Document auditing method, device and equipment based on large model and medium

Abstract

The invention provides an official document auditing method, device and equipment based on a large model and a medium, and the method comprises the following steps: preprocessing an official document to realize format and language standardization; using a retrieval enhancement generation (RAG) technology to complete vectorization conversion and fragmentation processing; then grammar and semantic analysis is carried out through a text analysis module fused with the large model, and matching detection is carried out according to a preset rule base in combination with an automatic rule engine module; then, deep auditing is conducted by means of a machine learning auditing model, content problems are recognized, manual rechecking is triggered in case of high uncertainty or key content, and an auxiliary tool is provided; receiving an audit opinion optimization model and a rule base; and finally outputting a result, and feeding back the problem by generating a report without annotation. The document auditing efficiency and accuracy can be improved, continuous optimization is supported, and the method is suitable for multiple scenes.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a document review method, apparatus, equipment and medium based on a large model. Background Technology

[0002] With the deepening of digital transformation in government affairs and enterprises, official documents, as the core carrier of policy implementation, decision transmission, and business communication, have fully entered the stage of electronic processing. The number of various official documents (such as policy documents, project plans, and corporate reports) generated is growing exponentially, and the content involves key information such as policy compliance, trade secrets, and personal privacy, placing extremely high demands on the professionalism, accuracy, and timeliness of the review process. Traditional document review relies on manual work, requiring reviewers to possess extensive industry knowledge, legal expertise, and document writing experience. This not only results in high labor costs and low review efficiency but is also prone to inconsistencies in review results due to differences in individual subjective judgment and oversights caused by fatigue, making it difficult to meet the demand for rapid review of massive amounts of official documents.

[0003] In recent years, the rapid development of artificial intelligence technology, especially large-scale modeling technology, has made it possible to upgrade official document review to intelligent levels. While some existing text review solutions incorporate natural language processing (NLP) technology for basic grammatical error correction or sensitive word detection, they generally have limitations: firstly, they lack deep adaptation to the specific knowledge of official document domains, making it difficult to accurately identify semantic contradictions and compliance risks unique to official documents (such as policy wording deviations or inconsistencies with industry standards); secondly, they have not formed a closed-loop mechanism of "preprocessing-analysis-review-feedback," making it impossible to dynamically optimize models and rules based on the review scenario. Furthermore, deficiencies exist in key areas such as data security protection and human-machine collaborative review, resulting in insufficient reliability of review results and making it difficult to truly replace manual review and meet the stringent requirements of official document processing. Summary of the Invention

[0004] In view of this, the embodiments of this application provide a document review method, apparatus, equipment and medium based on a large model, which can improve the efficiency and accuracy of document review, support continuous optimization and is applicable to multiple scenarios.

[0005] The technical solution of this application embodiment is implemented as follows: In a first aspect, embodiments of this application provide a document review method based on a large model, the method comprising: Perform preprocessing operations on official documents that require review, standardize document format and language, and adapt them to subsequent review processes; Based on the search-enhanced generation of RAG, the preprocessed official documents are vectorized and segmented to obtain document vector fragments and segmented data. The system integrates a large-scale model with a text analysis module and an automatic rule engine module. The text analysis module performs syntactic and semantic analysis on the fragmented data of official documents through natural language processing, while the automatic rule engine module performs rule matching and detection based on a preset official document review rule base. Based on a machine learning-based review model, combined with text analysis results and rule matching results, the system conducts in-depth reviews of official documents to identify potential inconsistencies, errors, and compliance issues. When uncertainty exceeds the uncertainty threshold or critical document content occurs during the review process, manual review is triggered, prompting reviewers with information on content requiring review and providing decision support tools. Receive review comments submitted by reviewers and users through the feedback interface, and iterate and optimize the rule base of the machine learning review model and the automatic rule engine module based on these comments; Output the review results of the official document content. If the review is approved, a review approval report will be generated. If the review is not approved, the problem type and location will be marked and feedback will be sent to the relevant personnel.

[0006] Secondly, embodiments of this application also provide a document review device based on a large model, the device comprising: The preprocessing module is used to perform preprocessing operations on official documents that need to be reviewed, and to standardize the document format and language to adapt to the subsequent review process; The processing module is used to perform vectorization and fragmentation on the preprocessed official documents based on the retrieval enhancement-generated RAG, so as to obtain document vector fragments and fragmented data. The building module is used to build a text analysis module and an automatic rule engine module that integrate large models. The text analysis module performs syntactic and semantic analysis on the fragmented data of official documents through natural language processing, while the automatic rule engine module performs rule matching and detection based on a preset official document review rule library. The review module is used to conduct in-depth reviews of official documents based on machine learning review models, combined with text analysis results and rule matching results, to identify potential inconsistencies, errors and compliance issues. When uncertainty exceeds the uncertainty threshold or critical document content occurs during the review process, manual review is triggered, prompting reviewers with the content that needs to be reviewed and providing decision support tools. The receiving module is used to receive review opinions submitted by reviewers and users through the feedback interface, and to iteratively optimize the rule base of the machine learning review model and the automatic rule engine module based on these opinions. The output module is used to output the review results of the official document content. If the review is approved, an approval report is generated; if the review is not approved, the problem type and location are marked and feedback is sent to the relevant personnel.

[0007] Thirdly, embodiments of this application also provide an electronic device, including: a processor, a storage medium, and a bus, wherein the storage medium stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor communicates with the storage medium via the bus, and the processor executes the machine-readable instructions to perform the document review method based on a large model as described in any of the first aspects.

[0008] Fourthly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the document review method based on a large model as described in any of the first aspects.

[0009] The embodiments of this application have the following beneficial effects: Through a collaborative end-to-end process involving preprocessing, RAG vectorization and sharding, text analysis and rule engine integration with large models, deep machine learning review, manual review, feedback optimization, and result output, this system not only achieves multi-dimensional coverage of document review from format standardization to semantic compliance, and from machine intelligence to human-machine collaboration, significantly improving review efficiency and accuracy, and effectively identifying inconsistencies, errors, and compliance issues, but also continuously optimizes review capabilities through a feedback iteration mechanism to adapt to the review needs of different types of documents while ensuring data security. This provides a highly efficient, accurate, intelligent, and sustainably iterative solution for document review, fully meeting the requirements of document review in terms of standardization, reliability, adaptability, and security. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a flowchart illustrating steps S101-S106 provided in the embodiments of this application; Figure 2a This is a flowchart illustrating steps S201-S203 provided in the embodiments of this application; Figure 2b This is a schematic diagram illustrating the vectorization and fragmentation of the document under review using RAG technology, as provided in an embodiment of this application. Figure 3 This is a flowchart illustrating steps S301-S303 provided in the embodiments of this application; Figure 4 This is a flowchart illustrating steps S401-S403 provided in the embodiments of this application; Figure 5 This is a flowchart illustrating steps S501-S503 provided in the embodiments of this application; Figure 6a This is a flowchart illustrating steps S601-S602 provided in the embodiments of this application; Figure 6b This is a diagram of the safety barrier architecture provided in an embodiment of this application; Figure 7 This is a schematic diagram of the structure of the document review device based on a large model provided in the embodiments of this application; Figure 8 This is a schematic diagram of the composition structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the accompanying drawings in this application are for illustrative and descriptive purposes only and are not intended to limit the scope of protection of this application. Furthermore, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in this application illustrate operations implemented according to some embodiments of this application. It should be understood that the operations in the flowcharts may not be implemented in sequence, and steps without logical contextual relationships may be reversed or implemented simultaneously. In addition, those skilled in the art, guided by the content of this application, may add one or more other operations to the flowcharts, or remove one or more operations from the flowcharts.

