Large model-based standard document automatic generation and multi-dimensional auditing method and system

Abstract

The invention provides a standard document automatic generation and multi-dimensional auditing method and system based on a large model, and relates to the technical field of artificial intelligence, and the method comprises the steps: 1, building a distributed database of a multi-source document, and analyzing a heterogeneous text through natural language processing to obtain a standardized knowledge network; step 2, extracting index elements based on the standardized knowledge network, and forming a structured parameter library through verification and verification; and step 3, based on the structured parameter library, constructing a template library, analyzing user demands in combination with semantic matching, and automatically generating a standard document outline. The document generation efficiency and quality are improved, the manual auditing cost is reduced, and the auditing comprehensiveness and accuracy are enhanced.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method and system for automatically generating and multi-dimensionally reviewing standard documents based on a large model. Background Art

[0002] In the process of writing standard documents, traditional technology has some defects. Traditional standard writing requires manual completion of a large amount of data collection, organization and analysis. Taking the writing of new energy vehicle charging interface standards as an example, the writers have to screen useful information from various documents, industry reports, and past related standards. Just collecting the charging interface technical parameters of different car companies, the country's existing electrical safety standards and other information may take several weeks, affecting the progress of standard writing.

[0003] In addition, due to differences in the professional level and experience of the writers, the quality of the standards varies. Experienced electrical engineers may fully consider key safety indicators such as the interface's voltage resistance and leakage protection, while inexperienced writers may ignore some important parameters, causing the written standards to have safety hazards in actual applications and unable to effectively guarantee the safety of the charging process. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method and system for automatic generation and multi-dimensional review of standard documents based on a large model, thereby improving the efficiency and quality of standard document generation, and at the same time promoting accuracy through multi-dimensional review.

[0005] In order to solve the above technical problems, the technical solutions of the present invention are as follows: In the first aspect, a method for automatically generating and multi-dimensionally reviewing standard documents based on a large model includes: Step 1: Establish a distributed database of multi-source documents and analyze heterogeneous texts through natural language processing to obtain a standardized knowledge network; Step 2: Based on the standardized knowledge network, extract indicator elements and form a structured parameter library through verification and validation; Step 3: Based on the structured parameter library, a template library is constructed, user needs are analyzed in combination with semantic matching, and a standard document outline is automatically generated; Step 4: Based on the standard document outline, similar document fragments are retrieved from the knowledge base, and the standard chapter content is written in combination with semantic generation to generate a draft standard document; Step 5: Based on the draft standard document, convert key data into point cloud data, determine the reference points, and delineate the inspection area. Set inspection points inside and outside the area, construct an arc inspection path, and generate path correction parameters to obtain audit conclusions and modification suggestions. Step 6: Integrate the review conclusions and modification suggestions into the knowledge base, and adjust the parameter weights and generation rules through machine learning mechanisms.

[0006] Secondly, the standard document automatic generation and multi-dimensional review system based on the big model includes: The knowledge network module is used to build a distributed database of multi-source documents and analyze heterogeneous texts through natural language processing to obtain a standardized knowledge network; Parameter extraction module, which is used to extract indicator elements based on the standardized knowledge network and form a structured parameter library through verification and validation; The document outline module is used to build a template library based on the structured parameter library, analyze user needs in combination with semantic matching, and automatically generate a standard document outline; The document content module is used to retrieve similar document fragments from the knowledge base based on the standard document outline, and complete the standard chapter content writing in combination with semantic generation to generate a standard document draft; The audit and assessment module is used to convert key data into point cloud data based on the draft standard document, determine the reference points and delineate the inspection area; set inspection points inside and outside the area, construct an arc inspection path and generate path correction parameters to obtain audit conclusions and modification suggestions; The knowledge feedback module is used to integrate the review conclusions and modification suggestions into the knowledge base, and adjust the parameter weights and generation rules through machine learning mechanisms.

[0007] According to a third aspect, a computing device includes: one or more processors; The storage device is used to store one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors implement the method.

[0008] In a fourth aspect, a computer-readable storage medium stores a program, which implements the method when executed by a processor.

[0009] The above solution of the present invention includes at least the following beneficial effects: By building a standardized knowledge network, the automated integration of multi-source heterogeneous texts such as policy documents, patent documents, and experimental reports can be achieved, alleviating the tedious process of traditional manual collection and screening of information one by one; with the help of structured parameter libraries and semantic matching technology, intelligent drive is achieved in the entire process from demand analysis to outline generation and chapter content writing, shortening the cycle from conception to draft formation of standard documents, and alleviating the core pain points of low efficiency and time-consuming traditional manual writing; through format verification and cross-document cross-validation mechanisms, conflicting descriptions in multi-source data can be accurately identified and corrected to consistent core parameter benchmark values ​​to ensure the data accuracy of standard content; an innovative design of the detection point-sub-area-dynamic adjustment value audit framework conducts automated audits from multiple dimensions, including the completeness of necessary elements (scope definition, core clauses, etc.), the standardization of professional expressions (terminology, symbols, etc.), and the logical relevance of citations (file, figure numbers, etc.), to improve the structural rationality and professional rigor of standard texts.

[0010] By associating and mapping structured parameters with template libraries, a set of term definitions, quantitative indicator clauses, and verification method clauses that comply with the standards are automatically generated, converting complex standardization rules into visual generation logic, reducing dependence on the professional experience and industry background of the writers; the standardization knowledge network integrates cross-domain and cross-type standardization resources, and realizes efficient reuse of parameter systems through case matching and fragment reorganization technology to avoid duplication of work; knowledge feedback continuously integrates audit results, manual revision tracks, and error distribution data through machine learning, dynamically adjusts conflict weight thresholds, semantic matching rules, and template mapping relationships, and can adapt to different industry standards and new specification requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a flowchart of a method for automatic generation and multi-dimensional review of standard documents based on a large model provided by an embodiment of the present invention.

[0012] Figure 2 It is a schematic diagram of a large-model-based standard document automatic generation and multi-dimensional review system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0013] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0014] like Figure 1 As shown, an embodiment of the present invention proposes a method for automatic generation and multi-dimensional review of standard documents based on a large model, the method comprising the following steps: Step 1: Establish a distributed database of multi-source documents and analyze heterogeneous texts through natural language processing to obtain a standardized knowledge network; Step 2: Based on the standardized knowledge network, extract indicator elements and form a structured parameter library through verification and validation; Step 3: Based on the structured parameter library, a template library is constructed, user needs are analyzed in combination with semantic matching, and a standard document outline is automatically generated; Step 4: Based on the standard document outline, similar document fragments are retrieved from the knowledge base, and the standard chapter content is written in combination with semantic generation to generate a draft standard document; Step 5: Based on the draft standard document, convert key data into point cloud data, determine the reference points, and delineate the inspection area. Set inspection points inside and outside the area, construct an arc inspection path, and generate path correction parameters to obtain audit conclusions and modification suggestions. Step 6: Integrate the review conclusions and modification suggestions into the knowledge base, and adjust the parameter weights and generation rules through machine learning mechanisms.

[0015] In an embodiment of the present invention, a distributed database is established by integrating multiple types of documents such as policies, patents, and experimental reports, thereby solving the problem of decentralized storage and difficulty in linkage of traditional documents and realizing centralized management of cross-domain knowledge. Natural language processing is used to complete entity recognition, semantic association, and data fusion, so that originally isolated texts can form an interconnected knowledge network, avoiding the one-sidedness of content caused by incomplete information. Core indicator elements are extracted based on the knowledge network, and the data presentation form is standardized through format verification to reduce the problem of missing fields or format confusion. Cross-document cross-validation is used to identify and correct conflicting descriptions in multi-source data to form unified benchmark parameters, providing structured, high-quality basic data for standard content generation, and ensuring the accuracy of the standard from the source.

