Thought chain-based bidding auxiliary evaluation and difference review method and system

Abstract

The present application belongs to the technical field of document processing and bid evaluation decision, and aims to solve the technical problems of low accuracy, lack of traceable records, and the difference between auxiliary scoring and expert scoring not forming a review optimization mechanism in the prior art. The method of the present application comprises: obtaining a procurement document and identifying an evaluation standard table of the procurement document, outputting evaluation standard data; obtaining response data of each supplier, and outputting a supplier response data set based on the response data; dividing all evaluation item data into objective evaluation items and subjective evaluation items based on the evaluation standard data, generating objective item evaluation thinking chains based on the objective evaluation items, and generating corresponding subjective item evaluation thinking chains based on the subjective evaluation items; based on the supplier response data set, combining the objective item and subjective item evaluation thinking chains, obtaining comprehensive evaluation data of each participating supplier and generating an evaluation ranking result; obtaining actual scoring data input by bid evaluation experts and performing difference analysis based on the comprehensive evaluation data to generate a difference review report.

Description

Technical Field

[0001] This invention belongs to the field of document processing and bid evaluation decision-making technology, specifically involving a bidding auxiliary evaluation and difference review method and system based on thinking chain. Background Technology

[0002] The bidding and evaluation process is the core of the procurement process. The comprehensive scoring method requires evaluation experts to score each supplier's technical solutions, business terms, price, etc., and finally determine the winning supplier. The quality of the evaluation directly affects the fairness and rationality of the procurement results.

[0003] Existing methods for assisting in bid evaluation mainly involve evaluation experts scoring each supplier and each item according to the scoring criteria in the procurement documents. This method is inefficient, and the scores given by different experts vary greatly, making it difficult to ensure consistency. Alternatively, electronic bid evaluation systems based on preset scoring criteria allow experts to score online, and the system automatically calculates the total score. While this method achieves online scoring and automatic calculation of the total score, it only records the final score and not the scoring reasoning process, making it impossible to review and optimize. Some methods utilize natural language processing technology to conduct preliminary analysis of the bid documents and provide scoring references. However, this method does not distinguish between the processing of objective and subjective items, nor does it establish a scoring process recording mechanism or a difference review mechanism. Summary of the Invention

[0004] (a) Purpose of the invention The purpose of this invention is to provide a bidding and tendering auxiliary evaluation and difference review method and system based on the thinking chain, which aims to solve the technical problems existing in the prior art, such as low accuracy of evaluation standard table parsing, insufficient efficiency and accuracy due to mixed processing of subjective and objective evaluation items, lack of traceable reasoning records in the scoring process, and lack of an automated review and optimization mechanism for the difference between auxiliary scoring and expert scoring.

[0005] (II) Technical Solution To address the aforementioned problems, a first aspect of the present invention provides a method for auxiliary evaluation and discrepancy review in bidding and tendering based on a thought chain approach, the method comprising: Obtain the procurement documents and identify the evaluation criteria table in the procurement documents; extract evaluation item data based on the evaluation criteria table and output the evaluation criteria data. Obtain the response data of each supplier, which includes quotation data, deviation data, performance data, certification data and complete text of tender documents, and output supplier response dataset based on the response data; Based on the evaluation standard data, all the evaluation item data are divided into objective evaluation items and subjective evaluation items. An objective evaluation thinking chain is generated based on the objective evaluation items, and a corresponding subjective evaluation thinking chain is generated based on the subjective evaluation items. Based on the supplier response dataset and the objective and subjective evaluation thought processes, comprehensive evaluation data for each participating supplier is obtained and an evaluation ranking result is generated. Obtain the actual scoring data entered by the evaluation experts and conduct a difference analysis based on the comprehensive evaluation data to generate a difference review report.

[0006] Preferably, the step of acquiring the procurement documents and identifying the evaluation criteria table of the procurement documents, extracting evaluation item data based on the evaluation criteria table, and outputting evaluation criteria data includes: Obtain the Word document of the procurement document and identify all table objects in the Word document; Based on each of the table objects, the text data of each cell in the first row is read, and the text data is matched with the preset evaluation standard table header feature word library to obtain the matching degree data; When the matching data exceeds a preset matching threshold, the corresponding table object is marked as an evaluation standard table; Extract the evaluation item name data, score data, evaluation details data, and weight allocation data row by row from the evaluation criteria table; The evaluation criteria data are output based on the extracted data.

[0007] Preferably, the method further includes: Based on the evaluation method field data in the evaluation standard table, the evaluation method type is identified through preset keyword matching rules; When the assessment method is a comprehensive assessment method, technical weight data, business weight data, and price weight data are extracted from the assessment standard table; When the evaluation method is the lowest valuation method, the price weight data is set to the preset maximum value, and the technical weight data and business weight data are set to zero. Write the evaluation method type data and the corresponding weight data into the evaluation standard data.

[0008] Preferably, the step of generating the objective evaluation thought chain based on the objective evaluation item includes: For each of the objective evaluation items, the corresponding quantitative evaluation rule type is determined based on its evaluation details data. The quantitative evaluation rule types include numerical comparison type, validity period verification type, and deviation detection type. Based on the quantitative evaluation rule type, a corresponding objective item evaluation thinking chain is generated. The objective item evaluation thinking chain includes data acquisition nodes, data verification nodes, rule matching nodes, and conclusion output nodes arranged in sequence.

[0009] Preferably, the step of generating the corresponding subjective evaluation thought chain based on the subjective evaluation item includes: For each of the subjective evaluation items, load the corresponding evaluation rule template according to its evaluation details data; Based on the evaluation dimensions defined in the evaluation rule template, a corresponding subjective item evaluation thought chain is generated. The subjective item evaluation thought chain includes text location nodes, feature extraction nodes, dimension matching nodes, weighted calculation nodes, and interval output nodes arranged in sequence.

[0010] Preferably, the comprehensive evaluation data for each participating supplier is obtained by combining the supplier response dataset with the objective evaluation thought process and the subjective evaluation thought process, wherein the processing of objective evaluation items includes: According to the node sequence defined in the objective evaluation thought chain, each processing node is executed sequentially. For numerical comparison-type evaluation items, first numerical data is extracted from the text layer data of the complete text of the tender document, and second numerical data is extracted from the optical character recognition result data of the complete text of the tender document. The first numerical data and the second numerical data are compared for consistency. Based on the comparison results, valid numerical data is determined. The valid numerical data is compared with the preset threshold data to generate score data. For validity period verification-type assessment items, the validity period expiration date data is extracted from the certification certificate data, the current system date data is obtained, the validity period expiration date data is compared with the current system date data, and the score data is generated based on the comparison result; For deviation detection type evaluation items, the deviation data is scanned based on the preset negative deviation keyword matching rules to detect whether there are negative deviation records, and score data is generated based on the detection results; The evaluation calculation process for each objective evaluation item generates evaluation process record data.

