AI-driven form intelligent generation and compliance verification method

Abstract

This invention provides an AI-driven intelligent form generation and compliance verification method, addressing the industry pain points of form compilation in high-compliance fields such as engineering construction. It overcomes the technical limitations of traditional template-driven form generation and unconstrained generation of general large models, constructing a dual-channel layered knowledge base consisting of a standard specification layer and a historical form layer. This achieves the separate modeling and collaborative utilization of authoritative compliance evidence and engineering practice experience. It pioneers a dual-channel step-by-step retrieval mechanism that first matches relevant standard clauses based on engineering natural language descriptions, and then retrieves corresponding historical practice structure templates, thus establishing a deep connection between compliance evidence and engineering practice. Furthermore, it proposes a semantic representation method based on the average value of the content vectors under the inspection project, achieving deduplication optimization and structured integration of multi-source templates through clustering and merging.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an AI-driven method for intelligent form generation and compliance verification. Background Technology

[0002] With the continuous improvement of the industry regulatory system and the increasing complexity of engineering projects, the requirements for compiling supervision and inspection forms are becoming increasingly stringent. Not only do they need to cover all aspects of the inspection content throughout the entire process and at all stages, but each inspection item also needs to have clear standard specifications or mature engineering practices to support it.

[0003] The creation of traditional supervisory and inspection forms relies entirely on manual operation by professionals. Experienced individuals must manually identify inspection items, compile inspection content, and design form structures, drawing upon current standards, company policies, and past project experience. This model has inherent and insurmountable flaws: firstly, form creation is extremely inefficient, consuming significant manpower and time, and is highly susceptible to human error, resulting in omissions, inconsistent wording, and incorrect standard references; secondly, the quality of form creation depends entirely on the individual experience and professional skills of the creator, failing to achieve standardized compliance control. It commonly suffers from unclear basis for inspection items and untraceable sources, making it difficult to meet the stringent requirements of high-compliance fields for form standardization and auditability.

[0004] To address the pain points of traditional manual form generation, those skilled in the art have successively developed various intelligent form processing technologies. Currently, these technologies mainly fall into three categories: template-driven automated form generation technology, large language model-based form generation technology, and intelligent governance technology for form data, as detailed below: 1. To address the need for digital form processing in the engineering field, existing technology discloses an intelligent form generation system and method for construction project management (Publication No.: CN108255799A). This solution constructs an integrated form generation architecture including a template editor, a template interpreter, a business logic generator, and a permission controller. Form templates are generated through a visual editing tool, and the template interpreter parses the templates into Web programming language-related data and database structure-related data. After processing by the business logic generator and combined with the field-level operation permissions set by the permission controller, the form is finally displayed and processed in web page format. This solution automates the generation of engineering forms from templates to usable interfaces, improving the efficiency of digital processing of traditional engineering forms to a certain extent. Its field-level permission management and approval process adaptation design also have strong practicality in engineering scenarios.

[0005] However, this solution is essentially a template-driven form rendering solution. Its form generation capabilities are entirely limited to manually pre-built form templates. It lacks the ability to understand the semantics of natural language requirements and generate content, and cannot automatically generate new inspection items and form structures based on the project feature descriptions input by the user. At the same time, the solution does not set up an external knowledge support system and compliance verification mechanism. When industry standards are updated, compliance requirements are adjusted, or new types of engineering projects emerge, the template library must be manually edited and updated. The system's scenario adaptability and flexibility are extremely poor, and it cannot meet the dynamic and diverse engineering management compliance needs.

[0006] 2. With the rapid development of large language model technology, existing technologies have emerged that apply generative AI to form generation, specifically a method and apparatus for application generation based on a large language model (Publication No.: CN118092908A). This solution predefines form generation specifications and component generation specifications, uses a fine-tuning dataset conforming to the specifications to supervise and fine-tune the large language model, obtaining a structural design model with structural design capabilities and a form design model with form design capabilities. Through these models, semantic analysis is performed on the user-input system functional requirement description, converting the requirement description into a data structure conforming to the generation specifications. Finally, a low-code generator converts the data structure into usable forms and functional components, achieving automated application generation. This solution breaks through the strong dependence of traditional template-driven solutions on manually pre-made templates, realizing end-to-end generation from natural language requirements to form interfaces, and possesses a certain level of intelligence.