[0016] Relying on the structured parameter library and standardization specifications to build a template library, it accurately interprets user needs through semantic matching, and automatically generates a standard outline containing term definitions and core clause frameworks, avoiding the problems of arbitrary outline structure and non-compliance with industry standards in traditional manual writing; ensuring that the outline not only meets the actual needs of users, but also meets the basic requirements of standardization, reducing the time cost of repeated outline adjustments; based on the outline decomposition and retrieval of similar document fragments, combining semantics to complete the chapter content writing, replacing the traditional manual chapter-by-chapter reference and manual writing process, shortening the generation cycle from outline to draft; through case matching, it achieves the reuse of high-quality experience, avoids repeated creation, and improves Improve the professionalism and writing efficiency of the standard content; focus on the audit points through fixed detection points and sub-area segmentation, and conduct automated audits from three dimensions: element integrity, professional expression standardization, and reference logical relevance, combined with adjustment values, to accurately locate problems such as missing elements, symbol errors, and reference confusion in the draft; compared with traditional manual review, it reduces subjective omissions, provides clear revision opinions, and improves the audit efficiency and standardization of standard documents; integrates audit results, manual feedback and other information, and adjusts parameter weights and generation rules through machine learning, which can continuously absorb experience from actual applications and dynamically adapt to the writing needs of different industries and different types of standards.

[0017] In a preferred embodiment of the present invention, the above step 1, establishing a distributed database of multi-source documents and analyzing heterogeneous texts through natural language processing to obtain a standardized knowledge network, may include: In an embodiment of the present invention, for five major categories of text, namely policy documents, patent documents, experimental reports, standard documents and internal enterprise documents, original document collection is completed through targeted collection (such as policy columns on government official websites, patent databases, internal enterprise document management systems, etc.) and batch import; the collected documents are classified and labeled by type, and metadata information such as the source, release time, and applicable fields of each type of document is recorded to form an initial document collection; a distributed storage architecture is designed according to document type, scale and access requirements, and the initial document collection is allocated to different storage nodes by type or field. Each node independently stores document data of the corresponding category, and at the same time, a centralized index node is established to record the storage location, metadata and access rights of each document; a data synchronization mechanism between nodes is used to ensure the consistency and accessibility of data on each storage node, forming a distributed database covering multiple types of text, and realizing centralized management and efficient retrieval of documents.

[0018] Perform unified preprocessing on heterogeneous texts in distributed databases. For documents in different formats (such as PDF, Word, TXT, etc.), extract plain text content through format conversion tools and remove redundant information (such as headers and footers, irrelevant symbols, and repeated paragraphs). Clean the extracted plain text, including unified text encoding, correcting typos, and standardizing the use of punctuation to ensure the readability and consistency of the text content, and provide regular text materials for subsequent natural language processing. Based on the preprocessed text materials, combined with domain features and text semantics, establish entity recognition by manually defining entity types (such as regulatory names and responsible parties in policy documents; technical terms and inventors in patent documents; indicator names and experimental equipment in experimental reports, etc.). Rule base; Scan the text sentence by sentence based on the rule base to identify and extract key entities that meet the type definition, and record the appearance position, context and document type of each entity in the text to form a preliminary entity list; For the identified entity list, establish association clues by analyzing the contextual logical relationship between entities in the text (such as "reference", "inclusion", "cause and effect", "implementation object", etc.); for policy documents, analyze the regulatory relationship between regulatory names and responsible entities; for patent documents, analyze the correspondence between technical terms and application scenarios; for experimental reports, analyze the association between indicator names and test methods, etc. By summarizing the association clues between entities in different documents, sort out the multi-dimensional association relationships between entities and form a preliminary entity association map.

[0019] For different expressions of the same entity in different documents (such as full name and abbreviation, professional term and common name), the contextual semantic similarity and consistency of the entity are compared to determine whether they are the same entity and merge them for annotation; for conflicts in entity associations in different documents (such as the same term is associated with different test methods in patents and experimental reports), the document authority and domain adaptability are combined to make judgments, retain reasonable associations, and finally organize the integrated entities and associations according to a unified structure to form a standardized knowledge network including entity types, attributes, and multi-dimensional associations.

[0020] In a preferred embodiment of the present invention, the above step 2, extracting indicator elements based on the standardized knowledge network and forming a structured parameter library through verification and validation, may include: Step 220 , based on the standardized knowledge network, semantic analysis is used to identify and extract the three core elements of the knowledge network: quantitative indicators, performance parameters, and test methods, to obtain descriptive text fragments; Step 221 , matching the descriptive text snippet with a preset standardized format, checking for missing fields and triggering an alarm; Step 222 , for the same core element that has completed verification, the similarity of the descriptive text segments in different source documents is calculated, and description conflicts whose conflict weights exceed a preset threshold are identified and recorded; Step 223 , for description conflicts, cross-validation is performed based on context relevance and preset rules to determine a benchmark value and perform numerical correction to generate a consistent description; In step 224 , the core elements and the consistent description are integrated, the hierarchical attribution is calculated, and classification labels are added to form a structured parameter library.

[0021] In an embodiment of the present invention, all documents in the standardized knowledge network are first scanned in full text, and the text content is split into paragraphs. In the semantic analysis stage, characteristic words in the text are identified sentence by sentence. For quantitative indicators, the focus is on capturing expressions containing features such as "value + unit", "percentage", "multiple", and "range". For example, "standby current: 50mA±5mA" is extracted from "the standby current of the equipment is 50mA±5mA" as a descriptive text segment of the quantitative indicator; for performance parameters, the focus is on performance keywords such as "speed", "precision", "efficiency", and "strength". For example, "rotation accuracy: 0.01mm" is extracted from "the bearing rotation accuracy reaches 0.01mm" as a descriptive text segment of the performance parameter; for test methods, process expressions such as "test steps", "measurement conditions", and "operation procedures" are locked. For example, the complete sentence from "when testing, first adjust the voltage to 220V, keep it for 30 minutes, and then record the data" is extracted as a descriptive text segment of the test method. After the extraction is completed, the name, chapter and page number information of the source document are marked for each text segment.

[0022] The preset standardized format specifies field requirements for different core element types. For example, quantitative indicators must include five fields: "indicator name, benchmark value, fluctuation range, measurement unit, and measurement environment". The extracted descriptive text fragments are compared with these fields one by one. First, check whether there is a clear statement of the corresponding field in the text. If a text fragment only mentions "pressure is 10MPa" and does not mention the measurement environment and fluctuation range, then the two fields of "fluctuation range" and "measurement environment" are missing. The missing field ratio is calculated as "number of missing fields ÷ total number of fields", that is, 2 ÷ 5 = 40%. If the threshold is set to 20%, an alarm will be triggered if 40% exceeds the threshold, and a prompt will pop up to mark the missing field. If a text fragment is missing one field, the missing ratio is 20%, which does not exceed the threshold, then only the missing field name will be marked in red.

