A method for automatically extracting and comparing key information in a bid document

Abstract

This invention discloses a method for automatically extracting and comparing key information in tender documents, belonging to the field of information comparison technology. The method includes obtaining the text of the tender document, extracting technical parameters and qualification requirements as clauses to be analyzed; performing semantic comparison of the clauses to be analyzed with historical procurement data and industry standard data, and predicting a list of interested companies; constructing a heterogeneous graph with bidders as nodes; calculating the correlation strength between different bidder nodes using a graph neural network; and outputting a collusion risk index. This invention identifies biased clauses and predicts a list of interested companies at the source, filling a technological gap in risk identification during the bidding process; it overcomes the shortcomings of traditional methods that can only detect explicit equity connections, achieving penetrating identification of implicit collusion networks; and it provides review experts and regulatory agencies with a high-precision, interpretable decision support tool, significantly improving the level of intelligent risk monitoring throughout the entire bidding process.

Description

Technical Field

[0001] This invention relates to the field of information comparison technology, and in particular to a method for automatically extracting and comparing key information in tender documents. Background Technology

[0002] With the widespread adoption of electronic bidding, the amount of document data generated in bidding activities is growing exponentially. A complete bid document often contains hundreds or even thousands of pages of text, tables, images, and other modal data, containing key information such as company qualifications, technical parameters, commercial quotations, and performance certificates. How to quickly and accurately extract key information from massive amounts of bid documents and effectively compare the information of different bidders has become a core technical challenge in the bidding review process. Traditional bid document review mainly relies on manual reading and comparison, which is not only inefficient and costly, but also prone to omissions and errors due to differences in reviewers' experience and fatigue. More seriously, manual review struggles to detect bid rigging and collusion hidden within large amounts of text. To address these issues, the industry has proposed several automated information extraction and comparison methods. Existing technologies mainly adopt the following technical approaches: keyword extraction based on rule templates, which extracts specific fields from text using preset rules; and optical character recognition (OCR) based... Document parsing for identification, converting scanned documents into editable text for information extraction; and similarity calculation based on string matching, comparing the same field in different bid documents—these methods improve information processing efficiency to some extent, but still have the following technical shortcomings: First, a lack of risk identification capability for the bidding documents themselves; some bidding parties may create a competitive advantage for specific bidders by setting technical parameters that deviate from industry standards or historical procurement patterns. Second, difficulty in uncovering deep relationships between bidders. Third, insufficient semantic understanding capability; existing technologies mostly remain at the keyword matching level and cannot understand the semantic connotations of technical parameters, qualification requirements, and other clauses. Fourth, a prominent data silo problem; the authenticity verification of bid documents often requires cross-comparison with external knowledge bases such as historical procurement data, industry standard data, and enterprise qualification data, but due to heterogeneous data formats and inconsistent interface standards, this data usually exists in silos and is difficult to effectively utilize during the review process.

[0003] However, current common solutions have many drawbacks, including: the single-dimensional, superficial, and mechanical processing mode of existing technologies; focusing only on the tender documents in terms of review dimensions while ignoring the source identification of the semantic tendencies of the tender documents themselves; limiting the correlation analysis to the explicit equity structure and failing to penetrate the implicit collusion network; relying solely on keyword matching in terms of semantic understanding, making it difficult to accurately verify the deep connotations; and outputting only scattered information in terms of risk warning, lacking the ability to collude and integrate multi-dimensional data for comprehensive judgment, resulting in low accuracy and high underreporting rate in identifying bid rigging and collusion, and failing to meet the technical requirements for intelligent supervision of the entire bidding process. Summary of the Invention

[0004] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0005] In view of the problems existing in the current method for automatically extracting and comparing key information in tender documents, this invention is proposed.

