An artificial intelligence-based large-scale bidding document intelligent review method

Abstract

The application discloses a kind of based on artificial intelligence large-scale tender file intelligent examination method, it is related to big data processing technical field, including, tender project file and log are written into big data storage;Tender file is carried out layout analysis and structured extraction, generates field record and binds the evidence anchor point of page number and coordinate;Tender file clause is compiled into executable review rule, and output does not meet item and its evidence anchor point;Tender file is cut into evidence fragment according to page number area, and the merkle tree root value is calculated;When generating review report, only disclose the minimum evidence fragment corresponding to non-compliance item and its verification path;While calculating the residual error degree of cooperative degree and behavior homology of tender offer, generating risk classification of surrounding string mark and writing into report.The application realizes the source verifiable review under the condition of minimum evidence disclosure, and gives consideration to commercial secret protection and regulatory requirements, and the joint description of offer deviation and process behavior improves the interpretability and adoptability of surrounding string mark identification.

Description

Technical Field

[0001] This invention relates to the field of big data processing technology, and in particular to a method for intelligent review of large-scale tender documents based on artificial intelligence. Background Technology

[0002] In bidding activities, the review process typically relies on the terms of the bidding documents to compare and verify the eligibility requirements, substantive responses, technical specifications, commercial terms, and price consistency of the bids. Current engineering practices often employ big data storage to hold project documents and process data, combined with layout parsing, field extraction, and rule base validation to achieve batch pre-review. The verification results are then provided in report form to evaluation experts or regulatory authorities to support evaluation decisions and ensure compliance documentation.

[0003] In large-scale scenarios, the above-mentioned conventional practices still have room for improvement in terms of evidence review and risk assessment: First, the evidence cited in the report often relies on the presentation or sharing of the entire tender document, making it difficult to balance the need for commercial confidentiality and verification of sources; Second, the assessment of collusion often focuses on the static similarity of text or price, and rarely integrates the behavioral homogeneous information of expected price residuals and process event logs, thus limiting the explanatory power of collaborative manipulation. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a large-scale intelligent review method for tender documents based on artificial intelligence to solve the problems of existing technologies, such as the reliance on full document disclosure for evidence verification making it difficult to balance confidentiality and verifiability, and the insufficient explanatory power of collaborative manipulation due to the limited number of feature dimensions in the analysis of bid rigging.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a large-scale intelligent review method for tender documents based on artificial intelligence, which includes: acquiring the tender documents, each bid document, and the event log of the bidding process for the same bidding project, writing them into big data storage, and retaining the original immutable records; The layout of each tender document is parsed and the structure is extracted to obtain field records and bind page numbers and coordinates as evidence anchors to each field record; The terms of the tender documents are compiled into executable review rules, and each tender document is checked against a uniform standard, outputting non-conformities and evidence anchors for non-conformities; The tender documents are divided into evidence segments according to page number areas, the Merkel root value of the evidence segments is calculated and written into an immutable evidence storage medium; When generating the review report, only the smallest piece of evidence corresponding to the non-compliance and the corresponding verification path are disclosed; Based on historical similar bidding projects and bidding process event logs, calculate the coherence of bid price residuals and the homogeneity of behavior; Based on the synergy of bid price residuals and the homogeneity of behavior, a risk classification for bid rigging is generated and written into the report.

[0007] As a preferred embodiment of the large-scale intelligent review method for tender documents based on artificial intelligence described in this invention, the page layout parsing refers to obtaining the original tender documents of the bidding entities from the bidding group, generating read-only processing task records, and establishing a page index table; The tender document is rendered as a page image, and a text-first extraction strategy is executed to obtain a sequence of character objects. The sequence of character objects from the page image is then input into the layout detection model to obtain a set of layout blocks.

[0008] As a preferred embodiment of the large-scale intelligent review method for tender documents based on artificial intelligence described in this invention, the structured extraction refers to constructing a directed graph with page blocks as nodes, performing topological sorting on the directed graph to obtain the reading order, identifying chapter title blocks in the page blocks according to the reading order, constructing a chapter node tree, and attaching each text block and table block to the corresponding chapter node. Perform table structure restoration on the table block and retrieve the field records corresponding to the table block; After segmenting the text block by sentence, field extraction is performed to obtain the field records corresponding to the text block.

[0009] As a preferred embodiment of the large-scale intelligent review method for tender documents based on artificial intelligence described in this invention, the evidence anchor point for binding page numbers and coordinates to each field record includes: for each field record, the page number is taken as the page number of the character object, the coordinate is taken as the minimum bounding rectangle of the character object's bounding box, and the original text fragment is taken as the normalized text within the bounding box's coverage area. The page number and coordinates are bound to the field to obtain the anchor point for each field record. Write the anchor point, field name, and field value of each field record into the structured result table to obtain the evidence anchor point for each field.

[0010] As a preferred embodiment of the large-scale intelligent review method for tender documents based on artificial intelligence described in this invention, wherein: the output non-conformities and the evidence anchors for the non-conformities include, Obtain the original bidding documents from the bidding project, analyze the layout, and obtain the set of bidding side layouts; Generate tender-side evidence anchors for the character records of each text block and table block; The tender document text is segmented according to a chapter tree to obtain the clause sequence; Perform field mapping and constraint element extraction on the text of each clause to obtain the target fields and constraint relationships; The constraints on the target fields corresponding to the clauses are broken down into sub-rules according to their order of appearance in the clauses; Each sub-rule is compiled into an executable expression, and conflicts are resolved according to priority to obtain the executable review rules for the tender documents. For each tender document's field set and field evidence anchors, each rule is evaluated according to the executable review rules. When a rule is deemed unmet, non-compliance items and their evidence anchors are generated.

