Bidding document auxiliary review method and system based on artificial intelligence

Abstract

The invention discloses a bidding document auxiliary review method and system based on artificial intelligence, and relates to the technical field of bidding and bidding, and the method comprises the steps: obtaining a bidding document and a bidding document corresponding to the bidding document, and determining a corresponding bidding demand according to the bidding document; respectively obtaining text information about the bidding content and picture information about the bidding qualification in the bidding file; and respectively identifying the text information and the picture information, and matching and comparing the identification result with the bidding demand so as to inspect the bidding document. According to the invention, the problems of manual bid evaluation and poor accuracy in the prior art are solved.

Description

A method and system for assisting in the review of tender documents based on artificial intelligence Technical Field

[0001] This invention relates to the field of bidding technology, and in particular to a method and system for assisting in the review of bid documents based on artificial intelligence. Background Technology

[0002] In bidding activities, bid evaluation, as the core link between bidding and winning the bid, is a key means to ensure the rational allocation of resources and maintain fair market competition. It is widely used in various public resource transactions and business operations, such as engineering construction, medical equipment procurement, and information technology services. Manual bid evaluation is the traditional review model that the industry has long relied on.

[0003] However, with the expansion of bidding projects, the increase in the complexity of requirements, and the growing number of bidders, modern bidding activities often involve single bid documents that are hundreds of pages long, and large projects can have more than 10 bidders, resulting in the total amount of text and image information that evaluators need to process exceeding several thousand pages. Furthermore, the diversity of bidding requirements places extremely high demands on the accuracy of review, but manual operation is susceptible to various judgment biases due to physiological fatigue and limitations in experience. Therefore, the traditional manual bid evaluation model suffers from low efficiency and accuracy. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide an artificial intelligence-based method and system for assisting in the review of tender documents, aiming to solve the problems of poor efficiency and accuracy of manual bid evaluation in the prior art.

[0005] This invention proposes an artificial intelligence-based method for assisting in the review of tender documents. The method comprises: obtaining the tender documents and the corresponding bidding documents; determining the corresponding bidding requirements based on the bidding documents; obtaining textual information about the bidding content and image information about the bidding qualifications from the tender documents; recognizing the textual and image information respectively; and matching and comparing the recognition results with the bidding requirements to review the tender documents.

[0006] Furthermore, the aforementioned AI-based bid document auxiliary review method includes the following steps for recognizing textual and image information: using a pre-trained language model to perform semantic segmentation on the text representing bidding requirements, extracting corresponding requirement items, and labeling each requirement item with constraint type, data type, and association weight to construct a structured bidding requirement knowledge graph; extracting key information from the bid document through entity recognition, and then identifying the relationships between key information through relation extraction, wherein semantic completion is performed on ambiguous or unclear statements using the domain knowledge graph; converting the recognition results into a structured data format corresponding to the bidding requirement knowledge graph; extracting key areas from the image information through image preprocessing, and recognizing the text content within the key areas using OCR and layout analysis techniques; recognizing the anti-counterfeiting texture of the certificate border in the image information through edge detection, verifying the authenticity of the official seal in the image information through a seal recognition model, and determining whether the image has been photoshopped using a tamper detection algorithm; and associating the text content recognized in the image with the recognition results of the text information in the bid document.

[0007] Furthermore, the aforementioned AI-based bid document review method further includes the following steps for reviewing bid documents by matching and comparing the identification results with bidding requirements: comparing the values ​​and formats of rigid constraints in the identification results, specifically comparing the values, codes, and dates in the bid documents with the threshold ranges of the bidding requirements field-by-field to generate a binary judgment of pass or fail, and automatically marking deviation values; encoding the contextual semantics of bidding requirements and corresponding bid responses using a Transformer architecture for flexible descriptive items in the identification results, and calculating matching scores by combining domain word vectors; mapping entities and relationships in the bid information to the association paths in the bidding requirements knowledge graph, generating association matching scores through path integrity verification, and verifying the consistency of the bid information; and labeling the risk levels of matching items close to the threshold and fuzzy response items in the identification results, and automatically generating supplementary inquiry suggestions.

[0008] Furthermore, the aforementioned AI-based bid document assisted review method, after the steps of recognizing text and image information respectively and matching and comparing the recognition results with bidding requirements to review the bid documents, further includes: assigning a unique identifier to each requirement item, associating it with its hierarchical path, requirement item type, original matching score, and deviation value in the bidding requirement knowledge graph; recording the original position of the bid document, key text fragments, and attention weight distribution of flexible description items corresponding to each requirement item; converting the multi-level structure of the bidding requirement knowledge graph into a horizontal axis coordinate, where the first-level category occupies the basic coordinate position, and the next-level category is displayed indented under the previous-level category; defining the vertical axis according to the matching score, using the matching score as a quantitative indicator, and mapping each requirement item to the corresponding vertical and horizontal axis positions according to its matching score and corresponding level; setting different color mappings according to different matching scores and performing special state rendering for special requirement items to obtain a heatmap review report; when the mouse hovers over the coordinate point of a requirement item's heatmap, triggering a pop-up window to display the text comparison between the original requirement text and the original bid document, the automatically calculated deviation value, and the evidence location.

