Two-layer semantic content elimination and cross-validation surrounding string flag detection method and system

Abstract

The application discloses a two-layer semantic content elimination and cross-validation surrounding bid detection method and system. The application belongs to the technical field of surrounding bid detection, and aims to solve the problems of large workload, weak pertinence, single judgment basis and inaccurate judgment results caused by full-text comparison in the existing detection technology. The method comprises the following steps: obtaining original text corpus of a bidding document and bidding documents of multiple bidding enterprises in a bidding project; performing paragraph segmentation on the text of the bidding documents to obtain a first paragraph set; eliminating contents with similarity greater than or equal to a preset first threshold in the first paragraph set to obtain a second paragraph set; eliminating template paragraphs in the second paragraph set to obtain a third paragraph set; performing similarity calculation on the third paragraph sets corresponding to any two bidding enterprises respectively to obtain semantic similarity scores; performing multi-dimensional auxiliary evidence feature cross-validation detection based on the bidding documents to obtain cross-validation detection results; and judging whether the multiple bidding enterprises have surrounding bid risks based on the semantic similarity scores and the cross-validation detection results, and outputting risk judgment results.

Description

Technical Field

[0001] This invention belongs to the field of token rigging detection technology, specifically relating to a token rigging detection method and system based on two-layer semantic content removal and cross-validation. Background Technology

[0002] Bid rigging, also known as collusion, refers to the illegal act of multiple suppliers manipulating bidding results by colluding on prices and preparing bid documents in a unified manner. It has long been a key target of governance in the field of bidding supervision. It not only seriously undermines the fair competition market environment, but also easily leads to a series of problems such as inferior suppliers winning bids, inflated procurement costs, and loss of control over project quality, thereby harming the legitimate rights and interests of the procuring party and other compliant suppliers.

[0003] From the perspective of bid document characteristics, collusion in bidding is generally accompanied by a high degree of similarity in bid document content. Colluding parties typically compile bid documents in bulk by copying and pasting identical technical solutions, after-sales service commitments, integrity pledges, project case studies, etc., resulting in extremely strong homogeneity at the text level. Therefore, bid document similarity detection based on text similarity analysis is one of the core technical approaches for intelligent identification of collusion in bidding.

[0004] Existing text similarity-based bid-rigging detection technologies convert bid documents into text vectors and then use cosine similarity to determine the degree of similarity. However, this approach fails to distinguish between template content and self-written content, leading to false positives. Deep semantic analysis solutions based on natural language processing can identify semantic rewriting-based similarities by calculating similarity through pre-trained model encoding. However, this requires full content comparison, which is labor-intensive and lacks specificity. Furthermore, relying solely on text similarity to assess bid-rigging risk provides a single criterion, resulting in a high false positive rate and failing to offer effective reference. Summary of the Invention

[0005] (a) Purpose of the invention The purpose of this invention is to provide a method and system for detecting tag-rigging based on two-layer semantic content removal and cross-validation, which aims to solve the problems of large workload, weak targeting, single judgment basis and inaccurate judgment results caused by full-text comparison in existing detection technologies.

[0006] (II) Technical Solution To address the aforementioned problems, a first aspect of the present invention provides a method for detecting cross-labeling based on two-layer semantic content removal and cross-validation, the method comprising: Obtain the original text of the bidding documents for the bidding project and the bid documents of multiple bidding companies; Based on the text of the tender document, paragraph segmentation is performed to obtain the first paragraph set; Based on the similarity between the original text corpus and the first paragraph set, content with a similarity greater than or equal to a preset first threshold in the first paragraph set is removed to obtain the second paragraph set; Based on the second set of paragraphs, the template paragraphs in the second set of paragraphs are removed to obtain the third set of paragraphs; The similarity score is obtained by calculating the similarity of the third paragraph sets corresponding to any two bidding companies. Based on the tender documents, multi-dimensional auxiliary evidence feature cross-validation detection is performed to obtain cross-validation detection results. The cross-validation detection includes the detection of identical typos, grammatical habit feature detection, and non-qualification image feature detection. Based on the semantic similarity score and cross-validation detection results, it is determined whether there is a risk of bid rigging among the multiple bidding companies, and the risk assessment result is output.

[0007] Preferably, obtaining the original text corpus of the bidding documents in the bidding project includes: Extract a first candidate paragraph with a length greater than a preset number of characters from the first format document of the tender document; Sentences are segmented from the second format document of the tender document according to punctuation marks or line breaks, and second candidate paragraphs with a length within a preset character range are selected. The original text corpus is constructed by deduplicating the first candidate paragraph and the second candidate paragraph.

[0008] Preferably, the first paragraph set obtained by segmenting the text of the tender document includes: Divide the text into initial segments based on line breaks and punctuation marks; The merged segment is obtained by merging adjacent initial segments based on the initial segment, and the paragraph length of the merged segment is between a preset minimum paragraph length and a preset maximum paragraph length; Discard any segments in the merged fragments whose length is less than the minimum paragraph length to obtain the first paragraph set.

[0009] Preferably, the step of removing content from the first paragraph set with a similarity greater than or equal to a preset first threshold based on the similarity between the original text corpus and the first paragraph set to obtain a second paragraph set includes: Calculate the similarity between each paragraph in the first paragraph set and each corpus entry in the original text corpus; When the similarity between any piece of corpus and the current paragraph is greater than or equal to the preset first threshold, the current paragraph is determined to be a copy of the original text of the tender document. The copied text is then removed to obtain the second paragraph set.

[0010] Preferably, the step of removing template paragraphs from the second paragraph set to obtain a third paragraph set includes: Identify template paragraphs in the second set of paragraphs, the template paragraphs including technical specification table paragraphs, response table paragraphs, and standardization commitment template paragraphs; The template paragraphs in the second paragraph set are removed to obtain the third paragraph set.

[0011] Preferably, the step of removing template paragraphs from the second paragraph set further includes: A preset set of template noise words, which includes at least one of the following: power of attorney, legal representative, authorized agent, authorized representative, no sub-authorization allowed, and photocopy of ID card; When a paragraph matches a term in the template noise word set, it is determined to be template noise and the current paragraph is removed.

[0012] Preferably, the step of calculating the similarity of the third paragraph sets corresponding to any two bidding companies to obtain a semantic similarity score includes: For each paragraph in the third paragraph set of the first bidding company, traverse all paragraphs in the third paragraph set of the second bidding company and calculate the sequence matching similarity. Record the best-match similarity and best-match paragraph for each paragraph of the first bidding company; Collect highly similar paragraph pairs whose best matching similarity is not lower than a preset second threshold; The semantic similarity score is calculated based on the highly similar paragraphs; The semantic similarity score includes the highest paragraph similarity, the number of highly similar paragraph pairs, and the average similarity of highly similar paragraph pairs.

[0013] Preferably, the step of determining whether multiple bidding companies pose a risk of bid rigging based on the semantic similarity score and cross-validation detection results, and outputting the risk determination result, includes: When the highest paragraph similarity is greater than or equal to a preset third threshold and the number of highly similar paragraph pairs is greater than or equal to a preset fourth threshold, it is determined to be semantically high risk; When the highest paragraph similarity is greater than or equal to the third threshold and the number of highly similar paragraph pairs is less than the fourth threshold, it is determined to be semantically risky; When the highest paragraph similarity is less than the third threshold, it is determined to be semantically low risk; When the semantic risk is high and there is at least one matching auxiliary evidence feature, it is determined to be of extremely high risk of collusion. When the semantics are high-risk and there is no supporting evidence to match, or when the semantics are medium-risk and there are at least two supporting evidence features to match, it is determined to be a high-risk case of collusion. When the semantics are of medium risk and there is a matching piece of auxiliary evidence, or when the semantics are of low risk and there is a matching piece of the same misspelling feature, it is determined to be of medium risk in the collusion.