[0011] Preferably, the comprehensive evaluation data for each participating supplier is obtained by combining the supplier response dataset with the objective evaluation thought process and the subjective evaluation thought process, wherein the processing of subjective evaluation items specifically includes: According to the node sequence defined in the subjective evaluation thought chain, each processing node is executed sequentially. At the text location node, the text paragraph data corresponding to each evaluation dimension is located from the complete text of the tender document according to the text paragraph location rules corresponding to each evaluation dimension. At the feature extraction node, content features are extracted from the text paragraph data based on a preset feature lexicon and semantic matching rules to generate feature matching result data. At the dimension matching node, the feature matching result data is judged according to the matching judgment criteria of each evaluation dimension to generate matching degree data for each dimension. At the weighted calculation node, based on the matching degree data of each dimension and the preset dimension weight parameters, suggested score range data is generated through weighted calculation. The suggested score range data includes suggested lower limit value and suggested upper limit value. At the output node of the interval, the feature matching result data and dimension weight parameters of each evaluation dimension are assembled into evaluation reason data.

[0012] Preferably, the step of obtaining the actual scoring data entered by the evaluation experts and performing a difference analysis based on the comprehensive evaluation data to generate a difference review report includes: Receive actual scoring data entered by evaluation experts through the interactive interface; For each evaluation item, calculate the difference between the auxiliary suggestion score and the actual score. Filter out the evaluation items whose difference values ​​exceed a preset difference threshold, and mark them as evaluation items to be reviewed; For the items to be reviewed and evaluated, the causes of the differences are classified based on preset difference attribution rules; The difference review report is generated based on the difference cause classification results. The difference review report includes a list of difference items, difference value data, difference cause classification results, and optimization suggestion data.

[0013] Preferably, the classification of the causes of differences based on preset difference attribution rules includes: For the evaluation item to be reviewed, obtain the evaluation process record data corresponding to the evaluation item; Based on the preset data integrity detection rules, check whether there are missing or abnormal markers in the data source of the data recorded in the evaluation process. When there are missing or abnormal markers, the reasons for the difference are classified as data missing reasons. When the data source in the evaluation process record data is complete and there are no abnormal markers, based on the preset rule matching deviation detection rule, the feature matching result data in the evaluation process record data is compared with the actual scoring data. When the feature matching result data is inconsistent with the actual scoring trend, the reason for the difference is classified as the reason for rule understanding deviation. When the difference values ​​of the same evaluation item across multiple suppliers all show the same direction of deviation, and the deviation exceeds the preset preference threshold, the reason for the difference will be classified as the reason for expert preference mode. Write the data of the differences and their causes into the difference review report.

[0014] Another aspect of the present invention provides a bidding assistance evaluation and difference review system based on the thinking chain, the system comprising: The evaluation criteria parsing module is used to obtain the procurement documents and identify the evaluation criteria table of the procurement documents, extract evaluation item data based on the evaluation criteria table and output the evaluation criteria data. The data query module is used to obtain the response data of each supplier. The response data includes quotation data, deviation data, performance data, certification data and complete text of tender documents. Based on the response data, the module outputs a supplier response dataset. The thought chain generation module divides all the evaluation item data into objective evaluation items and subjective evaluation items based on the evaluation standard data, generates an objective item evaluation thought chain based on the objective evaluation items, and generates a subjective item evaluation thought chain based on the subjective evaluation items. The comprehensive evaluation module, based on the supplier response dataset and combining the objective and subjective evaluation thought processes, obtains comprehensive evaluation data for each participating supplier and generates evaluation ranking results. The discrepancy review module acquires the actual scoring data entered by the evaluation experts and performs discrepancy analysis based on the comprehensive evaluation data to generate a discrepancy review report.

[0015] (III) Beneficial Effects The above-described technical solution of the present invention has the following beneficial technical effects: 1. Generate a corresponding thought chain record for each evaluation. The thought chain fully covers all aspects from data input and rule matching to conclusion output. Each step has a clear record and basis. When the evaluation results are questioned or require compliance audit, the complete reasoning process of the corresponding evaluation item can be directly retrieved as supporting material. There is no need to manually organize the process documents in a piecemeal manner, which improves the efficiency of objection response and audit verification. At the same time, it can also effectively constrain the arbitrariness of the evaluation process, ensure the fairness and impartiality of procurement activities, and enhance the credibility of the evaluation results.

[0016] 2. For objective evaluation items with clear quantitative standards, automated quantitative verification logic is adopted to achieve fully automated and highly accurate result calculation, replacing the inefficient labor of manually verifying data and avoiding omissions and biases in manual verification. For subjective evaluation items requiring professional judgment, fixed scores are not directly output. Instead, reference ranges with complete reasoning are generated through semantic analysis, providing systematic analytical support for experts while fully preserving the independent review rights of the evaluation experts. Both models leverage their respective technological advantages, resulting in significant improvements in evaluation efficiency and result accuracy compared to mixed processing solutions.

[0017] 3. A complete capability iteration loop is formed through the difference review process. The differences between the auxiliary evaluation results and the actual expert evaluation contain a large amount of high-value information such as data bias, rule adaptation bias, and review bias. Through automated difference analysis and attribution, this information can be transformed into actionable optimization directions, continuously feeding back into the optimization of evaluation rules and data processing mechanisms. Attached Figure Description

[0018] Figure 1 This is a flowchart of a bidding auxiliary evaluation and difference review method based on the thinking chain provided by the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and the accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0020] Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0021] Combination Figure 1 The first aspect of this invention provides a method for auxiliary evaluation and difference review in bidding and tendering based on a thought chain, the method comprising: Obtain the procurement documents and identify the evaluation criteria table in the procurement documents; extract evaluation item data based on the evaluation criteria table and output the evaluation criteria data. Obtain the response data of each supplier, including quotation data, deviation data, performance data, certification data and complete tender document text, and output supplier response dataset based on the response data; Based on the evaluation standard data, all evaluation items are divided into objective evaluation items and subjective evaluation items. An objective evaluation thought chain is generated based on the objective evaluation items, and a corresponding subjective evaluation thought chain is generated based on the subjective evaluation items. Based on the supplier response dataset, combined with the objective and subjective evaluation thought processes, comprehensive evaluation data for each participating supplier is obtained and evaluation ranking results are generated. Obtain the actual scoring data entered by the evaluation experts and conduct a difference analysis based on the comprehensive evaluation data to generate a difference review report.

[0022] Specifically, firstly, document structure parsing technology transforms unstructured procurement documents into machine-processable structured evaluation standard data; secondly, it unifies and aggregates scattered multi-source supplier response data to form a complete evaluation data foundation; thirdly, based on evaluation rule attributes, evaluation items are divided into objective and subjective categories, respectively adapting to quantitative verification and semantic analysis reasoning chains, with both chains recording the reasoning process in a chain structure—the evaluation thought chain; subsequently, a comprehensive evaluation result and ranking are generated based on unified weighted calculation rules; finally, by comparing actual expert evaluation data with auxiliary evaluation results, it automatically performs discrepancy analysis and outputs review conclusions, achieving a closed-loop iteration of evaluation capabilities.

