A system and method for predicting the probability of winning bids for enterprises based on multidimensional public data.

By generating and adjusting explicit reference relationships in the electronic clause sequence of bidding documents, and establishing directed back-pointing edges, the instability of the relationship network in predicting the probability of winning bids for enterprises is solved, resulting in more stable output of the probability of winning bids and traceability of data.

CN123118221APending Publication Date: 2026-11-17BEIJING QINGMENG SHUHAI TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611234947.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-14
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In predicting the probability of winning bids for enterprises, the existing system has difficulty in reliably preserving the order and explicit reference relationships between the qualification requirements and scoring content in the bidding documents. This makes the relationship network input susceptible to the influence of similar text and duplicate records, resulting in unstable output of the probability of winning bids.

Method used

By collecting, parsing, and standardizing multi-source public data, a sequence of electronic clauses in the tender documents is generated. Explicit reference texts of qualification clauses and scoring clauses are identified, directed back-pointing edges for threshold back-pointing values ​​are established, and back-pointing messages are adjusted in graph neural network message passing to construct a project-enterprise-condition relationship network and output the probability of winning the bid.

Benefits of technology

When multiple sources of publicly available electronic data, enterprise capability profiles, and historical records are used together for inference, the system can retain existing cross-chapter relationships within the bidding documents, reduce relationship loss and duplicate entries, and improve the stability and traceability of the winning probability output.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118221A_ABST
    Figure CN123118221A_ABST
Patent Text Reader

Abstract

This invention relates to the field of intelligent bidding technology, and discloses a system and method for predicting the probability of winning bids for enterprises based on multi-dimensional public data. The system includes: collecting and standardizing bidding documents, enterprise capability profiles, enterprise public participation records, and enterprise historical bidding records; generating an electronic clause sequence in the bidding documents; identifying qualification clauses, scoring clauses, and explicit citation text; establishing explicit clause pairs when scoring clauses cite qualification clauses; generating a threshold backreference value based on clause order information and explicit citation information; and establishing directed backreference edges that write this value in the project-enterprise-condition relationship network; using a graph neural network to adjust backreference messages using the threshold backreference value; generating a relationship network representation; and integrating enterprise capability profiles, condition matching results, historical change representations, and target project text representations to output the probability of the target enterprise winning the target bidding project. This invention is applicable to public bidding data processing and decision support.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of intelligent bidding technology, and more specifically, to a system and method for predicting the probability of winning bids for enterprises based on multidimensional public data. Background Technology

[0002] In public bidding systems, project announcements, qualification review documents, scoring rules, company registration information, qualification certificates, historical participation records, and winning bid records typically originate from multiple public platforms. The data access layer generally acquires electronic files through API calls, webpage downloads, batch imports, and scheduled tasks. These files are then converted into field-based data readable by subsequent programs using file parsing, OCR, and table extraction services. For traceability, the implementation also saves source identifiers, publication dates, summary verification values, and project identifiers. The original files are written to object storage, while the field-based data is written to a relational database or document database.

[0003] In scenarios involving predicting the probability of winning bids for enterprises, common systems break down tender documents into qualification-related content, scoring content, supporting document requirements, and project text. They then combine this with enterprise capability profiles, public participation records, and historical bidding records to perform rule matching, text similarity calculation, temporal coding, and probability inference. Graph neural networks are also frequently used for modeling the relationships between projects, enterprises, conditions, and historical projects, in order to merge information from multiple data sources into a relational network representation oriented towards the target enterprise.

[0004] In practice, qualification-related and scoring content in tender documents is often distributed across different chapters or table areas. Scoring content may reference previously mentioned qualification-related content, but after the document is split into independent records, the system often only retains the record text and general similarity results, making it difficult to stably preserve the existing sequence and explicit citation relationships within the electronic file. This results in the same requirement being processed separately in qualification review and scoring calculation, and the input of relationship networks is easily affected by similar text and duplicate records. Summary of the Invention

[0005] This invention provides a system and method for predicting the probability of winning bids for enterprises based on multidimensional public data, which solves the technical problems mentioned in the background art.

[0006] To achieve the above-mentioned objectives, the present invention provides the following technical solution.

[0007] A method for predicting the probability of winning bids for enterprises based on multidimensional publicly available data, executed by a computer system, includes:

[0008] Collect, analyze, and standardize multi-source publicly available electronic data to form a standardized dataset that includes tender documents, enterprise capability profiles, enterprise public participation records, and enterprise historical bidding records;

[0009] Generate an electronic sequence of tender document clauses in the natural order of the tender documents, and identify qualification clauses, scoring clauses, and explicit references;

[0010] When a scoring clause references a qualifying clause through an explicit reference text, an explicit clause pair is established, and a threshold reference value is generated based on the clause order information and the explicit reference information. The clause order information reflects the relative order of the qualifying clause and the scoring clause, and the explicit reference information reflects the occurrence of the explicit reference text. This allows the threshold reference value to characterize the direction of the scoring clause's reference to the qualifying clause, the clause interval, and the number of explicit references.

[0011] Construct a project-enterprise-condition relationship network, and establish directed back-pointing edges for writing threshold back-pointing values ​​between scoring condition nodes and qualification condition nodes;

[0012] Based on the enterprise's capability profile and qualification and scoring terms, condition matching results are generated; based on the enterprise's public participation records and historical bidding records, historical change representations are generated; and based on the target bidding project text, target project text representations are generated.

[0013] In graph neural network message passing, the threshold back-pointing value is used to adjust the back-pointing messages passed along the directed back-pointing edge, generating a relational network representation. The relational network representation, enterprise capability profile, condition matching results, historical change representation, and target project text representation are then integrated to output the probability of the target enterprise winning the target bidding project.

[0014] Compared with the prior art, the present invention has the following substantial features and significant progress:

[0015] This invention incorporates the order of directly readable clauses and explicit references from the tender documents into the computer processing flow. It generates directed back-pointing edges with threshold back-pointing values ​​between scoring condition nodes and qualification condition nodes, and adjusts the strength of the back-pointing messages during graph neural network message passing. As a result, the system can retain existing cross-chapter relationships within the tender documents when multiple sources of publicly available electronic data, enterprise capability profiles, and historical records are used in the inference process. This reduces relationship loss, duplicate entries, and interference from similar text caused by the splitting of qualification requirements and scoring content, resulting in a more consistent relationship network representation, a more stable output of the winning probability, and easier model input tracing through source identification. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the system architecture and threshold retracement adjustment mechanism of the present invention. Detailed Implementation

[0017] The following description is provided in conjunction with the accompanying drawings. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the invention; those skilled in the art can make equivalent substitutions or combinations for the specific implementations.

