Large model native driven bidding full-link intelligent management and control system

Abstract

The application discloses a large model native driving whole-link intelligent management and control system for bidding, relates to the technical field of electronic bidding data processing, and performs structured processing on a project objective file set corresponding to a same bidding project or a same bid section, extracts objective requirements capable of constraining supplier response content, and constructs an objective anchor point set; a large model is called to analyze each supplier electronic bidding file, response content is mapped to the objective anchor point, response action parameters and response support parameters are extracted, and an anchor point response unit is generated; the anchor point response unit is connected in the order of the objective anchor point to form a response chain sequence, a dynamic time warping algorithm with anchor point gating constraints is used to calculate semantic difference cost, and response chain isomorphism risk is identified; and according to review results, a semantic difference cost function, an abnormal response fragment library and an isomorphism threshold are corrected, so that the method can identify rewritten accompanying bidding risk with low literal similarity but high response logic isomorphism.

Description

Technical Field

[0001] This invention relates to the field of electronic bidding and tendering data processing technology, and more specifically, to a large-model native-driven intelligent management and control system for the entire bidding and tendering process. Background Technology

[0002] In electronic bidding scenarios, tenderers, agents, or regulatory platforms typically need to identify similarity risks in electronic bid documents submitted by different suppliers to detect risks such as direct copying, template application, duplicate attachments, and abnormally consistent responses. Existing technologies can detect full-text duplication rate, keyword overlap, paragraph edit distance comparison, template similarity analysis, attachment hash consistency detection, and fixed-format material comparison. These methods can, to a certain extent, detect explicit similarities such as large sections of text being completely identical, key paragraphs being highly repetitive, fixed templates being identical, or attachment materials being completely identical. They are suitable for risk identification scenarios such as directly copied bid documents, bid documents with the same template that have not been significantly modified, and duplicate submissions of attachments.

[0003] The existing technology has the following shortcomings: Suppliers may reduce the literal similarity of their bid documents by using methods such as paraphrasing, paragraph rearrangement, title replacement, table splitting, attachment repositioning, merging multiple responses, or adjusting supplementary commitments. This makes it difficult for traditional duplication rate detection and paragraph similarity detection to identify potential commonalities. In the same bidding project or the same bid section, different suppliers may have different wording, but they may form the same response actions, the same supporting evidence, the same incorrect citations, the same missing responses, the same unreasonable commitments, or the same unnecessary limitations under the same bidding requirements. Existing algorithms usually lack a response positioning mechanism based on the objective requirements of the bidding, and they also lack a process for structured decomposition, response chain sorting, and isomorphic alignment of the bid response content. As a result, the risk of rewritten dummy bids with low literal similarity but high response chain isomorphism is difficult to detect accurately.

[0004] To address the above problems, this invention proposes a solution. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a large-model natively driven intelligent management and control system for the entire bidding and tendering process, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The large-scale model-driven intelligent management and control system for the entire bidding process includes: an anchor point construction module, a response unit generation module, a homogeneous risk identification module, and a cost function correction module, with signal connections between the modules. Anchor point construction module: Obtain the set of objective documents of the project and perform document preprocessing to obtain the set of structured text of the project. Perform explicit similarity identification processing to generate explicit similarity risk results and supplier pairs to be analyzed. Extract objective requirements and generate an objective anchor point set by combining rule positioning, large model verification and minimum response unit decomposition. Generate an objective anchor point identifier and anchor point hierarchy relationship for each objective anchor point. Response Unit Generation Module: Calls the large model to perform semantic parsing on the electronic bid documents of each supplier. Through candidate recall, semantic judgment, main response determination and supplementary response merging, the bid content is mapped to the set of objective anchor points. The response action parameters and response support parameters are extracted and combined to generate anchor point response units, and consistency verification and deduplication merging are performed. Homogeneity Risk Identification Module: Connects anchor response units into a response chain sequence according to the objective anchor order. Uses a dynamic time warping algorithm with anchor gating constraints to perform non-equal length alignment on the response chain sequences of any two suppliers, calculates the semantic difference cost, and identifies abnormal response segments with the same missing response, the same incorrect reference, the same unreasonable commitment, and the same unnecessary limitation. When the minimum alignment cost is lower than the homogeneity threshold and there is at least one abnormal response segment, a rewritten sham bidding risk result is generated. Cost function correction module: Receives the verification results, marks the anchor response units that are confirmed to be identical sample pairs as identical sample pairs, marks the normally similar ones as normal sample pairs, uses a metric learning algorithm, corrects the semantic mapping model in the semantic difference cost function through the triple loss function, and updates the abnormal response fragment library, the normal similar fragment library and the isomorphism threshold.

[0007] In a preferred embodiment, the anchor point construction module includes the following steps: Document preprocessing includes document format recognition, text layer extraction, table content restoration, directory level recognition, paragraph boundary segmentation, attachment reference relationship recognition, page number positioning, clause number recognition, and document version identifier generation. The explicit similarity detection process includes detecting duplicate sentences, keyword overlap, template similarity, paragraph edit distance, and attachment hash consistency in the electronic bid documents of different suppliers; If two tender documents meet at least one of the following conditions: large sections of text are completely identical, fixed templates are completely identical, the repetition rate of key paragraphs exceeds the explicit similarity threshold, or the attachment materials have the same hash value, then an explicit similarity risk result will be generated. If the explicit similarity threshold is not reached, the two corresponding suppliers will be identified as supplier pairs to be analyzed. Rule location is used to identify clauses with explicit numbering, headings, table fields, or mandatory expressions from a collection of structured text for a project; Large model validation is used to determine whether the text obtained from rule-based localization constitutes an objective requirement that the supplier must respond to; Minimal response unit splitting is used to break down clauses containing multiple objective requirements into multiple objective anchors; Objective anchor point identifiers should include at least the project number, bid section number, source document type, source chapter, clause number, anchor point type, anchor point text, and anchor point sequence number; Anchor point hierarchy is used to determine the primary anchor point and parent anchor point.