[0007] However, the generation logic of this solution only stays at the shallow mapping level of "functional requirements - interface components". It does not carry out in-depth domain adaptation and modeling for the "inspection items - inspection content" hierarchical structure unique to the field of engineering supervision and inspection, nor does it design a structured reuse mechanism for historical engineering practice data. More importantly, the solution does not set up any compliance control and traceability mechanism. The generated form content does not establish an explicit relationship with the current standards, specifications and compliance requirements, which cannot guarantee the compliance of the generated content and does not have audit traceability. It completely fails to meet the core requirements of traceable form basis and verifiable content in high compliance fields such as engineering quality management.

[0008] 3. Regarding the data consistency management requirements during form filling, existing technology also discloses an intelligent form processing method, device, medium, and equipment (Publication No.: CN119358522A). This solution pre-configures the form's table style and cells, concatenating IDs from different dimensions to generate unique data IDs for each cell. During the form filling stage, a form is generated based on the pre-configured template, and data IDs are assigned to corresponding cells. After submission, conflicting data across forms and periods is identified using the data IDs. Simultaneously, form data can be quickly extracted and retrieved based on the data IDs. This solution effectively solves the data consistency problem in multi-stage, multi-source form filling, improving the efficiency and reliability of form data governance, and is suitable for multi-period, multi-entity form data comparison and integration scenarios.

[0009] However, the technical focus of this solution is entirely concentrated on the data governance of existing forms, without involving intelligent form generation technologies. It lacks natural language semantic understanding, knowledge retrieval, and content generation capabilities, and cannot automatically construct inspection items and form structures based on user-input engineering characteristics. It also cannot extract structured generation templates from standards, specifications, and historical forms. It has a fundamental functional deficiency in the intelligent form generation dimension and cannot solve the core pain points of efficiency and compliance in the form preparation process.

[0010] In summary, existing form-related technical solutions have significant shortcomings: template-driven form generation solutions cannot function without manually pre-made templates, resulting in low levels of intelligence, poor scenario adaptability, and an inability to cope with dynamically changing compliance requirements; general generation solutions based on large language models lack engineering-specific structural modeling and end-to-end compliance control, leaving generated content without verifiable data and failing to meet the audit traceability requirements of high-compliance scenarios; and form data governance solutions completely neglect the intelligent form generation stage, failing to address the core pain points of form creation. Therefore, developing a technical solution that can achieve efficient and intelligent form generation while simultaneously completing end-to-end compliance verification, ensuring traceable content sources, and guaranteeing compliance auditability has become an urgent technical challenge in this field. Summary of the Invention

[0011] This invention provides an AI-driven intelligent form generation and compliance verification method, which solves the core technical problem that existing intelligent form generation technologies cannot simultaneously address the intelligent generation and content compliance traceability issues in high-compliance scenarios.

[0012] To solve the above-mentioned technical problems, the technical solution adopted by this invention is: an AI-driven intelligent form generation and compliance verification method, comprising the following steps: S1. Construct a dual-channel layered knowledge base for the engineering management field. The dual-channel layered knowledge base includes a logically independent standard specification layer and a historical form layer. The standard specification layer systematically stores standard specification technical clauses. Each clause is associated with a number, clause content and its subject category, and encoded as a vector representation. The historical form layer archives structured data of historical supervision and inspection forms, clearly recording the hierarchical relationship between inspection items and their subordinate inspection contents, and adding project source and time information. S2. Perform dual-channel step-by-step retrieval and structural clustering operations. Receive the natural language description of the target project. First, extract the set of standard clauses related to the input description in the standard specification layer through semantic matching. Then, based on the matched standard clauses, reversely search for the corresponding inspection items-inspection content structure templates in the historical form layer. After vectorizing and clustering inspection items with similar semantics, merge and remove duplicates to generate an optimized template set. S3. Based on the optimized template set, a large language model is driven to generate a supervision and inspection form. Multi-level compliance verification is performed simultaneously. Each generated inspection item is marked with a traceability attribute. After the verification is passed, a compliance form is output, realizing a closed-loop control of compliance upon generation.

