An intelligent bidding document auditing method and system based on artificial intelligence

Abstract

This invention relates to the field of bidding document processing technology, and discloses an intelligent bidding document review method and system based on artificial intelligence. The invention includes: acquiring and parsing bidding documents and bids to be reviewed; extracting review elements from the bidding documents and constructing a structured model of the review elements; calculating a rule generation deterministic index for each review element based on the model; filtering review elements that can automatically generate rules based on the index; dynamically constructing a review rule model; using the review rule model to perform rule-driven matching and judgment on the bids, generating structured review result data, and outputting a review report. This invention can adaptively generate review rules according to different bidding documents, improving automation while ensuring review accuracy, and is applicable to various bidding scenarios.

Description

Technical Field

[0001] This invention relates to the field of bidding document processing technology, and in particular to an intelligent bidding document review method and system based on artificial intelligence. Background Technology

[0002] In the bidding process, bid document review is a crucial step in ensuring bid quality, reducing the risk of rejection, and improving review efficiency. As the types of bidding documents, project scale, and review rules become increasingly complex, relying solely on manual item-by-item verification, while still flexible and experience-based, often requires significant time and manpower when handling large volumes of documents, complex parameter constraints, and multi-level scoring rules. Therefore, the industry has gradually introduced technologies such as document parsing, rule engines, natural language processing, and text matching to assist in the review of bid documents' completeness, compliance, parameter response, and scoring items. These technical solutions typically enable text extraction, chapter identification, element extraction, rule comparison, and result output from bidding documents and tender documents, playing a positive role in improving review efficiency and reducing repetitive work, and have become an important development direction for the digital processing of bidding documents.

[0003] However, in existing technologies, many automated review solutions still primarily rely on pre-set rule bases or fixed matching methods based on keywords and templates to construct review logic. While this approach works well for tender documents with relatively standardized structures and uniform expression, it often requires manual pre-configuration or adjustment of review rules when dealing with differences in expression habits, clause levels, chapter organization, and the presentation of scoring details between different projects. The rule generation process remains heavily dependent on human experience. Especially when tender documents contain multiple heterogeneous review elements such as mandatory constraints, conditional constraints, numerical comparison constraints, and scoring details, there is still room for optimization in how to automatically determine which review elements are suitable for direct rule generation and which are more suitable for manual confirmation, based on the clarity of expression, semantic ambiguity, and chapter structure characteristics of the current tender document, thereby forming a dynamic review rule model for the current project. Summary of the Invention

[0004] The technical problem to be solved by this invention is that the existing technology uses a preset rule base or a fixed matching method based on keywords and templates to construct the review logic, which requires configuration and adjustment in actual use. To address this, we propose an intelligent tender review method and system based on artificial intelligence.

[0005] To achieve the above objectives, this application adopts the following technical solution: an intelligent tender document review method based on artificial intelligence, comprising the following steps: Step S1: acquiring and parsing the tender documents and the tender documents to be reviewed, performing document parsing, text extraction, and structure recognition; Step S2: extracting review elements from the tender documents, identifying at least mandatory material elements, technical parameter elements, commercial clause elements, and scoring item elements, and extracting element content, element type, and constraint attributes for each review element; Step S3: converting the review elements extracted in Step S2 into a unified data structured model of review elements, wherein the model includes element identifiers, element types, constraint conditions, and scoring score information; Step S4: based on the structured model of review elements, targeting the current... The tender documents are dynamically generated using a non-preset review rule model. For each review element, a deterministic index is generated based on the element's constraint strength, textual ambiguity, and the information entropy of its section, according to a preset nonlinear function calculation rule. The magnitude of this index determines whether to automatically generate review rules or proceed to manual confirmation. The review rule model includes at least one of integrity rules, compliance rules, parameter comparison rules, and scoring rules. Step S5: Based on the review rule model, the tender documents are matched and judged using rule-driven methods, and the matching results for each review element are output. Step S6: Structured review result data is generated, including matching results, a list of non-compliant items, and scoring results. Step S7: An review report and visualization results are generated based on the structured review result data.

[0006] Preferably, the specific methods for extracting review elements from the tender documents in step S2 include: identifying constraint words in the tender documents through natural language processing to locate constraint attributes; identifying various review elements based on chapter titles and paragraph structures; and further extracting scoring criteria, score weights, and scoring details for scoring item elements.