[0018] Example 1:

[0019] This implementation method is applicable to scenarios involving open bidding data processing and intelligent bidding decision support. The computer system receives multi-source publicly available electronic data, including historical winning bid announcements, bidding announcements, pre-qualification documents, scoring rules, enterprise business registration information, publicly available qualification certificate information, enterprise public participation records, and enterprise historical winning bid records. After the target bidding project is released, it generates a traceable probability output for the target enterprise.

[0020] like Figure 1 As shown, the system, along the data flow direction, sequentially includes a multi-source publicly available electronic data access unit 100, an electronic clause sequence generation unit 200, a clear clause pair identification unit 300, a threshold index generation unit 400, a project-enterprise-condition relationship network construction unit 500, a graph neural network message adjustment unit 600, and a winning probability output unit 700. Each unit can be deployed on the same server or on multiple servers connected through network communication units.

[0021] This implementation method does not process manually rewritten bidding rules, but rather publicly available electronic documents and data. The electronic clause sequence of the bidding documents retains the natural order of qualification clauses, scoring clauses, table rows, and explicit citations in the electronic document; explicit clauses are formed only when the scoring clauses contain qualification clause numbers or unique headings; the threshold reference value is generated based on the qualification clause number, scoring clause number, and the number of explicit citations.

[0022] In one embodiment, the computer system includes a processor, a memory, a network communication unit, a database, an object storage unit, a model file storage unit, a task scheduling unit, a log recording unit, and a terminal display unit. The memory stores a computer program, which the processor executes to implement the steps of this embodiment. The database stores structured records and model output results; the object storage unit stores read-only snapshots of the original files; and the model file storage unit stores the text-encoded model version, graph neural network model parameters, temporal network model parameters, a field dictionary, graph relation types, and the training data cutoff time.

[0023] Example 2:

[0024] The multi-source publicly available electronic data access unit 100 collects multi-source publicly available electronic data according to a preset cycle or manually triggered instructions. This multi-source publicly available electronic data includes historical winning bid announcements, tender announcements, pre-qualification documents, scoring rules, enterprise registration information, publicly available qualification certificate information, enterprise participation records, and enterprise historical winning bid records. For ease of implementation, historical winning bid announcements must at least include the project name, project number, issuing unit, candidate enterprises, final winning bidder, and announcement date; enterprise registration information must at least include the unified social credit code, enterprise name, and registration status; and publicly available qualification certificate information must at least include the qualification category, qualification level, and validity period.

[0025] When collecting publicly available electronic data from multiple sources, the multi-source publicly available electronic data access unit 100 saves the source identifier, publication time, download time, file type, file size, summary checksum, and project identifier for each file or record. The source identifier is used to link back to the source of the publicly available data; the publication time is used to determine whether the data has been published before the announcement time of the target bidding project; the summary checksum is used to detect duplicate files or file changes; and the project identifier is used to aggregate the multi-source publicly available electronic data to the corresponding project node. All of the above fields are standard electronic data collection fields.

[0026] For PDF, DOCX, HTML, XLSX, JSON, or image files downloaded from web pages, the multi-source publicly available electronic data access unit 100 first writes the original file to object storage in read-only snapshot mode, and then writes the file index and summary checksum to the database. For structured records obtained through public data interfaces, the multi-source publicly available electronic data access unit 100 saves the interface address, request time, returned record summary, and field version. For files imported manually in batches, the multi-source publicly available electronic data access unit 100 saves the import batch, file path, operator identifier, and import time.

[0027] The heterogeneous file parsing and unified format conversion module handles the standardization dataset formation process. This module identifies file types such as structured files, semi-structured files, layout files, tabular files, and image files, and respectively calls a conventional parser to extract the main text, headers and footers, tables, and appendices. For image pages, the module calls a text recognition program to obtain the text content, retaining page numbers and text box coordinates. For tabular files, the module saves cell records according to row and column coordinates, without directly rewriting the table content into free text.

[0028] During the unified format conversion, the heterogeneous file parsing and unified format conversion module standardizes character encoding to UTF-8, converts dates, monetary units, qualification level formats, and full and abbreviated company names into a unified field format, and preserves the original page numbers, paragraph positions, and table positions. When a company name changes, the system does not delete the old name but saves the name change time and maps the old name to the same company identifier. The resulting standardized dataset can be directly read by subsequent programs and can also locate the original content of the publicly available electronic data through source identifiers, page numbers, paragraph positions, and table positions.

[0029] When fields are missing in publicly available electronic data from multiple sources, the heterogeneous file parsing and unified format conversion module writes a missing field flag, without replacing the missing field with a speculative value. Field type, date range, file integrity, and source backlinks are checked using standard validation procedures. If the file title, project number, announcement date, and summary validation value are all the same for the same project, the system can mark it as a duplicate record; if the summary validation values ​​are different, multiple versions are retained, and the download time is recorded for each, so that previously published data at the corresponding time segment can be used during model training and target project inference.

[0030] Example 3:

[0031] The electronic clause sequence generation unit 200 reads the tender documents from a standardized dataset. The tender documents may include a tender notice, prequalification documents, scoring rules, and appendices containing qualification or scoring clauses. The electronic clause sequence generation unit 200 generates a sequence of electronic clauses for the tender documents according to the page order, paragraph order, heading level, and table row order in the electronic files.

[0032] The electronic clause sequence of the tender document consists of a title, qualification clauses, scoring clauses, supporting document requirements, table rows, and explicit citations. The title is used to determine the chapter level; the qualification clauses express the mandatory qualification requirements or qualification review requirements; the scoring clauses express the review scores, scoring instructions, and scoring conditions; the supporting document requirements specify the supporting documents that the company needs to submit; the table rows are used to retain the row and column positions in the tabular scoring details; and the explicit citations are used to express the qualification clause number or unique title that appears directly in the scoring clauses.