[0013] Preferably, in step S1, the content included in the standard specification layer includes technical clauses in national standards, industry regulations, and enterprise management systems. A pre-trained semantic coding model is used to convert the text of each standard clause into a dense vector representation with fixed dimensions.

[0014] Preferably, in step S1, the structured data archived in the historical form layer comes from the supervision and inspection forms of completed engineering projects. The project type and completion time information corresponding to the inspection project are extracted and stored simultaneously. A pre-trained semantic coding model with the same source as the standard clauses is used to convert the text of each inspection content into a dense vector representation of the same dimension.

[0015] Preferably, in step S2, the semantic matching of the standard specification layer specifically involves: converting the natural language description of the target project into a query vector through a pre-trained semantic coding model, calculating the cosine similarity between the query vector and the standard clause vectors in the standard specification layer, and selecting the top K standard clauses with the highest similarity to form a set of standard clauses related to the input description, where K is a preset positive integer.

[0016] Preferably, in step S2, the specific operation of reverse search for the structure template is as follows: for each standard clause in the standard clause set, calculate the cosine similarity between the vector of the standard clause and the vector of each inspection content in the historical form layer, filter out the inspection content with a similarity greater than the preset similarity threshold δ, restore the complete "inspection item-inspection content" hierarchical structure to which the inspection content belongs, and form a structure template set corresponding to the standard clause.

[0017] Preferably, in step S2, the average vector of all inspection contents under a single inspection item is used as the semantic representation vector of that inspection item, and the calculation formula is as follows: ; in, This represents the semantic vector of the k-th inspection item; This indicates the set of inspection items included in the inspection item; Indicates the number of items in the set; The vector representation of the l-th inspection item.

[0018] Preferably, in step S2, clustering and merging specifically involves: based on the semantic representation vectors of all inspection items, using an unsupervised clustering algorithm to cluster inspection items with similar semantics, merging and deduplicating inspection items within the same cluster, retaining all complete inspection content within the cluster, and finally generating an optimized template set without redundancy and with a complete structure.

[0019] Preferably, in step S3, the form generation adopts a structured prompt control mechanism, using the optimized template set as the context prompt input large language model, and generating the form according to the hierarchical structure of "inspection item - inspection content" through prompt word constraint model, and forcing the model to mark the source attribute for each generated inspection content: content originating from standard clauses is marked with the corresponding standard number and clause number, and content referencing historical practices is marked with the corresponding historical project name.

[0020] Preferably, in step S3, the multi-level compliance verification includes three levels: the first level is traceability label verification, which verifies whether the traceability labels of each inspection item are complete and valid; the second level is semantic consistency verification, which compares the inspection items on which the labeling standards are based with the corresponding original standard clauses semantically to verify the consistency of their requirement strength and coverage; the third level is structural integrity verification, which verifies whether the hierarchical structure of the form's "inspection items - inspection content" is complete and without logical omissions.

[0021] Preferably, in step S3, for forms that pass all levels of compliance verification, the system automatically extracts the source evidence for all inspection contents, generates a compliance explanation report with a complete list of references, and finally outputs a supervision and inspection form with a synchronized compliance explanation report.

[0022] This invention provides an AI-driven method for intelligent form generation and compliance verification, which has the following beneficial effects: 1. This invention constructs a dual-channel layered knowledge base consisting of a standard specification layer and a historical form layer, which realizes the separation and collaborative utilization of authoritative compliance basis and engineering practice experience. It not only solves the problem of lack of authoritative basis support in traditional template-driven solutions, but also avoids the defect of general large model generating content that is divorced from engineering practice, thus ensuring the compliance and operability of the generated content from the knowledge base.