[0009] Furthermore, the aforementioned AI-based bid document assisted review method, after the step of triggering a pop-up window to display the text comparison between the original requirement and the original bid document, automatically calculated deviation value, and evidence location when the mouse hovers over the coordinate point of a requirement item's heatmap, further includes: constructing a multi-bid entity association graph for all bid documents of the same bidding project using a graph neural network, wherein the nodes of the multi-bid entity association graph are the core information in the bid documents, and the edges of the multi-bid entity association graph are associated based on the similarity of information; setting an abnormal association threshold, and when an abnormal association exceeding the threshold is detected, adding a warning mark next to the corresponding requirement item in the heatmap review report; clicking the warning mark triggers a collaborative review view, synchronously displaying the comparison of corresponding information in multiple associated bid documents.

[0010] Furthermore, the aforementioned AI-based bid document assisted review method, wherein the step of setting an abnormal association threshold and adding a warning mark next to the corresponding requirement item in the heatmap review report when an abnormal association exceeding the threshold is detected includes: using historical bid documents without abnormalities as positive samples, generating negative samples through data augmentation, and training a model that can identify normal association features and abnormal association features; for the multi-bid entity association graph of new bidding projects, determining abnormal associations by calculating the abnormal confidence of each association edge through the model; extracting common erroneous expressions or rare overlapping phrases in the text of abnormal associations through keyword alignment and semantic role labeling, and generating a word vector similarity heatmap; and using a pixel-level difference algorithm to match the images of qualification certificates and contract scans involved in abnormal associations with forgery features. The system generates a difference mask image and a forgery feature comparison report; retrieves the upload logs, modification records, and IP address trajectories of the tender documents to generate an operation timeline, marking abnormal time-series features such as multiple files uploaded consecutively under the same IP and highly overlapping modification timestamps; it links to the enterprise credit database to extract implicit association data of equity associations, historical cooperation projects, and legal representatives of abnormal bidders; it uses a causal inference algorithm to perform correlation analysis on word vector similarity heatmaps, difference mask images, forgery feature comparison reports, abnormal time-series features, and implicit association data, marking the causal and correlation relationships between evidence; and it constructs a combined interface in the collaborative review view, where the combined interface uses abnormal association points as core nodes, radiating evidence branches outward according to evidence strength, and using solid lines to mark causal relationships and dashed lines to mark correlation relationships.

[0011] Furthermore, the aforementioned AI-based method for assisting in the review of tender documents, wherein the step of using a causal inference algorithm to perform correlation analysis on word vector similarity heatmaps, difference mask maps, forgery feature comparison reports, abnormal time-series features, and implicit correlation data, and marking the causal and correlation relationships between evidence, includes: standardizing the word vector similarity heatmaps, difference mask maps, forgery feature comparison reports, abnormal time-series features, and implicit correlation data; using an attention mechanism to weightedly fuse the standardized features to generate a unified evidence feature vector; extracting typical causal relationship rules in the bidding field based on historical review cases, and constructing a causal rule base for the bidding field; using strong abnormal features in the unified evidence feature vector as triggering conditions, matching corresponding causal rules from the rule base to generate a candidate hypothesis set consisting of multiple candidate hypotheses; retaining the top 3 high-priority hypotheses for the verification stage according to the strength of the abnormal features corresponding to the candidate hypotheses; simulating intervention operations on the dependent variables in the high-priority hypotheses, and observing... The process involves observing whether the result variable changes to determine if a causal relationship exists; eliminating the interference of confounding variables, such as those common to the industry and those specific to the project, on the association results; constructing a Bayesian network containing the dependent variable, result variable, and confounding variables; calculating the confidence level of the causal relationship and the correlation coefficient of the correlation relationship to label the causal and correlation relationships between evidence; and standardizing the word vector similarity heatmap, difference mask map, forgery feature comparison report, abnormal time-series features, and implicit association data, including: extracting BERT semantic vectors of common erroneous statements and rare overlapping phrases from the word vector similarity heatmap, and calculating phrase-level and paragraph-level two-layer similarity matrices; extracting forgery feature vectors from the difference mask map and forgery feature comparison report, quantifying the matching degree of forgery features, and generating an image abnormal feature matrix; converting the operation timeline into a time-series event sequence, and constructing a time-series feature vector using time difference quantification indicators and IP association degree; and converting implicit association data into association strength values ​​to form a structured association matrix.

[0012] Another objective of this invention is to provide an artificial intelligence-based bid document review system, the system comprising: an acquisition module for acquiring bid documents and corresponding tender documents, and determining corresponding tender requirements based on the tender documents; a determination module for acquiring textual information about the bid content and image information about the bid qualifications in the bid documents; and a review module for recognizing the textual and image information, and matching and comparing the recognition results with the tender requirements to review the bid documents.

[0013] Another object of the present invention is to provide a readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0014] Another object of the present invention is to provide an electronic device including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the program to implement the steps of the method described above.

[0015] This invention obtains the tender documents and their corresponding bidding documents, and determines the corresponding bidding requirements based on the bidding documents. It then acquires textual information about the bidding content and image information about the bidding qualifications from the tender documents. The textual and image information are then identified, and the identification results are matched and compared with the bidding requirements to review the tender documents. This automated identification process improves efficiency and avoids errors inherent in human judgment. It solves the problems of poor efficiency and accuracy in manual bid evaluation in existing technologies. Attached Figure Description

[0016] Figure 1 is a flowchart of the artificial intelligence-based bid document auxiliary review method in the first embodiment of the present invention; Figure 2 is a structural block diagram of the artificial intelligence-based bid document auxiliary review system in the third embodiment of the present invention.