[0014] Preferably, the method further includes: Count the number of paragraphs in the third paragraph set for each bidding company; When the number of paragraphs for any of the bidding companies is less than the preset minimum paragraph count threshold, the semantic similarity score involving that bidding company will be set to zero, and the sample will be marked as insufficient.

[0015] Another aspect of the present invention provides a system for detecting scrambling and cross-validation based on two-layer semantic content removal and cross-validation, the system comprising: The data acquisition module is used to acquire the original text of the bidding documents and the bid documents from multiple bidding companies in the bidding project; The paragraph segmentation module is used to segment the text of the tender document into paragraphs to obtain a first set of paragraphs. The first-layer semantic removal module is used to remove content in the first paragraph set whose similarity is greater than or equal to a preset first threshold based on the similarity between the original text corpus and the first paragraph set, so as to obtain the second paragraph set. The second-layer semantic elimination module is used to eliminate template paragraphs from the second paragraph set to obtain the third paragraph set. The similarity calculation module is used to calculate the similarity of the third paragraph sets corresponding to any two bidding companies to obtain a semantic similarity score; The cross-validation detection module is used to perform multi-dimensional auxiliary evidence feature cross-validation detection based on the tender documents to obtain cross-validation detection results. The cross-validation detection includes the detection of identical typos, the detection of grammatical habit features, and the detection of non-qualification image features. The risk assessment module is used to determine whether there is a risk of bid rigging among multiple bidding companies based on the semantic similarity score and cross-validation detection results, and output the risk assessment results.

[0016] (III) Beneficial Effects The above-described technical solution of the present invention has the following beneficial technical effects: 1. Through a two-layer semantic removal mechanism, the original copied content of the tender documents and various standard template content are separated one after the other. The similarity is calculated only for the content that the supplier actually writes independently. This completely eliminates the false high similarity caused by general templates, so that the similarity results can truly reflect the degree of similarity of the supplier's original content, reduce the amount of invalid manual review work, and enable the detection results to truly reflect whether there is plagiarism or collusion among bidding companies.

[0017] 2. Three independent auxiliary evidence dimensions are introduced: identical misspellings, grammatical habits, and non-qualified images. These three dimensions corroborate the relevance of the bidding documents from three different perspectives: consistency of writing errors, commonality of writing style, and relevance of material reuse. They complement each other and form a chain of evidence. When the semantic similarity is in a fuzzy range, the auxiliary evidence can provide independent judgment support, avoid the limitations of a single dimension, and improve the accuracy of the judgment on bid rigging.

[0018] 3. A progressive processing logic of coarse filtering, fine screening, and verification is adopted. First, the first layer quickly removes the most prevalent copied content, significantly reducing the amount of text to be processed. Then, the second layer accurately identifies relatively hidden standard templates. Finally, only the remaining small amount of self-written content is subjected to fine similarity calculation and multi-dimensional detection. This avoids performing complex calculations on the entire text and concentrates computing power on the most valuable self-written content, thus reducing computing costs while ensuring detection accuracy. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method for detecting cross-label collusion based on two-layer semantic content removal and cross-validation provided by the present invention; Figure 2 This is a schematic diagram of the structure of the tag-matching detection system based on two-layer semantic content removal and cross-validation provided by the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and the accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0021] The accompanying drawings illustrate layer structure diagrams according to embodiments of the present invention. These drawings are not to scale, and some details have been enlarged for clarity, and some details may have been omitted. The shapes of the various regions and layers shown in the drawings, as well as their relative sizes and positional relationships, are merely exemplary and may deviate from reality due to manufacturing tolerances or technical limitations. Furthermore, those skilled in the art can design regions / layers with different shapes, sizes, and relative positions as needed.

[0022] Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0023] In the description of this invention, it should be noted that the terms "first" to "fifth" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0024] Combination Figure 1 The first aspect of the present invention provides a method for detecting tag-rigging based on two-layer semantic content removal and cross-validation, the method comprising: Obtain the original text of the bidding documents for the bidding project and the bid documents of multiple bidding companies; The first set of paragraphs is obtained by segmenting the text of the tender document. Based on the similarity between the original text corpus and the first paragraph set, content with a similarity greater than or equal to a preset first threshold in the first paragraph set is removed to obtain the second paragraph set; Based on the second set of paragraphs, the template paragraphs in the second set of paragraphs are removed to obtain the third set of paragraphs; The similarity scores are calculated for the sets of third paragraphs corresponding to any two bidding companies to obtain semantic similarity scores. Based on the tender documents, cross-validation detection of multi-dimensional auxiliary evidence features is performed to obtain cross-validation detection results. Cross-validation detection includes detection of identical typos, detection of grammatical habit features, and detection of non-qualification image features. Based on semantic similarity scoring and cross-validation detection results, the system determines whether multiple bidding companies are at risk of collusion and outputs the risk assessment results.

[0025] Specifically, the process begins by collecting the text of all tender documents for the target project. The original text refers to unprocessed raw language materials, serving as foundational resources for language research, model training, or translation reference. Simultaneously, the text of all participating companies' tender documents is extracted. Standardized paragraph segmentation is performed on each tender document, organizing the disorganized raw text into uniform paragraph units, resulting in the first paragraph set. Next, a first-level semantic removal process is performed, comparing the similarity of each paragraph with the original tender document text. Paragraphs reaching a threshold are identified as directly copied from the original tender document and removed, resulting in the second paragraph set. A second-level semantic removal process is then performed on the remaining paragraphs after the first-level removal, identifying and removing standardized template paragraphs such as technical specification tables, response forms, and standardized commitment templates, resulting in a third paragraph set containing only content written by the suppliers. Based on this third paragraph set, pairwise paragraph similarity calculations are performed between any two bidding companies, generating multi-dimensional semantic similarity scores. Simultaneously, three auxiliary evidence checks are conducted on the tender documents: identical typos, grammatical errors, and non-qualification images, yielding cross-validation results. Finally, by combining semantic similarity scores and cross-validation results, the risk level of collusion between the two bidding companies is output according to the preset judgment rules.

[0026] This method employs a two-layer semantic removal mechanism to first separate copied content from the original tender documents from various standardized templates, calculating similarity only for content genuinely written by the suppliers. This completely eliminates false high similarity caused by generic templates, ensuring that the similarity results truly reflect the degree of similarity between the suppliers' original content. This reduces the workload of ineffective manual review and allows the detection results to accurately reflect whether plagiarism or collusion exists among bidding companies. Three independent dimensions of auxiliary evidence are introduced: identical typos, grammatical habits, and non-qualified images. These three dimensions corroborate the relevance of the tender documents from three different perspectives: consistency of writing errors, similarity of writing style, and relevance of material reuse. These complementary evidence chains provide independent support when semantic similarity is in an ambiguous range, avoiding the limitations of a single dimension and improving the accuracy of conclusions regarding bid rigging. The system employs a progressive processing logic of coarse filtering, fine screening, and verification. First, it quickly removes the most prevalent copied content from the original text through the first layer, significantly reducing the amount of text to be processed. Then, it accurately identifies relatively hidden standard templates through the second layer. Finally, it performs fine similarity calculations and multi-dimensional detection only on the remaining small amount of self-written content. This avoids performing complex calculations on the entire text and concentrates computing power on the most valuable self-written content, thus reducing computational costs while ensuring detection accuracy.