[0023] 2. This invention designs a dual-channel step-by-step retrieval mechanism. First, it matches relevant standard clauses based on the target project description, and then it searches in reverse to find the practical application structure of the standard clauses in historical projects. This breaks the shallow mapping logic of "requirement-component" in the existing technology and realizes a deep association between compliance basis, practical structure and generated content. This ensures from the source that the generated form content has clear standard support and meets the actual operation requirements of the project.

[0024] 3. This invention proposes a semantic representation method based on the average vector of the content under the inspection item. By using this semantic vector, similar inspection items are clustered and merged, which effectively solves the problems of content redundancy and structural fragmentation caused by differences in expression in multi-source historical templates. It avoids the limitations of traditional keyword matching and significantly improves the integrity, consistency and generation efficiency of the form structure.

[0025] 4. This invention constructs a closed-loop mechanism for form generation and multi-level compliance verification. By forcing the generation of content to be labeled with traceability attributes through structured prompts, it simultaneously completes full-dimensional verification of traceability tags, semantic consistency, and structural integrity. Finally, it generates a compliance statement report with an attached list of references, truly achieving "generation equals compliance". It completely solves the core pain point of existing technologies that generate content without evidence and cannot be audited or traced, and fully meets the rigid requirements of high compliance fields for forms.

[0026] 5. The overall solution of this invention does not require manual pre-made form templates. It can complete the end-to-end intelligent generation of forms based on natural language descriptions, which greatly reduces the dependence of form preparation on human professional experience. It shortens the form preparation work that originally required several hours to several days to minutes, while significantly improving the comprehensiveness and standardization of form content. It has strong engineering application value and promotion prospects. Attached Figure Description

[0027] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0028] like Figure 1 As shown, the AI-driven intelligent form generation and compliance verification method includes the following steps: S1. Construct a dual-channel layered knowledge base for the engineering management field. The dual-channel layered knowledge base includes a logically independent standard specification layer and a historical form layer. The standard specification layer systematically stores standard specification technical clauses. Each clause is associated with a number, clause content and its subject category, and encoded as a vector representation. The historical form layer archives structured data of historical supervision and inspection forms, clearly recording the hierarchical relationship between inspection items and their subordinate inspection contents, and adding project source and time information. S2. Perform dual-channel step-by-step retrieval and structural clustering operations. Receive the natural language description of the target project. First, extract the set of standard clauses related to the input description in the standard specification layer through semantic matching. Then, based on the matched standard clauses, reversely search for the corresponding inspection items-inspection content structure templates in the historical form layer. After vectorizing and clustering inspection items with similar semantics, merge and remove duplicates to generate an optimized template set. S3. Based on the optimized template set, a large language model is driven to generate a supervision and inspection form. Multi-level compliance verification is performed simultaneously. Each generated inspection item is marked with a traceability attribute. After the verification is passed, a compliance form is output, realizing a closed-loop control of compliance upon generation.

[0029] Preferably, in step S1, the content included in the standard specification layer includes technical clauses in national standards, industry regulations, and enterprise management systems. A pre-trained semantic coding model is used to convert the text of each standard clause into a dense vector representation with fixed dimensions.

[0030] Preferably, in step S1, the structured data archived in the historical form layer comes from the supervision and inspection forms of completed engineering projects. The project type and completion time information corresponding to the inspection project are extracted and stored simultaneously. A pre-trained semantic coding model with the same source as the standard clauses is used to convert the text of each inspection content into a dense vector representation of the same dimension.

[0031] Preferably, in step S2, the semantic matching of the standard specification layer specifically involves: converting the natural language description of the target project into a query vector through a pre-trained semantic coding model, calculating the cosine similarity between the query vector and the standard clause vectors in the standard specification layer, and selecting the top K standard clauses with the highest similarity to form a set of standard clauses related to the input description, where K is a preset positive integer.

