Feature fusion and black and white list-based bidding document risk intelligent early warning method

Abstract

The application provides a method for risk intelligent early warning of recruitment documents based on feature fusion and black and white lists, comprising: step 1, multi-modal analysis; step 2, marking high-similarity candidate file pairs; step 3, calculating file-level image similarity; step 4, fusing comprehensive similarity; step 5, marking risk levels; step 6, correcting risk levels; and step 7, generating a structured risk report. Through a text fingerprint extraction mechanism based on Burrows-Wheeler transformation and Nilsimsa hash fusion, an image feature extraction mechanism based on random projection and Hellinger distance fusion, and a cross-modal risk comprehensive evaluation mechanism based on multi-modal feature collaborative optimization, the application solves the problems of excessive calculation amount in pairwise comparison of large-scale document sets and multi-modal feature fusion and accurate risk identification.

Description

Technical Field

[0001] This invention relates to the field of computer data processing and intelligent analysis technology, and in particular to a method for intelligent early warning of risks in procurement documents based on feature fusion and blacklists / whitelists. Background Technology

[0002] In bidding activities, illegal and irregular activities such as bid rigging and collusion undermine the principle of fair market competition and harm the legitimate rights and interests of the bidding parties and the public interest. Traditional bid review methods rely heavily on manual, item-by-item comparison, which is inefficient and prone to omissions when dealing with a large number of bid documents, making it difficult to meet the dual requirements of timeliness and accuracy in modern bidding processes.

[0003] With the development of information technology, some electronic plagiarism detection tools have begun to be used in the bidding and tendering field. However, existing technologies still have the following shortcomings.

[0004] Existing plagiarism detection tools are mostly limited to simple text comparison, and their text processing methods are relatively rudimentary. Some solutions only use locality-sensitive hashing algorithms such as Simhash to directly calculate text similarity, lacking a fast indexing and candidate set filtering mechanism for large-scale document sets. When the number of tender documents is large, the computational cost increases exponentially. Although some solutions use the Burrows-Wheeler transform to process text, they only use the Burrows-Wheeler transform for text classification and authentication, without addressing fast candidate set filtering in large-scale document sets.

[0005] Existing technologies cannot effectively handle multimodal information such as images and tables contained in scanned portable document format files. The content of scanned portable document format files exists in image form, and traditional text comparison tools cannot directly extract the text information, let alone compare the similarity of the image content itself. While some solutions involve random projection and image-aware hashing, they use Hamming distance as a similarity metric. Hamming distance is sensitive to the numerical distribution of feature vectors and does not satisfy the triangle inequality property, limiting its matching accuracy under complex image transformation conditions.

[0006] Current technologies lack the ability to identify deep-seated bid-rigging and collusion. Bid-rigging and collusion manifest in various forms, including not only identical text content but also deliberate use of identical typos, highly similar seals, and abnormal file metadata—all of which are covert characteristics. Existing solutions primarily focus on text-level similarity detection, failing to comprehensively assess risk characteristics across multiple dimensions.

[0007] Existing technologies lack effective filtering for templated content. Bidding documents contain a large amount of standardized templated content such as legal provisions and industry standards. It is normal for this content to be repeated between different bids, but existing technologies cannot effectively distinguish between templated repetition and substantive abnormal repetition, which can easily lead to false alarms.

[0008] Existing technologies struggle to form a complete chain of risk evidence. Traditional solutions typically only output a similarity score or risk level, lacking precise location of duplicate content, detailed comparison, and evidence archiving capabilities, thus failing to provide sufficient traceability for subsequent review and handling. Summary of the Invention

[0009] This invention provides an intelligent risk warning method for procurement documents based on feature fusion and blacklists / whitelists. It solves the problems of excessive computational cost for pairwise comparisons in large-scale document sets and the problems of multimodal feature fusion and accurate risk identification by using a text fingerprint extraction mechanism based on Burrows-Wheeler transform and Nilsimsa hash fusion, an image feature extraction mechanism based on random projection and Hellinger distance fusion, and a cross-modal risk comprehensive assessment mechanism based on multimodal feature collaborative optimization.