[0007] Preferably, the structured model of the audit elements in step S3 includes at least the following fields: element identifier, element type, constraint text, constraint attribute quantification value, position pointer of the element in the original tender document, and score weight for the scoring item element; wherein the constraint attribute quantification value is assigned according to the predefined assignment rules based on the type and intensity of the constraint words.

[0008] Preferably, in step S4, when calculating the deterministic index, the strength of the constraint words is quantified based on the types of constraint words and negative words appearing in the review elements; the text ambiguity is calculated based on the standard deviation of the word vector of the sentence containing the element and normalized to the [0,1] interval; the chapter information entropy is calculated based on the distribution probability of the part-of-speech category within the chapter according to the Shannon entropy formula; and the maximum value of the information entropy of each chapter in the entire tender document is used as the normalization benchmark; the above three parameters are fused according to the preset nonlinear combination rules to obtain the deterministic index.

[0009] Preferably, the specific method for determining whether to automatically generate rules or transfer to manual confirmation based on the value of the certainty index generated by the rules in step S4 is as follows: a certainty threshold is preset. When the certainty index is greater than or equal to the threshold, it is determined that the audit element has sufficient rule generation reliability, and it is automatically converted into the corresponding category of audit rules and added to the audit rule model. When the certainty index is less than the threshold, the audit element is marked as an item to be manually confirmed and pushed to the manual review interface. After the reviewer confirms or corrects it, a rule is generated, or it is determined to be an invalid element. The audit rule model is a project-specific rule model dynamically generated for the current tender document, and its rule conditions and rule actions are all derived from the audit elements extracted in step S2.

[0010] Preferably, in step S4, different categories of audit rules are automatically generated based on the element type of the audit element. Specifically, these include: if the element type is a required field, an integrity rule is generated to check whether the corresponding chapter or document exists in the tender document; if the element type is a technical requirement, a compliance rule is generated to semantically match the tender document text with the tender requirements; if the element type contains numerical constraints, a parameter comparison rule is generated to compare the numerical parameters and constraint values ​​in the tender document; if the element type is a scoring item, a scoring rule is generated to automatically calculate the score according to the scoring details; and if the element type is a commercial clause, a commercial response rule is generated to check the tender document's response to each commercial clause.

[0011] Preferably, step S5 involves rule-driven matching and judgment of the tender documents based on the review rule model, specifically including: for completeness rules, locating the corresponding chapters or attachments in the tender documents and checking their existence and non-emptiness of content; for compliance rules, extracting the corresponding paragraph text in the tender documents, matching them with the bidding requirements using semantic similarity calculation methods, and judging the degree of compliance; for parameter comparison rules, extracting numerical parameters from the tender document text and unifying the units, and then performing greater than, less than, and equal to comparison operations with the constraint values; for scoring rules, matching the tender response content with the descriptions of each score level in the scoring details to determine the score for that item.

[0012] Preferably, the structured audit result data generated in step S6 includes the following fields: audit element identifier, element type, matching status, description of reasons for non-compliance, suggested modifications, risk level identifier, and obtained score; wherein, the risk level identifier is divided into high risk, medium risk, and low risk according to the importance of the audit element and the severity of non-compliance.

[0013] Preferably, the method further includes manual review and confirmation and final audit conclusion generation steps: after generating the audit report, a manual review interface is provided for auditors to view and modify the audit results item by item; after all audit items have been manually confirmed, the system automatically generates the final audit conclusion based on the matching status and risk level of each audit item. The final audit conclusion includes "pass", "conditionally pass" or "fail"; the system saves the manual review modification records and the final audit conclusion to the audit log database.

[0014] This invention proposes another technical solution: an intelligent tender document review system based on artificial intelligence, comprising the following modules: a document parsing module: used to acquire and parse tender documents and tenders to be reviewed, performing text extraction and structure recognition; a review element extraction module: used to extract review elements from tender documents, identifying various review elements and their constraint attributes; a structured model construction module: used to convert the extracted review elements into a structured model of review elements with a unified data structure; a rule generation module: used to dynamically generate a review rule model based on the structured model of review elements, including calculating the rule generation determinism index of each review element, and determining whether to automatically generate rules or transfer them to manual confirmation based on the index; a matching and judgment module: used to perform rule-driven matching and judgment of tenders based on the review rule model; a result output module: used to generate structured review result data and review reports; and a manual review interface module: used to receive elements to be confirmed whose rule generation determinism index is lower than a threshold, and provide a manual confirmation or correction interface.