[0017] The following detailed description, in conjunction with the accompanying drawings, will further illustrate the present invention. Detailed Implementation

[0018] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Several embodiments of the invention are illustrated in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0019] It should be noted that when a component is said to be "fixed to" another component, it can be directly on the other component or there may be an intervening component. When a component is said to be "connected to" another component, it can be directly connected to the other component or there may be an intervening component. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0021] Please refer to Figure 1 for Example 1, which shows the artificial intelligence-based bid document auxiliary review method in the first embodiment of the present invention. The method includes steps S10 to S12.

[0022] Step S10: Obtain the tender documents and the corresponding bidding documents, and determine the corresponding bidding requirements based on the bidding documents.

[0023] This process involves obtaining the tender documents to be reviewed and their corresponding bidding documents. By analyzing the bidding documents, the core requirements of the bidding party (such as qualification thresholds, technical parameters, and commercial terms) are clarified. The tender documents are the objects to be reviewed, and the bidding documents are the basis for the review. By analyzing the bidding documents, the bidding requirements are extracted, which is equivalent to establishing evaluation criteria for subsequent reviews.

[0024] Step S11: Obtain the text information about the bidding content and the image information about the bidding qualifications from the bid documents.

[0025] The text information of the tender documents includes unstructured text such as technical solutions, performance descriptions, and business responses, while the image information of the tender qualifications includes visual materials such as business licenses, qualification certificates, and scanned copies of contracts. In view of the heterogeneous nature of the tender documents, they are split into text information and image information to achieve information classification and processing. The purpose is to adapt to the recognition capabilities of different forms of information and avoid recognition errors caused by information mixing.

[0026] Specifically, the steps for recognizing textual and image information respectively include: using a pre-trained language model to perform semantic segmentation on the text representing bidding requirements, extracting corresponding requirement items, and labeling each requirement item with constraint type, data type, and association weight to construct a structured bidding requirement knowledge graph; extracting key information from the bid documents through entity recognition, and then recognizing the association between key information through relation extraction, wherein semantic completion is performed on ambiguous or unclear expressions using the domain knowledge graph; converting the recognition results into a structured data format corresponding to the bidding requirement knowledge graph; extracting key areas from the image information through image preprocessing, and recognizing the text content within the key areas using OCR and layout analysis techniques; recognizing the anti-counterfeiting texture of the certificate border in the image information through edge detection, verifying the authenticity of the official seal in the image information through a seal recognition model, and determining whether the image has any Photoshop traces through a tampering detection algorithm; and associating the text content recognized in the image with the recognition results of the text information in the bid documents.

[0027] Specifically, for text information recognition, a pre-trained language model is used to perform semantic segmentation on the bidding requirement text, extract core requirement items and label constraint types, data types and association weights, and construct a structured bidding requirement knowledge graph to record the logical relationships between requirements; at the same time, key information is extracted from the bidding text through entity recognition, the relationships between information are identified through relation extraction, and semantic completion is performed on fuzzy expressions in combination with the domain knowledge graph, and finally the recognition results are converted into a structured format corresponding to the bidding requirement knowledge graph.

[0028] For image information recognition, the solution extracts key areas through image preprocessing and recognizes text within the areas by combining OCR and layout analysis technologies. It further recognizes the anti-counterfeiting texture of certificate borders through edge detection, verifies the authenticity of official seals through seal recognition models, and judges PS traces through tamper detection algorithms, thus achieving dual processing of text extraction and anti-counterfeiting verification of image information. Finally, the recognized text in the image is correlated with the recognition results of the bidding text.

[0029] Specifically, BERT uses semantic segmentation to accurately extract core requirements such as "registered capital ≥ 10 million" and "possessing Class A qualification," labeling constraint types (rigid / flexible / related), data types (numerical / text / date), and association weights. It then constructs a knowledge graph in an "entity-relationship" format to record the logical dependencies between requirements. Furthermore, it extracts key entities such as company name, qualification level, and performance amount through relation extraction, identifying associations such as "performance implementation time is later than qualification acquisition time." For vague expressions like "similar project experience," it combines the domain knowledge graph to complete them into quantitative information. Finally, it converts the identification results into a structured data format corresponding to the bidding requirements knowledge graph, achieving data alignment between bidding and requirements, and uniformly outputting JSON-LD, etc. The format ensures that the bidding entity and the demand entity can be directly compared; image preprocessing includes denoising, correction, and region segmentation (such as splitting the title area, content area, and seal area of ​​the certificate); layout analysis locates the text position; OCR (such as PP-OCRv4) accurately recognizes text such as "Qualification Level: Grade A" and "Valid until 2026"; edge detection compares the similarity between the certificate border and the filing texture; the seal recognition model verifies the consistency between the official seal text, shape, and filing; tamper detection identifies PS traces through noise distribution differences; the text content recognized in the image is associated with the recognition results of the text information in the bidding documents, for example, associating the "Grade A" recognized in the qualification certificate image with the "Grade A" claimed in the bidding text, forming a two-way binding of "text statement - image evidence".

[0030] Furthermore, in some optional embodiments of the present invention, the step of reviewing the bid documents by matching and comparing the identification results with the bidding requirements further includes: comparing the values / formats of the rigid constraint items in the identification results, wherein the values, codes, and dates in the bid documents are compared field by field with the threshold range of the requirements of the bidding requirements to generate a binary judgment of pass or fail, and the deviation value is automatically marked; encoding the contextual semantics of the bidding requirements and the corresponding bid responses for the flexible description items in the identification results using the Transformer architecture, and calculating the matching score by combining domain word vectors; mapping the entities and relationships in the bid information to the association paths in the bidding requirements knowledge graph for the association requirements, generating an association matching score through path integrity verification, and verifying the consistency of the bid information; and marking the risk level of the matching items and fuzzy response items close to the threshold in the identification results, and automatically generating supplementary inquiry suggestions.