[0010] To achieve the above objectives, the present invention adopts the following technical solution: A method for intelligent early warning of risks in procurement documents based on feature fusion and blacklists / whitelists includes: Step 1: Perform multimodal parsing on the tender documents and bid documents to extract text content, image content and table content. Clean the text content to form a structured text content tree and perform structured transformation on the table content to form structured table data. Step 2: Process the structured text content tree to extract a set of effective words and key indicator words, generate a fusion fingerprint vector based on the set of effective words, filter out a set of candidate similar tender documents based on the fusion fingerprint vector, and mark high similarity candidate document pairs according to the distance metric between the fusion fingerprint vectors; Step 3: Perform visual feature extraction and dimensionality reduction on the image content to obtain visual feature vectors; calculate the probability distribution distance metric between the visual feature vectors of the images in the high similarity candidate file pair and convert it into an image similarity score; calculate the file-level image similarity based on the image similarity score. Step 4: Calculate sentence-level text similarity for the text content in the high-similarity candidate file pairs and adjust the file-level text similarity using the key indicator words. Compare the structured table data to calculate file-level table similarity. Then, through multimodal feature collaborative optimization, fuse the file-level text similarity, the file-level image similarity, and the file-level table similarity into a comprehensive similarity. Step 5: Mark the file pairs whose comprehensive similarity exceeds the first threshold as highly similar file pairs, perform deep analysis on the highly similar file pairs, calculate the comprehensive risk score of bid rigging and collusion in the bid documents based on the deep analysis results, and mark the risk level. Step 6: Perform blacklist word search on the tender documents marked with risk level to calculate blacklist penalty value, perform whitelist content matching to remove similarity contribution from whitelist content, recalculate the comprehensive risk score based on the refined similarity after removal and superimpose it with the blacklist penalty value to obtain the final risk score, and correct the risk level according to the final risk score. Step 7: Organize all processing results into a report data model, and generate a structured risk report based on the report data model.

[0011] In summary, the present invention has at least the following beneficial effects: This invention employs a text fingerprint extraction mechanism that fuses the Burrows-Wheeler transform and Nilsimsa hashing. It combines the run-length feature of the Burrows-Wheeler transform with the locality-sensitive property of Nilsimsa hashing to construct a 192-dimensional fused fingerprint vector. The Burrows-Wheeler transform highlights the structural features of the text by grouping similar words together, while Nilsimsa hashing generates similar hash summaries for similar texts through multiple hash mappings of word sequences. The fusion of these two methods enhances the ability to characterize text structural features while maintaining locality sensitivity. Furthermore, random projection and position-sensitive hashing indexing based on the fused fingerprint vector transform the selection of candidate file pairs from a full pairwise comparison to a frequency band query based on the projected signature vector, significantly reducing the time complexity of text similarity calculation. This enables the invention to efficiently complete deduplication tasks in large-scale sets of tender documents.

[0012] This invention employs an image feature extraction mechanism that fuses random projection and Hellinger distance. The original 416-dimensional visual features are reduced to 64 dimensions through random projection before similarity measurement using Hellinger distance. Random projection preserves the distance structure in the original feature space during dimensionality reduction, ensuring that the reduced feature vectors effectively reflect the similarity relationships between the original images. Hellinger distance, as a similarity measure between probability distributions, is naturally suited to feature vector normalization and satisfies the triangle inequality property, which is beneficial for accelerating nearest neighbor search in image databases. For scanned images, scale-invariant feature transformation features are further fused, enabling this invention to maintain high matching accuracy even under the common perspective distortion and resolution variations in scanned images. This effectively solves the technical challenge of comparing image content in scanned portable document formats.

[0013] This invention employs a multimodal feature collaborative optimization mechanism, using an iterative approach to collaboratively optimize the weights of text, image, and table modalities. The deviation between the average similarity of each modality and the overall average similarity drives the adaptive adjustment of weights, ensuring that modalities contributing more in each iteration receive higher weights, and modalities contributing less receive lower weights. This collaborative optimization mechanism does not rely on preset fixed weights but dynamically adjusts the weight allocation based on the actual performance of each modality in a specific plagiarism detection task. It achieves optimal overall similarity calculation results across different types of file sets, realizing adaptive fusion of cross-modal features.