[0023] This method generates a corresponding thought chain record for each evaluation, fully covering all stages from data input and rule matching to conclusion output. Each step has clear records and evidence. When the evaluation results are questioned or require compliance audits, the complete reasoning process of the corresponding evaluation item can be directly retrieved as supporting material, eliminating the need for manually compiling fragmented process documents. This improves the efficiency of responding to objections and audit verification, while also effectively constraining the arbitrariness of the evaluation process, ensuring the fairness and impartiality of procurement activities, and enhancing the credibility of the evaluation results. For objective evaluation items with clear quantitative standards, automated quantitative verification logic is used to achieve fully automated and highly accurate result calculation, replacing the inefficient labor of manually verifying data and avoiding omissions and biases in manual verification. For subjective evaluation items requiring professional judgment, fixed scores are not directly output. Instead, reference ranges with complete reasoning evidence are generated through semantic analysis, providing experts with systematic analytical support while fully preserving the independent review rights of the evaluation experts. Both models leverage their respective technological advantages, significantly improving evaluation efficiency and result accuracy compared to mixed processing solutions. The difference review process forms a complete capability iteration loop. The differences between the auxiliary evaluation results and the actual expert evaluation contain a large amount of high-value information, such as data bias, rule adaptation bias, and review bias. Through automated difference analysis and attribution, this information can be transformed into actionable optimization directions, continuously feeding back into the optimization of evaluation rules and data processing mechanisms.

[0024] In a preferred embodiment, the procurement documents are obtained and the evaluation criteria table in the procurement documents is identified. Evaluation item data is extracted based on the evaluation criteria table, and evaluation criteria data is output, including: Retrieve the Word document containing the procurement documents and identify all table objects within the Word document; Based on each table object, the text data of each cell in the first row is read, and the text data is matched with the preset evaluation standard table header feature word library to obtain the matching degree data; When the matching score exceeds the preset matching threshold, the corresponding table object is marked as the evaluation standard table; Extract the data of evaluation item names, scores, evaluation details, and weight allocation from the evaluation criteria table row by row; The extracted data is used to output evaluation criteria data.

[0025] Specifically, by leveraging the fixed header format of the evaluation standard table in the procurement documents, the system traverses the original table objects in the document, reads the first row of text, and calculates the matching degree with a preset feature word library to accurately locate the evaluation standard table, eliminating interference from the main text paragraphs and other business tables. After locating the table, the system extracts the core fields of the evaluation items row by row according to the column mapping relationship and outputs structured data.

[0026] This method directly identifies evaluation criteria based on native table objects in Word documents, rather than performing a fuzzy search for evaluation standards within the entire text. This avoids the problem of confusion between paragraph text and table content at the data source level. Furthermore, by calculating the matching degree of table header feature words, it effectively filters out a large number of non-evaluation tables in procurement documents, such as procurement requirement tables, qualification condition tables, and parameter comparison tables, significantly reducing the probability of misidentification and omission, and effectively reducing the workload of subsequent manual corrections. After locating the target table, the system automatically extracts the core fields of all evaluation items row by row according to column mapping relationships. It can directly convert unstructured table data formatted in Word into structured item data that can be directly accessed by the machine. There is no need for manual entry of evaluation items, scores, and details, significantly shortening the time spent on standard entry before bidding. It also avoids typos such as incorrect score input, omission of details copying, and omission of items that may occur during manual entry, ensuring the accuracy of the basic data for subsequent evaluation calculations from the source.

[0027] In a preferred embodiment, the method further includes: For each row in the evaluation criteria table, read the text data of the evaluation details cell; The text data of the evaluation details cell is scanned for patterns based on a preset delimiter rule base, which includes number sequence pattern, semicolon delimiter pattern and line break delimiter pattern. When multiple sub-items that conform to the delimiter rule library are detected in the text data of the evaluation details cell, the multiple sub-items will be created as independent evaluation item data records respectively; For each independent evaluation item data record, the evaluation item name data and score data of the original table row are inherited, and the score data is split according to the number of sub-items and the preset proportion allocation rules. The split evaluation item data records are written into the evaluation standard data.

[0028] Specifically, due to layout requirements, actual procurement documents often place multiple sub-evaluation items in the same cell of the same table row. If they are not split, they can only be evaluated in a coarse-grained manner and cannot accurately match the corresponding response content. By scanning the text with preset rules for delimiters such as numbering, semicolons, and line breaks, composite sub-items can be automatically identified and broken down into independent evaluation units. At the same time, the scores can be split according to the rules to achieve finer evaluation granularity.

[0029] This method, by splitting composite sub-items within the same cell into independent evaluation units, allows each sub-item to be independently matched with supplier responses, generate its own thought chain, and calculate its own evaluation result. This avoids the matching ambiguity caused by mixed evaluations of composite items. For example, when a cell contains both performance requirements and certificate requirements, after splitting, each can be independently verified against performance and qualification data. It prevents the entire composite item from receiving full marks simply because one sub-item meets the requirement, and also avoids deducting points for not meeting the requirement. The scoring and rule matching are more accurate, and the evaluation results are more closely aligned with the original rules set in the procurement documents. It can automatically identify and split hidden sub-items within cells without requiring manual adjustments to the table format beforehand. While preserving the original procurement document layout structure, it achieves a complete breakdown of evaluation rules. The automatic splitting mechanism can fully identify all evaluation sub-rules, ensuring that each scoring detail corresponds to a specific evaluation logic, avoiding rule omissions, improving processing efficiency, and guaranteeing the compliance of score allocation.

[0030] In a preferred embodiment, the method further includes: Based on the evaluation method field data in the evaluation criteria table, the evaluation method type is identified through preset keyword matching rules; When the assessment method is a comprehensive assessment method, extract the technical weight data, business weight data, and price weight data from the assessment standard table; When the assessment method is the lowest appraised value method, the price weight data is set to the preset maximum value, and the technical weight data and commercial weight data are set to zero; Write the data on the evaluation method type and the corresponding weight data into the evaluation standard data.

[0031] Specifically, the evaluation method field is located in the header or description row of the evaluation criteria table, and the corresponding text content is read. The evaluation method type is identified using preset keyword matching rules. If the text contains keywords such as "comprehensive evaluation" or "comprehensive score," it is determined to be a comprehensive evaluation method; if it contains keywords such as "lowest bid price" or "lowest evaluation price," it is determined to be a minimum evaluation price method. When determined to be a comprehensive evaluation method, the specific values ​​of technical weight, commercial weight, and price weight are automatically extracted from the table. When determined to be a minimum evaluation price method, the price weight is automatically set to the preset maximum value, and the technical and commercial weights are set to zero. Finally, the evaluation method type and corresponding weight data are written into the structured evaluation criteria data.