[0031] Specifically, for rigid constraints, a "field-by-field comparison of numerical values ​​and formats" is adopted, comparing the numerical values, codes, and dates in the bidding information with the required threshold range to generate a "pass / fail" binary judgment and mark the deviation value. For flexible descriptive items (such as technical solution fit and service commitments), the Transformer architecture is used to encode the contextual semantics of the bidding requirements and bid responses, and a matching score is calculated by combining domain word vectors to achieve deep semantic comparison. For related requirements, bidding entities and relationships are mapped to the association paths in the bidding requirements knowledge graph, and an association matching score is generated through path integrity verification to verify the logical consistency of information. In addition, the solution also marks the risk level of matching items close to the threshold (such as a flexible score of 68 points and a threshold of 70 points) and ambiguous response items (such as "basically meets the requirements") and automatically generates supplementary inquiry suggestions.

[0032] Specifically, for rigid items (such as registered capital and certificate number), precise comparison rules are designed. Numerical fields calculate the deviation of "bid value - demand threshold" (e.g., "5 million - 10 million = -5 million"). Encoded fields verify format consistency (e.g., whether the certificate number conforms to the "industry code - year - serial number" rule). Date-based fields determine whether the data is within the demand range (e.g., whether the performance is within the last 3 years). The final output is a binary result and deviation details. For flexible items (such as technical solutions and service commitments), semantic comparison uses a Transformer architecture (e.g., RoBERTa) to encode contextual semantics to capture deep relationships. Domain word vectors (e.g., "smart park" terminology vectors) improve industry adaptability, and cosine similarity is used to further enhance the comparison. Calculate a matching score from 0-100% (e.g., the similarity between the requirement "overall design" and the bid "design and implementation" is 85%); perform logical verification on related items (e.g., "qualification + performance" and "joint bidding + agreement")—map the bidding entities and relationships to the requirement knowledge graph, verify whether the path is complete (e.g., whether both "qualification" and "performance" exist) and whether the logic is consistent (e.g., whether "joint bidding" is associated with the "agreement" entity). A complete path scores 100 points, a missing entity scores 50 points, and a logical contradiction scores 0 points; items close to the threshold (e.g., a flexible score of 68 points and a threshold score of 70 points) and fuzzy response items (e.g., "basically compliant") are marked as "medium risk," and inquiry suggestions are generated (e.g., "please provide scanned copies of contracts for similar projects").

[0033] Step S12: Recognize the text information and image information respectively, and match and compare the recognition results with the bidding requirements to review the bid documents.

[0034] Finally, artificial intelligence technology is used to identify and analyze the text and image information respectively, and the extracted key information is matched and compared with the previously determined bidding requirements to complete the review of the compliance and suitability of the bid documents.

[0035] In summary, the AI-based bid document review method described in the above embodiments of the present invention obtains the bid document and the corresponding tender document, and determines the corresponding tender requirements based on the tender document; it obtains text information about the bid content and image information about the bid qualifications from the bid document; it identifies the text information and image information respectively, and matches and compares the identification results with the tender requirements to review the bid document. The automated identification improves efficiency and avoids the judgment errors inherent in human intervention. This solves the problems of poor efficiency and accuracy of manual bid evaluation in existing technologies.

[0036] Example 2 This example also proposes an AI-based method for assisting in the review of tender documents. The difference between this AI-based method and the one in Example 1 lies in the following: after the steps of recognizing text and image information respectively, and matching and comparing the recognition results with bidding requirements to review the tender documents, the method further includes: assigning a unique identifier to each requirement item, associating it with its hierarchical path, requirement item type, original matching score, and deviation value in the bidding requirements knowledge graph; recording the original text location, key text fragments, and flexible description of each requirement item in the tender document. The attention weight distribution of the items is described; the multi-level structure of the bidding requirement knowledge graph is converted into a horizontal axis coordinate, where the first-level category occupies the basic coordinate position, and the next-level category is displayed indented under the previous-level category; the vertical axis is defined according to the matching score, and each requirement item is mapped to the corresponding vertical and horizontal axis positions according to its matching score and corresponding level; different color mappings are set according to different matching scores, and special state rendering is performed for special requirement items to obtain a heat map review report; when the mouse hovers over the coordinate point of a requirement item's heat map, a pop-up window is triggered to display the text comparison between the original requirement and the original bid, the automatically calculated deviation value, and the evidence location.

[0037] First, assign a unique identifier to each requirement item and associate it with its hierarchical path in the bidding requirement knowledge graph (such as "business requirements, qualification threshold, registered capital"), requirement type (rigid / flexible / related type), original matching score (from rigid / flexible / related type score) and deviation value; at the same time, record the original text location of each requirement item in the tender document (such as "P5 paragraph 2"), key text fragments (such as "registered capital of 5 million") and attention weight distribution of flexible description items to retain evidence clues.