[0008] In a preferred embodiment, the response unit generation module includes the following steps: Candidate recall is based on clause number, title keywords, technical parameter names, delivery deadline expressions, acceptance terms, qualification terms, attachment references, and semantic similarity judgment of the large model, and selects fragments related to objective anchor points from the candidate response fragment set; Semantic determination is used to confirm whether a candidate response fragment constitutes a substantive response to an objective anchor point; The main response determination rule is: select the segment that corresponds to the objective anchor clause number or title; If none exist, select the segment containing a clear action verb and a response object; if multiple segments still exist, select the segment that contains both the response action and the supporting content. The response action parameters include action type, action object, action condition, and action result; Supporting parameters for the response include the source, supporting documentation, performance conditions, and commitment boundaries; consistency verification includes: If the response action parameter is empty but the response support parameter exists, it is marked as a response unit with missing action. If the response action parameters exist but the response support parameters are empty, it is marked as an insufficiently supported response unit. If both are empty, mark it as a missing response unit; If both exist, mark it as a complete response unit; Deduplication includes incorporating duplicate response content with the same objective anchor point into the main response unit, and recording the source of the same segment when the same bid content corresponds to multiple objective anchor points.

[0009] In a preferred embodiment, the isomorphism risk identification module includes the following steps: If two anchor point response units originate from the same objective anchor point, parent-child anchor point, or the same anchor point cluster, they are allowed to enter the alignment candidate. If the source is an irrelevant anchor, the semantic difference cost is set to a prohibited matching value or the matching cost is increased. The dynamic time warping algorithm calculates the minimum alignment cost between two response chain sequences recursively. During the recursion process, the current alignment cost is obtained by adding the semantic difference cost between the current two anchor response units to the minimum cumulative alignment cost of the previous step. The rules for identifying abnormal response fragments include: If two suppliers fail to respond effectively under the same objective anchor point, it is recorded as an identical missing response. References to the same error standard, error clause number, or obviously inapplicable specifications are recorded as identical error references. Making the same commitments that exceed the requirements of the tender documents and are unnecessary shall be recorded as the same unreasonable commitments. Adding the same non-essential constraint, even if the content is not required by the tender document, is considered as the same non-essential constraint. The isomorphism threshold is determined based on the final minimum alignment cost distribution of confirmed identical sample pairs and normal sample pairs in the historical review samples.

[0010] In a preferred embodiment, the cost function correction module includes the following steps: The review results include the review and confirmation of suspected supplier pairs, abnormal response segments, objective anchor point locations, original tender documents, alignment paths, and risk statements; Anchor response units that are confirmed to be rewritten dummy targets, highly similar, or have a common source rewriting relationship are marked as identical sample pairs; Anchor point response units that are normally similar due to the same tender document, the same technical specification, the same industry standard, or the same legal format will be marked as normal sample pairs. The metric learning algorithm uses a triplet loss function to correct the semantic mapping model: Construct triplet samples from the review results, including benchmark anchor response units, identical anchor response units, and normal anchor response units, so that the distance between identical sample pairs in the semantic mapping space is less than the distance between normal sample pairs, and they are separated by a constant interval. The abnormal response fragments in the verified identical samples are written into the abnormal response fragment library, the corresponding expressions in the normal samples excluded from verification are written into the normal similar fragment library, and the isomorphism threshold is updated according to the final minimum alignment cost distribution of the newly added verification samples.

[0011] The technical effects and advantages of the large-model natively driven intelligent management and control system for the entire bidding and tendering process of this invention are as follows: By constructing an objective anchor point set, the objective requirements in bidding documents, technical specifications, contract delivery terms, acceptance standards, and qualification documents are transformed into a unified basis for response comparison. This allows bids from different suppliers to be positioned and analyzed around the same objective requirements, avoiding omissions caused by judging solely based on full-text duplication rate, keyword overlap, or paragraph editing distance. Through large-scale model generation of anchor point response units, scattered paragraphs, tables, appendix descriptions, and commitment statements in the bid documents are transformed into structured objects containing response action parameters and response support parameters. This provides a calculable, sortable, and alignable analytical basis for the bid response content. The algorithm employs a dynamic time warping method with anchor-gated constraints to identify response chain isomorphism risks among different suppliers, even when bid documents have undergone text rewriting, paragraph rearrangement, title changes, attachment repositioning, or multi-anchor-point merging responses. These risks stem from variations in response actions, supporting evidence, missing items, incorrect citations, and unnecessary limitations. By introducing a semantic difference cost function based on review results, an abnormal response fragment library, a normal similar fragment library, and an isomorphism threshold, the algorithm can gradually distinguish between abnormal similarities resulting from source rewriting and normal similarities caused by industry-standard expressions, legally mandated formats, or mandatory requirements of the bidding documents. This improves the accuracy and interpretability of identifying rewritten bid-rigging risks. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the structure of the intelligent management and control system for the entire bidding and tendering process driven by the large model natively, as described in this invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] Example Please see Figure 1 As shown, this invention discloses a large-model native-driven intelligent management and control system for the entire bidding and tendering process, including: an anchor point construction module, a response unit generation module, a homogeneous risk identification module, and a cost function correction module, with signal connections between the modules; Anchor point construction module: Obtain the set of objective documents of the project and perform document preprocessing to obtain the set of structured text of the project. Perform explicit similarity identification processing to generate explicit similarity risk results and supplier pairs to be analyzed. Extract objective requirements and generate an objective anchor point set by combining rule positioning, large model verification and minimum response unit decomposition. Generate an objective anchor point identifier and anchor point hierarchy relationship for each objective anchor point. Response Unit Generation Module: Calls the large model to perform semantic parsing on the electronic bid documents of each supplier. Through candidate recall, semantic judgment, main response determination and supplementary response merging, the bid content is mapped to the set of objective anchor points. The response action parameters and response support parameters are extracted and combined to generate anchor point response units, and consistency verification and deduplication merging are performed. Homogeneity Risk Identification Module: Connects anchor response units into a response chain sequence according to the objective anchor order. Uses a dynamic time warping algorithm with anchor gating constraints to perform non-equal length alignment on the response chain sequences of any two suppliers, calculates the semantic difference cost, and identifies abnormal response segments with the same missing response, the same incorrect reference, the same unreasonable commitment, and the same unnecessary limitation. When the minimum alignment cost is lower than the homogeneity threshold and there is at least one abnormal response segment, a rewritten sham bidding risk result is generated. Cost function correction module: Receives the verification results, marks the anchor response units that are confirmed to be identical sample pairs as identical sample pairs, marks the normally similar ones as normal sample pairs, uses a metric learning algorithm, corrects the semantic mapping model in the semantic difference cost function through the triple loss function, and updates the abnormal response fragment library, the normal similar fragment library and the isomorphism threshold.

