Intelligent checking and scoring optimization method and system for bidding documents based on double-case library iterative learning

Abstract

The application discloses a kind of based on double case library iterative learning's tender file intelligent checking and scoring optimization method and system, method includes: import and parse tender document and tender document, generate structured comparison baseline;Based on enterprise historical case library and industry public case library, depth learning training is carried out, and double-layer checking model suitable for the adaptation of enterprise characteristics and industry general specification is generated;Structured comparison baseline is input into double-layer checking model and is carried out compliance inspection, and grading checking report is generated;According to the feedback data generated after artificial marking to grading checking report, the model is iteratively trained, and the updated double-layer checking model is generated;Based on the updated double-layer checking model, scoring optimization report is generated.Using the embodiment of the application, the closed-loop processing from risk compliance checking to objective scoring optimization of tender document can be realized, and the checking efficiency is improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically a method and system for intelligent verification and scoring optimization of tender documents based on iterative learning with a dual case library. Background Technology

[0002] Bidding and tendering activities are a core part of a company's acquisition of business opportunities, and the quality of the bid documents directly affects the probability of winning the bid. Currently, bid documents for large-scale engineering projects are often thousands of pages long, covering multiple dimensions such as business, technology, and qualifications. Manual verification is not only inefficient but also prone to overlooking risk factors for bid rejection and scoring errors. Existing electronic bid evaluation tools mostly use keyword matching and template filling, which can only complete format checks and basic compliance verification, failing to deeply understand the relationship between the semantic requirements of the bidding documents and the actual response of the bid documents. Meanwhile, companies have accumulated a large number of bid rejection and scoring error cases from their past bidding experiences, and publicly available industry case databases contain rich evaluation experience and common scoring patterns, but this valuable knowledge is almost entirely unused in existing verification tools. More importantly, after verification, existing tools cannot provide systematic score statistics and scoring error diagnosis based on objective scoring rules, lacking a closed-loop capability from "problem discovery" to "scoring optimization," making it difficult to substantially improve the quality of bid documents. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for intelligent verification and scoring optimization of tender documents based on iterative learning with a dual case library, in order to overcome the shortcomings of the existing technology, realize the closed-loop processing of tender documents from risk compliance verification to objective scoring optimization, and improve verification efficiency.

[0004] One embodiment of this application provides a method for intelligent verification and scoring optimization of tender documents based on iterative learning with a dual-case library, the method comprising: Import and parse the tender documents and bid documents, integrate all the scoring clauses and hard clauses for rejection in the tender documents, establish a mapping relationship with the content of the bid documents, and generate a structured comparison baseline; Deep learning training is conducted based on the enterprise's historical case library and the industry's publicly available case library to transform the reasons for bid rejection and the points lost in the cases into standardized verification rules, generating a two-layer verification model that adapts to the characteristics of the enterprise and the general industry standards. The structured comparison baseline is input into the two-layer verification model for compliance checks. The detected issues are marked with the corresponding clauses, document locations, and risk levels, and a graded verification report is generated. Based on the feedback data generated after manually annotating the graded verification report, the model is iteratively trained to generate an updated two-layer verification model. Based on the updated two-layer verification model, the objective scoring details in the bidding documents are decomposed and matched with the bidding documents. The objective scores, deduction points and missing data items are statistically analyzed, and a scoring optimization report is generated.

[0005] Another embodiment of this application provides an intelligent verification and scoring optimization system for tender documents based on iterative learning with a dual-case library, the system comprising: The parsing module is used to import and parse the bidding documents and tender documents, integrate all the scoring clauses and hard clauses for rejection in the bidding documents, establish a mapping relationship with the content of the tender documents, and generate a structured comparison baseline. The training module is used for deep learning training based on the enterprise's historical case library and the industry's public case library. It transforms the reasons for bid rejection and the points lost in the cases into standardized verification rules, and generates a two-layer verification model that adapts to the enterprise's characteristics and the industry's general standards. The verification module is used to input the structured comparison baseline into the two-layer verification model for compliance checks, mark the corresponding relationship of clauses, document location and risk level of the detected issues, and generate a graded verification report; The optimization module is used to iteratively train the model based on the feedback data generated after manually annotating the graded verification report, and generate an updated two-layer verification model. The scoring module is used to decompose the objective scoring details in the bidding documents and match them with the bidding documents based on the updated two-layer verification model, to calculate the objective scores, deduction points and missing data items, and to generate a scoring optimization report.

[0006] Another embodiment of this application provides a storage medium storing a computer program, wherein the computer program is configured to execute the method described in any of the preceding claims when running.

[0007] Another embodiment of this application provides an electronic device including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the method described in any of the preceding claims.

[0008] Compared with existing technologies, the present invention provides an intelligent verification and scoring optimization method for tender documents based on iterative learning of dual case libraries, which can realize closed-loop processing of tender documents from risk compliance verification to objective scoring optimization, thereby improving verification efficiency. Attached Figure Description

[0009] Figure 1 A hardware structure block diagram of a computer terminal for an intelligent verification and scoring optimization method for tender documents based on iterative learning with a dual case library, provided in an embodiment of the present invention; Figure 2A flowchart illustrating an intelligent verification and scoring optimization method for tender documents based on iterative learning with a dual-case library, provided for an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an intelligent verification and scoring optimization system for tender documents based on iterative learning with a dual case library, provided in an embodiment of the present invention. Detailed Implementation

[0010] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0011] This invention first provides a method for intelligent verification and scoring optimization of tender documents based on iterative learning with a dual case library. This method can be applied to electronic devices, such as computer terminals, specifically ordinary computers.

[0012] The following detailed explanation uses a computer terminal as an example. Figure 1 This is a hardware structure diagram of a computer terminal for an intelligent verification and scoring optimization method for tender documents based on iterative learning with a dual-case library, provided as an embodiment of the present invention. Figure 1 As shown, the computer device includes a processor, memory, and network interface connected via a system bus, wherein the memory may include non-volatile storage media and internal memory.

[0013] See Figure 2 The embodiments of the present invention provide a method for intelligent verification and scoring optimization of tender documents based on iterative learning with a dual-case library, which may include the following steps: S201: Import and parse the tender documents and bid documents, integrate all the scoring clauses and hard clauses for rejection in the tender documents, establish a mapping relationship with the content of the bid documents, and generate a structured comparison baseline. Specifically, the system can receive user-uploaded tender documents and bid documents through a batch import interface. The tender documents include the main tender announcement, supplementary announcement, Q&A documents, and specification documents, while the bid documents include a credit book, a technical book, and a commercial book, generating a multi-source heterogeneous file set. The core of this step is to establish a unified document access channel, comprehensively collect original bidding documents from different sources and of different types, complete the classification, labeling, and validity verification of the documents, and form a set of original documents covering all bidding rules and bid response content. This provides complete basic materials for subsequent document parsing and clause comparison. The specific implementation method is as follows: The batch import interface is a standardized file upload channel open to users, supporting both single file upload and batch folder upload modes. The interface has built-in file format and integrity verification mechanisms, automatically identifying the legality of file extensions and blocking corrupted, encrypted, or unsupported files. It also limits the size of uploaded files, with a maximum single file size of 200MB and a maximum of 50 files per batch, ensuring the stability of file transmission and subsequent parsing. User-uploaded files are divided into two main categories: tender documents and bid documents. Tender documents include four core types: the main tender announcement, which is the core original document of the tender rules, clearly defining the basic project requirements, review rules, and rejection clauses; supplementary announcements, which revise and supplement the main announcement, typically including information such as parameter adjustments and time limit changes; Q&A documents, which are official responses to bidders' questions and have the same rule-making effect as the main announcement; and specification documents, which are reference documents such as industry technical standards and acceptance specifications that the project must follow. The tender documents are divided into three volumes: the creditworthiness volume, which displays the company's qualifications, performance, personnel, honors and other creditworthiness documents; the technical volume, which presents the technical response content such as technical solutions, implementation plans, personnel configuration, quality assurance and other technical response content; and the commercial volume, which provides commercial content such as price details, commercial terms and payment methods.

[0014] In the example, a user's single upload of tender documents includes 1 main tender announcement, 2 supplementary announcements, 3 Q&A documents, and 4 supporting specification documents. The bid documents include 1 credit document, 1 technical document, and 1 commercial document, totaling 12 original documents. After the interface completes upload verification, it assigns a unique file identifier to each file, and also marks the file's category, subcategory, upload time, file size, and version number. All valid files are then uniformly collected, ultimately generating a multi-source heterogeneous file set. "Multi-source" refers to files originating from different publishing entities and serving different functional roles; "heterogeneous" refers to differences in file format and content structure. This set fully preserves the ontology information and attribute tags of all original files without modifying the file content, ensuring that subsequent parsing work is based on the original, authentic file content.

[0015] Each file in the multi-source heterogeneous file set is processed for format identification and classification. A multimodal document parsing algorithm is used to extract the text layer of PDF format, the paragraph structure of Word format, and the table data of Excel format, respectively, to generate the structured text content of each file. The core of this step is to adopt an adaptive parsing strategy for documents of different formats, transforming unstructured and semi-structured raw documents into standardized structured text with positional tags. This eliminates data barriers caused by format differences, fully preserving the text content, hierarchical structure, and table information, and providing a unified data foundation for subsequent clause integration and semantic matching. The specific implementation method is as follows: The multimodal document parsing algorithm is an integrated document processing algorithm that combines multiple format parsing engines. It can automatically identify the file format type and match the corresponding parsing logic without requiring manual switching of parsing tools. The overall text recognition accuracy can reach over 98%, and the table structure restoration accuracy is no less than 95%. "Multimodal" means that the algorithm adapts to various document content formats, covering editable text, scanned text, structured tables, and other content. For PDF files, the algorithm prioritizes extracting the built-in editable text layer, reading the text content page by page, and recording the page number, line number, and coordinate position of each paragraph. It automatically filters redundant information such as headers, footers, and page watermarks. For PDF pages containing scanned documents, the algorithm calls the built-in optical character recognition module for text extraction, ensuring that scanned documents can also be parsed. For Word files, the algorithm focuses on restoring the document's paragraph hierarchy, recognizing the heading styles of first-level, second-level, and third-level headings, preserving the complete table of contents logic and chapter relationships, and extracting formatting marks such as bold and underline within paragraphs, corresponding to key requirements in tender terms and response identifiers in bid documents. For Excel files, the algorithm extracts cell data worksheet by worksheet, preserving the table's row and column structure and cell merging relationships. It identifies the table header row and column, converting discrete cell data into structured table text with row and column labels. Simultaneously, it records the worksheet name and cell coordinates corresponding to each data item, facilitating subsequent location and tracing.