[0015] The technical effects and advantages of this invention are as follows: By performing structured parsing of tender documents and automatically extracting review elements, this invention can dynamically generate review rule models based on the specific content of the current tender documents, avoiding the insufficient adaptability problem caused by relying on preset rule bases in traditional methods. By introducing a rule generation determinism index, the system can determine the generability of different review elements, enabling it to automatically generate review rules when constraints are clear and semantics are clear, while reverting to manual confirmation when expressions are ambiguous or uncertain, thus achieving a balance between automation efficiency and review accuracy. At the same time, by converting review elements into a unified structured model and combining it with integrity rules, compliance rules, parameter comparison rules, and scoring rules, this invention achieves multi-dimensional automatic review of tender documents, improving the consistency and standardization of the review process. Attached Figure Description

[0016] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts:

[0017] Figure 1 This is an overall flowchart of the present invention; Figure 2This is a flowchart of the review process for this invention; Figure 3 This is a timing diagram of module interactions in this invention. Detailed Implementation

[0018] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0019] Example 1: Refer to Figures 1-2 As shown, the present invention provides a technical solution: an intelligent tender document review method based on artificial intelligence. This method automatically extracts review elements and constructs a structured model by performing structured parsing of the tender document, and then dynamically generates a review rule model corresponding to the current tender document. Finally, the tender document is automatically reviewed based on the rule model. The method includes steps S1 to S7, which are described in detail below.

[0020] Step S1. Obtain and parse the tender documents and bids to be reviewed: The system receives the tender documents and bids to be reviewed uploaded by the user. Tender documents and bids are typically in PDF, Word, or plain text format. The system calls a document parsing engine to process the documents as follows: Format parsing: Select the corresponding parser based on the file extension. For PDF format, use pdfplumber or a similar library to extract text and table content. For Word format, use python-docx or a similar library to extract paragraphs, tables, and embedded objects. Text extraction: The parsed text content is structured and stored at the granularity of page, paragraph, line, and table cell, removing interfering information such as headers, footers, and watermarks. Structure recognition: By analyzing the document's table of contents, heading levels (such as "Chapter 1", "1.1", "1.", etc.), font size, bolding, and other style information, the system automatically constructs the document's chapter tree, recording the start and end positions and level depth of each chapter. The output of this step is the structured text data of the tender documents and bids, which will be used in subsequent steps.

[0021] Step S2. Extracting Review Elements from the Tender Documents: The goal of this step is to identify the various elements that need to be reviewed in the tender documents. The system uses natural language processing technology to analyze the structured text of the tender documents obtained in Step S1 and identify at least the following four types of review elements:

[0022] Required materials elements: Documents or supporting materials that the tender documents explicitly require bidders to provide, such as "letter of tender", "legal representative's identity certificate", "letter of authorization", "business license", "safety production license", "financial statements of recent years", "similar project performance certificate", "after-sales service commitment letter", etc.; locate these elements by identifying constraint words such as "required", "should be attached", "must be included" and keywords such as "qualification requirements" and "qualification certificate" in the chapter titles.

[0023] Technical parameter elements: Technical specifications of products, equipment or services specified in the tender documents, such as "rated power not less than XXkW", "accuracy error ≤0.5%", "response time less than XX milliseconds", "material is XX stainless steel", etc. The system extracts this type of element by recognizing the combination of numerical values ​​and units (such as "XXmm", "XXkg"), comparison operators ("≥", "≤", ">", "<", "=") and constraint words.

[0024] Commercial terms and conditions: Requirements in the tender document regarding pricing, delivery, payment, warranty, and acceptance, such as "the bid price shall not exceed the budgeted price," "the delivery period is within XX days after the contract is signed," "the warranty period is not less than XX years," and "the payment method is 90% payment upon acceptance of the goods." The system extracts these terms by recognizing keywords such as "price," "delivery," "payment," "warranty," and "acceptance" and their context.

[0025] Scoring elements: The scoring items listed in the bidding document's evaluation method, such as "Technical solution (30 points)", "Enterprise performance (20 points)", "Personnel configuration (15 points)", "Bid price (40 points)", etc.; Each scoring element also includes detailed scoring rules, such as "Complete and scientifically reasonable technical solution gets 20-30 points; basically reasonable gets 10-19 points; obvious defects get 0-9 points"; The system extracts the scoring elements and their weights (scores) by recognizing keywords such as "scoring", "score", "points" and structures such as tables and numbered lists.