[0027] In a preferred scenario, the original text of the bidding documents for the bidding project is obtained, including: Extract the first candidate paragraph with a length greater than the preset number of characters from the first format document of the tender document; Sentences are segmented from the second format document of the tender document according to punctuation marks or line breaks, and second candidate paragraphs with a length within a preset character range are selected. The original text corpus was constructed by removing duplicates from the first and second candidate paragraphs.

[0028] Specifically, for the first format tender document, the original paragraph structure of the document is directly read, all paragraphs are traversed, and paragraphs with a length greater than a preset character limit are selected as the first candidate paragraphs. For the second format tender document, the entire text content is first extracted, and then it is divided into independent sentences using period, semicolon, exclamation mark, and other sentence-ending punctuation and line breaks as delimiters. Sentences with a length within a preset character range are selected as the second candidate paragraphs. The first and second candidate paragraphs are merged, and deduplication is performed according to the text content to eliminate duplicate sentences and paragraphs, ultimately obtaining a well-organized original text corpus of the tender document, which is used for subsequent matching and comparison in the first layer of semantic removal. In one optional scenario, the first format document is a Word document, and the second format document is a PDF document; the preset character limit is ten characters, and the preset character range is 10 to 500 characters.

[0029] This method employs adaptive extraction strategies for different formats, directly extracting paragraphs from Word documents and reconstructing sentences from PDFs through punctuation. It comprehensively collects original text content from tender documents in various formats, avoiding corpus gaps due to format limitations. This ensures that all copied paragraphs are fully identified during the first-level semantic removal, preventing template content omissions. By setting a character count threshold to filter candidate paragraphs, it ensures that each piece of text in the corpus has sufficient semantic discriminative power. This reduces invalid computations in subsequent sequence matching, improves the efficiency of the first-level removal, and avoids false removals caused by broad matching of short texts, guaranteeing the retention rate of original content and the accuracy of corpus matching. Simultaneously, by simplifying the corpus through deduplication, the matching computation in the first-level semantic removal stage is significantly reduced without affecting the coverage of the original text, improving detection speed. The deduplicated corpus also avoids redundant matching results caused by duplicate text, making the similarity judgment logic simpler and more reliable, facilitating subsequent threshold optimization and result analysis.

[0030] In a preferred embodiment, the text of the tender document is segmented into paragraphs to obtain a first set of paragraphs, including: Divide the text into initial segments based on line breaks and punctuation marks; The merged segment is obtained by merging adjacent initial segments based on the initial segment. The paragraph length of the merged segment is between the preset minimum paragraph length and maximum paragraph length. Discard any segments in the merged sequence whose length is less than the minimum paragraph length to obtain the first paragraph set.

[0031] Specifically, the entire text of the tender document is first read. Using line breaks and sentence-ending punctuation such as periods, semicolons, and exclamation marks as delimiters, the continuous text is divided into several initial text segments, each corresponding to a minimum semantic unit. Then, a merging strategy is used to traverse all initial segments from left to right, maintaining a current segment buffer. Adjacent initial segments are progressively appended to the buffer until the buffer length is close to but does not exceed the preset maximum segment length. At this point, the buffer content is output as a complete segment, the buffer is cleared, and subsequent segments are processed. After the traversal is complete, all merged segments are counted, and segments shorter than the preset minimum segment length are discarded. The remaining content constitutes the first set of segments with uniform granularity.

[0032] In one implementation, adjacent initial segments are merged to obtain a merged segment. A greedy merging algorithm is then used to merge adjacent segments, keeping the segment length within the range of 50 to 500 characters. Segments shorter than 50 characters are discarded. The greedy merging algorithm works as follows: traverse the initial segments from left to right, maintain a current segment buffer, and when the length of the buffer plus the length of the next segment does not exceed the maximum segment length, append the next segment to the buffer; otherwise, output the contents of the buffer as a segment, clear the buffer, and start again from the next segment. After the traversal is complete, output the remaining contents of the buffer as the last segment.

[0033] This method, which involves first segmenting and then merging, standardizes all bid document text into uniform length ranges, forming consistent paragraph units. This ensures that similarity calculations across different suppliers are performed at a unified text granularity, guaranteeing the objectivity and fairness of the results and avoiding biases caused by formatting differences. Simultaneously, by merging adjacent segments, semantic units truncated by formatting are reassembled into complete paragraphs, restoring the original semantic structure of the text. This addresses the fragmentation issue in PDF-format bid documents, allowing subsequent semantic removal and similarity calculations to be based on complete semantic units, thus improving the reliability of the detection results. By setting a minimum paragraph length and discarding excessively short segments, invalid text noise is effectively filtered out. This not only reduces the computational load of the subsequent two layers of semantic removal and similarity calculation but also avoids false high similarity caused by broad matching of short segments, further reducing the probability of false positives and allowing the detection results to focus more on meaningful semantic content.

[0034] In a preferred embodiment, based on the similarity between the original text corpus and the first paragraph set, content in the first paragraph set with a similarity greater than or equal to a preset first threshold is removed to obtain the second paragraph set, which includes: Calculate the similarity between each paragraph in the first paragraph set and each corpus entry in the original text corpus; When the similarity between any piece of corpus and the current paragraph is greater than or equal to a preset first threshold, the current paragraph is determined to be a copy of the original text of the tender document. The copied text is then removed, and the second paragraph set is obtained.

[0035] Specifically, the completed original text corpus of the tender document is loaded, and each target paragraph in the first paragraph set is traversed. For each target paragraph, the sequence similarity is calculated with each original text corpus in the corpus. During the corpus traversal, if any corpus has a similarity to the target paragraph greater than or equal to a preset first threshold, such as a similarity greater than or equal to 0.8, the subsequent corpus traversal of that paragraph is immediately terminated, and the paragraph is marked as a copy of the original text of the tender document. After the matching judgment of all paragraphs is completed, all paragraphs marked as copies of the original text are removed, and the remaining paragraphs form the second paragraph set, which enters the second layer of semantic removal.

[0036] This method employs full-scale segment-by-segment matching to ensure comprehensive removal of original text content, eliminating blind spots and completely stripping away all copied content. It avoids residual template content caused by scattered or obscure original text locations, clearing away the primary interference for subsequent second-layer removal and similarity calculation. For each target segment, subsequent corpus traversal terminates as soon as any matching text meeting the threshold is found, eliminating the need to traverse the entire corpus. This significantly reduces the amount of matching computation, improving the efficiency of the first-layer removal. It ensures thorough removal, preventing omissions due to partial matching, and optimizes algorithm efficiency through early termination of matching, making it suitable for processing large volumes of tender documents. By first removing the template content that accounts for the largest proportion in the first layer, the number of paragraphs entering the second layer of semantic removal can be significantly reduced, thus reducing the processing pressure of template recognition in the second layer. The two-layer progressive removal architecture, which is from easy to difficult and from coarse to fine, is more efficient than the solution of performing all template recognition at once, and it is also easier to improve the overall removal accuracy through layer-by-layer optimization. At the same time, the two layers run independently, and the rule adjustment of one layer will not affect the other layer, which facilitates subsequent continuous optimization.