[0033] The electronic clause sequence generation unit 200 can segment the tender document according to chapter titles, numbering levels, punctuation, and table rows. For the main text paragraphs, the system determines individual condition records based on paragraph number, indentation, title format, and punctuation position. For tabular scoring rules, the system maps fields based on scoring items, scoring instructions, scores, and supporting materials in the table header, and retains the row and column numbers of each table cell. For the image page text obtained from OCR, the system saves the recognized text along with the page number and text box coordinates, and uses the recognition confidence score as the extraction confidence score record.

[0034] When generating the electronic clause sequence of the tender document, the electronic clause sequence generation unit 200 assigns a natural sequence number to each qualification clause and each scoring clause. The natural sequence number reflects the order in which the clauses appear in the electronic document. The qualification clause sequence number is the... The natural order of the qualification clauses within the electronic clauses sequence of the tender document is denoted as follows: The rating clause number is [number missing]. The natural order of the scoring clauses within the electronic clause sequence of the tender document is denoted as: .

[0035] The electronic clause sequence generation unit 200 also extracts condition fields. These condition fields may include enterprise type, qualification name, qualification level, certificate validity period, performance duration, performance quantity, personnel certificates, equipment requirements, geographical requirements, scoring values, and supporting document names. Condition field extraction can employ conventional text processing methods such as keyword dictionaries, regular expressions, named entity recognition models, and pre-trained text encoding models. The above processing is only used to convert clauses in publicly available documents into computer-readable records.

[0036] Example 4:

[0037] After receiving the electronic tender document clause sequence, the explicit clause identification unit 300 first distinguishes between qualification clauses, scoring clauses, and supporting documentation requirements. If a scoring clause contains a qualification clause number or unique title, and this qualification clause number or unique title can uniquely locate a qualification clause in the electronic tender document clause sequence, then the explicit clause identification unit 300 records an explicit clause pair. Explicit clause pairs are used to express the explicit reference relationship between the scoring clause and the qualification clause, and are not used to express ordinary textual similarity relationships.

[0038] Explicit clause pairs include the clause pair number, eligibility clause number, scoring clause number, explicit reference count, scoring clause source identifier, and eligibility clause source identifier. The clause pair number is used to distinguish a scoring clause from its explicitly referenced eligibility clause and is denoted as [example number would be inserted here]. Explicit citation count refers to the number of times the corresponding eligibility clause number or unique heading appears explicitly in the scoring clause, denoted as _____. If the scoring terms do not explicitly quote the text, then explicit terms will not generate explicit terms pairs for identification unit 300.

[0039] Explicit citation counts only include instances where the text is explicitly cited. For example, if a corresponding eligibility clause number appears once in a scoring clause, the explicit citation count is [number missing]. If the corresponding eligibility clause number or unique title appears twice in the scoring terms, the number of explicit citations is [number missing]. If the scoring clause and the eligibility clause are only similar in wording, but there is no direct reference in the form of an eligibility clause number or unique heading, then the express reference count is [number missing]. Furthermore, it does not generate explicit clause pairs. This restriction prevents the system from substituting semantic guesses for references that do not exist in publicly available electronic documents.

[0040] The threshold reference value generation unit 400 generates a threshold reference value based on the explicit clauses, corresponding to the clause order information and explicit reference information. The clause order information reflects the relative order of the eligibility clauses and the scoring clauses, while the explicit reference information reflects the occurrence of explicitly cited texts. The threshold reference value simultaneously characterizes the direction of the scoring clause's reference to the eligibility clause, the clause interval, and the number of explicit references. For the first... A clear clause states that the threshold return value is denoted as... The calculation formula is:

[0041]

[0042] in, The qualification clause number indicates the first one. The clear terms correspond to the natural order of the qualification terms in the electronic terms sequence of the tender document. The electronic clause sequence generation unit 200 reads directly from the clause order of the standardized document.

[0043] The rating clause number indicates the number of the rating clause. The clear terms correspond to the natural order of the scoring terms in the electronic terms sequence of the tender document. Similarly, the electronic clause sequence generation unit 200 reads directly from the clause order of the standardized document.

[0044] To indicate the number of citations, this indicates the number of times the scoring clause explicitly cites the corresponding eligibility clause number or unique heading. The score is obtained by counting the explicit references in the scoring clauses by the explicit clause identification unit 300.

[0045] This is a subsequent assessment result. When the scoring clause follows the eligibility clause, the subsequent assessment result takes precedence. When the scoring clause is not placed after the eligibility clause, the subsequent judgment result shall be applied. This post-judgment result ensures that the threshold reference value only applies to document relationships where the later-appearing scoring clause explicitly references the earlier-appearing eligibility clause.

[0046] constant in the denominator Used to avoid denominator being This ensures that the threshold retracement value remains bounded. The threshold retracement value satisfies the following range:

[0047]

[0048] When the number of explicit citations is the same, the smaller the difference between the scoring clause number and the eligibility clause number, the smaller the clause interval, and the larger the threshold retracement value; when the clause interval is the same, the more explicit citations, the larger the threshold retracement value.

[0049] In a specific example, the qualification clause number is The rating clause number cited is The eligibility clause number is clearly stated in the scoring terms. Then the threshold backreference value is:

[0050]

[0051] The threshold retrace value is written into the directed retrace edge between the scoring condition node and the eligibility condition node. During message passing in the graph neural network, the retrace message sent from the eligibility condition node to the scoring condition node is first multiplied by... Then, it participates in the aggregation of neighbor messages of the scoring condition node.

[0052] like Figure 1 The explicit clause identification unit 300 and the threshold retracement value generation unit 400 are shown in the diagram. The explicit clause identification unit 300 outputs explicit clause pairs, and the threshold retracement value generation unit 400 receives the explicit clause pairs and outputs threshold retracement values. The threshold retracement value generation unit 400 outputs threshold retracement values ​​to the project-enterprise-condition relationship network construction unit 500 for writing directed retracement edges; on the other hand, it outputs threshold retracement values ​​to the graph neural network message adjustment unit 600 for adjusting retracement messages.