[0026] For each extracted audit element, the system further extracts the following information: Element content: The original text fragment describing the element in the tender document. Element type: One of the four categories mentioned above. Constraint attributes: Including constraint words (such as "must", "should", "appropriate", "must not", "not less than", etc.) and their negative forms, as well as the upper and lower limits of numerical constraints.

[0027] The specific implementation of this step includes: keyword matching based on a predefined constraint word dictionary; using dependency parsing to identify the modification relationship between constraint words and target objects; and using a table parser to extract the row and column correspondence for parameter requirements in tabular form.

[0028] Step S3. Construct a structured model of audit elements: Convert each audit element extracted in Step S2 into a unified data structure to form a structured model of audit elements. This model is stored in the form of a structured data table or a JSON object. Each audit element corresponds to one record, containing the following fields: Element Identifier: A unique ID assigned to the element, such as "REQ_001" or "REQ_002", used for subsequent traceability and reference. Element Type: One of the following: required materials, technical parameters, business terms, or scoring items. Constraint Text: The original constraint description text of the element in the tender document. Constraint Attribute Quantification Value: Quantified according to the predefined assignment rules based on the type and strength of the constraint words; for example, "must", "strictly prohibited", and "must not" are assigned a value of 3; "should" and "shall" are assigned a value of 2; "preferably" and "recommended" are assigned a value of 1; "may" and "may" are assigned a value of 0.5; no constraint words are assigned a value of 0. If there is a negative word (such as "not" or "not") before the constraint word, the quantification value is multiplied by -1. For numerical constraints, comparison operators and thresholds are also recorded. Weight: For scoring items, record the corresponding score; for non-scoring items, the weight field is set to 0 or empty. Original Location Pointer: Records the chapter number, page number, paragraph index, or table coordinates of the element in the original tender document, used for quick location when generating the review report. Through this structured model, the scattered and unstructured review requirements in the tender document are converted into a computer-readable and processable form, laying the data foundation for subsequent dynamic rule generation.

[0029] Step S4. Dynamically generate the audit rule model: This step is based on the structured audit element model generated in step S3. It dynamically generates a set of non-preset, project-specific audit rule models for the current tender document. The rule model is not retrieved from the pre-stored rule base, but is generated in real time entirely from the audit elements of the current tender document. Specifically, it includes the following sub-steps.

[0030] S41. Calculate the rule generation certainty index for each audit element: Not all audit elements extracted from the tender documents are suitable for automatic rule generation; there may be elements in the tender documents that are vaguely described, have unclear constraints, or lack sufficient contextual information. If rules are forcibly generated automatically, it is easy to lead to misjudgment. Therefore, this invention uses a rule generation certainty index to quantitatively evaluate the reliability of automatically generated rules for each audit element.

[0031] The formula for calculating the certainty index D is as follows: The physical meaning and acquisition method of each parameter are as follows: Constraint strength C: taken from the constraint attribute quantification value of the audit element in step S3; the value range of C depends on the constraint quantification rule, usually an integer or half-integer between [-3,3]; the larger the value of C, the more mandatory and clear the requirements of the tender document for the element are, and the higher the reliability of the automatically generated rule; a negative value of C indicates a negative constraint (such as "must not"), which also has a clear judgment condition.

[0032] Textual Ambiguity E: Represents the degree of semantic ambiguity in the text paragraph containing the review element. The value of E ranges from [0,1], with a higher value indicating greater ambiguity. In this implementation, E is preferentially calculated using the following method: The sentence containing the element is segmented into words, and a pre-trained Chinese word embedding model (such as the shallow output of Word2Vec or BERT) is used to obtain the vector representation of each word. The standard deviation of these vectors is calculated, and then the standard deviation is normalized to the [0,1] interval as E. When the word vectors in the sentence are concentrated, the standard deviation is small, and E approaches 0, indicating clear semantics. When the word vectors are dispersed, the standard deviation is large, and E approaches 1, indicating semantic ambiguity. As an optional implementation, a statistical method can also be used: The number of conjunctions (such as "or", "and", "etc."), modifiers (such as "generally", "usually", "approximately"), and the density of commas and semicolons in the sentence are counted, and these statistics are weighted and normalized to obtain E. Both methods can achieve the purpose of this invention.