[0011] As a preferred embodiment of the large-scale intelligent review method for tender documents based on artificial intelligence described in this invention, the specific steps of segmenting the tender document into evidence fragments according to page number regions, calculating the Merkle root value of the evidence fragments, and writing them into an immutable evidence storage medium are as follows: Determine the page coordinate system based on the original tender documents and page index table of the bidding entity; Create a candidate region grid for each page and obtain the candidate region set; The candidate region set is segmented, and evidence fragments are generated for each candidate region; The evidence fragments are standardized to generate a standardized string of evidence fragments; The standardized string of evidence fragments is concatenated with page numbers and coordinate frames to form a leaf input; Based on the leaf input, the leaf digest value is calculated using a collision-resistant hash function; For all leaf summary values ​​in the tender document, a binary Merck tree is used to concatenate the summaries pairwise from bottom to top to generate the parent node summary, until the root value is obtained; Write the Merkel root value of the tender document into an immutable storage medium.

[0012] As a preferred embodiment of the large-scale intelligent review method for tender documents based on artificial intelligence described in this invention, the step of disclosing only the smallest piece of evidence corresponding to the non-conformity and the corresponding verification path includes, Cluster the anchor points of the bid-side evidence for each non-compliance item to obtain candidate anchor point groups, and generate a minimum disclosure area for each candidate anchor point group according to the minimum disclosure method; For each minimum disclosure area, extract the content covered by the area from the structured results of the tender documents, perform normalization processing, and obtain the disclosure text string; Based on the page number and coordinate frame of the disclosure area, the evidence fragment with the highest spatial overlap is retrieved from the fragment index. For each selected evidence fragment leaf summary, the sibling node summary sequence from the leaf to the root is extracted from the Merkle tree structure to form a verification path. The verification path, root value, and evidence receipt number are encapsulated into a verification material package, and the verification material package is bound to the disclosed evidence fragments; When generating the review report, for each non-compliance item, output the rule number, the tendering side evidence anchor point, the minimum disclosure evidence fragment index, the disclosure text string, and the verification material package index.

[0013] As a preferred embodiment of the large-scale intelligent review method for tender documents based on artificial intelligence described in this invention, the bid price residual coordination degree includes, Extract historical data of completed bid evaluations from historical bidding projects to generate a historical bid sample set; Extract collectable feature vectors and bid prices from each historical bid sample set to form training samples; A linear expected price model is established based on training samples. The expected price of any bidding entity is calculated based on the linear expected price model to obtain the expected price. For each bidder in the current bidding project, calculate the bid price residual; Based on the benchmark of bid price residual scale, the bid price residual is standardized, and for any two bidders in the current bidding project, the bid price residual synergy is calculated according to the standardized bid price residual.

[0014] As a preferred embodiment of the large-scale intelligent review method for tender documents based on artificial intelligence described in this invention, the behavioral homology includes: Obtain the event sequence for each bidding entity from the event log of the bidding process; For each bidding entity, the number of transitions of the same event type in the statistical event sequence is calculated, and the transition probability distribution is obtained. Based on the transition probability distribution, calculate the overlap ratio of the transition distributions of any two bidding entities; Calculate the time interval sequence for adjacent events, and calculate the rhythm overlap ratio for any two bidding entities; The overlap ratio of transition distribution and the overlap ratio of rhythm are fused to obtain the behavioral homology.

[0015] As a preferred embodiment of the large-scale intelligent review method for tender documents based on artificial intelligence described in this invention, the step of generating a risk classification of bid rigging and writing it into a report includes performing risk mapping on the synergy of the bid residuals and the homogeneity of behavior of any two bidding entities to obtain risk scores for the two bidding entities. When the risk scores of the two bidding entities are not less than the edge threshold, an edge is established in the relationship graph of the two bidding entities. Extract the connected components of the current bidding entity from the relationship graph as candidate groups, and calculate the group risk value for any candidate group based on the average of the pairwise risk scores within the group. Based on the group's risk value, a risk classification for bid-rigging is generated and written into the review report.

[0016] The beneficial effects of this invention are as follows: By disclosing only the smallest piece of evidence corresponding to non-conformities and simultaneously outputting the verification path when generating the review report, the authenticity verification of evidence can be completed without disclosing the entire tender document. This is beneficial for balancing the protection of trade secrets in tender documents with the verifiable requirements of regulatory review, and reduces the risk of unnecessary information dissemination. By calculating the bid price residual based on historical data and combining it with the event log of the bidding process to generate the common source characteristics of bid rigging behavior, the joint characterization and interpretable hierarchical early warning of bid price deviation process behavior can be achieved. This is beneficial for improving the effectiveness of discovering bid rigging clues and enhancing the interpretability and acceptability of risk assessment. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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.

[0018] Figure 1 This is a flowchart of a large-scale intelligent review method for tender documents based on artificial intelligence.

[0019] Figure 2 This is a flowchart for page layout analysis.

[0020] Figure 3 A flowchart for compiling clauses into executable review rules.

[0021] Figure 4 This is a flowchart for verifying each non-compliant item. Detailed Implementation

[0022] 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.

[0023] 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.

[0024] 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.

[0025] Reference Figures 1-4This is one embodiment of the present invention, which provides a large-scale intelligent review method for tender documents based on artificial intelligence, including the following steps: S1. Obtain the tender documents, bid documents, and bidding process event logs for the same bidding project, write them to big data storage, and retain the original immutable records.

[0026] Furthermore, when a bidding project is created, the access terminal generates a project number for the bidding project, and generates a bidder number for each bidder under the project number.

[0027] All tender documents, bid documents, and event logs of the tendering process should be archived under the project number.

[0028] It should be noted that for each file upload request, the project number, bidder number, file type identifier, upload timestamp, and uploader's identity identifier must be included.

[0029] If any required field is missing, the write operation will be rejected and an error code will be returned to avoid creating an untraceable orphan file.

[0030] Furthermore, the access end uses an encrypted transmission channel to receive file streams and performs fixed-size chunked uploads for large files.