[0015] In the anchor point construction module, the objective document set of the project is obtained and preprocessed to obtain a structured text set of the project. Explicit similarity identification is performed, generating explicit similarity risk results and supplier pairs to be analyzed. Objective requirements are extracted, and an objective anchor point set is generated using a combination of rule-based localization, large model validation, and minimum response unit decomposition. For each objective anchor point, an objective anchor point identifier and anchor point hierarchy are generated. Specific content includes: Obtain the set of objective project documents corresponding to the same bidding project or the same bid section. The set of objective project documents is used to provide an objective basis for comparing the content of suppliers' bid responses and serves as the data source for generating anchor response units and response chain sequences in the future. For example, the set of objective documents for a project may include tender documents, technical specifications, contract delivery terms, acceptance criteria, qualification documents, and electronic tender documents submitted by each supplier; wherein, the tender documents are used to provide the procurement object, procurement scope, functional requirements, and business scope; the technical specifications are used to provide technical parameters, performance indicators, interface conditions, compatibility requirements, and applicable standards; the contract delivery terms and acceptance criteria are used to provide the delivery cycle, deliverables, acceptance conditions, response time limits, and service boundaries; the qualification documents are used to provide supplier qualifications, performance conditions, certification requirements, personnel configuration requirements, and equipment capability requirements; and the electronic tender documents are used to provide the supplier's response to the above objective requirements; After obtaining the objective document set of the project, the objective document set of the project is preprocessed to obtain the structured text set of the project. Document preprocessing includes document format recognition, text layer extraction, table content restoration, directory hierarchy recognition, paragraph boundary segmentation, attachment reference relationship recognition, page number positioning, clause number recognition, and document version identifier generation. For electronic bid documents containing scanned copies, image-based tables, or compressed attachment packages, they are converted into searchable text while retaining the original page numbers, chapter titles, table row and column positions, attachment names, and file hash values. For cases where the same supplier submits multiple versions of bid documents, the last valid version before the bid deadline is used as the main document version, and the remaining versions are used as version difference reference data. After obtaining the structured text set of the project, the existing electronic tender documents are subjected to similarity identification processing to obtain explicit similarity risk results and supplier pairs to be analyzed; Specifically, the electronic tender documents from different suppliers are tested for duplicate sentences, keyword overlap, template similarity, paragraph edit distance, and attachment hash consistency. If two tender documents contain large sections of identical text, have completely identical fixed templates, have a key paragraph repetition rate exceeding the preset explicit similarity threshold, or have attachment materials with the same hash value, then an explicit similarity risk result will be generated. If two tender documents do not reach the explicit similarity threshold, they will not be directly deemed to be free of similarity risk. Instead, they will be identified as supplier pairs to be analyzed and will proceed to the subsequent objective anchor point construction and response chain isomorphism identification process. Extract objective requirements that can constrain the content of supplier bid responses from the project's structured text set, and generate an objective anchor point set. The objective anchor point set is the set of objective requirements that suppliers need to respond to explicitly in the current bidding project or the current bid section. For example, the set of objective anchor points may include bidding requirements, technical constraints, delivery and acceptance items, and qualification boundary items. Bidding requirements are used to indicate the procurement object, procurement scope, functional requirements, and business application scope; technical constraints are used to indicate technical parameters, performance indicators, interface conditions, compatibility requirements, and applicable standards; delivery and acceptance items are used to indicate the delivery cycle, delivery results, acceptance conditions, response time limit, and service boundaries; qualification boundary items are used to indicate supplier qualifications, performance conditions, certification requirements, personnel configuration, and equipment capability requirements. It should be noted that the tender requirements, technical constraints, delivery and acceptance items and qualification boundary items are only used to form an objective anchor set. The core calculation objects are anchor response units, response chain sequences and semantic difference costs, which can make full use of the objective requirements in the tender documents, technical specifications, delivery and acceptance clauses and qualification conditions documents. When extracting the set of objective anchor points, a combination of rule-based localization, large model verification, and minimum response unit decomposition is adopted. Rule localization is used to identify clauses with explicit numbering, headings, table fields, or mandatory expressions from a set of structured text for a project, such as scope of supply, technical parameters, delivery deadlines, acceptance requirements, eligibility criteria, response time, and compatibility standards. Large model validation is used to determine whether the text obtained through rule-based localization constitutes an objective requirement that the supplier must respond to; If a certain text is only part of the project background, policy introduction, definition explanation or non-mandatory prompt, it will not be used as an objective anchor point. The smallest response unit is used to handle the situation where a clause contains multiple objective requirements at the same time. For example, if a clause contains interface requirements, delivery deadline requirements, and acceptance report requirements, it is broken down into multiple objective anchors, so that each objective anchor can be individually matched with the specific response content in the tender document. Generate an objective anchor point identifier for each objective anchor point; An objective anchor identifier must include at least the project number, bid section number, source document type, source chapter, clause number, anchor type, anchor text, and anchor sequence number. The anchor sequence number is determined according to the business logic order in the project's objective document set. If different files contain the same, inclusive, or related objective requirements, the more specific objective requirement is used as the primary anchor point, and the higher-level objective requirement is used as the parent anchor point, forming an anchor point hierarchy. This hierarchy is used for anchor point gating constraints in subsequent steps. Specifically, anchor point response units under the same anchor point, parent-child anchor points, or the same anchor point cluster are allowed to enter the alignment candidate, while anchor point response units under unrelated anchor points are not allowed to match directly. Output the set of objective anchor points, denoted as: ;in, This refers to the set of objective anchor points for the current bidding project or the current bid section. For the first An objective anchor point; The number of objective anchor points; This is an objective anchor point number, with a value range of one to... Each objective anchor point All include objective anchor point identifiers, source documents, clause locations, anchor point text, anchor point types, anchor point sequence numbers, and anchor point hierarchical relationships. The set of objective anchor points is used in subsequent steps to locate, extract in a structured manner, and sort the response chain in the supplier's electronic bid documents. In this step, the explicit similarity threshold is determined by the repetition rate distribution of historically confirmed direct copy type similar samples and normal non-similar samples. Specifically, the lower quantile of the duplication rate of directly copied identical file pairs and the upper quantile of the duplication rate of normal file pairs in the historical samples are statistically analyzed, and the boundary value between the two that can take into account both the false negative rate and the false positive rate is selected as the explicit duplication threshold. If the historical samples are insufficient, the explicit duplication threshold is initialized with the first batch of manually labeled samples and continuously corrected in the subsequent review and feedback process. This step establishes an objective anchor point set for the current bidding project or bid section based on the objective document set of the project. It retains the explicit similarity risks that can be identified by the existing electronic bid document similarity identification capabilities. By establishing an objective positioning basis between the bidding project documents and the bid response content through the objective anchor point set, subsequent steps can compare the response logic of different suppliers around the same objective requirements, thereby providing a data foundation for identifying rewritten dummy bid documents with low literal similarity but high response chain isomorphism. This step structures the set of objective documents for the same bidding project or the same bid section, and extracts objective requirements that can constrain the supplier's response from the bidding documents, technical specifications, contract delivery terms, acceptance standards, qualification documents, and electronic bid documents. This forms an objective anchor point set with objective anchor point identifiers, clause locations, anchor point types, anchor point sequence numbers, and anchor point hierarchical relationships. This establishes a unified positioning basis between the objective requirements of the bidding project and the supplier's bid response, enabling subsequent steps to compare the response logic of different suppliers around the same objective anchor point. It also provides a data foundation for the response unit generation module to generate anchor point response units and for the isomorphic risk identification module to set anchor point gating constraints.