[0032] Upon receiving the natural language description of the target project, the system performs a dual-channel step-by-step retrieval and structural clustering operation to generate a high-quality, structured context template. First, semantic matching is performed at the standard specification layer to extract the set of standard clauses most relevant to the input description: ; in, This represents the set of the top K standard clauses that are most relevant to the input description q; This represents the set of all standard clauses in the standard specification layer; This represents the vector representation of the input description q after it has been encoded by the semantic model. The text vector representing the i-th standard clause; This represents the cosine similarity function; TopK represents selecting the K results with the highest similarity.

[0033] Preferably, in step S2, the specific operation of reverse search for the structure template is as follows: for each standard clause in the standard clause set, calculate the cosine similarity between the vector of the standard clause and the vector of each inspection content in the historical form layer, filter out the inspection content with a similarity greater than the preset similarity threshold δ, restore the complete "inspection item-inspection content" hierarchical structure to which the inspection content belongs, and form a structure template set corresponding to the standard clause.

[0034] Based on the matching standard clauses, the expression of the clauses in actual application is searched back in the historical form layer to identify the corresponding "inspection item - inspection content" structure template: ; in, Representation and Standard Clauses A collection of relevant historical inspection structure templates; This represents the k-th inspection item in the j-th historical form; This represents the set of inspection items under this inspection item; This indicates the specific inspection content of item l; A vector representation of the content being inspected; This is a preset semantic similarity threshold used to determine whether content is relevant.

[0035] Preferably, in step S2, the average vector of all inspection contents under a single inspection item is used as the semantic representation vector of that inspection item, and the calculation formula is as follows: ; in, This represents the semantic vector of the k-th inspection item; This indicates the set of inspection items included in the inspection item; Indicates the number of items in the set; The vector representation of the l-th inspection item.

[0036] Preferably, in step S2, clustering and merging specifically involves: based on the semantic representation vectors of all inspection items, using an unsupervised clustering algorithm to cluster inspection items with similar semantics, merging and deduplicating inspection items within the same cluster, retaining all complete inspection content within the cluster, and finally generating an optimized template set without redundancy and with a complete structure.

[0037] The final optimized template set is deduplicated, complete, and clearly structured. ; in, This represents the optimized set of structured context templates; ClusterMerge represents a semantic vector-based clustering and merging algorithm used to merge similar inspection items, eliminate redundancy, and retain complete inspection content. This step realizes the transformation from discrete text to structured compliance templates, providing high-fidelity, inheritable context support for large model generation.

[0038] Preferably, in step S3, the form generation adopts a structured prompt control mechanism, using the optimized template set as the context prompt input large language model, and generating the form according to the hierarchical structure of "inspection item - inspection content" through prompt word constraint model, and forcing the model to mark the source attribute for each generated inspection content: content originating from standard clauses is marked with the corresponding standard number and clause number, and content referencing historical practices is marked with the corresponding historical project name.

[0039] Preferably, in step S3, the multi-level compliance verification includes three levels: the first level is traceability label verification, which verifies whether the traceability labels of each inspection item are complete and valid; the second level is semantic consistency verification, which compares the inspection items on which the labeling standards are based with the corresponding original standard clauses semantically to verify the consistency of their requirement strength and coverage; the third level is structural integrity verification, which verifies whether the hierarchical structure of the form's "inspection items - inspection content" is complete and without logical omissions.

[0040] Preferably, in step S3, for forms that pass all levels of compliance verification, the system automatically extracts the source evidence for all inspection contents, generates a compliance explanation report with a complete list of references, and finally outputs a supervision and inspection form with a synchronized compliance explanation report.