[0038] Secondly, coordinate mapping is designed: the multi-level structure of the knowledge graph is converted into a horizontal axis coordinate (first-level categories are positioned at the base, and lower-level categories are displayed with indentation). For example, first-level categories (such as "Business Requirements") occupy 0-100 pixels on the horizontal axis, second-level categories (such as "Qualification Threshold") are indented by 10-90 pixels, and third-level categories (such as "Registered Capital") are indented by 20-80 pixels, intuitively presenting the demand logic. The matching score is defined as the vertical axis coordinate (0-100% quantitative indicator), and each demand item is mapped to the corresponding coordinate point according to its level and score. Thirdly, a heatmap review report is generated through color mapping and special state rendering, such as green for 80-100 points, yellow for 50-79 points, red for <50 points, red flashing for rigid failure items, and orange borders for items close to the threshold.

[0039] Furthermore, in some preferred embodiments of the present invention, when converting the multi-level knowledge graph structure into a horizontal axis coordinate, if the coordinates of multi-level categories overlap due to dense demand items, the categories in the overlapping area are intelligently folded, displaying only the core category names; the coordinate areas corresponding to high-weight demand items are split, and their display width is expanded to ensure that key information is not obscured. When using the matching score vertical axis mapping, outliers in the matching scores (such as extremely high / low scores due to recognition errors) can be identified through box plot analysis. These outliers are standardized and corrected before being mapped to the vertical axis, and correction markers and the basis for correction are annotated next to the coordinates to avoid extreme values ​​interfering with review and judgment.

[0040] In addition, in some optional embodiments of the present invention, the demand item trajectory of each bid document can be obtained based on the coordinates of the demand items in the heat map report, and the demand item trajectory of all bidders' bid documents under the same bidding project can be obtained using the same method. The similarity of the demand item trajectories of any different bidders can be calculated using the DTW algorithm, and the presence of bid rigging in the current bid document can be determined based on the magnitude of the similarity.

[0041] Finally, the interactive functionality was designed: a pop-up window is triggered when the mouse hovers over the coordinates, displaying a comparison between the requirements and the original tender document, automatically calculated deviation values, and evidence location (e.g., "Tender document P5, paragraph 2"). A "Jump to P5" button is provided to locate the evidence.

[0042] In addition, in some optional embodiments of the present invention, after the step of triggering a pop-up window to display the text comparison between the original requirement and the original bid when the mouse hovers over the coordinate point of the heatmap of a requirement item, automatically calculating the deviation value, and locating the evidence, the method further includes: constructing a multi-bid entity association graph for all bid documents of the same bidding project through a graph neural network, wherein the nodes of the multi-bid entity association graph are the core information in the bid documents, and the edges of the multi-bid entity association graph are associated based on the similarity of the information; setting an abnormal association threshold, and when an abnormal association exceeding the threshold is detected, adding a warning mark next to the corresponding requirement item in the heatmap review report; clicking the warning mark triggers a collaborative review view, which synchronously displays the comparison of the corresponding information in the multiple associated bid documents.

[0043] Specifically, for all bid documents of the same bidding project, based on the extracted core information (such as company name, qualification certificate number, performance description, etc.) as "nodes," the similarity of information between different bid documents is calculated as "edges." A multi-bid entity association graph is constructed through a graph neural network to intuitively present the implicit associations between multiple documents. Secondly, an abnormal association threshold is set (such as text similarity ≥90%, certificate number similarity ≥85%), and the "edges" in the association graph are scanned. Abnormal associations exceeding the threshold (such as the performance description similarity of A and B being 95%) are mapped to the heat map review report, and a warning mark (such as an orange triangle exclamation mark) is added next to the relevant requirement item (such as "project performance"). For example, the step of setting the abnormal association threshold and adding a warning mark next to the corresponding requirement item in the heat map review report when an abnormal association exceeding the threshold is detected includes: using historical bid documents without abnormalities as positive samples, generating negative samples through data augmentation, and training a model that can identify normal association features and abnormal association features; for the multi-bid entity association graph of the new bidding project, the model calculates the similarity of information between each association edge. Anomaly confidence levels determine anomalous associations; through keyword alignment and semantic role labeling, common erroneous expressions or rare overlapping phrases in the text of anomalous associations are extracted to generate word vector similarity heatmaps; for images of qualification certificates and contract scans involved in anomalous associations, pixel-level difference algorithms are used to match them with forgery features, generating difference mask maps and forgery feature comparison reports; upload logs, modification records, and IP address trajectories of tender documents are retrieved to generate an operation timeline, marking anomalous temporal features such as multiple files being uploaded consecutively under the same IP and highly overlapping modification timestamps; the enterprise credit database is linked to extract implicit association data of equity associations, historical cooperation projects, and legal representatives of anomalous bidders; causal inference algorithms are used to perform correlation analysis on word vector similarity heatmaps, difference mask maps, forgery feature comparison reports, anomalous temporal features, and implicit association data, marking causal and correlational relationships between evidence; a combined interface is constructed in the collaborative review view, where the combined interface uses anomalous association points as core nodes, radiating evidence branches outward according to evidence strength, with solid lines marking causal relationships and dashed lines marking correlational relationships.

