated to remove duplicate rules, that is, the supplementary review comments or the real-time review comments are randomly removed.

[0131] S4042: If the similarity is less than the similarity threshold, the accuracy of the supplementary review opinion and the real-time review opinion is evaluated based on the quality score and number of rules of each secondary review rule, as well as the quality score and number of rules of each target review rule. The review opinion with higher accuracy after the accuracy evaluation is taken as the final standardized review result.

[0132] If the similarity between the supplementary review comments and the real-time review comments is below the similarity threshold, it indicates a significant discrepancy in the core judgments of the two comments. In this case, it is necessary to assess the accuracy of both comments by combining key indicators of the rule dimension, and finally select the review comment with higher accuracy as the standardized result. The assessment criteria are: the number of rules supporting each review comment (the number of secondary review rules corresponds to the supplementary review comments, and the number of target review rules corresponds to the real-time review comments), and the average quality score of the rules (the average quality score of secondary review rules corresponds to the supplementary review comments, and the average quality score of target review rules corresponds to the real-time review comments). The principle of "the more rules and the higher the average quality score, the higher the accuracy of the corresponding review comments" is followed to ensure the reliability of the final conclusion.

[0133] For example, if the similarity threshold is set to 70%, and the similarity between the supplementary review comments and the real-time review comments is 60% (below the threshold), then the accuracy assessment stage begins.

[0134] The secondary review rules supporting the supplementary review comments consist of 3 rules, which are the core rules with quality scores of 95, 92 and 98 respectively. The average quality score is (95+92+98)÷3≈95.

[0135] The target review rules supporting real-time review comments consist of 8 rules, with quality scores of 75, 80, 82, 78, 85, 70, 83, and 76 respectively. The average quality score is (75+80+82+78+85+70+83+76)÷8=78.625.

[0136] Evaluation results: The number of rules in the real-time review comments (8) far exceeded the number of supplementary review comments (3), but the average quality score of the rules in the supplementary review comments (95 points) was significantly higher than that of the real-time review comments (78.625 points). Considering both the "quantity + average quality" dimensions, the supplementary review comments were more accurate, and therefore they were used as the final standardized review results.

[0137] This secondary review mechanism can better adapt to complex review scenarios and improve the overall efficiency and accuracy of the review process.

[0138] In one possible implementation, to improve the standardization of document management and the efficiency of audit traceability, the method also includes a metadata annotation step: the system first obtains the metadata information corresponding to the target document and the standard rule file respectively (the metadata specifically covers three core contents: document / file name, upload time, and the business scenario to which it belongs (similar to the audit scenario, such as the "compliance audit of bank micro and small enterprise loan contracts")), and then accurately annotates this metadata in the corresponding target document and standard rule file. The annotated metadata can be directly associated with the basic attributes of the document / file, which not only facilitates quick retrieval and location of relevant documents, but also clearly traces the document source, applicable scenario, and upload time during the audit process, providing support for the traceability of audit results and management efficiency.

[0139] In one possible implementation, to ensure the accuracy of subsequent rule extraction, document review, and other processes, and to avoid interference from irrelevant information, the method also includes a specific data cleaning process for the target document and standard rule files:

[0140] The specific cleaning process focuses on removing invalid content and correcting text defects: on the one hand, meaningless characters (such as non-functional special symbols, placeholder content in blank pages) and redundant information (such as duplicate headers and footers, watermark text that does not affect the core content, and explanations in document comments that are not related to business); on the other hand, text formatting issues are corrected, including resolving garbled characters caused by abnormal character encoding and fixing sentence segmentation deviations caused by formatting errors (such as incorrectly splitting complete sentences and semantic fragmentation caused by misplaced punctuation), ultimately outputting clean and formatted text, laying a reliable data foundation for subsequent processing.

[0141] For example, if the standard rule file contains the watermarked phrase "Internal Document," duplicate file names in the header, or meaningless special symbols like "★★★" in the target document, or if formatting errors cause errors such as splitting "Small and micro enterprise credit business requires an account manager system. If no designated person is specified, the review will not be approved" into "Small and micro enterprise credit business requires an account manager system. If no designated person is specified" and "If no designated person is specified, the review will not be approved," the data cleaning process will remove watermarks, redundant header content, and special symbols, correct punctuation errors, and restore the complete and standardized text.