[0041] Based on the optimized structured template set This system drives a large language model to generate monitoring and inspection forms and simultaneously executes multi-level compliance checks, achieving a closed-loop control of "compliance upon generation." The generation process is controlled by a structured prompting mechanism, requiring the model to generate specific and actionable inspection content for each inspection item and mandating that each piece of content be labeled with its source attribute: content originating from standard clauses is labeled "Basis: Standard Number {Clause Number}", and content referencing historical practices is labeled "Reference: ${Project Name}". Based on this, a compliance check module is activated to compare the generated content with the original clauses in the standard specification layer for semantic consistency, verifying the strength and scope of the requirements; simultaneously, it checks whether the form structure is complete and whether the traceability tags are complete. For forms that pass the check, the system automatically generates a compliance explanation report with an attached list of references. This mechanism ensures that the final output form is not only structurally sound and complete in content, but also that every item is traceable and verifiable.

[0042] In practice, the first channel is the "standards and specifications layer," which includes authoritative texts such as national standards and industry standards. Each standard clause is encoded into a text vector using a semantic model. The second channel is the "historical form layer," which aggregates supervision and inspection forms from multiple completed projects, extracts the "inspection item - inspection content" structure, and encodes each inspection content into a semantic vector. This enables the structured storage of practical experience.

[0043] After receiving the engineering description input by the user, the system first performs semantic matching at the standard specification layer and calculates the input vector. With each standard clause The similarity is used to select the top-K relevant criteria to form a matching set. Subsequently, for each matching criterion... Search for inspection content with similar semantics in the historical form layer (i.e. ), and restore its complete inspection structure to form a set. Next, for each inspection item Calculate its semantic vector Based on this vector, semantic clustering is performed on all candidate items, merging items with different expressions but similar semantics, and finally generating a deduplicated and complete optimized template set. .

[0044] Will As contextual prompts, the large language model is guided to generate a well-structured and comprehensive initial draft of a supervisory checklist, while simultaneously performing compliance checks: verifying the traceability of each generated item. The standard basis or The historical templates in the database are used to generate compliance reports, ensuring that the output is both professional and reliable, and has audit traceability.

[0045] This invention provides an AI-driven intelligent form generation and compliance verification method, addressing the industry pain points of form compilation in high-compliance fields such as engineering construction. It overcomes the technical limitations of traditional template-driven form generation and unconstrained generation of general large models, constructing a dual-channel layered knowledge base consisting of a standard specification layer and a historical form layer. This achieves the separate modeling and collaborative utilization of authoritative compliance evidence and engineering practice experience. It pioneers a dual-channel step-by-step retrieval mechanism that first matches relevant standard clauses based on engineering natural language descriptions, and then retrieves corresponding historical practice structure templates, thus establishing a deep connection between compliance evidence and engineering practice. It proposes a semantic representation method based on the average value of the content vectors under the inspection project, achieving deduplication optimization and structured integration of multi-source templates through clustering and merging. Finally, it constructs a closed loop for form generation and multi-level compliance verification with mandatory traceability constraints, achieving full-link traceability and verifiability of form content. This truly achieves end-to-end intelligent form generation where "generation equals compliance" in high-compliance scenarios, realizing a systematic innovation across the entire chain from knowledge base, retrieval logic, structural optimization to compliance control.

[0046] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. An AI-driven method for intelligent form generation and compliance verification, characterized in that, Includes the following steps: S1. Construct a dual-channel layered knowledge base for the engineering management field. The dual-channel layered knowledge base includes a logically independent standard specification layer and a historical form layer. The standard specification layer systematically stores standard specification technical clauses. Each clause is associated with a number, clause content and its subject category, and encoded as a vector representation. The historical form layer archives structured data of historical supervision and inspection forms, clearly recording the hierarchical relationship between inspection items and their subordinate inspection contents, and adding project source and time information. S2. Perform dual-channel step-by-step retrieval and structural clustering operations. Receive the natural language description of the target project. First, extract the set of standard clauses related to the input description in the standard specification layer through semantic matching. Then, based on the matched standard clauses, reversely search for the corresponding inspection items-inspection content structure templates in the historical form layer. After vectorizing and clustering inspection items with similar semantics, merge and remove duplicates to generate an optimized template set. S3. Based on the optimized template set, a large language model is driven to generate a supervision and inspection form. Multi-level compliance verification is performed simultaneously. Each generated inspection item is marked with a traceability attribute. After the verification is passed, a compliance form is output, realizing a closed-loop control of compliance upon generation.