[0023] For the same core element that passes the format check (such as "engine power"), collect the descriptive text fragments in all source documents. When calculating similarity, first remove function words such as "of" and "in" from the text, and retain core words such as "power", "120kW", and "rated speed". Count the number of overlapping core words in different text fragments. The proportion of the overlapping number to the total number of core words is the similarity. For example, if both texts contain "power", "120kW", and "rated speed 3000r / min", and the core word overlap rate is 80%, then the similarity is 80%. The conflict weight calculation starts from four aspects. The numerical difference is calculated as "|text 1 value - text 2 value| ÷ text 1 value". For example, the difference rate between 120kW and 110kW is (10 ÷ 120) ≈ 8.3%. The range difference is calculated as "|upper limit difference| + |lower limit difference| ÷ total length of the standard range". For example, the difference value between 20 - 30°C and 22 - 28°C is (2 + 2) ÷ 10 = 40%. If the unit difference is between similar units (such as m and cm), calculate the numerical difference after conversion. For different units (such as kg and m), directly record it as a 100% difference. The conditional attribute difference is calculated as the proportion of the number of different conditions to the total number of conditions. For example, if one test environment is "room temperature" and the other is "high temperature", the difference rate is 50%. Multiply the difference rates of the four aspects by the preset weights (numerical 0.3, range 0.2, unit 0.3, condition 0.2) and then sum them. If the result exceeds the 50% threshold, record it as a description conflict.

[0024] For conflict descriptions, first calculate the context relevance by counting the frequency of association of this core element with other elements in the knowledge network. For example, the associations of "engine power" with "fuel consumption" and "speed" appear 10 times, and the relevance is calculated as "actual association frequency ÷ maximum possible association frequency" and is 80%. Combining the conflict resolution rules (such as "authoritative manuals take precedence over ordinary reports" and "the latest version takes precedence over the old version"), if one conflict text comes from an authoritative manual in 2024 (relevance 80%) and the other comes from an ordinary report in 2022 (relevance 60%), then select the value in the authoritative manual as the reference value, correct the part that does not match the reference value in the conflict description. For example, correct "110kW" in the ordinary report to "120kW", and add a note "The original 110kW was corrected to the reference value 120kW due to the lower authority of the source" to generate a consistent description.

[0025] After integrating the core elements and consistent descriptions, the hierarchical attribution is calculated according to predefined classification rules (such as "discipline field → technical category → specific parameters"), and the elements are checked to see whether they contain keywords in the fields of "mechanics" and "electronics", and matched to the field of "mechanical engineering" (matching degree 90%); then it is determined whether they belong to the major categories such as "power system" and "transmission system", and matched to "power system" (matching degree 85%); finally, the specific level is determined and the label "mechanical engineering-power system-engine power" is added. At the same time, the missing field marks in the verification stage and the correction records in the conflict resolution stage are entered, and all information is arranged according to the fixed structure of "element name → consistent description → verification value → conflict record → classification label" to form a structured parameter library.

[0026] Through the precise calculation of the missing field ratio in the format check, the completeness of the core element description can be strictly controlled. When the missing ratio exceeds the threshold, an alarm will be issued in time to prevent incomplete information from entering the parameter library and ensure that each parameter contains the necessary key information. The conflict weight calculation analyzes the differences from multiple dimensions such as value, range, unit, and condition, and determines the benchmark value based on the context relevance and resolution rules, and makes targeted corrections to the conflict description. This process eliminates the description contradictions of documents from different sources, makes the parameter values ​​closer to the actual situation, and reduces data errors. The description differences of the same elements are identified through similarity calculation, and then a consistent description is generated through cross-validation, so that the same core element has only a unique standardized expression in the parameter library, avoiding the parameters caused by different document sources. Chaos is eliminated, ensuring the unity of data in the entire knowledge network; hierarchical classification labels are calculated based on clear rules, so that parameters are arranged in order according to logical levels. Users can quickly locate the required parameters through classification labels, reducing retrieval time and improving the convenience of using the parameter library; the structured parameter library contains information such as check values ​​and conflict resolution records, which clearly records the entire process from parameter extraction, verification to correction. Users can trace the source and processing details of each parameter, enhancing the credibility and transparency of the data. The automated calculation process replaces the traditional tedious work of manual sentence-by-sentence comparison and verification. The alarm mechanism and automatic marking function reduce the workload of manual investigation. The application of conflict resolution rules also reduces the difficulty of manual judgment and improves the efficiency of parameter library construction.

[0027] In a preferred embodiment of the present invention, the above step 3, based on the structured parameter library, constructs a template library, analyzes user needs in combination with semantic matching, and automatically generates a standard document outline, which may include: Step 330: Based on the classification labels and the parameter system, matching the preset chapter framework rules, and establishing a mapping relationship table between parameter classification labels and document chapters; Step 331: extract domain keywords, standard type codes, and core constraint condition expressions based on the mapping relationship table, calculate the matching degree with the template, and locate the target chapter structure; Step 332 , based on the target chapter structure, automatically generate a term definition set for the standard document by indexing the parameter system that matches the domain keywords and the standard type code from the structured parameter library; Step 333: Generate quantitative indicator threshold clauses and verification method clauses based on the core constraint expression, combined with the performance parameters and test method data; In step 334 , the term definition set is integrated with the core clause framework to generate a standard document outline.

[0028] In an embodiment of the present invention, first, all hierarchical classification labels are extracted from the structured parameter library. These labels are hierarchically divided according to fields, standard types, etc. For example, large technical fields are divided first, and then specific standard categories are divided under each field. At the same time, predefined chapter framework rules are obtained from the standardization basic specifications to clarify the subject scope, covered content and other information of each chapter; then, each hierarchical classification label is compared one by one with the chapters in the chapter framework rules. For each classification label, the core meaning and the content scope involved are analyzed, and then whether the corresponding chapter theme is related to it is checked. For example, if the classification label is "electronic equipment safety" and the chapter theme is "safety requirements", it is judged that the two are associated; then, the associated classification labels and chapters are recorded and a preliminary corresponding relationship is established. After that, the preliminary established corresponding relationship is verified to check whether there is a situation where a classification label corresponds to multiple chapters or a chapter corresponds to multiple classification labels. If so, the closeness of the association is further analyzed, and finally a unique mapping relationship is determined to form a mapping relationship table.

[0029] Based on the mapping relationship table generated in step 330, start extracting relevant information. First, analyze user needs and related data through semantic rules to extract domain keywords, standard type codes and core constraint expressions. Domain keywords are words that can reflect the field to which the current standard document belongs, such as "mechanical manufacturing" and "information technology"; standard type codes are specific codes that identify standard types, such as codes representing national standards, industry standards, etc.; core constraint expressions are descriptions of the core requirements of the standards.

[0030] When calculating the matching score between domain keywords and chapter topics, first determine the key words contained in the chapter topic, then compare the domain keywords with these key words, count the number of words in the domain keywords that are identical or semantically similar to the chapter topic keywords, and then divide it by the total number of domain keywords to get a basic ratio. Then adjust it according to the importance of the words. If important words are successfully matched, a certain score will be added to finally get the matching score.

[0031] When calculating the compatibility weight between the standard type code and the template type, first clarify the standard type code range corresponding to different template types in the template library, and compare the extracted standard type code with the code range corresponding to each template type to see which template type's code range the code belongs to. If it completely belongs to the code range of a certain template type, the compatibility weight is 1; if there is partial overlap, the weight is determined according to the degree of overlap. The higher the proportion of the overlapping part, the greater the weight, and vice versa; then, perform a weighted calculation, multiply the matching score by the corresponding weight, and then multiply the compatibility weight of the standard type code and the template type by the corresponding coefficient. The two are added together to obtain a comprehensive score. The chapter structures in the template library are sorted from high to low according to the comprehensive score, and the one with the highest score is the target chapter structure.

