A multi-agent collaborative standard document review system

Abstract

The application discloses a multi-agent collaborative standard document compilation and review system, which comprises an infrastructure layer, an agent layer and an application service layer. The infrastructure layer is provided with a large language model base, a vector database, an OCR engine and a standard document special layout analysis model. The agent layer comprises four types of independent business agents, namely, an analysis agent, an extraction agent, a compilation agent and a review agent. The application relates to the technical field of artificial intelligence and standard digital engineering. The multi-agent collaborative standard document compilation and review system abandons the general large model general text question and answer mode, designs a dependency syntax and local condition graph double-driven index extraction logic, and adds a graph isomorphism verification closed loop mechanism. Compared with the general basic large model, the index extraction accuracy and recall rate are greatly improved, the core numerical parameter extraction error rate is greatly reduced, the absolute precision requirement of standard compilation for parameters is fully met, and the huge workload of manual item-by-item checking of indexes is eliminated.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and standard digitization engineering technology, specifically to a multi-agent collaborative standard document editing and review system. Background Technology

[0002] The drafting and review of standardized documents (national standards, industry standards, and group standards) must strictly adhere to the "GB / T1.1 Standardization Work Guidelines." The document hierarchy, terminology, technical specifications, and rules for referencing documents are all subject to stringent regulations. Currently, standard drafting and review within the industry primarily relies on manual labor combined with general word processing tools. This existing technology suffers from three major shortcomings: First, the manual compilation of standard documents is labor-intensive and prone to errors. Standard documents contain multi-level clauses, multi-page tables, specialized terminology, and limiting technical parameters. When written manually, problems such as skipped chapter numbering levels, non-standard font formatting, inconsistent wording of technical terms, and misaligned rows and columns in tables are very likely to occur. Commonly used Word and PDF editing software only provide basic formatting functions and cannot perform automated verification for standard-specific formatting rules. Manually correcting the format and refining the structure requires a lot of time, resulting in a long standard compilation cycle.

[0003] Second, existing document review tools can only perform shallow text verification and cannot complete cross-document technical indicator logical conflict verification. Existing text proofreading tools only support typos and basic grammar checks, and do not have the ability to identify technical indicator entities or parse constraints. The core review requirement for standard documents is to verify whether there are numerical or constraint conflicts between the technical thresholds of the document and the referenced mandatory national and industry standards. Existing tools cannot automatically extract quantitative parameters, test conditions, and veto clauses in the document, nor can they establish cross-standard indicator comparison logic. All indicator conflict verification work can only rely on reviewers to manually compare each item. When facing tens of thousands of existing standard libraries, the workload increases exponentially, and it is easy to miss or mis-review.

[0004] Third, the application of general-purpose large language models alone in standard compilation and review suffers from insufficient professionalism, significant numerical illusions, and workflow mismatches. Currently, general-purpose large models are only suitable for general text continuation and general question-and-answer scenarios, failing to break down the entire business chain for standard compilation. When directly using general-purpose large models to generate standard clauses, the models are prone to tampering with indicator values, omitting environmental constraints, and fabricating non-existent standard reference documents, creating serious "illusions." Furthermore, a single model cannot separate the independent business processes of parsing, extracting, compiling, and reviewing, lacking a process-based verification loop. The output content cannot be directly used for formal standard compilation, requiring manual secondary comprehensive correction, making it unsuitable for industrial standard compilation and review scenarios.

[0005] In summary, existing technologies lack an automated solution that can cover the entire process of standard documents, from import, structured parsing, indicator extraction, drafting assistance, and multi-dimensional compliance review. It cannot simultaneously address the three major industry pain points: standard document format standardization, cross-standard indicator logic verification, and the illusion of large model output. Summary of the Invention

[0006] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a multi-agent collaborative standard document editing and review system. It solves the problems of inefficient reuse of historical standard data and time-consuming and labor-intensive formatting during the standard drafting process. It enables intelligent drafting assistance, overcomes the limitations of existing review tools, solves the problem of difficulty in automatically extracting and accurately comparing core technical indicators between cross-standard documents, and also addresses the problem of insufficient ability of a single model to handle complex standard business. It constructs a multi-agent collaborative network with clear division of labor and controllable processes, covering the entire lifecycle from project initiation to final review.