[0053] Example 5:

[0054] The Enterprise Capability Profile Time-Based Encoding Module uses the Unified Social Credit Code as the enterprise owner identifier, linking enterprise registration information, publicly available qualification certificates, licenses, personnel certificates, publicly available performance records, and historical bidding announcements. If the enterprise name changes, the module saves the name change time and maps the old name to the same enterprise identifier. The enterprise identifier is used to generate enterprise nodes in the project-enterprise-condition relationship network and is used for subsequent condition matching results, historical change representation, and bidding probability output.

[0055] The enterprise capability profile time-based coding module extracts enterprise capability profiles based on the announcement time of the target bidding project. For qualification certificates, licenses, and personnel certificates, the system checks their publication date and validity period, using only information that was published before the target bidding project's announcement time and has not yet expired. For the enterprise's historical bidding records and public participation records, the system also only uses records that were published before the target bidding project's announcement time to prevent information that appeared after the announcement from being written into the model input.

[0056] Category fields use a unified dictionary encoding, quantity fields are marked for missing values ​​and scaled using standard numerical methods, and text fields are converted into vectors from a pre-trained text encoding model. Enterprise capability profiles can include data blocks such as basic enterprise information, qualifications and capabilities, personnel capabilities, historical projects, and publicly available relationships. Each data block stores the source time and source identifier, enabling the output of the winning bid probability to be traced back to the corresponding publicly available electronic data.

[0057] The qualification threshold and scoring item hierarchical matching module receives enterprise capability profiles, qualification terms, and scoring terms, and generates condition matching results. For directly comparable conditions such as qualification level, certificate validity period, performance volume, and number of personnel, the qualification threshold and scoring item hierarchical matching module uses a rule engine to generate rule compliance scores. Rule compliance scores are denoted as follows: ,in Number the company For condition numbering.

[0058] The enterprise ID is used to distinguish the enterprise being evaluated and is mapped from the unified social credit code. The condition ID is used to distinguish individual qualification conditions or scoring conditions in the bidding documents and is obtained from the sequence number of the standardized condition record. The rule compliance degree indicates the enterprise's compliance with the specified rule. The degree to which a directly comparable condition is met can be obtained by comparing the field values ​​in the company's capability profile with the threshold values ​​in the qualification or scoring clauses.

[0059] For textual conditions such as project experience descriptions, technical skill descriptions, and scoring instructions, the qualification threshold and scoring item hierarchical matching module generates semantic closeness scores. Semantic closeness scores are denoted as... This indicates that the company's document text is consistent with the first... The degree of similarity between the conditional texts. Semantic similarity can be calculated by encoding the enterprise file text and the conditional text separately using a pre-trained text encoding model, and then calculating the similarity using conventional methods.

[0060] The qualification threshold and scoring item hierarchical matching module generates a comprehensive matching value based on rule compliance and semantic similarity. The comprehensive matching value is denoted as... The calculation formula is:

[0061]

[0062] in, The overall matching value indicates the relationship between the company and the... The combined matching results of the conditions. To determine compliance with the rules, the rule engine compares the company's capability profile with the threshold value. The semantic similarity is obtained from the vector similarity between the enterprise file text and the conditional text. The rule proportion represents the percentage of rule compliance in the overall matching value, which can be determined by the validation set using the conventional parameter selection method.

[0063] It should be noted that both rule conformity and semantic proximity are converted to a dimensionless and unidirectional unified scale, and are both limited to... The interval; where This indicates a mismatch or inaccuracy in the text. This indicates a perfect match or the closest match in text. The rule percentage is a dimensionless parameter, satisfying... The rule proportion is determined by the validation set based on a comprehensive evaluation of the area under the receiver operating characteristic (AUC) curve, precision, recall, and probability error. Once determined, this proportion is written into the model version and remains unchanged during training and inference within the same model version. If the original calculation results of rule conformity or semantic proximity are not... The interval, qualification threshold and scoring item hierarchical matching module first maps them to the normalized parameters fixed during the training phase. Then calculate the comprehensive matching value within the range.

[0064] When the hard qualification thresholds are not met, the qualification threshold and scoring item hierarchical matching module saves the qualification failure mark and the reason for failure. The qualification failure mark does not delete the enterprise node or the original matching evidence, but is input into the subsequent fusion process as part of the conditional matching result. For supporting materials that must be submitted but for which there is no corresponding public record in the enterprise capability file, the system records them as missing fields and does not automatically assume that they are met.

[0065] Example 6:

[0066] The Project-Enterprise-Condition Relationship Network Construction Unit 500 constructs a project-enterprise-condition relationship network based on standardized datasets, enterprise capability profiles, condition matching results, explicit clause pairs, and threshold reference values. This relationship network includes at least project nodes, enterprise nodes, qualification condition nodes, and scoring condition nodes, and may include historical project nodes when publicly available data exists.

[0067] The project node represents the target bidding project or a historical bidding project. The enterprise node represents the enterprise being evaluated or a historically participating enterprise. The qualification condition node represents the qualification clauses in the bidding documents. The scoring condition node represents the scoring clauses in the bidding documents. The historical project node represents historical projects for which winning bids have been published and which can be used for training or relationship modeling. The initial input for the above nodes includes standardized fields and text vectors.

[0068] The Project-Enterprise-Condition Relationship Network Construction Unit 500 establishes the inclusion relationship between project nodes and qualification condition nodes, the inclusion relationship between project nodes and scoring condition nodes, the possession relationship between enterprise nodes and qualification condition nodes, and the participation or bidding relationship between enterprise nodes and historical projects. Participation or bidding relationships are used to generate a relationship network representation and save the occurrence time and source project. For publicly disclosed investment or control relationships in business registration records, the system can establish publicly related edges if the publicly disclosed record exists; otherwise, it will not create publicly related edges.

[0069] For explicit clause pairs, Project-Enterprise-Condition Relationship Network Construction Unit 500 establishes directed backlinks between the scoring condition node and its explicitly referenced qualification condition node, and writes the threshold backlink value into the directed backlink. The directed backlink only reproduces the explicit reference relationships already existing in the tender document; it does not add new clauses, adjust the order of clauses, or replace the original node text. If a scoring clause does not have an explicit reference text, Project-Enterprise-Condition Relationship Network Construction Unit 500 does not establish directed backlinks for that scoring clause and qualification clause.

