A multi-stage intelligent tender automatic generation system and method

Abstract

This invention provides a multi-stage intelligent automated tender document generation system and method, belonging to the field of automated tender document generation technology. The method includes: collecting and preprocessing bidding information and auxiliary information to generate structured bidding information; constructing a multi-level tender document directory based on the structured bidding information, generating a qualified multi-level tender document directory after verification by matching degree evaluation rules, and constructing an overall tender document framework; performing format recognition and differential analysis on objective content, generating objective fill-in content and associating it with the overall tender document framework; using the final-level directory as a generation unit, generating subjective content through enhanced retrieval generation technology and constraint rules, generating subjective fill-in content through quality evaluation and iterative correction, and associating it with the overall tender document framework; and performing format processing on the overall tender document framework to output an intelligent tender document file. This achieves automation and intelligence throughout the entire tender document generation process, balancing generation efficiency and content quality, effectively avoiding the risk of bid rejection, and improving the success rate of bidding.

Description

Technical Field

[0001] This invention relates to the field of automated tender document generation technology, and in particular to a multi-stage intelligent automated tender document generation system and method. Background Technology

[0002] As the vehicle through which bidders respond to tender requirements, the quality and efficiency of their preparation directly impact the bidding results. With increasingly frequent bidding activities and rising project complexity, traditional manual bid preparation methods can no longer meet companies' demands for efficiency, standardization, and accuracy. Automated bid preparation tools have become an inevitable trend in the industry.

[0003] In existing technologies, automated tender document generation solutions mainly fall into two categories. The first category is based on template engines and rule bases, which breaks down the tender document into fixed modules and relies on manually preset mapping rules for content assembly. This solution has significant shortcomings: the maintenance cost of templates and rules is extremely high; new industry or non-standard projects require redevelopment; and it lacks versatility. It can only achieve shallow personalization at the parameter replacement level, unable to handle complex technical indicators and customized needs; compliance verification is limited to keyword matching, failing to identify logical conflicts and regulatory detail deviations; and module assembly easily leads to content redundancy or logical gaps. The second category is based on large models combined with retrieval generation technology. While this eliminates the reliance on fixed templates, the controllability and standardization of the generated content are severely lacking. Large models are prone to "illusions," leading to the generation of fictitious clauses or neglect of legal format requirements, and the expression is overly colloquial. Simultaneously, large models provide vague descriptions of professional technical indicators, and the generated scoring criteria lack clear scoring basis, resulting in poor applicability in bid evaluation and requiring professionals to spend considerable time revising and refining them.

[0004] Therefore, it is necessary to provide a multi-stage intelligent automated tender document generation system and method to solve the above-mentioned technical problems. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a multi-stage intelligent automated tender document generation system and method, which solves the problems of low efficiency and poor format compliance in existing tender document generation schemes, as well as the uncontrollable content and difficulty in guaranteeing quality of large model generation.

[0006] This invention provides a multi-stage intelligent automated tender document generation system, the system comprising: The data acquisition module is used to collect bidding requirement information and auxiliary reference information, preprocess the bidding requirement information and auxiliary reference information, and generate structured bidding information. The directory framework construction module is used to construct a multi-level bid document directory based on the structured bidding information, verify the multi-level bid document directory against the structured bidding information according to the matching degree evaluation rules, generate a qualified multi-level bid document directory, and construct the overall bid document framework based on the qualified multi-level bid document directory. The objective content generation module is used to perform format recognition and differential parsing on the pre-stored objective content of the tender document, generate objective fill content for the tender document, and associate the objective fill content of the tender document with the overall framework of the tender document; The subjective content generation module is used to take the last-level directory in the qualified multi-level tender document directory as the generation unit, generate the subjective content of the tender document based on the structured bidding information, and generate the subjective content of the tender document through enhanced retrieval generation technology and preset constraint rules. The module performs quality assessment and iterative correction on the subjective content of the tender document, generates subjective filling content of the tender document, and associates and fills the subjective filling content of the tender document into the overall framework of the tender document. The tender document output module is used to format and standardize the overall framework of the completed tender document, and output an intelligent tender document file.

[0007] Preferably, the directory framework construction module is used to construct a multi-level bid document directory based on the structured bidding information, verify the multi-level bid document directory against the structured bidding information according to the matching degree evaluation rules, generate a qualified multi-level bid document directory, and construct an overall bid document framework based on the qualified multi-level bid document directory, specifically including: The first-level directory generation unit is used to extract the structured bidding information, generate bidding keywords, match the bidding keywords with a preset first-level directory template library, and generate the first-level directory of the tender document. The secondary directory generation unit is used to perform semantic parsing on the review clauses and scoring items in the structured bidding information, generate review points, and generate a secondary directory of the tender document belonging to the primary directory of the tender document based on the correspondence between the review points and the primary directory of the tender document. The directory matching degree evaluation unit is used to integrate the first-level directory of the tender documents and the second-level directory of the tender documents to construct the multi-level tender document directory, and to use the directory matching degree algorithm to quantitatively verify the multi-level tender document directory and the structured bidding information to generate a directory matching degree score. The qualification determination unit is used to determine the multi-level tender document directory as the qualified multi-level tender document directory when the directory matching score is greater than or equal to the preset qualification threshold; when the directory matching score is less than the preset qualification threshold, it receives the user's directory adjustment instruction to adjust the multi-level tender document directory, and redetermines the directory matching score based on the adjusted multi-level tender document directory until the qualified multi-level tender document directory is generated. The framework construction unit is used to construct the overall framework of the tender document based on the hierarchical relationship of the qualified multi-level tender document directory, and to associate the attachment content in the structured bidding information with the corresponding attachment position of the overall framework of the tender document.

[0008] Preferably, a directory matching algorithm is used to quantitatively verify the first-level directory and the second-level directory of the tender document against the structured bidding information, generating a first matching degree and a second matching degree. The first matching degree and the second matching degree are then weighted and summed to generate a directory matching degree score. The corresponding calculation formula is as follows: In the formula, This indicates the total number of bidding keywords extracted from structured bidding information; This indicates the total number of review points extracted from the structured bidding information; This indicates the number of successfully matched bidding keywords in the first-level directory of the tender document; This indicates the number of review points successfully matched in the second-level directory of the tender document; The weight coefficient representing the first degree of matching. The weight coefficients representing the second degree of matching satisfy... and .

[0009] Preferably, the objective content generation module is used to perform format recognition and differential parsing on the pre-stored objective content of the tender document, generate objective fill-in content for the tender document, and fill the objective fill-in content into the overall framework of the tender document, specifically including: An objective content recognition unit is used to call the pre-stored objective content of the tender document, perform format recognition on the objective content of the tender document, and determine the file type of the objective content of the tender document, wherein the file type includes scanned files and electronic files; An objective content parsing unit is used to extract the text information of the objective content of the tender document using optical character recognition technology when the document type is a scanned document, and to perform structured parsing of the text information based on preset field mapping rules to generate the objective content to fill in the tender document; when the document type is an electronic document, it directly extracts the structured information of the electronic document as the objective content to fill in the tender document. An objective content mapping unit is used to perform semantic matching between the last-level directory and the objective content of the tender document based on the directory characteristics of the last-level directory in the qualified multi-level tender document directory, establish a mapping relationship between the last-level directory and the objective content of the tender document, and generate an objective content filling mapping table. An objective content filling unit is used to associate and fill the objective content of the tender document with the overall framework of the tender document according to the objective content filling mapping table, and to perform format adaptation and content quality verification on the filled objective content of the tender document.