[0031] Upon arrival of each fragment, the fragment length is first verified, the format whitelist is verified, and malicious content is detected.

[0032] After all fragments are received, they are reassembled to obtain the original byte sequence.

[0033] The original byte sequence is processed using SHA-256, a type of collision-resistant hash, to calculate the content digest value, which is then written into the metadata record of the current file.

[0034] Furthermore, the original byte sequence of the file is written to the object storage using the content digest value as the object key, while an immutable retention strategy is enabled for the object storage.

[0035] The immutable retention policy specifically includes enabling object version control and object locking; prohibiting overwrites and in-place modifications, allowing only the addition of new versions; and immediately writing retention policy parameters, writer identity, write node identifier, and write completion timestamp to form auditable metadata after the object is written.

[0036] This ensures that the original record is immutable, and any subsequent read will locate the same original content using the object key, while the storage side refuses to delete or modify it.

[0037] Furthermore, a manifest record is written for each file in the distributed metadata table.

[0038] The fields recorded in the list include project number, bidder number, object key, file name, file size, submission batch number, uploader identity, access end timestamp, storage end confirmation timestamp, and retention policy parameters.

[0039] At the same time, the inventory records are appended to the audit log topic, such as a message queue or log append file, and a chained summary is formed for the log records.

[0040] Furthermore, for process events generated by the bidding platform, a capture method based on event subscription is used to obtain real-time access to the process events.

[0041] Process events include, but are not limited to, login, download, upload, withdrawal, resubmission, and Q&A download.

[0042] Write each process event into a unified event structure.

[0043] The event structure includes event type, event timestamp, subject number, project number, terminal identifier, network identifier, request digest, and result code.

[0044] Process events are written to the append write log and placed in the archive area under the same immutable storage policy.

[0045] Furthermore, once the bidding entity has completed the submission of all documents under the current bidding project, a closed-loop marker for the submission batch is generated, and the list of object keys and the event log range index contained in the submission batch are written into the batch index table.

[0046] When the reviewer reads the document, the original document object, the corresponding metadata list, and the current batch event sequence are replayed according to the project number and batch number.

[0047] S2. Perform layout parsing and structured extraction on each tender document, obtain field records, and bind page numbers and coordinates as evidence anchors for each field record.

[0048] Furthermore, the original bid documents of the bidding entities are obtained from the bidding group, a read-only processing task record is generated, and a page index table is established.

[0049] The page index table includes a record of the page number and page size for each page.

[0050] Render PDF or Word-type tender documents page by page into page images and implement a text-first extraction strategy.

[0051] The text-first extraction strategy involves extracting the text content of each character and its bounding box in the page coordinate system if the tender document contains directly parsable text. If the page is a scanned image or the text layer is missing, character recognition is performed on the page image, outputting character sequences, character confidence scores, and character bounding boxes. The output is then uniformly normalized into a sequence of character objects, where each character object includes character content, page number, bounding box, and character confidence score.

[0052] Input each page image into the layout detection model to obtain a set of layout blocks.

[0053] Each page block outputs the block type and block bounding box, along with the block confidence score.

[0054] The block types include, but are not limited to, title blocks, body blocks, table blocks, header and footer blocks, and seal / signature blocks.

[0055] After the block detection output is completed, the block bounding boxes are deduplicated and merged.

[0056] Specifically, in the deduplication and merging process, if the overlap between two blocks exceeds the merging threshold, the block with the higher confidence level is retained and the two covered areas are merged.

[0057] It should be noted that the acquisition of the layout detection model specifically involves extracting representative bidding documents and tender documents pages from historical bidding documents. After rendering each page of the bidding documents as an image, the layout areas such as title blocks, body text blocks, table blocks, header and footer blocks, and seal and signature blocks in the documents are selected one by one using an annotation tool and assigned a category, generating a data sample set with page images, region coordinates, and category labels. Using the data sample set as the training set, the layout detection model is trained using an instance segmentation framework under supervised learning, and iteratively adjusted on the validation set. Finally, a layout detection model that can output layout region categories, region coordinate boxes, and confidence scores for input page images is solidified. The layout detection model consists of a backbone network that extracts multi-scale features from page images, a feature pyramid that integrates multi-scale features, and a detection head that outputs layout region categories, bounding boxes, and confidence scores.

[0058] The merging threshold is determined by performing a gridded evaluation on the labeled validation set for different overlap thresholds. The threshold that optimizes the sum of the detection metrics of the blocks, such as the compromise between localization accuracy and recall, and the F1 score, while minimizing the false merging rate, is selected as the merging threshold. The value range is generally [0.3-0.7].

[0059] Furthermore, a directed graph is constructed with page blocks as nodes. Specifically, edges represent reading order constraints from top to bottom and from left to right. A topological sort is performed on the directed graph to obtain the reading order. Chapter title blocks are identified based on the reading order, and a chapter tree is constructed. Each text block and table block is attached to its respective chapter node.

[0060] The table structure is restored for the table block. Specifically, the table grid lines and cell boundaries are identified by aggregating pixel edges and line segments. Character objects are placed into the corresponding cell areas according to the bounding box, and the row and column structure and cell text are output. Field name normalization mapping is performed on the table header row, such as mapping the project period / planned project period / contract project period to the same field name, and generating field records for the corresponding value cells of each field name in the same row.

[0061] After segmenting the main text into sentence blocks, candidate fields are extracted using a field dictionary with fixed values.

[0062] Among them, field dictionary setting refers to locating field prompt words through the field dictionary, such as security deposit, warranty period, project manager, etc.; and then extracting field values, such as date, amount, certificate number, percentage, enumeration items, etc., through the setting mode; if the same field appears multiple times in the same chapter, one is selected as the main record based on the credibility of the evidence anchor, and the rest are kept as alternative records.