[0037] In a preferred embodiment, based on the second paragraph set, template paragraphs are removed from the second paragraph set to obtain a third paragraph set, including: Identify template paragraphs in the second paragraph set, which include technical specification table paragraphs, response table paragraphs, and standardization commitment template paragraphs; Remove the template paragraphs from the second paragraph set to obtain the third paragraph set.

[0038] Specifically, for the second set of paragraphs after the first layer of elimination, each paragraph undergoes three types of template feature identification. For technical specification tables and response tables, matching and identification are performed using features such as numbering format, technical parameter patterns, and standardized terminology. For standardized commitment template paragraphs, identification is performed using a dual rule combining Chinese numbering format and commitment keywords. As long as a paragraph matches the identification rules of any one type of template, it is marked as a template paragraph. After completing the identification of all paragraphs, all marked template paragraphs are eliminated, and the remaining content constitutes the third set of paragraphs composed of content written by the supplier.

[0039] This method, with its second layer specifically targeting and eliminating these standardized templates, fills the gaps in the first layer's coverage. It further isolates non-self-written standardized content, ensuring the final set of retained paragraphs more closely reflects the supplier's actual self-created content. This reduces interference from template similarity and improves detection accuracy. The elimination scope includes three typical standardized template types: technical specification tables, response forms, and standardized commitment templates. This covers the vast majority of non-original template types in tender documents, adapting to the structure of tender documents for different categories of projects, including engineering, goods, and services, demonstrating strong versatility and scenario adaptability. In this two-layer elimination architecture, each layer focuses on a specific type of template feature: the first layer focuses on overlapping features in the original text, and the second layer focuses on standardized format features. The recognition rules for each layer can be precisely designed for the corresponding features, reducing the probability of misjudgments due to feature overlap. Furthermore, after the first layer eliminates a large amount of content, the second layer processes smaller paragraphs, facilitating the use of more refined recognition rules without excessively impacting efficiency, thus achieving a balance between elimination accuracy and operational efficiency.

[0040] In a preferred embodiment, removing template paragraphs from the second paragraph set also includes: A preset set of template noise words, which includes at least one of the following: power of attorney, legal representative, their authorized agent, their authorized representative, no sub-authorization allowed, and photocopy of ID card; When a paragraph matches a term in the template noise word set, it is identified as template noise and the current paragraph is removed.

[0041] Specifically, a template noise word set is pre-built, which includes common standard expressions commonly found in bidding documents, such as terms related to power of attorney and expressions related to legal representative identity certificates, which are standard content used across the industry. When processing the second paragraph set, keyword matching verification is performed on each paragraph. If the core content of a paragraph completely matches the words in the noise word set, it is determined to be a template noise paragraph. The paragraphs determined to be template noise are removed together with other template paragraphs to complete the second layer of semantic removal process.

[0042] This method pre-defines a set of template noise words, specifically filtering out standardized expressions common across projects. This fills the gap in removing general templates beyond project-specific templates, further reducing interference from non-autonomous content in similarity calculations and allowing similarity results to focus more on the supplier's personalized content. Simultaneously, in the two-layer semantic removal architecture, this lightweight filtering method can serve as a supplementary filtering step, further cleaning up residual fragmented template content without significantly increasing the overall computational load, achieving a balance between improved removal accuracy and operational efficiency.

[0043] In the preferred scenario, similarity is calculated for the third paragraph sets corresponding to any two bidding companies to obtain a semantic similarity score, including: For each paragraph in the third paragraph set of the first bidding company, iterate through all paragraphs in the third paragraph set of the second bidding company and calculate the sequence matching similarity. Record the best-match similarity and best-match paragraph for each paragraph of the first bidding company; Collect highly similar paragraph pairs with the best matching similarity score not lower than a preset second threshold; Calculate semantic similarity scores based on highly similar paragraphs; The semantic similarity score includes the highest paragraph similarity, the number of highly similar paragraph pairs, and the average similarity of highly similar paragraph pairs.

[0044] In one implementation, two sets of self-written content paragraphs from any two bidding companies are selected, denoted as set A and set B, respectively. Each paragraph in set A is traversed, and the sequence matching similarity is calculated with all paragraphs in set B. The highest similarity value corresponding to the paragraph and the corresponding paragraph in set B are recorded, which is the best matching result for the paragraph. After traversal, all paragraph pairs with the best matching similarity greater than or equal to a preset second threshold are selected. For example, if the best matching similarity is greater than or equal to 0.6, the paragraph pair is regarded as a highly similar paragraph pair. Based on the highly similar paragraph pairs, three core indicators are calculated: the highest similarity value among all highly similar paragraph pairs, the total number of highly similar paragraph pairs, and the average similarity of all highly similar paragraph pairs. These three indicators together constitute the semantic similarity score of this comparison.

[0045] This method not only more accurately measures the degree of similarity between self-produced content but also directly identifies specific similar paragraphs, providing a clear direction for subsequent manual review and making the detection results more practical. It directly supports the review work of evaluation experts and regulatory personnel, meeting the actual business needs of bidding and tendering supervision. Simultaneously, it outputs three indicators: highest paragraph similarity, number of highly similar paragraph pairs, and average similarity. These indicators characterize similarity features from three dimensions: the highest degree of similarity in a single paragraph, the coverage of similar content, and the average overlap of similar content. This provides a more comprehensive reflection of the similarity between the two suppliers' bids. Different combinations of indicators correspond to different similarity patterns, such as extremely high single-paragraph similarity but few instances, or a large number of paragraphs with moderate similarity. These multiple indicators provide a more sufficient basis for subsequent risk classification, supporting more refined risk grading. The mechanism of traversing paragraphs one by one to find the best match does not depend on the order of paragraphs. It can find the most similar corresponding content in all paragraphs of another supplier. It is not affected by paragraph restructuring or order adjustment. It can also effectively identify similarities in tender documents that have been slightly rewritten or structurally adjusted, thus improving the ability to detect covert bid-rigging behavior.

[0046] In a preferred embodiment, the detection of identical typos includes: Compare the words in the paragraphs of the tender document with a standard dictionary; Identify words that are not in the standard dictionary but have an edit distance of one from a word in the standard dictionary, and record them as candidate misspellings. When two bidding companies have at least one identical misspelled word and correct word pair in their candidate word sets, it is determined to be a match of the same misspelled word feature.

[0047] Specifically, the third paragraph set of each bidding company is processed using Chinese word segmentation to obtain a complete word set. This word set is then compared with a standard modern Chinese dictionary, and words not included in the dictionary are selected as candidate words. The edit distance between each candidate word and a word in the dictionary is calculated, and the dictionary word with an edit distance of 1 is selected as the correct form corresponding to that candidate word, forming candidate pairs of misspelled words and correct words. The misspelled word candidate pairs of the two bidding companies are compared. If there is at least one set of completely identical misspelled words and corresponding correct words, it is determined to be a match of identical misspelled word features.