[0044] The model uses historically unaffected bid documents as positive samples and generates negative samples through data augmentation. This trains a model capable of identifying both normal and abnormal association features, providing the foundation for anomaly detection. Positive samples are verified compliant documents, while negative samples are generated through data augmentation methods such as randomly modifying performance records and forging certificates. This trains a GNN-based anomaly detection model, learning the differences in features between "normal associations (e.g., 30% text similarity)" and "abnormal associations (e.g., 95% text similarity)." For the multi-bid entity association graph of new bidding projects, the model calculates the anomaly confidence score for each association edge to determine abnormal associations, replacing the traditional fixed threshold. The model outputs each... The anomaly confidence level (0-100%) of the edges is used. For example, if the text association confidence level between A and B is 98%, it is still considered anomaly even if it does not reach the preset 90% threshold. Through keyword alignment and semantic role labeling, common errors or rare overlapping phrases in the texts with abnormal associations are extracted to generate word vector similarity heatmaps. For example, by collecting text evidence, extracting "commonly misspelled owner unit names" and "rare combinations of technical terms," ​​heatmaps are generated to display high-similarity text regions. For images of qualification certificates and contract scans involved in abnormal associations, pixel-level difference algorithms are used to match them with forgery features, generating difference mask maps and forgery feature comparison reports. For example, by collecting image evidence... According to the data, the overlapping seal areas of certificates A and B are marked with pixel-level differential markers, and forgery feature matching is used to identify common PS tampering traces, generating a mask image and report; upload logs, modification records, and IP address trajectories of the tender documents are retrieved to generate an operation timeline, marking abnormal time sequence features such as multiple files being uploaded consecutively under the same IP and highly overlapping modification timestamps, collecting time sequence evidence, such as timeline markings of "the same IP uploaded files A and B within 5 minutes" and "both modification timestamps are 10:05:30", etc.; linked to the enterprise credit database, implicit association data of equity relationships, historical cooperation projects, and legal representatives of abnormal bidders are extracted to collect structured evidence. For example, it extracts implicit data such as the 30% overlap in equity between A and B and the kinship between their legal representatives; it uses causal inference algorithms to perform correlation analysis on word vector similarity heatmaps, difference mask maps, forgery feature comparison reports, abnormal time sequence features, and implicit related data, marking the causal and correlation relationships between evidence, analyzing the logic of evidence, and distinguishing between "same IP operation, text similarity (causation)" and "text similarity and simultaneous upload (correlation)"; it constructs a combined interface in the collaborative review view, in which the combined interface takes abnormal correlation points as core nodes, radiates evidence branches outward according to the strength of evidence, uses solid lines to mark causal relationships and dashed lines to mark correlation relationships, and presents the evidence chain.

[0045] For example, the steps of using a causal inference algorithm to perform correlation analysis on word vector similarity heatmaps, difference mask maps, forgery feature comparison reports, abnormal time-series features, and implicit correlation data, and to label the causal and correlation relationships between evidence, include: standardizing the word vector similarity heatmaps, difference mask maps, forgery feature comparison reports, abnormal time-series features, and implicit correlation data; using an attention mechanism to weightedly fuse the standardized features to generate a unified evidence feature vector; extracting typical causal relationship rules in the bidding field based on historical review cases, and constructing a causal rule base for the bidding field; using strong abnormal features in the unified evidence feature vector as triggering conditions, matching corresponding causal rules from the rule base to generate a candidate hypothesis set consisting of multiple candidate hypotheses; retaining the top 3 high-priority hypotheses for the verification stage based on the strength of the abnormal features corresponding to the candidate hypotheses; simulating intervention operations on the dependent variable in the high-priority hypotheses and observing whether the result variable changes. The process involves determining whether a causal relationship exists; eliminating the interference of confounding variables, such as those common to the industry and those specific to the project, on the association results; constructing a Bayesian network containing dependent, effect, and confounding variables; calculating the confidence level of the causal relationship and the correlation coefficient of the correlation relationship to label the causal and correlation relationships between evidence; and standardizing the word vector similarity heatmap, difference mask map, forgery feature comparison report, abnormal time-series features, and implicit association data, including: extracting BERT semantic vectors of common erroneous statements and rare overlapping phrases from the word vector similarity heatmap, and calculating phrase-level and paragraph-level two-layer similarity matrices; extracting forgery feature vectors from the difference mask map and forgery feature comparison report, quantifying the matching degree of forgery features, and generating an image abnormal feature matrix; converting the operation timeline into a time-series event sequence, and constructing a time-series feature vector using time difference quantification indicators and IP correlation; and converting implicit association data into association strength values ​​to form a structured association matrix.

[0046] The process involves standardizing word vector similarity heatmaps, difference mask maps, forgery feature comparison reports, abnormal temporal features, and implicit association data. An attention mechanism is used to weightedly fuse the standardized features, generating a unified evidence feature vector to achieve evidence unification. Specifically, text standardization involves extracting BERT semantic vectors and calculating phrase / paragraph two-layer similarity matrices; image standardization involves extracting forgery feature vectors and generating an image anomaly feature matrix; temporal standardization involves converting the data into event sequences and constructing temporal feature vectors; and structured standardization involves converting the data into association strength values ​​to form a structured association matrix. These are then weighted and fused into a unified vector using an attention mechanism (e.g., image 0.3, text 0.35, temporal 0.2, structured 0.15). Based on historical review cases, typical causal relationship rules in the bidding field are extracted, and a causal rule library for the bidding field is constructed to establish business logic constraints. Using strong abnormal features in the unified evidence feature vector as triggering conditions, corresponding causal rules are matched from the rule library to generate a candidate hypothesis set consisting of multiple candidate hypotheses. This generates directional hypotheses, such as the triggering rule "text similarity 95%, same IP," which generates "H1: same IP, text similarity 95%, same IP." The hypothesis of "similarity" is adopted. Based on the strength of the abnormal features corresponding to the candidate hypotheses, the top 3 high-priority hypotheses are retained to enter the verification stage. Valid hypotheses are screened, and logically contradictory hypotheses (such as "H2: Text similarity → Same IP" which violates the time sequence) are eliminated according to the strength of the features. The top 3 are retained. The intervention operation on the dependent variable in the high-priority hypotheses is simulated to observe whether the result variable changes in order to determine whether there is a causal relationship. The intervention verification is carried out. For example, if the "IP" of H1 is different, the text similarity drops from 95% to 30%, and the causal relationship is determined to exist. The interference of confounding variables with industry commonality and project characteristics on the correlation results is eliminated. For example, the text similarity of engineering texts is generally high. The confounding variable is eliminated by matching normal projects in the same industry. A Bayesian network containing dependent variable, result variable and confounding variable is constructed. The confidence of causal relationship and the correlation coefficient of correlation relationship are calculated to mark the causal and correlation relationships between evidence and quantify the relationship strength. The Bayesian network calculates the causal confidence of H1 as 92% (strong causality) and the correlation coefficient of "text similarity and simultaneous upload" as 0.8 (strong correlation). The results are associated with the collaborative review view to complete the visualization presentation.