[0070] like Figure 1 As shown in the Project-Enterprise-Condition Relationship Network Construction Unit 500, this unit contains project nodes, enterprise nodes, qualification condition nodes, and scoring condition nodes. A directed back-pointing edge is marked between the scoring condition nodes and the qualification condition nodes, and a threshold back-pointing value is written next to the directed back-pointing edge. The Project-Enterprise-Condition Relationship Network Construction Unit 500 outputs a relationship network containing back-pointing edges to the Graph Neural Network Message Adjustment Unit 600, enabling the Graph Neural Network Message Adjustment Unit 600 to identify which edge needs to be adjusted by the threshold back-pointing value.

[0071] For the project to be predicted, the Project-Enterprise-Condition Relationship Network Construction Unit 500 uses the announcement time of the target bidding project as the cutoff time and extracts a partial subgraph from the Project-Enterprise-Condition Relationship Network that connects to the target enterprise, similar historical projects, and potential competitors. The partial subgraph retains project nodes, enterprise nodes, qualification condition nodes, scoring condition nodes, inclusion relationships, possession relationships, participation relationships, winning bid relationships, and directed back-pointing edges. Except for the newly added directed back-pointing edges and their threshold back-pointing values, the other node types, relationship types, and time truncation methods remain consistent with the original multi-source publicly available electronic data prediction process.

[0072] Example 7:

[0073] The graph neural network message conditioning unit 600 receives a relation network containing back-pointing edges and threshold back-pointing values. The unit uses the standardized fields and text vectors of project nodes, enterprise nodes, qualification condition nodes, and scoring condition nodes as initial node inputs, and employs a graph convolutional network or graph attention network to aggregate neighbor information. Different relation types can be distinguished by relation type numbering or independent linear transformations.

[0074] In each layer of message passing, the graph neural network message adjustment unit 600 first determines whether a neighbor message enters the scoring condition node from the eligibility condition node along a directed backlink edge. If the neighbor message does not belong to a directed backlink edge, the graph neural network message adjustment unit 600 aggregates it according to the original relation type processing method. If the neighbor message belongs to a directed backlink edge, the graph neural network message adjustment unit 600 reads the threshold backlink value written on the directed backlink edge and adjusts the backlink message using the threshold backlink value.

[0075] In one implementation, the eligibility condition node is at the... The node representation of a layer is denoted as The scoring condition node is at the 1st The node representation of a layer is denoted as The threshold retracement value corresponding to the directed retracement edge is The linear transformation matrix corresponding to the relation type is The retrace message propagated along a directed retrace edge can be represented as:

[0076]

[0077] in, For the first The layer passes back the message from the eligibility condition node to the scoring condition node. The threshold reference value is calculated from the qualification clause number, the scoring clause number, and the number of explicit references. For the first The layer is a linear transformation matrix for directed back-pointing edge relationships. For the first The node representation of the qualification condition node.

[0078] When aggregating neighbor messages, the scoring condition node performs summation or attention aggregation on the retracement messages adjusted for the threshold retracement value along with other neighbor messages. In one summation aggregation implementation, the next-level node of the scoring condition node can be represented as follows:

[0079]

[0080] in, For the first The node representation of the layer scoring condition node. It is a non-linear activation function. For the first The layer is a linear transformation matrix for the rating condition node itself. This is the set of neighboring nodes of the scoring condition node. For the first Layer neighbor nodes The message passed to the scoring condition node. For messages entering the scoring condition node along the directed back-pointing edge... The form after adjusting the threshold retracement value is adopted; for other types of edges, The original relation type processing method is adopted.

[0081] When determining the dimension of message passing in a graph neural network, the first... Layer node representation dimension is , No. Layer node representation dimension is The node representation of the eligibility criteria node. Node representation of scoring condition nodes All Dimensionless vector; linear transformation matrix and linear transformation matrix All will be the first Layer node representation mapping to The message space has a matrix dimension of 1. . Retrace message and neighboring nodes Messages passed to the scoring condition node All A dimensionless message vector. All message vectors entering the same scoring condition node are transformed into the same... After determining the dimensions, perform summation or attention aggregation.

[0082] Threshold retracement value When the threshold retrieval value is greater than 1, the corresponding retrieval message is not transmitted. At that time, the graph neural network message adjustment unit 600 adjusts the retracement message proportionally according to the threshold retracement value. This processing enables the system to distinguish between two relationships: similar ordinary text and scoring clauses that explicitly retrace eligibility clauses, reducing relationship loss, duplicate counting, and interference from similar texts caused by the splitting of eligibility requirements and scoring content.

[0083] like Figure 1 As shown in the diagram, the neural network message adjustment unit 600 internally incorporates a back-pointing message adjustment process. This process receives a relational network containing back-pointing edges from the project-enterprise-condition relational network construction unit 500, and a threshold back-pointing value from the threshold back-pointing value generation unit 400. Internally, messages pointing from qualification condition node information to scoring condition node information are adjusted according to the threshold back-pointing value, and then the relational network representation is output.

[0084] The relationship network representation is a vector representation obtained by the graph neural network message conditioning unit 600, which aggregates the project nodes, enterprise nodes, qualification condition nodes, scoring condition nodes, and related historical project relationships of the target enterprise in the local subgraph of the current project. The relationship network representation does not determine the winning probability independently, but participates in the subsequent multi-dimensional representation fusion as a data block representation in the winning probability output unit 700.

[0085] Example 8:

[0086] The historical record time-series representation module generates a historical change representation based on the company's public participation records and historical bidding records. This module extracts the company's public participation records and historical bidding records according to the announcement time of the target bidding project, arranges the records by fixed time windows such as months or quarters, and counts the number of participations, the number of successful bids, the distribution of project categories, the status of qualification changes, and the continuous vacancy period within each time window.

[0087] A consecutive vacancy period is defined as the number of time slices in which a company has neither publicly participated in nor won bids in a series of consecutive time slices. If a company has no publicly participated records and no historical winning bid records in multiple consecutive time slices, the historical record time series representation module accumulates these time slices into a consecutive vacancy period. Each time slice stores a missing value marker to avoid misinterpreting filled values ​​as actual records. The historical record time series representation module can use a Long Short-Term Memory (LSTM) network, gated recurrent units (GRUs), or a one-dimensional convolutional network to encode the time slice sequence, obtaining a representation of historical changes.