2. The AI-driven intelligent form generation and compliance verification method according to claim 1, characterized in that, In step S1, the content included in the standard specification layer includes technical clauses in national standards, industry regulations, and enterprise management systems. A pre-trained semantic coding model is used to convert the text of each standard clause into a dense vector representation with fixed dimensions.

3. The AI-driven intelligent form generation and compliance verification method according to claim 1, characterized in that, In step S1, the structured data archived in the historical form layer comes from the supervision and inspection forms of completed engineering projects. The project type and completion time information corresponding to the inspection project are extracted and stored synchronously. A pre-trained semantic coding model with the same source as the standard clauses is used to convert the text of each inspection content into a dense vector representation of the same dimension.

4. The AI-driven intelligent form generation and compliance verification method according to claim 1, characterized in that, In step S2, the semantic matching of the standard specification layer specifically involves: converting the natural language description of the target project into a query vector through a pre-trained semantic coding model; calculating the cosine similarity between the query vector and the standard clause vectors in the standard specification layer; selecting the top K standard clauses with the highest similarity to form a set of standard clauses related to the input description, where K is a preset positive integer.

5. The AI-driven intelligent form generation and compliance verification method according to claim 4, characterized in that, In step S2, the specific operation of reverse search for the structure template is as follows: for each standard clause in the standard clause set, calculate the cosine similarity between the vector of the standard clause and the vector of each inspection content in the historical form layer, filter out the inspection content with a similarity greater than the preset similarity threshold δ, restore the complete "inspection item-inspection content" hierarchical structure to which the inspection content belongs, and form a structure template set corresponding to the standard clause.

6. The AI-driven intelligent form generation and compliance verification method according to claim 1, characterized in that, In step S2, the average vector of all inspection contents under a single inspection item is used as the semantic representation vector of that inspection item. The calculation formula is as follows: ; in, This represents the semantic vector of the k-th inspection item; This indicates the set of inspection items included in the inspection item; Indicates the number of items in the set; The vector representation of the l-th inspection item.

7. The AI-driven intelligent form generation and compliance verification method according to claim 6, characterized in that, In step S2, clustering and merging specifically involves: based on the semantic representation vectors of all inspection items, using an unsupervised clustering algorithm to cluster inspection items with similar semantics, merging and deduplicating inspection items within the same cluster, retaining all complete inspection content within the cluster, and finally generating an optimized template set without redundancy and with a complete structure.

8. The AI-driven intelligent form generation and compliance verification method according to claim 1, characterized in that, In step S3, the form generation adopts a structured prompt control mechanism, using the optimized template set as the context prompt input large language model. The prompt word constrains the model to generate the form according to the hierarchical structure of "inspection item - inspection content", and forces the model to mark the source attribute for each generated inspection content: content originating from standard clauses is marked with the corresponding standard number and clause number, and content referencing historical practices is marked with the corresponding historical project name.

9. The AI-driven intelligent form generation and compliance verification method according to claim 8, characterized in that, In step S3, the multi-level compliance verification includes three levels: the first level is traceability label verification, which verifies whether the traceability labels of each checked item are complete and valid; The second level is semantic consistency verification, which compares the inspection content based on the standard with the corresponding original standard clauses to verify the consistency of the requirement strength and coverage. The third level is structural integrity verification, which verifies whether the "inspection item - inspection content" hierarchical structure of the form is complete and without logical omissions.

10. The AI-driven intelligent form generation and compliance verification method according to claim 9, characterized in that, In step S3, for forms that pass all levels of compliance verification, the system automatically extracts the source evidence for all inspection contents, generates a compliance explanation report with a complete list of references, and finally outputs a supervision and inspection form with a synchronized compliance explanation report.

Images