[0013] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0014] Furthermore, the described embodiments are merely some, not all, of the embodiments of this application. The components of the embodiments of this application described and illustrated herein can typically be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0015] In the following description, the terms "first, second, third" are used merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, third" may be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0016] It should be noted that the term "comprising" will be used in the embodiments of this application to indicate the presence of the features declared thereafter, but does not exclude the addition of other features.

[0017] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application and is not intended to limit this application.

[0018] See Figure 1 , Figure 1 This is a flowchart illustrating steps S101-S106 of the document review method based on a large model provided in this application embodiment, which will be combined with... Figure 1 Steps S101-S106 shown will be explained.

[0019] In step S101, preprocessing operations are performed on the official documents to be reviewed to standardize the document format and language and adapt them to the subsequent review process.

[0020] Here, to address potential format differences in official documents from different sources (such as the commonly used GB / T9704-2012 format for official documents, and custom formats for internal corporate reports), a format parsing tool (such as converting and adapting documents in Word, PDF, Excel, and other formats) is used to uniformly convert the documents into an intermediate format (such as a standardized text format) that meets the system's review requirements, thus avoiding missed or false checks due to format incompatibility.

[0021] Standardize the language used in official documents, including correcting typos, unifying the use of professional terminology (such as unifying "laws and regulations" with "laws and decrees," and correcting "budget" with "pre-budget"), and adjusting word order (such as changing "in the process of policy implementation, the handling of violations" to "the handling of violations in the process of policy implementation"). This ensures that the language in official documents conforms to the normative requirements of official document writing and reduces interference from non-standard language in subsequent semantic analysis.

[0022] In step S102, the preprocessed official document is vectorized and segmented based on the retrieval enhancement-generated RAG to obtain document vector fragments and segmented data.

[0023] The advantage of RAG technology lies in its combination of retrieval and generation capabilities. Targeting the characteristics of official documents, which are highly specialized and require precise matching of review rules, it transforms the official document text into a machine-understandable vector form through vectorization. At the same time, it breaks down long documents into short segments through fragmentation, avoiding the problems of low analysis efficiency and incomplete semantic understanding caused by excessively long documents.

[0024] The vectorization model (such as BERT, RoBERTa, etc., which are pre-trained based on official document domain corpora) is called to convert the preprocessed official document text into high-dimensional vectors. The vector dimension needs to be adapted to the subsequent vector database storage and retrieval requirements (such as 512-dimensional, 768-dimensional, etc.) to ensure that the vector can accurately represent the semantic information of the official document text and provide a foundation for subsequent similarity retrieval.

[0025] Set segmentation rules according to the content logic of the official document (such as chapter division, paragraph structure, theme switching, etc.), and break the official document into multiple segments (such as each segment is 500-1000 words, or it is divided into segments according to chapters such as "1. General Provisions - 2. Specific Requirements - 3. Implementation Steps"). Each segment must contain a complete semantic unit to avoid semantic breaks caused by segmentation. At the same time, add identification information (such as segment number, chapter to which it belongs, etc.) to each segment to facilitate the traceability of subsequent review results.

[0026] In step S103, a text analysis module and an automatic rule engine module that integrate the large model are constructed. The text analysis module performs syntactic and semantic analysis on the document fragment data through natural language processing, while the automatic rule engine module performs rule matching detection based on a preset document review rule base.

[0027] Based on the natural language processing capabilities of large models (such as the GPT series, ERNIE, and other finely tuned models adapted to the official document domain), multi-level analysis is performed on fragmented official document data: Detect grammatical errors in official documents, such as improper subject-verb-object collocation (correcting "the department implements the policy" to "the department implements the policy") and incorrect use of punctuation marks (mixing up book titles and quotation marks, misusing commas and semicolons, etc.).

[0028] To gain a deep understanding of the semantic logic of official documents, identify semantic contradictions (such as the contradiction between "the budget amount is 1 million yuan" in the preceding text and "the total budget is 1.5 million yuan" in the following text) and semantic ambiguity (such as the lack of a clear time range for "recently completed projects"). At the same time, analyze the relevance between the content of the official documents and the review objectives (such as whether policy documents conform to the policy guidance of higher authorities and whether business proposals match the company's development plan).

[0029] The module uses a pre-defined document review rule base as its core, which covers various hard standards for document review. It then uses rule matching algorithms (such as forward matching, reverse matching, and fuzzy matching) to perform rule checks on the segmented document data. For different types of document fragments, the system automatically matches the corresponding rule category in the rule base (e.g., policy documents are prioritized to match legal compliance rules, and corporate announcements are prioritized to match word accuracy rules). If a document fragment is detected to violate a rule (e.g., containing sensitive words or not conforming to writing standards), the system records the violation rule number, the location of the violation content, and the degree of violation, providing a basis for subsequent in-depth review.

[0030] In step S104, based on the machine learning review model, combined with the text analysis results and rule matching results, the document content is deeply reviewed to identify potential inconsistencies, errors and compliance issues. When uncertainty exceeds the uncertainty threshold or key document content occurs during the review process, manual review is triggered, prompting the reviewers with the content that needs to be reviewed and providing auxiliary decision-making tools.

[0031] The model takes text analysis results and rule matching results as input, and combines them with historical review data in the field of official documents to conduct a multi-dimensional and in-depth evaluation of the content of official documents: Content consistency audit: Check the logical consistency of the overall content of the official document (such as the connection logic between chapters and the consistency of data) and its consistency with external standards (such as whether it complies with industry norms).

[0032] Error identification: In addition to the basic errors detected by the text analysis module, further identify hidden errors (such as the incomplete time description caused by not specifying the month in "the project period is 2024-2025").

[0033] Compliance assessment: For the compliance requirements involved in the official document (such as data privacy compliance, industry regulatory compliance), assess whether there are any compliance risks in the content of the document (such as the failure to anonymize user privacy data in the company report).

[0034] The model uses a pre-set uncertainty assessment algorithm (such as an assessment based on predicted probability; if the model's predicted probability for a certain audit result is less than 80%, it is determined that the uncertainty is greater than the threshold). When uncertainty is detected to be greater than the preset threshold, the model automatically marks the content that needs to be reviewed. Through keyword matching (such as keywords such as "major decision", "core data", "legal responsibility") and content importance assessment (such as the core clauses and conclusions of the document), the model identifies key document content and triggers manual review. The model also pushes the necessary tools for review to the reviewers to ensure that manual review is efficient and accurate.

[0035] In step S105, the review opinions submitted by reviewers and users through the feedback interface are received, and the rule base of the machine learning review model and the automatic rule engine module is iteratively optimized based on the opinions.

[0036] The system sets up feedback interfaces (such as the "Submit Feedback" button on the reviewer's interface and the feedback form on the user's end) to collect feedback, including: machine review results errors (such as misjudging compliant content as non-compliant, or missed detection errors), unreasonable rules (such as a rule being too strict or outdated), and review efficiency issues (such as the review of a certain type of document taking too long). The feedback is then categorized and organized (such as by "model error," "rule issue," or "functional requirement"), and key information is extracted (such as the document fragments corresponding to the erroneous review results and the numbers of unreasonable rules) to provide a basis for subsequent optimization.