[0006] Therefore, the purpose of this invention is to provide a method for automatically extracting and comparing key information in tender documents. This method is applicable to solving the problems of the single-dimensional, superficial, and mechanical processing mode of existing technologies: in terms of review, it only focuses on the tender documents and ignores the source identification of the semantic tendencies of the tender documents themselves; in terms of correlation analysis, it only stays at the explicit equity structure and cannot penetrate the implicit collusion network; in terms of semantic understanding, it only relies on keyword matching and is difficult to achieve accurate verification of deep connotations; in terms of risk warning, it only outputs scattered information and lacks the ability to collude and integrate multi-dimensional data and make comprehensive judgments, resulting in low accuracy and high underreporting rate in identifying bid rigging and collusion, and failing to meet the technical requirements of intelligent supervision of the entire bidding process.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, embodiments of the present invention provide a method for automatically extracting and comparing key information in tender documents, comprising: acquiring the text of the tender document; extracting the technical parameters and qualification requirements therein as clauses to be analyzed; performing semantic comparison of the clauses to be analyzed with historical procurement data and industry standard data, calculating the semantic deviation degree, marking clauses exceeding a threshold as biased clauses, and predicting a list of interested companies; extracting entity information from the tender documents of multiple bidders and constructing a heterogeneous graph with bidders as nodes; calculating the correlation strength between different bidder nodes through a graph neural network; performing collision fusion between the list of interested companies and the bidder correlation graph, identifying bidder combinations that appear simultaneously in the list of interested companies and have strong correlations, and outputting a bid-rigging risk index.

[0008] As a preferred embodiment of the method for automatically extracting and comparing key information in a tender document as described in this invention, the entity information includes: legal representative, shareholder information, key management personnel, contact number, email address, office address, and information on partners recorded in historical performance projects; the heterogeneous graph includes bidder nodes, personnel nodes, contact information nodes, performance project nodes, and edges representing the relationships between each node.

[0009] As a preferred embodiment of the method for automatically extracting and comparing key information in tender documents according to the present invention, the step of semantically comparing the clauses to be analyzed with historical procurement data and industry standard data, and calculating the semantic deviation degree, specifically includes the following: vectorizing the clauses to be analyzed to obtain a first feature vector; vectorizing similar clauses in historical procurement data to obtain a second feature vector set; vectorizing corresponding clauses in the industry standard database to obtain a third feature vector set; calculating the average distance between the first feature vector and the second feature vector set to obtain a historical deviation value; calculating the average distance between the first feature vector and the third feature vector set to obtain an industry deviation value; and determining the semantic deviation degree based on the weighted sum of the historical deviation value and the industry deviation value.

[0010] As a preferred embodiment of the method for automatically extracting and comparing key information in tender documents as described in this invention, the step of predicting the list of potential companies includes the following: identifying the technical parameters or qualification requirements marked as biased clauses; matching the biased clauses with a pre-constructed enterprise capability knowledge graph, which includes the company's historical bidding records, qualification certificate range, and product parameter information; filtering out companies that meet the biased clauses, sorting them from high to low according to the matching degree, and generating a list of potential companies.

[0011] As a preferred embodiment of the method for automatically extracting and comparing key information in a tender document as described in this invention, the step of calculating the association strength between different bidder nodes using a graph neural network specifically includes: inputting the heterogeneous graph into a graph neural network model, aggregating the features of neighboring nodes of each node through a message passing mechanism; updating the feature representation of each node to obtain a node embedding vector that integrates association information; and calculating the similarity between the node embedding vectors corresponding to any two bidder nodes as the association strength.

[0012] As a preferred embodiment of the method for automatically extracting and comparing key information in a tender document as described in this invention, in the step of merging the list of prospective companies with the bidder association graph to identify bidder combinations that appear simultaneously in the list of prospective companies and have a strong association, the criteria for judging the strong association include at least one of the following: the association strength between two bidder nodes exceeds a preset threshold; there is an indirect association path within two degrees between two bidder nodes; two bidder nodes share the same personnel node or contact information node.

[0013] As a preferred embodiment of the method for automatically extracting and comparing key information in a tender document as described in this invention, after the step of outputting the risk index of bid rigging, a visual risk warning report is generated. The report indicates the content of the tender document marked as biased clauses, the predicted list of interested companies, the combination of bidders with strong correlations and their correlation paths, and the risk level assessment generated based on the above information.