[0142] See Figure 5 , 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 includes:

[0143] The acquisition unit 501 is used to acquire the document under review and the review scenario of the document under review, and to acquire the official rule file adapted to the review scenario; the content of the document under review and the official rule file are in either image format or portable document format PDF format;

[0144] The conversion unit 502 is used to convert the image information in the candidate document and the official rule file into an editable text format, respectively, to obtain the target document and the standard rule file;

[0145] The rule base construction unit 503 is used to extract rules from the standard rule file to obtain a rule base; the rule base includes multiple audit rules marked with quality scores; the audit rules include rule ID, rule content, rule application scenario and rule judgment criteria;

[0146] The rule traversal unit 504 is used to traverse the review rules in the rule base based on the review scenario to obtain multiple target review rules; the applicable scenario of the target review rules matches the review scenario.

[0147] The first review unit 505 is used to review the target document based on the multiple target review rules and obtain real-time review opinions on the candidate document.

[0148] In one possible implementation, the rule base construction unit 504 specifically includes:

[0149] The paragraph analysis unit is used to perform paragraph analysis on each text paragraph in the standard rule file to obtain multiple usable rules;

[0150] The scoring and labeling unit is used to calculate the quality score of each available rule and label the quality score of each available rule on the corresponding available rule. The quality score of a rule is calculated based on the rule executability score, the rule specificity score, and the rule scenario applicability score. The rule executability score is obtained by quantitatively evaluating the degree to which the rule meets the executability requirements. The rule specificity score is obtained by quantitatively evaluating the degree to which the rule meets the specificity requirements. The rule scenario applicability score is obtained by quantitatively evaluating the degree to which the rule meets the applicable scenario requirements.

[0151] The organization and storage unit is used to organize and store all available rules labeled with quality scores to obtain the rule base.

[0152] In one possible implementation, the paragraph analysis unit is specifically used for:

[0153] Based on the first screening requirement, each text paragraph in the standard rule file is initially screened to obtain multiple screened paragraphs; the first screening requirement includes screening out text paragraphs that contain at least one target keyword and conform to at least one target sentence structure.

[0154] The fit score of each selected paragraph is calculated, and the selected paragraphs with a fit score greater than the fit threshold are retained to obtain candidate paragraphs; the fit score is obtained by quantitative evaluation based on the number of target keywords and the number of target sentence structures included in the rule.

[0155] Based on preset rule extraction requirements, rule extraction is performed on each candidate paragraph to obtain multiple candidate rules; the rule extraction requirements are used to indicate the rule ID, rule content, rule application scenario, and rule judgment criteria in the extracted text;

[0156] Based on the second screening criteria, each candidate rule is subjected to quality screening to obtain the plurality of usable rules; the second screening criteria include screening out rules that simultaneously meet the executability requirement, the specificity requirement, and the applicable scenario requirement;

[0157] Specifically, if a rule has a clear judgment criterion, then the rule is determined to meet the enforceability requirement; if a rule has no vague expression, then the rule is determined to meet the specificity requirement; if a rule includes a clear applicable scenario, then the rule is determined to meet the applicable scenario requirement.

[0158] In one possible implementation, the device further includes:

[0159] The rule recall unit is used to recall the review rules in the rule base based on the review scenario to obtain at least one secondary review rule; the number of secondary review rules is less than the number of target review rules.

[0160] The second review unit is used to review the target document based on all secondary review rules and obtain supplementary review opinions on the candidate document.

[0161] A similarity calculation unit is used to calculate the similarity between the supplementary review comments and the real-time review comments;

[0162] The deduplication unit is used to deduplicatize the supplementary review comments and the real-time review comments if the similarity is greater than or equal to the similarity threshold.

[0163] The comprehensive unit, if the similarity is less than the similarity threshold, is used to evaluate the accuracy of the supplementary review opinions and the real-time review opinions based on the quality scores and the number of rules of each secondary review rule, as well as the quality scores and the number of rules of each target review rule. The review opinion with higher accuracy after the accuracy evaluation is taken as the final standardized review result.