[0063] Furthermore, evidence anchors are generated for each field record. Specifically, the page number is taken from the page number of the character object containing the field value; the coordinates are taken from the smallest bounding rectangle of the character object corresponding to the field value; the original text fragment is taken from the normalized text within the area covered by the bounding box; and the anchors, field names, and field values ​​are written together into the structured result table to obtain the evidence anchors for the field.

[0064] Specifically, the calculation of the confidence level of the field's evidence anchor point involves statistically analyzing the confidence levels of all characters within the coverage area of ​​the field's evidence anchor point into a histogram and normalizing it. A density function defined on the interval [0,1] is then obtained. Based on this density function, the confidence level of the field's evidence anchor point is calculated, expressed as: ; ; ; in, Indicates the confidence level of the character. This represents the confidence level independent variable. Represents the density function, Indicates the overlap ratio. This represents the area of ​​intersection between the evidence anchor region and the field prior region. This represents the area of ​​the union of the evidence anchor region and the field prior region. Indicates the credibility of the evidence anchor point of the field. This represents the block confidence weighting coefficient. This represents the block confidence score output by the block detection function. This represents the character confidence weight coefficient. This represents the overlap ratio weighting coefficient. This represents a bias constant.

[0065] It should be noted that the intersection area By representing the evidence anchor region and the field prior region as two polygons in a unified coordinate system on the same page, geometric clipping is performed on the two polygons to obtain the intersection polygon, and the intersection area is calculated using the vertex coordinates of the intersection polygon according to the polygon area formula; the union area is obtained. The geometric areas of the evidence anchor point region and the field prior region are calculated separately, and then the union area is obtained by the identity of union area = evidence anchor point region area + field prior region area - intersection area. The field prior region refers to the common occurrence position of a field in the tender document. It is a candidate positioning area defined in the page coordinate system, which is generated by the statistics of historical tender document samples and is used to constrain the search range of the field.

[0066] It should be noted that the block confidence weight coefficient, character confidence weight coefficient, and overlap ratio weight coefficient were obtained by sampling regularized logistic regression on historical bidding documents, using the correctness of evidence anchor points as labels, and training the block confidence weight coefficient, character confidence weight coefficient, and overlap ratio weight coefficient with maximum likelihood and determining the final coefficients using cross-validation.

[0067] S3. Compile the terms of the tender documents into executable review rules, and verify each tender document item by item according to a uniform standard, outputting non-conformities and evidence anchors for non-conformities.

[0068] Furthermore, the original bidding documents are obtained from the bidding project, and the same layout parsing as the bid documents is performed to obtain the chapter tree, text blocks, table blocks, page numbers, and coordinates.

[0069] Generate bidding-side evidence anchor points for each text block and table block, and bind and save the bidding-side evidence anchor points and corresponding text fragments.

[0070] The tender document text is segmented by chapter tree to obtain the clause sequence.

[0071] Perform a clause type determination on each clause in the clause sequence.

[0072] The types of terms include, but are not limited to, eligibility requirements, formal compliance, substantive response terms, technical specification boundaries, commercial terms, and quotation consistency requirements.

[0073] Generate a globally unique rule number for each clause and record the evidence anchor points of the clause in the tender documents to form a clause record.

[0074] For each clause, field mapping and constraint element extraction are performed using a field dictionary combined with fixed syntax to obtain the target field name and constraint relationship.

[0075] The target field names include the project duration, security deposit amount, qualification level, and whether the deviation table has been submitted.

[0076] Constraints include existence, numerical range, enumeration set, date validity period, and document location requirements.

[0077] When a clause corresponds to multiple target fields, it is split into multiple sub-rules according to the order of appearance in the clause, and each sub-rule only constrains one target field.

[0078] Furthermore, each sub-rule is compiled into an executable expression. The expression is stored in the form of an abstract syntax tree, and the node types are fixed as existence judgment, comparison judgment, set inclusion judgment, regular expression matching judgment, and semantic consistency judgment. The compilation output includes rule number, rule type, target field name, execution expression, output template, and bidding side evidence anchor point.

[0079] When multiple constraints exist for the same target field, conflict resolution is performed in order of priority.

[0080] Among these, conflict resolution and solidification are carried out in order of priority. Specifically, substantive response clauses take precedence over eligibility conditions, eligibility conditions take precedence over technical specifications, technical specifications take precedence over commercial terms, and commercial terms take precedence over format requirements.

[0081] Only the rule with the highest priority and the strictest constraint is retained as the main rule, and the remaining rules are retained as supplementary validation rules and marked as secondary prompts in the review report.

[0082] It should be noted that a substantive response clause refers to a key clause in the tender document that has a decisive impact on the validity of the bid, the evaluation, and the winning bid, and that the bidder must unconditionally meet and cannot reserve; otherwise, the bidder's bid will be deemed invalid.

[0083] Furthermore, for each tender document, obtain the field set and the evidence anchor points for each field, and execute the following expressions A1-A2 one by one in order of rule number: A1. For existence, interval, set, and regular expression rules, the field value is used directly for judgment.

[0084] A2. For semantic consistency rules, first locate the response fragments in the tender document that are related to the same chapter or field of the current clause text, obtain the evidence anchor points of the response fragments, and then perform semantic consistency judgment.

[0085] Specifically, semantic consistency judgment involves extracting from the clause text. For each keyword, obtain a keyword set, calculate the character set overlap score in the bid response fragment, and obtain... Each score sample is used to construct a probability density and calculate its overlap with the ideal density, thus obtaining the semantic deviation, expressed as: ; ; ; in, The probability density function represents the bid response score, and its value ranges from [0,1]. Indicates the score independent variable. Indicates the number of keywords. Indicates the smoothness scale. Indicates the first The character set overlap score of each keyword in the bid response fragment. Represents the ideal bid response density function. Indicates the degree of semantic deviation.