[0010] Preferably, the subjective content generation module is used to take the lowest-level directory in the qualified multi-level tender document directory as the generation unit, generate subjective content for the tender document based on the structured bidding information through enhanced retrieval generation technology and preset constraint rules, perform quality assessment and iterative correction on the subjective content, generate subjective content to fill in the tender document, and associate the subjective content to fill in the overall framework of the tender document, specifically including: The subjective content retrieval unit is used to use the last-level directory in the qualified multi-level tender document directory as the retrieval anchor point, retrieve multi-source tender information associated with the last-level directory from the structured tender information and the preset external knowledge base, and perform vectorization encoding on the multi-source tender information to generate retrieval enhancement information vectors. The subjective content constraint unit is used to generate initial subjective content corresponding to the last-level directory based on the retrieval enhancement information vector and the preset constraint rules, through the enhanced retrieval generation technology, and to perform constraint compliance detection and iterative correction on the initial subjective content, and output the tender subjective content. The subjective content evaluation and correction unit is used to quantitatively evaluate the subjective content of the tender document through a preset quality evaluation model, generate a subjective content quality score, perform hierarchical iterative correction on the subjective content of the tender document based on the subjective content quality score, and generate the subjective content to be filled in the tender document. The subjective content filling unit is used to associate and fill the subjective content of the tender document into the overall framework of the tender document according to the correspondence between the last-level directory and the subjective content of the tender document, and to perform format adaptation and content quality verification on the filled subjective content of the tender document.

[0011] Preferably, the subjective content constraint unit is used to generate initial subjective content corresponding to the last-level directory based on the retrieval enhancement information vector and the preset constraint rules, using the enhanced retrieval generation technology, and to perform constraint compliance detection and iterative correction on the initial subjective content, and output the tender subjective content, specifically including: The constraint prompt construction unit is used to fuse the retrieval enhancement information vector with the semantic vector of the last-level directory to construct a prompt context, construct constraint control instructions according to the length constraint rules, logical structure constraint rules and key point coverage constraint rules in the preset constraint rules, embed the constraint control instructions into the prompt context, and generate constrained prompts. The constraint-guided generation unit is used to input the constraint-based generation prompts into a preset enhanced retrieval generation model and generate initial subjective content corresponding to the last-level directory. The constraint compliance detection unit is used to perform constraint compliance detection on the initial subjective content, and calculate the first compliance degree of the initial subjective content with the length constraint rule, the second compliance degree of the initial subjective content with the logical structure constraint rule, and the third compliance degree of the initial subjective content with the key point coverage constraint rule. The constraint correction iteration unit is used to determine the constraint rules corresponding to compliance degrees less than the preset sub-item threshold as constraints to be corrected, and to correct the initial subjective content based on the constraints to be corrected until all compliance degrees are greater than or equal to the preset sub-item threshold. The corrected initial subjective content is then used as the tender subjective content and output.

[0012] Preferably, the subjective content of the tender document is quantitatively evaluated using a preset quality assessment model to generate a subjective content quality score. The corresponding calculation formula is as follows: In the formula, T represents the semantic vector representation of the final-level directory; C represents the semantic vector representation of the subjective content of the tender document; and R represents the set of key tender information points retrieved from the structured tender information that corresponds to the final-level directory. Indicates the semantic similarity between the final-level directory and the subjective content of the tender document; This indicates the coverage of key bidding information points within the main content of the bid document; This indicates the content quality score of the subjective content of the tender document; These represent the weighting coefficients for semantic similarity, coverage, and content quality score, respectively, satisfying... and .

[0013] Preferably, the step of performing graded iterative correction on the subjective content of the tender document based on the subjective content quality score to generate the subjective content filler for the tender document specifically includes: The subjective content quality score is graded and determined by setting a first score interval, a second score interval, and a third score interval, wherein the scores of the first score interval, the second score interval, and the third score interval decrease sequentially. When the subjective content quality score is within the first score range, the subjective content of the tender document will be directly used as the subjective content to fill in the tender document. When the subjective content quality score is in the second scoring range, the semantic similarity, the coverage, and the content quality score are compared with the corresponding preset dimension thresholds. Defect dimensions that are less than the preset dimension thresholds are identified. Based on the defect dimensions, the corresponding preset correction template is called to perform targeted completion or logical reorganization of the subjective content of the tender document, and the subjective content quality score is recalculated until the subjective content quality score is in the first scoring range. When the subjective content quality score is within the third score range, it is determined that the subjective content of the tender document has structural defects. The retrieval range of the enhanced retrieval information vector and the constraint parameters of the preset constraint rules are adjusted, triggering the enhanced retrieval generation technology to regenerate the subjective content of the tender document. The regenerated subjective content of the tender document is then subjected to quality evaluation and iterative correction until the subjective content of the tender document is generated.

[0014] A multi-stage intelligent method for automatically generating tender documents, the method comprising: Collect bidding requirements information and auxiliary reference information, preprocess the bidding requirements information and auxiliary reference information, and generate structured bidding information; A multi-level tender document catalog is constructed based on the structured tender information. The multi-level tender document catalog is verified against the structured tender information according to the matching degree evaluation rules to generate a qualified multi-level tender document catalog. The overall framework of the tender document is then constructed based on the qualified multi-level tender document catalog. The format of the pre-stored objective content of the tender document is identified and differentiated, generating objective fill content for the tender document, and the objective fill content is then associated with and filled into the overall framework of the tender document. Using the lowest level directory in the qualified multi-level tender document directory as the generation unit, based on the structured bidding information, the subjective content of the tender document is generated by enhancing the retrieval generation technology and preset constraint rules. The subjective content of the tender document is then evaluated for quality and iteratively corrected to generate subjective content to fill in the tender document. Finally, the subjective content to fill in the tender document is associated with and filled into the overall framework of the tender document. The overall framework of the completed tender document is formatted and standardized to output an intelligent tender document file.

[0015] Compared with related technologies, the multi-stage intelligent automated tender document generation system and method provided by this invention has the following beneficial effects: This invention includes a data acquisition module for collecting bidding requirements information and auxiliary reference information, preprocessing the bidding requirements information and auxiliary reference information to generate structured bidding information; a directory framework construction module for constructing a multi-level bid document directory based on the structured bidding information, verifying the multi-level bid document directory against the structured bidding information according to matching evaluation rules, generating a qualified multi-level bid document directory, and constructing an overall bid document framework based on the qualified multi-level bid document directory; and an objective content generation module for performing format recognition and differential parsing on pre-stored objective content of the bid document, generating objective content to fill in the bid document, and associating the objective content to fill in the overall bid document framework. The system comprises a framework and a subjective content generation module. The subjective content generation module uses the lowest-level directory in the qualified multi-level tender document directory as the generation unit. Based on structured bidding information, it generates subjective content for the tender document through enhanced retrieval generation technology and preset constraint rules. It then performs quality assessment and iterative correction on the subjective content, generates subjective fill-in content, and integrates this content into the overall tender document framework. The tender document output module performs format standardization processing on the completed overall tender document framework and outputs an intelligent tender document file. This enables automation and intelligence throughout the entire tender document generation process, balancing generation efficiency and content quality, effectively mitigating the risk of bid rejection, and improving the success rate of bids.