[0014] Secondly, to further address the aforementioned technical problems, this invention provides an automatic extraction and comparison system for key information in tender documents, comprising: a text acquisition module for acquiring the text of the tender document and extracting the technical parameters and qualification requirements as clauses to be analyzed; a semantic comparison module for performing semantic comparison between the clauses to be analyzed and historical procurement data and industry standard data, calculating the semantic deviation, marking clauses exceeding a threshold as biased clauses, and predicting a list of interested companies; an information extraction module for extracting entity information from the tender documents of multiple bidders and constructing a heterogeneous graph with bidders as nodes; a network calculation module for calculating the correlation strength between different bidder nodes using a graph neural network; and a collision analysis module for performing collision fusion between the list of interested companies and the bidder correlation graph, identifying bidder combinations that appear simultaneously in the list of interested companies and have strong correlations, and outputting a collusion risk index.

[0015] Thirdly, embodiments of the present invention provide a computer device, including a memory and a processor, wherein the memory stores a computer program, and the computer program, when executed by the processor, implements any step of the method for automatic extraction and comparison of key information in a tender document as described in the first aspect of the present invention.

[0016] Fourthly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the method for automatically extracting and comparing key information in a tender document as described in the first aspect of the present invention.

[0017] The beneficial effects of this invention are as follows: By constructing a three-in-one technical architecture of "semantic analysis at the bidding end + correlation mining at the bidding end + data collision at both ends," this invention breaks through the limitation of existing technologies that only focus on bid documents in terms of review. By semantically comparing bidding terms with historical procurement data and industry standard data, it identifies biased terms and predicts the list of potential companies from the source, filling the technical gap in risk identification in the bidding process. In terms of correlation mining, it overcomes the defect of traditional methods that can only discover explicit equity relationships. By constructing a heterogeneous graph containing multi-dimensional entity information and using graph neural networks for deep computing, it achieves penetrating identification of implicit collusion networks. In terms of judgment capability, it merges the list of potential companies analyzed at the bidding end with the correlation graph mined at the bidding end to output a quantitative risk index of bid rigging and a visualized early warning report. This provides review experts and regulatory agencies with a high-precision and interpretable decision support tool, significantly improving the level of intelligent risk monitoring throughout the entire bidding process. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart illustrating the implementation of the present invention in Example 1. Detailed Implementation

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0020] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0021] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0022] Example 1 Reference Figure 1 This is the first embodiment of the present invention, which provides a method for automatically extracting and comparing key information in tender documents, including the following steps: S1: Obtain the text of the tender document and extract the technical parameters and qualification requirements as clauses to be analyzed.

[0023] Preferably, the steps for obtaining the tender document text include: using OCR technology to perform text recognition on the scanned paper tender document or directly parsing the underlying text data of the electronic tender document (PDF / DOCX format); for PDF files, using a PDF parsing library to extract the text content and retain the chapter structure information; for image-type PDFs, first calling the OCR engine to perform text recognition, and then using a layout analysis algorithm to restore the reading order and hierarchical relationship of the text.

[0024] Furthermore, the step of extracting technical parameters and qualification requirements adopts an information extraction method based on a pre-trained language model: a named entity recognition model that integrates a bidirectional long short-term memory network and a conditional random field is constructed and fine-tuned on a corpus of tender documents labeled with technical parameters and qualification requirements, so as to achieve automatic identification and extraction of key clauses in the tender documents.

[0025] Furthermore, technical parameters include, but are not limited to, quantifiable technical requirements such as product specifications, performance indicators, material requirements, process standards, and service period; qualification requirements include, but are not limited to, entry conditions such as enterprise qualification level, registered capital scale, number of similar project performances, and personnel qualification requirements.

[0026] Specifically, to improve the accuracy of extraction, rule templates can be used as an auxiliary means: different keyword libraries and extraction rules can be preset for different types of bidding projects (such as engineering, goods and services) to verify and correct the model recognition results.