[0086] Set the semantic deviation If the semantic deviation is not greater than the semantic deviation threshold, the current response fragment is considered semantically consistent; if the semantic deviation is less than the threshold, the response fragment is considered semantically consistent. If the deviation exceeds the semantic deviation threshold, the current response fragment is determined to be semantically inconsistent, and non-compliant items and bidding side evidence anchor points are output.

[0087] It should be noted that the semantic deviation threshold is determined by scanning the semantic deviation on historical labeled samples of similar bidding projects to ensure that the false positive rate and false negative rate of semantics meet the automatic review allowable range specified by the bidding party or regulator, and the optimal value of the comprehensive index is used as the semantic deviation threshold.

[0088] When a rule is determined to be unmet, a non-compliant item is generated.

[0089] Non-compliance items include anchor points of evidence from the bidding side, anchor points of evidence from the tendering side, and the field values ​​or response fragment summaries that are being judged.

[0090] If a field in the tender document is missing, making it impossible to locate the anchor point of the tender evidence, then the anchor point of the tender evidence will be set to unchecked, along with a description of the search scope. If the search fails to find the target within the specified section, the reviewer will be able to reproduce the search path.

[0091] All non-conformities are sorted in a fixed order. In this embodiment, a fixed sorting order is provided, namely, qualification category, substantive response category, technical specification category, commercial category, and format category.

[0092] Generate a review report field for each non-compliance item.

[0093] The review report fields include rule number, rule type, tendering-side anchor point, bidder-side anchor point, summary of judgment basis, and summary of field value or response fragment.

[0094] The review report fields are written to the big data store as structured results.

[0095] S4. Divide the tender documents into evidence segments according to page number areas, calculate the Merkel root value of the evidence segments, and write it into an unalterable evidence storage medium.

[0096] Furthermore, obtain the original tender documents and page index table of the bidding entity to determine the page coordinate system.

[0097] Specifically, the page coordinate system is determined by taking the bottom left corner of the page as the origin, the horizontal axis as the x-axis, and the vertical axis as the y-axis, with the coordinate unit being pixels or print points; the width, height, resolution, and rendering scaling parameters of each page are written into the coordinate system parameter table.

[0098] For each page, a candidate region grid is created, and the generation of the candidate region grid follows the B1-B4 rules as follows: B1. Obtain the set of page blocks from the page layout analysis results, and use the bounding box of each page block as a first-level candidate region.

[0099] B2. For the table block, further split the table into secondary candidate areas according to cell boundaries.

[0100] B3. For the main text block, split it into secondary candidate regions according to paragraph boundaries.

[0101] B4. For adjacent candidate regions on the same page, if the distance between them is less than the distance threshold and they belong to the same chapter, they are merged into the same candidate region.

[0102] Therefore, the set of candidate regions for each page is obtained.

[0103] Each candidate region includes a page number, a coordinate frame, and a region type.

[0104] It should be noted that the spacing threshold is obtained by statistically analyzing the spacing distribution of adjacent blocks in historical bid document samples and then fixing it with quantiles such as the 95th percentile value as the threshold.

[0105] Furthermore, the candidate region set is segmented, and evidence fragments are generated for each candidate region.

[0106] The evidence fragment consists of page numbers, coordinate frames (i.e., the coordinates of the lower left and upper right corners of the area), and standardized text or structured content summaries within the area.

[0107] It should be noted that the table area uses the serialized result of field name-field value.

[0108] At the same time, a fragment number is generated for each piece of evidence. The fragment number is composed of the project number, the bidding entity number, the page number, and the area number.

[0109] It should be noted that, in order to avoid changes in the root value due to format differences, the evidence fragments are standardized.

[0110] Content standardization includes, but is not limited to, unifying full-width and half-width characters and capitalization, unifying number formats, compressing consecutive whitespace into single spaces, ensuring stable output of table content in "row-major order, column-major order", and deleting non-essential content such as headers and footers.

[0111] The output will be normalized to generate a normalized string of evidence fragments.

[0112] Furthermore, a leaf input string is constructed for each evidence fragment. Specifically, the normalized string of the evidence fragment is concatenated with the page number and coordinate frame to form an unambiguous leaf input; and the leaf summary value is calculated through an anti-collision hash function.

[0113] Each leaf summary corresponds one-to-one with a segment number, and the segment number, page number, coordinates, and leaf summary are written into the segment index table.

[0114] After sorting all leaf summary values ​​under the same tender document in the following order C1-C2, construct a Merkle tree: C1. The sorting keys are as follows: page number ascending, area bottom left corner x coordinate ascending, area bottom left corner y coordinate ascending, area area ascending; C2. Then, a binary Merkle tree is used to concatenate the digests pairwise from bottom to top to generate the parent node digest, until the root value is obtained; if the number of leaves is odd, the last leaf is self-concatenated to generate the parent node.

[0115] Write the Merkel root value of the tender document into an immutable storage medium.

[0116] The evidence storage medium can be blockchain, trusted timestamp service, or audit ledger with the characteristic that it cannot be modified after being written.

[0117] The content to be written should include at least the project number, bidder number, file object key, Merkle root value, number of leaves, summary of coordinate system parameter table, write timestamp, and write confirmation receipt number.

[0118] After successful evidence storage, the receipt number will be written back to the data lake metadata table.

[0119] Furthermore, when the same tender document is further segmented to generate leaf summaries and Merkel tree root values ​​are constructed, the new root values ​​and the existing root values ​​are compared.

[0120] If the new root value and the existing root value are inconsistent, an alarm will be triggered and the subsequent automatic review process of the current tender document will be frozen, and manual verification will be initiated.

[0121] S5. When generating the review report, only disclose the smallest piece of evidence corresponding to the non-compliance and the corresponding verification path.

[0122] Furthermore, after completing the rule-by-rule verification, the set of non-conformities for each tender document is obtained.