[0037] In step S106, the review result of the document content is output. If the review is passed, a review pass report is generated. If the review is not passed, the problem type and location are marked and feedback is sent to the relevant personnel.

[0038] If the review is approved, a review report will be generated, including the review conclusion ("Review Approved"), review dimensions (such as "Syntactic Compliance", "Semantic Consistency", "Rule Matching", "Compliance Standard Achieved"), review time, and reviewer (if manual review is involved). The report format must meet the requirements of official document management and can be directly used for document archiving or circulation. If the review is not approved, the problem type (such as "Sensitive Word Violation", "Syntactic Error", "Compliance Risk"), problem location (such as document page number, paragraph number, specific sentence), and problem description (such as "Page 3, Paragraph 2 contains the sensitive word 'XXX', violating Article 5 of the 'Rules for the Management of Sensitive Words in Official Documents') will be provided, along with preliminary modification suggestions (such as "It is recommended to replace it with 'XXX'"). Feedback will be given to the document drafter or relevant person in charge to facilitate subsequent modification and improvement.

[0039] In some embodiments, see Figure 2a , Figure 2a This is a flowchart illustrating steps S201-S203 provided in the embodiments of this application. The vectorization and fragmentation of the preprocessed official document based on the retrieval enhancement-generated RAG can be achieved through steps S201-S203, which will be explained in conjunction with each step.

[0040] In step S201, the vector database set is loaded, and historical vector data related to document review is obtained as a reference.

[0041] The vector database (such as Milvus, Elasticsearch, and other databases adapted for document vector storage) pre-stores historical review data vectors in the document domain (such as vectors of approved / unapproved documents, vectors of typical error fragments, and vectors of compliant standard text). The system loads the corresponding data set through the database interface. The loading of the set must meet the requirements of data integrity (such as no missing vectors) and data timeliness (such as excluding outdated historical data) to ensure the reliability of subsequent reference data. The acquired historical vector data is used to compare the similarity with the document vector of the current document to help determine the rationality of the current document content (for example, if the similarity between a certain segment vector of the current document and the historical violation fragment vector is higher than 90%, it is initially determined that the segment has a violation risk and needs to be reviewed more closely). At the same time, it provides verification for the accuracy of the vectorization conversion results (if the current document vector differs too much from the vector of compliant documents of the same type, it is necessary to check whether the vectorization model is compatible).

[0042] In step S202, the vectorization model is invoked to convert the preprocessed official document into document vectors. At the same time, the official document is segmented according to the preset segmentation rules to obtain multiple document segmentation data.

[0043] Here, the vectorized model used needs to be fine-tuned using corpora in the official document domain (e.g., pre-trained using corpora of official documents, corporate reports, and files) to ensure that the model can accurately understand the professional terminology and writing logic of official documents, avoiding vector representation bias caused by insufficient adaptability of general models to the official document domain; the segmentation rules need to be set in conjunction with the document type, content length, and semantic logic, for example: Document: Divide into sections according to the chapter structure of "General Provisions - Chapter Clauses - Supplementary Provisions" to ensure that each section contains complete clause content; Enterprise reports are segmented according to the logical structure of "Summary - Core Data - Analysis Conclusions - Recommendations" to avoid disrupting data coherence through splitting; For short official documents (such as notices and announcements): If the length is less than 1,000 words, they can be divided into paragraphs or not, to avoid semantic breaks caused by excessive splitting.

[0044] In step S203, the fragmented data is filtered and optimized according to the type and complexity of the official documents, invalid fragments are removed, and valid fragmented data is retained for subsequent review.

[0045] Screening criteria are set based on the type of official document (e.g., documents, business proposals, project plans) and the complexity of its content (e.g., simple notices, complex research reports): Type-related filtering: exclude segments that are not related to the core content of the official document (such as irrelevant attachment segments in policy documents, and duplicate table segments in corporate reports). Complexity filtering: Segments with low content complexity (such as those containing only "attachment list" or "contact information") are removed if no review is required; segments with high content complexity (such as those containing core data or key decisions) are retained and marked as segments requiring key review. Optimize operations: Standardize the format of retained valid segments (such as unifying the text encoding and line break format of segments) and remove redundant information (such as deleting whitespace characters and duplicate expressions in segments) to ensure that the segment data is concise and standardized, thereby improving the efficiency of subsequent review.

[0046] For example, see Figure 2b , Figure 2b This is a schematic diagram illustrating the principle of vectorization and fragmentation of the audited document using RAG technology, as provided in the embodiments of this application. Figure 2b As shown, this process realizes the entire lifecycle from user question submission to obtaining the answer. The core lies in combining Retrieval Augmentation (RAG) technology with large model inference to ensure the accuracy and relevance of the answer.

[0047] Users submit questions (such as "What is a specific clause?") through the "Web frontend / interface," which then enters the system processing stage. The system performs semantic understanding and intent recognition on the user's question to determine the question type (such as policy inquiry, business consultation, etc.), preparing for subsequent retrieval and generation. The system retrieves relevant data from two dimensions to ensure comprehensive information: RAG-related fragments: Retrieve knowledge base document fragments related to the question semantics from Milvus & Elasticsearch databases (vector database + full-text search database) through vector similarity (based on vector data pre-stored in the "slice vectorization" step).

[0048] Plugin-related data: Retrieve plugin interface data related to the problem from the OceanBase database (such as API information of business systems, preset rules, etc.).

[0049] The recalled "RAG-related fragments" and "plugin-related data" are sorted by relevance, and the Top N pieces of information that best match the user's question are selected to provide high-quality input for large model inference.

[0050] Input the "user question + top-ranked document / plugin information" into the inference model (such as a large language model adapted to the business scenario), and the model will combine the retrieved information to generate accurate and detailed answers.

[0051] The final answer is returned to the user through the "Web front-end / interface", completing a question-and-answer interaction.

[0052] This process enables the import and preprocessing of "knowledge base documents" and "plugins," providing a data foundation for retrieval in the question-and-answer phase.

[0053] Users can upload knowledge base documents (such as policy documents and product manuals) or plugins and knowledge base documents (such as functional plugins for business systems and customized knowledge documents) through the "Web front-end / interface".

[0054] Uploaded files are stored in the system to ensure long-term data accessibility. Plugin-type data is split into interfaces, and knowledge base documents are sliced ​​(e.g., into smaller segments by chapter or paragraph) to facilitate subsequent vectorization and retrieval. The sliced ​​data is then vectorized to generate semantic vectors recognizable by vector databases. Knowledge base document data: The knowledge base document slices are converted into vectors using vectorization models (such as BERT, Sentence-BERT) and stored in Milvus & Elasticsearch databases.

[0055] Plugin interface data: Convert plugin interface information into structured data and store it in the OceanBase database. Vectors can also be generated as needed to assist in retrieval.

[0056] The above embodiments solve the "illusion problem" of large models (generating unfounded content) by combining "RAG + large model + multi-source database". At the same time, they realize dynamic updates of the knowledge base (automatic preprocessing after document / plugin upload) and accuracy of question answering (retrieval enhancement + model reasoning). They are suitable for intelligent question answering scenarios that require the combination of enterprise private knowledge and business data (such as government affairs consultation, internal enterprise knowledge question answering, etc.).