[0032] According to the target chapter structure determined in step 331, the fields and standard types involved in the chapter structure are clarified, and the parameter systems related to the current field keywords are screened out from the structured parameter library. This requires comparing the description information in the parameter system with the field keywords, and retaining those parameter systems whose descriptions contain field keywords or are semantically related. At the same time, according to the standard type coding, the parameter systems are further screened, and only the parameter systems under the standard type corresponding to the coding are retained. The screened parameter systems are analyzed and the key terms therein are extracted. These terms are words that appear repeatedly in the parameter system and have specific meanings. Then, a definition is determined for each key term. The content of the definition comes from the explanation and description of the term in the parameter system. These terms and corresponding definitions are organized into a set to form a normative term definition set. During the organization process, the accuracy and consistency of the term definitions must be ensured to avoid ambiguity.

[0033] First, parse the core constraint expression to clarify the performance parameters, test methods, and constraint requirements involved. Obtain performance parameter data related to the core constraint expression from the structured parameter library. These data include the value range and common values ​​of the performance parameter under different circumstances. According to the constraint requirements in the core constraint expression, determine the quantitative indicator threshold. For example, if the constraint is "the operating temperature of the equipment shall not exceed 80 degrees Celsius", then the quantitative indicator threshold clause is "the operating temperature threshold of the equipment is 80 degrees Celsius". When determining the threshold, refer to the performance parameter data in the parameter library to ensure the rationality of the threshold. If the parameter The common maximum value of this performance parameter in the library is 75 degrees Celsius. Combined with the constraints, the threshold may be set at 80 degrees Celsius. For the verification method clause, it is generated based on the core constraint expression and the test method data in the parameter library. The test steps, test equipment, test environment and other information applicable to the current performance parameters in the test method data are analyzed, and this information is organized into standardized clause content to determine how to verify the quantitative indicator threshold. For example, if the performance parameter is temperature, and the test method data includes a method of measuring at a specific location with a thermometer, the verification method clause will describe in detail the steps and requirements for measuring at the specified location using the thermometer.

[0034] Sort out the term definition set generated in step 332, arrange them according to the importance and logical order of the terms, and ensure that they are presented clearly in the outline. At the same time, organize the core clause framework generated in step 333, including quantitative indicator threshold clauses and verification method clauses, and clarify the logical relationship between the clauses. For the scope definition part, determine it according to the target chapter structure and field keywords, and clarify the scope of application of the standard document, such as applicable product types, application scenarios, etc., which are obtained by analyzing the scope covered by the subject of the target chapter structure and the field keywords; then, integrate the scope definition, term definition set and core clause framework in a certain logical order, first list the scope definition to clarify the applicable boundaries of the document; then the term definition set to ensure that readers have a unified understanding of the key terms in the document; finally the core clause framework, list the contents of each clause in detail. During the integration process, check whether the connection between the parts is smooth and whether the content is complete, and finally form the outline of the standard document.

[0035] Through detailed calculation processes, such as the precise calculation of the matching degree between domain keywords and chapter themes, and the matching degree between standard type codes and template types, the target chapter structure can be accurately located to ensure that the generated outline is highly consistent with user needs and standard specifications, reducing deviations caused by human judgment errors. When generating a set of normative term definitions, the accuracy and consistency of term definitions are ensured based on the parameter system that has been screened and verified in the structured parameter library, avoiding the problem of different definitions of the same term in different documents, making it easier for readers to understand and use standard documents. When generating quantitative indicator threshold clauses and verification method clauses, the performance parameters and test method data in the structured parameter library are referenced. , combined with the core constraint expression, the quantitative indicator threshold is consistent with the actual situation, the verification method is operational, and the scientificity and practicality of the standard document are improved. The entire process reduces manual intervention through automated calculation and analysis steps. From establishing mapping relationships to integrating and generating outlines, each link has a clear calculation process, which can quickly process large amounts of data and information, shortening the generation time of the standard document outline. Through a step-by-step integration process, the scope definition, term definition set and core clause framework are orderly integrated to ensure that the standard document outline contains all necessary components, and the logic between each part is clear and the connection is smooth, which improves the overall quality of the standard document.

[0036] In a preferred embodiment of the present invention, the above step 4, based on the standard document outline, searching for similar document fragments in the knowledge base, and completing the standard chapter content writing in combination with semantic generation to generate a standard document draft, may include: Step 440 , breaking down the standard document outline into chapters, generating chapter feature vectors and logical position identifiers, and calculating similarity scores between the segment set and the current chapter based on the chapter feature vectors; Step 441 , adjusting the segment set with similarity scores exceeding a threshold value in combination with the logical position identifier to obtain a reorganized clause unit; Step 442 , matching the reorganized clause unit with the structured parameter library, retrieving missing parameter items and injecting quantitative index thresholds and test conditions to obtain clause content with complete parameters; Step 443: Perform industry terminology compliance verification on the parameter completion clauses to generate a complete standard document draft.

[0037] In an embodiment of the present invention, first, the standard document outline is disassembled into chapters, and the title, keywords and core content description of each chapter are extracted. For each chapter, its feature vector is generated. This process requires analyzing the text content in the chapter to determine the key concepts, terms and the relationship between them. For example, the feature vector is constructed by analyzing the high-frequency words, important phrases and their semantic associations in the chapter; at the same time, a logical position identifier is generated for each chapter. This identifier contains the hierarchical information of the chapter in the entire outline, the association relationship between the previous and next chapters, etc. For example, the identifier may contain information such as the first-level title and second-level title to which the chapter belongs, as well as the sequential position of the chapter in chapters of the same level.

[0038] Based on the generated chapter feature vector, a search is performed in the standardized knowledge base to find matching clause fragments. For each clause fragment, its similarity score with the current chapter is calculated. This calculation process involves semantic analysis of the clause fragment and the chapter feature vector. First, the keywords and topics in the clause fragment are determined, and then these keywords and topics are compared with the content in the chapter feature vector. The number of identical or semantically similar keywords in the two is counted, and the importance of these keywords is taken into consideration. Successful matching of important keywords will obtain a higher score, and finally the similarity score between the clause fragment and the current chapter is obtained.

[0039] For the set of fragments whose similarity scores exceed the threshold, contextual coherence adjustment is performed in combination with the logical position identifier. First, the expected position and contextual relationship of these fragments in the standard document are determined based on the logical position identifier. For example, it is determined whether the fragment belongs to the introduction, main content or conclusion, and its logical coherence with the previous and following content. Then, the fragment is adjusted so that it can naturally integrate into the context of the current chapter. This may include modifying some expressions in the fragment to make it consistent with the overall style and terminology of the chapter; or rearranging the order of the fragments to ensure logical coherence. For example, if some fragments in the fragment set have a semantically progressive relationship, and their order in the original knowledge base may not conform to this progressive relationship, their order needs to be adjusted.

[0040] Match the reorganized clause units with the structured parameter library and retrieve missing parameter items. This process requires a detailed analysis of the content in the clause unit to determine the parts that require parameter support. For example, the clause may mention that "the operating temperature of the equipment should meet the requirements", but does not specify the specific temperature range. This requires retrieving the corresponding temperature parameters from the structured parameter library. Once the missing parameter items are determined, the corresponding quantitative indicator thresholds and test conditions are extracted from the structured parameter library. For example, for the temperature parameters mentioned above, the specific temperature thresholds are obtained from the parameter library, such as "the operating temperature range of the equipment is -20℃ to 60℃", and the corresponding test conditions, such as "under standard atmospheric pressure, use a thermometer with an accuracy of ±0.5℃ for measurement". These retrieved parameter items are injected into the clause unit to ensure that the clause content contains all necessary parameter information and obtain clause content with complete parameters. During the injection process, the accuracy and applicability of the parameters must be ensured to match the context of the clause.