[0123] For each non-compliance item, obtain the bid-side evidence anchor point. Aggregate non-compliance items that are on the same page, spatially adjacent, and belong to the same rule number to obtain a candidate evidence anchor point group.

[0124] For each candidate evidence anchor group, a minimum disclosure area is generated using the minimum disclosure method.

[0125] Specifically, the minimum disclosure method uses the smallest bounding rectangle of all coordinate frames within the candidate evidence anchor point group as the initial disclosure frame. The initial disclosure frame is then expanded with fixed margins to cover the relative relationships between field names and field values. If the expanded disclosure frame overlaps with headers, footers, or other chapter content, the disclosure frame is shrunk according to the page block type constraint, retaining only the text block or table block area directly related to the non-compliance item. If overlap still exists, the disclosure frame is split into multiple non-overlapping sub-disclosure frames, and the final minimum disclosure area set is obtained with the minimum number of sub-disclosure frames as the optimization objective.

[0126] For each minimum disclosure area, extract the text or table serialization content covered by the area from the structured results of the tender documents, perform normalization processing, and obtain the disclosure text string.

[0127] It also saves the page number, coordinate frame, and source type of the disclosed segment, such as paragraph or table.

[0128] Based on the page number and coordinate frame of the disclosed area, retrieve the evidence fragment with the highest spatial overlap from the fragment index.

[0129] If the disclosure area consists of multiple sub-disclosure boxes, then for each sub-disclosure box, the corresponding evidence fragments are retrieved and multiple leaf summaries are obtained.

[0130] When there are ties in the search, the unique evidence fragment is selected according to the lexicographical order of the fragment number.

[0131] For each selected evidence fragment leaf summary, extract the sibling node summary sequence from the leaf to the root from the Merkle tree structure to form the verification path.

[0132] The verification path, root value, and evidence receipt number are packaged into a verification material package and bound to the disclosure fragment.

[0133] The verification material package includes the disclosure text string, page number, coordinate frame, fragment number, leaf summary, verification path, Merkel root value, evidence receipt number, and coordinate system parameter summary.

[0134] When generating the review report, for each non-compliance item, output the rule number, the tendering side evidence anchor point, the minimum disclosure fragment index of the bidding side, the disclosure text string, and the verification material package index.

[0135] Furthermore, tiered disclosure is set up for different roles. For example, evaluation experts can view the disclosure text string and page coordinates; supervision or review agencies can additionally view the verification path and root value receipt, while ordinary participants can only view the summary of non-compliance items, in order to avoid unnecessary information dissemination.

[0136] Furthermore, after receiving the disclosure text string and verification material package, the review end generates a disclosure fragment commitment string according to the same normalization rules and calculates the leaf summary.

[0137] The reconstructed root value (new root value) is obtained by calculating from bottom to top using the verification path, and then the reconstructed root value is compared with the root value (proof root value) in the proof receipt.

[0138] If they match, the disclosed fragment is determined to be an evidence fragment from the original tender document, and the verifiable review is completed; if they do not match, the evidence chain is broken, and the process of reading the original document and manual verification is triggered.

[0139] S6. Based on historical similar bidding projects and bidding process event logs, calculate the consistency degree of bid price residuals and the homogeneity degree of behavior.

[0140] Furthermore, historical data of completed bid evaluations are obtained from historical bidding projects, and a sample set of similar historical bids is generated according to the screening criteria. The screening criteria include bidding category, region, bid section type, project scale range, and release time window.

[0141] For each historical bidding sample, extract the collectable feature vector and the corresponding bid price to form a training sample record set.

[0142] A linear expected price model is established based on the training sample record set. The structure of the linear expected price model is fixed as the intercept term and the feature weighted sum, and the obtained coefficients are fixed in the parameter table.

[0143] The expected bid price for any bidding entity is calculated and obtained as follows: ; in, Indicates the expected price. This represents the intercept coefficient of the expected price model. The expected price model is represented by the first... dimensional eigenvalues, Indicates the first Dimensionally collectable feature values, Indicates the feature dimension.

[0144] It should be noted that the intercept coefficient of the expected bid model is obtained by taking the logarithm of each bid price and inputting it along with the normalized features into the regression solution process on a sample set of similar historical bidding projects, thus obtaining the corresponding constant intercept coefficient; the expected bid model... The dimensional characteristic coefficients are obtained by applying the first characteristic coefficient in the same regression solution process. The weights of the normalized features are obtained by minimizing the objective function of "the error between the predicted logarithmic quote and the actual logarithmic quote"; the 3rd... Dimensional feature values ​​can be collected by extracting the first feature value from the business data of the current bidding project and the bidding entity. The dimensions are obtained after numerical conversion.

[0145] Furthermore, for each bidding entity in the current bidding project, the bid price is obtained, and collectable feature values ​​are constructed, which are then substituted into the linear expected price model to obtain the expected price and calculate the bid price residual.

[0146] A benchmark for bid price residuals is statistically analyzed from a set of similar historical bid samples, and the bid price residuals are standardized based on this benchmark.

[0147] For any two bidders in the current bidding project, obtain their corresponding standardized bid price residuals, calculate the bid price residual coherence, and express it as:

[0148] in, This indicates the degree of consistency in the residual bid prices between the two bidding entities. This represents the standardized bid price residual of the first bidding entity. This represents the standardized bid price residual of the second bidding entity.

[0149] When the bid price residual coordination degree When the value is close to 1, it indicates that the deviations of the two entities from the expected bids are highly similar, which can be used as evidence of bid-rigging risk.

[0150] Furthermore, the event sequence for each bidding entity is obtained from the event log of the bidding process; the events include logging in, downloading the bidding documents, downloading the Q&A documents, uploading, withdrawing, and resubmitting.

[0151] Each event is assigned a unified event type code and an occurrence timestamp, and then the events are ordered in ascending order by occurrence timestamp to form an event sequence.