[0164] Among them, the more rules there are and the higher the average quality score of the rules, the more accurate the review comments are.

[0165] In one possible implementation, the device further includes:

[0166] The metadata acquisition and annotation unit is used to acquire the metadata corresponding to the target document and the standard rule file respectively, and to annotate the target document and the standard rule file; the metadata includes the name of the document or file, the upload time, and the business scenario.

[0167] In one possible implementation, the device further includes:

[0168] The cleaning unit is used to perform data cleaning on the target document and the standard rule file respectively;

[0169] The data cleaning process includes: removing meaningless characters and redundant information from the document, and correcting text garbled characters and sentence breaks.

[0170] In addition, this application embodiment also provides a document review device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the document review method as described above.

[0171] In addition, this application embodiment also provides a computer-readable storage medium storing instructions that, when executed on a terminal device, cause the terminal device to perform the document review method as described above.

[0172] This application's embodiments firstly solve the adaptation problem of different document formats by accurately matching the review scenario with the official rule document, combined with the image-to-editable text preprocessing step. This completely avoids the problems of information omission and low verification efficiency caused by complex formats in manual review. Secondly, by systematically extracting rules to form a standardized rule library containing quality scores, rule IDs, applicable scenarios, and judgment criteria, it not only saves a lot of time cost of manually comparing rules word by word (such as significantly shortening the review cycle of hundreds of pages of documents) and meets the efficiency requirements of rapid business progress, but also eliminates the risk of misjudgment caused by subjective understanding bias and fatigue in manual review through the standardized definition of rule elements and quality screening mechanism. Finally, by relying on the review scenario to achieve accurate recall of target review rules, it avoids the abuse and misuse of rules, and ensures the consistency and accuracy of review results for different scenarios and different documents. This resolves the compliance risks hidden in the manual review model from the source and achieves a dual improvement in review efficiency and quality.

[0173] The foregoing has provided a detailed description of a document review method, apparatus, device, and storage medium provided in this application. The various embodiments in the specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section. It should be noted that those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of the claims of this application.

[0174] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0175] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

Claims

1. A document review method, characterized in that, The method includes: Obtain the pending documents and the review scenarios for the pending documents, and obtain the official rule files that are compatible with the review scenarios; the content of the pending documents and the official rule files are in either image format or portable document format PDF format; The image information in the candidate document and the official rule file are converted into editable text format to obtain the target document and the standard rule file, respectively. The standard rule file is used to extract rules to obtain a rule base; the rule base includes multiple audit rules marked with quality scores; the audit rules include rule identifier ID, rule content, rule application scenario and rule judgment criteria; Based on the audit scenario, the audit rules in the rule base are traversed to obtain multiple target audit rules; the applicable scenarios of the target audit rules match the audit scenario. The target document is reviewed based on the multiple target review rules to obtain real-time review comments on the candidate document.

2. The method according to claim 1, characterized in that, The step of extracting rules from the standard rule file to obtain a rule base includes: Each text paragraph in the standard rule file is analyzed to obtain multiple usable rules; Calculate the quality score of each available rule and mark the quality score of each available rule on the corresponding available rule; the quality score of the rule is calculated based on the rule executability score, the rule specificity score, and the rule scenario applicability score. The rule executability score is obtained by quantitatively evaluating the degree to which the rule meets the executability requirements. The rule specificity score is obtained by quantitatively evaluating the degree to which the rule meets the specificity requirements. The rule scenario applicability score is obtained by quantitatively evaluating the degree to which the rule meets the applicable scenario requirements. All available rules labeled with quality scores are organized and stored to obtain the rule base.