[0032] This method automatically identifies the evaluation method type, allowing the same system to simultaneously support two mainstream evaluation scenarios: comprehensive evaluation and lowest evaluated price. There's no need to develop separate system versions for different evaluation methods. Procurement projects across different industries and product categories can directly use the same auxiliary evaluation system; simply uploading the corresponding procurement documents automatically adapts the correct evaluation logic, improving the system's versatility and applicability, and reducing deployment costs across multiple scenarios. Automatically identifying the evaluation method based on the original procurement document content avoids the risk of selecting the wrong method during manual operation, ensuring complete consistency between the evaluation logic and the procurement document requirements. This mechanism significantly reduces the probability of human error, especially in scenarios with a large number of projects and complex operators, guaranteeing the compliance of evaluation results. Furthermore, all weight data originates directly from the original procurement documents, possessing complete traceability. In case of disputes, the source can be directly traced back to the original table content, ensuring the rigor and verifiability of the evaluation process.

[0033] In a preferred embodiment, the response data from each supplier is obtained, including: Based on the procurement project identification data, retrieve all associated supplier records from the data storage system; For each supplier record, query its quotation data table, deviation data table, performance data table, and certification data table separately; Retrieve the complete text of each supplier's bid document from the file storage system. The complete text of the bid document includes text layer data and optical character recognition result data. The retrieved data is linked and assembled according to the supplier dimension to generate a supplier response dataset.

[0034] Specifically, supplier response data is stored in different systems: structured business data is stored in the business database, unstructured tender documents are stored in the file system, and image recognition data is stored in the OCR system. By using a unified retrieval logic to pull data by project dimension and assemble it by supplier dimension, a complete response dataset can be formed, providing a unified and complete data foundation for subsequent evaluation.

[0035] This method unifies supplier data scattered across different storage systems and formats into a standardized dataset, categorized by supplier. All evaluation stages are conducted using the same complete dataset, eliminating the need for repeated data retrieval across systems. This improves evaluation efficiency and avoids biases caused by asynchronous multi-source data, ensuring fairness by basing all evaluation dimensions on the same set of data for the same supplier. All response data is automatically queried and assembled via system interfaces, eliminating the need for manual downloading of bid documents and organization of performance and certificate data. This reduces pre-evaluation data preparation workload, especially significant for large projects with numerous suppliers and complex qualification materials. Automated data collection significantly shortens the evaluation preparation cycle. Furthermore, automatic data association using unique identifiers avoids supplier data mismatches and material omissions that occur during manual organization, significantly improving data accuracy. Simultaneously acquiring both the original text layer data and OCR recognition result layer data from the bid documents provides necessary data support for subsequent dual cross-validation of objective items. This dual-layer data enables cross-validation, effectively filtering erroneous data and improving the reliability of objective item evaluations—a technical advantage that solutions relying solely on single text data cannot achieve.

[0036] In a preferred embodiment, an objective evaluation thought chain is generated based on the objective evaluation items, including: For each objective evaluation item, the corresponding quantitative evaluation rule type is determined based on its evaluation details data. The quantitative evaluation rule types include numerical comparison type, validity period verification type, and deviation detection type. Based on the quantitative evaluation rule type, a corresponding objective item evaluation thought chain is generated. The objective item evaluation thought chain includes data acquisition nodes, data verification nodes, rule matching nodes, and conclusion output nodes arranged in sequence.

[0037] Specifically, all objective evaluation items are quantitative rules, but the verification logic for different subcategories differs significantly. We first categorize these rules into three typical types: numerical comparison, validity period verification, and deviation detection. Then, we uniformly break them down into four standardized processing nodes: data acquisition, data verification, rule matching, and conclusion output. This ensures that different types are adapted to specific verification logic while standardizing the thought process structure, facilitating unified traceability and management. Specifically, the data acquisition node records the identifiers of the data fields to be acquired and the data source path; the data verification node records the data integrity verification rules and anomaly handling strategies; the rule matching node records the quantitative comparison rules and threshold parameters; and the conclusion output node records the score calculation logic and process record data generation rules.

[0038] This method breaks down the evaluation process of objective items into four standardized chain nodes. Each node records clearly defined input data, processing logic, and output results, making the entire evaluation process transparent and traceable. When questions arise about the evaluation results, the process can be traced back layer by layer along the thought chain to accurately pinpoint whether the error occurred in the data acquisition, verification, or rule matching stages. This eliminates the need to rerun the entire evaluation process, significantly improving problem-solving efficiency and freeing the evaluation process from a black box, thus enhancing its auditability. Dedicated verification logic is tailored to three different types of objective evaluation items: numerical comparison focuses on cross-validation, validity period verification focuses on date comparison, and deviation detection focuses on semantic keyword scanning. Each of these three logics precisely corresponds to a business rule, avoiding rule adaptation errors caused by a one-size-fits-all approach and improving the overall accuracy of objective item evaluation.

[0039] In a preferred embodiment, a corresponding subjective evaluation thought chain is generated based on the subjective evaluation item, including: For each evaluation item in the subjective evaluation items, load the corresponding evaluation rule template based on its evaluation details data; Based on the evaluation dimensions defined in the evaluation rule template, a corresponding subjective item evaluation thought chain is generated. The subjective item evaluation thought chain includes text location nodes, feature extraction nodes, dimension matching nodes, weighted calculation nodes, and interval output nodes arranged in sequence.

[0040] Specifically, subjective evaluation items lack absolute quantitative standards and require comprehensive judgment from multiple dimensions. By breaking down pre-set rule templates into standardized evaluation dimensions, and further decomposing the evaluation process into five sequentially executed processing nodes—text localization, feature extraction, dimension matching, weighted calculation, and interval output—it transforms ambiguous subjective judgments into a reproducible and traceable structured reasoning process, while retaining expert discretion in the output score range. Specifically, the text localization node records the text paragraph localization rules corresponding to each evaluation dimension; the feature extraction node records the feature vocabulary and semantic matching rules; the dimension matching node records the matching criteria and dimension weight parameters for each evaluation dimension; the weighted calculation node records the weighted summation formula and confidence calculation rules; and the interval output node records the logic for generating the upper and lower limits of the suggested score range and the assembly rules for the evaluation rationale data.

[0041] This method breaks down subjective judgments, which previously relied heavily on individual expert experience, into five standardized processing nodes. Each node has clear and unified processing rules, making the auxiliary evaluation of subjective items reproducible. The evaluation criteria remain consistent across different suppliers and projects, avoiding standard fluctuations caused by factors such as expert condition and time constraints during manual evaluation. This improves the consistency of subjective item evaluations and provides experts with a stable and unified reference benchmark. Based on the breakdown of evaluation details into multiple independent evaluation dimensions, each dimension undergoes independent matching analysis, rather than making a general overall evaluation of subjective items. This results in finer-grained evaluation, more specific evaluation reasons, improved efficiency of expert review, and more persuasive auxiliary recommendations.