[0016] This invention utilizes a data acquisition module to perform structured preprocessing of bidding requirements and auxiliary reference information, effectively improving information extraction accuracy and reducing human interpretation bias. Based on structured bidding information, this invention automatically constructs a multi-level bid document catalog, combining it with matching degree evaluation rules for quantitative verification. This ensures a high degree of alignment between the catalog hierarchy and bidding requirements, significantly reducing the time spent manually building the catalog framework. It also features a human-machine collaborative adjustment mechanism for insufficient catalog matching, further guaranteeing the reliability and relevance of the catalog construction. Furthermore, this invention performs format recognition and differential parsing on pre-stored objective content of the bid documents, automatically adapting the parsing path to scanned or electronic versions based on file type. This enables accurate extraction, structured mapping, and error-free filling of fixed content such as company qualifications and performance certificates, fundamentally avoiding the risks of manual input errors and non-standard formats. This invention uses the lowest-level directory as the generation unit, integrating enhanced retrieval generation technology and preset constraint rules to guide the generation of subjective content in the tender document. Through a quality assessment model and a hierarchical iterative correction mechanism, it optimizes the semantic relevance, key point coverage, and intrinsic quality of the content in multiple dimensions, ensuring that the generated content is both professional and strictly adheres to bidding specifications. This overcomes the shortcomings of traditional large-scale models, such as uncontrollable content generation and vague professional details. The invention also completes global format standardization processing through the tender document output module, outputting intelligent tender document files that meet bidding requirements. Attached Figure Description

[0017] Figure 1 This is a system block diagram of a multi-stage intelligent automated tender document generation system according to the present invention; Figure 2 This is a flowchart of a multi-stage intelligent automated tender document generation method according to the present invention. Detailed Implementation

[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments. Example

[0019] like Figure 1 As shown, a multi-stage intelligent automated tender document generation system includes: The data acquisition module is used to collect bidding requirement information and auxiliary reference information, preprocess the bidding requirement information and auxiliary reference information, and generate structured bidding information. The bidding requirements information includes original bidding documents, clarifications, and addenda. Supplementary reference information refers to external supporting materials such as historical bidding materials and industry standards. Preprocessing includes natural language processing analysis, key element extraction, and redundant formatting cleanup. Structured bidding information refers to a standardized collection of information organized according to a pre-defined data model, where key elements of the bidding project are labeled as field tags and hierarchical relationships are established for easy machine reading.

[0020] Understandably, the data acquisition module, as the system's information entry point, is responsible for extracting key elements from original bidding documents in various formats and combining them with historical bidding materials and industry standards to form supplementary reference information. The preprocessing process transforms scattered bidding terms into structured data representations by performing semantic parsing, field annotation, and format cleaning on unstructured text, reducing human interpretation bias and improving information reuse efficiency.

[0021] The directory framework construction module is used to construct a multi-level bid document directory based on the structured bidding information, verify the multi-level bid document directory against the structured bidding information according to the matching degree evaluation rules, generate a qualified multi-level bid document directory, and construct the overall bid document framework based on the qualified multi-level bid document directory. The matching degree evaluation rule is a quantitative judgment standard that calculates and weights the matching ratio between the first-level bid document directory and the bidding keywords, and between the second-level bid document directory and the review points. This is used to measure the degree to which the directory covers the bidding requirements. A qualified multi-level bid document directory refers to a directory structure that, after verification by the matching degree evaluation rule, achieves a matching degree score that reaches a preset qualified threshold or meets the depth of the bidding response after manual adjustment. The overall bid document framework is a bid document skeleton built based on the hierarchical relationship of the qualified multi-level bid document directory, and embeds the appendix content from the structured bidding information.

[0022] Understandably, the directory framework construction module automatically extracts bidding keywords and review points from structured bidding information, generating first-level and second-level directories for the bidding document. It then uses matching evaluation rules to quantitatively verify the completeness of the directory coverage, ensuring that the directory hierarchy accurately aligns with the bidding requirements. When the verification results do not meet expectations, manual intervention is used to adjust the directory structure until a qualified multi-level bidding document directory is generated. Based on this, the module constructs the overall framework of the bidding document according to the hierarchical relationship of the qualified multi-level bidding document directory and places the bidding attachments into their corresponding positions within the overall framework.

[0023] Through the above methods, the catalog framework construction module automates the process of interpreting bidding information and building the catalog. It quantitatively verifies the completeness of the catalog coverage based on the matching degree evaluation rules, ensuring that the generated multi-level bidding catalog is highly consistent with the bidding requirements, effectively reducing the time spent on manually sorting out bidding terms and compiling the catalog level by level.

[0024] The objective content generation module is used to perform format recognition and differential parsing on the pre-stored objective content of the tender document, generate objective fill content for the tender document, and associate the objective fill content of the tender document with the overall framework of the tender document; The objective content of the tender document refers to fixed factual documents that do not require secondary creation, such as business licenses, qualification certificates, personnel certificates, and performance contracts. Format recognition is used to determine whether the pre-stored content is a scanned or electronic version. Differential parsing refers to using optical character recognition technology to extract text information or directly read structured fields based on the file type, generating objective content that can be filled in the tender document, and then using semantic matching to associate it with the corresponding directory location in the overall framework of the tender document.

[0025] Understandably, the objective content generation module handles fixed information that doesn't require secondary creation. It identifies the file format of pre-stored objective content in the tender document, distinguishing between scanned and electronic versions. It then uses optical character recognition (OCR) technology to extract and structure the text information, or directly reads the structured fields from the electronic file, generating the objective content to fill in the tender document. Subsequently, this module performs semantic matching based on the directory characteristics of the lowest-level directory in the qualified multi-level tender document directory. This accurately associates the objective content with the corresponding position in the overall tender document framework, and performs format adaptation and integrity checks on the objective content. This avoids the inefficiency and input errors caused by manually searching, copying, and pasting company qualifications and supporting documents repeatedly.

[0026] The subjective content generation module is used to take the last-level directory in the qualified multi-level tender document directory as the generation unit, generate the subjective content of the tender document based on the structured bidding information, and generate the subjective content of the tender document through enhanced retrieval generation technology and preset constraint rules. The module performs quality assessment and iterative correction on the subjective content of the tender document, generates subjective filling content of the tender document, and associates and fills the subjective filling content of the tender document into the overall framework of the tender document. The enhanced retrieval and generation technology refers to using the bottom-level directory as a retrieval anchor point to retrieve multi-source bidding information from structured bidding information and external knowledge bases, and then performing vectorized encoding. This is combined with preset constraint rules to generate the subjective content of the bid document. The preset constraint rules are used to limit the form and scope of the generated content. Quality assessment and iterative correction refer to using a preset quality assessment model to quantitatively score the subjective content of the bid document, and performing targeted completion, logical restructuring, or regeneration operations based on the scoring results until qualified subjective content for the bid document is output.

[0027] Through the above approach, the subjective content generation module uses the lowest-level directory as the smallest response unit. It leverages enhanced retrieval and generation technology to obtain multi-source bidding information from structured bidding information and external knowledge bases. Combined with preset constraint rules, this guides the generation of subjective content in the bid documents, preventing the subjective content from deviating from the actual bidding process. Simultaneously, through quality assessment and tiered iterative correction, it ensures that the subjective content of the bid documents meets the specifications in terms of semantic relevance, key point coverage, and intrinsic quality, significantly reducing the cost of manual writing and review.

[0028] The tender document output module is used to format and standardize the overall framework of the completed tender document, and output an intelligent tender document file.

[0029] Understandably, after the tender document output module completes the filling, it performs global format standardization processing on the overall framework of the tender document, including unifying the font size, line spacing, and page margins of the body text, automatically arranging consecutive page numbers according to the table of contents, marking the reserved space for signatures, and exporting the processed complete content as an intelligent tender document file that meets the requirements for tender submission. It supports multiple output formats such as portable document format and text processing format, reducing the workload of manual typesetting and format adjustment.