[0016] In the example, a 36-page PDF master tender notice was parsed, extracting 217 valid text paragraphs. Each paragraph was labeled with its corresponding page number and starting line number, and 12 instances of duplicate header and footer information were filtered out. A Word document containing a tender technical manual was parsed, restoring the complete hierarchical structure of 8 first-level chapters and 32 second-level chapters, resulting in 1246 extracted paragraphs. An Excel-formatted price list was parsed, restoring a 6-column, 28-row table structure, fully preserving the name, value, and unit of each price item. All parsed content was uniformly categorized by file, with location metadata appended to each text fragment, ultimately generating structured text content for each document. This content retains all textual information from the original document while possessing standardized location and structural markers, allowing for direct use in subsequent clause merging and semantic matching.

[0017] The structured text content of all bidding documents is merged and deduplicated, old clauses that have been covered by supplementary announcements are identified and removed, the latest valid scoring details and hard clauses for rejection are retained, and a set of integrated bidding clauses is generated. The core of this step is to merge and clean the rule content of multiple bidding documents, eliminating duplicates, conflicts, and outdated content between different versions, and filtering out the core scoring and rejection clauses to form a unified, accurate, and unambiguous set of bidding rules. This avoids deviations in subsequent verification caused by conflicts between multiple versions of rules. The specific implementation method is as follows: The deduplication process first sorts all tender documents by their publication date. Using the main tender announcement as the base text, supplementary announcements, Q&A documents, and specification documents are included sequentially from earliest to latest publication date to ensure accurate rule updates. Clause deduplication employs a semantic similarity matching mechanism. Each clause to be matched is transformed into a 768-dimensional semantic vector using a semantic vector model. The semantic overlap between two clauses is then calculated using the cosine similarity formula: sim(v_1,v_2)=(v_1·v_2) / (|v_1|×|v_2|), where v_1 and v_2 represent the semantic vectors of the two clauses, and sim represents the semantic similarity between the two clauses, ranging from 0 to 1. A value closer to 1 indicates a higher degree of semantic overlap. The similarity threshold is set to 0.9. When the similarity calculation result of two clauses is greater than or equal to 0.9, the two clauses are determined to be different expressions of the same rule. In this case, the clause version with the later publication time is retained and the older clause with the earlier publication time is removed, so as to realize the automatic deduplication and version update of clauses.

[0018] After deduplication, all retained clauses are categorized and labeled into two main categories: scoring details and mandatory rejection clauses. Scoring details are clauses corresponding to the review scoring items, each labeled with its corresponding score, scoring criteria, and scoring conditions. Mandatory rejection clauses are clauses that directly trigger rejection if not met, labeled with the triggering conditions and consequences of violation. In the example, the original clause in the main tender announcement was "Bidders must provide no less than 5 similar project performance records in the past 3 years," corresponding to a score of 3 points. The first supplementary announcement updated this requirement to "Bidders must provide no less than 8 similar project performance records in the past 3 years," with the corresponding score remaining unchanged. The semantic similarity between the two clauses was calculated to be 0.92, exceeding the 0.9 threshold. Therefore, the system retained the 8 versions from the supplementary announcement and removed the 5 older versions from the original announcement. After all processing was completed, a total of 72 valid scoring details and 18 mandatory rejection clauses were retained. Each clause is accompanied by a unique number, clause type, source of publication, original content, and corresponding score, ultimately generating an integrated set of tender clauses. This set eliminates conflicts and redundancy among multiple versions of files and serves as the sole rule benchmark for the entire verification process.

[0019] The integrated set of bidding terms is semantically matched with the structured text content of the bid documents. An address mapping is established between each scoring clause and rejection clause and the corresponding chapter, paragraph, and attachment in the bid documents, ultimately generating a structured comparison baseline.

[0020] The core of this step is to establish a one-to-one correspondence between bidding rules and bid response content through semantic matching. This accurately locates the response position in the bid document for each clause, forming a comparison benchmark that binds clauses and response content. This allows subsequent verification work to directly locate the corresponding paragraph without manual traversal, significantly improving the accuracy and efficiency of verification. The specific implementation method is as follows: Semantic association matching employs a bidirectional semantic retrieval mechanism. First, each clause in the integrated set of bidding terms is sequentially converted into a semantic vector. Simultaneously, all paragraphs and table cells in the structured text of the bid document are also converted into semantic vectors of the same dimension. Then, the cosine similarity between each bidding clause vector and all bid paragraph vectors is calculated. These are sorted from highest to lowest similarity, and the paragraph with the highest similarity (greater than or equal to 0.75) is selected as the corresponding response content for that clause. If the highest similarity value is lower than 0.75, it is determined that there is currently no corresponding bid response for that clause, and it is marked as an item to be verified. The 0.75 threshold is the optimal matching threshold verified by a large number of samples, which avoids mismatches of irrelevant content while ensuring that valid response content with semantically different expressions is not missed. After content matching is completed, a corresponding address mapping is established for each bidding clause. The mapping information includes the bid section to which the response content belongs (i.e., the information section, technical section, or business section), the corresponding page number, paragraph number, and chapter path. If the response content is tabular data, the worksheet name and cell range are added; if there are corresponding supporting documents, the attachment name and attachment location are added. Address mapping uses a unique path identifier to ensure that each clause can be quickly located to its specific position in the tender document through the mapping information, with a positioning accuracy down to the single paragraph level.

[0021] In the example, the tender clause "Bidders must provide a valid ISO9001 quality management system certification" is a mandatory clause for bid rejection. After semantic matching, its similarity to the content of paragraph 3 on page 12 of the tender document's qualification brochure is 0.89. The corresponding attachment is a scanned copy of the fourth certificate in the qualification certificate package. The system establishes an address mapping for this clause, labeling the brochure as the qualification brochure, page number 12, paragraph 3, and attachment as qualification attachment 4. After matching all clauses, a corresponding mapping was established for 72 scoring details and 18 bid rejection clauses. Among them, 12 clauses matched multiple response locations, and 3 clauses did not match any valid response content and were marked as suspected missing items. The integrated tender clause set, all address mapping relationships, and unmatched item marking information are integrated into a single structured comparison baseline. This baseline is the sole input benchmark for the subsequent two-layer verification model. All compliance checks and scoring calculations are based on this baseline to ensure consistent comparison standards and accurate positioning.

[0022] S202 uses deep learning training based on the enterprise's historical case library and the industry's public case library to transform the reasons for bid rejection and the points lost in the cases into standardized verification rules, generating a two-layer verification model that adapts to the characteristics of the enterprise and the general industry standards. Specifically, historical bid rejection cases and score deduction cases can be imported from the enterprise's local database. For each case, the reasons for rejection, the names of missing materials, the specific terms violated, and the points deducted can be marked to generate an enterprise-marked case dataset. The core of this step is to collect the company's past failures and point deductions in bidding projects. Through standardized annotation, these scattered historical cases are transformed into structured data that can be used for model training. This anchors the company's own high-frequency and error-prone risk points, providing a dedicated training data source for the company's characteristic layer of the two-layer verification model. The specific implementation method is as follows: The company's local database stores complete records of all its bidding projects over the past five years. This includes both failed bids that did not pass the qualification review and those that were ultimately rejected, as well as winning bids that resulted in point deductions during the evaluation process. During the import process, data cleaning was performed first, removing invalid cases with incomplete project information or missing documentation, and retaining valid cases with complete bidding terms, bid documents, evaluation results, and explanations of rejection or point deductions. In this example, a total of 326 valid historical cases were imported, including 49 rejected bids and 277 cases with point deductions, covering all of the company's main business areas, such as building construction, municipal utilities, and electromechanical installation.

[0023] The annotation process employs a unified four-dimensional annotation standard, annotating each valid case item by item. The first dimension is the reason for rejection, used to categorize the core causes of case failure, specifically divided into five categories: incompatibility of qualifications, missing supporting documents, violation of format specifications, failure to meet performance requirements, and abnormal pricing. Each case corresponds to a unique core reason label, facilitating subsequent model learning of violation patterns. The second dimension is the name of missing documents. For rejection and point deduction scenarios involving missing documents, the specific missing documents are clearly recorded, such as the project manager's registration certificate, social security payment certificates for the past three months, key pages of similar performance contracts, bid bond transfer vouchers, etc., accurately locating missing materials and providing a basis for refining subsequent verification rules. The third dimension is the specific clause violated, corresponding to the rejection or scoring clause number explicitly stated in the bidding documents, along with excerpts of the original clause text, establishing a correspondence between the violation and the bidding clause, allowing the model to learn which behavior corresponds to which clause violation. The fourth dimension is the deduction points. For cases of deduction, the specific deduction points and reasons are recorded. For example, 3 points are deducted for the lack of specific construction measures in the technical solution, and 2 points are deducted for the company's number of similar performances being less than 1. The consequences of deduction for violations are quantified.

[0024] After all annotations are completed, an annotation consistency check is performed. Two senior bidding specialists cross-check 20% of the randomly selected cases, and the annotation consistency rate must reach 0.9 or higher. Inconsistent annotations are corrected by the review team to ensure the accuracy and standardization of the annotation data. Finally, all annotated cases are stored in a unified format to generate an enterprise annotation case dataset. This dataset is closely aligned with the enterprise's bidding habits, data organization patterns, and frequently encountered errors, and forms the core data foundation for building the enterprise-specific verification layer.