[0041] Performing industry terminology compliance verification on clauses with complete parameters requires establishing an industry terminology standard library, which contains standard terms and abbreviations commonly used in the target industry and their correct usage. Compare the terms in the clauses with the standard library to check whether there are any term expressions that do not comply with industry standards; if non-compliant terms are found, make corresponding corrections, for example, replace the non-standard synonyms used in the clauses with standardized terms in the industry; or check the abbreviations of the terms to ensure that they comply with industry conventions, generate chapters and paragraphs that comply with the expression standards of the target industry, and integrate these chapters and paragraphs in the order of the chapters in the standard outline. During the integration process, ensure that the transition between chapters is natural and smooth, and the logic is coherent. Check whether the reference relationship between chapters is correct, and whether there are any duplicate or contradictory content, and finally generate a complete draft of the standard document.

[0042] Through the automated chapter decomposition, matching retrieval and content generation process, the time and workload of manually writing standard documents are reduced, relevant content can be quickly found and integrated from the knowledge base, and the speed of document generation is improved; parameter injection based on the structured parameter library ensures that the quantitative indicators and test conditions contained in the document are accurate. At the same time, through industry terminology compliance verification, the use of terminology in the document is guaranteed to comply with industry standards, improving the consistency and professionalism of the document content; contextual connection adjustment ensures the logical coherence of the document content and makes the document structure more reasonable; parameter completeness processing ensures that the clauses in the document are complete and contain all necessary information, thereby improving the overall quality of the document; through case matching retrieval of similar document fragments in the knowledge base, the accumulated knowledge and experience in the past can be fully utilized to avoid duplication of work. At the same time, it also ensures that the generated draft standard document has a certain degree of authority and reliability. The generated draft standard document is based on a standardized outline and knowledge base, with a clear structure and standardized content, which makes document maintenance and updating more convenient. A new version of the standard document can be quickly generated by simply modifying the relevant content in the knowledge base.

[0043] In a preferred embodiment of the present invention, step 5 above, based on the draft standard document, converts key data into point cloud data, determines reference points, and delineates a detection area; sets detection points within and outside the area, constructs an arc detection path, and generates path correction parameters to obtain an audit conclusion and modification suggestions, which may include: Step 550 , based on the draft standard document, extract technical parameters, test methods, and performance indicators as key data, and map them to a three-dimensional coordinate system to form a point cloud set; Step 551 , based on the point cloud set, using the point cloud of the industry mandatory compliance indicator as a reference point, construct a baseline and a sector-shaped detection area; Step 552: setting dual detection points within the sector-shaped detection area to verify the synergistic relationship between parameters and the matching degree between the parameters and the test method; Step 553: Based on the coordinates of the reference point, the dual detection point, and the single detection point, they are connected according to the original time sequence to generate an arc detection path, and the path curvature change rate is analyzed to generate path correction parameters; Step 554 , integrating the parameter fluctuation information reflected by the path correction parameters with the content of the draft standard document, performing a comprehensive verification to obtain a multi-dimensional verification result; In step 555 , the path correction parameters are converted into parameter compliance risk levels, and combined with the logic conflict items, terminology deviation items, and structure missing items identified in the verification results to obtain the audit conclusion and modification suggestions.

[0044] In an embodiment of the present invention, a text recognition tool is used to comprehensively comb through the draft of the standard document chapter by chapter and section by section. During the combing process, the technical parameters in the document are accurately located and extracted, such as product specifications, material performance parameters, etc.; test methods, including experimental processes, operating specifications, etc.; performance indicators, such as product qualification rate and service life, etc., are determined as key data; then, clear mapping rules are formulated for each attribute of these key data. For example, the numerical size of the technical parameters, the complexity of the test methods, the compliance requirements of the performance indicators and other attributes are all corresponding to the X-axis, Y-axis, and Z-axis of the three-dimensional coordinate system. According to this rule, the various attributes of each key data are converted into specific coordinate values ​​in the three-dimensional coordinate system, so that each key data corresponds to a point in the three-dimensional space; when all key data have completed such a conversion, these points are gathered together to form a point cloud set, and each point clearly represents a key data in the document and all its attribute characteristics.

[0045] From the already formed point cloud collection, according to the mandatory compliance standards within the industry, select the point clouds corresponding to the indicators that must be strictly followed, and identify these point clouds as reference points, because they are the core reference for document compliance. The cuckoo algorithm is introduced, and each point in the point cloud collection is regarded as a "bird's nest", and the reference point is the initial and final "bird's nest". The process of the cuckoo searching for a high-quality nest is simulated, and the degree of correlation between each point cloud and the reference point is calculated. Specifically, the correlation between the key data represented by each point cloud and the mandatory compliance indicators represented by the reference point is analyzed, such as whether the parameters revolve around the mandatory indicators. , whether the test method is based on mandatory indicators, etc. Through this analysis, those adjacent points with very prominent correlation with the benchmark points are screened out; then, these adjacent points with prominent correlation are connected with the benchmark points in sequence to construct a baseline. This baseline reflects the main correlation path between key data and core compliance indicators. Then, with this baseline as the central axis, the preset threshold range is expanded to the surrounding area. This threshold range is determined based on the industry's conventional error tolerance range and the degree of data correlation. After expansion, a fan-shaped detection area is formed, which concentrates all key data points that are closely related to the benchmark points.

[0046] Two inspection points are set up inside the sector-shaped detection area. The workflow of the first inspection point is to collect the technical parameters represented by all point cloud data in the area, and then conduct a comprehensive analysis of these parameters. For example, it analyzes whether the numerical matching between different technical parameters is reasonable, and whether there is a lack of coordination caused by parameters that are too high or too low, so as to verify the coordination relationship between the parameters and ensure that the various parameters can cooperate with each other and operate normally in actual applications; the second inspection point also collects point cloud data in the area, focusing on the technical parameters and the corresponding test methods, and analyzing whether the test method can accurately test the technical parameters. For example, for a certain accuracy parameter, whether the measurement accuracy of the test method matches it, and whether the test steps can effectively reflect the actual situation of the parameter, so as to verify the matching degree of the parameter and the test method; outside the sector-shaped detection area, according to the distribution of point clouds outside the area, a representative location is selected to set a single inspection point. This single inspection point is mainly used to cover key data that are not in the core detection area but are equally important, to ensure that all key data in the document can be included in the detection range and avoid detection loopholes.

[0047] First, accurately record the specific coordinate positions of the reference points, double-check points, and single-check points in the three-dimensional coordinate system. Then, review the draft standard document to determine the order in which the key data represented by these points appear in the document, that is, the original time sequence. According to this original time sequence, the reference points, double-check points, and single-check points are connected in sequence. Because these points have different positions in three-dimensional space and appear in a certain order, they naturally form an arc-shaped detection path after connection. This path fully reflects the presentation logic and order of the key data in the document. Next, analyze the generated arc-shaped detection path, focusing on the change in the curvature of the path. The curvature change rate reflects the change in the degree of curvature of the path. When the curvature change rate fluctuates abnormally, it indicates that there may be problems with the key data at the corresponding location. Drawing on the idea of ​​optimizing low-quality bird nests in the cuckoo algorithm, the reasons for these abnormal curvature change rates are analyzed, such as whether the key data is inconsistent or the logic is incoherent. Based on the analysis results, path correction parameters are generated. These parameters record the location in the path that needs to be corrected, the direction of correction, and the degree of correction.