[0088] The target project text representation generation module generates a target project text representation based on the target bidding project text. The target bidding project text can come from bidding announcements, prequalification documents, scoring rules, and project attachments. The module text-encodes the title, basic project information, technical requirements, delivery requirements, qualification clauses, scoring clauses, and supporting documentation requirements, and uses the encoded results as the target project text representation. This target project text representation, along with the relationship network representation, company capability profile, condition matching results, and historical change representation, is input into the winning probability output unit 700.

[0089] The winning probability output unit 700 converts the condition matching results, enterprise capability profiles, relationship network representations, historical change representations, and target project text representations into data block representations of the same dimension. The data block numbers are denoted as follows: The total number of data blocks is denoted as Data blocks are represented as The percentage of data blocks is denoted as... This is calculated by the regular attention layer. Enterprise convergence is denoted as... The calculation formula is:

[0090]

[0091] in, For enterprise fusion representation, it represents the enterprise representation after weighted aggregation of multiple data blocks. The total number of data blocks is directly read from the model input configuration. Number the data blocks to distinguish input blocks such as condition matching results, enterprise capability profiles, relationship network representations, historical change representations, and target project text representations. The percentage of data blocks indicates the percentage of the first data block. The percentage of each data block in this fusion. The data block is represented as the first data block of the enterprise. Vector representation of input blocks.

[0092] The winning bid probability output unit 700 inputs the enterprise fusion representation into a fully connected output layer and obtains the winning bid probability of the target enterprise for the target bidding project through a probability compression function. The winning bid probability is denoted as... The calculation formula is:

[0093]

[0094] in, , where represents the probability of winning the bid, indicating the model's output on the likelihood of a company winning a bid for a target tender project. This is a probability compression function. The output weights are obtained through supervised training. This refers to the integration of enterprises. The output offset is obtained through supervised training.

[0095] like Figure 1 As shown in the diagram, the winning bid probability output unit 700 receives the relational network representation output by the graph neural network message conditioning unit 600, and integrates the enterprise capability profile, condition matching results, historical change representation, and target project text representation. After multi-dimensional representation fusion, it outputs the winning bid probability. The output record includes at least the project identifier, enterprise identifier, winning bid probability, model version, inference time, data cutoff time, qualification threshold check results, main matching items, missing fields, and original data source.

[0096] In one implementation, the target company can be specified by the user, or the system can generate a list of candidate companies from a set of companies with major publicly disclosed qualifications according to the same rules. The generation of the candidate company list is only used for batch inference.

[0097] Example 9:

[0098] The supervised learning sample generation module uses each historical project with published bidding results as a time segment, constructing project inputs before the bidding results are published. If the announcement publicly discloses candidate companies or bidding companies, these companies are used as sample companies; if only the winning companies are disclosed, control samples are drawn from companies in the same region and category with key qualifications according to standard rules. Labels are then written for the final winning companies. Write tags for other sample companies in the same project. .

[0099] The sample uses only enterprise capability profiles, public participation records, historical bidding records, and project-enterprise-condition relationship networks prior to the project announcement date, removing any information from bidding results or subsequent changes. The training, validation, and test sets are divided according to the project announcement time to avoid the same project appearing in multiple sets. When the number of categories differs significantly, conventional category weights or random undersampling can be used.

[0100] The model training, calibration, and version fixation modules use the training set to jointly train the text encoding layer, graph neural network, temporal network, attention fusion layer, and output layer, employing binary cross-entropy loss, batch gradient updates, and early stopping on the validation set. For a fixed general text encoding model, only its upper-layer networks can be trained. During training, the threshold retracement value participates in the graph neural network message passing as a fixed-edge gating quantity of the directed retracement edge.

[0101] On the validation set, the model training, calibration, and version consolidation modules check the area under the receiver operating characteristic (ROC) curve, precision, recall, and probability error, and employ logistic calibration or monotonic regression calibration to ensure that the probability range output by the model remains largely consistent with the actual proportions in historical samples. After training, the model training, calibration, and version consolidation modules solidify the model parameters, field dictionary, text-encoded model version, graph relation type, time window, training data cutoff time, and program summary verification value, forming a reproducible model version.

[0102] When inferring the target project, the computer system calls the pre-installed model version. If new winning bid announcements or publicly available company documents are generated, the system appends them to the original data storage and regenerates samples and trains the model according to a preset cycle. The new model replaces the old model version after it has been validated, while the old model version and its input fields are retained for tracing historical inferences.

[0103] Specifically, at the start of a complete control cycle, the multi-source public electronic data access unit 100 first collects historical bidding announcements, tender announcements, prequalification documents, scoring rules, enterprise registration information, publicly available qualification certificate information, enterprise public participation records, and enterprise historical bidding records, and saves the source identifier, publication time, download time, file type, file size, summary check value, and project identifier. The original files are written to object storage in read-only snapshot mode, and structured records are written to the database.

[0104] Subsequently, the heterogeneous file parsing and unified format conversion module identifies the file type, extracts the main text and tables, performs text recognition on the image pages, and converts the date, monetary unit, qualification level format, and full and abbreviated company name into a unified field format. The electronic clause sequence generation unit 200 reads the tender document, generates the electronic clause sequence of the tender document in natural order, and retains the page numbers, paragraph positions, and table positions.

[0105] Next, the explicit clause identification unit 300 searches the electronic clause sequence of the tender document for explicit references to the qualification clause number or unique title of the scoring clause. When a scoring clause contains explicit reference text, the explicit clause identification unit 300 records the explicit clause pair and outputs the qualification clause number, the scoring clause number, and the number of explicit references. The threshold reference value generation unit 400 generates a threshold reference value based on the qualification clause number, the scoring clause number, and the number of explicit references.

[0106] Then, the system generates enterprise capability profiles, condition matching results, historical change representations, and target project text representations according to the announcement time of the target bidding project, and constructs a project-enterprise-condition relationship network with the threshold back-index value.

[0107] Next, the system executes graph neural network message adjustment to generate a relational network representation, and merges the relational network representation with other data blocks to output the probability of the target company winning the target bidding project.