[0025] We collected publicly available cases of rejected bids from the public service platform for construction bidding by web crawling, and used a text summarization algorithm to extract the key elements of each case, including the key points of the rejection decision, the description of the violation, and the legal clause number, to generate a dataset of publicly available industry cases. The core of this step is to supplement the company's own case coverage. By collecting publicly available cases of rejected bids from across the industry, standardized violation characteristics are extracted to expand the risk identification boundaries of the model, covering common industry risks that the company has not yet encountered, thus avoiding verification blind spots. The specific implementation method is as follows: The web crawler conducted targeted data collection from the bid rejection announcement sections of official bidding and tendering public service platforms nationwide and in various provinces and cities. The collection scope covered all publicly released bid rejection announcements for engineering projects in the fields of housing construction, municipal engineering, water conservancy, and transportation over the past three years. The crawler adopted an incremental collection mechanism, only crawling newly published announcements to avoid duplicate collection. It also automatically filtered out non-bid rejection information such as announcements of successful bidders and notices of failed bids, retaining only the official announcement text that clearly stated a bid rejection decision. During the collection process, the platform's data access specifications were strictly adhered to, and the request frequency was controlled to avoid interfering with the platform's normal operation. In this example, a total of 1320 original bid rejection announcements were collected, covering projects at different levels in 27 provinces and cities across the country, including bid rejection cases from different entities such as central state-owned enterprises, local state-owned enterprises, and private enterprises, ensuring sufficient industry representativeness of the sample.

[0026] The original bid rejection notices are unstructured government documents, lengthy and inconsistently formatted. Directly using them for model training would introduce a large amount of redundant information. Therefore, an extractive text summarization algorithm is used to extract key information from each notice. This algorithm can locate the core information sentences in the notice, eliminate procedural and clichéd redundant content, and accurately extract three types of key elements. The first type is the key points of the bid rejection decision, that is, the official conclusive statement on the bid rejection result, such as "Because the bidder's qualifications do not meet the requirements of the bidding documents, its bid is deemed invalid." This element clarifies the final conclusion and nature of the bid rejection. The second type is the description of the violation, that is, the specific facts of the behavior that led to the bid rejection, such as "The construction engineer certificate of the project manager provided by the bidder has expired and there is no valid renewal certificate." This element is the core content for the model to learn the violation behavior pattern, corresponding to the specific errors in the bid documents. The third type is the basis for the clause number, that is, the specific clause number of the bidding documents on which the bid rejection decision is based. Some notices also include the corresponding legal clause number. This element is used to establish a mapping relationship between the violation behavior and the basis for the decision, forming a unified dimension with the clause annotations in the enterprise case.

[0027] After element extraction, all cases were deduplicated and cleaned, removing invalid cases with incomplete elements or vague descriptions, ultimately yielding 1140 valid cases. All cases were stored according to a unified element structure, generating an industry-wide publicly available case dataset. This dataset covers various typical violation scenarios across the industry, including many risk types that companies themselves have not yet encountered. It effectively fills the coverage gaps in companies' historical cases and provides ample training data support for the industry-wide general rule layer of the two-layer verification model.

[0028] The enterprise-labeled case dataset and the industry-public case dataset are merged and input into the large language model for instruction fine-tuning training. The model learns the causal relationship between violation patterns and clauses in the cases and generates a set of verification rule parameters that integrate enterprise-specific pain points and industry-wide risks. The core of this step is to leverage the semantic understanding and knowledge induction capabilities of a large language model, and through a lightweight training method of fine-tuning instructions, allow the model to learn the violation judgment logic in bidding scenarios from two types of case data, establish the causal relationship between violations and bidding terms, and finally output a parameterized verification rule system. The specific implementation method is as follows: Instruction fine-tuning is a mainstream training method for optimizing large language models for specific domain tasks. Unlike full-parameter retraining, this method only adjusts a small portion of the model's parameters. While preserving the general semantic understanding capabilities of the large model, it allows the model to quickly adapt to the specific scenario of bidding and tendering verification. It has low training costs, fast convergence speed, and strong adaptability. Before training, the merged case dataset is converted into a unified instruction fine-tuning sample format. Each sample contains three parts: instruction, input, and output. The instruction part is fixed as "Based on the given bidding terms and tender response content, determine whether the tender content is compliant, explain the reasons for the violation and the corresponding risk level," which clarifies the model's task objective. The input part is the original text of the bidding terms corresponding to the case and the fragment of the violation content in the tender document, recreating the real verification scenario. The output part is the compliance judgment result, the reason for the violation, the corresponding clause number, and the risk level, which correspond one-to-one with the elements labeled in the case. In the example, the merged dataset generated a total of 1466 valid training samples, including 326 enterprise cases and 1140 industry cases. These were divided into training set, validation set and test set in a ratio of 8:1:1 to ensure the verifiability of the training results.

[0029] The training process employs a hierarchical learning rate strategy, setting the learning rate for the bottom-level parameters to 1e-5 and the top-level output layer to 2e-5. This ensures that the underlying general language capabilities are not compromised while allowing the top-level task adaptation layer to quickly learn domain knowledge. The total training epochs are set to three. After each epoch, the accuracy is calculated on the validation set. Training is stopped early if the validation set accuracy improves by less than 0.5% for two consecutive epochs to avoid overfitting. The core objective of the training is to enable the model to learn the causal relationship between violation patterns and clauses, rather than simply matching keywords. For example, the model should understand that "social security payment duration less than 6 months" constitutes a violation pattern of "unmet eligibility requirements," corresponding to a violation of the personnel eligibility clause in the tender document, rather than simply identifying the keyword "social security" and determining a violation. The model automatically summarizes the common characteristics of different types of violations, establishing a complete reasoning chain of "behavioral characteristics - clause requirements - risk consequences."

[0030] After training, the model's performance was validated using a test set. In the example, the accuracy rate for identifying invalid items reached 92.3%, and the accuracy rate for identifying deducted items reached 89.7%. Compared to the untuned general-purpose model, the accuracy rate improved by 31.5%, and the false positive rate decreased by 27.8%. The final output of the training, the set of verification rule parameters, includes the model's attention weight parameters, violation feature embedding vectors, and judgment thresholds for different risk levels. It is an iterative and optimizable parameterized rule system, rather than fixed hard-coded rules, which can support subsequent iterative updates and performance optimization.

[0031] The set of verification rule parameters is encapsulated into a two-layer verification rule library. The first layer is the enterprise-specific rule layer, which is used to capture the enterprise's historical error-prone points. The second layer is the industry-general rule layer, which is used to cover typical risk points in public cases. Finally, a two-layer verification model that adapts to enterprise characteristics and industry-general standards is generated.

[0032] The core of this step is to modularly encapsulate the trained parameterized rules according to a two-layer architecture of "enterprise-specific - industry-general", forming a verification model that has both customized accuracy and full-scenario coverage. The layered architecture also facilitates subsequent iterative updates, improving the model's maintenance efficiency. The specific implementation method is as follows: The encapsulation process first involves feature splitting of the verification rule parameter set. Based on the source of the training samples and the applicable scope of the rules, the parameters are divided into two independent modules. The first layer is the enterprise-specific rule layer, corresponding to the exclusive rule parameters trained from the enterprise's labeled case dataset. The core function of this layer is to accurately capture the enterprise's historical error-prone points. Since the training data for this layer all come from real bidding cases of the enterprise, it is highly sensitive to the types of errors that the enterprise frequently makes. For example, high-frequency error points that the enterprise has repeatedly made in the past, such as "similar performance contracts lacking signature pages of both parties", "inconsistent page numbers and table of contents annotations in technical solutions", and "inconsistent amounts in the bid letter", will have stricter judgment thresholds set in this layer, triggering identification first, ensuring that the most common errors made by the enterprise can be accurately detected and avoiding repeated mistakes. In the example, the enterprise-specific rule layer covers 76 types of enterprise-specific high-frequency risk points, and the recall rate for identifying historical error-prone items of the enterprise reaches over 96%, which can effectively avoid repetitive errors.

[0033] The second layer is the industry-wide general rules layer, corresponding to the general rule parameters trained from publicly available industry case datasets. The core function of this layer is to cover typical bid rejection and penalty risks across the entire industry, filling the gaps in coverage of enterprise cases. This layer includes judgment rules for various common violation scenarios, such as "bid security deposit received later than the deadline required by the tender documents," "joint bid without a joint bidding agreement," "expired qualification certificate without renewal proof," and "legal representative's signature replaced with a seal without authorization," etc. These are common risk points across the industry, and this layer can effectively identify them regardless of whether the enterprise has encountered similar problems in the past. The judgment thresholds of the industry-wide general rules layer balance accuracy and coverage, avoiding both excessive leniency leading to missed detections and excessive strictness resulting in numerous false positives. In the example, the industry-wide general rules layer covers 214 industry-wide risk points, covering more than 90% of common violation types in mainstream bidding scenarios.

[0034] The two-layer rule base employs a serial reasoning mechanism. The content to be verified first undergoes an initial judgment by the enterprise-specific rule layer, followed by a second round of full scanning by the industry-general rule layer. The judgment results from both layers are automatically deduplicated and merged to avoid repeatedly labeling the same issue. The advantage of this layered architecture is that subsequent iterations can update the rule parameters of a single layer without affecting the judgment performance of the other layer. For example, if a company adds a new type of error-prone scenario, only the enterprise-specific rule layer needs to be updated, without retraining the entire model. The final encapsulated two-layer verification model possesses both enterprise-customized accuracy and industry-level coverage. Compared to traditional single-layer rule verification models, it reduces the false negative rate by 34% and the false positive rate by 28%, providing reliable model support for subsequent compliance checks.