[0057] In some embodiments, see Figure 3 , Figure 3 This is a flowchart illustrating steps S301-S303 provided in the embodiments of this application. The preset document review rule base is constructed through steps S301-S303, and will be explained in conjunction with each step.

[0058] In step S301, document writing standards, terminology standards, sensitive word lists, and legal compliance requirements from different industries and organizations are collected as basic data for the rule base.

[0059] The rule base's basic data collection covers various standards and requirements involved in official document review, specifically including: Writing guidelines: such as the requirements for document structure, font format, punctuation, and title expression in the "Regulations on Official Document Processing" and "Official Document Format"; Terminology standards: Standards for official terminology used in official documents of various industries and organizations (such as expression norms, terminology requirements for "risk control" and "compliance review" in financial enterprises), and lists of prohibited / restricted words; List of sensitive words: covering privacy-sensitive words (such as the expression of personal ID numbers and mobile phone numbers), industry-sensitive words, and internal organizational sensitive words (such as the expression of strategic data that is not disclosed by the company). Compliance requirements: Laws and regulations, industry regulatory provisions, and internal rules and regulations of the organization related to the content of the official document; The comprehensiveness of basic data is ensured through methods such as manual sorting (e.g., organizing professional auditors to sort out current standards), system crawling (e.g., obtaining publicly available official documents and standard documents from official websites and industry association platforms), and extraction of historical audit data (e.g., summarizing implicit rules that are not explicitly written into the standards but must be followed from past audit cases).

[0060] In step S302, the basic data is classified and organized into multiple rule categories, such as writing standardization rules, word accuracy rules, sensitive word detection rules, and legal compliance rules.

[0061] Here, rules are categorized by review dimensions and application scenarios to ensure targeted rule invocation. The specific categories are as follows: Document formatting rules: rules concerning the format, structure, and expression logic of official documents (such as "the title of an official document must include the issuing authority, subject matter, and document type, such as 'Notice from XX regarding XX matters'" and "the first line of each paragraph in the body text must be indented by 2 characters"). Rules on word accuracy: Rules for the use of terminology and vocabulary in official documents (such as "'must' is used for mandatory requirements, 'should' is used for obligatory requirements, and they shall not be used interchangeably" and "Internet slang and non-standard abbreviations are prohibited"). Sensitive word detection rules: Rules for the use of sensitive words (such as "The use of sensitive word 'XXX' is prohibited" and "Statements involving personal privacy information must be anonymized, such as stating 'ID number 110101XXXXXX' as 'ID number 110101********'"). Compliance rules: Rules concerning the compliance of official document content (such as "The content of the document must comply with Article X" or "The data disclosure in the company's report must comply with the requirements of the 'Information Disclosure Management Measures'"). The basic data is categorized using a tagging method, and category tags (such as "writing style - title format" and "sensitive word detection - privacy protection") are added to each rule to facilitate subsequent rule retrieval and invocation.

[0062] In step S303, rule encoding technology is used to convert various rules into machine-recognizable rule codes, which are then entered into the rule base and a rule retrieval index is established to facilitate subsequent rule matching and invocation.

[0063] Here, rules expressed in natural language are converted into machine-readable and executable code logic. For example, regular expression encoding (e.g., "matching sensitive words such as 'XXX' and 'YYY'; if they appear in the official document text, a violation warning will be triggered"); format validation code encoding (e.g., "checking whether the official document title contains the keyword 'about'; if not, it is determined to be a title format violation"). The encoding must ensure that the rule logic is unambiguous and has high execution efficiency, avoiding rule matching errors due to code vulnerabilities; indexing technology (such as inverted index and hash index) is used to build a retrieval index for the rule base. The index keywords include rule category, applicable document type, and core rule requirements. For example, a "policy document - legal compliance" index is built for "rules applicable to legal compliance of policy documents". When reviewing policy documents, the system can quickly locate and call relevant rules through the index, improving rule matching efficiency.

[0064] In some embodiments, see Figure 4 , Figure 4 This is a flowchart illustrating steps S401-S403 provided in the embodiments of this application. The machine learning audit model is trained through steps S401-S403, and will be explained in conjunction with each step.

[0065] In step S401, sample data of official documents with marked audit results are collected, covering samples of different document types, different error types, and compliance issues.

[0066] The sample must cover a variety of official document review scenarios to ensure the model's generalization ability, specifically including: Document types: policy documents, project plans, corporate reports, business proposals, notices and announcements, meeting minutes, and other common official documents.

[0067] Error types: various errors that may occur during the review process, such as grammatical errors, semantic contradictions, word violations, sensitive word issues, formatting errors, and compliance risks.

[0068] The review results include approved samples (documents with no errors or whose errors have been corrected) and unapproved samples (documents with various errors). The ratio of the two types of samples should be balanced (e.g., 1:1 to 1:1.5) to avoid the model being biased towards the prediction of a certain review result.

[0069] Each sample must be annotated in detail, including: sample document text, document type, error location (if any), error type (if any), error description (if any), review conclusion (pass / fail), annotator, and annotation time. Annotations must ensure accuracy (e.g., error type matches error content) and consistency (different annotators provide consistent annotation results for the same sample). Annotation quality can be ensured through cross-annotation by multiple people and verification of annotation results (e.g., annotation accuracy rate must reach 95% or higher).

[0070] In step S402, data augmentation processing is performed on the sample data, including text replacement, word order adjustment, and adding similar samples to expand the size of the training dataset.

[0071] The purpose of augmentation is to expand the size of the training dataset through data augmentation, avoid overfitting caused by insufficient sample size, and improve the model's adaptability to different representations.

[0072] Without altering the semantics and error attributes of the official document, replace some non-core words (e.g., replace "conduct project research" with "perform project research", and replace "violate regulations" with "violate regulations"); adjust the word order of sentences (e.g., change "to ensure policy implementation, supervision needs to be strengthened" to "supervision needs to be strengthened to ensure policy implementation") to ensure that the semantics and error attributes remain unchanged after the adjustment; generate similar samples based on existing samples (e.g., for "sensitive word violation" samples, replace different sensitive words to generate new violation samples; for "format error" samples, adjust the position of the error format to generate new error samples).

[0073] The enhanced samples are validated to ensure that there are no semantic breaks, no new errors (such as no new syntax errors after text replacement), and that the error attributes are consistent with the original samples. Samples that pass the validation are included in the training dataset.

[0074] In step S403, the enhanced training dataset is input into the initial machine learning model, the model training parameters are set, and the gradient descent algorithm is used to iteratively train the model until the model's audit accuracy reaches the preset threshold, thus completing the model training.

[0075] Here, a model architecture suitable for official document text analysis (such as Transformer, BERT, GPT, etc.) is selected. The initial model needs to be pre-trained on general texts, and then fine-tuned using official document domain corpora to ensure that the model has basic semantic understanding capabilities for official documents. Reasonable parameters are set according to the model architecture and sample characteristics, including: Batch size: Set according to hardware computing power (such as GPU memory), usually 16, 32, 64, etc., to ensure stability during training; Learning Rate: The initial learning rate is set to 1e-5 to 1e-4, and a learning rate decay strategy (such as cosine annealing decay) is adopted to avoid excessive learning rate in the later stages of training, which may cause model oscillation. Training epochs: Set to 10-20 epochs, and combine with an early stopping strategy. Stop training when the validation set accuracy does not improve for 3 consecutive epochs to avoid overfitting.