[0048] By means of this method, taking identical typos as one of the auxiliary evidences can provide strong evidential support for the determination of collusive bidding. Especially when the semantic similarity is at a critical value, the existence of identical typos can greatly improve the confidence of the determination, make the detection conclusion more convincing, and better meet the rigor requirements of supervision and evidence collection. By defining the screening rule with an edit distance of 1, it is possible to accurately identify typical typos caused by writing errors and input method association errors. Such errors conform to the general error rule of writing, and can effectively eliminate the interference of non-error uncommon words such as professional terms and industry-specific expressions, avoid misjudging normal professional expressions as typos, ensure the accuracy of typo detection, reduce false positives in the auxiliary evidence dimension, and avoid increasing the risk level due to false typo matching errors.

[0049] In a preferred case, the grammatical habit feature detection includes: For the independently written content paragraphs of each bidding enterprise, count the usage frequency ratio of structural auxiliary words, the usage preference of specific connecting words and the occurrence frequency of uncommon description methods, so as to construct a grammatical habit feature vector; When the cosine similarity of the grammatical habit feature vectors of two bidding enterprises is not lower than a preset fifth threshold, it is determined that the grammatical habit features match.

[0050] Specifically, the feature dimensions of grammatical habits are predetermined, including the proportion of usage frequency of structural auxiliary words, the usage preference of preset high-frequency connecting words, and the occurrence frequency of preset uncommon description sentences, wherein the structural auxiliary words include "de (of)", "de (adjective suffix)", and "de (adverb suffix)". Full-text statistics are performed on the third paragraph set of each bidding enterprise, the value of each feature dimension is calculated, and the grammatical habit feature vector is constructed after normalization processing. For any two bidding enterprises, calculate the cosine similarity of their grammatical habit feature vectors; if the cosine similarity is greater than or equal to the preset fifth threshold, for example, the cosine similarity is greater than or equal to 0.85, it is determined that the grammatical habit features of the two suppliers match.

[0051] This method provides independent evidence from the perspective of writing style, enabling the identification of subtle collusion through minor rewriting. It compensates for the limitations of pure text similarity detection, complementing content-level evidence and enriching the cross-validation evidence system to enhance the detection capability of covert collusion. Abstract grammatical habits are decomposed into multiple quantifiable statistical features. Feature vectors are constructed through objective frequency statistics, and then quantitative comparisons are performed using cosine similarity. The entire process is based entirely on data calculation, eliminating subjective judgment bias and ensuring stable and consistent identification of writing styles from the same source. It also facilitates parameter optimization, allowing adjustment of feature dimensions and thresholds based on the text characteristics of different industries to adapt to various detection needs. Furthermore, statistical analysis is conducted on all self-written content after two layers of filtering, covering all of the supplier's self-created texts. The sufficient sample size ensures that the statistically obtained features accurately reflect the supplier's overall writing style, avoiding feature distortion caused by local text bias. Simultaneously, since the statistics only target self-written content, interference from uniform expressions in template content is eliminated, ensuring that the statistically obtained grammatical features are entirely derived from the supplier's own writing habits, thus improving the effectiveness of feature comparison.

[0052] In a preferred embodiment, non-qualified image feature detection includes: Extract non-qualification-related image files from the tender documents; The extracted non-certified image files are subjected to duplicate detection, which includes at least one of precise duplicate detection, internet source trace detection, and perceptual hash supplementation detection.

[0053] Specifically, the bidding documents of each bidding company are analyzed, and all image files are extracted. Images of standardized qualification certificates such as business licenses, qualification certificates, and system certifications are excluded by filtering based on filename, attribute features, and content classification, while non-qualification images such as equipment photos, construction photos, and project case photos are retained. Non-qualification images undergo a triple detection process: first, the hash value of the image file is calculated and compared to see if there are any exactly duplicate images; second, the image metadata is read to check for traces of the same internet URL source; and third, the perceptual hash value of inaccurately duplicated images is calculated and compared to determine if they are similar images. If any of these detections results in a duplicate or homologous image, it is determined to be a non-qualification image feature match. Furthermore, the precise duplication detection includes calculating the hash value of the non-qualification image files. When two bidding companies have image files with the same hash value, they are determined to be precise duplications. The internet source trace detection includes detecting whether the metadata of the non-qualification image files contains internet URL traces. When the metadata of the image files of two bidding companies contains the same internet URL source, they are determined to be images from the same source. The perceptual hash supplementation detection includes using perceptual hashing to calculate the similarity of non-precise duplication images. When the similarity is not lower than a preset sixth threshold, they are determined to be similar images.

[0054] This method supplements the evidence chain from the visual material dimension by detecting non-qualified images. It forms a multimodal cross-validation with textual evidence such as text similarity, typos, and grammatical habits, further improving the evidence system. It can identify collusion behaviors that cannot be detected by textual detection alone, enhancing the comprehensiveness of collusion detection and preventing colluders from evading detection by replacing text or retaining images. Precise duplicate detection targets scenarios where the original image is directly copied; internet source trace detection targets scenarios where the same download source exists; and perceptual hash supplementary detection targets slightly modified images. This three-tiered mechanism progressively covers most typical scenarios of image reuse, ensuring that simple image modifications do not prevent identification and guaranteeing the coverage and detection rate of image-level detection. Standardized qualification images are first filtered out, and detection is only conducted on self-prepared images such as non-qualified project materials and on-site photos. This ensures that the detected images reflect the supplier's self-prepared work, and any duplicates or similar sources have a clear correlation, avoiding invalid matches caused by standardized qualification images and ensuring the reference value and evidentiary validity of the image detection results.

[0055] In the preferred scenario, based on semantic similarity scoring and cross-validation detection results, it is determined whether multiple bidding companies pose a risk of bid rigging, and the risk assessment result is output, including: When the highest paragraph similarity is greater than or equal to the preset third threshold and the number of highly similar paragraph pairs is greater than or equal to the preset fourth threshold, it is judged as semantically high risk; When the highest paragraph similarity is greater than or equal to the third threshold and the number of highly similar paragraph pairs is less than the fourth threshold, it is judged as semantically medium risk; When the highest paragraph similarity is less than the third threshold, it is judged as semantically low risk; When the semantic risk is high and there is at least one matching auxiliary evidence feature, it is judged as having an extremely high risk of collusion. When the semantic risk is high and there is no supporting evidence to match, or when the semantic risk is medium and there are at least two supporting evidence features to match, it is judged as a high risk of collusion. When there is a semantic risk and an auxiliary evidence matching, or a semantic risk and a matching of the same misspelling feature, it is judged as a risky collusion.

[0056] Specifically, firstly, based on the highest paragraph similarity and the number of highly similar paragraph pairs in the semantic similarity score, and comparing them with preset third and fourth thresholds (e.g., the third threshold is 0.8 and the fourth threshold is 5), the semantic risks of the two bidding companies are divided into three levels: high, medium, and low. That is, if the highest similarity meets the standard and the number of highly similar paragraph pairs meets the standard, it is considered high semantic risk; if only the highest similarity meets the standard, it is considered medium semantic risk; and if the highest similarity does not meet the standard, it is considered low semantic risk. Then, the number of matched auxiliary evidence items is counted, and the classification is performed according to the semantic risk level judgment rules. High semantic risk with at least one piece of auxiliary evidence is considered extremely high risk of collusion; high semantic risk with no auxiliary evidence, or medium semantic risk with at least two pieces of auxiliary evidence is considered high risk of collusion; medium semantic risk with one piece of auxiliary evidence, or low semantic risk with matching identical misspelling features is considered medium risk of collusion; all other cases are judged as low risk of collusion. Finally, the corresponding risk level result is output.