[0152] Calculate the time interval sequence for adjacent events.

[0153] For each bidding entity, the number of transitions from event type to event type in the event sequence is statistically analyzed and normalized into a transition probability distribution.

[0154] The overlap ratio of transfer distributions for any two bidding entities is calculated and expressed as: ; in, Indicates the overlap ratio of the transfer distribution. Indicates the number of transition probability components. The first bidder is represented by the [number]th [unit]. Each transition probability, The second bidding entity is represented by the first Each transition probability.

[0155] Construct a time interval probability density function for the time interval sequence of each bidding entity.

[0156] The overlap ratio of rhythm distributions for any two bidding entities is calculated and expressed as: ; in, Indicates the overlap ratio of rhythm distribution. This indicates the upper limit of the time interval integration. The probability density function representing the time interval of the first bidding entity. The probability density function representing the time interval of the second bidding entity. This represents the independent variable in the integral.

[0157] The overlap ratios of transition distributions and rhythm distributions are fused with equal weights to obtain behavioral homology. , is represented as: ; S7. Based on the synergy of bid price residuals and the homogeneity of behavior, generate a risk classification of bid rigging and write it into the report.

[0158] Furthermore, for any two bidding entities, the residual synergy and behavioral homogeneity are mapped to risk to obtain the risk scores of the two bidding entities, expressed as: ; in, This indicates the risk level of the two bidding entities. This represents the risk mapping intercept coefficient. Represents the residual collaborative weighting coefficient. This represents the homogeneity weighting coefficient for behaviors.

[0159] It should be noted that the risk mapping intercept coefficient was obtained by fitting a maximum likelihood logistic regression on training samples labeled with historical bid documents indicating whether or not collusion occurred, with the real labels of the two bidders as the target and residual synergy and behavioral homology as independent variables; the residual synergy weight coefficient was obtained by estimating the regression coefficient corresponding to the residual synergy through maximum likelihood in the same logistic regression fitting process; and the behavioral homology weight coefficient was obtained by estimating the regression coefficient corresponding to the behavioral homology through maximum likelihood in the same logistic regression fitting process.

[0160] When the risk scores of the two bidding entities are not lower than the edge threshold, an edge is established in the relationship graph.

[0161] Connected components are extracted from the relationship graph as candidate groups.

[0162] For any candidate group, the group risk value is calculated by averaging the risk scores of each pair of groups.

[0163] When the group's risk value is not less than the high-risk threshold, output high risk; when the group's risk value is less than the high-risk threshold but not less than the medium-risk threshold, output medium risk; otherwise, output low risk.

[0164] For each candidate group, the review report should include a list of group members, risk level, group risk value, and a summary of triggering factors. The summary of triggering factors includes standardized residual statistics, overlap ratio of transfer distribution, overlap ratio of rhythm distribution, and the maximum and mean of the risk scores of the two bidding entities.

[0165] It also provides an event log range index and a bid price source index.

[0166] It should be noted that the edge threshold is calculated by distributing the risk scores of the two bidding entities on the marked samples of historical bid documents, and the risk score value that makes the recall rate of the collusion sample pair not lower than the allowable lower limit and the false positive rate the lowest is selected as the edge threshold; the high-risk threshold is calculated by calculating the risk value of candidate groups in historical bidding projects, and the minimum risk value that makes the precision rate of high-risk groups reach the set target and covers the main confirmed groups is selected as the high-risk threshold; the medium-risk threshold is selected from the same calculation result, and the risk value that makes the recall rate of medium-risk groups reach the set target and controls the proportion of low-risk false positives within the allowable upper limit is selected as the medium-risk threshold, and the medium-risk threshold should be less than the high-risk threshold.

[0167] In summary, this invention achieves evidence authenticity verification without disclosing the entire tender document by disclosing only the smallest piece of evidence corresponding to non-conformities and simultaneously outputting the verification path when generating the review report. This is beneficial for balancing the protection of trade secrets in tender documents with the verifiability requirements of regulatory review, and reduces the risk of unnecessary information dissemination. By calculating bid price residuals based on historical data and combining them with the event logs of the bidding process to generate common characteristics of bid rigging behavior, this invention achieves joint characterization and interpretable hierarchical early warning of bid price deviation behavior, which is beneficial for improving the effectiveness of discovering bid rigging clues and enhancing the interpretability and credibility of risk assessment.

[0168] 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 large-scale intelligent review method for tender documents based on artificial intelligence, characterized in that: include, Obtain the tender documents, all bid documents, and event logs of the same bidding project, write them to big data storage, and retain the original immutable records; The layout of each tender document is parsed and the structure is extracted to obtain field records and bind page numbers and coordinates as evidence anchors to each field record; The terms of the tender documents are compiled into executable review rules, and each tender document is checked against a uniform standard, outputting non-conformities and evidence anchors for non-conformities; The tender documents are divided into evidence segments according to page number areas, the Merkel root value of the evidence segments is calculated and written into an immutable evidence storage medium; When generating the review report, only the smallest piece of evidence corresponding to the non-compliance and the corresponding verification path are disclosed; Based on historical similar bidding projects and bidding process event logs, calculate the coherence of bid price residuals and the homogeneity of behavior; Based on the synergy of bid price residuals and the homogeneity of behavior, a risk classification for bid rigging is generated and written into the report.

2. The large-scale intelligent review method for tender documents based on artificial intelligence as described in claim 1, characterized in that: The page layout parsing refers to obtaining the original bid documents of the bidding entities from the bidding group, generating read-only processing task records, and establishing a page index table; The tender document is rendered as a page image, and a text-first extraction strategy is executed to obtain a sequence of character objects. The sequence of character objects from the page image is then input into the layout detection model to obtain a set of layout blocks.