[0033] Chapter information entropy H: Indicates the degree of content disorder in the chapter containing the review element. First, the text of the chapter is segmented and tagged with parts of speech, and the frequency of each part of speech (noun, verb, adjective, adverb, conjunction, etc.) is counted to obtain the probability distribution of the part-of-speech categories. Then, according to the Shannon entropy formula... Calculate the information entropy of this chapter; the larger the H value, the more even the part-of-speech distribution, the more mixed the content, and the less concentrated the basis for rule generation; the smaller the H value, the more singular the content and the clearer the theme of this chapter. This is the maximum value of the information entropy of all chapters in the entire tender document, used for normalization.

[0034] Index Term This factor reflects the impact of chapter information entropy on the determinism index. When H equals (That is, when the chapter is most chaotic) the exponent term approaches 0; when H is much smaller than When (i.e., when the content of this chapter is concentrated), the exponential term approaches The exponential function guarantees that the degree of chapter disorder has a nonlinear suppression effect on the deterministic exponent, and the degree of suppression intensifies as the degree of disorder increases.

[0035] Overall product: Multiplying this by the exponential factor yields the final deterministic exponent D; This reflects the combined effect of constraint strength and textual ambiguity: the stronger the constraint and the lower the ambiguity, the larger the ratio; conversely, the higher the ambiguity, the larger the denominator, and the lower the ratio. Multiplying this by the chapter entropy factor further reduces the certainty index of elements in chapters with chaotic content.

[0036] This formula allows the system to provide an objective quantitative score for the reliability of rule generation for each audit element.

[0037] S42. Triage based on deterministic index: The system presets a deterministic threshold. This threshold can be configured by the administrator according to the audit accuracy requirements, with a typical value range of 0.6 to 0.8.

[0038] Automatic rule generation: When If the system determines that the audit element has sufficient reliability for rule generation, it automatically converts it into an audit rule of the corresponding category based on the element type and adds the generated rule to the audit rule model of the current tender document. See step S43 for details on rule generation.

[0039] Transfer to manual confirmation: When If the system determines that the reliability of the rule generation for the audit element is insufficient, it will not automatically generate a rule. Instead, it will mark the element as "awaiting manual confirmation" and push it to the manual review interface. On the manual review interface, the system displays the element's original text, extracted constraint attributes, calculated D value, and its sub-parameters (C, E, H) for review. Reviewers can choose to: confirm the automatically generated rule (if there are preliminary results); manually modify the rule conditions and then confirm; or determine the element as invalid and not generate any rules.

[0040] S43. Generate different categories of review rules based on element type: For review elements that require automatic rule generation, the system generates corresponding categories of rules according to their element type. This invention supports, but is not limited to, the following five types of rules. Each type of rule dynamically constructs rule conditions and actions by calling the corresponding rule generation logic: Completeness rule: Generated when the element type is "Required Material". This rule is used to check whether there is a corresponding chapter, document, or supporting material in the tender document and to verify that its content is not empty. The condition part of the rule is "whether there is a chapter or attachment pointing to this element in the tender document", and the action part is "if it exists, mark it as satisfied; otherwise, mark it as not satisfied". Compliance rule: Generated when the element type is "Technical Parameter" or general requirement. This rule is used to extract the corresponding paragraph text in the tender document, perform semantic matching with the requirement text in the tender document, and determine the degree of satisfaction. The condition part of the rule includes a semantic similarity threshold, and the action part is "if the similarity is ≥ the threshold, it is satisfied; otherwise, it is not satisfied or partially satisfied". Semantic matching can use a vector space model based on cosine similarity or similarity calculation based on a pre-trained language model, but is not limited to these. Parameter Comparison Rule: Generated when the element type is "Technical Parameter" and contains explicit numerical constraints (such as "≥", "≤", ">", "<", "="). This rule is used to extract numerical parameters from the tender document text, standardize the units, and compare them with the constraint values. The condition part of the rule consists of comparison operators and thresholds, and the action part determines whether the constraint is met or not based on the comparison result. Scoring Rule: Generated when the element type is "Scoring Item Element". This rule semantically matches the tender response content with the descriptions of each score level according to the scoring details in the tender document, and automatically calculates the score for that item. The condition part of the rule is the correspondence between the level descriptions and scores in the scoring details, and the action part is the score corresponding to the level with the highest matching degree. Business Response Rule: Generated when the element type is "Business Terms". This rule is used to check whether the tender responds to each business term in the tender document and identifies deviations. The condition part of the rule is "whether the tender contains a response statement for the corresponding business terms", and the action part is a three-category judgment: "complete response, partial response, and no response". All generated rules are stored in a structured format. Each rule contains fields such as rule ID, rule type, rule conditions, rule actions, and associated audit element IDs. The collection of these rules constitutes the audit rule model for the current tender document.