[0014] This invention employs a multi-level risk indicator system and a blacklist / whitelist collaborative filtering mechanism. It integrates five risk indicators—excessive similarity, repeated long texts, common error features, seal similarity, and metadata anomalies—with preset weights to form a risk score. The "common error features," as an independent risk indicator, effectively identifies covert collusion tactics where bidders deliberately create identical typos to circumvent conventional plagiarism checks. This multi-indicator weighted fusion approach eliminates reliance on single-dimensional threshold judgments for risk identification, instead relying on a comprehensive assessment of multiple evidentiary dimensions, significantly improving the accuracy of identifying covert bid-rigging behavior. Simultaneously, a blacklist penalty mechanism and a whitelist filtering mechanism are introduced for bidirectional correction. The blacklist penalizes bids containing sensitive or prohibited words, while the whitelist filters out templated content such as legal provisions and industry standards. This synergistic effect ensures the risk score more accurately reflects the substantive risk level of the bids, effectively reducing false positives.

[0015] This invention automatically generates structured risk reports and archives evidence chains, presenting the plagiarism detection results in five levels: project overview, document comparison, risk analysis, detailed evidence, and blacklist / whitelist filtering. Each risk indicator that triggers an alert is accompanied by a complete evidence chain, including precise location of duplicate content, similarity value, screenshot of the original content, and screenshot of the comparison result. This provides sufficient traceable evidence for subsequent review and handling, solving the problem that existing technologies struggle to form a complete risk evidence chain. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of an intelligent early warning method for risks in procurement documents based on feature fusion and blacklists / whitelists.

[0017] Figure 2 This is a schematic diagram of the text fingerprint extraction and fast candidate set selection process.

[0018] Figure 3 This is a schematic diagram of the image feature extraction and similarity calculation process.

[0019] Figure 4 This is a schematic diagram of the multimodal feature collaborative optimization and risk warning process. Detailed Implementation

[0020] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0021] like Figure 1 As shown in the figure, this embodiment provides a method for intelligent early warning of risks in procurement documents based on feature fusion and blacklists / whitelists, including the following steps.

[0022] Step S1: Multimodal parsing and structured feature extraction of bidding documents. The system receives the set of bidding documents uploaded by users through the front-end interface. and the collection of tender documents Collection of tender documents Include A collection of tender documents and bid documents Include The tender documents, of which and All values ​​are positive integers greater than or equal to 1. Each file is a Portable Document Format (Portable Document Format) file conforming to the International Organization for Standardization (ISO) 32000-2 standard, i.e., Portable Document Format 2.0. The system processes each input file... Perform format verification. Format verification includes checking whether the file is encrypted and whether it is corrupted. If the file is encrypted, the system will display a prompt window on the front-end interface and request the user to provide the decryption key. If the file is corrupted, the system will record the abnormal status of the file and terminate further processing of that file. Files that pass format verification are sequentially sent to the parsing queue.

[0023] The system sorts each file according to its metadata and content stream structure. The file is classified as either an electronic portable document format file or a scanned portable document format file. The classification method is as follows: The system extracts all text objects from the file content stream and calculates the total number of text characters. Calculate the total page area of ​​the file. The unit is square centimeters. Calculate the text content density. ,in For text content density, The total number of text characters. This represents the total page area of ​​the file. If... Greater than the threshold It is then determined to be an electronic file, where the threshold is... The value is 0.5 characters per square centimeter. The unit is per square centimeter. If Less than or equal to If the ratio of the number of image objects to the total number of pages is greater than 0.8, the file is considered a scanned document. If none of the above conditions are met, it is considered an electronic document by default. For electronic documents, the system directly extracts text objects, image objects, and table objects from the file content stream. For scanned documents, the system uses an optical character recognition engine to perform character recognition on the images of each page and then extracts the text content.