[0048] Key data fluctuations reflected by path correction parameters, such as abnormal changes in parameter values ​​and irregularities in logical relationships, are mapped and integrated with the specific content of the draft standard document. A comprehensive verification process is conducted, starting with rule verification, comparing the document's content to relevant industry standards, specifications, and regulations to check whether it complies with these requirements. For example, whether technical parameters meet industry minimum standards and whether test methods comply with industry operating specifications. Next, terminology verification is conducted to verify whether the professional terminology used in the document is consistent with common industry terminology, and to check whether there are any confusion or errors in terminology, ensuring the accuracy and standardization of terminology. Finally, structural analysis is conducted to analyze whether the document's overall framework is reasonable, whether the chapter arrangement is logically clear, whether the content is complete, and whether there are any structural gaps or disorganized chapters. The results of rule verification, terminology verification, and structural analysis are integrated with parameter fluctuation information to obtain comprehensive, multi-dimensional verification results.

[0049] Based on the path correction parameters and combined with the industry's risk assessment standards, parameter fluctuations are converted into specific parameter compliance risk levels. If the parameter fluctuation is small and within the allowable range, the corresponding risk level is low. If the parameter fluctuation is large, exceeding a certain range but still adjustable, the risk level is medium. If the parameter fluctuation is extremely large and seriously fails to meet the requirements, the risk level is high. Combined with the logical conflicts identified in the multi-dimensional verification results, such as logical contradictions between technical parameters, logical mismatches between test methods and performance indicators, terminology deviations, such as incorrect terminology use and non-standard terminology expressions, and structural omissions, such as missing important chapters and incomplete content, the overall compliance and completeness of the document are evaluated based on the parameter compliance risk level and these specific problem items, and a graded review conclusion is generated, such as passed, minor modification required, major modification required, etc. For each problem item, combined with the correction direction and degree indicated by the path correction parameters, specific and feasible modification suggestions are formulated to form a corresponding set of modification suggestions, which clearly indicate the content that needs to be modified, the modification method and the modified goals.

[0050] By converting key data into point clouds and constructing three-dimensional coordinates, combined with the cuckoo algorithm for precise screening and correlation analysis of key data points, it is possible to deeply explore the intrinsic relationship between key data. This method breaks through the limitations of traditional manual review that checks the text surface, and can accurately discover deep-seated problems such as improper parameter coordination and mismatched test methods, thereby improving the accuracy of the review; setting reference points, double detection points and single detection points, constructing fan-shaped detection areas and arc-shaped detection paths, forming a comprehensive detection network, which not only focuses on key data in core compliance areas, but also covers important data outside the area. In addition, the comprehensive verification of rule verification, terminology detection and structural analysis ensures a comprehensive inspection of all aspects of the document content and avoids audit omissions. After the introduction of the cuckoo algorithm, a certain degree of automated analysis and screening is achieved in the processes of point cloud clustering, detection point optimization and path correction, reducing manual one-by-one The tedious work of comparison and analysis, at the same time, the structured audit steps and clear verification standards enable auditors to carry out their work more efficiently, shorten the audit cycle, and improve the audit efficiency. The generated set of modification suggestions is based on specific parameter fluctuation information, logical conflict items, terminology deviation items and structural missing items, combined with the parameter compliance risk level, and is highly targeted. Each suggestion clearly points out the problem and the direction of modification, so that document compilers can clearly know what needs to be modified and how to modify it, which greatly enhances the practicality of the modification suggestions. It combines point cloud technology, time series analysis and cuckoo algorithm to build an intelligent standard document review system. This system changes the traditional audit model that relies on manual experience, realizes the digitization and intelligence of the audit process, provides new technical means and methods for the field of standard document audit, and promotes the intelligent development of industry audit work.

[0051] In a preferred embodiment of the present invention, the above step 6, integrating the audit conclusions and modification suggestions into the knowledge base, and adjusting the parameter weights and generation rules through a machine learning mechanism, may include: Step 660 : Integrate the review results with the manual revision feedback to generate a fused feedback dataset, and associate the revision opinions with the conflict description nodes of the standardized knowledge network to generate associated node information; Step 661: extract the error type distribution and revision operation traces in the audit record based on the associated node information to generate an error distribution dataset; Step 662: Store the error distribution dataset into the knowledge base to form a review case set with weight labels, and analyze it through a machine learning mechanism to generate conflict weight threshold parameters; Step 663: Based on the conflict weight threshold parameter, dynamically update the terminology compliance verification rules and adjust the mapping relationship weight between the chapter framework and the parameter system in the template library.

[0052] In an embodiment of the present invention, a detailed weighted fusion calculation is performed on the audit results and manual revision feedback based on the previously generated correction parameters. The correction parameters comprehensively consider multiple factors such as the semantic density and conflict records of each sub-region. It is like an "intelligent regulator" that can assign different weights to the audit results and manual revision feedback according to the actual conditions of different sub-regions. The distribution of weights has a strict basis. For those sub-regions with high semantic density and few conflict records, it means that the content quality of the region is relatively high. The weight of manual revision feedback in these regions is appropriately reduced, because these regions may only require a small amount of manual intervention to achieve good results; on the contrary, for those sub-regions with more conflict records and lower semantic density, the weight of manual revision feedback is increased, because these regions require more manual expertise to be corrected and improved. In the weighted fusion calculation process, each piece of information in the audit results and manual revision feedback is deeply analyzed and compared. For the erroneous information found in the audit results and the revision suggestions put forward, the views and handling methods of these issues in the manual revision feedback are carefully studied.

[0053] If the manual revision feedback is consistent with the audit results, the credibility of this information will be enhanced and it will be retained as an important reference. If the manual revision feedback is inconsistent with the audit results, we will not rush to make a judgment, but will further analyze the reasons and refer to more data to see which treatment method is more effective in similar situations. We will also use domain knowledge to evaluate the rationality of the two opinions from a professional perspective. Through this rigorous analysis and comparison, we can ensure the accuracy and reliability of the fused feedback data. After completing the weighted fusion calculation, we will generate a fused feedback data set. This data set contains the integrated and verified audit results and manual revision feedback information. The revision opinions in the fused feedback data set are associated with the conflict description node of the standardization knowledge network. The standardization knowledge network is a huge and complex knowledge system, like a "knowledge network" that organically connects various domain knowledge and standard specifications.

[0054] The conflict description nodes are like the "key points" on this network, recording various possible conflict situations and corresponding solutions. Through semantic analysis and pattern matching methods, the revision opinions are accurately associated with the conflict description nodes. During the association process, associated node information is generated. This information records in detail the association relationship between the revision opinions and the conflict description nodes, including the strength of the association, the basis of the association, etc. The strength of the association reflects the closeness between the revision opinions and the conflict description nodes, and the basis of the association explains why they are associated together. These associated node information provide important basis for knowledge integration and rule adjustment, just like providing a detailed "map" for decision-making.

[0055] The associated node information is deeply integrated with the fusion feedback dataset. During the integration process, advanced data matching and merging technologies are used to accurately match and organically merge each record in the associated node information with the corresponding record in the fusion feedback dataset. For example, for a certain revision opinion, carefully find the conflict description node information corresponding to it in the associated node information, and add this information to the fusion feedback dataset to form a more complete and rich record. Through this integration, the revision opinions can be organically combined with the relevant knowledge in the standardization knowledge network, and the error type distribution and revision operation trajectory can be extracted from the audit records. The error type distribution refers to the number and proportion of various error types found during the audit process, and each error information in the audit record is carefully classified and counted.