[0035] S203, input the structured comparison baseline into the two-layer verification model for compliance check, mark the corresponding relationship of clauses, document location and risk level of the detected issues, and generate a graded verification report; Specifically, all tender clauses and tender document mappings can be extracted from the structured comparison baseline, and then input into the compliance check engine of the two-layer verification model one by one to generate the original compliance status vector of each clause. The core of this step is to complete the standardized extraction and model adaptation of the verification input data. Independent bidding terms and corresponding bid response content are extracted from the structured comparison baseline and uniformly transformed into vectorized data that can be directly identified and calculated by the two-layer verification model. This provides a standardized, complete and omission-free input basis for subsequent two-layer rule-based progressive judgment, avoiding judgment bias and missed detection caused by inconsistent data formats.

[0036] The structured comparison baseline forms the complete data foundation built through prior document parsing and mapping. It contains all valid tender clauses after deduplication and integration, along with address mappings and content snapshots of the corresponding chapters, paragraphs, and attachments in the tender documents for each clause. The extraction process traverses by clause type, first extracting all mandatory clauses for rejected bids, then extracting all scoring clauses. During extraction, supplementary information such as clause number, chapter, and clause type tag are retained to ensure the completeness and traceability of each clause's attributes. In this example, a total of 128 valid clauses were extracted from the tender documents, including 32 mandatory clauses for rejected bids, 41 commercial scoring clauses, and 55 technical scoring clauses. All clauses are arranged in the original tender document order, and the corresponding mapped tender content covers all relevant chapters and attachments in the three volumes: the credit book, the technical book, and the commercial book.

[0037] The compliance check engine is the pre-input processing module of the two-layer verification model, responsible for transforming the natural language requirements of the terms and conditions and the content of the bid response into standardized numerical vectors. The engine first decomposes each bidding clause into keywords and rule elements, extracting core judgment dimensions such as qualification requirements, quantity requirements, validity period requirements, format requirements, and signature and seal requirements. Then, it matches these dimensions with the response content in the bid documents, quantifying and scoring it from four dimensions: semantic matching degree, data completeness, format compliance, and time validity, ultimately generating a multi-dimensional raw compliance status vector. The general expression for the original compliance status vector is S=[s_type,s_match,s_complete,s_format], where s_type is the clause type identifier, with a value of 1 for rejected clauses and a value of 0 for scoring clauses; s_match is the semantic matching score, ranging from 0 to 1, with higher values ​​indicating a higher semantic fit between the bid response content and the clause requirements; s_complete is the data completeness score, ranging from 0 to 1, with higher values ​​indicating more complete supporting data; and s_format is the format compliance score, ranging from 0 to 1, with higher values ​​indicating that the data format better meets the bidding requirements.

[0038] Taking a disqualifying clause, "The bid security must be provided with a bank guarantee with a validity period of not less than 90 days," as an example, the corresponding bid response content is a scanned copy of the bank guarantee on page 12 of the business brochure. After the engine deconstructs and judges the clause, it obtains a semantic matching degree of 0.85, a data completeness score of 0.92, and a format compliance score of 0.78. Combined with the clause type identifier, the final generated original compliance status vector is [1, 0.85, 0.92, 0.78]. After all 128 clauses have been vectorized, a set of original compliance status vectors corresponding one-to-one with each clause is formed. The entire process adopts a batch automatic processing mechanism, with processing errors controlled within 0.1%, ensuring no clause omissions and no vector mismatches, providing accurate input for subsequent two-level judgment.

[0039] The two-layer verification model sequentially calls the enterprise-specific rule layer and the industry-general rule layer to make dual judgments on the original compliance status vector, identify the deviation between the content of the bid response and the clause requirements, and generate a set of deviation features. The core of this step is to comprehensively determine the compliance status of the clauses through a progressive verification of a two-layer rule system, from two dimensions: enterprise-specific historical risks and industry-wide compliance risks. This not only covers the enterprise's frequently occurring error-prone pain points, but also fills in the industry-wide common risk points that the enterprise has not experienced. It accurately identifies various explicit non-compliance items and implicit deviations, generates a complete set of deviation features, and minimizes the probability of missed detections.

[0040] The two-layer verification model adopts a progressive judgment logic of "first specific, then general." The first layer is the enterprise-specific rule layer, and the second layer is the industry-general rule layer. The two layers of rules are independent yet complementary. The judgment logic of the enterprise-specific rule layer is entirely based on training with historical enterprise cases, focusing on capturing the points of deduction and rejection that the enterprise has repeatedly encountered in past bidding. It is adapted to the enterprise's own compilation habits and common oversights. For example, if an enterprise has repeatedly incurred deductions for "insufficient social security payment period for project managers" or "performance contracts without acceptance certificates," the enterprise-specific rule layer will increase the judgment weight for such clauses, amplifying the sensitivity to identify subtle deviations. The judgment logic of the industry-general rule layer is based on training with publicly available rejection cases across the entire industry, covering compliance risks and typical reasons for rejection that are common in the construction bidding field. For example, issues such as "the bid documents are not stamped with the required seal across the seam," "the power of attorney for the legal representative is missing a signature or date," and "the qualification certificate has expired" are frequently encountered problems across the industry, filling in risk blind spots that enterprises have not encountered.

[0041] During the judgment process, the model first inputs the original compliance status vector into the enterprise-specific rule layer to complete the first round of deviation identification and generate preliminary deviation features. Then, the same vector is input into the industry-general rule layer to complete the second round of deviation identification and generate supplementary deviation features. Finally, the deviation features obtained from the two rounds of identification are deduplicated and merged to avoid the same deviation being counted repeatedly. Each deviation feature contains four core attributes: deviation type, deviation degree, judgment basis, and rule source. Deviation types are divided into four categories: missing data, qualification mismatch, format violation, and time limit exceeding. The deviation degree ranges from 0 to 1, with higher values ​​indicating more severe deviations. The rule source indicates whether the deviation was detected at the enterprise level or the industry level, facilitating the location of the rule level during subsequent iterations and optimizations.

[0042] Taking the scoring clause requiring "at least three similar project performances in the past three years" as an example, the original compliance status vector is [0, 0.72, 0.65, 0.88]. The enterprise characteristic rule layer first makes a judgment, matching three cases in the company's history where points were deducted for "performances without supporting acceptance certificates." Therefore, the completeness of the performance supporting materials is carefully checked, identifying that two of the five submitted performances lack corresponding acceptance reports, generating two deviation features of missing data. Subsequently, the industry general rule layer makes a supplementary judgment, matching the industry's general rule that "the performance time range is determined based on the contract signing date." It is found that among the two performances lacking acceptance reports, one of the contracts was signed more than three years after the bidding date, making it an invalid performance, generating an additional deviation feature of timeliness exceeding the limit. Ultimately, this clause yields a total of three valid deviation features, all of which are included in the deviation feature set. After completing the dual assessment of all 128 clauses, a total of 42 valid deviation features were identified. Compared with single-layer rule assessment, the detection rate increased by 37%, effectively covering enterprise-specific risks and industry-wide risks, and avoiding major missed detections.

[0043] For each deviation item in the deviation feature set, perform multi-dimensional annotation processing, and attach the original text of the corresponding bidding clause, the hit page number in the bid document, and the name of the responsible person or document involved in the clause to generate an annotated deviation record. The core of this step is to supplement the deviation characteristics obtained from purely technical judgment with business information. Through multi-dimensional standardized labeling, each deviation has a clear judgment basis, accurate document location, and clear responsibility attribution, eliminating the information gap between technical judgment results and actual business rectification, and ensuring that subsequent rectification work can be implemented, traced, and assigned, without the need for repeated manual cross-checking of bidding documents and tender documents.

[0044] The multi-dimensional annotation comprises three core dimensions: clause basis annotation, document location annotation, and responsible entity annotation. All annotation information is automatically matched and extracted from the previously generated structured data, eliminating the need for manual input and achieving an accuracy rate of over 99%. The first dimension is the original text of the corresponding bidding clause. The complete original text of the clause is retrieved directly from the structured comparison baseline's set of bidding clauses and appended to the corresponding deviation item, clearly indicating the bidding requirements violated by the deviation. Users can understand the basis for judgment without consulting the original bidding documents, avoiding misunderstandings of the rules. The second dimension is the hit page number in the bid document. The specific location of the corresponding bid content is extracted from the address mapping data of the structured comparison baseline, accurate to the volume, page number, and paragraph number, such as "page 23, paragraph 2 of the credit book," along with a snapshot of the corresponding content. Users can directly locate the specific position of the deviation, quickly verify the problem, and make corrections. The third dimension is the name of the responsible personnel or documents involved in the clause. It is automatically matched based on the company's internal pre-set bidding preparation division of labor mapping table. For example, personnel qualification clauses correspond to credit preparation specialists, construction plan clauses correspond to technical preparation specialists, and quotation clauses correspond to commercial preparation specialists. At the same time, the names of the documents that need to be rectified are marked, which makes it convenient to directly assign tasks according to responsibilities during the rectification stage and improve rectification efficiency.

[0045] Taking the deviation item "Insufficient social security payment period for project manager" as an example, the complete record after multi-dimensional annotation includes three parts of supplementary information. The clause basis section indicates that the corresponding original tender text is "The project manager must provide proof of social security payment for more than 6 consecutive months, and the paying unit must be completely consistent with the bidding unit; otherwise, no points will be awarded for this item." The document location section is annotated as "Page 23 of the credit register, pages 1 to 5 of the scanned copy of the social security payment certificate," along with a snapshot showing "payment period as 5 months." The responsible entity section is annotated as "Corresponding responsible personnel: Credit Compilation Specialist, Corresponding rectification materials: Project Manager's social security payment certificate." All annotation information is bound to the deviation characteristics one by one, and each deviation record generates a unique deviation number. The numbering rule is "Environmental type code - Clause type code - Serial number," for example, the deviation number for the 3rd item in the rejected bid category is "FB-003," which facilitates subsequent hierarchical management and traceability verification.

[0046] After all 42 deviation features were annotated in multiple dimensions, a unified set of annotated deviation records was formed. All records were formatted in a standardized manner and contained complete information, including both the technical judgment attributes of the deviation itself and all the auxiliary information required for business rectification. This completely connected the model judgment results with actual business operations, providing a standardized data foundation for subsequent risk classification and report generation.