[0007] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: a multi-agent collaborative standard document editing and review system, comprising an infrastructure layer, an agent layer, and an application service layer, wherein the infrastructure layer deploys a large language model base, a vector database, an OCR engine, and a standard document-specific layout analysis model; The intelligent agent layer includes four independent business intelligent agents: parsing intelligent agent, extraction intelligent agent, compilation intelligent agent, and review intelligent agent. The four types of intelligent agents are connected through a message bus to form a business pipeline for document parsing, indicator extraction, auxiliary drafting, and compliance review. The application service layer includes a front-end block-based editor, a review report generation module, and a standard knowledge base management backend.

[0008] Preferably, the parsing intelligent agent includes a layout analysis and recognition submodule, a high-precision ROI OCR submodule, and a hierarchical structure conversion submodule; The layout analysis and recognition submodule identifies 10 types of exclusive layout tags in standard documents based on the optimized LayoutLMv3 model, and outputs the bounding box coordinates of each tag area. The ROI high-precision OCR submodule performs local cropping and image enhancement of the title label area, and uses a standard terminology Trie tree dictionary to guide decoding for zero-error recognition of professional terms. The hierarchical structure transformation submodule converts the recognized text mapping into a Markdown intermediate file that retains the complete hierarchy.

[0009] Preferably, the extraction agent includes a structured prompt word engineering submodule, a local condition graph construction submodule, a graph isomorphism anti-hallucination verification submodule, and an index vector storage submodule; The local condition graph construction submodule performs dependency parsing on the standard text and constructs a local condition graph A with indicator entities as nodes and constraint logic as directed edges. The map isomorphism anti-hallucination verification submodule performs the following steps: S1. Input the graph A into the large model to generate structured JSON data B; S2. Reconstruct the verification graph A' from the reverse JSON; S3. The rule engine verifies whether the topology, node attributes, and constraint values ​​of graph A and A' are completely matched. S4. If the verification does not match, the conflict point is used as a negative sample and re-inputted into the model for iterative generation until the map isomorphism is matched. S5. The structured indicators that pass the verification are vectorized and stored in the vector database to build a standard indicator rule base that can be calculated by machines.

[0010] Preferably, the extraction agent is equipped with a standard domain fine-tuning large model, and the standard domain fine-tuning large model adopts a Decoder-only architecture, configured with GQA group query attention, RoPE rotation position encoding and SwiGLU activation function. The loss function of the standard domain fine-tuning large model includes cross-entropy base loss and numerical bias penalty term. During the training phase, a LoRA low-rank adaptation fine-tuning strategy is adopted, and the context window length is set to 8192 tokens.

[0011] Preferably, the writing agent includes an automatic format alignment submodule, a semantic association recommendation submodule, and a block-based human-computer interaction submodule; The automatic format alignment submodule has a built-in GB / T 1.1 standard rule base, which automatically corrects chapter numbers, font formats, and clause standard terms; The semantic association recommendation submodule is based on semantic retrieval from a vector database and recommends standard terms, current reference standards, and index value ranges for different scenarios. The block-based human-computer interaction submodule divides the terms into independent editing blocks, supports real-time streaming editing suggestions and retention of multiple version revision traces.

[0012] Preferably, the auditing agent includes a formal review submodule, an indicator logic conflict verification submodule, a reference standard verification submodule, and an audit report generation submodule; The formal review submodule completes document hierarchical skipping, empty chapter detection, circular reference detection, chart layout distance verification, and full-text terminology consistency detection. The indicator logic conflict verification submodule performs semantic matching between the set of indicators in the draft to be reviewed and the mandatory standard indicator library to verify whether there are conflicts between the indicator values ​​and the veto clauses. The reference standard verification submodule calls the standard public service platform API to identify obsolete and expired reference standards and push alternative versions. The review report generation submodule outputs a tiered list of issues, suggested amendments to clauses, and a comparison table of indicators.

[0013] Preferably, the layout analysis and recognition submodule uses a joint loss function during the training phase: ; Among them, L cls For classification cross-entropy loss, L reg The bounding box Smooth L1 regression loss is used. , To balance the weights, For real labels, To predict the class probability distribution for the model, Here, t represents the true bounding box coordinate offset, and t represents the model-predicted bounding box coordinate offset. The optimizer used is AdamW, with an initial learning rate of 5×10⁻⁶. −5 Batch size 16, iterative training 50 rounds.