[0076] The model is iteratively trained using gradient descent algorithms (such as Adam and SGD optimizers). After each training round, the model's review accuracy is evaluated on the validation set (e.g., accuracy in correctly identifying error types and accuracy in correctly determining review results). Training stops when the accuracy reaches a preset threshold (e.g., above 90%, with the specific threshold set according to the accuracy requirements of document review; for example, the threshold for documents involving legal compliance can be set to above 95%). The trained model needs to be finally validated on the test set to ensure that the model can maintain high accuracy on unseen samples.

[0077] In some embodiments, the decision support tools provided to auditors include at least one of the following tools: The document issue location tool is used to mark the specific page number, paragraph, and line number of the content requiring review in the document; The historical case reference tool is used to search and display historical document review cases and their processing results that are similar to the current content that needs to be reviewed; The rule explanation tool is used to provide detailed explanations of the violations detected by the automatic rule engine module, including the meaning of the rule, applicable scenarios, and consequences of the violation.

[0078] The document issue location tool is used to quickly pinpoint the specific location of content requiring review within a document, avoiding time-consuming manual searches by reviewers and improving review efficiency. It accurately marks the location of content requiring review using page numbers, paragraph numbers, and sentence numbers (e.g., "page 5, paragraph 3, sentence 2"). This location is also visually displayed in the document preview interface through highlighting and red highlighting. Reviewers can click on the mark to jump to the corresponding content. In addition to the marked content, the tool also displays the context (e.g., the preceding three sentences, the following three sentences) to help reviewers understand the semantic background and avoid misjudgments caused by viewing content in isolation. When the machine reviewer marks a passage as having "semantic contradictions" or "suspected violations of sensitive words," reviewers can use this tool to quickly locate the passage and determine whether the problem truly exists based on the context.

[0079] The historical case reference tool provides auditors with similar historical cases for reference, helping them to standardize their audits and reduce subjective judgment discrepancies. Based on the semantic features (such as keywords and semantic vectors) of the content currently requiring review, it retrieves historical audit cases stored in the vector database, filtering out cases with a similarity higher than 85% (e.g., if the current review content is "unclear project budget description," it retrieves historical audit cases with "vague budget description"). It displays the document type, content requiring review, audit conclusion (e.g., "determined to require modification, it is recommended to clarify budget details"), audit basis (e.g., "violation of Article 8 of the 'Rules for Document Data Expression'"), and an example of the modified content. Auditors can refer to the handling methods of historical cases to make judgments on the current content. When auditors are unsure about the judgment of a certain content requiring review (e.g., "uncertainty about whether a certain expression is a sensitive word"), they can improve the accuracy of their judgment by searching for similar historical cases and referring to past audit conclusions.

[0080] The rule explanation tool is used to explain the violation rules cited by machine reviewers to auditors, helping them understand the meaning and applicable scenarios of the rules and ensuring that the review judgments comply with the rule requirements. When machine review marks content as violating a rule, the tool displays the complete description of the rule, the basis for its formulation (e.g., "formulated according to Article X of Document XX"), the applicable scenario (e.g., "applies to all official documents involving this description, including documents and meeting minutes"), and an explanation of the consequences of the violation: explaining the possible consequences of violating the rule (e.g., "may cause ambiguity in policy interpretation, affecting the effectiveness of document implementation" or "may violate XX, posing a compliance risk"), helping auditors to pay attention to rule implementation; and providing examples of expressions for compliance and violation (e.g., compliance: "implementing relevant policies"; violation: "providing a personalized interpretation of the policy"), helping auditors intuitively understand the boundaries of the rules. When machine review cites a rule that auditors are unfamiliar with (e.g., industry-specific compliance rules), auditors use this tool to understand the rule content and determine whether the current content actually violates the rule.

[0081] In some embodiments, see Figure 5 , Figure 5 This is a flowchart illustrating steps S501-S503 provided in the embodiments of this application. Iterative optimization of the rule base of the machine learning review model and the automatic rule engine module can be achieved through steps S501-S503, which will be explained in conjunction with each step.

[0082] In step S501, the feedback opinions are classified and statistically analyzed to determine the frequency and severity of the review issues corresponding to different types of opinions.

[0083] The statistics are categorized according to the nature and scope of impact of feedback, and key indicators for each category are also statistically analyzed to provide a basis for prioritization. The categorization dimensions include: model-related feedback (e.g., "model misjudged compliant content as non-compliant," "model missed semantic contradictions," "model has low accuracy in reviewing certain types of documents"); rule-related feedback (e.g., "ambiguous rule descriptions leading to misjudgments," "rules are outdated (e.g., not complying with the latest laws and regulations)," "rules are missing (e.g., no corresponding review rules for certain types of new documents)"); function-related feedback (e.g., "slow review speed," "inconvenient operation of the review tool"); and statistical indicators: frequency (e.g., "model misjudgment" appeared 20 times this month), severity (e.g., "missed compliance risk issues" is high severity, "format error misjudgment" is medium severity), and document type involved (e.g., "model has low accuracy in reviewing business proposals").

[0084] Based on the statistical results, optimization priorities are determined. For example, opinions with high severity and high frequency (such as "compliance risk issues that the model missed") are optimized first, while opinions with low severity and low frequency (such as "misjudgment of individual format errors") are optimized gradually in the future.

[0085] In step S502, for the machine learning review model, the official document data corresponding to the feedback opinions are used as new training samples to incrementally train the model and adjust the model parameters to improve the review accuracy.

[0086] Compared to full retraining, incremental training uses only the newly added samples corresponding to the feedback, reducing training time and computational cost while preventing the model from forgetting its existing review capabilities. The official document data corresponding to the feedback (such as document fragments misjudged by the model or document texts with missed errors) is labeled (with error type and correct review conclusion) and used as new training samples. These new samples are then input into the already trained machine learning review model, and incremental training algorithms (such as gradient-based incremental training) are used to adjust some model parameters (such as output layer and last few network parameters) to retain the model's original correct review capabilities while correcting errors. After incremental training, the model's accuracy is evaluated on the validation set (including both existing and new samples). If the accuracy improves (e.g., from 90% to 92%) without significant overfitting, model optimization is complete. If the accuracy does not improve or there is overfitting, training parameters (such as learning rate and the number of new samples) need to be adjusted and retrained. When feedback such as "model misjudgment" or "model missed detection" occurs frequently, incremental training is used to optimize the model and improve review accuracy.

[0087] In step S503, for the rule base of the automatic rule engine module, corresponding rules are added, modified, or deleted according to feedback, and the rule retrieval index is updated.

[0088] Rule update types include: New rule: When feedback indicates that "there is no corresponding review rule for a certain type of new document" (e.g., there is no relevant review rule for a newly added document type "digital transformation solution"), collect the review standards for this type of document (e.g., industry standards, organizational requirements), formulate new rules according to the construction process, and enter them into the rule base; Modify the rules: When feedback points out that the rules are "vague" or "too strict", revise the wording of the rules (e.g., change "official document data must be accurate" to "core data of official documents must be accurate to two decimal places, and non-core data must be accurate to integer places") and adjust the scope of application of the rules (e.g., change "applies to all official documents" to "applies to corporate reports and project proposals"). Deleting rules: When feedback indicates that a rule is "outdated" (e.g., the laws and regulations upon which a rule is based have been repealed), delete the rule to avoid matching errors caused by redundancy in the rule base.