[0016] In the response unit generation module, a large model is invoked to perform semantic parsing on the electronic bid documents of each supplier. Through candidate recall, semantic judgment, main response determination, and supplementary response merging, the bid content is mapped to an objective anchor point set. Response action parameters and response support parameters are extracted and combined to generate anchor point response units, and consistency verification and deduplication merging are performed. Specific content includes: The large model is invoked to perform semantic parsing on the electronic bid documents submitted by each supplier, and the paragraphs, tables, attachment descriptions, supporting material descriptions and commitment statements in the bid documents are mapped to the set of objective anchor points constructed by the anchor point construction module. Through the anchor point response unit, the bid documents are transformed from natural language text into a computable, sortable and alignable structured response object. Let the set of suppliers participating in the current bidding project or the current bid section be: ;in, For supplier aggregation; For the first One supplier; Number of suppliers; The supplier serial number has a range of values. to For any supplier First, recall response fragments from the electronic bid documents to form a set of candidate response fragments; It should be noted that candidate response fragments may include body paragraphs, technical parameter tables, service commitment tables, qualification certificate descriptions, attachment index descriptions, contract response descriptions, acceptance response descriptions, and supplementary commitment descriptions. For each candidate response fragment, the text content, chapter title, page number, paragraph number, table location, attachment source, context boundaries, and citation relationships should be preserved. For each objective anchor point In the supplier The search process involves finding the response content corresponding to the objective anchor point from the set of candidate response fragments. The search process includes four sub-steps: candidate recall, semantic determination, main response determination, and supplementary response merging. Candidate recall is used to filter fragments that may be related to the objective anchor from the set of candidate response fragments. Candidate recall can be completed by combining clause number, title keywords, technical parameter names, delivery deadline expressions, acceptance terms, qualification terms, attachment reference relationships, and semantic similarity judgment of the large model. Semantic judgment is used to confirm whether the candidate response fragment constitutes a substantial response to the objective anchor. Substantial response means that the fragment can explain what actions the supplier takes in response to the objective anchor, or explain the basis, conditions, materials, or commitment boundaries used by the supplier to meet the objective anchor. If a fragment only repeats the original text of the tender document without reflecting the supplier's own actions, basis, or commitment boundaries, it is marked as a low-substantial response fragment. When an objective anchor point corresponds to multiple candidate response fragments, the main response content is determined. The main response content is the fragment that has the most direct semantic relationship with the objective anchor point and can independently explain the supplier's response action. The rule for determining the main response content is: select the fragment that corresponds to the objective anchor point's clause number or title. If no corresponding segment with a number or title exists, select the segment containing a clear action verb and a response object; If multiple candidates still exist, select the segment that contains both the response action and the response support content. If multiple fragments meet the conditions, the fragment with the chapter level closest to the anchor point type is selected, and the remaining fragments are incorporated into the response support parameters as supplementary response content. If no response content corresponding to a certain objective anchor point is found in the tender document, a missing response marker is generated. After determining the response content, the large model is used to extract two extended parameters from the main response content and the supplementary response content: response action parameters and response support parameters. The response action parameter is used to represent the technical actions, implementation actions, service actions, management actions, or configuration actions actually proposed by the supplier for this objective anchor point; The response action parameters can be further composed of four fields: action type, action object, action condition, and action result. Action type indicates the category of actions such as adoption, configuration, delivery, testing, maintenance, training, acceptance cooperation, and quality control; action object indicates the equipment, service, interface, personnel, material, or process affected by the action; action conditions indicate the time, scenario, standards, qualifications, or resource conditions required for the action to occur; action result indicates the expected direct result of the action. The response support parameters are used to indicate the basis, conditions and commitment boundaries for the supplier to perform the response action. The response support parameters can be composed of four fields: source of basis, supporting materials, performance conditions and commitment boundaries. The source of information refers to the standards, specifications, technical documents, internal policies, or contract terms cited by the supplier; supporting documents refer to the certificates, performance records, personnel information, equipment information, test reports, or supporting materials provided by the supplier; performance conditions refer to the personnel, equipment, time, processes, or resources on which the response action depends; commitment boundaries refer to the delivery results, acceptance results, service response results, quality assurance results, or qualification compliance results that the supplier promises to achieve. The response support parameters combine the previously scattered sources of information and commitments into a unified result. The objective anchor point identifier, response action parameters, and response support parameters are combined to generate the anchor point response unit for the supplier. For objective anchor points The resulting anchor point response unit is denoted as The data structure of the anchor point response unit is as follows: ;in, For suppliers For the An objective anchor point The resulting anchor point response unit; For the first An objective anchor point identifier; For suppliers For objective anchor points Response action parameters; For suppliers For objective anchor points The response support parameters, objective anchor point identifiers are used for positioning and sorting, and response action parameters and response support parameters are used for subsequent semantic difference cost calculation; To ensure that the generated results of the large model meet the requirements of being publicly available and executable, consistency verification is performed on the anchor point response units; If the response action parameter is empty but the response support parameter exists, mark the anchor point response unit as a response unit with missing action, indicating that the supplier has provided proof, materials or commitments but has not specified the specific response action. If the response action parameters exist but the response support parameters are empty, mark it as an insufficiently supported response unit, indicating that the supplier has proposed a response action but has not provided the basis, conditions, or commitment boundaries; if both are empty, mark it as a missing response unit. If both exist, mark it as a complete response unit; Deduplication and merging of anchor response units within the same supplier; If the same objective anchor point is repeatedly responded to in multiple chapters of the tender document, the main anchor point response unit is generated with the main response content as the core, and the supplementary content is incorporated into the response support parameters. If the same section of the tender content responds to multiple objective anchor points at the same time, the main anchor point and subordinate anchor points are determined according to the anchor point hierarchy. When the content has a substantial response to multiple anchor points, it is allowed to generate multiple anchor point response units, but the same fragment source is recorded in each response unit to avoid repeatedly amplifying the influence of the fragment in subsequent isomorphic identification. The output of this step is a set of anchor response units for each supplier. Each anchor response unit has a clear source, clear fields, clear response status, and traceable original text. This structured result solves the problem that the paragraphs of the tender documents in the existing algorithm cannot correspond to the tender requirements. It also solves the problem that the response logic cannot be compared after the literal similarity is reduced due to the rewriting of the large model. This step, based on the set of objective anchor points, calls a large model to parse the electronic tender documents of each supplier. Through candidate recall, semantic judgment, main response determination, and supplementary response merging, it extracts response action parameters and response support parameters to generate anchor response units with source, field, status, and original text backtracking relationship. This transforms the scattered natural language response content into a computable, sortable, and alignable structured object, enabling subsequent comparison of the response actions and supporting evidence of different suppliers under the same objective anchor point. It also provides the smallest analysis unit for the isomorphic risk identification module to construct response chain sequences and calculate semantic difference costs.