[0014] Preferably, the total loss function of the standard domain fine-tuning large model is: ; Among them, the first item Generate cross-entropy loss for autoregressive text, the latter term For the predicted numerical token Compared with the actual value Calculate the deviation penalty. As a balancing factor, the model is forced to remain absolutely faithful to the limit thresholds in the standard.

[0015] Preferably, the execution logic of the indicator logic conflict verification submodule is as follows: for indicator d in the draft... i With mandatory standard index m j If the two application domains overlap, the semantic similarity is multiplied by the domain overlap weight to obtain the matching score; when the score is higher than the 0.85 threshold, numerical logic comparison is carried out, and three types of verification results are output: consistent, conflicting, and manually reviewed.

[0016] (III) Beneficial Effects This invention provides a standard document editing and review system for multi-agent collaboration. Compared with existing technologies, it has the following advantages: (1) This multi-agent collaborative standard document editing and review system abandons the general large model general text question answering mode, designs a dual-drive index extraction logic of dependency syntax and local condition graph, and adds a graph isomorphism verification closed-loop mechanism. After verification by 1,000 mechanical, electronic and chemical industry standard test sets, compared with the general basic large model, the accuracy and recall of index extraction are greatly improved, the error rate of core numerical parameter extraction is greatly reduced, and it fully meets the hard requirement of absolute accuracy of parameters in standard compilation, eliminating the huge workload of manually checking each index.

[0017] (2) The multi-agent collaborative standard document editing and review system is based on the optimized LayoutLMv3 page analysis model and is equipped with a standard terminology-specific OCR decoding strategy. It optimizes the visual recognition logic for national and industry standard-specific multi-level headings, cross-page tables, and footnote scenarios. The page element recall rate is greatly improved, the standard document structure restoration fidelity is high, the manual typesetting time is greatly reduced, and after importing the PDF original, it automatically outputs Markdown files with complete hierarchical alignment and no table misalignment, without the need for manual reorganization of the document structure.

[0018] (3) The multi-agent collaborative standard document editing and review system has a three-layer hierarchical architecture and a message bus asynchronous communication mechanism to achieve complete decoupling of the four types of agents. It can support each agent to provide API interfaces to the outside world and adapt to the existing digital platform of the enterprise. It can also complete the entire process of standard import, parsing, indicator extraction, drafting and review through the workflow engine. The whole system covers the entire life cycle of standard project drafting, internal review and final review and archiving. The multi-agent hierarchical decoupling architecture is flexible and expandable, realizes the closed loop of full-process automated editing and review, replaces the traditional manual editing and review mode, and greatly shortens the overall cycle of small and medium-sized standard compilation projects.

[0019] (4) The standard document editing and review system of this multi-agent collaboration only supports shallow text correction through existing proofreading tools and cannot complete cross-document technical indicator comparison. This invention stores all existing standard indicators through a vector database, and automatically compares the draft with the mandatory national standard constraints based on the semantic matching algorithm. It automatically identifies numerical conflicts, repealed standard references, and terminology misuse issues, and automatically outputs a graded review report with modification basis. It realizes automatic compliance verification across standard systems, fills the technical gap of existing review tools, and solves the industry pain point of manual omission and wrong review in cross-review of multiple standard systems. Attached Figure Description

[0020] Figure 1 This is a block diagram illustrating the structural principle of the system of the present invention. Detailed Implementation

[0021] 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 the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figure 1 The present invention provides two technical solutions: a standard document editing and review system for multi-agent collaboration, specifically including the following embodiments: Example 1: A standard document editing and review system for multi-agent collaboration. The system adopts a three-layer decoupled layered architecture, consisting of an infrastructure layer, an agent layer, and an application service layer from bottom to top. Each layer asynchronously transmits structured JSON data through a standardized message bus. Each agent is independently decoupled and can be called individually or run in a pipeline. Infrastructure layer: Deploys a large language model base, vector database, high-precision OCR engine, and standard document-specific layout analysis model LayoutLMv3; the infrastructure layer provides underlying support for visual recognition, text generation, vector storage, and numerical computation. The intelligent agent layer contains four independent business intelligent agents: parsing intelligent agent, extraction intelligent agent, compilation intelligent agent, and review intelligent agent. The four types of intelligent agents communicate with each other through a message bus and are connected in sequence to form a complete business pipeline of "document parsing - indicator extraction - drafting assistance - compliance review". Application service layer: includes a front-end block-based compilation editor, a review report generation module, and a standard knowledge base management backend; provides a visual interactive interface for compilers and reviewers, and outputs structured compilation and review deliverables and hierarchical review reports.