[0089] After a rule is added, modified, or deleted, the retrieval index of the rule base is updated synchronously (e.g., adding corresponding category tags and index keywords for new rules, and removing the corresponding index for deleted rules) to ensure that the latest rule can be accurately located when the rule is called. After the rule base is updated, the accuracy of rule matching is verified through a test set (containing official document samples that need to apply the new rules and the modified rules). If the rule matching error rate is less than 3%, the rule base optimization is completed; if the error rate is high, the rule content and coding logic need to be re-examined.

[0090] In some embodiments, see Figure 6a , Figure 6a This is a flowchart illustrating steps S601-S602 provided in the embodiments of this application. In the document preprocessing stage and the review result output stage, data security protection steps S601-S602 are also included, which will be explained in conjunction with each step.

[0091] In step S601, during the preprocessing stage, data encryption technology is used to encrypt the official documents to prevent information leakage during document transmission and storage.

[0092] Here, to prevent official documents from being illegally obtained or tampered with during transmission (e.g., from the document drafting end to the review system) and storage (e.g., temporarily stored on the system server after preprocessing), and to protect the confidentiality and integrity of the document content, encryption technologies that comply with information security standards are adopted, specifically including: Encryption during transmission: The SSL / TLS protocol is used to encrypt the document transmission process to ensure that the transmitted data is not eavesdropped or intercepted; Storage encryption: The pre-processed official documents are encrypted and stored using a symmetric encryption algorithm (such as AES-256). The encryption key is managed uniformly through a key management system (KMS) and is changed regularly (such as once a month) to avoid the risk of document decryption due to key leakage.

[0093] The encryption scope covers the entire content of the official document (including text, tables, and attachments), and also encrypts the document's identification information (such as document number and drafter information) to ensure the security of all data. Only authorized reviewers (through account password, identity authentication such as facial recognition, or USBKey authentication) can decrypt the document during the review process. The decrypted data is only temporarily stored in memory, and the decrypted data in memory is immediately cleared after the review is completed to avoid data residue.

[0094] In step S602, during the audit result output stage, the personal privacy information and sensitive business data in the audit results are desensitized, and the sensitive information is removed or replaced before the final report is generated and fed back.

[0095] Here, to prevent the leakage of personal privacy information and sensitive business data in the audit results (such as unauthorized personnel viewing the audit results during the circulation process), it is necessary to anonymize the data. The objects of anonymization include personal privacy information: including personal name, ID number, mobile phone number, email address, home address, etc. (e.g., "Auditor: Zhang San" is anonymized to "Auditor: Zhang XX", and "ID number: 110101XXXXXX" is anonymized to "ID number: 110101********"); sensitive business data: including unpublished strategic data of the enterprise (such as "2025 market expansion plan"), core financial data (such as "annual net profit of 10 million yuan"), and confidential information (such as internal decision-making data), etc. (e.g., "annual net profit of 10 million yuan" is anonymized to "annual net profit of XXX million yuan", and "internal decision-making data" is anonymized to "confidential information, encrypted").

[0096] For desensitization methods, the following can be used: Replacement and desensitization: For personal information such as names and mobile phone numbers, use the method of "replacing some characters with XX" (e.g., "Mobile phone number: 138XXXX1234"); Masking and desensitization: For sensitive business data and classified information, the method of "masking key content + explanation" is adopted (e.g., "2025 Market Expansion Plan: [Classified, see internal document for details]"). Encryption and desensitization: Highly sensitive data (such as core financial data) is encrypted using asymmetric encryption algorithms (such as RSA), and only authorized personnel (such as the company's financial officer or personnel with access to confidential information) can decrypt and view it.

[0097] After the anonymization process, check whether the sensitive information in the audit results has been completely anonymized (e.g., no complete ID number or mobile phone number), and whether the anonymized information does not affect the understanding of the audit results (e.g., "Auditor: Zhang XX" still clearly identifies the person responsible for the audit, and "Annual net profit of XXX million yuan" still reflects the "financial data compliance" in the audit conclusion), to ensure a balance between the anonymization effect and the usability of the audit results.

[0098] For an example, please see Figure 6b , Figure 6b This is a diagram of the safety barrier architecture provided in an embodiment of this application, such as... Figure 6b As shown, through modules such as "Input Security Agent," "Output Security Agent," and "Security Barrier," end-to-end security control over the input and output of large models is achieved. Specifically: Perform compliance review on user input (such as questions and instructions), checking for sensitive information or prohibited content (such as politically sensitive words or illegal information). Identify and block "prompt attacks" targeting large models (such as inducing the model to output harmful or confidential content through malicious prompts). If the input fails the review, directly "reject" the request (marked "Review failed, rejected"); if the review passes, proceed to the "security guardrail" and "user security protection strategy" stages, where further verification is performed using a "sensitive word database".

[0099] As the core security layer, the security barrier manages input and output with global security policies (such as access control, traffic monitoring, and abnormal behavior interception). User security protection policies are based on a "sensitive word database" (storing politically sensitive words, privacy-related words, and prohibited terms) to perform fine-grained security rule matching on input and output, ensuring that content meets security requirements.

[0100] Output security agent chain includes: Content review: The output content generated by the large model is reviewed again for compliance to prevent the model from outputting illegal information due to misjudgment.

[0101] Personal privacy information anonymization: Anonymize personal privacy information (such as name, ID number, mobile phone number, etc.) in the output (e.g., replace with "XXX" "*" etc.).

[0102] Output after verification: After the final output has undergone "verification processing", it is returned to the user through the "input & output" process to ensure that the output content is safe, compliant and privacy is well protected.

[0103] The above embodiments prevent malicious content and attack commands from entering the system through full-lifecycle control of "input-link-output," avoiding model "contamination" or inducement to output harmful information; they achieve fine-grained security policy execution through security guardrails and sensitive word libraries; and the output side ensures that the model outputs no illegal content and no privacy leaks. This is suitable for large-scale model application scenarios with high security and compliance requirements, such as those in finance and enterprises (e.g., intelligent document review, intelligent financial customer service, government Q&A systems, etc.).

[0104] In summary, the embodiments of this application have the following beneficial effects: This method achieves standardized adaptation of document formats and language through preprocessing, enhances the structure and accuracy of review data through vectorization and segmentation using RAG technology, and implements dual verification of deep semantic understanding and rigorous rule matching using a text analysis module that integrates large models and an automatic rule engine module. Further, a machine learning review model is used for in-depth review, coupled with a human review mechanism to ensure reliability. A feedback optimization mechanism enables dynamic iteration of the model and rule base, and comprehensive data security measures ensure the security of document information. This approach not only significantly improves the efficiency and accuracy of document review, effectively identifying inconsistencies, errors, and compliance issues, but also adapts to the review needs of different types of documents through human-machine collaboration and continuous optimization, balancing intelligence and flexibility in the review process. It comprehensively meets the requirements of document review in terms of standardization, security, and adaptability, providing efficient and reliable technical support for document management in enterprises and other fields.

[0105] Based on the same inventive concept, this application also provides a document review device based on a large model, which corresponds to the document review method based on a large model in the first embodiment. Since the principle of the device in this application is similar to the above-mentioned document review method based on a large model, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be described again.