[0108] Finally, the system writes the project identifier, company identifier, probability of winning the bid, model version, inference time, data cutoff time, qualification threshold check results, main matching items, missing fields, and original data source into the database, and displays them through the terminal display unit. Since each input data block saves the source identifier and time information, users can trace the publicly available electronic data on which the probability of winning bid output is based.

[0109] As an optional implementation, under normal operating conditions, the tender notice, prequalification documents, and scoring details of the target bidding project can all be parsed. The electronic clause sequence generation unit 200 generates a sequence of electronic clauses in the tender documents in natural order. If a scoring clause in the scoring details explicitly references a preceding qualification clause number, the explicit clause pair identification unit 300 forms an explicit clause pair and records the qualification clause number, the scoring clause number, and the number of explicit references. Under this operating condition, the system sequentially completes the generation of threshold backreference values, the writing of directed backreference edges, the adjustment of backreference messages, and the output of the probability of winning the bid. The explicit reference relationship between the scoring clause and the qualification clause is preserved in the relationship network representation.

[0110] As an optional implementation, in the threshold feedback value triggering condition, the scoring clause is located after the qualification clause, and the corresponding qualification clause number or unique title appears multiple times in the scoring clause. In this case, the post-judgment result is... Explicitly cited more than The threshold retracement value is greater than When the graph neural network message adjustment unit 600 sends the qualification condition node information to the scoring condition node information along the directed back-pointing edge, it first multiplies it by the threshold back-pointing value before participating in the neighbor message aggregation. The purpose of this operation is that when the qualification clause and the scoring clause have an explicit reference relationship in the electronic document, the system no longer relies solely on semantic proximity or the trained ordinary graph attention weights to infer their relationship. Instead, it transforms the existing explicit reference relationship in the document into a fixed-edge gating quantity. If two clauses are textually similar but do not have explicit reference text, the explicit clause pair identification unit 300 does not generate an explicit clause pair, and the project-enterprise-condition relationship network construction unit 500 does not establish a directed back-pointing edge.

[0111] As an optional implementation, in the case of no explicit reference text, scoring clauses and qualification clauses may contain similar expressions, but the scoring clauses do not have qualification clause numbers or unique titles. The explicit clause pair identification unit 300 does not generate explicit clause pairs, the threshold retracement value generation unit 400 does not generate threshold retracement values, and the project-enterprise-condition relationship network construction unit 500 does not establish corresponding directed retracement edges. In this case, the graph neural network message adjustment unit 600 can still pass messages according to other relationships between project nodes, enterprise nodes, qualification condition nodes, and scoring condition nodes, and the qualification threshold and scoring item hierarchical matching module can still generate rule compliance, semantic proximity, and comprehensive matching values. The system will not construct explicit clause pairs out of thin air due to text similarity, thereby avoiding the erroneous treatment of ordinary similar text as explicit reference relationships.

[0112] As an optional implementation, in the case of missing fields and time truncation, the enterprise capability profile may lack a record of a certain supporting document, or the information of a certain publicly disclosed qualification certificate may only be made public after the announcement time of the target bidding project. The enterprise capability profile time-based encoding module truncates the time according to the announcement time of the target bidding project, retaining only information that has been made public before the announcement time and has not yet expired. The qualification threshold and scoring item hierarchical matching module writes the supporting documents lacking corresponding public records into the missing field, and saves the qualification failure mark and the reason for failure when the hard qualification threshold is not met. In this case, the historical record time sequence representation module also only reads the enterprise's public participation records and the enterprise's historical bidding records before the announcement time of the target bidding project. The resulting historical change representation will not include the bidding results or subsequent changes of the target bidding project. The bidding probability output unit 700 can maintain the consistency of the time section and can be traced through source identification and data cutoff time.

[0113] Example 10:

[0114] A system for predicting the probability of winning bids for enterprises based on multi-dimensional public data includes a computer system for processing multi-source publicly available electronic data. This system includes modules for data acquisition and raw data retention, heterogeneous file parsing and unified format conversion, bidding condition decomposition and threshold retracement value generation, enterprise capability profile time-based encoding, hierarchical matching of qualification thresholds and scoring items, historical time-series representation, target project text representation generation, project-enterprise-condition relationship network and retracement edge construction, graph neural network relationship representation learning, and attention fusion and probability output.

[0115] The data acquisition and raw data retention module corresponds to the multi-source publicly available electronic data access unit 100, which is used to acquire multi-source publicly available electronic data and save source identifiers. The heterogeneous file parsing and unified format conversion module is used to generate standardized datasets. The bidding condition decomposition and threshold index generation module corresponds to the electronic clause sequence generation unit 200, the explicit clause pair identification unit 300, and the threshold index generation unit 400, which are used to generate the electronic clause sequence of the bidding documents, identify explicit clause pairs, and generate threshold index values.

[0116] The Enterprise Capability Profile Time-Based Encoding Module generates enterprise capability profiles. The Qualification Threshold and Scoring Item Hierarchical Matching Module generates conditional matching results. The Historical Record Time-Series Representation Module generates historical change representations based on the enterprise's public participation records and historical bidding records. The Target Project Text Representation Generation Module generates target project text representations. The Project-Enterprise-Condition Relationship Network and Back-Point Edge Construction Module corresponds to Project-Enterprise-Condition Relationship Network Construction Unit 500, used to construct the Project-Enterprise-Condition Relationship Network and establish directed back-point edges that write threshold back-point values.

[0117] The graph neural network relationship representation learning module corresponds to the graph neural network message conditioning unit 600, which is used to adjust the retracement message using the threshold retracement value and generate a relationship network representation. The attention fusion and probability output module corresponds to the winning probability output unit 700, which is used to output the winning probability based on the relationship network representation, enterprise capability profile, condition matching results, historical change representation, and target project text representation.

[0118] The aforementioned modules can be implemented by a computer program stored on a computer-readable storage medium. When executed by a processor, this computer program can perform methods and steps such as multi-source public electronic data acquisition and standardization, generation of electronic clause sequences in tender documents, identification of explicit clause pairs, generation of threshold back-index values, construction of project-enterprise-condition relationship networks, graph neural network message conditioning, multi-dimensional representation fusion, and output of winning probability. The computer-readable storage medium can be a hard disk, solid-state drive, read-only memory, random access memory, flash memory, or server storage medium.