[0017] In the isomorphic risk identification module, anchor response units are connected into a response chain sequence according to the objective anchor point order. A dynamic time warping algorithm with anchor point gating constraints is used to perform non-equal length alignment on the response chain sequences of any two suppliers. The semantic difference cost is calculated, and anomalous response segments with identical missing responses, identical misreferences, identical unreasonable promises, or identical unnecessary constraints are identified. When the minimum alignment cost is below the isomorphic threshold and at least one anomalous response segment exists, a rewritten sham bidding risk result is generated, specifically including: The anchor response units are connected according to the order of the anchors in the objective anchor set to construct the response chain sequence of each supplier. The dynamic time warping algorithm with anchor gating constraints is used to identify whether there is a risk of response chain isomorphism between different supplier bid documents. For any supplier It reads all anchor response units from its anchor response unit set and follows the objective anchor set in the anchor construction module. Arranged in order; If the anchor response unit corresponding to a certain objective anchor point is a complete response unit, then it is directly included in the response chain sequence; If the corresponding anchor point response unit is an action-missing response unit, a support-insufficient response unit, or a missing response unit, it is still included in the response chain sequence, and the corresponding response status mark is retained. Rewritten bid documents form a homogeneity with the complete response content and have the same omissions, the same insufficient support, the same incorrect references, or the same unreasonable promises under the same objective anchor points. If the abnormal or missing response units are directly deleted, important clues of common origin will be lost. Suppliers The response chain sequence is denoted as ;in, For suppliers The response chain sequence; For suppliers For objective anchor points Anchor point response unit; If there is no corresponding response content for a certain objective anchor point, the anchor point response unit at that position is a missing response unit, and each supplier's bid document is converted into a response chain sequence arranged according to the objective anchor points. For any two suppliers and its response chain sequence and When performing isomorphism identification, since different suppliers may circumvent ordinary plagiarism checks by rearranging paragraphs, rewriting titles, splitting content, swapping attachments, or merging responses with multiple anchor points, the number of effective response units and the expression organization of the two response chain sequences may not be completely consistent. A dynamic time warping algorithm is used to align non-equal length sequences, and an anchor gating constraint is introduced. The anchor gating constraint is used to control which anchor response units can enter the matching candidate. If two anchor point response units originate from the same objective anchor point, parent-child anchor point, or the same anchor point cluster, they are allowed to enter the alignment candidate; If two anchor response units originate from unrelated anchors, set their semantic difference cost to a prohibited matching value or significantly increase their matching cost. Let the response chain sequence of the first supplier be... The response chain sequence of the second supplier is as follows: ;in, The response chain sequence of the first supplier; For the response chain sequence of the second supplier; The first supplier One anchor point response unit; For the second supplier One anchor point response unit; and These represent the number of anchor response units in the two response chain sequences, respectively. and Given the unit indices in the two response chain sequences, calculate the semantic difference cost between the two anchor response units. The semantic difference cost is determined by the semantic mapping difference between the response action parameters and the response support parameters and is subject to objective anchor gating constraints. Specifically, if two anchor response units correspond to the same objective anchor point, and the action type, action object, action condition and action result in the response action parameters are highly similar, and the basis source, supporting materials, performance conditions and commitment boundaries in the response support parameters are also highly similar, then the semantic difference cost between the two is low. If two anchor response units correspond to unrelated objective anchors, or if the response actions and response support are significantly different, then the semantic difference between the two is costly. If both have the same missing response, the same incorrect reference, the same unreasonable commitment, the same unnecessary limitation, or the same abnormal attachment organization under the same objective anchor point, mark the pair of anchor point response units as candidates for abnormal response fragments. The dynamic time warping recursive process with anchor gating constraints is as follows: ;in, Before becoming the first supplier Each anchor point response unit and the second supplier before Minimum alignment cost between anchor response units; Anchor point response unit With anchor point response unit The cost of semantic differences between them; This represents the cumulative alignment cost when the first supplier's current anchor point response unit is skipped. This represents the cumulative alignment cost when the current anchor response unit of the second supplier is skipped; This represents the cumulative alignment cost when the response units of the two anchor points are matched. and Without passing the anchor point gating constraint, Set to a prohibited match value so that the dynamic time warping path will not prioritize this irrelevant match; After dynamic time warping, the minimum alignment path and final minimum alignment cost between the two response chain sequences are obtained. The final minimum alignment cost is used to represent the overall difference between the two supplier bids at the objective anchor point response chain level. Further checks are performed to determine if there are identical anomalous response segments in the minimum alignment path. The rules for identifying anomalous response segments include: If two suppliers fail to respond effectively under the same objective anchor point, it is recorded as an identical missing response. If two suppliers cite the same erroneous standard, erroneous clause number, or obviously inapplicable specification under the same objective anchor point, it is considered that they have the same erroneous reference. If two suppliers make the same commitments under the same objective anchor point that exceed the requirements of the tender documents and are unnecessary in business, it shall be recorded as the same unreasonable commitment. If two suppliers add the same brand, model, process, attachment organization method or non-essential limiting conditions under the same objective anchor point, and such content is not required by the tender document, it is recorded as the same non-essential limiting condition. If two suppliers lack action descriptions or supporting materials in the same way under the same objective anchor point, it is recorded as the same response defect. When the final minimum alignment cost is lower than the preset isomorphism threshold, and there is at least one identical abnormal response fragment in the minimum alignment path, a rewritten scrambling risk result is generated. The isomorphism threshold is determined by extracting confirmed identical sample pairs and confirmed normal sample pairs from historical review samples, calculating their final minimum alignment cost distributions, and selecting the threshold value that enables the recall rate of identical sample pairs to reach the preset requirements and the false alarm rate of normal sample pairs to be lower than the preset requirements as the isomorphism threshold. If the historical review samples are insufficient, the isomorphic threshold is initialized with the first batch of manually labeled samples and updated in the cost function correction module based on the newly added review results; This step connects the anchor response units of each supplier into a response chain sequence according to the objective anchor order. It then uses a dynamic time warping algorithm with anchor gating constraints to perform non-equal length alignment on the response chain sequences of different suppliers, calculates the semantic difference cost, and identifies abnormal response fragments such as identical missing responses, identical incorrect references, identical unreasonable promises, identical unnecessary limitations, and identical response defects. It can identify the risk of response chain isomorphism even when text is rewritten, paragraphs are rearranged, or attachments are swapped. Suspected supplier pairs, alignment paths, and abnormal response fragments are then passed to the cost function correction module for review and correction.