[0106] like Figure 7 As shown, Figure 7 This is a schematic diagram of the structure of the document review device 700 based on a large model provided in this application embodiment. The document review device 700 based on a large model includes: The preprocessing module 701 is used to perform preprocessing operations on official documents that need to be reviewed, and to standardize the document format and language to adapt to the subsequent review process. Processing module 702 is used to perform vectorization and fragmentation on the preprocessed official document based on retrieval enhancement to generate RAG, and obtain document vector fragments and fragmented data; Module 703 is used to build a text analysis module and an automatic rule engine module that integrate large models. The text analysis module performs syntactic and semantic analysis on the document fragment data through natural language processing, and the automatic rule engine module performs rule matching and detection based on a preset document review rule library. The review module 704 is used to conduct in-depth reviews of official document content based on machine learning review models, combined with text analysis results and rule matching results, to identify potential inconsistencies, errors and compliance issues. When uncertainty exceeds the uncertainty threshold or key document content occurs during the review process, manual review is triggered, prompting reviewers with the content that needs to be reviewed and providing decision support tools. The receiving module 705 is used to receive review opinions submitted by reviewers and users through the feedback interface, and to iteratively optimize the rule base of the machine learning review model and the automatic rule engine module based on these opinions. Output module 706 is used to output the review results of official document content. If the review is passed, an approval report is generated; if the review is failed, the problem type and location are marked and feedback is sent to the relevant personnel.

[0107] Those skilled in the art should understand that Figure 7 The functions of each unit in the large-scale model-based document review device 700 shown can be understood by referring to the relevant description of the aforementioned large-scale model-based document review method. Figure 7 The functions of each unit in the large-scale document review device 700 shown can be implemented by a program running on a processor or by specific logic circuits.

[0108] In one possible implementation, the processing module 702 generates a RAG based on retrieval enhancement and performs vectorization and fragmentation processing on the preprocessed official document, including: Load the vector database collection and obtain historical vector data related to official document review as a reference; The vectorization model is invoked to convert the preprocessed official documents into document vectors. At the same time, the official documents are divided into fragments according to the preset fragmentation rules to obtain multiple fragmented data of official documents. Based on the type and complexity of the official documents, the fragmented data is filtered and optimized, invalid fragments are removed, and valid fragmented data is retained for subsequent review.

[0109] In one possible implementation, the preset document review rule base is constructed in the following way: Collect official document writing standards, terminology standards, sensitive word lists, and legal compliance requirements from different industries and organizations as the basic data for the rule base; The basic data is categorized and organized into several rule categories, including rules on writing standards, rules on word accuracy, rules on sensitive word detection, and rules on legal compliance. By employing rule encoding technology, various rules are converted into machine-readable rule codes, which are then entered into the rule base and a rule retrieval index is established to facilitate subsequent rule matching and invocation.

[0110] In one possible implementation, the machine learning auditing model is trained in the following manner: Collect sample data of official documents with marked review results, covering samples of different document types, different error types, and compliance issues; Data augmentation is performed on the sample data, including text replacement, word order adjustment, and adding similar samples, to expand the size of the training dataset; The enhanced training dataset is input into the initial machine learning model, the model training parameters are set, and the gradient descent algorithm is used to iteratively train the model until the model's audit accuracy reaches the preset threshold, thus completing the model training.

[0111] In one possible implementation, the decision support tools provided to auditors include at least one of the following tools: The document issue location tool is used to mark the specific page number, paragraph, and line number of the content requiring review in the document; The historical case reference tool is used to search and display historical document review cases and their processing results that are similar to the current content that needs to be reviewed; The rule explanation tool is used to provide detailed explanations of the violations detected by the automatic rule engine module, including the meaning of the rule, applicable scenarios, and consequences of the violation.

[0112] In one possible implementation, the receiving module 705 iteratively optimizes the rule base of the machine learning auditing model and the automatic rule engine module, including: The feedback was categorized and statistically analyzed to determine the frequency and severity of the review issues corresponding to different types of feedback. For the machine learning review model, the official document data corresponding to the feedback opinions are used as new training samples to incrementally train the model and adjust the model parameters to improve the review accuracy. For the rule base of the automatic rule engine module, add, modify or delete relevant rules based on feedback, and update the rule retrieval index.

[0113] In one possible implementation, the output module 706 further includes data security protection steps during the document preprocessing stage and the review result output stage: During the preprocessing stage, data encryption technology is used to encrypt official documents to prevent information leakage during document transmission and storage. During the audit results output stage, personal privacy information and sensitive business data in the audit results are anonymized, sensitive information is removed or replaced, and then the final report is generated and fed back.

[0114] The aforementioned large-scale model-based document review device achieves standardized adaptation of document formats and language through preprocessing. Combined with RAG technology's vectorization transformation and segmentation processing, it enhances the structure and accuracy of review data. Leveraging a text analysis module and an automatic rule engine module that integrates the large-scale model, it achieves dual verification through deep semantic understanding and rigorous rule matching. Furthermore, it employs a machine learning review model for in-depth review, coupled with a human review mechanism to ensure reliability. Simultaneously, it utilizes a feedback optimization mechanism to achieve dynamic iteration of the model and rule base, and incorporates comprehensive data security protection measures to ensure document information security. This method not only significantly improves the efficiency and accuracy of document review, effectively identifying inconsistencies, errors, and compliance issues, but also adapts to the review needs of different types of documents through human-machine collaboration and continuous optimization. It balances the intelligence and flexibility of the review process, comprehensively meeting the requirements of document review in terms of standardization, security, and adaptability, providing efficient and reliable technical support for document management in enterprises and other fields.

[0115] like Figure 8 As shown, Figure 8 This is a schematic diagram of the composition structure of the electronic device 800 provided in the embodiments of this application. The electronic device 800 includes: The device includes a processor 801, a storage medium 802, and a bus 803. The storage medium 802 stores machine-readable instructions that can be executed by the processor 801. When the electronic device 800 is running, the processor 801 communicates with the storage medium 802 via the bus 803. The processor 801 executes the machine-readable instructions to perform the steps of the document review method based on a large model as described in the embodiments of this application.

[0116] In practical applications, the various components in the electronic device 800 are coupled together via bus 803. It can be understood that bus 803 is used to achieve communication between these components. In addition to a data bus, bus 803 also includes a power bus, a control bus, and a status signal bus. However, for clarity, in... Figure 8 The general labeled all buses as Bus 803.

[0117] The aforementioned electronic equipment achieves standardized adaptation of document formats and languages ​​through preprocessing, enhances the structure and accuracy of review data through vectorization and segmentation processing using RAG technology, and achieves dual verification of deep semantic understanding and rigorous rule matching by leveraging a text analysis module that integrates large models and an automatic rule engine module. Further, in-depth review is conducted through a machine learning review model, coupled with a human review mechanism to ensure review reliability. Simultaneously, a feedback optimization mechanism enables dynamic iteration of the model and rule base, and comprehensive data security protection measures ensure the security of document information. This method not only significantly improves the efficiency and accuracy of document review, effectively identifying inconsistencies, errors, and compliance issues, but also adapts to the review needs of different types of documents through human-machine collaboration and continuous optimization, balancing intelligence and flexibility in review. It comprehensively meets the requirements of document review in terms of standardization, security, and adaptability, providing efficient and reliable technical support for document management in enterprises and other fields.

[0118] This application also provides a computer-readable storage medium storing executable instructions. When the executable instructions are executed by at least one processor 801, the document review method based on a large model described in this application is implemented.

[0119] In some embodiments, the storage medium may be a magnetic random access memory (FRAM), a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a flash memory, a magnetic surface memory, an optical disc, or a compact disc read-only memory (CD-ROM), etc.; or it may be a device that includes one or any combination of the above-mentioned memories.