[0023] 1. Analyzing the intelligent agent The parsing agent receives standard original PDF / Word documents imported from external sources, performs layout visual recognition, directional high-precision OCR recognition, and hierarchical faithful Markdown structured conversion, and outputs standardized intermediate files to provide a unified input data source for downstream agents. It includes a layout analysis and recognition submodule, an ROI high-precision OCR submodule, and a hierarchical structured conversion submodule.

[0024] The layout analysis and recognition submodule uses the optimized LayoutLMv3 deep learning model to recognize visual elements in standard documents. The model's backbone network integrates a ResNet-50 visual feature extractor with a Transformer encoding layer, introducing relative spatial location encoding to adapt for cross-page tables and scattered footnotes. The model's pre-training dataset contains 50,000 publicly available national and industry standard PDFs, manually annotated with 10 categories of standard-specific layout labels, totaling 2 million bounding boxes. A joint loss function is used during the training phase. ; Among them, L cls For classification cross-entropy loss, L reg The bounding box Smooth L1 regression loss is used. , To balance the weights, For real labels, To predict the class probability distribution for the model, Here, t represents the true bounding box coordinate offset, and t represents the model-predicted bounding box coordinate offset. The optimizer used is AdamW, with an initial learning rate of 5×10⁻⁶. −5 Batch size 16, iterative training 50 rounds.

[0025] doc_title: Standard document title (standard number + name), extracted with high priority, and used as a unique index key in the database; paragraph_title: Section, subsection, and item level headings; extract the numerical numbers to construct the document hierarchy tree. normative_text: The standard and normative text, distinguishing between requirement clauses (must / must not) and statement clauses; table_cell: A table cell that records row and column coordinates and marks cells that have been merged across pages; footnote: Document footnotes, bound to the corresponding main clause to retain citation association; the layout analysis model outputs the pixel bounding box coordinates of each label area, distinguishing the independent areas of titles, body text, tables, charts, and footnotes within the document.

[0026] The ROI high-precision OCR submodule performs local precise text recognition on the doc_title and paragraph_title tag areas output by the layout model, eliminates background and watermark interference, and achieves zero-error recognition of standard terms. The execution steps are: (1) Dynamic region cropping: cropping the local ROI of the original image according to the bounding box coordinates and stripping the irrelevant page background; (2) Image adaptive enhancement: binarizing the cropped image and correcting the text tilt with Radon transform; (3) Trie tree terminology guided decoding: fine-tuning the OCR model based on TrOCR and attaching a standard terminology prefix tree dictionary in the decoding stage; when the model outputs easily confused units and professional words, the terminology spelling is forcibly corrected through the prefix tree forward probability penalty mechanism to eliminate character recognition errors such as MPa and NPa.

[0027] The hierarchical structure conversion submodule converts OCR-recognized text into Markdown intermediate files by combining the layout label hierarchy mapping. The conversion rules are as follows: doc_title maps to first-level headings, and paragraph_title maps to second- to fourth-level headings according to the numbering hierarchy; table cells retain row and column merging marks; after conversion, the data is automatically written to the file system for storage and used as a unified input file for extracting intelligent agents.

[0028] 2. Extracting the intelligent agent Extract the Markdown standard file output by the intelligent agent, complete the structured extraction of quantitative parameters, limiting conditions, and veto clauses within the standard, and build a machine-computable index vector library. This intelligent agent is the core innovative module of the system, including a structured prompt word engineering submodule, a local condition graph construction submodule, a graph isomorphism anti-hallucination verification submodule, and an index vector library entry submodule.

[0029] The structured prompt word engineering submodule design standard uses a dedicated indicator extraction template, which forces the large model to switch to a precise entity extraction mode. The template limits the extraction objects to four types: quantitative numerical indicators, environmental constraints, prohibited veto items, and test and detection methods. It defines a standardized output schema, which forces the output of a unique indicator identifier, the original indicator text, standardized numerical value, unit, related clause, and confidence level field, and prohibits free text output.

[0030] The local condition graph construction submodule performs dependency parsing on long and complex sentences in standard text, splitting them into indicator entities, logical constraint relationships, and preconditions. The parsed entities are used as directed graph nodes, and the constraint logic is used as directed edges to construct a local condition graph A. The graph fully records all the preconditions of the indicators, solving the problem of mismatch between conditions and indicators caused by pure text sequence extraction.