[0018] In the cost function correction module, the verification results are received, anchor response units confirmed as identical sample pairs are marked as identical sample pairs, and normally similar ones are marked as normal sample pairs. A metric learning algorithm is used to correct the semantic mapping model in the semantic difference cost function through a triplet loss function. Simultaneously, the abnormal response fragment library, the normal similar fragment library, and the isomorphism threshold are updated. Specific details include: The response chain isomorphism risk results are reviewed, and the review results are used to correct the semantic difference cost function in the isomorphism risk identification module. The analysis is conducted based on the output of suspected supplier pairs, the identified abnormal response fragments, the objective anchor point locations, the original tender documents, the alignment paths, and the risk statements. If the review confirms that two bids are rewritten dummy bids, highly similar, or have obvious rewriting relationships, the corresponding anchor point response units in the minimum alignment path will be marked as identical copy pairs. If the review determines that the two tender documents are only similar due to the same tender document, the same technical specification, the same industry standard or the same legal format, the corresponding anchor point response unit will be marked as a normal sample pair. If the verification fails to confirm, it will be marked as a sample to be observed. Samples to be observed will not participate in the semantic difference cost function correction in this round. The semantic difference cost function is corrected by a metric learning algorithm. Specifically, the anchor response unit is input into the semantic mapping model to obtain the semantic mapping representation of the anchor response unit. The semantic mapping representation is jointly determined by the response action parameters and the response support parameters, and the objective anchor identifier is retained as gating constraint information. For anchor response units that are confirmed to be identical sample pairs after review, their semantic mapping distance is reduced in subsequent calculations; For anchor response units that are confirmed as normal sample pairs after review, their semantic mapping distance is increased in subsequent calculations; anchor point response unit and The semantically mapped representations are denoted as follows: and Then the semantic difference cost in the isomorphic risk identification module can be expressed as: ;in, Anchor point response unit With anchor point response unit The cost of semantic differences between them; Anchor point response unit Semantic mapping representation; Anchor point response unit Semantic mapping representation; Indicates Euclidean distance; The input fields of the semantic mapping representation include objective anchor identifiers, action type, action object, action condition and action result in response action parameters, and basis source, supporting materials, performance conditions and commitment boundaries in response support parameters. Objective anchor identifiers are not directly involved in similarity accumulation as ordinary numerical features, but are used as gating constraints to determine whether matching is allowed. Response action parameters and response support parameters are the main inputs of the semantic mapping model. The triplet metric learning approach is adopted to construct triplet samples from the review results. The triplet samples include benchmark anchor response units, identical anchor response units, and normal anchor response units. The benchmark anchor response units and identical anchor response units come from identical sample pairs confirmed by the review, and the benchmark anchor response units and normal anchor response units come from normal sample pairs confirmed by the review. The semantic mapping model is updated through the triplet loss function to make the distance between identical sample pairs smaller than the distance between normal sample pairs. The reference anchor point response unit is denoted as... The similar anchor point response unit is denoted as The normal anchor point response unit is denoted as Let the interval constant be denoted as The triplet loss function is: ;in, This is the triplet loss value; As the reference anchor point response unit; Anchor point response units that form a similar template pair with the reference anchor point response units; Anchor point response units that form a normal sample pair with the reference anchor point response units; , , These are the semantic mapping representations of the corresponding anchor point response units; The interval constant is used to separate Rado sample pairs and normal sample pairs. The distance is determined by the difference between the mean distance of identical sample pairs in the historical review samples and the mean distance of normal sample pairs. If the historical review samples are insufficient, the initial sample is initialized with the first batch of manual review samples, and the distance is re-estimated after the number of review samples reaches the preset number. After correcting the semantic difference cost function, the updated semantic mapping model will be used for the next batch of bidding projects or the next round of bid document similarity identification. If subsequent projects exhibit response actions and support combinations similar to those of the confirmed copy, the semantic difference cost will be reduced, making it easier to enter the response chain isomorphism risk outcome. If subsequent projects encounter response combinations that are confirmed to be normal industry-standard expressions, their semantic difference cost will be increased or their anomaly identification priority will be reduced, thereby reducing false alarms. The abnormal response fragment library and the normal similar fragment library are updated. For the similar samples that have been verified, the abnormal response fragments with the same source characteristics are written into the abnormal response fragment library. For example, for normal samples that are excluded from review due to the same incorrect references, the same technical omissions, the same unreasonable commitments, the same unnecessary limitations, the same abnormal attachment organization, and the same response defects, their corresponding expressions are written into the normal similar fragment library, such as industry-standard expressions, legally formatted expressions, and necessary similar expressions caused by mandatory requirements of bidding documents. The isomorphism threshold is periodically adjusted, and after each review cycle, the final minimum alignment cost distribution of newly added identical sample pairs and normal sample pairs is statistically analyzed. If new samples show that the current isomorphism threshold leads to an increase in the false alarm rate, then the trigger strength of the abnormal response fragment for risk output should be appropriately reduced or the isomorphism threshold requirement should be increased. If new samples show that the current isomorphism threshold is causing an increase in the false negative rate, then the isomorphism threshold should be appropriately lowered or the recognition priority of specific abnormal response segments should be increased. This step receives the response chain isomorphism risk results and verification results output by the isomorphism risk identification module. Anchor response units confirmed to be rewritten from the same source are marked as identical sample pairs. Normal similarity due to industry standards, legal formats, or mandatory requirements of bidding documents is marked as normal sample pairs. A metric learning algorithm is used to correct the semantic difference cost function, the abnormal response fragment library, the normal similar fragment library, and the isomorphism threshold. Based on the verification feedback, false positives and false negatives can be reduced, and the correction results are fed back to subsequent projects or the next round of identification, forming a closed-loop processing relationship for continuous optimization.