[0027] For example, taking the "Intelligent Traffic Signal System Procurement Project" of a municipal government procurement center as an example, the text of the tender document is obtained, and the technical parameters "signal light response time ≤ 50ms" and the qualification requirement "having ISO9001 quality management system certification" are extracted from it through a PDF parsing library as clauses to be analyzed.

[0028] S2: Perform semantic comparison between the terms to be analyzed and historical procurement data and industry standard data, calculate the semantic deviation, mark terms that exceed the threshold as biased terms, and predict the list of interested companies.

[0029] Preferably, the step of semantically comparing the terms to be analyzed with historical procurement data and industry standard data, and calculating the semantic deviation, is as follows: The terms to be analyzed are vectorized to obtain the first feature vector; Vectorize similar terms in historical procurement data to obtain a second feature vector set; The corresponding clauses in the industry standard database are vectorized and encoded to obtain the third feature vector set; Calculate the average distance between the first feature vector and the second feature vector set to obtain the historical deviation value; Calculate the average distance between the first and third feature vector sets to obtain the industry deviation value; The semantic deviation is determined by a weighted sum of historical deviations and industry deviations.

[0030] Furthermore, a pre-trained language model is used to convert the terms to be analyzed into dense vectors of fixed dimensions to obtain the first feature vector. Similarly, similar terms in historical procurement data are vectorized in batches to obtain the second feature vector set. The corresponding terms in the industry standard database are vectorized to obtain the third feature vector set.

[0031] Specifically, the historical procurement data comes from the bidding records of the bidding entities over the past 3-5 years, including the specific wording of the corresponding clauses in each bidding; the industry standard database comes from publicly available documents such as national standards, industry specifications, and group standards, as well as the average technical requirements of similar projects obtained by crawling mainstream bidding platforms.

[0032] Furthermore, the cosine or Euclidean distance between the first feature vector and each vector in the second feature vector set is calculated, and the average value is taken as the historical deviation value; the average distance between the first feature vector and each vector in the third feature vector set is calculated as the industry deviation value.

[0033] Furthermore, the semantic deviation is determined based on the weighted sum of historical deviation values ​​and industry deviation values. The weight coefficients can be dynamically adjusted according to the actual application scenario: for industries with rapid technological updates, the weight of historical deviation values ​​can be appropriately reduced; for industries with standardized specifications, the weight of industry deviation values ​​can be appropriately increased.

[0034] Specifically, the formula for calculating semantic deviation is as follows: ; in, This refers to semantic deviation. Historical deviation value; This is the industry deviation value; and For the preset weighting coefficients, and .

[0035] Furthermore, a dynamic threshold mechanism is set: the threshold is not fixed, but dynamically determined based on the distribution characteristics of historical data. For example, "mean + 2 standard deviations" is used as the threshold. When the semantic deviation exceeds this threshold, the clause is marked as a biased clause.

[0036] Specifically, the steps for predicting the list of interested companies are as follows: Identify the technical parameters or eligibility requirements marked as biased terms; The preferential terms are matched with a pre-built enterprise capability knowledge graph, which includes the enterprise's historical bidding records, qualification certificate scope, and product parameter information. Companies that meet the preferred criteria are selected, sorted from highest to lowest matching degree, and a list of potential companies is generated.

[0037] Furthermore, identify the technical parameters or eligibility requirements marked as biased terms, and extract the key qualifiers and numerical ranges therein.

[0038] Furthermore, the preferential terms are matched with a pre-built enterprise capability knowledge graph. The enterprise capability knowledge graph is a heterogeneous graph constructed with enterprises as nodes and enterprise capability attributes as edges. It contains multi-dimensional information such as the enterprise's historical bidding records (bid project name, bid time, bid amount), qualification certificate scope (qualification type, level, validity period), and product parameter information (product model, performance indicators, test reports).