[0031] The graph isomorphism anti-illusion verification submodule (core anti-illusion closed loop) serializes the local condition graph A and inputs it into the large language model along with the standard text block. The model outputs structured JSON data B. The verification process is as follows: (1) Graph logic assertion conversion: Graph A is converted into a set of machine-readable logic assertions, which fully records all entity nodes and constraint edge relationships; (2) JSON reverse reconstruction of the verification graph: The conversion module parses the model output JSON and reverse generates the verification graph A'; (3) Rule engine bidirectional isomorphism verification: The consistency of node entities, constraint edge direction, numerical threshold, and logical symbols (≥, ≤) are verified synchronously; If the topology and attributes of graph A and A' are completely matched, the verification is passed; If there are numerical omissions, condition tampering, or entity missing, the verification is judged to fail; (4) Penalty iterative regeneration: When the verification is inconsistent, the rule engine adds the conflict difference points as negative samples to the prompt words, re-inputs the large model to generate JSON, and performs loop verification until the graph isomorphism matches, thus completely suppressing the numerical illusion of the model.

[0032] The structured indicator data that passes the verification of the indicator vector storage submodule is encoded into 768-dimensional semantic vectors using a self-developed domain BGE vectorization model and stored in the vector database. The vector database stores fields including indicator semantic vectors, indicator original text, standardized structured parameters, source standard number, clause path, and indicator classification (safety / performance / environmental protection). This completes the conversion of unstructured text into a computable logical rule base, providing indicator data support for subsequent compilation and review of intelligent agents.

[0033] This agent is equipped with a dedicated standard domain-specific fine-tuned large model. The model is based on a decoder-only architecture and employs GQA grouped query attention, RoPE rotational position encoding, SwiGLU activation function, and Pre-RMSNorm layer normalization. Training consists of two stages: domain-specific continuous pre-training and instruction fine-tuning. The dataset contains 500,000 standard original text-index quadruple samples. The loss function adds a numerical bias penalty term to the traditional autoregressive cross-entropy. ; Among them, the first item Generate cross-entropy loss for autoregressive text, the latter term For the predicted numerical token Compared with the actual value Calculate the deviation penalty. As a balancing factor, the model is forced to remain absolutely faithful to the limit thresholds in the standard.

[0034] 3. Compile intelligent agents The intelligent agent receives standard outlines and draft fragments input from front-end users, calls and parses the intelligent agent format rule library and extracts the intelligent agent indicator vector library, and realizes automatic alignment of GB / T 1.1 specifications, intelligent recommendation of clauses, and human-computer collaborative drafting. It includes an automatic format alignment submodule, a semantic association recommendation submodule, and a block-based human-computer interaction submodule.

[0035] The GB / T 1.1 format auto-alignment submodule has a built-in complete GB / T 1.1-2020 format rule library. It scans the user's draft Markdown text in real time and automatically completes multiple standard corrections: detects skipped chapter numbers and empty chapters and prompts for completion in a pop-up window; automatically distinguishes between boldface for terms and Song typeface for body text; standardizes the writing format of normative reference documents; and unifies the terminology of clauses, replacing the colloquial "should" with the standard normative term "should".

[0036] The semantic association recommendation submodule uses semantic retrieval based on a vector library to achieve intelligent recommendations for different scenarios: when drafting a terminology chapter, it retrieves high-frequency terms and standard definitions from historical standards in the same field; when drafting normative reference documents, it connects to the API of the National Standards Information Public Service Platform to recommend currently valid standards and mark obsolete and replacement versions; when drafting a technical requirements chapter, it retrieves the statistical distribution of similar standard indicators to provide users with reasonable numerical ranges for reference.

[0037] The modular human-computer interaction submodule front-end editor divides each standard clause into an independent editing block; when the user enters text, the intelligent agent outputs editing suggestions in real time, providing three interaction modes: adopt, ignore, and adopt after modification; the system automatically retains all revision traces and supports comparison of differences between multiple versions of documents.