[0042] In the preferred embodiment, both the objective item evaluation thought chain and the subjective item evaluation thought chain are stored using a preset structured chain data format. The structured chain data format includes: The evaluation item identifier field is used to uniquely identify the current evaluation item; The assessment type field is used to identify whether the current assessment item is an objective assessment item or a subjective assessment item; The chain node sequence field is used to store the processing node data arranged in the execution order. Each processing node includes a node number, node type, input data identifier, processing logic description, output data identifier, and execution timestamp. Based on the data fields, it is used to store the standard terms data and supplier response data summaries on which the evaluation is based; The conclusion field is used to store the final score data or the suggested score range data; The confidence score field stores the confidence score data of the evaluation results. The confidence score data is calculated based on data integrity and rule matching degree. The anomaly marker field is used to mark data anomalies detected during the evaluation process.

[0043] Specifically, a unified structured storage format is the foundation for the traceability, statistics, analysis, and optimization of the thinking chain. Both objective and subjective thinking chains share the same storage structure, differing only in the content of the chain nodes. This enables unified management, querying, and retrieval of all thinking chains, reducing system complexity. Furthermore, the comprehensive field design supports diverse subsequent data analysis needs. All objective and subjective evaluation thinking chains are stored using a unified structured data format, containing seven core fields: Evaluation Item Identifier: stores a unique ID to associate with the corresponding evaluation item; Evaluation Type: identifies whether the item is objective or subjective; Chain Node Sequence: stores detailed information about all processing nodes in execution order, with each node including number, type, input data identifier, processing logic description, output data identifier, and timestamp; Basis Data: stores the referenced evaluation criteria clauses and response data summary; Conclusion: stores the final score or suggested score range; Confidence: calculated based on data integrity and rule matching degree, characterizing the reliability of the evaluation results; Anomaly Marker: records information such as data anomalies and identification failures detected during the evaluation process.

[0044] This approach ensures that both objective and subjective evaluation thought chains utilize a completely unified storage format, differing only in their internal node content. The system manages all thought chains through a single set of storage, query, and export interfaces, eliminating the need to develop two separate storage and query logics. This significantly reduces system development and maintenance costs. Furthermore, when outputting evaluation reports or retrieving supporting materials for objections, the system can output thought chain content in a unified format, providing a more consistent experience and higher retrieval efficiency. Seven fields comprehensively cover all dimensions, from identifiers, basis, process, conclusions to anomaly markers, forming a complete chain of evidence. From the attribution of evaluation items to the standard clauses of the evaluation basis, to the processing details of each step, the final conclusion, and even data anomalies, everything is clearly recorded. When addressing bidding challenges or compliance audits, complete structured data can be directly retrieved as supporting evidence, eliminating the need for manually organizing fragmented process documents, greatly improving verification efficiency, and fully ensuring the compliance and traceability of the bidding process.

[0045] In the preferred scenario, based on the supplier response dataset and combining objective and subjective evaluation thought processes, comprehensive evaluation data for each participating supplier is obtained. The processing of objective evaluation items includes: Execute each processing node sequentially according to the node sequence defined in the objective item evaluation thought chain; For numerical comparison-type evaluation items, first numerical data is extracted from the text layer data of the complete text of the tender document, and second numerical data is extracted from the optical character recognition result data of the complete text of the tender document. The first and second numerical data are compared for consistency. Based on the comparison results, valid numerical data is determined. The valid numerical data is compared with the preset threshold data to generate score data. For validity period verification-type assessment items, the validity period expiration date data is extracted from the certification certificate data, the current system date data is obtained, the validity period expiration date data is compared with the current system date data, and the score data is generated based on the comparison result. For deviation detection type evaluation items, the deviation data is scanned based on the preset negative deviation keyword matching rules to detect whether there are negative deviation records, and score data is generated based on the detection results; The evaluation calculation process for each objective evaluation item generates evaluation process record data.

[0046] Specifically, the process is executed sequentially according to the node sequence defined in the thought chain. The numerical comparison type uses two-layer data cross-validation to ensure numerical accuracy, the validity period verification type determines the validity of the qualification through date comparison, and the deviation detection type identifies negative deviations through keyword scanning. All execution processes are recorded synchronously in the thought chain, and a fault tolerance mechanism is set up for abnormal scenarios such as recognition failure.

[0047] Furthermore, a consistency comparison is performed on the first and second numerical data, and the valid numerical data is determined based on the comparison results, specifically including: When the first numerical data matches the second numerical data, the matching numerical data will be considered as valid numerical data. When the first numerical data is inconsistent with the second numerical data, the first numerical data extracted from the text layer data shall be used as the standard, and the data inconsistency information shall be marked in the data recorded during the evaluation process. When optical character recognition result data acquisition fails, a boundary processing strategy is executed: a preset base score is assigned, and the data source is marked as boundary processing in the data recorded during the evaluation process. The base score is greater than zero and less than the full score.

[0048] For deviation detection-type assessment items, score data is generated based on the detection results, specifically including: When there are no negative deviation records in the deviation data, assign full marks; When there are negative deviation records in the deviation data, count the number of negative deviation records and calculate the score data according to the preset deduction rules based on the number of negative deviation records. When the optical character recognition result of the deviation data fails to be acquired, a boundary processing strategy is executed: a preset base score is assigned, and the data source is marked as boundary processing in the data recorded during the evaluation process. The base score is greater than zero and less than the full score.

[0049] For validity period verification-type assessment items, score data is generated based on the comparison results, specifically including: If the expiration date is later than the current system date, the certificate is deemed valid and assigned full marks. If the expiration date is earlier than the current system date, the certificate is deemed expired and assigned a score of zero. When the time difference between the expiration date data and the current system date data is less than the preset expiration threshold, the certificate is determined to be near expiration, and a portion of the score data is assigned according to the preset expiration deduction rules. The certificate validity determination results and date comparison data are written into the evaluation process record data.

[0050] This approach employs a dual-data source cross-validation mechanism (text layer and OCR layer) for numerical comparison-based evaluation items, effectively filtering out abnormal data and selecting reliable and valid values ​​for evaluation. This reduces the error rate of numerical evaluations and provides double assurance for the accuracy of objective evaluations. Each of the three types of objective evaluation items corresponds to a dedicated verification logic, perfectly tailored to its respective business characteristics. Numerical comparison focuses on data verification, validity period verification focuses on time dimension judgment, and deviation detection focuses on semantic recognition. This ensures that the execution logic of each objective evaluation corresponds one-to-one with the scoring details in the procurement documents, guaranteeing that the evaluation results fully comply with the procurement rules. Clear fault-tolerance rules are set up for abnormal scenarios such as data inconsistency and recognition failures. This ensures that the overall evaluation work can continue without project delays due to individual data anomalies, and also highlights abnormal items through anomaly markers, prompting evaluation experts to pay close attention, thus improving the system's robustness and adaptability to actual business needs.