[0119] Although embodiments of this application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of this application, and all such modifications and variations fall within the scope defined by the appended claims.

Claims

1. A method for predicting the probability of winning bids for enterprises based on multidimensional publicly available data, executed by a computer system, characterized in that: include: Collect, analyze, and standardize multi-source publicly available electronic data to form a standardized dataset that includes tender documents, enterprise capability profiles, enterprise public participation records, and enterprise historical bidding records; Generate an electronic sequence of tender document clauses in the natural order of the tender documents, and identify qualification clauses, scoring clauses, and explicit references; When a scoring clause references a qualifying clause through an explicit reference text, an explicit clause pair is established, and a threshold reference value is generated based on the clause order information and the explicit reference information. The clause order information reflects the relative order of the qualifying clause and the scoring clause, and the explicit reference information reflects the occurrence of the explicit reference text. This allows the threshold reference value to characterize the direction of the scoring clause's reference to the qualifying clause, the clause interval, and the number of explicit references. Construct a project-enterprise-condition relationship network, and establish directed back-pointing edges for writing threshold back-pointing values ​​between scoring condition nodes and qualification condition nodes; Based on the enterprise's capability profile and qualification and scoring terms, condition matching results are generated; based on the enterprise's public participation records and historical bidding records, historical change representations are generated; and based on the target bidding project text, target project text representations are generated. In graph neural network message passing, the threshold back-pointing value is used to adjust the back-pointing messages passed along the directed back-pointing edge, generating a relational network representation. The relational network representation, enterprise capability profile, condition matching results, historical change representation, and target project text representation are then integrated to output the probability of the target enterprise winning the target bidding project.

2. The method for predicting the probability of winning a bid for an enterprise based on multidimensional public data according to claim 1, characterized in that, The multi-source publicly available electronic data includes historical winning bid announcements, tender announcements, prequalification documents, scoring rules, enterprise business registration information, publicly available qualification certificate information, enterprise public participation records, and enterprise historical winning bid records. When collecting the multi-source publicly available electronic data, the source identifier, publication time, summary verification value, and project identifier used to aggregate the multi-source publicly available electronic data to the corresponding project node are saved.

3. The method for predicting the probability of winning a bid for an enterprise based on multidimensional public data according to claim 1, characterized in that, Generating the electronic clause sequence of the tender document includes identifying file types such as structured documents, semi-structured documents, layout documents, table documents, and image files; extracting the main text and tables; performing text recognition on image pages; and retaining page numbers, paragraph positions, and table positions to complete the electronic clause sequence of the tender document and locate its source.

4. The method for predicting the probability of winning a bid for an enterprise based on multidimensional public data according to claim 1, characterized in that, Identifying the explicit clause pairs involves segmenting the tender document by chapter title, numbering hierarchy, punctuation, and table rows, distinguishing between qualification clauses, scoring clauses, and supporting documentation requirements, and recording the explicit clause pairs when the scoring clauses contain qualification clause numbers or unique headings.

5. The method for predicting the probability of winning a bid for an enterprise based on multidimensional public data according to claim 1, characterized in that, Generating the threshold reference value includes generating the threshold reference value using the eligibility clause number, the scoring clause number, and the number of explicit references when the explicit reference text exists, and not generating the explicit clause pair when the explicit reference text does not exist.

6. The method for predicting the probability of winning a bid for an enterprise based on multidimensional public data according to claim 1, characterized in that, Constructing the project-enterprise-condition relationship network includes establishing project nodes, enterprise nodes, qualification condition nodes, and scoring condition nodes; establishing inclusion relationships between project nodes and qualification condition nodes or scoring condition nodes; possess relationships between enterprise nodes and qualification condition nodes; and participation relationships or winning relationships between enterprise nodes and historical projects. Directed back edges are established only for the explicitly stated terms. The participation relationships or winning relationships are used to generate the relationship network representation.

7. The method for predicting the probability of winning a bid for an enterprise based on multidimensional public data according to claim 1, characterized in that, Generating the conditional matching result includes: The rule engine generates rule compliance scores for directly comparable conditions in the qualification clauses or scoring clauses, generates semantic proximity scores for textual conditions in the qualification clauses or scoring clauses, and generates a comprehensive matching value based on the rule compliance scores and semantic proximity scores. When a hard qualification threshold is not met, a qualification failure flag and the reason for failure are saved.

8. The method for predicting the probability of winning a bid for an enterprise based on multidimensional public data according to claim 1, characterized in that, The generation of the historical change representation includes extracting the company's public participation records and the company's historical bidding records according to the announcement time of the target bidding project, and statistically analyzing the number of participations, the number of successful bids, the distribution of project categories, the status of qualification changes, and the continuous vacancy period according to a fixed time window. The statistical results are then used as time slice features for time-series encoding. The continuous vacancy period is the number of time slices in which there are consecutive absences of the company's public participation records and the company's historical bidding records.

9. A system for predicting the probability of winning bids for enterprises based on multi-dimensional publicly available data, comprising a computer system for processing multi-source publicly available electronic data, characterized in that, The system includes: The data acquisition and raw data retention module is used to acquire the multi-source publicly available electronic data and save the source identifier; The heterogeneous file parsing and unified format conversion module is used to generate standardized datasets; The tender condition decomposition and threshold index generation module is used to generate an electronic clause sequence in the tender document, identify explicit clause pairs, and generate threshold index values. The enterprise capability profile time-based coding module is used to generate enterprise capability profiles; The qualification threshold and scoring item hierarchical matching module is used to generate condition matching results; The historical time-series representation module is used to generate historical change representations based on the company's public participation records and historical bidding records; The target project text representation generation module is used to generate a text representation of the target project. The Project-Enterprise-Condition Relationship Network and Back-Point Edge Construction Module is used to construct the Project-Enterprise-Condition Relationship Network and establish directed back-point edges that write the threshold back-point value. The graph neural network relation representation learning module is used to adjust the retracement message using the threshold retracement value and generate a relation network representation. The attention fusion and probability output module is used to output the winning probability based on the relationship network representation, the enterprise capability profile, the condition matching result, the historical change representation, and the target project text representation.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements a method for predicting the probability of winning a bid for an enterprise based on multidimensional public data, as described in any one of claims 1 to 8.