[0019] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0020] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0021] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and inventive constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0022] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0023] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0024] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A large-scale, natively driven intelligent management and control system for the entire bidding and tendering process, characterized in that: Signal connections between modules; Anchor point construction module: Obtain the set of objective documents of the project and perform document preprocessing to obtain the set of structured text of the project. Perform explicit similarity identification processing to generate explicit similarity risk results and supplier pairs to be analyzed. Extract objective requirements and generate an objective anchor point set by combining rule positioning, large model verification and minimum response unit decomposition. Generate an objective anchor point identifier and anchor point hierarchy relationship for each objective anchor point. Response Unit Generation Module: Calls the large model to perform semantic parsing on the electronic bid documents of each supplier. Through candidate recall, semantic judgment, main response determination and supplementary response merging, the bid content is mapped to the set of objective anchor points. The response action parameters and response support parameters are extracted and combined to generate anchor point response units, and consistency verification and deduplication merging are performed. Homogeneity Risk Identification Module: Connects anchor response units into a response chain sequence according to the objective anchor order. Uses a dynamic time warping algorithm with anchor gating constraints to perform non-equal length alignment on the response chain sequences of any two suppliers, calculates the semantic difference cost, and identifies abnormal response segments with the same missing response, the same incorrect reference, the same unreasonable commitment, and the same unnecessary limitation. When the minimum alignment cost is lower than the homogeneity threshold and there is at least one abnormal response segment, a rewritten sham bidding risk result is generated. Cost function correction module: Receives the verification results, marks the anchor response units that are confirmed to be identical sample pairs as identical sample pairs, marks the normally similar ones as normal sample pairs, uses a metric learning algorithm, corrects the semantic mapping model in the semantic difference cost function through the triple loss function, and updates the abnormal response fragment library, the normal similar fragment library and the isomorphism threshold.