[0030] Through the above methods, the tender document output module performs unified format adjustments, page numbering, and signature / seal placement markings on the completed tender document framework, outputting an intelligent tender document that conforms to the tender submission specifications. This eliminates the repetitive work and formatting errors caused by manual page-by-page formatting and manual marking of signature / seal areas, ensuring the standardization and submission efficiency of the final tender document.

[0031] The directory framework construction module is used to construct a multi-level bid document directory based on the structured bidding information, verify the multi-level bid document directory against the structured bidding information according to the matching degree evaluation rules, generate a qualified multi-level bid document directory, and construct an overall bid document framework based on the qualified multi-level bid document directory, specifically including: The first-level directory generation unit is used to extract the structured bidding information, generate bidding keywords, match the bidding keywords with a preset first-level directory template library, and generate the first-level directory of the tender document. The secondary directory generation unit is used to perform semantic parsing on the review clauses and scoring items in the structured bidding information, generate review points, and generate a secondary directory of the tender document belonging to the primary directory of the tender document based on the correspondence between the review points and the primary directory of the tender document. The directory matching degree evaluation unit is used to integrate the first-level directory of the tender documents and the second-level directory of the tender documents to construct the multi-level tender document directory, and to use the directory matching degree algorithm to quantitatively verify the multi-level tender document directory and the structured bidding information to generate a directory matching degree score. The qualification determination unit is used to determine the multi-level tender document directory as the qualified multi-level tender document directory when the directory matching score is greater than or equal to the preset qualification threshold; when the directory matching score is less than the preset qualification threshold, it receives the user's directory adjustment instruction to adjust the multi-level tender document directory, and redetermines the directory matching score based on the adjusted multi-level tender document directory until the qualified multi-level tender document directory is generated. The framework construction unit is used to construct the overall framework of the tender document based on the hierarchical relationship of the qualified multi-level tender document directory, and to associate the attachment content in the structured bidding information with the corresponding attachment position of the overall framework of the tender document.

[0032] The first-level directory template library is a pre-built set of bid document chapter titles based on different bidding types and industry standards, used to map extracted bidding keywords into standardized first-level directory entries. Review clauses refer to the mandatory provisions in the bidding documents regarding bidder qualifications, compliance, and conditions for rejection. Scoring items are specific evaluation indicators for the technical review, commercial review, and price review stages. Review points are response elements extracted from review clauses and scoring items through semantic parsing, used to accurately match and generate the second-level directory of the bid document. The directory matching degree algorithm is an evaluation method that quantitatively verifies the fit between the multi-level bid document directory and the bidding information, comprehensively calculating the directory coverage completeness. The preset qualification threshold is the benchmark score for determining whether the multi-level bid document directory meets the bidding response requirements. The hierarchical relationship refers to the subordinate and sorting logical relationships between the multi-level bid document directories. Attachments are standard forms, format templates, supplementary explanations, and other supporting materials that need to be embedded in the bid document.

[0033] Understandably, after the secondary directory generation unit semantically parses the review clauses and scoring items to generate review points, it assigns these review points to matching primary directories in the tender document using preset correspondence rules. This ensures that each review point has a corresponding secondary directory entry, avoiding omissions of tender requirements. When the directory matching score fails to reach the preset threshold, the qualification judgment unit proactively sends adjustment prompts to the user and receives user instructions to adjust directory names, hierarchical affiliations, or add / remove entries. Based on this, it recalculates the directory matching score, forming a closed-loop verification mechanism involving human and machine collaboration, until the directory structure highly matches the tender requirements, effectively balancing automation efficiency with the reliability of manual review.

[0034] A directory matching algorithm is used to quantitatively verify the first-level directory and the second-level directory of the tender document against the structured bidding information, generating a first matching degree and a second matching degree. The first matching degree and the second matching degree are then weighted and summed to generate a directory matching degree score. The corresponding calculation formula is as follows: In the formula, This indicates the total number of bidding keywords extracted from structured bidding information; This indicates the total number of review points extracted from the structured bidding information; This indicates the number of successfully matched bidding keywords in the first-level directory of the tender document; This indicates the number of review points successfully matched in the second-level directory of the tender document; The weight coefficient representing the first degree of matching. The weight coefficients representing the second degree of matching satisfy... and .

[0035] The first matching degree characterizes the degree of alignment between the first-level headings of the tender document and the bidding keywords. The second matching degree characterizes the degree of alignment between the second-level headings of the tender document and the review criteria. Weighted summation combines the two matching degrees according to their corresponding weight coefficients to obtain a comprehensive heading matching degree score. Weight coefficients are used to calculate this score. , These are used to allocate the calculation proportions of the two levels of matching degree, and their sum is 1. Value less than This highlights the crucial role of the second-level tender document catalog in directly corresponding to the review and scoring requirements, ensuring that the catalog construction accurately aligns with the bidding review rules and improving the overall fit between the multi-level tender document catalog and bidding needs.

[0036] In practical applications, taking government procurement information service projects as an example, a total of 8 bidding keywords and 15 review points were extracted from structured bidding information. The first-level bid document directory successfully matched 7 of these bidding keywords, and the second-level bid document directory successfully matched 14 review points. (Weight coefficient) Take 0.4, A weighted score of 0.6 is applied. If the weighted score for the directory matching degree reaches the passing standard, the multi-level tender directory is directly deemed qualified. This weighting method highlights the priority of the secondary directory's response to the review points, ensuring that the directory accurately covers core bidding requirements such as qualification review and scoring items, thus preventing omissions or deviations in the response of the tender documents from the outset.

[0037] The objective content generation module is used to perform format recognition and differential parsing on the pre-stored objective content of the tender document, generate objective fill-in content for the tender document, and fill the objective fill-in content into the overall framework of the tender document, specifically including: An objective content recognition unit is used to call the pre-stored objective content of the tender document, perform format recognition on the objective content of the tender document, and determine the file type of the objective content of the tender document, wherein the file type includes scanned files and electronic files; An objective content parsing unit is used to extract the text information of the objective content of the tender document using optical character recognition technology when the document type is a scanned document, and to perform structured parsing of the text information based on preset field mapping rules to generate the objective content to fill in the tender document; when the document type is an electronic document, it directly extracts the structured information of the electronic document as the objective content to fill in the tender document. An objective content mapping unit is used to perform semantic matching between the last-level directory and the objective content of the tender document based on the directory characteristics of the last-level directory in the qualified multi-level tender document directory, establish a mapping relationship between the last-level directory and the objective content of the tender document, and generate an objective content filling mapping table. An objective content filling unit is used to associate and fill the objective content of the tender document with the overall framework of the tender document according to the objective content filling mapping table, and to perform format adaptation and content quality verification on the filled objective content of the tender document.

[0038] The objective content to be filled in the tender document consists of information that has been parsed and organized to conform to the tender document specifications and can be directly filled in. Scanned files are image scans of qualification certificates. Electronic files are electronic documents with directly readable field information. Optical character recognition (OCR) technology is used to convert the image text in scanned files into editable text. Preset field mapping rules are pre-defined standards for the correspondence between certificate fields and tender document filling positions. Structured parsing organizes the extracted fragmented text into standardized field formats. Directory features are the attributes and response requirements of the lowest-level directory in a qualified multi-level tender document directory. Semantic matching accurately matches the corresponding objective content to be filled in the tender document based on the semantics of the qualified multi-level tender document directory. The objective content filling mapping table is a list recording the correspondence between the lowest-level directory and the objective content to be filled in the tender document. Format adaptation adjusts the objective content to be filled in the tender document to a unified tender document format. Content quality verification checks the completeness, accuracy, and compliance of the objective content to be filled in the tender document.