[0039] Specifically, the matching degree between each enterprise and the preferential terms is calculated. The matching degree calculation formula is as follows: ; in, The degree of matching between each enterprise and the preferential terms; The semantic similarity between the enterprise's product parameters and the terms and conditions. This indicates whether the company has a history of winning bids for similar projects. This indicates whether the company possesses the qualifications required by the terms and conditions. , , These are the weighting coefficients.

[0040] Furthermore, companies with a matching degree exceeding a preset threshold are selected, sorted from highest to lowest matching degree, and a list of potential companies is generated.

[0041] Furthermore, the enterprise capability knowledge graph needs to be updated regularly by crawling public data sources such as business registration information, bidding and winning announcements, and qualification certification announcements to maintain the timeliness and accuracy of the graph.

[0042] For example, the above clause is semantically compared with historical procurement data (the response time of traffic lights procured by the center in the past 3 years was 80-100ms) and industry standard data (the Ministry of Public Security's industry standard requires ≤100ms). The semantic deviation of "traffic light response time ≤50ms" is calculated to be 0.92, which exceeds the threshold of 0.8 and is marked as a biased clause. The clause is matched with the enterprise capability knowledge graph and it is found that only the product parameters of Company D and Company E can meet the 50ms requirement. Moreover, Company D has won other projects of the center before, with a matching degree of 0.95. Company E has a matching degree of 0.90. A list of potential companies [Company D, Company E] is generated.

[0043] S3: Extract entity information from the bid documents of multiple bidders and construct a heterogeneous graph with bidders as nodes.

[0044] Preferably, the entity information includes: legal representative, shareholder information, key management personnel, contact number, email address, office address, and information on partners recorded in historical performance projects.

[0045] Furthermore, the heterogeneous graph includes bidder nodes, personnel nodes, contact information nodes, performance project nodes, and edges representing the relationships between the nodes.

[0046] Specifically, the method for constructing heterogeneous graphs is as follows: Graph structure definition: Heterogeneous graph G=(V,E), where node type V includes bidder nodes (B type), personnel nodes (P type, including legal persons, shareholders, and managers), contact information nodes (C type, including telephone numbers and email addresses), and performance project nodes (R type, including historical performance project names); edge type E includes: Job Relationship Edge: Connects the personnel node to the bidder node, indicating the position the personnel holds in the bidder; Shareholding relationship edge: connects shareholder nodes and bidder nodes, indicating the shareholding relationship; Contact information edge: Connects the contact information node to the bidder node, indicating that the contact information belongs to the bidder; Performance Relationship Edge: Connects the performance project node and the bidder node, indicating that the bidder has participated in the project; Partnership edge: connects two performance project nodes, representing different participants in the same project; Shared relationship edge: connects two bidder nodes, indicating that they share the same personnel node or contact information node (automatically generated implicit edge).

[0047] Furthermore, to handle large-scale bidding scenarios, a distributed graph storage architecture can be adopted to store heterogeneous graphs in graph databases (such as Neo4j and JanusGraph), supporting efficient node and relationship queries.

[0048] Specifically, after the graph is constructed, necessary graph preprocessing is performed, including filtering isolated nodes, initializing edge weights (initial weights can be preset according to the strength of association), and initializing feature vectors (assigning initial features to each node, such as using enterprise business information codes to represent bidder nodes and name vectors to represent personnel nodes).

[0049] Furthermore, to improve the robustness of the model, the node features of the heterogeneous graph can be normalized, and the edges can be sampled to control the graph size.

[0050] Specifically, the feature dimensions of the graph neural network input are uniformly mapped to 128 or 256 dimensions for subsequent calculations.

[0051] Furthermore, while constructing the heterogeneous graph, the metadata information of each node is recorded, including node ID, node type, original text information, etc., for subsequent visualization and tracing.

[0052] For example, entity information was extracted from the bid documents of 5 bidders to construct a heterogeneous graph. It was found that the technical person in charge of Company D was Zhang San, and the supervisor of Company E was Li Si. Zhang San and Li Si had worked together at Company F. At the same time, the contact number of Company D was 138****1234, which also appeared in the bid document of Company G as an emergency contact.