[0047] In summary, the AI-based bid document review method described in the above embodiments of the present invention obtains the bid document and the corresponding tender document, and determines the corresponding tender requirements based on the tender document; it obtains text information about the bid content and image information about the bid qualifications from the bid document; it identifies the text information and image information respectively, and matches and compares the identification results with the tender requirements to review the bid document. The automated identification improves efficiency and avoids the judgment errors inherent in human intervention. This solves the problems of poor efficiency and accuracy of manual bid evaluation in existing technologies.

[0048] Please refer to Figure 2 for Embodiment 3, which shows the AI-based bid document auxiliary review system proposed in the third embodiment of the present invention. The system includes: an acquisition module 100, used to acquire the bid document and the corresponding tender document, and determine the corresponding tender requirements based on the tender document; a determination module 200, used to acquire text information about the bid content and image information about the bid qualifications in the bid document; and a review module 300, used to identify the text information and image information respectively, and match and compare the identification results with the tender requirements to review the bid document.

[0049] The functions or operation steps implemented by the above modules are largely the same as those in the above method embodiments, and will not be repeated here.

[0050] Example 4: Another aspect of the present invention provides a readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the steps of the method described in any one of Examples 1 to 2 above.

[0051] Example 5: Another aspect of the present invention provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the program, it implements the steps of the method described in any one of Examples 1 to 2 above.

[0052] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0053] Those skilled in the art will understand that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequential list of executable instructions for implementing logical functions, and can be embodied in any computer-readable storage medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included 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 storage medium" can mean 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.

[0054] More specific examples (a non-exhaustive list) of computer-readable storage media 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 storage media can even be paper or other suitable media on which the program can be printed, since 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.

[0055] 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.

[0056] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0057] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for assisting in the review of tender documents based on artificial intelligence, characterized in that, The method includes: obtaining the tender documents and the corresponding bidding documents, and determining the corresponding bidding requirements based on the bidding documents; obtaining text information about the bidding content and image information about the bidding qualifications in the tender documents; recognizing the text information and image information respectively, and matching and comparing the recognition results with the bidding requirements to review the tender documents.

2. The artificial intelligence-based bid document auxiliary review method according to claim 1, characterized in that, The steps for recognizing textual and image information respectively include: using a pre-trained language model to perform semantic segmentation on the text representing bidding requirements, extracting corresponding requirement items, and labeling each requirement item with constraint type, data type, and association weight to construct a structured bidding requirement knowledge graph; extracting key information from the bid documents through entity recognition, and then identifying the association between key information through relation extraction, wherein semantic completion is performed on ambiguous or unclear expressions using the domain knowledge graph; converting the recognition results into a structured data format corresponding to the bidding requirement knowledge graph; extracting key areas from the image information through image preprocessing, and recognizing the text content within the key areas using OCR and layout analysis techniques; recognizing the anti-counterfeiting texture of the certificate border in the image information through edge detection, verifying the authenticity of the official seal in the image information through a seal recognition model, and determining whether the image has been photoshopped through a tamper detection algorithm; and associating the text content recognized in the image with the recognition results of the text information in the bid documents.

3. The artificial intelligence-based bid document auxiliary review method according to claim 2, characterized in that, The step of reviewing the bid documents by matching and comparing the identification results with the bidding requirements further includes: comparing the values ​​and formats of rigid constraint items in the identification results, wherein the values, codes, and dates in the bid documents are compared field by field with the threshold range of the requirements in the bidding requirements, generating a binary judgment of pass or fail, and automatically marking the deviation value; encoding the contextual semantics of the bidding requirements and the corresponding bid responses using the Transformer architecture for flexible descriptive items in the identification results, and calculating the matching score by combining domain word vectors; mapping the entities and relationships in the bid information to the association paths in the bidding requirements knowledge graph, generating association matching scores through path integrity verification, and verifying the consistency of the bid information; and marking the risk level of matching items and fuzzy response items close to the threshold in the identification results, and automatically generating supplementary inquiry suggestions.