[0047] Based on the severity of risk of each deviation item in the marked deviation record, they are classified into hard rejection items, data missing deduction items, or non-critical defects, and the output is organized according to category to finally generate a graded verification report.

[0048] The core of this step is to categorize all deviations according to their risk levels, clearly distinguish between fatal bid rejection risks, routine point deduction risks, and minor flaw risks, and organize the output in order of priority from high to low to form a hierarchical verification report with clear levels and highlighting key points. This helps users prioritize high-risk issues, allocate rectification efforts reasonably, and avoid direct bid rejection due to the omission of fatal risks.

[0049] The risk grading criteria perfectly match the actual risk levels of the bidding process. The three categories have clear boundaries and no ambiguous overlap. The first category, mandatory rejection items, represents the highest risk level. These deviations directly trigger mandatory rejection clauses in the bidding documents. If such issues exist and are not rectified, the bid will be rejected, rendering it completely invalid. The criteria are that the deviation item corresponds to a mandatory rejection clause in the bidding documents, and the deviation level is 0.7 or higher. In other words, the non-compliance directly meets the rejection conditions. Common mandatory rejection items include insufficient bid security amount or validity period, missing signature and seal on the legal representative's authorization letter, failure to meet basic requirements for the bidding entity's qualifications, and key personnel's qualifications not meeting mandatory requirements. The second category is for items with missing data, which is of medium risk. These won't directly trigger bid rejection, but will cause deviations in the objective scoring criteria. This type of issue affects the final bid score but won't directly lead to bid elimination. The judgment rule is that the deviation item corresponds to a scoring clause in the tender, and the deviation prevents the bidder from obtaining the full score for that item. Common items with missing data include insufficient performance records, missing award certificates, missing annual inspection pages for personnel qualification certificates, and non-compliant format of supporting materials. The third category is for non-critical defects, which is of low risk. These are minor issues that do not affect bid compliance or cause any deductions. They are only slight deficiencies in expression or layout, and do not affect the validity or score of the bid. Common non-critical defects include minor typos that do not affect semantic understanding, non-standard punctuation, slight inconsistencies in layout format, and inconsistent font sizes in non-critical positions.

[0050] The classification process is completed automatically by the system. Based on the clause type label and deviation severity value of each deviation record, the system automatically matches the corresponding risk level and performs a secondary verification of the classification results to avoid misclassification of high-risk deviations as low-risk. In the example, after the 42 deviations were classified, there were 3 hard rejection items: insufficient validity period of the bid bond, unsigned power of attorney of the legal representative, and no seal across the seam of the bid documents. All of these are serious issues that directly trigger the rejection clauses. There were 18 items with missing data, including 5 items with missing acceptance certificates for performance, 7 items with missing annual inspection pages for personnel qualification certificates, and 6 items with missing scanned copies of original award certificates, with a total deduction of approximately 12 points. There were 21 non-critical defects, mainly minor issues such as formatting and wording, which do not affect the score or compliance.

[0051] After the grading is completed, the report content is organized according to the priority order of "hard rejection items first, missing data deduction items in the middle, and non-critical flaws last." The report begins with a risk overview module, which counts the number of deviations in the three categories and their corresponding risk impacts, clearly marking the core message: "The three rejection risks must be rectified first, and the tender documents cannot be submitted until the rectification is completed." Subsequently, each category lists the deviation records one by one, including complete information such as deviation description, clause basis, document location, and rectification suggestions. The report can be exported to a common document format for easy distribution and collaborative rectification within the team. The final generated graded verification report has a clear structure and highlights the key points, which can help users quickly identify critical risks and systematically sort out all deductions and flaws, providing a complete guide for subsequent rectification and optimization of tender documents.

[0052] S204. Based on the feedback data generated after manually annotating the graded verification report, the model is iteratively trained to generate an updated two-layer verification model. Specifically, a manual annotation port for graded verification reports can be opened to users to receive their judgment results for each detected issue, including confirmation of validity, marking as false detection, or supplementary entry of missed detection items, and generate manual annotation feedback data. The core of this step is to establish a two-way interactive channel between human professional experience and algorithmic models. This transforms the industry experience of evaluation experts and bidding specialists into structured feedback data that can be learned by the model, correcting the model's judgment biases and supplementing the model with hidden violation scenarios not covered by the model. This provides accurate training material for subsequent incremental iterations. The specific implementation method is as follows: The manual annotation portal is embedded in the display page of the tiered verification report, corresponding one-to-one with each detected deviation item. Users can complete the annotation operation directly without switching systems. The portal offers three standardized annotation options for each issue: Confirm Valid, Mark False Detection, and Supplement Missed Detection Items. Each option has a clear business meaning. Confirm Valid means the user acknowledges the issue detected by the model; the deviation does indeed violate the bidding terms and is a real risk point. Mark False Detection means the model's judgment is incorrect; the detected issue does not actually exist, usually due to misunderstandings of the terms, lack of synchronization of supplementary announcements, or overly strict formatting standards leading to false alarms. Supplement Missed Detection Items represent violations that the model failed to identify but actually exist; users need to manually enter the issue description, corresponding bidding terms, bid document location, risk level, and other information to fill the model's blind spots.

[0053] To ensure the standardization and usability of feedback data, the system provides structured input templates for supplementary entries of missed items. It mandates the inclusion of five core fields: violation clause number, bid document page number, specific problem description, risk level, and rectification suggestions. This prevents ambiguous free-text descriptions from hindering the model's ability to extract effective features. The system also incorporates built-in annotation verification logic, providing real-time alerts for duplicate annotations within the same report, missing fields, and mismatches between risk levels and problem severity, guiding users to complete standardized annotations. In the example, a tiered verification report for a construction project bid document identified 32 issues. After reviewing each issue, the user confirmed 22 as valid and marked 6 format-related issues as false positives. These false positives occurred because the model incorrectly identified "attached" supporting documents in the bidding documents as missing, even though the documents were already in the appendix section of the business booklet. Additionally, the user added 4 missed items: insufficient social security payment period for the project manager, contract amount for similar projects not meeting requirements, lack of official seal on qualification certificates, and incorrect signature placement in the bid letter. All four types of issues fall under implicit judgment rules not covered by the model. After all annotation operations are completed, the system automatically associates and integrates the annotation results with the original deviation items and corresponding clause information. Each record contains eight fields: unique issue ID, original text of the bidding clause, location of the tender document, original judgment result, type of manual annotation, annotation description, annotation personnel, and annotation time. Finally, complete manual annotation feedback data is formed, providing standardized input for subsequent feature extraction and incremental training.

[0054] Extract the features of issues marked as false positives and the features of missed positives added by users from the manually labeled feedback data, construct the negative sample feature set and the newly added positive sample feature set, and generate incremental training data for the model. The core of this step is to perform feature processing on unstructured human feedback text, transforming false positives and false negatives into vector features that the model can recognize. These features correspond to two types of training data: negative samples and newly added positive samples. This provides accurate feature basis for adjusting model parameters, ensuring that iterative updates can accurately hit the model's judgment blind spots and bias points. The specific implementation method is as follows: Feature extraction employs a combination of semantic vector extraction and rule-based feature matching, simultaneously extracting multi-dimensional features for each feedback data point to ensure that the features fully represent the core attributes of the problem. The extracted feature dimensions include four categories: First, clause semantic features, which transform the text content of the corresponding bidding clauses into 768-dimensional semantic vectors, representing the semantic connotation and requirement type of the clauses, distinguishing different categories such as qualification requirements, personnel requirements, performance requirements, and format requirements; second, text matching features, extracting features such as similarity, keyword overlap, and numerical deviation rate between the bid response content and clause requirements, quantifying the degree of matching between the response content and the requirements; third, rule triggering features, recording which verification rule in the model was triggered by the problem, as well as the rule's original judgment threshold, locating the specific judgment node of the model; and fourth, risk level features, corresponding to the risk level marked by the user, distinguishing between three levels: rejection level, deduction level, and defect level.

[0055] For issues marked as false positives, their corresponding feature combinations constitute negative samples, labeled "non-violation," indicating that the content under this feature pattern does not fall under any violation category, and the model's previous judgment was incorrect. The features of all false positives are aggregated to form a negative sample feature set, used to subsequently adjust the model's judgment boundaries, reducing the probability of over-judgment and minimizing false alarms. For user-added missed detections, their corresponding feature combinations constitute new positive samples, labeled "violation," indicating that this feature pattern belongs to a real violation scenario that the model previously failed to identify. The features of all missed detections are aggregated to form a new positive sample feature set, used to supplement the model's violation pattern library, covering previously unlearned scenarios and improving recall.

[0056] To avoid impacting the performance of the original model during incremental training, the number of negative samples and newly added positive samples needs to be controlled within a reasonable range. The number of incremental samples in a single iteration should not exceed 5% of the total number of samples in the original training set, ensuring small-step iterations and smooth optimization. In the example, the manual annotation feedback contained 6 false positives and 4 false negatives, resulting in the extraction of 6 sets of negative sample feature vectors and 4 sets of newly added positive sample feature vectors. The negative samples mainly focused on over-judgment of format-related clauses, characterized by high keyword matching but semantic inclusion relationships, with the original model classifying "can be provided" as a "must be provided" requirement. The newly added positive samples mainly covered three types of implicit rules: personnel social security details, performance amount judgment, and signature location requirements. The feature characteristics were that the clause requirements were implicit in the text description without explicit keyword identification. The two feature sets were paired and encapsulated with corresponding labels to form the incremental training data for the model. The total data volume was 10 sets of samples, accounting for approximately 2.1% of the total number of samples in the original training set, which meets the small-sample update requirement of incremental learning and can correct model biases without interfering with the original knowledge system.