3. The method according to claim 2, characterized in that, The paragraph analysis performed on each text paragraph in the standard rule file yields multiple usable rules, including: Based on the first screening requirement, each text paragraph in the standard rule file is initially screened to obtain multiple screened paragraphs; the first screening requirement includes screening out text paragraphs that contain at least one target keyword and conform to at least one target sentence structure. The fit score of each selected paragraph is calculated, and the selected paragraphs with a fit score greater than the fit threshold are retained to obtain candidate paragraphs; the fit score is obtained by quantitative evaluation based on the number of target keywords and the number of target sentence structures included in the rule. Based on preset rule extraction requirements, rule extraction is performed on each candidate paragraph to obtain multiple candidate rules; the rule extraction requirements are used to indicate the rule ID, rule content, rule application scenario, and rule judgment criteria in the extracted text; Based on the second screening criteria, each candidate rule is subjected to quality screening to obtain the plurality of usable rules; the second screening criteria include screening out rules that simultaneously meet the executability requirement, the specificity requirement, and the applicable scenario requirement; Specifically, if a rule has a clear judgment criterion, then the rule is determined to meet the enforceability requirement; if a rule has no vague expression, then the rule is determined to meet the specificity requirement; if a rule includes a clear applicable scenario, then the rule is determined to meet the applicable scenario requirement.

4. The method according to claim 1, characterized in that, The method further includes: Based on the aforementioned review scenario, the review rules in the rule base are retrieved to obtain at least one secondary review rule; the number of secondary review rules is less than the number of target review rules. The target document is reviewed based on all secondary review rules to obtain supplementary review comments for the document under review; Calculate the similarity between the supplementary review comments and the real-time review comments; If the similarity is greater than or equal to the similarity threshold, then the supplementary review comments and the real-time review comments are deduplicated. If the similarity is less than the similarity threshold, the accuracy of the supplementary review opinions and the real-time review opinions is evaluated based on the quality score and number of rules of each secondary review rule, as well as the quality score and number of rules of each target review rule. The review opinion with higher accuracy after the accuracy evaluation is taken as the final standardized review result. Among them, the more rules there are and the higher the average quality score of the rules, the more accurate the review comments are.

5. The method according to claim 1, characterized in that, The method further includes: Obtain the metadata corresponding to the target document and the standard rule file respectively, and annotate them in the target document and the standard rule file; the metadata includes the name of the document or file, the upload time, and the business scenario.

6. The method according to claim 1, characterized in that, The method further includes: Data cleaning is performed on the target document and the standard rule file respectively; The data cleaning process includes: removing meaningless characters and redundant information from the document, and correcting text garbled characters and sentence breaks.

7. A document review device, characterized in that, The device includes: The acquisition unit is used to acquire the candidate document and the review scenario of the candidate document, and to acquire the official rule file adapted to the review scenario; the content of the candidate document and the official rule file are in either image format or portable document format PDF format; The conversion unit is used to convert the image information in the candidate document and the official rule file into an editable text format, respectively, to obtain the target document and the standard rule file; The rule base construction unit is used to extract rules from the standard rule file to obtain a rule base; the rule base includes multiple audit rules marked with quality scores; the audit rules include rule ID, rule content, rule application scenario and rule judgment criteria; The rule traversal unit is used to traverse the review rules in the rule base based on the review scenario to obtain multiple target review rules; the applicable scenario of the target review rules matches the review scenario. The first review unit is used to review the target document based on the multiple target review rules and obtain real-time review opinions on the candidate document.

8. The apparatus according to claim 7, characterized in that, The rule base construction unit specifically includes: The paragraph analysis unit is used to perform paragraph analysis on each text paragraph in the standard rule file to obtain multiple usable rules; The scoring and labeling unit is used to calculate the quality score of each available rule and label the quality score of each available rule on the corresponding available rule. The quality score of a rule is calculated based on the rule executability score, the rule specificity score, and the rule scenario applicability score. The rule executability score is obtained by quantitatively evaluating the degree to which the rule meets the executability requirements. The rule specificity score is obtained by quantitatively evaluating the degree to which the rule meets the specificity requirements. The rule scenario applicability score is obtained by quantitatively evaluating the degree to which the rule meets the applicable scenario requirements. The organization and storage unit is used to organize and store all available rules labeled with quality scores to obtain the rule base.

9. A document review device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the document review method as described in any one of claims 1-6.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed on a terminal device, cause the terminal device to perform the document review method as described in any one of claims 1-6.

Images