[0039] Understandably, the preset field mapping rules pre-configure the correspondence between extracted fields and the locations to be filled in the tender document for different document types. For example, business licenses are mapped to fields such as "Company Name," "Unified Social Credit Code," "Legal Representative," "Registered Capital," "Date of Establishment," and "Business Scope," while qualification certificates are mapped to fields such as "Certificate Name," "Certificate Number," "Qualification Level," "Issuing Authority," and "Validity Period." This ensures that the scattered text extracted by optical character recognition in the scanned document can be categorized by field and transformed into structured information. When performing semantic matching, the objective content mapping unit calculates the similarity between the directory features of the final-level directory and the field tags of the objective content to be filled in the tender document, automatically establishing an association between the "Qualification Certificate" directory and the corresponding qualification certificate content, avoiding the tedious manual operation of searching for each item.

[0040] In practical applications, a construction company participating in a public tender needs to submit objective documents such as its business license, construction enterprise qualification certificate, safety production license, project manager's registered construction engineer certificate, and similar performance contracts from the past three years. The objective content recognition unit retrieves pre-stored data and determines that the business license is an electronic file, while the other four types are scanned documents. The objective content parsing unit performs optical character recognition (OCR) to extract text from the scanned documents and structures fields such as "qualification category and level," "certificate number," and "expiration date" from the qualification certificate according to preset field mapping rules. For example, the extracted information on the first-level qualification for general contracting of municipal public works construction is organized to generate the objective content to be filled in the tender document. The objective content mapping unit performs semantic matching based on the directory features of the final-level directory "Qualification Review Materials - Qualification Certificates," automatically establishing mapping relationships and filling them into the corresponding positions. Calculations show that this module reduces the traditional manual process of searching, verifying, and pasting approximately 45 supporting documents to within 2 minutes, with a filling accuracy rate of over 98%, effectively avoiding the risks of document misalignment and information omission.

[0041] The subjective content generation module is used to take the lowest-level directory in the qualified multi-level tender document directory as the generation unit, generate subjective content for the tender document based on the structured bidding information, through enhanced retrieval generation technology and preset constraint rules, perform quality assessment and iterative correction on the subjective content, generate subjective content to fill in the tender document, and associate the subjective content to fill in the overall framework of the tender document. Specifically, it includes: The subjective content retrieval unit is used to use the last-level directory in the qualified multi-level tender document directory as the retrieval anchor point, retrieve multi-source tender information associated with the last-level directory from the structured tender information and the preset external knowledge base, and perform vectorization encoding on the multi-source tender information to generate retrieval enhancement information vectors. The subjective content constraint unit is used to generate initial subjective content corresponding to the last-level directory based on the retrieval enhancement information vector and the preset constraint rules, through the enhanced retrieval generation technology, and to perform constraint compliance detection and iterative correction on the initial subjective content, and output the tender subjective content. The subjective content evaluation and correction unit is used to quantitatively evaluate the subjective content of the tender document through a preset quality evaluation model, generate a subjective content quality score, perform hierarchical iterative correction on the subjective content of the tender document based on the subjective content quality score, and generate the subjective content to be filled in the tender document. The subjective content filling unit is used to associate and fill the subjective content of the tender document into the overall framework of the tender document according to the correspondence between the last-level directory and the subjective content of the tender document, and to perform format adaptation and content quality verification on the filled subjective content of the tender document.

[0042] The search anchor uses the lowest-level directory as the search basis to accurately locate and match the required reference information. The external knowledge base is a pre-set reference resource library including bidding regulations, industry standards, historical high-quality bids, and project cases. Multi-source bidding information is a comprehensive set of information, including bidding requirements, technical specifications, and review rules, obtained from structured bidding information and the external knowledge base. Vectorization encoding converts text information into vector data, improving the accuracy of information retrieval and matching. The enhanced retrieval information vector is a set of reference information after vectorization processing.

[0043] Following this, the initial subjective content is the unverified response content initially generated based on preset constraint rules. Constraint compliance testing verifies whether the initial subjective content meets the preset constraint requirements. Iterative correction involves repeatedly adjusting and optimizing substandard initial subjective content. The quality assessment model is a standard model for quantitatively evaluating the quality of the initial subjective content. The subjective content quality score is a quantitative result measuring the pass rate of the initial subjective content. Graded iterative correction implements different correction strategies according to the subjective content quality score level. The final subjective content to be filled in the tender document is the verified and qualified content that can be directly filled in.

[0044] Specifically, format adaptation refers to automatically unifying the subjectively filled content of the tender document with the preset formatting specifications such as font, font size, line spacing, and paragraph spacing of the overall tender document framework. This eliminates potential layout differences after content filling and ensures a consistent overall style. Content quality verification involves performing a secondary check on the subjectively filled content after completion. This check focuses on detecting anomalies such as field overflow, garbled special characters, and line breaks across pages. It also verifies the accuracy of the correspondence between the filled position and the final-level directory. If the verification finds incompleteness or format conflicts, it promptly alerts the user, thereby eliminating the risk of content misalignment and visual defects before output, ensuring the standardization and submitability of the intelligent tender document.

[0045] The subjective content constraint unit is used to generate initial subjective content corresponding to the last-level directory based on the retrieval enhancement information vector and the preset constraint rules, using the enhanced retrieval generation technology, and to perform constraint compliance detection and iterative correction on the initial subjective content, and output the tender subjective content, specifically including: The constraint prompt construction unit is used to fuse the retrieval enhancement information vector with the semantic vector of the last-level directory to construct a prompt context, construct constraint control instructions according to the length constraint rules, logical structure constraint rules and key point coverage constraint rules in the preset constraint rules, embed the constraint control instructions into the prompt context, and generate constrained prompts. The constraint-guided generation unit is used to input the constraint-based generation prompts into a preset enhanced retrieval generation model and generate initial subjective content corresponding to the last-level directory. The constraint compliance detection unit is used to perform constraint compliance detection on the initial subjective content, and calculate the first compliance degree of the initial subjective content with the length constraint rule, the second compliance degree of the initial subjective content with the logical structure constraint rule, and the third compliance degree of the initial subjective content with the key point coverage constraint rule. The constraint correction iteration unit is used to determine the constraint rules corresponding to compliance degrees less than the preset sub-item threshold as constraints to be corrected, and to correct the initial subjective content based on the constraints to be corrected until all compliance degrees are greater than or equal to the preset sub-item threshold. The corrected initial subjective content is then used as the tender subjective content and output.

[0046] The semantic vector of the final-level directory is a numerical vector that transforms the textual meaning of the smallest heading unit in a qualified multi-level tender document directory into a machine-recognizable vector, used to accurately represent the semantic information of the directory. The generated prompt context is a set of basic information generated by fusing the retrieval enhancement information vector and the semantic vector. The length constraint rule limits the word count range of the initial subjective content. The logical structure constraint rule standardizes the logical framework for content writing. The key point coverage constraint rule requires the content to fully respond to the core review points of the tender. The constraint control instructions are operational instructions that transform the three types of constraint rules into executable instructions for the enhanced retrieval generation model. The constrained generation prompts are complete generation guidance information embedded with the constraint control instructions.

[0047] Furthermore, the enhanced retrieval generation model is an intelligent generation model used to generate initial subjective content that aligns with the final-level directory and bidding requirements. First conformity, second conformity, and third conformity represent the degree of matching between the initial subjective content and the constraints on length, logical structure, and key point coverage, respectively. Preset sub-thresholds are benchmark values ​​for determining whether the conformity of various constraints meets the standards. Constraints requiring correction are those that do not meet the conformity standards and need targeted adjustments.