[0057] The incremental training data of the model is merged with the original training dataset. The weight parameters of the two-layer verification model are updated using an incremental learning algorithm. The decision boundary of the rule layer is adjusted to eliminate false detections and cover missed detection scenarios, generating an intermediate model with optimized parameters. The core of this step is to replace full retraining with a lightweight incremental learning algorithm. While preserving the original knowledge system of the model and avoiding catastrophic forgetting, the model weights and decision thresholds are fine-tuned based on newly added feedback samples to accurately correct false positives and false negatives. At the same time, it significantly reduces the time cost and computing power consumption of iterative training. The specific implementation method is as follows: Incremental learning algorithms are machine learning methods specifically designed for continuous iterative optimization of models. Unlike traditional full-scale training, which requires relearning all data, incremental learning only fine-tunes local parameters of the model based on new samples. Simultaneously, it retains the ability to recognize historical samples through a knowledge replay mechanism. This allows for rapid absorption of new knowledge without compromising original performance, adapting to the iterative needs of bidding verification scenarios where rules are continuously updated and enterprise requirements are constantly changing. Before training, the dataset is first mixed by merging the incremental training data with a small portion of the original training dataset. This extracted portion of the original samples is called replay samples. Its purpose is to allow the model to continuously review historical knowledge during training, preventing the forgetting of old judgment rules after learning new samples. The proportion of replay samples is set to 5% of the original training set, using stratified sampling to ensure coverage of all typical scenarios under both enterprise-specific and industry-wide rules, guaranteeing the representativeness of the replay samples.

[0058] For the two-layer architecture of the dual-layer verification model, a layered update strategy is adopted, with different parameter update magnitudes and logics at different levels to ensure that the core positioning of each layer does not shift. For the first layer, the enterprise-specific rules layer, since false positives and false negatives are mostly related to the enterprise's own document preparation habits, common expressions, and historical error-prone points, the parameter update magnitude for this layer is larger, with a learning rate set to 0.002. The focus is on adjusting the judgment thresholds for error-prone scenarios and optimizing the semantic matching logic for enterprise-specific expressions, making the model more closely reflect the actual characteristics of enterprise bid document preparation. For the second layer, the industry-general rules layer, this layer is responsible for covering industry-wide standards and risks, requiring higher stability. Therefore, only parameters are supplemented for general violation scenarios involved in false negatives, with a learning rate set to 0.001. The core parameters of the original general rules are not adjusted to avoid compromising the accuracy of the general rules' judgments.

[0059] During parameter updates, gradient descent is used to iteratively optimize weights, continuously reducing the error between the model's judgment and the labeled values. Simultaneously, parameter variation constraints are set, ensuring that the core parameters do not change by more than 10% in a single iteration to prevent significant parameter fluctuations from degrading the overall model stability. Training terminates when the accuracy of incremental samples reaches 98% or higher, and the accuracy of replay samples decreases by no more than 2%. At this point, the model has absorbed new knowledge without forgetting old knowledge, and training converges. In this example, the incremental training reached convergence after 86 iterations. Eleven judgment thresholds in the enterprise characteristic rule layer were fine-tuned. Specifically, the judgment threshold for format-related clauses increased from 0.72 to 0.79, effectively reducing the probability of over-judgment; the judgment threshold for personnel performance-related clauses decreased from 0.81 to 0.75, improving the sensitivity to implicit requirements. Three new violation pattern features were added to the industry-general rule layer without altering the original rule parameters. After the update, an intermediate model with optimized parameters was formed. According to preliminary calculations, the false detection rate of the intermediate model on the feedback samples decreased from 18.8% to 0, and the false negative rate decreased from 12.5% ​​to 0. At the same time, the recognition accuracy of historical samples remained at 96.7%, without any obvious performance degradation, achieving the expected optimization goal.

[0060] The intermediate model with optimized parameters is subjected to regression testing to verify that the original verification capability has not been weakened. After passing the test, it is deployed as a new verification rule layer, and finally an updated two-layer verification model is generated.

[0061] The core of this step is to verify, through a standardized regression testing process, that the original core verification capabilities have not degraded after model iteration, avoiding the catastrophic forgetting problem caused by incremental updates, and ensuring that the new model retains all the original reliable functions while fixing known issues. After successful verification, formal deployment is completed, forming a usable updated version of the model. The specific implementation method is as follows: Regression testing is a mandatory verification step before model iteration and release. The test uses a pre-built historical validation set as benchmark data. This validation set is independent of the training set and the current incremental data, covering all confirmed typical violation scenarios, normal response scenarios, and boundary cases, comprehensively measuring the model's overall verification capabilities. The historical validation set contains 500 labeled samples, including 82 rejected bid samples covering various hard rejection scenarios such as ineligibility, missing signatures, and expired validity; 168 deducted bid samples covering common deductions such as insufficient performance data, ineligible personnel qualifications, and non-standard formatting; and 250 normal bid samples covering various compliant bid responses, used to test the model's false positive rate.

[0062] During testing, the intermediate model and the original model were run simultaneously in the same testing environment, with identical historical validation set data as input. The output verification results were then compared, and two types of core performance metrics were compared. The first metric is recall, which measures the model's ability to identify real violations, representing the degree of missed detections. The second metric is precision, which measures the accuracy of the model's detection, representing the degree of false positives. The third metric is overall accuracy, which measures the model's comprehensive judgment ability. The acceptance criteria are: an overall accuracy decrease of no more than 2%, a recall rate no lower than the original version, a precision decrease of no more than 3%, and all high-risk invalid samples being correctly identified without omission. If the test results meet the acceptance criteria, the intermediate model is considered to have passed the regression test, and its original verification ability has not been weakened. If not, the proportion of replay samples and the learning rate during incremental training need to be adjusted, the intermediate model retrained, and the test repeated until the acceptance requirements are met.

[0063] After passing the test, the model is deployed using a hot update method, replacing only the rule layer parameters in the two-layer verification model. The overall model architecture, inference engine, and interface logic remain unchanged, allowing for updates without system downtime and ensuring continuous system availability. Upon deployment, the system automatically generates a new model version number, synchronously updates the version description, records the optimization content, parameter changes, test results, and version effective time of this iteration, and retains the complete model file of the previous version as a rollback backup. If anomalies occur during the operation of the new version, a quick switch back to a historical stable version can be made. In the example, the regression test results show that the overall accuracy of the intermediate model is 96.9%, slightly improved from the original version's 96.6%; the recall rate is 93.8%, an improvement of 1.5 percentage points from the original version's 92.3%; the precision rate is 97.1%, a decrease of 0.2 percentage points from the original version's 97.3%, far below the 3% threshold requirement; all 82 rejected samples were correctly identified, with an accuracy rate of 100%, fully meeting the acceptance criteria. After deployment, the updated two-layer verification model with version number v2.2.0 was officially generated. This model retains all the original verification capabilities and fixes the false detection and missed detection issues reported in this report. The overall performance has been steadily improved and it can be directly used for subsequent bid document verification work.

[0064] S205, based on the updated dual-layer verification model, decompose the objective scoring details in the bidding documents and match them with the tender documents, calculate the objective scores, deduction points and missing data items, and generate a scoring optimization report.

[0065] Specifically, all scoring details with objective judgment criteria can be extracted from the structured comparison baseline in the bidding documents. These objective judgment criteria include certificate type, quantity requirements, validity period, and format specifications, generating a set of objective scoring details. The core of this step is to select quantifiable, objective scoring items from the complete bidding terms system, eliminating scoring items that rely on the subjective judgment of the judges, such as technical solutions and service commitments. This provides clear judgment objects for subsequent item-by-item matching and verification, ensuring that the scoring calculation process is based on evidence and free from subjective bias. The specific implementation method is as follows: Objective scoring criteria are those items with clearly defined thresholds in the bidding document's scoring method. Their core characteristic is that the scoring result depends entirely on whether the supporting materials meet the preset conditions, without requiring judges to make judgments based on experience. In contrast, subjective scoring items include aspects such as the rationality of the solution, technological advancement, and service response speed, which rely on subjective judgment. The extraction process relies on semantically annotated bidding clauses in a structured comparison baseline. It uses rule matching combined with semantic classification to complete the screening. First, objective judgment expressions such as "must provide," "possess," "validity period," "quantity," and "official seal" are matched by keyword recognition to initially screen candidate items. Then, a semantic model is used to eliminate items containing subjective evaluation terms such as "excellent," "good," "reasonable," and "relatively strong," ultimately yielding all objective scoring criteria.

[0066] The four objective judgment criteria each have a clear purpose and function. Certificate type limits the attributes and levels of supporting materials, distinguishing between different categories such as system certification, qualification certificates, personnel qualification certificates, and performance certificates, while specifying level requirements, such as Level 1 qualification or senior professional title. Quantity requirements set the minimum or maximum number of materials required to meet the conditions, such as no fewer than three similar performance records or no fewer than two certified personnel, and also include a threshold for the value of a single document, such as a single contract amount of no less than 10 million yuan. Validity period limits the time validity of the materials, including the validity period of the certificate itself, the time range of the performance record, and the period of social security contributions for personnel. The core judgment node is the bid closing date, and all materials must remain valid at that date. Format specifications limit the presentation format and compliance requirements of supporting materials, including document format, signature requirements, content completeness, and attachment requirements, such as requiring scanned copies with an official seal, screenshots of official website queries, and inclusion of key pages of the contract.

[0067] In the example, the structured comparison baseline for a municipal bidding project contains 42 scoring clauses. After screening, 27 objective scoring details were extracted, including 11 clauses related to creditworthiness and qualifications, 7 clauses related to project performance, 6 clauses related to personnel configuration, and 3 clauses related to format compliance. Objective scores account for 65% of the total score. Each detail was broken down into dimensions. For example, "2 points for possessing a valid ISO45001 Occupational Health and Safety Management System certification" is broken down into: ISO45001 Occupational Health and Safety Management System certification, a requirement of 1 certificate, validity period up to the bid closing date, and a scanned copy of the certificate bearing the bidder's official seal. All the decomposed details were sorted from highest to lowest score, uniformly labeled with a unique number and the corresponding clause text, ultimately forming a set of objective scoring details, which serves as the benchmark for subsequent material matching and score calculation.