[0051] In the preferred scenario, based on the supplier response dataset and combining objective and subjective evaluation thought processes, comprehensive evaluation data for each participating supplier is obtained. The processing of subjective evaluation items specifically includes: Execute each processing node sequentially according to the node sequence defined in the subjective evaluation thought chain; At the text location node, the text paragraph data corresponding to each evaluation dimension is located from the complete text of the tender document according to the text paragraph location rules corresponding to each evaluation dimension. At the feature extraction node, content features are extracted from the text paragraph data based on the preset feature lexicon and semantic matching rules to generate feature matching result data. At the dimension matching node, the feature matching result data is judged according to the matching judgment criteria of each evaluation dimension, and the matching degree data of each dimension is generated. At the weighted calculation node, based on the matching degree data of each dimension and the preset dimension weight parameters, the suggested score range data is generated through weighted calculation. The suggested score range data includes the suggested lower limit value and the suggested upper limit value. At the interval output node, the feature matching result data and dimension weight parameters of each evaluation dimension are assembled into evaluation reason data.

[0052] Specifically, following a semantic processing pipeline from location to extraction and from matching to calculation, the process first accurately identifies the corresponding content paragraphs to eliminate irrelevant interference, then extracts content features and matches them dimension by dimension, and finally generates a score range and assembles evaluation reasons through weighted calculation. The entire process is based on the actual content of the tender document, and all steps are traceable. The process executes node by node in the subjective evaluation thought chain: the text location node accurately identifies the specific paragraph range for the corresponding evaluation dimension from the entire tender document based on location rules such as chapter titles; the feature extraction node extracts content features from the identified paragraphs based on a pre-set feature lexicon and semantic matching rules, generating feature matching result data; the dimension matching node judges the feature matching results according to the judgment criteria of each dimension and outputs the matching degree value for each dimension; the weighted calculation node combines the matching degree of each dimension with the pre-set dimension weights to obtain the benchmark score through weighted calculation and generates a suggested score range including upper and lower limits; the range output node assembles the matching situation of each dimension and the weight parameters into a structured evaluation reason text.

[0053] This method first identifies specific content paragraphs for each evaluation dimension by locating text nodes, then performs feature matching. This effectively eliminates interference from irrelevant chapters. For example, when evaluating technical solutions, only the technical solution section is analyzed, excluding business or pricing information. This results in more accurate matching and evaluation results that better align with the scoring criteria, avoiding misattribution and enhancing the reference value of supplementary suggestions. Through feature extraction and dimension matching, specific matching results are generated for each evaluation dimension. The final evaluation reasoning can be refined to the content coverage and matching degree of each dimension. Experts can quickly grasp the strengths and weaknesses of the bidding proposal in each dimension through the evaluation reasoning, assisting them in efficiently completing their review and judgment. It also provides clear supporting evidence for the suggested scores, significantly enhancing their persuasiveness.

[0054] In the preferred scenario, comprehensive evaluation data for each participating supplier is obtained, and an evaluation ranking result is generated, including: Extract all supplier quotation data from the supplier response dataset; The benchmark price data is determined based on the preset benchmark price calculation rules. The benchmark price calculation rules include taking the arithmetic mean of all the bid data, taking the lowest value among all the bid data, or taking the arithmetic mean after removing the highest and lowest bids. For each supplier, price evaluation data is calculated according to a preset price evaluation formula. The price evaluation formula is: price evaluation data equals the benchmark price data divided by the supplier's quoted price data multiplied by 100% and then multiplied by the price weight data. For each supplier, the comprehensive evaluation data is calculated according to the preset comprehensive evaluation formula. The comprehensive evaluation formula is: the comprehensive evaluation data equals the original technical score data multiplied by the technical weight data, plus the original commercial score data multiplied by the commercial weight data, plus the original price score data multiplied by the price weight data. The evaluation results are generated by sorting the comprehensive evaluation data from highest to lowest.

[0055] Specifically, the valid quotation data of all participating suppliers is extracted from the supplier response dataset; the benchmark price data is calculated according to the project's preset benchmark price rules, such as the arithmetic mean of all quotations, the lowest value, and the average value after removing the highest and lowest values; the price evaluation data is calculated for each supplier according to the price evaluation formula; the original scores of the technical and commercial sections are then summarized separately, and combined with the corresponding weights, the comprehensive evaluation data of each supplier is calculated according to the comprehensive evaluation formula; finally, the comprehensive evaluation data is sorted from high to low, generating an evaluation ranking result that includes the ranking, supplier name, scores of each sub-item, and comprehensive score.

[0056] This method supports multiple benchmark price calculation rules, adapting to the price review requirements of different industries and procurement categories. Users only need to configure parameters to switch rules. Different regions and organizations have different requirements for benchmark price calculation methods. The flexible rule configuration capability allows the system to be widely adapted to various scenarios, avoiding the narrow applicability of a single benchmark price rule and improving the system's versatility. Price scores and comprehensive scores are all automatically calculated by the system according to preset formulas. In large projects with many suppliers and complex scoring items, automated calculation completely avoids arithmetic errors that are prone to occur in manual calculation. The calculation process and parameters are fully traceable and auditable, allowing for verification at any time in case of doubt, ensuring the accuracy of the scoring results.

[0057] In the preferred scenario, the actual scoring data entered by the evaluation experts is obtained, and a difference analysis is performed based on the comprehensive evaluation data to generate a difference review report, including: Receive actual scoring data entered by evaluation experts through the interactive interface; For each assessment item, calculate the difference between the suggested auxiliary score and the actual score. Filter out evaluation items whose difference values ​​exceed a preset difference threshold and mark them as evaluation items to be reviewed; For the review and evaluation items, the causes of the differences are classified based on the preset difference attribution rules; A difference review report is generated based on the difference cause classification results. The difference review report includes a list of difference items, difference value data, difference cause classification results, and optimization suggestion data.

[0058] Specifically, the system receives actual scoring data from evaluation experts via an interactive interface; for each evaluation item, it calculates the difference between the auxiliary suggestion score and the actual score; it filters out evaluation items whose difference exceeds a preset difference threshold and marks them as evaluation items to be reviewed; for all evaluation items to be reviewed, it automatically performs difference cause classification based on preset difference attribution rules; and finally, it integrates the list of difference items, difference value data, difference cause classification results, and corresponding optimization suggestions to generate a structured difference review report.

[0059] By calculating discrepancy values ​​and setting thresholds for filtering, this method allows for the precise identification of key items with significant differences from a large number of evaluation items. This enables optimization efforts to focus on core issues, improving the efficiency of post-mortem optimization and achieving maximum consistency improvement with minimal optimization input. It also avoids the resource waste caused by indiscriminate optimization. From discrepancy calculation and selection of items to be reviewed to cause classification, the entire process is automated, reducing the manual workload of post-mortem reviews. Automated post-mortem reviews can output complete conclusions in a very short time.