[0057] This method categorizes scenarios into three levels: extremely high risk (corresponding to highly suspicious scenarios with sufficient evidence), high risk (corresponding to highly suspicious scenarios at the semantic level), medium risk (corresponding to scenarios with some doubts), and low risk (corresponding to scenarios with no obvious anomalies). This grading system better aligns with the actual workflow of bidding and tendering supervision. Regulators and evaluation experts can take differentiated measures based on the risk level. For example, extremely high risk can be directly investigated, high risk can be listed as a key review target, and medium risk can be included in routine monitoring, improving the efficiency and targeting of supervision. For semantically high-risk scenarios, supplementary evidence upgrades the risk level to extremely high risk, strengthening the completeness of the evidence chain. Similarly, for scenarios with less prominent semantics but sufficient supplementary evidence, the risk level can be increased accordingly to avoid overlooking or concealing collusion. This linkage mechanism avoids the one-sidedness of single-dimensional judgment, ensuring that the final risk conclusion is based on both core textual similarity features and independent-dimensional evidence support, improving the rigor and comprehensiveness of the judgment logic. Furthermore, the judgment rules clearly state that when the semantic risk is low, only identical misspellings can trigger a medium-risk judgment, which reflects the weight difference of different evidence, conforms to the objective law of evidence strength, avoids the situation where weak evidence triggers a high-risk judgment on its own, ensures the rigor of risk judgment, and reduces misjudgment; at the same time, strong evidence can still play a role when the semantic similarity is low, which also avoids strong evidence being ignored, and achieves a reasonable correspondence between evidence strength and risk level.

[0058] In a preferred embodiment, the method further includes: Count the number of paragraphs in the third paragraph set for each bidding company; When the number of paragraphs for any bidding company is less than the preset minimum paragraph count threshold, the semantic similarity score involving that bidding company will be set to zero, and the sample will be marked as insufficient.

[0059] Specifically, after completing two layers of semantic removal and obtaining the third paragraph set, the total number of paragraphs in the third paragraph set corresponding to each bidding company is first counted. The statistical results are compared with the preset minimum paragraph number threshold. If the number of paragraphs of any bidding company is less than the threshold, such as less than 10, then all pairwise similarity comparisons involving that supplier will directly set the semantic similarity score to zero and simultaneously mark the insufficient sample. Subsequent paragraph-level similarity calculations will not be performed, and the process will directly enter the risk assessment stage.

[0060] This approach employs an insufficient sample protection mechanism, proactively invalidating similarity results when the sample size falls below statistical requirements. It avoids risk assessments based on unreliable samples, mitigating statistical bias caused by small samples and ensuring the reliability of similarity scores. This prevents suppliers from being wrongly accused of bid rigging due to overly concise bids, while also making the test results more rigorous and interpretable. This facilitates understanding of the underlying meaning by regulators and evaluation experts, allowing for alternative verification methods for suppliers with insufficient samples rather than outright exclusion, thus balancing misjudgment prevention and risk control. For suppliers with concise bids and minimal independent content, extensive matching calculations are unnecessary, allowing them to proceed directly to the next stage. This saves computational resources and improves the overall efficiency of the testing process. The pre-verification design aligns with engineering optimization logic, performing lightweight checks to eliminate invalid scenarios before performing complex calculations on valid scenarios, thus balancing accuracy and efficiency.

[0061] Combination Figure 2 Another aspect of the present invention provides a system for detecting scrambling and cross-validation based on two-layer semantic content removal and cross-validation, the system comprising: The data acquisition module is used to acquire the original text of the bidding documents and the bid documents from multiple bidding companies in the bidding project; The paragraph segmentation module is used to segment the text of the tender document into paragraphs to obtain the first set of paragraphs. The first-level semantic removal module is used to remove content in the first paragraph set with a similarity greater than or equal to a preset first threshold based on the similarity between the original text corpus and the first paragraph set, so as to obtain the second paragraph set. The second-layer semantic removal module is used to remove template paragraphs from the second paragraph set to obtain the third paragraph set. The similarity calculation module is used to calculate the similarity of the third paragraph sets corresponding to any two bidding companies and obtain a semantic similarity score. The cross-validation detection module is used to perform multi-dimensional auxiliary evidence feature cross-validation detection based on the tender documents and obtain the cross-validation detection results. The cross-validation detection includes the detection of identical typos, grammatical habit features, and non-qualification image features. The risk assessment module is used to determine whether there is a risk of collusion among multiple bidding companies based on semantic similarity scores and cross-validation detection results, and outputs the risk assessment results.

[0062] In one specific implementation, during the detection of bid-rigging in a power equipment charging and discharging module procurement project, the detection targets were six suppliers, labeled A to F, with a total of 2506 pages of commercial and technical documents. All six suppliers' bids were in PDF format, with a total size of approximately 524 megabytes. Key parameters included: a first-level elimination threshold of 0.8, a minimum paragraph length of 50 characters, a maximum paragraph length of 500 characters, a high similarity threshold of 0.6, a high-risk similarity threshold of 0.8, a high-risk paragraph count threshold of 5, and a minimum paragraph count threshold of 10.

[0063] The specific process in the implementation method is as follows: Construct a corpus of original texts of procurement documents, including Word format procurement documents, technical specification PDFs, etc., and extract them to build the original text corpus.

[0064] Paragraph segmentation results: Supplier A: 22 original paragraphs; Supplier B: 15 original paragraphs; Supplier C: 42 original paragraphs; Supplier D: 20 original paragraphs; Supplier E: 178 original paragraphs; Supplier F: 63 original paragraphs. Technical document segmentation results: Supplier A: 96 paragraphs; Supplier B: 84 paragraphs; Supplier C: 182 paragraphs; Supplier D: 42 paragraphs; Supplier E: 251 paragraphs; Supplier F: 198 paragraphs.

[0065] The results of the first-level elimination: After the first-level elimination, the number of paragraphs in the commercial documents of the six suppliers were 14, 10, 36, 14, 168, and 47, respectively. The number of paragraphs in the technical documents after the first-level elimination were 35, 43, 86, 19, 202, and 144, respectively. The first-level elimination mainly removed paragraphs copied from the original procurement documents.

[0066] The results of the second-level elimination: After the second-level elimination, the number of paragraphs in the commercial documents of the six suppliers were 14, 10, 36, 14, 168, and 47, respectively. The number of paragraphs in the technical documents after the second-level elimination were 35, 43, 86, 19, 202, and 144, respectively. The second-level elimination mainly removed the technical specification response form paragraphs and commitment template paragraphs. Taking Supplier A's technical document as an example, the original document had 96 paragraphs; after the second-level elimination, 30 paragraphs were removed, leaving 35 paragraphs of self-written content. Supplier C's technical document originally had 182 paragraphs; after the second-level elimination, 96 paragraphs were removed, leaving 86 paragraphs.

[0067] Statistics on the effectiveness of the two-tiered elimination process: the average elimination rate for technical documents was approximately 63%, and the average elimination rate for commercial documents was approximately 36%. The eliminated content mainly consisted of copied sections from the original procurement documents, sections from the technical specification response forms, and sections from the standardized commitment templates.

[0068] Sample check: After two rounds of elimination, all six suppliers had more than 10 paragraphs in their self-written content, thus passing the sample check.