4. The artificial intelligence-based bid document auxiliary review method according to claim 3, characterized in that, The steps of recognizing text and image information separately, and matching and comparing the recognition results with bidding requirements to review the bid documents, further include: assigning a unique identifier to each requirement item, associating it with its hierarchical path, requirement item type, original matching score, and deviation value in the bidding requirement knowledge graph; recording the original position of the bid document, key text fragments, and attention weight distribution of flexible description items corresponding to each requirement item; converting the multi-level structure of the bidding requirement knowledge graph into a horizontal axis coordinate, where the first-level category occupies the basic coordinate position, and the next-level category is displayed indented under the previous-level category; defining the vertical axis according to the matching score, using the matching score as a quantitative indicator, and mapping each requirement item to the corresponding vertical and horizontal axis positions according to its matching score and corresponding level; setting different color mappings according to different matching scores and rendering special states for special requirement items to obtain a heatmap review report; when the mouse hovers over the coordinate point of a requirement item's heatmap, a pop-up window is triggered to display the text comparison between the original requirement text and the original bid document, the automatically calculated deviation value, and the evidence location.

5. The artificial intelligence-based bid document auxiliary review method according to claim 4, characterized in that, The step of triggering a pop-up window to display a text comparison between the original requirement and the original bid when the mouse hovers over the coordinates of a requirement item's heatmap, automatically calculating the deviation value, and locating evidence, further includes: constructing a multi-bid entity association graph for all bid documents of the same bidding project using a graph neural network, where the nodes of the multi-bid entity association graph are the core information in the bid documents, and the edges of the multi-bid entity association graph are associated based on the similarity of the information; setting an abnormal association threshold, and when an abnormal association exceeding the threshold is detected, adding a warning mark next to the corresponding requirement item in the heatmap review report; clicking the warning mark triggers a collaborative review view, synchronously displaying a comparison of the corresponding information in multiple associated bid documents.

6. The artificial intelligence-based bid document auxiliary review method according to claim 5, characterized in that, The step of setting an abnormal association threshold and adding a warning mark next to the corresponding requirement item in the heatmap review report when an abnormal association exceeding the threshold is detected includes: using historical bid documents without abnormalities as positive samples, generating negative samples through data augmentation, and training a model that can identify normal association features and abnormal association features; for the multi-bid entity association graph of new bidding projects, determining abnormal associations by calculating the abnormal confidence of each association edge through the model; extracting common erroneous expressions or rare overlapping phrases in the text of abnormal associations through keyword alignment and semantic role labeling, and generating a word vector similarity heatmap; for images of qualification certificates and contract scans involved in abnormal associations, matching them with forgery features using a pixel-level difference algorithm, generating a difference mask image and comparing it with the forgery features. The report retrieves upload logs, modification records, and IP address traces of the tender documents to generate an operation timeline, marking abnormal time-series features such as multiple files uploaded consecutively from the same IP and highly overlapping modification timestamps; it also links to the enterprise credit database to extract implicit association data on equity relationships, historical cooperation projects, and legal representatives of abnormal bidders; it employs causal inference algorithms to perform correlation analysis on word vector similarity heatmaps, difference mask maps, forgery feature comparison reports, abnormal time-series features, and implicit association data, marking the causal and correlation relationships between evidence; and it constructs a combined interface in the collaborative review view, where the combined interface uses abnormal association points as core nodes, radiating evidence branches outward according to evidence strength, with solid lines marking causal relationships and dashed lines marking correlation relationships.

7. The artificial intelligence-based bid document auxiliary review method according to claim 6, characterized in that, The steps for using a causal inference algorithm to perform correlation analysis on word vector similarity heatmaps, difference mask maps, forgery feature comparison reports, abnormal time-series features, and implicit correlation data, and to label the causal and correlation relationships between evidence, include: standardizing the word vector similarity heatmaps, difference mask maps, forgery feature comparison reports, abnormal time-series features, and implicit correlation data; using an attention mechanism to weightedly fuse the standardized features to generate a unified evidence feature vector; extracting typical causal relationship rules in the bidding field based on historical review cases to construct a causal rule base for the bidding field; using strong abnormal features in the unified evidence feature vector as triggering conditions, matching corresponding causal rules from the rule base to generate a candidate hypothesis set consisting of multiple candidate hypotheses; retaining the top 3 high-priority hypotheses for the verification stage based on the strength of the abnormal features corresponding to the candidate hypotheses; simulating intervention operations on the dependent variable in the high-priority hypotheses and observing whether the result variable changes to determine... Whether a causal relationship exists; excluding the interference of confounding variables of industry commonality and project characteristics on the association results; constructing a Bayesian network containing dependent variable, effect variable, and confounding variables, and calculating the confidence level of causality and the correlation coefficient of correlation to label the causal and correlation relationships between evidence; wherein, the steps of standardizing word vector similarity heatmap, difference mask map and forgery feature comparison report, abnormal time series features, and implicit association data include: extracting BERT semantic vectors of common erroneous statements and rare overlapping phrases from word vector similarity heatmap, and calculating phrase-level and paragraph-level two-layer similarity matrices; extracting forgery feature vectors from difference mask map and forgery feature comparison report, quantifying the matching degree of forgery features, and generating an image abnormal feature matrix; converting the operation timeline into a time series event sequence, and constructing a time series feature vector using time difference quantification indicators and IP correlation; converting implicit association data into association strength values ​​to form a structured association matrix.

8. A tender document review system based on artificial intelligence, characterized in that, The system includes: an acquisition module for acquiring tender documents and corresponding bidding documents, and determining the corresponding bidding requirements based on the bidding documents; a determination module for acquiring text information about the bidding content and image information about the bidding qualifications in the tender documents; and a review module for recognizing the text information and image information, and comparing the recognition results with the bidding requirements to review the tender documents.

9. A readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the program, implements the steps of the method as described in any one of claims 1 to 7.

Images