[0060] In a preferred embodiment, the causes of differences are classified based on preset difference attribution rules, including: For each item to be reviewed and evaluated, obtain the corresponding evaluation process record data. Based on the preset data integrity detection rules, check whether there are missing or abnormal markers in the data sources of the data recorded in the evaluation process. When there are missing or abnormal markers, the reasons for the discrepancies are classified as data missing reasons. When the data source in the evaluation process record data is complete and there are no abnormal markers, based on the preset rule matching deviation detection rule, the feature matching result data in the evaluation process record data is compared with the actual scoring data. When the feature matching result data is inconsistent with the actual scoring trend, the reason for the difference is classified as the reason for rule understanding deviation. When the difference values ​​of the same evaluation item across multiple suppliers all show the same direction of deviation, and the deviation exceeds the preset preference threshold, the reason for the difference will be classified as the reason for expert preference mode. Write the data from the classification of the causes of the differences into the difference review report.

[0061] Specifically, for each evaluation item to be reviewed, the corresponding complete evaluation process record data is retrieved. First, a data integrity check is run to check whether there are any missing or abnormal data markers in the process record. If so, it is directly classified as a data missing cause. If the data is complete and there are no abnormalities, the trend of feature matching results is further compared with the actual scoring trend. If the two trends are inconsistent, it is classified as a rule comprehension deviation cause. Finally, the scoring deviation of the same expert for multiple suppliers on the same evaluation item is statistically analyzed. If all deviations show the same direction and exceed the preference threshold, it is classified as an expert preference pattern cause. Finally, all classification results are written into the difference review report.

[0062] This approach involves a progressive, layered investigation, following the order of data, rules, and experts. Each layer has clear and quantifiable criteria, allowing for precise identification of the specific level at which discrepancies arise. Data issues are addressed by optimizing data sources, rule issues by refining matching logic, and expert preference issues by recording appropriate adjustments. Each type of cause corresponds to a dedicated optimization path, ensuring that optimization measures directly target the root cause and improve the return on investment. All attribution decisions are based on objective data recorded during the evaluation process, such as anomaly markers, feature matching results, and historical scoring data. These are all objective records retained by the system, ensuring high reliability of the attribution results. Furthermore, the entire attribution process is traceable, allowing for direct verification of the corresponding process records when there are doubts about the attribution results.

[0063] In a preferred embodiment, after generating the difference review report based on the difference cause classification results, the process also includes: When the reason for the discrepancy is missing data, data validation enhancement suggestion data is generated, which includes configuration parameters to enhance the dual data validation mechanism; When the reason for the difference is a deviation in rule understanding, rule optimization suggestion data is generated. The rule optimization suggestion data includes the feature matching rules and dimension weight parameters that need to be adjusted. When the reason for the difference is due to the expert preference pattern, preference pattern record data is generated. The preference pattern record data includes expert identifier, evaluation item identifier and preference offset data. Data validation enhancement suggestions, rule optimization suggestions, and preference pattern records are stored in the optimization suggestion database for use in subsequent evaluation rounds.

[0064] Specifically, for reasons related to missing data, data validation enhancement suggestions are generated, including optimized configuration parameters for a dual data validation mechanism; for reasons related to rule comprehension bias, rule optimization suggestions are generated, including feature matching rules and dimension weight parameters that need adjustment; for reasons related to expert preference patterns, preference pattern record data is generated, including expert identifiers, evaluation item identifiers, and preference offset data. All optimization data is stored in a structured format in the optimization suggestion database, and the corresponding parameters can be directly called in subsequent evaluation projects to achieve automatic optimization and adaptation of evaluation rules.

[0065] This approach generates specific optimization measures for three different types of discrepancies. Each measure is precisely matched to the root cause of the problem. Data-level issues are reinforced with a robust verification mechanism, rule-level issues are adjusted with matching logic, and expert-level issues are recorded and adapted to ensure that the optimization measures effectively address the corresponding discrepancies and further improve optimization results. All optimization suggestions and parameters are stored in a structured data format in a database, forming a reusable optimization experience library. When conducting evaluations of similar projects in the future, the optimized rules and parameters can be directly called upon, eliminating the need for re-analysis and adjustment each time. This ensures that the results of each review continue to be valuable, steadily improving the system's evaluation capabilities and creating a virtuous cycle of positive iteration.

[0066] Another aspect of the present invention provides a bidding assistance evaluation and difference review system based on the thinking chain, the system comprising: The evaluation criteria parsing module is used to obtain the procurement documents and identify the evaluation criteria table in the procurement documents, extract evaluation item data based on the evaluation criteria table and output the evaluation criteria data. The data query module is used to obtain the response data of each supplier. The response data includes quotation data, deviation data, performance data, certification data and complete text of tender documents. Based on the response data, the module outputs a supplier response dataset. The thought chain generation module divides all evaluation item data into objective evaluation items and subjective evaluation items based on the evaluation standard data. It generates an objective evaluation thought chain based on the objective evaluation items and a subjective evaluation thought chain based on the subjective evaluation items. The comprehensive evaluation module, based on the supplier response dataset and combining objective and subjective evaluation thought processes, obtains comprehensive evaluation data for each participating supplier and generates evaluation ranking results. The discrepancy review module acquires the actual scoring data entered by the evaluation experts and performs discrepancy analysis based on the comprehensive evaluation data to generate a discrepancy review report.

[0067] It should be understood that the specific embodiments described above are merely illustrative or explanatory of the principles of the invention and do not constitute a limitation thereof. Therefore, any modifications, equivalent substitutions, improvements, etc., made without departing from the spirit and scope of the invention should be included within the protection scope of the invention. Furthermore, the appended claims are intended to cover all variations and modifications falling within the scope and boundaries of the appended claims, or equivalent forms of such scope and boundaries.

Claims

1. A method for auxiliary evaluation and variance review in bidding and tendering based on thought chain, characterized in that, The method includes: Obtain the procurement documents and identify the evaluation criteria table in the procurement documents; extract evaluation item data based on the evaluation criteria table and output the evaluation criteria data. Obtain the response data of each supplier, which includes quotation data, deviation data, performance data, certification data and complete text of tender documents, and output supplier response dataset based on the response data; Based on the evaluation standard data, all the evaluation item data are divided into objective evaluation items and subjective evaluation items. An objective evaluation thinking chain is generated based on the objective evaluation items, and a corresponding subjective evaluation thinking chain is generated based on the subjective evaluation items. Based on the supplier response dataset and the objective and subjective evaluation thought processes, comprehensive evaluation data for each participating supplier is obtained and an evaluation ranking result is generated. Obtain the actual scoring data entered by the evaluation experts and conduct a difference analysis based on the comprehensive evaluation data to generate a difference review report.