[0069] Similarity calculation results: Pair A and B have a highest similarity of 0.909, with 3 pairs of highly similar paragraphs and an average similarity of 0.9; Pair A and C have a highest similarity of 0.917, with 3 pairs of highly similar paragraphs and an average similarity of 0.85; Pair A and D have a highest similarity of 0.816, with 1 pair of highly similar paragraphs; Pair A and E have a highest similarity of 0.672, with 1 pair of highly similar paragraphs; Pair A and F have a highest similarity of 0.697, with a pair of highly similar paragraphs; Pair B and C have a highest similarity of 0.86, with 3 pairs of highly similar paragraphs; Pair B and D have a highest similarity of 0.878, with 1 pair of highly similar paragraphs. Yes; B and E have a highest similarity of 0.686, with 1 pair of highly similar paragraphs; B and F have a highest similarity of 0.76, with 1 pair of highly similar paragraphs; C and D have a highest similarity of 0.811, with 2 pairs of highly similar paragraphs; C and E have a highest similarity of 0.866, with 1 pair of highly similar paragraphs; C and F have a highest similarity of 0.781, with 3 pairs of highly similar paragraphs; D and E have a highest similarity of 0.586, with no highly similar paragraphs; D and F have a highest similarity of 0.813, with 1 pair of highly similar paragraphs; E and F have a highest similarity of 0.634, with 1 pair of highly similar paragraphs.

[0070] Similarity calculation results: Pair A and B have a highest similarity of 1.0, 14 pairs of highly similar paragraphs, and an average similarity of 0.897; Pair A and C have a highest similarity of 1.0, 15 pairs of highly similar paragraphs, and an average similarity of 0.728; Pair A and D have a highest similarity of 0.459, with no highly similar paragraphs; Pair A and E have a highest similarity of 0.442, with no highly similar paragraphs; Pair A and F have a highest similarity of 0.449, with no highly similar paragraphs; Pair B and C have a highest similarity of 1.0, 12 pairs of highly similar paragraphs, and an average similarity of 0.755; Pair B and D have a highest similarity of 0.70. 5. One pair of highly similar paragraphs; B and E have a maximum similarity of 0.43 and no highly similar paragraphs; B and F have a maximum similarity of 0.368 and no highly similar paragraphs; C and D have a maximum similarity of 0.655 and one pair of highly similar paragraphs; C and E have a maximum similarity of 0.401 and no highly similar paragraphs; C and F have a maximum similarity of 0.582 and no highly similar paragraphs; D and E have a maximum similarity of 0.616 and one pair of highly similar paragraphs; D and F have a maximum similarity of 0.398 and no highly similar paragraphs; E and F have a maximum similarity of 0.518 and no highly similar paragraphs.

[0071] Examples of highly similar paragraphs: In technical documents A and B, 14 pairs of highly similar paragraphs include equipment model descriptions, such as the DC shore power interface box AZ1 independent type (similarity 1.0); equipment list headers (similarity 0.955); warranty clauses (similarity 0.973); and GB standard lists (similarity 0.935), etc. After manual verification, these highly similar paragraphs were all written by the supplier and are not from the procurement document template.

[0072] Feature detection of auxiliary evidence: Item 1: Detection of identical typos: Typo detection is performed on the content independently prepared by six suppliers, and no identical typos across different suppliers are found.

[0073] Item 2: Detection of grammatical habit features: The usage frequency ratio of "de / pronoun 'of' / structural particle 'de'" among the six suppliers is counted, which is as follows: for supplier A is 178:71:3, supplier B is 198:66:3, supplier C is 415:93:20, supplier D is 186:11:2, supplier E is 369:641:11, and supplier F is 1332:123:49. The frequency ratios of various suppliers differ greatly, and no grammatical habit matching with a cosine similarity not less than 0.85 is found. Among them, the usage frequency of the structural particle "de" (indicating location) of supplier E is abnormally high, which is significantly different from the other five suppliers.

[0074] Item 3: Feature detection of unqualified images: A total of 6931 images are extracted from six suppliers, among which 704 are meaningful images. No exactly MD5 duplicated images across suppliers, no perceptually hash similar images, and no traces of Internet URL sources are found.

[0075] Risk judgment: Pair of supplier A and supplier B: the maximum similarity of technical documents is 1.0, which is greater than 0.8, and the number of highly similar paragraph pairs is 14, which is greater than 5, so it is determined as high semantic risk. There is no feature matching of auxiliary evidence, so it is determined as high risk of collusive bidding.

[0076] Pair of supplier A and supplier C: the maximum similarity of technical documents is 1.0, which is greater than 0.8, and the number of highly similar paragraph pairs is 15, which is greater than 5, so it is determined as high semantic risk. There is no feature matching of auxiliary evidence, so it is determined as high risk of collusive bidding.

[0077] Pair of supplier B and supplier C: the maximum similarity of technical documents is 1.0, which is greater than 0.8, and the number of highly similar paragraph pairs is 12, which is greater than 5, so it is determined as high semantic risk. There is no feature matching of auxiliary evidence, so it is determined as high risk of collusive bidding.

[0078] Pairs of other suppliers: the maximum similarity is less than 0.8 or the number of highly similar paragraph pairs is less than 5, so it is determined as low risk of collusive bidding.

[0079] In another specific implementation, in the collusive bidding detection of a cable procurement project, the detection object is the cable procurement project involving five suppliers, among which only two suppliers have analyzable texts, and the total number of commercial and technical documents is 568.

[0080] Implementation process: Paragraph segmentation result: supplier A has 8 commercial paragraphs and 34 technical paragraphs; supplier B has 107 commercial paragraphs and 469 technical paragraphs.

[0081] The number of independent content paragraphs after two-layer elimination: supplier A has 6 commercial paragraphs and 28 technical paragraphs; supplier B has 89 commercial paragraphs and 385 technical paragraphs.

[0082] Sample inspection: After removing 6 segments from Supplier A's business documents, which is less than the threshold of 10 segments, the sample size is insufficient, and the similarity score for the business documents is set to zero. The technical documents have 28 segments, which is more than 10 segments, and therefore pass the inspection.

[0083] Similarity calculation results: The highest similarity between pairs A and B is 0.828, there are 6 pairs of highly similar paragraphs, and the average similarity is 0.683.

[0084] Supporting evidence feature detection: No identical typos were found, no grammatical conventions were found to match, and no duplicate images were found.

[0085] Risk Assessment: The highest similarity score between technical documents A and B is 0.828, which is greater than 0.8. Six pairs of highly similar paragraphs have a similarity score greater than 5, indicating a high semantic risk. Manual verification revealed that all six pairs of highly similar paragraphs are direct references to technical specifications, GB standard clauses, and terminology definitions. These are template content from the procurement documents, not content written by the supplier. A two-layer elimination strategy has identified and removed these paragraphs as template content. However, a small number of technical specification references were not completely eliminated. Combined with the results of supplementary evidence detection, the final assessment indicates a low risk of bid-rigging.

[0086] Result: Suppliers A and B were determined not to have engaged in bid rigging.

[0087] This embodiment verifies the effectiveness of the two-layer elimination strategy in distinguishing between technical specification citations and self-written content. Existing full-text similarity methods classify the full-text similarity between A and B as 0.87, which is considered high risk and a false positive. The detection method of this invention eliminates false positives through two-layer elimination.

[0088] In another specific implementation, based on a cable procurement project in the previous implementation, during the insufficient sample protection verification, supplier A's business documents were almost entirely composed of procurement document template content.