2. The intelligent management and control system for the entire bidding and tendering process driven by a large model as described in claim 1, characterized in that, Document preprocessing includes document format recognition, text layer extraction, table content restoration, directory level recognition, paragraph boundary segmentation, attachment reference relationship recognition, page number positioning, clause number recognition, and document version identifier generation.

3. The intelligent management and control system for the entire bidding and tendering process driven by a large model as described in claim 2, characterized in that, The explicit similarity detection process includes detecting duplicate sentences, keyword overlap, template similarity, paragraph edit distance, and attachment hash consistency in the electronic bid documents of different suppliers; If two tender documents meet at least one of the following conditions: large sections of text are completely identical, fixed templates are completely identical, the repetition rate of key paragraphs exceeds the explicit similarity threshold, or the attachment materials have the same hash value, then an explicit similarity risk result will be generated. If the explicit similarity threshold is not reached, the two corresponding suppliers will be identified as supplier pairs to be analyzed.

4. The intelligent management and control system for the entire bidding and tendering process driven by a large model as described in claim 2, characterized in that, Rule location is used to identify clauses with explicit numbering, headings, table fields, or mandatory expressions from a collection of structured text for a project; Large model validation is used to determine whether the text obtained from rule-based localization constitutes an objective requirement that the supplier must respond to; Minimal response unit splitting is used to break down clauses containing multiple objective requirements into multiple objective anchors; Objective anchor point identifiers should include at least the project number, bid section number, source document type, source chapter, clause number, anchor point type, anchor point text, and anchor point sequence number; Anchor point hierarchy is used to determine the primary anchor point and the parent anchor point.

5. The intelligent management and control system for the entire bidding and tendering process driven by a large model as described in claim 1, characterized in that, Candidate recall is based on clause number, title keywords, technical parameter names, delivery deadline expressions, acceptance terms, qualification terms, attachment references, and semantic similarity judgment of the large model, and selects fragments related to objective anchor points from the candidate response fragment set; Semantic determination is used to confirm whether a candidate response fragment constitutes a substantive response to an objective anchor point; The main response determination rule is: select the segment that corresponds to the objective anchor clause number or title; If none exists, select the segment containing a clear action verb and a response object; if multiple segments still exist, select the segment that contains both the response action and supporting content.

6. The intelligent management and control system for the entire bidding and tendering process driven by a large model as described in claim 5, characterized in that, The response action parameters include action type, action object, action condition, and action result; Supporting parameters for the response include the source, supporting documentation, performance conditions, and commitment boundaries; consistency verification includes: If the response action parameter is empty but the response support parameter exists, it is marked as a response unit with missing action. If the response action parameters exist but the response support parameters are empty, it is marked as an insufficiently supported response unit. If both are empty, mark it as a missing response unit; If both exist, mark it as a complete response unit; Deduplication includes incorporating duplicate response content with the same objective anchor point into the main response unit, and recording the source of the same segment when the same bid content corresponds to multiple objective anchor points.

7. The intelligent management and control system for the entire bidding and tendering process driven by a large model as described in claim 1, characterized in that, Anchored gating constraints refer to: If two anchor point response units originate from the same objective anchor point, parent-child anchor point, or the same anchor point cluster, they are allowed to enter the alignment candidate. If the source is an irrelevant anchor, the semantic difference cost is set to a prohibited matching value or the matching cost is increased. The dynamic time warping algorithm calculates the minimum alignment cost between two response chain sequences recursively. During the recursion process, the current alignment cost is obtained by adding the semantic difference cost between the current two anchor response units to the minimum cumulative alignment cost of the previous step.

8. The intelligent management and control system for the entire bidding and tendering process driven by a large model as described in claim 7, characterized in that, The rules for identifying abnormal response fragments include: If two suppliers fail to respond effectively under the same objective anchor point, it is recorded as an identical missing response. References to the same error standard, error clause number, or obviously inapplicable specifications are recorded as identical error references. Making the same commitments that exceed the requirements of the tender documents and are unnecessary shall be recorded as the same unreasonable commitments. Adding the same non-essential constraint, even if the content is not required by the tender document, is considered as the same non-essential constraint. The isomorphism threshold is determined based on the final minimum alignment cost distribution of confirmed identical sample pairs and normal sample pairs in the historical review samples.

9. The intelligent management and control system for the entire bidding and tendering process driven by a large model as described in claim 1, characterized in that, The review results include the review and confirmation of suspected supplier pairs, abnormal response segments, objective anchor point locations, original tender documents, alignment paths, and risk statements; Anchor response units that are confirmed to be rewritten dummy targets, highly similar, or have a common source rewriting relationship are marked as identical sample pairs; Anchor point response units that are normally similar due to the same tender document, the same technical specification, the same industry standard, or the same legal format will be marked as normal sample pairs.

10. The intelligent management and control system for the entire bidding and tendering process driven by a large model as described in claim 9, characterized in that, The metric learning algorithm uses a triplet loss function to correct the semantic mapping model: Construct triplet samples from the review results, including benchmark anchor response units, identical anchor response units, and normal anchor response units, so that the distance between identical sample pairs in the semantic mapping space is less than the distance between normal sample pairs, and they are separated by a constant interval. The abnormal response fragments in the verified identical samples are written into the abnormal response fragment library, the corresponding expressions in the normal samples excluded from verification are written into the normal similar fragment library, and the isomorphism threshold is updated according to the final minimum alignment cost distribution of the newly added verification samples.

Images