2. The method according to claim 1, characterized in that, The process of obtaining procurement documents and identifying the evaluation criteria table of the procurement documents, extracting evaluation item data based on the evaluation criteria table, and outputting evaluation criteria data includes: Obtain the Word document of the procurement document and identify all table objects in the Word document; Based on each of the table objects, the text data of each cell in the first row is read, and the text data is matched with the preset evaluation standard table header feature word library to obtain the matching degree data; When the matching data exceeds a preset matching threshold, the corresponding table object is marked as an evaluation standard table; Extract the evaluation item name data, score data, evaluation details data, and weight allocation data row by row from the evaluation criteria table; The evaluation criteria data are output based on the extracted data.

3. The method according to claim 2, characterized in that, The method further includes: Based on the evaluation method field data in the evaluation standard table, the evaluation method type is identified through preset keyword matching rules; When the assessment method is a comprehensive assessment method, technical weight data, business weight data, and price weight data are extracted from the assessment standard table; When the evaluation method is the lowest valuation method, the price weight data is set to the preset maximum value, and the technical weight data and business weight data are set to zero. Write the evaluation method type data and the corresponding weight data into the evaluation standard data.

4. The method according to claim 1, characterized in that, The generation of the objective evaluation thought chain based on the objective evaluation items includes: For each of the objective evaluation items, the corresponding quantitative evaluation rule type is determined based on its evaluation details data. The quantitative evaluation rule types include numerical comparison type, validity period verification type, and deviation detection type. Based on the quantitative evaluation rule type, a corresponding objective item evaluation thinking chain is generated. The objective item evaluation thinking chain includes data acquisition nodes, data verification nodes, rule matching nodes, and conclusion output nodes arranged in sequence.

5. The method according to claim 4, characterized in that, The process of generating the corresponding subjective evaluation thought chain based on the subjective evaluation item includes: For each of the subjective evaluation items, load the corresponding evaluation rule template according to its evaluation details data; Based on the evaluation dimensions defined in the evaluation rule template, a corresponding subjective item evaluation thought chain is generated. The subjective item evaluation thought chain includes text location nodes, feature extraction nodes, dimension matching nodes, weighted calculation nodes, and interval output nodes arranged in sequence.

6. The method according to claim 5, characterized in that, The comprehensive evaluation data for each participating supplier is obtained by combining the supplier response dataset with the objective evaluation thought process and the subjective evaluation thought process. The processing of objective evaluation items includes: According to the node sequence defined in the objective evaluation thought chain, each processing node is executed sequentially. For numerical comparison-type evaluation items, first numerical data is extracted from the text layer data of the complete text of the tender document, and second numerical data is extracted from the optical character recognition result data of the complete text of the tender document. The first numerical data and the second numerical data are compared for consistency. Based on the comparison results, valid numerical data is determined. The valid numerical data is compared with the preset threshold data to generate score data. For validity period verification-type assessment items, the validity period expiration date data is extracted from the certification certificate data, the current system date data is obtained, the validity period expiration date data is compared with the current system date data, and the score data is generated based on the comparison result; For deviation detection type evaluation items, the deviation data is scanned based on the preset negative deviation keyword matching rules to detect whether there are negative deviation records, and score data is generated based on the detection results; The evaluation calculation process for each objective evaluation item generates evaluation process record data.

7. The method according to claim 6, characterized in that, The comprehensive evaluation data for each participating supplier is obtained by combining the supplier response dataset with the objective evaluation thinking chain and the subjective evaluation thinking chain. The processing of subjective evaluation items specifically includes: According to the node sequence defined in the subjective evaluation thought chain, each processing node is executed sequentially. At the text location node, the text paragraph data corresponding to each evaluation dimension is located from the complete text of the tender document according to the text paragraph location rules corresponding to each evaluation dimension. At the feature extraction node, content features are extracted from the text paragraph data based on a preset feature lexicon and semantic matching rules to generate feature matching result data. At the dimension matching node, the feature matching result data is judged according to the matching judgment criteria of each evaluation dimension to generate matching degree data for each dimension. At the weighted calculation node, based on the matching degree data of each dimension and the preset dimension weight parameters, suggested score range data is generated through weighted calculation. The suggested score range data includes suggested lower limit value and suggested upper limit value. At the output node of the interval, the feature matching result data and dimension weight parameters of each evaluation dimension are assembled into evaluation reason data.

8. The method according to claim 7, characterized in that, The process of obtaining the actual scoring data entered by the evaluation experts and conducting a difference analysis based on the comprehensive evaluation data to generate a difference review report includes: Receive actual scoring data entered by evaluation experts through the interactive interface; For each evaluation item, calculate the difference between the auxiliary suggestion score and the actual score. Filter out the evaluation items whose difference values ​​exceed a preset difference threshold, and mark them as evaluation items to be reviewed; For the items to be reviewed and evaluated, the causes of the differences are classified based on preset difference attribution rules; The difference review report is generated based on the difference cause classification results. The difference review report includes a list of difference items, difference value data, difference cause classification results, and optimization suggestion data.

9. The method according to claim 8, characterized in that, The classification of the causes of differences based on preset difference attribution rules includes: For the evaluation item to be reviewed, obtain the evaluation process record data corresponding to the evaluation item; Based on the preset data integrity detection rules, check whether there are missing or abnormal markers in the data source of the data recorded in the evaluation process. When there are missing or abnormal markers, the reasons for the difference are classified as data missing reasons. When the data source in the evaluation process record data is complete and there are no abnormal markers, based on the preset rule matching deviation detection rule, the feature matching result data in the evaluation process record data is compared with the actual scoring data. When the feature matching result data is inconsistent with the actual scoring trend, the reason for the difference is classified as the reason for rule understanding deviation. When the difference values ​​of the same evaluation item across multiple suppliers all show the same direction of deviation, and the deviation exceeds the preset preference threshold, the reason for the difference will be classified as the reason for expert preference mode. Write the data of the differences and their causes into the difference review report.

10. A bidding and tendering auxiliary evaluation and difference review system based on thinking chain, characterized in that, The system includes: The evaluation criteria parsing module is used to obtain the procurement documents and identify the evaluation criteria table of the procurement documents, extract evaluation item data based on the evaluation criteria table and output the evaluation criteria data. The data query module is used to obtain the response data of each supplier. The response data includes quotation data, deviation data, performance data, certification data and complete text of tender documents. Based on the response data, the module outputs a supplier response dataset. The thought chain generation module divides all the evaluation item data into objective evaluation items and subjective evaluation items based on the evaluation standard data, generates an objective item evaluation thought chain based on the objective evaluation items, and generates a subjective item evaluation thought chain based on the subjective evaluation items. The comprehensive evaluation module, based on the supplier response dataset and combining the objective and subjective evaluation thought processes, obtains comprehensive evaluation data for each participating supplier and generates evaluation ranking results. The discrepancy review module acquires the actual scoring data entered by the evaluation experts and performs discrepancy analysis based on the comprehensive evaluation data to generate a discrepancy review report.

Images