[0048] Understandably, after receiving constrained generation prompts, the enhanced retrieval generation model does not rely solely on its own parameter memory for content generation. Instead, it uses the vectorized retrieval enhancement information vector as a reference source for the generation context, ensuring that the generation process is always anchored to structured bidding information and authentic materials provided by external knowledge bases. When generating the subjective content of the tender document word by word, the model dynamically weighs the relevance weights of each part in the retrieval enhancement information vector through an attention mechanism, prioritizing reference fragments with high semantic similarity to the final-level directory. This ensures fluency while significantly suppressing illusions such as fabricated clauses and technical parameters, guaranteeing that the output content is based on evidence and is professionally accurate.

[0049] In this way, the subjective content constraint unit embeds three types of constraint rules—length, logical structure, and key point coverage—into the generation prompt context, guiding the enhanced retrieval generation model to produce initial subjective content that conforms to the specification framework. Through multi-dimensional compliance detection and targeted iterative correction, it ensures that the output content meets the preset standards in terms of controllable length, rigorous structure, and no omissions of key points, effectively overcoming the defects of generated content that deviates from bidding requirements, has loose logic, or omits key review points.

[0050] The subjective content of the tender document is quantitatively evaluated using a preset quality assessment model to generate a subjective content quality score. The corresponding calculation formula is as follows: In the formula, T represents the semantic vector representation of the final-level directory; C represents the semantic vector representation of the subjective content of the tender document; and R represents the set of key tender information points retrieved from the structured tender information that corresponds to the final-level directory. Indicates the semantic similarity between the final-level directory and the subjective content of the tender document; This indicates the coverage of key bidding information points within the main content of the bid document; This indicates the content quality score of the subjective content of the tender document; These represent the weighting coefficients for semantic similarity, coverage, and content quality score, respectively, satisfying... and .

[0051] The semantic vector representation of the subjective content of the tender document is a numerical feature obtained by homologous encoding of the subjective content. The key points set of bidding information is a list of response clauses corresponding to the final-level directory extracted from the structured bidding information. Semantic similarity is calculated using cosine distance to determine the directional consistency between the final-level directory and the subjective content of the tender document in the semantic space. Coverage measures the proportion of each clause in the key points set of bidding information that is explicitly responded to in the subjective content of the tender document. The content quality score is used to comprehensively evaluate the logical coherence, linguistic standardization, and lack of redundancy in the subjective content of the tender document. Weighting coefficients reflect the differences in the contribution of each dimension to the subjective content quality score.

[0052] Understandably, the weighting coefficient for semantic similarity is greater than that for coverage, which in turn is greater than that for content quality score. Semantic similarity has the highest weighting coefficient, aiming to ensure that the generated content closely matches the theme of the final-level directory, fundamentally preventing the content from deviating from the direction of the tender response. Coverage has the next highest weighting coefficient, emphasizing a comprehensive response to the key points of the tender information, preventing the omission of crucial review clauses. The weighting coefficient for content quality score, while ensuring the correct content direction and completeness of key points, also considers the fluency and standardization of the text. This weighting method effectively curbs the problem of simply pursuing flowery language while neglecting substantive content response.

[0053] In practical application, for a section of subjective content generated from the final-level directory "Technical Solution Response," the final-level directory and the subjective content are first encoded into semantic vectors, and their semantic similarity is calculated, resulting in a score of 0.92. Simultaneously, eight key bidding information points related to "Technical Solution Response" are extracted from the structured bidding information. Verification shows that the subjective content actually covers seven of these points, resulting in a coverage of 0.875. The content quality score, assessed for logical coherence and language standardization, is 0.90. Using preset weighting coefficients of 0.45, 0.35, and 0.20, the subjective content quality score is 0.90025, equivalent to approximately 90 points. Since the subjective content quality score falls within the first scoring range, this content requires no correction and can be directly used as the subjective content for the bid. If a piece of content has a semantic similarity of only 0.70 but acceptable coverage and quality scores, the overall score will be significantly lower due to the highest weight of semantic similarity, triggering a subsequent graded correction process to ensure that the final output of the subjective content strictly adheres to the requirements of the bidding directory.

[0054] The step of performing graded iterative correction on the subjective content of the tender document based on the subjective content quality score to generate the subjective content to fill in the tender document specifically includes: The subjective content quality score is graded and determined by setting a first score interval, a second score interval, and a third score interval, wherein the scores of the first score interval, the second score interval, and the third score interval decrease sequentially. When the subjective content quality score is within the first score range, the subjective content of the tender document will be directly used as the subjective content to fill in the tender document. When the subjective content quality score is in the second scoring range, the semantic similarity, the coverage, and the content quality score are compared with the corresponding preset dimension thresholds. Defect dimensions that are less than the preset dimension thresholds are identified. Based on the defect dimensions, the corresponding preset correction template is called to perform targeted completion or logical reorganization of the subjective content of the tender document, and the subjective content quality score is recalculated until the subjective content quality score is in the first scoring range. When the subjective content quality score is within the third score range, it is determined that the subjective content of the tender document has structural defects. The retrieval range of the enhanced retrieval information vector and the constraint parameters of the preset constraint rules are adjusted, triggering the enhanced retrieval generation technology to regenerate the subjective content of the tender document. The regenerated subjective content of the tender document is then subjected to quality evaluation and iterative correction until the subjective content of the tender document is generated.

[0055] The grading system is a method of determining content quality based on the numerical range of subjective content quality scores, classifying content into acceptable levels and implementing corresponding processing strategies. The first scoring range represents content that meets quality standards and can be directly used for content filling; the second scoring range represents content with minor flaws requiring targeted correction; and the third scoring range represents content with serious problems requiring regeneration. Preset dimension thresholds are the baseline values ​​for determining whether semantic similarity, coverage, and content quality scores meet requirements. Defect dimensions are those that fail to reach the preset dimension thresholds and are considered unacceptable.

[0056] Pre-defined correction templates are standardized correction models pre-designed for different defect dimensions. Targeted completion involves supplementing missing bidding points, while logical restructuring involves reorganizing logically confused content. Structural defects refer to fundamental deviations between the overall content framework, response logic, and bidding requirements. The search scope defines the boundaries for acquiring enhanced information vectors, and constraint parameters are the specific numerical limits of pre-defined constraint rules.

[0057] Understandably, the main purpose of tiered iterative correction is to implement differentiated optimization strategies based on the subjective content quality score's range, avoiding over-processing of high-quality content or insufficient correction of low-quality content. When the subjective content quality score is in the second range, the system does not rewrite the entire content, but rather precisely identifies the defective dimensions. For example, if the coverage is insufficient, the preset correction template will automatically extract unresponsive clauses from the key points of the bidding information set, generate supplementary paragraphs according to the preset wording format, and embed them in the appropriate positions in the original text. If the logical structure is loose, the paragraph order and hierarchical relationship will be rearranged according to the logical structure constraint rules. This targeted correction method retains the effective parts of the original content, only optimizing the shortcomings, and the correction efficiency is significantly improved compared to a complete rewrite. When the subjective content quality score is in the third range, it is determined that the content can no longer meet the standard through partial repair. The system automatically expands the retrieval scope of the enhanced information vector or relaxes the constraints of the parameters, triggering enhanced retrieval generation technology to regenerate the content, improving the generated quality from the source.