[0038] Example 2: The technical solution of Example 1 of the present invention differs from that of Example 1 in that: the auditing intelligent agent Once the standard draft is completed, the review agent is triggered to simultaneously perform formal review, cross-standard indicator logical conflict verification, and reference document validity check. Finally, a hierarchical structured review report is output, which includes a formal review submodule, an indicator logical conflict verification submodule, a reference standard verification submodule, and a review report generation submodule.

[0039] The specific processing steps of the formal review submodule are as follows: (1) Document hierarchy integrity verification: Extract Markdown title level tuples, traverse the sequence to detect skipping levels, empty chapters, and circular references; if the difference between the serial numbers before and after the title tuple is greater than 1, it is determined to skip levels; if the reference link forms a closed loop, it is marked as a circular reference; if there is no text content in the title interval, it is marked as an empty chapter; (2) Chart layout consistency verification: Match the text with the chart number in the document; calculate the physical distance between the text referenced and the chart page, and mark the layout as unreasonable if it spans more than two pages; (3) Terminology uniformity verification: Extract term definitions to build a local dictionary, use vector to search for synonyms in the full text, and highlight the problem of mixed use of undefined terms.

[0040] The indicator logic conflict verification submodule calls the intelligent agent to synchronously parse the entire indicator set D of the draft to be reviewed, and performs a full comparison with the mandatory national standard indicator library M in the vector database. The execution logic is as follows: for each indicator d in the draft... i traverse the forced standard index m j If the two application domains overlap, calculate the semantic similarity × domain overlap weight to obtain the matching score; when the score is higher than the 0.85 threshold, conduct numerical logic comparison: quantitative indicators verify whether the draft values ​​meet the mandatory standard constraint threshold; veto items verify whether the prohibitive clauses conflict; finally, output three types of results: consistent, conflicting, and manually reviewed.

[0041] The standard verification submodule uses regular expressions to extract all normative standard reference numbers from the draft, calls the standard public service platform API to query the current status of the standards, marks obsolete and expired standards, and automatically pushes the official replacement standard number to the review report.

[0042] The review report generation submodule automatically generates a graded review report, categorizing issues into three types based on severity: errors, warnings, and alerts. Each conflict item includes the original text of the conflicting clause, the basis for the mandatory standard, and directly adoptable modification suggestions. A visual table comparing indicators is provided to intuitively display the differences between the draft indicators and the mandatory standard values.

[0043] Complete data flow process of the system: 1. Import standard PDF / Word documents from external sources, input the parsing intelligent agent, and output a Markdown structured intermediate file after layout recognition, high-precision OCR, and hierarchical conversion; 2. Extract the agent from the Markdown file, construct the local conditional graph, generate JSON from the large model, verify the isomorphism of the graph, and store the structured indicators in the vector database. 3. Users draft standard drafts in the front-end editor, and the intelligent agent calls the format rule library and indicator vector library to output specification corrections and content recommendations in real time; 4. After the draft is finalized, the review agent is activated to simultaneously complete formal review, cross-standard indicator conflict verification, and validity check of cited documents; 5. The system outputs a tiered review report and simultaneously supports exporting the final draft in a standardized format; 6. The workflow supports two operating modes: independent API call mode, in which four types of intelligent agents expose RESTful interfaces to the outside world; and fully automated pipeline mode, in which the workflow engine automatically orchestrates the complete chain of parsing, extraction, compilation, and review.

[0044] Furthermore, any content not described in detail in this specification is existing technology known to those skilled in the art.

[0045] It should 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 process, method, article, or apparatus.