[0041] Step S5. Matching and Judging Bids Based on the Audit Rule Model: After obtaining the audit rule model, the system applies it to the bids to be audited. This step performs the following operations for each rule: Content Location: Based on the original position pointer of the audit element associated with the rule, locate the corresponding chapter, paragraph, or table in the bid. If the structure of the bid is not completely consistent with the tender document, the system uses a fuzzy matching method based on title similarity to find the most relevant bid chapter. Rule Matching: For integrity rules: Check if the located bid chapter exists and if its content is not empty; if it is a document requirement, check if the bid attachment contains the specified document. For compliance rules: Extract the text of the corresponding paragraph in the bid, use a semantic matching method (such as calculating the cosine similarity between text vectors) to calculate the similarity between the bid text and the tender requirement text, and compare the similarity with a preset threshold. For parameter comparison rules: Extract numerical parameters from the bid text, use regular expressions to match numbers and units, convert the units to standard units (such as "kW", "kg", "mm", etc.), and then perform numerical comparison operations. For the scoring rules: Semantically match the content of the tender response with the descriptions of each score level in the scoring details, and take the score corresponding to the highest matching level as the score for that item. For the business response rules: Check whether there are any "response" or "deviation" statements in the tender document, or determine whether the tender document explicitly agrees to the clause through semantic analysis. Result determination: Output a determination result for each rule, including "satisfied", "not satisfied", or "partially satisfied"; for the scoring rules, output the specific score value.

[0042] Step S6. Generate Structured Audit Result Data: Summarize all judgment results obtained in Step S5 to form structured audit result data; each result data should contain at least the following fields: Audit Element Identifier: Associated with the element ID in Step S3. Element Type: Required material type, technical parameter type, etc. Matching Status: Satisfied / Unsatisfied / Partially Satisfied. Reason for Unsatisfiedness: For items that are unsatisfied or partially satisfied, the system generates a natural language description of the reason, such as "The tender document does not provide financial audit reports for the past three years" or "The rated power in the technical parameters is stated as 45kW, which is lower than the 50kW required by the tender." Suggested Modifications: Based on the reason for unsatisfiedness, the system generates modification suggestions, such as "Please provide financial audit reports for 2021-2023" or "Please increase the rated power to above 50kW." Risk Level Identification: Based on the importance of the review elements and the severity of any non-compliance, risk levels are categorized as high, medium, and low. High risk typically corresponds to defects such as missing required fields, unmet key parameters, or expired qualifications that could lead to bid rejection. Medium risk corresponds to deviations from non-critical clauses or incomplete materials. Low risk corresponds to flaws such as non-standard formatting, layout, or terminology. Score: For each scoring element, the actual score is recorded.

[0043] Step S7. Generate audit report and visualization results:

[0044] Based on the structured audit results data from step S6, the system automatically generates an audit report. The report can be exported as PDF, Word, or HTML, and includes: Audit Summary: Total number of audit items, number of passed items, number of non-compliant items, estimated total score, etc. Issue List: Sorted by risk level, listing all audit elements that are non-compliant or partially compliant, along with reasons for non-compliance, suggested modifications, and the issue's location in the tender document (page number, chapter). Scoring Details: Listing the name, maximum score, actual score, and scoring basis for each scoring item. Visualization Charts: Displaying the audit pass rate for each dimension in the form of radar charts, bar charts, etc. After the report is generated, it is displayed on the user interface and supports downloading and printing.

[0045] Example 2: This example further describes the manual review and final audit conclusion based on Example 1.

[0046] After automatically generating the audit report, the system provides a manual review function. Auditors can view the results of the automatic audit item by item on the review interface, modify the judgment results, adjust the risk level, and supplement or correct the modification suggestions; the system records the original value and the modified value of each manual modification, forming an audit log.