[0056] Errors are divided into different types, such as terminology errors, format errors, logical errors, data errors, etc. Then, the frequency of each type of error is calculated to analyze which types of errors are more common and which types of errors are more rare. Through this analysis, we can understand the key and difficult issues in the audit process. The revision operation trajectory refers to the detailed record of various operations performed during the revision process. A comprehensive and detailed record is made of each revision operation in the audit record, including the time of the operation, the content of the operation, the executor of the operation, the purpose of the operation and other information. By analyzing the revision operation trajectory, we can understand the behavioral habits and preferences in the revision process, such as which revision method is more inclined to adopt, and what different strategies will be adopted when dealing with different types of problems. The extracted error type distribution and revision operation trajectory information are integrated to generate an error distribution dataset. This dataset contains the distribution of various error types in the audit process and the trajectory information of the revision operation.

[0057] The error distribution dataset is stored in a knowledge base, forming a weighted audit case set. The knowledge base is a vast knowledge storage system, like a "knowledge library," storing a wide range of domain knowledge and audit experience. During the storage process, a weight label is added to each error record and revision operation. The size of the weight label is not determined arbitrarily, but is assigned based on the severity of the error and the effectiveness of the revision operation. Errors that seriously affect document quality or may lead to misunderstandings or incorrect applications are assigned higher weight labels because these errors require more attention and attention. Effective revision operations, such as revision methods that can quickly and accurately resolve issues, are also assigned higher weight labels because these operations have high reference value. Adding weight labels enables more refined management and analysis of audit cases, providing a clearer understanding of the value and significance of each case. The stored audit case set is analyzed through a machine learning mechanism to generate conflict weight threshold parameters. The machine learning mechanism uses a variety of advanced algorithms and technologies to deeply mine and analyze the data in the audit case set. First, the machine learning algorithm conducts a comprehensive and in-depth analysis of the error type distribution and revision operation trajectory, attempting to identify hidden patterns and regularities.

[0058] For example, the algorithm may find that certain types of errors often appear in specific chapters or areas, which may suggest that these chapters or areas have some common problems and require special attention; or the algorithm may find that certain revision operations are more effective in resolving specific types of errors. Based on the laws and patterns discovered, the machine learning algorithm will perform complex calculations and analysis, and ultimately generate weight threshold parameters for various conflict situations; the conflict weight threshold parameter is a very important indicator used to judge the severity of the conflict. When the weight of a conflict exceeds this threshold, the conflict will be considered serious and need to be dealt with in a focused manner, requiring more resources and energy to resolve it; conversely, when the weight of a conflict is lower than this threshold, the conflict will be considered minor and can be handled generally without excessive attention. Through continuous learning, the machine learning mechanism can dynamically adjust the conflict weight threshold parameter according to changes in the audit case set, so that it can more accurately reflect the actual situation.

[0059] The conflict weight threshold parameter is applied to the semantic matching process, and the term compliance verification rules are dynamically updated. During the semantic matching process, the similarity between the terms in the document and the industry standard terms will be carefully compared and analyzed. The conflict weight threshold parameter will have an important impact on the calculation method of this similarity. For some terms marked as high conflict weights, it means that these terms are prone to problems during use. The strictness of their matching is increased, and these terms are required to have a higher similarity with the industry standard terms. Only when a certain similarity threshold is reached will they be considered compliant. This can effectively reduce the problem of irregular use of terms and improve the professionalism and accuracy of the documents. The term compliance verification rules are dynamically adjusted according to the conflict weight threshold parameter. If the conflict weight of a term is high, it means that the term has a greater risk during use. The compliance verification of the term will be strengthened and more verification conditions and rules will be added.

[0060] Check whether the context in which the term is used complies with industry standards and whether there is any possibility of ambiguity or misunderstanding; check whether the logical relationship between the term and other related terms is correct. On the contrary, if the conflict weight of a term is low, it means that the term is relatively safe during use, which will simplify the compliance verification of the term, reduce unnecessary verification steps, and improve verification efficiency. By dynamically updating the verification rules, it is possible to more accurately identify term compliance issues in the document, improve the accuracy and efficiency of the verification, and ensure that the use of terminology in the document complies with industry standards and specifications.

[0061] Apply the conflict weight threshold parameter to the template library to adjust the mapping relationship weight between the chapter framework and the parameter system. The template library contains various standard document templates. Each template defines the chapter framework and parameter system of the document. The chapter framework stipulates the structure and chapter division of the document, and the parameter system contains various parameters and data required in the document. There is a certain mapping relationship between the chapter framework and the parameter system. This mapping relationship determines how to accurately map the content in the parameter system to the corresponding position in the chapter framework when generating the document. The conflict weight threshold parameter will affect the weight of this mapping relationship. For some chapters and parameters with more conflicts found during the review process, it means that there are certain problems in the matching between these chapters and parameters, and the weight of the mapping relationship between them should be increased.

[0062] When generating documents, we pay more attention to the matching degree between these chapters and parameters, and invest more energy to ensure that the parameter content can be accurately mapped to the appropriate chapter position to reduce the occurrence of conflicts. On the contrary, for some chapters and parameters with fewer conflicts, we reduce the mapping relationship weights between them. By adjusting the mapping relationship weights, we can optimize the structure and function of the template library and improve the quality and efficiency of document generation. When using the template library to generate documents, we can more reasonably allocate parameter content to each chapter according to the conflict weight threshold parameters, making the document structure more reasonable, the content more coherent, and the logic stronger, reducing the occurrence of conflicts, and improving the quality of the document and user experience.

[0063] By weightedly integrating audit results and manual revision feedback and linking them to a standardized knowledge network, the knowledge base can be continuously enriched and improved. The knowledge in the knowledge base will be more accurate and comprehensive, better reflecting industry standards and actual applications. The machine learning mechanism's analysis and learning of audit case sets will enable the continuous discovery of new patterns and rules, improving intelligence and decision-making capabilities. Parameter weights and generation rules can be dynamically adjusted based on data and current circumstances to better adapt to different document audit needs. The application of conflict weight threshold parameters will focus more on high-risk areas and issues during the audit process, improving the relevance and efficiency of audits. By dynamically updating terminology compliance verification rules and adjusting template library mapping relationship weights, potential issues can be discovered and resolved in advance, reducing manual intervention and audit costs. By verifying rules and template library mapping relationship weights, generated documents will be more compliant with industry standards and specifications, reducing errors and conflicts. This will enhance the logic, coherence, and accuracy of documents, thereby increasing their value and impact. Parameter weights and generation rules will be continuously adjusted based on accumulated audit cases and feedback.

[0064] like Figure 2 As shown, the embodiment of the present invention also provides a large model-based standard document automatic generation and multi-dimensional review system, including: The knowledge network module is used to build a distributed database of multi-source documents and analyze heterogeneous texts through natural language processing to obtain a standardized knowledge network; Parameter extraction module, which is used to extract indicator elements based on the standardized knowledge network and form a structured parameter library through verification and validation; The document outline module is used to build a template library based on the structured parameter library, analyze user needs in combination with semantic matching, and automatically generate a standard document outline; The document content module is used to retrieve similar document fragments from the knowledge base based on the standard document outline, and complete the standard chapter content writing in combination with semantic generation to generate a standard document draft; The audit and assessment module is used to convert key data into point cloud data based on the draft standard document, determine the reference points and delineate the inspection area; set inspection points inside and outside the area, construct an arc inspection path and generate path correction parameters to obtain audit conclusions and modification suggestions; The knowledge feedback module is used to integrate the review conclusions and modification suggestions into the knowledge base, and adjust the parameter weights and generation rules through machine learning mechanisms.