[0046] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A standard document editing and reviewing system for multi-agent collaboration, characterized in that: It includes an infrastructure layer, an intelligent agent layer, and an application service layer. The infrastructure layer deploys a large language model base, a vector database, an OCR engine, and a standard document-specific layout analysis model. The intelligent agent layer includes four independent business intelligent agents: parsing intelligent agent, extraction intelligent agent, compilation intelligent agent, and review intelligent agent. The four types of intelligent agents are connected through a message bus to form a business pipeline for document parsing, indicator extraction, auxiliary drafting, and compliance review. The application service layer includes a front-end block-based editor, a review report generation module, and a standard knowledge base management backend. The analytical agent includes a layout analysis and recognition submodule, a high-precision ROI OCR submodule, and a hierarchical structured conversion submodule; The layout analysis and recognition submodule identifies 10 types of exclusive layout tags in standard documents based on the optimized LayoutLMv3 model, and outputs the bounding box coordinates of each tag area. The ROI high-precision OCR submodule performs local cropping and image enhancement of the title label area, and uses a standard terminology Trie tree dictionary to guide decoding for zero-error recognition of professional terms. The hierarchical structured transformation submodule converts the recognized text mapping into a Markdown intermediate file that retains the complete hierarchy; The extraction agent includes a structured prompt word engineering submodule, a local conditional graph construction submodule, a graph isomorphism anti-hallucination verification submodule, and an index vector storage submodule. The local condition graph construction submodule performs dependency parsing on the standard text and constructs a local condition graph A with indicator entities as nodes and constraint logic as directed edges. The map isomorphism anti-hallucination verification submodule performs the following steps: S1. Input the graph A into the large model to generate structured JSON data B; S2. Reconstruct the verification graph A' from the reverse JSON; S3. The rule engine verifies whether the topology, node attributes, and constraint values ​​of graph A and A' are completely matched. S4. If the verification does not match, the conflict point is used as a negative sample and re-inputted into the model for iterative generation until the map isomorphism is matched. S5. The structured indicators that pass the verification are vectorized and stored in the vector database to build a standard indicator rule base that can be calculated by machines. The compilation agent includes an automatic format alignment submodule, a semantic association recommendation submodule, and a block-based human-computer interaction submodule. The automatic format alignment submodule has a built-in GB / T 1.1 standard rule base, which automatically corrects chapter numbers, font formats, and clause standard terms; The semantic association recommendation submodule is based on semantic retrieval from a vector database and recommends standard terms, current reference standards, and index value ranges for different scenarios. The block-based human-computer interaction submodule divides the terms into independent editing blocks, supports real-time streaming editing suggestions and retention of multiple version revision traces.

2. The standard document editing and reviewing system for multi-agent collaboration according to claim 1, characterized in that, The extraction agent is equipped with a standard domain fine-tuning large model, which adopts a Decoder-only architecture and is configured with GQA group query attention, RoPE rotation position encoding and SwiGLU activation function. The loss function of the standard domain fine-tuning large model includes cross-entropy base loss and numerical bias penalty term. During the training phase, a LoRA low-rank adaptation fine-tuning strategy is adopted and the context window length is set to 8192 tokens.

3. The standard document editing and reviewing system for multi-agent collaboration according to claim 1, characterized in that, The auditing agent includes a formal review submodule, an indicator logic conflict verification submodule, a reference standard verification submodule, and an audit report generation submodule. The formal review submodule completes document hierarchical skipping, empty chapter detection, circular reference detection, chart layout distance verification, and full-text terminology consistency detection. The indicator logic conflict verification submodule performs semantic matching between the set of indicators in the draft to be reviewed and the mandatory standard indicator library to verify whether there are conflicts between the indicator values ​​and the veto clauses. The reference standard verification submodule calls the standard public service platform API to identify obsolete and expired reference standards and push alternative versions. The review report generation submodule outputs a tiered list of issues, suggested amendments to clauses, and a comparison table of indicators.

4. A standard document editing and reviewing system for multi-agent collaboration according to claim 1, characterized in that: The layout analysis and recognition submodule uses a joint loss function during the training phase: ； Among them, L cls For classification cross-entropy loss, L reg The bounding box Smooth L1 regression loss is used. , To balance the weights, For real labels, To predict the class probability distribution for the model, Here, t represents the true bounding box coordinate offset, and t represents the model-predicted bounding box coordinate offset. The optimizer used is AdamW, with an initial learning rate of 5×10⁻⁶. −5 Batch size 16, iterative training 50 rounds.

5. A standard document editing and reviewing system for multi-agent collaboration according to claim 2, characterized in that: The total loss function of the standard domain fine-tuning large model is: ； Among them, the first item Generate cross-entropy loss for autoregressive text, the latter term Regarding the predicted values Compared with the actual value Calculate the deviation penalty. As a balancing factor, the model is forced to remain absolutely faithful to the limit thresholds in the standard.

6. A standard document editing and reviewing system for multi-agent collaboration according to claim 3, characterized in that: The execution logic of the indicator logic conflict verification submodule is as follows: For indicator d in the draft... i With mandatory standard index m j If the two application domains overlap, the semantic similarity is multiplied by the domain overlap weight to obtain the matching score; when the score is higher than the 0.85 threshold, numerical logic comparison is carried out, and three types of verification results are output: consistent, conflicting, and manually reviewed.

Images