[0047] After all audit items have been manually confirmed, the system automatically generates a final audit conclusion based on the matching status and risk level of each audit item. The final audit conclusion is divided into three types: Pass: All high-risk items are met, and the number of medium-risk items does not exceed a preset threshold (e.g., 3). Conditional Pass: A small number of medium-risk or low-risk items are not met, but this does not affect the validity of the bid; the system includes a list of rectification conditions. Fail: Any high-risk item is not met, or the number of medium-risk items exceeds the preset threshold. The system saves the final audit conclusion, the manual review modification records, and the audit report to the audit log database for subsequent quality traceability and model optimization.

[0048] Example 3: Figure 3As shown in the figure, this embodiment proposes an intelligent tender document review system based on artificial intelligence. The system includes the following modules: a document parsing module: used to acquire and parse tender documents and tenders to be reviewed, performing text extraction and structure recognition; this module supports common formats such as PDF, Word, and TXT, and can integrate OCR components to process scanned documents. A review element extraction module: used to extract review elements from tender documents, identifying various review elements and their constraint attributes; this module includes a constraint term dictionary, a natural language processing pipeline, and a table parser. A structured model construction module: used to convert the extracted review elements into a structured model of review elements with a unified data structure, and store it in a relational database or document database. A rule generation module: used to dynamically generate a review rule model based on the structured model of review elements; this module further includes: a deterministic index calculation unit: calculating the D value of each review element according to the aforementioned formula; a triage decision unit: deciding whether to automatically generate rules or transfer to manual confirmation based on the comparison result of the D value and the threshold; and a rule generation unit: calling the corresponding rule generation logic according to the element type, filling in parameters such as constraints, thresholds, and weights to generate specific review rules. Matching and Judgment Module: Used for rule-driven matching and judgment of tender documents based on the review rule model. This module implements functions such as content location, semantic matching, numerical comparison, and score calculation. Result Output Module: Used to generate structured review result data and review reports, supporting export in multiple formats. Manual Review Interface Module: Used to receive elements to be confirmed where the rule generation certainty index is below a threshold, and provides a manual confirmation or correction interface; it is also used to receive manual review operations for automatic review results.

[0049] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A method for intelligent review of tender documents based on artificial intelligence, characterized in that, Includes the following steps: Step S1: Obtain and parse the tender documents and bids to be reviewed, performing document parsing, text extraction, and structure recognition; Step S2: Extract review elements from the tender documents, identifying at least mandatory material elements, technical parameter elements, commercial clause elements, and scoring item elements, and extracting element content, element type, and constraint attributes for each review element; Step S3: Convert the review elements extracted in Step S2 into a unified data structured model of review elements, the model including element identifiers, element types, constraint conditions, and scoring information; Step S4: Based on the structured model of review elements, dynamically generate a non-preset review rule model for the current tender documents; wherein, for each The review elements are evaluated based on the strength of the constraint words, the degree of textual ambiguity, and the information entropy of the chapter they belong to. A deterministic index is generated according to a preset nonlinear function calculation rule, and the magnitude of the index determines whether to automatically generate review rules or transfer the review to manual confirmation. The review rule model includes at least one of the following: integrity rules, compliance rules, parameter comparison rules, and scoring rules. Step S5: Based on the review rule model, the tender documents are matched and judged in a rule-driven manner, and the matching results of each review element are output. Step S6: Structured review result data is generated, including matching results, a list of non-compliant items, and scoring results. Step S7: An review report and visualization results are generated based on the structured review result data.

2. The method for intelligent review of tender documents based on artificial intelligence according to claim 1, characterized in that, The specific methods for extracting review elements from the tender documents in step S2 include: identifying constraint words in the tender documents through natural language processing to locate constraint attributes; identifying various review elements based on chapter titles and paragraph structures; and further extracting scoring criteria, score weights, and scoring details for scoring items.

3. The method for intelligent review of tender documents based on artificial intelligence according to claim 1, characterized in that, The structured model of the audit elements in step S3 includes at least the following fields: element identifier, element type, constraint text, constraint attribute quantification value, position pointer of the element in the original tender document, and score weight for the scoring item element. The constraint attribute quantification value is assigned according to the predefined rules based on the type and strength of the constraint term.