[0068] The set of objective scoring rules is input into the updated two-layer verification model. Each rule is matched against the corresponding supporting documents in the tender documents to determine whether the documents are complete, valid, and compliant with the format requirements, and a sequence of matching judgment results is generated. The core of this step is to rely on the iteratively optimized two-layer verification model to complete a one-by-one comparison and verification of the bidding requirements and bidding materials. It quantitatively determines the degree to which each supporting document meets the requirements from three dimensions: completeness, validity, and compliance. This replaces the repetitive work of manual item-by-item verification, while avoiding omissions and oversights in manual verification. The specific implementation method is as follows: The updated dual-layer verification model has optimized the judgment boundary through multiple rounds of iterative training. The enterprise-specific rule layer prioritizes matching frequently occurring error-prone scenarios in the enterprise's past bidding, such as frequently missing certificate appendices, missing signature pages in performance contracts, and non-continuous social security payments for employees—these are the enterprise's unique pain points. During verification, these high-risk points will be identified in detail. The industry-general rule layer covers typical invalid situations in publicly available cases across the industry, such as expired certificates, performance time exceeding the limit, illegible official seals, and inverted content in scanned documents—these are common risks. The two layers of rules are verified in parallel, taking into account both personalized and general risks.

[0069] The matching and verification process uses the order of items in the objective scoring criteria set as a benchmark, calling the address mapping in the structured comparison baseline item by item to locate the corresponding chapter, page number, and attachment in the tender document, and then conducting a judgment in three dimensions. The first dimension is whether the materials are complete, verifying whether all supporting materials required for the scoring item are provided, including the main certificate, attachments, supporting materials, query screenshots, and all other required documents. The absence of any one item is considered as incomplete. The second dimension is whether the materials are valid, verifying the core elements of validity, including whether the certificate validity period covers the bid deadline, whether the performance time and amount meet the threshold requirements, whether the personnel qualifications correspond to the profession, and whether social security is paid normally. Failure of any core element is considered as invalid. The third dimension is whether the materials meet the format requirements, verifying the formal elements such as the signature, clarity, document format, and content completeness of the materials. For example, whether the bidder's official seal is affixed, whether the scanned copy is clear and legible, and whether the required key pages are included. Materials that do not meet the format requirements cannot be used as valid proof.

[0070] In the example, regarding the scoring rule of "2 points for each similar municipal project with a single contract value of over 15 million RMB in the past 5 years, up to a maximum of 6 points," the model located the 4 performance items submitted in the commercial brochure of the tender documents. After verifying each item, the following judgments were made: The first performance item, with a contract value of 18 million RMB, was signed in 2022, and the acceptance documents were complete and in a compliant format, thus it was deemed a valid performance item; the second performance item, with a contract value of 16.5 million RMB, was signed in October 2019, exceeding the 5-year time limit until the tender deadline of September 2025, thus it was deemed an invalid performance item; the third performance item, with a contract value of 21 million RMB, only provided the first and last pages of the contract, lacking key pages that reflected the contract content and amount, thus it was deemed incomplete; the fourth performance item, with a contract value of 14.8 million RMB, did not reach the 15 million RMB threshold and was deemed an invalid performance item. After all the detailed rules are matched, the judgment conclusion and specific judgment basis of each rule are output in the order of the entries, forming a sequence of matching judgment results. Each record in the sequence contains four types of information: rule number, judgment level, explanation of non-compliance, and corresponding file location. The judgment level is divided into three categories: fully compliant, partially compliant, and completely non-compliant.

[0071] Based on the matching judgment result sequence, calculate the score and deduction of each objective scoring rule, summarize the total objective score, the company's current valid score, the deduction value and the reason for each deduction point, and extract the zero score items due to missing data to generate a score statistics matrix. The core of this step is to quantify the objective score based on the matching judgment results. It presents a complete picture of the objective score of the current bid document from multiple dimensions, including total score, actual score, deduction items, and zero-score items. This allows users to intuitively grasp the score distribution and key areas of point deduction, providing a quantitative basis for subsequent rectification and optimization. The specific implementation method is as follows: The scoring follows the scoring rules in the tender documents. For items scored item by item, each valid requirement met earns a corresponding single point, with the cumulative score not exceeding the full score for that item. For qualification-related binary judgment items, full marks are awarded for complete fulfillment of the requirements, while zero marks are awarded for any core element not met. For partially compliant cases, the score is calculated proportionally based on the fulfilled requirements, and formatting defects are penalized according to preset standards. The entire calculation process is linked to the original tender terms, and all deductions have clear rules and no room for discretion.

[0072] The four core statistical indicators each have a clear meaning. The objective full score refers to the sum of the scores of all objective scoring details, representing the highest possible score for the objective part of the project, and serving as a benchmark for measuring the scoring level. The current effective score of the enterprise refers to the total objective score that can actually be obtained in the current state of the tender documents after matching and judgment, reflecting the actual scoring ability of the existing materials. Each deduction point refers to all scoring items that did not receive a full score. Each point corresponds to a clear deduction value and reason for deduction, accurately locating the specific item and cause of the loss of points. The zero-score item for missing materials refers to the item that has no supporting materials provided at all or has missing core materials, resulting in a score of zero for this item. This type of item is the highest priority optimization target for rectification.

[0073] In the example, the project has 27 objective scoring rules, with a total objective score of 65 points. The company's current effective score is 50.5 points, resulting in an overall objective score rate of 77.7%. There are 9 deduction points in total, deducting a total of 14.5 points. These include 6 points deducted for 3 invalid performance items, 3 points deducted for 1 completely missing special qualification certificate, 2 points deducted for 2 personnel certificates with insufficient validity, 1.5 points deducted for the lack of a certified public accountant's signature on the financial audit report, and 1 point and 1 point deducted for 2 incomplete performance documents. Among all deductions, 2 items resulted in zero points due to completely missing documents: the special qualification certificate and the environmental system certification, totaling 5 points deducted. These points can be fully recovered by providing supplementary materials. All items are sorted by scoring category, and the following are listed sequentially: rule number, preset score, actual score, deducted score, reason for deduction, and corresponding document location, forming a complete scoring statistics matrix. The data within the matrix corresponds one-to-one, allowing for direct traceability of the source of each point's acquisition and deduction.

[0074] Based on the missing data and zero-score items in the scoring statistics matrix, an automatic list of supplementary data for rectification is generated, and all deduction points and their basis are compiled. Finally, a scoring optimization report containing objective score statistics and rectification suggestions is generated.

[0075] The core of this step is to transform the scoring statistics into actionable rectification guidelines, clearly distinguish between optimizable and non-optimizable items, and provide targeted corrective suggestions and expected benefits. This helps users recover as many points as possible before the bidding deadline and improve their overall competitiveness. The specific implementation method is as follows: The supplementary information rectification checklist is generated for three categories of rectifyable issues: missing information, incomplete information, and non-standard format. Each rectification item clearly indicates the corresponding scoring item, the current deduction value, the details of the materials to be supplemented or corrected, the format requirements and precautions, and also indicates the expected points that can be recovered after rectification, making it convenient for users to allocate their efforts according to priority. For items with zero points due to complete absence, the checklist will clearly list the complete list of materials that need to be supplemented, such as "Supplement a scanned copy of the ISO14001 Environmental Management System Certification Certificate, which must be stamped with the bidder's official seal and ensure that the certificate's validity period covers the bid deadline. This item is worth 2 points, and the full amount can be recovered after supplementation." For items with deductions due to incomplete information, the specific pages or supporting materials that are missing are clearly stated, such as "Supplement the contract amount page and project acceptance certificate for item 3, which is worth 2 points and the full amount can be recovered after supplementation." For format-related issues, specific directions for correction are given, such as "The financial audit report needs to be supplemented with a scanned copy of the certified public accountant's signature page. Correction can recover 1.5 points."

[0076] The compilation of all deduction points will be synchronized with the corresponding bidding terms and conditions. Each deduction item will be accompanied by the original text of the bidding terms and condition number to ensure that the deductions are justified and facilitate user verification. Furthermore, all deduction items will be categorized into two types based on their feasibility for rectification: Rectifiable items are those that can be resolved by supplementing materials or correcting formatting, such as missing documents, formatting errors, or incomplete attachments; Unrectifiable items are hard defects that cannot be remedied within the bidding period, such as insufficient years of performance experience, substandard qualification levels, or insufficient personnel. These items will be marked with risk warnings, suggesting that users assess their overall competitiveness and adjust their bidding strategy if necessary.

[0077] The final scoring optimization report comprises four core modules. The first module provides an overall overview of the objective score, listing the total maximum objective score, current valid score, overall score rate, recoverable score, and irrecoverable score, clearly presenting the current score level and room for optimization. The second module details the deduction points, sorted by deduction amount from highest to lowest, listing the deduction value, reason for deduction, corresponding clause, and rectification attribute for each point. The third module is a rectification and optimization list, arranged by priority from highest to lowest, with each item including rectification requirements and expected benefits. The fourth module provides overall optimization suggestions, giving priority to rectification, precautions, and risk warnings. In the example, the rectification list for this project lists 8 items that can be optimized. Completing all of them can recover 11 points, and the objective score is expected to increase to 61.5 points, with the score rate increasing to 94.6%. Two irrecoverable items are also marked, resulting in a total deduction of 3.5 points. It is recommended to consider the project's weighting when evaluating bidding strategies.

[0078] Another embodiment of the present invention provides an intelligent verification and scoring optimization system for tender documents based on iterative learning with a dual-case library, see [link to relevant documentation]. Figure 3 The system may include: The parsing module 301 is used to import and parse the bidding documents and tender documents, integrate all the scoring clauses and hard clauses for rejection in the bidding documents, establish a mapping relationship with the content of the tender documents, and generate a structured comparison baseline. Training module 302 is used for deep learning training based on the enterprise's historical case library and the industry's public case library. It transforms the reasons for bid rejection and the points lost in the cases into standardized verification rules, and generates a two-layer verification model that adapts to the characteristics of the enterprise and the general industry standards. The verification module 303 is used to input the structured comparison baseline into the two-layer verification model for compliance checks, mark the corresponding relationship of clauses, document location and risk level of the detected issues, and generate a graded verification report; Optimization module 304 is used to iteratively train the model based on feedback data generated after manual annotation of the graded verification report, and generate an updated two-layer verification model. The scoring module 305 is used to decompose the objective scoring details in the bidding documents and match them with the bidding documents based on the updated two-layer verification model, to calculate the objective score, deduction points and missing data items, and to generate a scoring optimization report.