[0053] S4: Calculate the correlation strength between nodes of different bidders using a graph neural network.

[0054] Preferably, the step of calculating the correlation strength between different bidder nodes using a graph neural network specifically includes: The heterogeneous graph is input into the graph neural network model, and the features of each node's neighbor nodes are aggregated through the message passing mechanism; Update the feature representations of each node to obtain the node embedding vector that incorporates the associated information; Calculate the similarity between the node embedding vectors corresponding to any two bidder nodes, and use this as the association strength.

[0055] Specifically, select a graph neural network model suitable for heterogeneous graphs, such as R-GCN (Relational Graph Convolutional Network), HAN (Heterogeneous Graph Attention Network), or GAT (Graph Attention Network), and use the heterogeneous graph constructed by S3 as input, with the initial feature vector of each node as the model input feature.

[0056] Specifically, through a message passing mechanism, each node aggregates the feature information of its neighboring nodes. Taking R-GCN as an example, the update formula for node i is: ; in, For nodes In the Layer feature representation; For nodes In the Layer feature representation; Represents a set of relation types; Represents a node In relationship The set of neighboring nodes; This is the normalization constant; and These are trainable parameters; This is the activation function.

[0057] Specifically, after passing through an L-layer graph convolutional network, each node aggregates the information of its neighboring nodes within its L-hop range to obtain a node embedding vector that integrates multi-hop association information. Typically, L is 2 or 3, which can cover indirect associations within two or three degrees.

[0058] Specifically, for any two bidder nodes and Calculate the similarity between the corresponding node embedding vectors as the association strength. The similarity can be expressed as cosine similarity or vector dot product. ; in, For bidder nodes With nodes The strength of the correlation between them; For bidder nodes The final node embedding vector after being updated by the graph neural network; For bidder nodes The final node embedding vector after being updated by the graph neural network.

[0059] Specifically, the graph neural network model needs to be pre-trained. The training data can use known cases of bid rigging in historical bidding projects as positive samples, and randomly select unrelated combinations of bidders as negative samples. Contrastive learning or triple loss can be used to optimize the model.

[0060] Furthermore, in order to handle the feature heterogeneity of different types of nodes in heterogeneous graphs, a graph neural network guided by meta-paths can be used. Several meaningful meta-paths are preset (such as "bidder-personnel-bidder", "bidder-performance project-bidder"), the node representation under each meta-path is calculated separately, and then fused through an attention mechanism.

[0061] Specifically, after the model training is completed, the association strength is calculated for each pair of all bidder nodes to obtain an N×N association strength matrix, where N is the number of bidders.

[0062] Furthermore, to improve computational efficiency, for large-scale bidding scenarios (such as N>1000), an approximate nearest neighbor search algorithm can be used to calculate only the association strength of the K most relevant nodes to each node, without performing full pairwise calculations.

[0063] For example, the heterogeneous graph is input into the R-GCN model, and after two layers of convolution, the association strength between the bidder nodes is calculated. The association strength between Company D and Company E is 0.75 (association through personnel), the association strength between Company D and Company G is 0.89 (association through telephone), and the association strength between Company E and Company G is 0.12 (no obvious association).

[0064] S5: Collision and fusion of the list of prospective companies and the bidder association graph to identify bidder combinations that appear in the list of prospective companies and have strong associations, and output the risk index of bid rigging.

[0065] Specifically, in the step of merging the list of prospective companies with the bidder association graph to identify bidder combinations that appear simultaneously in the list of prospective companies and have strong associations, The criteria for determining strong association include at least one of the following: the association strength between two bidder nodes exceeds a preset threshold; there is an indirect association path within two degrees between two bidder nodes; or the two bidder nodes share the same personnel node or contact information node.