[0058] Through the above methods, the hierarchical iterative correction implements a differentiated processing strategy based on the interval of the subjective content quality score. It performs targeted completion or logical reorganization of local defective content, and regenerates structural defective content after adjusting the search scope and constraint parameters. This ensures that the subjective content filled in the output tender document meets the preset standards in terms of semantic fit, key point coverage and intrinsic quality, effectively balancing correction efficiency and content quality. Example

[0059] like Figure 2 As shown, a multi-stage intelligent method for automatically generating tender documents includes: S1, Collect bidding requirement information and auxiliary reference information, preprocess the bidding requirement information and auxiliary reference information, and generate structured bidding information; S2, construct a multi-level bid document directory based on the structured bidding information, verify the multi-level bid document directory and the structured bidding information according to the matching degree evaluation rules, generate a qualified multi-level bid document directory, and construct the overall framework of the bid document based on the qualified multi-level bid document directory; S3, perform format recognition and differential parsing on the pre-stored objective content of the tender document, generate objective fill content for the tender document, and associate the objective fill content of the tender document with the overall framework of the tender document; S4, taking the last-level directory in the qualified multi-level tender document directory as a generation unit, generating subjective content of the tender document based on the structured bidding information through enhanced retrieval generation technology and preset constraint rules, performing quality assessment and iterative correction on the subjective content of the tender document, generating subjective filling content of the tender document, and associating the subjective filling content of the tender document with the overall framework of the tender document. S5. After the overall framework of the tender document is filled in, perform format standardization processing and output intelligent tender document file.

[0060] As described in the above embodiments, the present invention includes a data acquisition module for collecting bidding requirement information and auxiliary reference information, preprocessing the bidding requirement information and auxiliary reference information, and generating structured bidding information; a directory framework construction module for constructing a multi-level bid document directory based on the structured bidding information, verifying the multi-level bid document directory against the structured bidding information according to matching degree evaluation rules, generating a qualified multi-level bid document directory, and constructing an overall bid document framework based on the qualified multi-level bid document directory; and an objective content generation module for performing format recognition and differential parsing on pre-stored objective content of the bid document, generating objective content to fill in the bid document, and associating the objective content to fill in the bid document. The system comprises the overall framework of the tender document; a subjective content generation module, which uses the lowest-level directory in the qualified multi-level tender document directory as the generation unit, generates subjective content based on structured bidding information through enhanced retrieval generation technology and preset constraint rules, performs quality assessment and iterative correction on the subjective content, generates subjective fill-in content, and associates the subjective fill-in content with the overall tender document framework; and a tender document output module, which performs format standardization processing on the completed overall tender document framework and outputs an intelligent tender document file. This enables automation and intelligence of the entire tender document generation process, balancing generation efficiency and content quality, effectively avoiding the risk of bid rejection, and improving the success rate of bidding.

[0061] This invention utilizes a data acquisition module to perform structured preprocessing of bidding requirements and auxiliary reference information, effectively improving information extraction accuracy and reducing human interpretation bias. Based on structured bidding information, this invention automatically constructs a multi-level bid document catalog, combining it with matching degree evaluation rules for quantitative verification. This ensures a high degree of alignment between the catalog hierarchy and bidding requirements, significantly reducing the time spent manually building the catalog framework. It also features a human-machine collaborative adjustment mechanism for insufficient catalog matching, further guaranteeing the reliability and relevance of the catalog construction. Furthermore, this invention performs format recognition and differential parsing on pre-stored objective content of the bid documents, automatically adapting the parsing path to scanned or electronic versions based on file type. This enables accurate extraction, structured mapping, and error-free filling of fixed content such as company qualifications and performance certificates, fundamentally avoiding the risks of manual input errors and non-standard formats. This invention uses the lowest-level directory as the generation unit, integrating enhanced retrieval generation technology and preset constraint rules to guide the generation of subjective content in the tender document. Through a quality assessment model and a hierarchical iterative correction mechanism, it optimizes the semantic relevance, key point coverage, and intrinsic quality of the content in multiple dimensions, ensuring that the generated content is both professional and strictly adheres to bidding specifications. This overcomes the shortcomings of traditional large-scale models, such as uncontrollable content generation and vague professional details. The invention also completes global format standardization processing through the tender document output module, outputting intelligent tender document files that meet bidding requirements.

[0062] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0063] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-Erasable Programmable Read-Only Memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium capable of carrying or storing data.

[0064] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

Claims

1. A multi-stage intelligent automated tender document generation system, characterized in that, The system includes: The data acquisition module is used to collect bidding requirement information and auxiliary reference information, preprocess the bidding requirement information and auxiliary reference information, and generate structured bidding information. The directory framework construction module is used to construct a multi-level bid document directory based on the structured bidding information, verify the multi-level bid document directory against the structured bidding information according to the matching degree evaluation rules, generate a qualified multi-level bid document directory, and construct the overall bid document framework based on the qualified multi-level bid document directory. The objective content generation module is used to perform format recognition and differential parsing on the pre-stored objective content of the tender document, generate objective fill content for the tender document, and associate the objective fill content of the tender document with the overall framework of the tender document; The subjective content generation module is used to take the last-level directory in the qualified multi-level tender document directory as the generation unit, generate the subjective content of the tender document based on the structured bidding information, and generate the subjective content of the tender document through enhanced retrieval generation technology and preset constraint rules. The module performs quality assessment and iterative correction on the subjective content of the tender document, generates subjective filling content of the tender document, and associates and fills the subjective filling content of the tender document into the overall framework of the tender document. The tender document output module is used to format and standardize the overall framework of the completed tender document, and output an intelligent tender document file.

2. The multi-stage intelligent automated tender document generation system according to claim 1, characterized in that, The directory framework construction module is used to construct a multi-level bid document directory based on the structured bidding information, verify the multi-level bid document directory against the structured bidding information according to the matching degree evaluation rules, generate a qualified multi-level bid document directory, and construct an overall bid document framework based on the qualified multi-level bid document directory, specifically including: The first-level directory generation unit is used to extract the structured bidding information, generate bidding keywords, match the bidding keywords with a preset first-level directory template library, and generate the first-level directory of the tender document. The secondary directory generation unit is used to perform semantic parsing on the review clauses and scoring items in the structured bidding information, generate review points, and generate a secondary directory of the tender document belonging to the primary directory of the tender document based on the correspondence between the review points and the primary directory of the tender document. The directory matching degree evaluation unit is used to integrate the first-level directory of the tender documents and the second-level directory of the tender documents to construct the multi-level tender document directory, and to use the directory matching degree algorithm to quantitatively verify the multi-level tender document directory and the structured bidding information to generate a directory matching degree score. The qualification determination unit is used to determine the multi-level tender document directory as the qualified multi-level tender document directory when the directory matching score is greater than or equal to the preset qualification threshold; when the directory matching score is less than the preset qualification threshold, it receives the user's directory adjustment instruction to adjust the multi-level tender document directory, and redetermines the directory matching score based on the adjusted multi-level tender document directory until the qualified multi-level tender document directory is generated. The framework construction unit is used to construct the overall framework of the tender document based on the hierarchical relationship of the qualified multi-level tender document directory, and to associate the attachment content in the structured bidding information with the corresponding attachment position of the overall framework of the tender document.

3. The multi-stage intelligent automated tender document generation system according to claim 2, characterized in that, A directory matching algorithm is used to quantitatively verify the first-level directory and the second-level directory of the tender document against the structured bidding information, generating a first matching degree and a second matching degree. The first matching degree and the second matching degree are then weighted and summed to generate a directory matching degree score. The corresponding calculation formula is as follows: In the formula, This indicates the total number of bidding keywords extracted from structured bidding information; This indicates the total number of review points extracted from the structured bidding information; This indicates the number of successfully matched bidding keywords in the first-level directory of the tender document; This indicates the number of review points successfully matched in the second-level directory of the tender document; The weight coefficient representing the first degree of matching. The weight coefficients representing the second degree of matching satisfy... and .