[0065] It should be noted that this system is a system corresponding to the above method, and all implementation methods in the above method embodiment are applicable to this embodiment and can achieve the same technical effects.

[0066] An embodiment of the present invention further provides a computing device comprising: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the above-described method. All implementations in the above-described method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0067] The embodiment of the present invention further provides a computer-readable storage medium storing instructions, which, when executed on a computer, causes the computer to execute the above-described method. All implementations in the above-described method embodiment are applicable to this embodiment and can achieve the same technical effects.

[0068] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. The method of automatic generation and multi-dimensional audit of standard documents based on large models is characterized by: The method comprises: Step 1: Establish a distributed database of multi-source documents and analyze heterogeneous texts through natural language processing to obtain a standardized knowledge network; Step 2: Based on the standardized knowledge network, extract indicator elements and form a structured parameter library through verification and validation; Step 3: Based on the structured parameter library, a template library is constructed, user needs are analyzed in combination with semantic matching, and a standard document outline is automatically generated; Step 4: Based on the standard document outline, similar document fragments are retrieved from the knowledge base, and the standard chapter content is written in combination with semantic generation to generate a draft standard document; Step 5: Based on the draft standard document, convert key data into point cloud data, determine the reference points, and delineate the inspection area. Set inspection points inside and outside the area, construct an arc inspection path, and generate path correction parameters to obtain audit conclusions and modification suggestions. Step 6: Integrate the review conclusions and modification suggestions into the knowledge base, and adjust the parameter weights and generation rules through machine learning mechanisms.

2. The method for automatic generation and multi-dimensional review of standard documents based on a large model according to claim 1 is characterized in that: Based on the standardized knowledge network, the indicator elements are extracted and verified to form a structured parameter library, including: Based on the standardized knowledge network, semantic analysis is used to identify and extract the three core elements of the knowledge network: quantitative indicators, performance parameters, and test methods, and obtain descriptive text fragments; Match descriptive text snippets with pre-set standardized formats, check for missing fields, and trigger alerts; For the same core element that has been verified, the similarity of descriptive text fragments in different source documents is calculated, and description conflicts whose conflict weight exceeds the preset threshold are identified and recorded; For description conflicts, cross-validation is performed based on context relevance and preset rules to determine the benchmark value and make numerical corrections to generate consistent descriptions. Integrate the core elements and consistent descriptions, calculate the hierarchical attribution and add classification labels to form a structured parameter library.

3. The method for automatic generation and multi-dimensional review of standard documents based on a large model according to claim 2 is characterized in that: Based on the structured parameter library, a template library is built. By combining semantic matching to analyze user needs, a standard document outline is automatically generated, including: Based on the classification labels and parameter system, match the preset chapter framework rules and establish a mapping relationship table between parameter classification labels and document chapters; Based on the mapping relationship table, domain keywords, standard type codes, and core constraint expressions are extracted, the matching degree with the template is calculated, and the target chapter structure is located; According to the target chapter structure, the parameter system of matching domain keywords and standard type codes is indexed from the structured parameter library to automatically generate the term definition set of the standard document; Based on the core constraint expression, combined with performance parameters and test method data, quantitative indicator threshold clauses and verification method clauses are generated; Integrate term definition sets with the core clause framework to generate a standard document outline.

4. The method for automatic generation and multi-dimensional review of standard documents based on a large model according to claim 3 is characterized in that: Based on the standard document outline, similar document fragments are retrieved from the knowledge base, and combined with semantic generation to complete the writing of standard chapter content and generate a draft standard document, including: The standard document outline is broken down into chapters, and chapter feature vectors and logical position identifiers are generated. Based on the chapter feature vectors, the similarity scores between the fragment set and the current chapter are calculated. For the segment sets whose similarity scores exceed the threshold, adjustments are made based on the logical position identifiers to obtain the reorganized clause units; Match the restructured clause units with the structured parameter library, retrieve missing parameter items, and inject quantitative indicator thresholds and test conditions to obtain clause content with complete parameters; Perform industry terminology compliance verification on parameter-complete clauses and generate a complete draft of the standard document.

5. The method for automatic generation and multi-dimensional review of standard documents based on a large model according to claim 4 is characterized in that: Step 5 includes: Based on the draft standard document, technical parameters, test methods, and performance indicators are extracted as key data and mapped to a three-dimensional coordinate system to form a point cloud set; Based on the point cloud collection, the point cloud of the industry mandatory compliance indicators is used as the reference point to construct the baseline and sector detection area; Set up dual detection points within the sector-shaped detection area to verify the synergistic relationship between parameters and the matching degree between parameters and test methods; Based on the coordinates of the reference point, double detection point and single detection point, they are connected according to the original time sequence to generate an arc detection path, and the path curvature change rate is analyzed to generate path correction parameters; Integrate the parameter fluctuation information reflected by the path correction parameters with the content of the draft standard document, perform a comprehensive verification, and obtain multi-dimensional verification results; Convert the path correction parameters into parameter compliance risk levels, and combine them with the logical conflicts, terminology deviations, and structural missing items identified in the verification results to obtain audit conclusions and modification suggestions.

6. The method for automatic generation and multi-dimensional review of standard documents based on a large model according to claim 5 is characterized in that: The comprehensive verification includes rule verification, calling the preset logical rule library to verify the causal rationality of technical terms; terminology detection, comparing with the standard terminology library to verify the consistency of professional terminology expressions; structure analysis, verifying the integrity of core chapters through the chapter weight mechanism.

7. The method for automatic generation and multi-dimensional review of standard documents based on a large model according to claim 6 is characterized in that: Integrate the review conclusions and modification suggestions into the knowledge base, and adjust parameter weights and generation rules through machine learning mechanisms, including: Integrate the review results with manual revision feedback to generate a fused feedback dataset, and associate the revision opinions with the conflict description nodes of the standardization knowledge network to generate associated node information; Based on the associated node information, the error type distribution and revision operation traces in the audit records are extracted to generate an error distribution dataset; The error distribution dataset is stored in the knowledge base to form a review case set with weight labels, and is analyzed through a machine learning mechanism to generate conflict weight threshold parameters; Based on the conflict weight threshold parameters, the terminology compliance verification rules are dynamically updated, and the mapping relationship weights between the chapter framework and the parameter system in the template library are adjusted.

8. A large-scale model-based standard document automatic generation and multi-dimensional review system, which implements the method according to any one of claims 1 to 7, characterized in that: include: The knowledge network module is used to build a distributed database of multi-source documents and analyze heterogeneous texts through natural language processing to obtain a standardized knowledge network; Parameter extraction module, which is used to extract indicator elements based on the standardized knowledge network and form a structured parameter library through verification and validation; The document outline module is used to build a template library based on the structured parameter library, analyze user needs in combination with semantic matching, and automatically generate a standard document outline; The document content module is used to retrieve similar document fragments from the knowledge base based on the standard document outline, and complete the writing of standard chapter content in combination with semantic generation to generate a standard document draft; The audit and assessment module is used to convert key data into point cloud data, determine benchmark points, and delineate inspection areas based on the draft standard document; Set up inspection points inside and outside the area, construct an arc inspection path and generate path correction parameters to obtain audit conclusions and modification suggestions; The knowledge feedback module is used to integrate the review conclusions and modification suggestions into the knowledge base, and adjust the parameter weights and generation rules through machine learning mechanisms.

9. A computing device, characterized in that include: one or more processors; A storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a program, which, when executed by a processor, implements the method according to any one of claims 1 to 7.

Images