[0066] Specifically, the formula for calculating the risk index R is as follows: ; in, The risk index for bid rigging comprehensively reflects the overall risk level of this bidding project; For each identified risky group, there is a portfolio of bidders. This refers to a specific risk group, which includes several bidder nodes; This represents the total number of bidders for this project. For gang The percentage of interested companies is the number of companies in the group that appear on the list of interested companies, divided by the size of the group.

[0067] Furthermore, the visualization report presents the relationship map in a graphical way, using different colors to highlight enterprise nodes and ordinary bidder nodes, using line thickness to indicate the relationship strength, and displaying a specific relationship path description when the mouse hovers over it, making it easier for review experts to intuitively understand the source of risk.

[0068] Furthermore, after outputting the risk index of bid rigging, a visual risk warning report is generated. The report indicates the contents of the tender documents marked as biased clauses, the predicted list of interested companies, the combination of bidders with strong correlations and their correlation paths, and the risk level assessment generated based on the above information.

[0069] For example, by merging the list of potential companies [Company D, Company E] with the association graph, it is identified that the association strength between Company D and Company G is 0.89 > the threshold of 0.8, and there is a second-degree association path between Company D and Company E (D-Zhang San-Li Si-E). At the same time, both Company D and Company E are in the list of potential companies. The collusion risk index R=0.82 is output, and a visualization report is generated, highlighting the biased clause "50ms", potential companies D and E, as well as the telephone association path of DG and the personnel association path of DE, and giving a "high risk" rating.

[0070] In summary, this invention constructs a three-in-one technical architecture of "semantic analysis at the bidding end + correlation mining at the bidding end + data collision at both ends." In terms of review, it breaks through the limitations of existing technologies that only focus on bid documents. By semantically comparing bidding terms with historical procurement data and industry standard data, it identifies biased clauses and predicts the list of potential companies from the source, filling the technical gap in risk identification during the bidding process. In terms of correlation mining, it overcomes the deficiency of traditional methods that can only discover explicit equity relationships. By constructing a heterogeneous graph containing multi-dimensional entity information and using graph neural networks for deep computation, it achieves penetrating identification of implicit collusion networks. In terms of judgment capabilities, it merges the list of potential companies analyzed at the bidding end with the correlation graph mined at the bidding end, outputting a quantitative risk index for bid rigging and a visualized early warning report. This provides review experts and regulatory agencies with a high-precision, interpretable decision support tool, significantly improving the level of intelligent risk monitoring throughout the entire bidding process.

[0071] Example 2, an embodiment of the present invention, provides an automatic extraction and comparison system for key information in tender documents, comprising: a text acquisition module for acquiring the text of the tender document and extracting the technical parameters and qualification requirements therein as clauses to be analyzed; a semantic comparison module for performing semantic comparison of the clauses to be analyzed with historical procurement data and industry standard data, calculating the semantic deviation degree, marking clauses exceeding a threshold as biased clauses, and predicting a list of interested companies; an information extraction module for extracting entity information from the tender documents of multiple bidders and constructing a heterogeneous graph with bidders as nodes; a network calculation module for calculating the correlation strength between different bidder nodes using a graph neural network; and a collision analysis module for performing collision fusion between the list of interested companies and the bidder correlation graph, identifying bidder combinations that appear simultaneously in the list of interested companies and have strong correlations, and outputting a collusion risk index.

[0072] Example 3 is an embodiment of the present invention, which differs from the previous embodiment in that: If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0073] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0074] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0075] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0076] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for automatically extracting and comparing key information in tender documents, characterized in that: include: Obtain the text of the tender document and extract the technical parameters and qualification requirements as clauses to be analyzed. The clauses to be analyzed are semantically compared with historical procurement data and industry standard data. The semantic deviation is calculated, and clauses exceeding the threshold are marked as biased clauses. A list of interested companies is predicted. Extract entity information from the bid documents of multiple bidders and construct a heterogeneous graph with bidders as nodes; The correlation strength between nodes of different bidders is calculated using a graph neural network; By merging the list of prospective companies with the bidder association graph, we can identify bidder combinations that appear in both the prospective company list and have strong associations, and output a bid-rigging risk index.