4. The multi-stage intelligent automated tender document generation system according to claim 1, characterized in that, The objective content generation module is used to perform format recognition and differential parsing on the pre-stored objective content of the tender document, generate objective fill-in content for the tender document, and fill the objective fill-in content into the overall framework of the tender document, specifically including: An objective content recognition unit is used to call the pre-stored objective content of the tender document, perform format recognition on the objective content of the tender document, and determine the file type of the objective content of the tender document, wherein the file type includes scanned files and electronic files; An objective content parsing unit is used to extract the text information of the objective content of the tender document using optical character recognition technology when the document type is a scanned document, and to perform structured parsing of the text information based on preset field mapping rules to generate the objective content to fill in the tender document; when the document type is an electronic document, it directly extracts the structured information of the electronic document as the objective content to fill in the tender document. An objective content mapping unit is used to perform semantic matching between the last-level directory and the objective content of the tender document based on the directory characteristics of the last-level directory in the qualified multi-level tender document directory, establish a mapping relationship between the last-level directory and the objective content of the tender document, and generate an objective content filling mapping table. An objective content filling unit is used to associate and fill the objective content of the tender document with the overall framework of the tender document according to the objective content filling mapping table, and to perform format adaptation and content quality verification on the filled objective content of the tender document.

5. The multi-stage intelligent automated tender document generation system according to claim 1, characterized in that, The subjective content generation module is used to take the lowest-level directory in the qualified multi-level tender document directory as the generation unit, generate subjective content for the tender document based on the structured bidding information, through enhanced retrieval generation technology and preset constraint rules, perform quality assessment and iterative correction on the subjective content, generate subjective content to fill in the tender document, and associate the subjective content to fill in the overall framework of the tender document. Specifically, it includes: The subjective content retrieval unit is used to use the last-level directory in the qualified multi-level tender document directory as the retrieval anchor point, retrieve multi-source tender information associated with the last-level directory from the structured tender information and the preset external knowledge base, and perform vectorization encoding on the multi-source tender information to generate retrieval enhancement information vectors. The subjective content constraint unit is used to generate initial subjective content corresponding to the last-level directory based on the retrieval enhancement information vector and the preset constraint rules, through the enhanced retrieval generation technology, and to perform constraint compliance detection and iterative correction on the initial subjective content, and output the tender subjective content. The subjective content evaluation and correction unit is used to quantitatively evaluate the subjective content of the tender document through a preset quality evaluation model, generate a subjective content quality score, perform hierarchical iterative correction on the subjective content of the tender document based on the subjective content quality score, and generate the subjective content to be filled in the tender document. The subjective content filling unit is used to associate and fill the subjective content of the tender document into the overall framework of the tender document according to the correspondence between the last-level directory and the subjective content of the tender document, and to perform format adaptation and content quality verification on the filled subjective content of the tender document.

6. The multi-stage intelligent automated tender document generation system according to claim 5, characterized in that, The subjective content constraint unit is used to generate initial subjective content corresponding to the last-level directory based on the retrieval enhancement information vector and the preset constraint rules, using the enhanced retrieval generation technology, and to perform constraint compliance detection and iterative correction on the initial subjective content, and output the tender subjective content, specifically including: The constraint prompt construction unit is used to fuse the retrieval enhancement information vector with the semantic vector of the last-level directory to construct a prompt context, construct constraint control instructions according to the length constraint rules, logical structure constraint rules and key point coverage constraint rules in the preset constraint rules, embed the constraint control instructions into the prompt context, and generate constrained prompts. The constraint-guided generation unit is used to input the constraint-based generation prompts into a preset enhanced retrieval generation model and generate initial subjective content corresponding to the last-level directory. The constraint compliance detection unit is used to perform constraint compliance detection on the initial subjective content, and calculate the first compliance degree of the initial subjective content with the length constraint rule, the second compliance degree of the initial subjective content with the logical structure constraint rule, and the third compliance degree of the initial subjective content with the key point coverage constraint rule. The constraint correction iteration unit is used to determine the constraint rules corresponding to compliance degrees less than the preset sub-item threshold as constraints to be corrected, and to correct the initial subjective content based on the constraints to be corrected until all compliance degrees are greater than or equal to the preset sub-item threshold. The corrected initial subjective content is then used as the tender subjective content and output.

7. The multi-stage intelligent automated tender document generation system according to claim 5, characterized in that, The subjective content of the tender document is quantitatively evaluated using a preset quality assessment model to generate a subjective content quality score. The corresponding calculation formula is as follows: In the formula, T represents the semantic vector representation of the final-level directory; C represents the semantic vector representation of the subjective content of the tender document; and R represents the set of key tender information points retrieved from the structured tender information that corresponds to the final-level directory. Indicates the semantic similarity between the final-level directory and the subjective content of the tender document; This indicates the coverage of key bidding information points within the main content of the bid document; This indicates the content quality score of the subjective content of the tender document; These represent the weighting coefficients for semantic similarity, coverage, and content quality score, respectively, satisfying... and .

8. The multi-stage intelligent automated tender document generation system according to claim 7, characterized in that, The step of performing graded iterative correction on the subjective content of the tender document based on the subjective content quality score to generate the subjective content to fill in the tender document specifically includes: The subjective content quality score is graded and determined by setting a first score interval, a second score interval, and a third score interval, wherein the scores of the first score interval, the second score interval, and the third score interval decrease sequentially. When the subjective content quality score is within the first score range, the subjective content of the tender document will be directly used as the subjective content to fill in the tender document. When the subjective content quality score is in the second scoring range, the semantic similarity, the coverage, and the content quality score are compared with the corresponding preset dimension thresholds. Defect dimensions that are less than the preset dimension thresholds are identified. Based on the defect dimensions, the corresponding preset correction template is called to perform targeted completion or logical reorganization of the subjective content of the tender document, and the subjective content quality score is recalculated until the subjective content quality score is in the first scoring range. When the subjective content quality score is within the third score range, it is determined that the subjective content of the tender document has structural defects. The retrieval range of the enhanced retrieval information vector and the constraint parameters of the preset constraint rules are adjusted, triggering the enhanced retrieval generation technology to regenerate the subjective content of the tender document. The regenerated subjective content of the tender document is then subjected to quality evaluation and iterative correction until the subjective content of the tender document is generated.

9. A multi-stage intelligent automated tender document generation method, applied to a multi-stage intelligent automated tender document generation system as described in any one of claims 1-8, characterized in that, The method includes: Collect bidding requirements information and auxiliary reference information, preprocess the bidding requirements information and auxiliary reference information, and generate structured bidding information; A multi-level tender document catalog is constructed based on the structured tender information. The multi-level tender document catalog is verified against the structured tender information according to the matching degree evaluation rules to generate a qualified multi-level tender document catalog. The overall framework of the tender document is then constructed based on the qualified multi-level tender document catalog. The format of the pre-stored objective content of the tender document is identified and differentiated, generating objective fill content for the tender document, and the objective fill content is then associated with and filled into the overall framework of the tender document. Using the lowest level directory in the qualified multi-level tender document directory as the generation unit, based on the structured bidding information, the subjective content of the tender document is generated by enhancing the retrieval generation technology and preset constraint rules. The subjective content of the tender document is then evaluated for quality and iteratively corrected to generate subjective content to fill in the tender document. Finally, the subjective content to fill in the tender document is associated with and filled into the overall framework of the tender document. The overall framework of the completed tender document is formatted and standardized to output an intelligent tender document file.

Images