3. The large-scale intelligent review method for tender documents based on artificial intelligence as described in claim 2, characterized in that: The structured extraction refers to constructing a directed graph with page blocks as nodes, performing topological sorting on the directed graph to obtain the reading order, identifying chapter title blocks in the page blocks according to the reading order, constructing a chapter node tree, and attaching each text block and table block to its respective chapter node. Perform table structure restoration on the table block and retrieve the field records corresponding to the table block; After segmenting the text block by sentence, field extraction is performed to obtain the field records corresponding to the text block.

4. The large-scale intelligent review method for tender documents based on artificial intelligence as described in claim 3, characterized in that: The evidence anchor point for binding page numbers and coordinates to each field record includes: for each field record, the page number is taken from the page number of the character object, the coordinates are taken from the smallest bounding rectangle of the character object's bounding box, and the original text fragment is taken from the normalized text within the bounding box's coverage area. The page number and coordinates are bound to the field to obtain the anchor point for each field record. Write the anchor point, field name, and field value of each field record into the structured result table to obtain the evidence anchor point for each field.

5. The large-scale intelligent review method for tender documents based on artificial intelligence as described in claim 4, characterized in that: The output non-compliance items and the evidence anchors for the non-compliance items include, Obtain the original bidding documents from the bidding project, analyze the layout, and obtain the set of bidding side layouts; Generate tender-side evidence anchors for the character records of each text block and table block; The tender document text is segmented according to a chapter tree to obtain the clause sequence; Perform field mapping and constraint element extraction on the text of each clause to obtain the target fields and constraint relationships; The constraints on the target fields corresponding to the clauses are broken down into sub-rules according to their order of appearance in the clauses; Each sub-rule is compiled into an executable expression, and conflicts are resolved according to priority to obtain the executable review rules for the tender documents. For each tender document's field set and field evidence anchors, each rule is evaluated according to the executable review rules. When a rule is deemed unmet, non-compliance items and their evidence anchors are generated.

6. The large-scale intelligent review method for tender documents based on artificial intelligence as described in claim 5, characterized in that: The specific steps for dividing the tender document into evidence segments according to page number areas, calculating the Merkel root value of each evidence segment, and writing it into an immutable evidence storage medium are as follows: Determine the page coordinate system based on the original tender documents and page index table of the bidding entity; Create a candidate region grid for each page and obtain the candidate region set; The candidate region set is segmented, and evidence fragments are generated for each candidate region; The evidence fragments are standardized to generate a standardized string of evidence fragments; The standardized string of evidence fragments is concatenated with page numbers and coordinate frames to form a leaf input; Based on the leaf input, the leaf digest value is calculated using a collision-resistant hash function; For all leaf summary values ​​in the tender document, a binary Merck tree is used to concatenate the summaries pairwise from bottom to top to generate the parent node summary, until the root value is obtained; Write the Merkel root value of the tender document into an immutable storage medium.

7. The large-scale intelligent review method for tender documents based on artificial intelligence as described in claim 6, characterized in that: The disclosure of only the minimum evidentiary fragment corresponding to the non-compliance item and the corresponding verification path includes: Cluster the anchor points of the bid-side evidence for each non-compliance item to obtain candidate anchor point groups, and generate a minimum disclosure area for each candidate anchor point group according to the minimum disclosure method; For each minimum disclosure area, extract the content covered by the area from the structured results of the tender documents, perform normalization processing, and obtain the disclosure text string; Based on the page number and coordinate frame of the disclosure area, the evidence fragment with the highest spatial overlap is retrieved from the fragment index. For each selected evidence fragment leaf summary, the sibling node summary sequence from the leaf to the root is extracted from the Merkle tree structure to form a verification path. The verification path, root value, and evidence receipt number are encapsulated into a verification material package, and the verification material package is bound to the disclosed evidence fragments; When generating the review report, for each non-compliance item, output the rule number, the tendering side evidence anchor point, the minimum disclosure evidence fragment index, the disclosure text string, and the verification material package index.

8. The large-scale intelligent review method for tender documents based on artificial intelligence as described in claim 7, characterized in that: The bid price residual consistency includes, Extract historical data of completed bid evaluations from historical bidding projects to generate a historical bid sample set; Extract collectable feature vectors and bid prices from each historical bid sample set to form training samples; A linear expected price model is established based on training samples. The expected price of any bidding entity is calculated based on the linear expected price model to obtain the expected price. For each bidder in the current bidding project, calculate the bid price residual; Based on the benchmark of bid price residual scale, the bid price residual is standardized, and for any two bidders in the current bidding project, the bid price residual synergy is calculated according to the standardized bid price residual.

9. The large-scale intelligent review method for tender documents based on artificial intelligence as described in claim 8, characterized in that: The behavioral homology includes, Obtain the event sequence for each bidding entity from the event log of the bidding process; For each bidding entity, the number of transitions of the same event type in the statistical event sequence is calculated, and the transition probability distribution is obtained. Based on the transition probability distribution, calculate the overlap ratio of the transition distributions of any two bidding entities; Calculate the time interval sequence for adjacent events, and calculate the rhythm overlap ratio for any two bidding entities; The overlap ratio of transition distribution and the overlap ratio of rhythm are fused to obtain the behavioral homology.

10. The large-scale intelligent review method for tender documents based on artificial intelligence as described in claim 9, characterized in that: The process of generating a risk classification for bid rigging and writing it into a report includes performing risk mapping on the degree of coordination of bid residuals and the degree of homogeneity of behavior between any two bidders to obtain risk scores for the two bidders. When the risk scores of the two bidding entities are not less than the edge threshold, an edge is established in the relationship graph of the two bidding entities. Extract the connected components of the current bidding entity from the relationship graph as candidate groups, and calculate the group risk value for any candidate group based on the average of the pairwise risk scores within the group. Based on the group's risk value, a risk classification for bid-rigging is generated and written into the review report.

Images