2. The method for automatically extracting and comparing key information in a tender document as described in claim 1, characterized in that: The entity information includes: legal representative, shareholder information, key management personnel, contact number, email address, office address, and information on partners recorded in historical performance projects; The heterogeneous graph includes bidder nodes, personnel nodes, contact information nodes, performance project nodes, and edges representing the relationships between the nodes.

3. The method for automatically extracting and comparing key information in a tender document as described in claim 1, characterized in that: The steps of semantically comparing the terms to be analyzed with historical procurement data and industry standard data, and calculating the semantic deviation, are detailed below: The terms to be analyzed are vectorized to obtain the first feature vector; Vectorize similar terms in historical procurement data to obtain a second feature vector set; The corresponding clauses in the industry standard database are vectorized and encoded to obtain the third feature vector set; Calculate the average distance between the first feature vector and the second feature vector set to obtain the historical deviation value; Calculate the average distance between the first and third feature vector sets to obtain the industry deviation value; The semantic deviation is determined based on the weighted sum of the historical deviation value and the industry deviation value.

4. The method for automatically extracting and comparing key information in a tender document as described in claim 3, characterized in that: The specific steps for predicting the list of interested companies are as follows: Identify the technical parameters or eligibility requirements marked as biased terms; The preferential terms are matched with a pre-built enterprise capability knowledge graph, which includes the enterprise's historical bidding records, qualification certificate scope, and product parameter information. Select companies that meet the aforementioned preference criteria, sort them from highest to lowest matching degree, and generate a list of potential companies.

5. The method for automatically extracting and comparing key information in a tender document as described in claim 4, characterized in that: The step of calculating the correlation strength between different bidder nodes using a graph neural network specifically includes: The heterogeneous graph is input into a graph neural network model, and the features of each node's neighbor nodes are aggregated through a message passing mechanism; Update the feature representations of each node to obtain the node embedding vector that incorporates the associated information; Calculate the similarity between the node embedding vectors corresponding to any two bidder nodes, and use this as the association strength.

6. The method for automatically extracting and comparing key information in a tender document as described in claim 5, characterized in that: In the step of performing collision and fusion analysis between the list of prospective companies and the bidder association graph to identify bidder combinations that appear simultaneously in the list of prospective companies and have a strong association, The criteria for determining strong association include at least one of the following: the association strength between two bidder nodes exceeds a preset threshold; there is an indirect association path within two degrees between two bidder nodes; or the two bidder nodes share the same personnel node or contact information node.

7. The method for automatically extracting and comparing key information in a tender document as described in claim 1, characterized in that: After the step of outputting the risk index of bid rigging, a visual risk warning report is generated. The report indicates the contents of the tender documents marked with biased clauses, the predicted list of interested companies, the combination of bidders with strong correlations and their correlation paths, and the risk level assessment generated based on the above information.

8. A system for automatically extracting and comparing key information in a bid document, based on the method for automatically extracting and comparing key information in a bid document as described in any one of claims 1 to 7, characterized in that: include, The text acquisition module is used to acquire the text of the tender document and extract the technical parameters and qualification requirements as clauses to be analyzed. The semantic comparison module is used to perform semantic comparison between the terms to be analyzed and historical procurement data and industry standard data, calculate the semantic deviation, mark terms that exceed the threshold as biased terms, and predict the list of interested companies. The information extraction module is used to extract entity information from the bid documents of multiple bidders and construct a heterogeneous graph with bidders as nodes. The network computing module is used to calculate the correlation strength between nodes of different bidders using a graph neural network; The collision analysis module is used to perform collision fusion between the list of prospective companies and the correlation graph of bidders, identify combinations of bidders that appear in the list of prospective companies and have strong correlations, and output a risk index of bid rigging.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the method for automatically extracting and comparing key information in a tender document as described in any one of claims 1 to 7.

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 the steps of the method for automatically extracting and comparing key information in a tender document as described in any one of claims 1 to 7.

Images