[0120] In some embodiments, executable instructions may take the form of a program, software, software module, script, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a standalone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0121] As an example, executable instructions may, but do not necessarily, correspond to files in the file system. They may be stored as part of a file that holds other programs or data, for example, in one or more scripts in a HyperText Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple collaborating files (e.g., a file that stores one or more modules, subroutines, or code sections).

[0122] As an example, executable instructions can be deployed to execute on a single computing device, or on multiple computing devices located in one location, or on multiple computing devices distributed across multiple locations and interconnected via a communication network.

[0123] The aforementioned computer-readable storage medium achieves standardized adaptation of document formats and languages ​​through preprocessing. Combined with RAG technology's vectorization conversion and fragmentation processing, it enhances the structure and accuracy of review data. A text analysis module integrating a large model and an automatic rule engine module achieves dual verification through deep semantic understanding and rigorous rule matching. Furthermore, a machine learning review model performs in-depth review, coupled with a human review mechanism to ensure reliability. Simultaneously, a feedback optimization mechanism enables dynamic iteration of the model and rule base, and comprehensive data security measures ensure the security of document information. This method not only significantly improves the efficiency and accuracy of document review, effectively identifying inconsistencies, errors, and compliance issues, but also adapts to the review needs of different types of documents through human-machine collaboration and continuous optimization. It balances the intelligence and flexibility of review, comprehensively meeting the requirements of document review in terms of standardization, security, and adaptability, providing efficient and reliable technical support for document management in enterprises and other fields.

[0124] In the several embodiments provided in this application, it should be understood that the disclosed methods and electronic devices can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components may be combined, or integrated into another system, or some features may be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed may be through some interfaces, and the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0125] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0126] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0127] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a platform server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks.

[0128] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A document review method based on a large model, characterized in that, The method includes: Perform preprocessing operations on official documents that require review, standardize document format and language, and adapt them to subsequent review processes; Based on the search-enhanced generation of RAG, the preprocessed official documents are vectorized and segmented to obtain document vector fragments and segmented data. The system integrates a large-scale model with a text analysis module and an automatic rule engine module. The text analysis module performs syntactic and semantic analysis on the fragmented data of official documents through natural language processing, while the automatic rule engine module performs rule matching and detection based on a preset official document review rule base. Based on a machine learning-based review model, combined with text analysis results and rule matching results, the system conducts in-depth reviews of official documents to identify potential inconsistencies, errors, and compliance issues. When uncertainty exceeds the uncertainty threshold or critical document content occurs during the review process, manual review is triggered, prompting reviewers with information on content requiring review and providing decision support tools. Receive review comments submitted by reviewers and users through the feedback interface, and iterate and optimize the rule base of the machine learning review model and the automatic rule engine module based on these comments; Output the review results of the official document content. If the review is approved, a review approval report will be generated. If the review is not approved, the problem type and location will be marked and feedback will be sent to the relevant personnel.

2. The method according to claim 1, characterized in that, The process of vectorizing and segmenting the preprocessed official documents based on retrieval-enhanced RAG generation includes: Load the vector database collection and obtain historical vector data related to official document review as a reference; The vectorization model is invoked to convert the preprocessed official documents into document vectors. At the same time, the official documents are divided into fragments according to the preset fragmentation rules to obtain multiple fragmented data of official documents. Based on the type and complexity of the official documents, the fragmented data is filtered and optimized, invalid fragments are removed, and valid fragmented data is retained for subsequent review.

3. The method according to claim 1, characterized in that, The preset document review rule base is constructed in the following way: Collect official document writing standards, terminology standards, sensitive word lists, and legal compliance requirements from different industries and organizations as the basic data for the rule base; The basic data is categorized and organized into several rule categories, including rules on writing standards, rules on word accuracy, rules on sensitive word detection, and rules on legal compliance. By employing rule encoding technology, various rules are converted into machine-readable rule codes, which are then entered into the rule base and a rule retrieval index is established to facilitate subsequent rule matching and invocation.

4. The method according to claim 1, characterized in that, The machine learning auditing model is trained in the following way: Collect sample data of official documents with marked review results, covering samples of different document types, different error types, and compliance issues; Data augmentation is performed on the sample data, including text replacement, word order adjustment, and adding similar samples, to expand the size of the training dataset; The enhanced training dataset is input into the initial machine learning model, the model training parameters are set, and the gradient descent algorithm is used to iteratively train the model until the model's audit accuracy reaches the preset threshold, thus completing the model training.

5. The method according to claim 1, characterized in that, The decision support tools provided to auditors include at least one of the following: The document issue location tool is used to mark the specific page number, paragraph, and line number of the content requiring review in the document; The historical case reference tool is used to search and display historical document review cases and their processing results that are similar to the current content that needs to be reviewed; The rule interpretation tool is used to provide detailed explanations of the violations detected by the automatic rule engine module, including the meaning of the rule, applicable scenarios, and consequences of the violation.

6. The method according to claim 1, characterized in that, The iterative optimization of the rule base for the machine learning auditing model and the automatic rule engine module includes: The feedback was categorized and statistically analyzed to determine the frequency and severity of the review issues corresponding to different types of feedback. For the machine learning review model, the official document data corresponding to the feedback opinions are used as new training samples to incrementally train the model and adjust the model parameters to improve the review accuracy. For the rule base of the automatic rule engine module, add, modify or delete relevant rules based on feedback, and update the rule retrieval index.

7. The method according to claim 1, characterized in that, The document preprocessing stage and the review result output stage also include data security protection steps: During the preprocessing stage, data encryption technology is used to encrypt official documents to prevent information leakage during document transmission and storage. During the audit results output stage, personal privacy information and sensitive business data in the audit results are anonymized, sensitive information is removed or replaced, and then the final report is generated and fed back.

8. A document review device based on a large model, characterized in that, The device includes: The preprocessing module is used to perform preprocessing operations on official documents that need to be reviewed, and to standardize the document format and language to adapt to the subsequent review process; The processing module is used to perform vectorization and fragmentation on the preprocessed official documents based on the retrieval enhancement-generated RAG, so as to obtain document vector fragments and fragmented data. The building module is used to build a text analysis module and an automatic rule engine module that integrate large models. The text analysis module performs syntactic and semantic analysis on the fragmented data of official documents through natural language processing, while the automatic rule engine module performs rule matching and detection based on a preset official document review rule library. The review module is used to conduct in-depth reviews of official documents based on machine learning review models, combined with text analysis results and rule matching results, to identify potential inconsistencies, errors and compliance issues. When uncertainty exceeds the uncertainty threshold or critical document content occurs during the review process, manual review is triggered, prompting reviewers with the content that needs to be reviewed and providing decision support tools. The receiving module is used to receive review opinions submitted by reviewers and users through the feedback interface, and to iteratively optimize the rule base of the machine learning review model and the automatic rule engine module based on these opinions. The output module is used to output the review results of the official document content. If the review is approved, an approval report is generated; if the review is not approved, the problem type and location are marked and feedback is sent to the relevant personnel.

9. An electronic device, characterized in that, include: The device includes a processor, a storage medium, and a bus, wherein the storage medium stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor communicates with the storage medium via the bus, and the processor executes the machine-readable instructions to perform the document review method based on a large model as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the document review method based on a large model as described in any one of claims 1 to 7.

Images