4. The method for intelligent review of tender documents based on artificial intelligence according to claim 1, characterized in that, In step S4, when the deterministic index is generated by the calculation rules, the strength of the constraint words is quantified based on the types of constraint words and negative words appearing in the review elements; the text ambiguity is calculated based on the standard deviation of the word vector of the sentence containing the element and normalized to the [0,1] interval; the chapter information entropy is calculated based on the distribution probability of the part-of-speech category within the chapter according to the Shannon entropy formula; and the maximum value of the information entropy of each chapter in the entire tender document is used as the normalization benchmark; the above three parameters are fused according to the preset nonlinear combination rules to obtain the deterministic index.

5. The intelligent tender document review method based on artificial intelligence according to claim 1, characterized in that, In step S4, the specific method for determining whether to automatically generate rules or transfer to manual confirmation based on the value of the certainty index generated by the rules is as follows: a certainty threshold is preset. When the certainty index is greater than or equal to the threshold, it is determined that the audit element has sufficient rule generation reliability, and it is automatically converted into the corresponding category of audit rules and added to the audit rule model. When the certainty index is less than the threshold, the audit element is marked as an item to be manually confirmed and pushed to the manual review interface. After the reviewer confirms or corrects it, a rule is generated, or it is determined to be an invalid element. The audit rule model is a project-specific rule model dynamically generated for the current tender document, and its rule conditions and rule actions are all derived from the audit elements extracted in step S2.

6. The method for intelligent review of tender documents based on artificial intelligence according to claim 1, characterized in that, In step S4, different categories of audit rules are automatically generated based on the element type of the audit element. Specifically, these include: if the element type is a required field, an integrity rule is generated to check whether the corresponding chapter or document exists in the tender document; if the element type is a technical requirement, a compliance rule is generated to semantically match the tender document text with the tender requirements; if the element type contains numerical constraints, a parameter comparison rule is generated to compare the numerical parameters and constraint values ​​in the tender document; if the element type is a scoring item, a scoring rule is generated to automatically calculate the score according to the scoring details; and if the element type is a commercial clause, a commercial response rule is generated to check the tender document's response to each commercial clause.

7. The method for intelligent review of tender documents based on artificial intelligence according to claim 1, characterized in that, Step S5 involves rule-driven matching and judgment of the tender documents based on the review rule model. Specifically, this includes: for completeness rules, locating the corresponding chapters or attachments in the tender documents and checking their existence and non-emptiness; for compliance rules, extracting the corresponding paragraph text in the tender documents, matching them with the bidding requirements using semantic similarity calculation methods, and determining the degree of compliance; for parameter comparison rules, extracting numerical parameters from the tender document text, unifying the units, and then performing greater than, less than, and equal to comparison operations with the constraint values; and for scoring rules, matching the tender response content with the descriptions of each score level in the scoring details to determine the score for that item.

8. The method for intelligent review of tender documents based on artificial intelligence according to claim 1, characterized in that, The structured audit result data generated in step S6 includes the following fields: audit element identifier, element type, matching status, description of reasons for non-compliance, suggested modifications, risk level identifier, and obtained score; wherein, the risk level identifier is divided into high risk, medium risk, and low risk according to the importance of the audit element and the severity of non-compliance.

9. The method for intelligent review of tender documents based on artificial intelligence according to claim 1, characterized in that, The method also includes manual review and confirmation and final audit conclusion generation steps: after generating the audit report, a manual review interface is provided for auditors to view and modify the audit results item by item; after all audit items have been manually confirmed, the system automatically generates the final audit conclusion based on the matching status and risk level of each audit item, and the final audit conclusion includes "pass", "conditionally pass" or "fail"; the system saves the manual review modification records and the final audit conclusion to the audit log database.

10. An AI-based intelligent tender document review system, used to implement the AI-based intelligent tender document review method as described in any one of claims 1-9, characterized in that, It includes the following modules: Document parsing module: used to acquire and parse bidding documents and tender documents to be reviewed, and to perform text extraction and structure recognition; Review element extraction module: used to extract review elements from bidding documents and identify various review elements and their constraint attributes; The module includes: a structured model building module, used to convert extracted audit elements into a structured model of audit elements with a unified data structure; a rule generation module, used to dynamically generate audit rule models based on the structured model of audit elements, including calculating the rule generation determinism index for each audit element and deciding whether to automatically generate rules or transfer them to manual confirmation based on the index; a matching and judgment module, used to perform rule-driven matching and judgment on tender documents based on the audit rule model; and a result output module, used to generate structured audit result data and audit reports. Manual review interface module: used to receive elements to be confirmed when the rule generation certainty index is lower than the threshold, and to provide a manual confirmation or correction interface.

Images