[0089] Implementation process: After two layers of elimination, Supplier A's business documents consisted of only 6 paragraphs, almost all of which were template content. After elimination, only 6 paragraphs of personalized content remained.

[0090] Supplier A's business document has 6 paragraphs, which is less than 10, indicating insufficient sample size. Therefore, the similarity score for all documents involving Supplier A's business document is set to 0.

[0091] Result: This example demonstrates the effectiveness of the insufficient sample protection mechanism by avoiding misjudgment caused by a small number of paragraphs in Supplier A's business documents.

[0092] It should be understood that the specific embodiments described above are merely illustrative or explanatory of the principles of the invention and do not constitute a limitation thereof. Therefore, any modifications, equivalent substitutions, improvements, etc., made without departing from the spirit and scope of the invention should be included within the protection scope of the invention. Furthermore, the appended claims are intended to cover all variations and modifications falling within the scope and boundaries of the appended claims, or equivalent forms of such scope and boundaries.

Claims

1. A method for detecting tag-rigging based on two-layer semantic content removal and cross-validation, characterized in that, The method includes: Obtain the original text of the bidding documents for the bidding project and the bid documents of multiple bidding companies; Based on the text of the tender document, paragraph segmentation is performed to obtain the first paragraph set; Based on the similarity between the original text corpus and the first paragraph set, content with a similarity greater than or equal to a preset first threshold in the first paragraph set is removed to obtain the second paragraph set; Based on the second set of paragraphs, the template paragraphs in the second set of paragraphs are removed to obtain the third set of paragraphs; The similarity score is obtained by calculating the similarity of the third paragraph sets corresponding to any two bidding companies. Based on the tender documents, multi-dimensional auxiliary evidence feature cross-validation detection is performed to obtain cross-validation detection results. The cross-validation detection includes the detection of identical typos, grammatical habit feature detection, and non-qualification image feature detection. Based on the semantic similarity score and cross-validation detection results, it is determined whether there is a risk of bid rigging among the multiple bidding companies, and the risk assessment result is output.

2. The method according to claim 1, characterized in that, The acquisition of the original text corpus of the bidding documents in the bidding project includes: Extract a first candidate paragraph with a length greater than a preset number of characters from the first format document of the tender document; Sentences are segmented from the second format document of the tender document according to punctuation marks or line breaks, and second candidate paragraphs with a length within a preset character range are selected. The original text corpus is constructed by deduplicating the first candidate paragraph and the second candidate paragraph.

3. The method according to claim 1, characterized in that, The text of the tender document is segmented into paragraphs to obtain a first set of paragraphs, including: Divide the text into initial segments based on line breaks and punctuation marks; The merged segment is obtained by merging adjacent initial segments based on the initial segment, and the paragraph length of the merged segment is between a preset minimum paragraph length and a preset maximum paragraph length; Discard any segments in the merged fragments whose length is less than the minimum paragraph length to obtain the first paragraph set.

4. The method according to claim 1, characterized in that, Based on the similarity between the original text corpus and the first paragraph set, content with a similarity greater than or equal to a preset first threshold in the first paragraph set is removed to obtain a second paragraph set, including: Calculate the similarity between each paragraph in the first paragraph set and each corpus entry in the original text corpus; When the similarity between any piece of corpus and the current paragraph is greater than or equal to the preset first threshold, the current paragraph is determined to be a copy of the original text of the tender document. The copied text is then removed to obtain the second paragraph set.

5. The method according to claim 1, characterized in that, The process involves removing template paragraphs from the second paragraph set to obtain a third paragraph set, which includes: Identify template paragraphs in the second set of paragraphs, the template paragraphs including technical specification table paragraphs, response table paragraphs, and standardization commitment template paragraphs; The template paragraphs in the second paragraph set are removed to obtain the third paragraph set.

6. The method according to claim 5, characterized in that, The process of removing template paragraphs from the second paragraph set also includes: A preset set of template noise words, which includes at least one of the following: power of attorney, legal representative, authorized agent, authorized representative, no sub-authorization allowed, and photocopy of ID card; When a paragraph matches a term in the template noise word set, it is determined to be template noise and the current paragraph is removed.

7. The method according to claim 1, characterized in that, The similarity calculation for the third paragraph sets corresponding to any two bidding companies is used to obtain a semantic similarity score, including: For each paragraph in the third paragraph set of the first bidding company, traverse all paragraphs in the third paragraph set of the second bidding company and calculate the sequence matching similarity. Record the best matching similarity and best matching paragraph for each paragraph of the first bidding company; Collect highly similar paragraph pairs whose best matching similarity is not lower than a preset second threshold; The semantic similarity score is calculated based on the highly similar paragraphs; The semantic similarity score includes the highest paragraph similarity, the number of highly similar paragraph pairs, and the average similarity of highly similar paragraph pairs.

8. The method according to claim 7, characterized in that, Based on the semantic similarity score and cross-validation detection results, the system determines whether multiple bidding companies pose a risk of bid rigging and outputs a risk assessment result, including: When the highest paragraph similarity is greater than or equal to a preset third threshold and the number of highly similar paragraph pairs is greater than or equal to a preset fourth threshold, it is determined to be semantically high risk; When the highest paragraph similarity is greater than or equal to the third threshold and the number of highly similar paragraph pairs is less than the fourth threshold, it is determined to be semantically risky; When the highest paragraph similarity is less than the third threshold, it is determined to be semantically low risk; When the semantic risk is high and at least one auxiliary evidence feature matches, it is determined to be of extremely high risk of collusion. When the semantics are high-risk and there is no supporting evidence to match, or when the semantics are medium-risk and there are at least two supporting evidence features to match, it is determined to be a high-risk case of collusion. When the semantics are of medium risk and there is a matching piece of auxiliary evidence, or when the semantics are of low risk and there is a matching piece of the same misspelling feature, it is determined to be of medium risk in the collusion.

9. The method according to claim 1, characterized in that, The method further includes: Count the number of paragraphs in the third paragraph set for each bidding company; When the number of paragraphs for any of the bidding companies is less than the preset minimum paragraph count threshold, the semantic similarity score involving that bidding company will be set to zero, and the sample will be marked as insufficient.

10. A system for detecting cross-label collusion based on two-layer semantic content removal and cross-validation, characterized in that, The system includes: The data acquisition module is used to acquire the original text of the bidding documents and the bid documents from multiple bidding companies in the bidding project; The paragraph segmentation module is used to segment the text of the tender document into paragraphs to obtain a first set of paragraphs. The first-layer semantic removal module is used to remove content in the first paragraph set whose similarity is greater than or equal to a preset first threshold based on the similarity between the original text corpus and the first paragraph set, so as to obtain the second paragraph set. The second-layer semantic elimination module is used to eliminate template paragraphs from the second paragraph set to obtain the third paragraph set. The similarity calculation module is used to calculate the similarity of the third paragraph sets corresponding to any two bidding companies to obtain a semantic similarity score; The cross-validation detection module is used to perform multi-dimensional auxiliary evidence feature cross-validation detection based on the tender documents to obtain cross-validation detection results. The cross-validation detection includes the detection of identical typos, the detection of grammatical habit features, and the detection of non-qualification image features. The risk assessment module is used to determine whether there is a risk of bid rigging among multiple bidding companies based on the semantic similarity score and cross-validation detection results, and output the risk assessment results.

Images