[0079] This invention also provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above method embodiments when running.

[0080] This invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0081] Specifically, the aforementioned electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the aforementioned processor, and the input / output device is connected to the aforementioned processor.

[0082] The above description, based on the embodiments shown in the figures, details the structure, features, and effects of the present invention. The above description is only a preferred embodiment of the present invention, but the present invention is not limited to the scope of implementation shown in the figures. Any changes made in accordance with the concept of the present invention, or equivalent embodiments modified to have equivalent changes, that do not exceed the spirit covered by the specification and figures, should be within the protection scope of the present invention.

Claims

1. A method for intelligent verification and scoring optimization of tender documents based on iterative learning with a dual-case library, characterized in that, The method includes: Import and parse the tender documents and bid documents, integrate all the scoring clauses and hard clauses for rejection in the tender documents, establish a mapping relationship with the content of the bid documents, and generate a structured comparison baseline; Deep learning training is conducted based on the enterprise's historical case library and the industry's publicly available case library to transform the reasons for bid rejection and the points lost in the cases into standardized verification rules, generating a two-layer verification model that adapts to the characteristics of the enterprise and the general industry standards. The structured comparison baseline is input into the two-layer verification model for compliance checks. The detected issues are marked with the corresponding clauses, document locations, and risk levels, and a graded verification report is generated. Based on the feedback data generated after manually annotating the graded verification report, the model is iteratively trained to generate an updated two-layer verification model. Based on the updated two-layer verification model, the objective scoring details in the bidding documents are decomposed and matched with the bidding documents. The objective scores, deduction points and missing data items are statistically analyzed, and a scoring optimization report is generated.

2. The method according to claim 1, characterized in that, The process of importing and parsing the tender documents and bid documents, integrating all scoring clauses and mandatory rejection clauses in the tender documents, establishing a mapping relationship with the content of the bid documents, and generating a structured comparison baseline includes: The system receives user-uploaded tender documents and bid documents through a batch import interface. The tender documents include the main tender announcement, supplementary announcement, Q&A documents, and specification documents. The bid documents include a credit book, a technical book, and a business book, generating a multi-source heterogeneous file set. Each file in the multi-source heterogeneous file set is processed for format identification and classification. A multimodal document parsing algorithm is used to extract the text layer of PDF format, the paragraph structure of Word format, and the table data of Excel format, respectively, to generate the structured text content of each file. The structured text content of all bidding documents is merged and deduplicated, old clauses that have been covered by supplementary announcements are identified and removed, the latest valid scoring details and hard clauses for rejection are retained, and a set of integrated bidding clauses is generated. The integrated set of bidding terms is semantically matched with the structured text content of the bid documents. An address mapping is established between each scoring clause and rejection clause and the corresponding chapter, paragraph, and attachment in the bid documents, ultimately generating a structured comparison baseline.

3. The method according to claim 2, characterized in that, The method involves deep learning training based on a database of historical enterprise cases and a database of publicly available industry cases. This training transforms the reasons for bid rejection and points lost in the cases into standardized verification rules, generating a two-layer verification model that adapts to both enterprise characteristics and general industry standards. This model includes: Import historical bid rejection and score deduction cases from the enterprise's local database, and annotate each case with the reason for rejection, the name of the missing document, the specific clauses violated and the points deducted, to generate an enterprise annotated case dataset; We collected publicly available cases of rejected bids from the public service platform for construction bidding by web crawling, and used a text summarization algorithm to extract the key elements of each case, including the key points of the rejection decision, the description of the violation, and the legal clause number, to generate a dataset of publicly available industry cases. The enterprise-labeled case dataset and the industry-public case dataset are merged and input into the large language model for instruction fine-tuning training. The model learns the causal relationship between violation patterns and clauses in the cases and generates a set of verification rule parameters that integrate enterprise-specific pain points and industry-wide risks. The set of verification rule parameters is encapsulated into a two-layer verification rule library. The first layer is the enterprise-specific rule layer, which is used to capture the enterprise's historical error-prone points. The second layer is the industry-general rule layer, which is used to cover typical risk points in public cases. Finally, a two-layer verification model that adapts to enterprise characteristics and industry-general standards is generated.

4. The method according to claim 3, characterized in that, The structured comparison baseline is input into the two-layer verification model for compliance checks. Detected issues are marked with clause correspondences, document locations, and risk levels, generating a tiered verification report, including: Extract all tender clauses and tender document mappings from the structured comparison baseline, input them one by one into the compliance check engine of the two-layer verification model, and generate the original compliance status vector for each clause; The two-layer verification model sequentially calls the enterprise-specific rule layer and the industry-general rule layer to make dual judgments on the original compliance status vector, identify the deviation between the content of the bid response and the clause requirements, and generate a set of deviation features. For each deviation item in the deviation feature set, perform multi-dimensional annotation processing, and attach the original text of the corresponding bidding clause, the hit page number in the bid document, and the name of the responsible person or document involved in the clause to generate an annotated deviation record. Based on the severity of risk of each deviation item in the marked deviation record, they are classified into hard rejection items, data missing deduction items, or non-critical defects, and the output is organized according to category to finally generate a graded verification report.

5. The method according to claim 4, characterized in that, The step of iteratively training the model based on feedback data generated after manually annotating the graded verification report to generate an updated two-layer verification model includes: Open the manual annotation port of the graded verification report to users, receive the user's judgment results for each detected problem, including confirming the validity, marking it as a false detection, or supplementing the missing items, and generate manual annotation feedback data; Extract the features of issues marked as false positives and the features of missed positives added by users from the manually labeled feedback data, construct the negative sample feature set and the newly added positive sample feature set, and generate incremental training data for the model. The incremental training data of the model is merged with the original training dataset. The weight parameters of the two-layer verification model are updated using an incremental learning algorithm. The decision boundary of the rule layer is adjusted to eliminate false detections and cover missed detection scenarios, generating an intermediate model with optimized parameters. The intermediate model with optimized parameters is subjected to regression testing to verify that the original verification capability has not been weakened. After passing the test, it is deployed as a new verification rule layer, and finally an updated two-layer verification model is generated.

6. The method according to claim 5, characterized in that, The updated two-layer verification model is used to decompose the objective scoring details in the bidding documents and match them with the bid document information. The objective scores, deduction points, and missing data items are statistically analyzed to generate a scoring optimization report, including: Extract all scoring details with objective judgment criteria from the structured comparison baseline in the bidding documents. The objective judgment criteria include certificate type, quantity requirements, validity period range and format specifications, and generate a set of objective scoring details. The set of objective scoring rules is input into the updated two-layer verification model. Each rule is matched against the corresponding supporting documents in the tender documents to determine whether the documents are complete, valid, and compliant with the format requirements, and a sequence of matching judgment results is generated. Based on the matching judgment result sequence, calculate the score and deduction of each objective scoring rule, summarize the total objective score, the company's current valid score, the deduction value and the reason for each deduction point, and extract the zero score items due to missing data to generate a score statistics matrix. Based on the missing data and zero-score items in the scoring statistics matrix, an automatic list of supplementary data for rectification is generated, and all deduction points and their basis are compiled. Finally, a scoring optimization report containing objective score statistics and rectification suggestions is generated.

7. A tender document intelligent verification and scoring optimization system based on dual-case library iterative learning, characterized in that, The system includes: The parsing module is used to import and parse the bidding documents and tender documents, integrate all the scoring clauses and hard clauses for rejection in the bidding documents, establish a mapping relationship with the content of the tender documents, and generate a structured comparison baseline. The training module is used for deep learning training based on the enterprise's historical case library and the industry's public case library. It transforms the reasons for bid rejection and the points lost in the cases into standardized verification rules, and generates a two-layer verification model that adapts to the enterprise's characteristics and the industry's general standards. The verification module is used to input the structured comparison baseline into the two-layer verification model for compliance checks, mark the corresponding relationship of clauses, document location and risk level of the detected issues, and generate a graded verification report; The optimization module is used to iteratively train the model based on the feedback data generated after manually annotating the graded verification report, and generate an updated two-layer verification model. The scoring module is used to decompose the objective scoring details in the bidding documents and match them with the bidding documents based on the updated two-layer verification model, to calculate the objective scores, deduction points and missing data items, and to generate a scoring optimization report.

8. The system according to claim 7, characterized in that, The parsing module is specifically used for: The system receives user-uploaded tender documents and bid documents through a batch import interface. The tender documents include the main tender announcement, supplementary announcement, Q&A documents, and specification documents. The bid documents include a credit book, a technical book, and a business book, generating a multi-source heterogeneous file set. Each file in the multi-source heterogeneous file set is processed for format identification and classification. A multimodal document parsing algorithm is used to extract the text layer of PDF format, the paragraph structure of Word format, and the table data of Excel format, respectively, to generate the structured text content of each file. The structured text content of all bidding documents is merged and deduplicated, old clauses that have been covered by supplementary announcements are identified and removed, the latest valid scoring details and hard clauses for rejection are retained, and a set of integrated bidding clauses is generated. The integrated set of bidding terms is semantically matched with the structured text content of the bid documents. An address mapping is established between each scoring clause and rejection clause and the corresponding chapter, paragraph, and attachment in the bid documents, ultimately generating a structured comparison baseline.

9. A storage medium, characterized in that, The storage medium stores a computer program, wherein the computer program is configured to execute the method of any one of claims 1-6 when it is run.

10. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method of any one of claims 1-6.

Images