[0024] The system parses each file. extract the page content stream and a collection of text fragments. .in Indicates the first A text fragment, This represents the total number of text segments in the file. Each text segment... It contains three attributes: text string content The coordinates of the text fragment's position on the page. and the page number to which the text fragment belongs. Location coordinates Using the top left corner of the page as the origin, and the unit as pixels, These are the horizontal coordinate values. These are the coordinate values ​​in the vertical direction.

[0025] During text extraction, the system simultaneously performs four cleaning operations. The first cleaning operation removes isolated carriage returns and redundant spaces, compressing consecutive whitespace characters into single whitespace characters, including spaces, tabs, and newlines. The second cleaning operation detects and removes abnormal characters that do not belong to the UTF-8 character set in the unified character encoding standard. The third cleaning operation merges visually belonging-to-the-line but formatted-to-multiple text fragments into a complete line of text based on coordinate position information; the merging condition is the vertical coordinate value of each text fragment. The absolute value of the difference is less than 5 pixels and the horizontal coordinate values ​​are continuous. The fourth cleaning operation, based on the vertical coordinate values ​​of the line text clustering and the continuity of the horizontal coordinate values, merges consecutive lines belonging to the same paragraph into paragraph text. The clustering distance threshold is set to 1.5 times the average line height on the page. After cleaning, the system sorts all text fragments of each document in ascending order by page number, and within the same page, in ascending order by vertical coordinate values, forming a structured text content tree. .

[0026] The system extracts data from each file. Extract image object collection ,in Indicates the first An image object, This represents the total number of image objects in the file. During the extraction process, the system extracts each image object... Read the following information: bitmap data, position coordinates on the page. Width dimensions and height dimensions The system detects the color space type of each image, supporting six color space formats: red-green-blue, red-green-blue alpha, blue-green-red, blue-green-red alpha, grayscale, and CIELAB. CIELAB is a color space standard defined by the International Commission on Illumination (ICI), where L represents the luminance channel, and A and B represent two contrasting color channels. The system discards invalid images based on the following conditions: if the standard deviation of the image's pixel values ​​is less than a threshold of 2.0, it is considered a solid color image; if the image uses indexed colors, it is considered an indexed color image; if all alpha channels of the image are transparent (i.e., all pixels have an alpha channel value of 0), it is considered a fully transparent image. All three types of invalid images are skipped and do not proceed to further processing. For valid images, the system records the image object, its location information, and its size information.

[0027] The system uses a computer vision-based table detection algorithm to detect each document. The algorithm identifies table regions and extracts the coordinate data of table cells. First, it uses Hough transform to detect horizontal and vertical lines on the page. The polar coordinate angle resolution of the Hough transform is 1 degree, the accumulator threshold is 100 pixels, and the minimum line segment length is 50 pixels. Then, it constructs a grid structure of candidate table regions based on the detected lines. For each detected table... The system then performs the following operations: Determines the number of rows in the table using line detection. Number of columns ,in For table indexing. Extract the text content of each cell from the corresponding position based on cell coordinates. and image content ,in For row index and , For column index and Convert the extracted tabular data into a structured format. , where each row This contains the contents of all cells in the row, each cell... Includes text content and image content The system only processes tables with complete borders. A table is considered to have complete borders when the number of detected horizontal lines is greater than or equal to the expected number of rows plus one, and the number of detected vertical lines is greater than or equal to the expected number of columns plus one. For tables where complete borders cannot be detected, the system records the table's location information but does not perform structured extraction.

[0028] The system is based on a text content tree. Identify and tag headers, footers, and table of contents. The header identification rule is: the text fragment is located in the top area of ​​the page and appears repeatedly across the page; the top area refers to the vertical coordinate value of the fragment. Less than 15% of the page height; repeated across pages refers to the same text appearing on three or more consecutive pages. The footer is identified as a text fragment located at the bottom of the page and repeating across pages; the bottom area refers to the fragment's vertical coordinate value. The text content must exceed 85% of the page height and appear on three or more consecutive pages. The table of contents is identified as follows: text blocks containing the keywords "table of contents," "contents," or "table of contents" and containing consecutive numerical page numbers. Text content marked as headers, footers, or tables of contents is excluded in all subsequent deduplication steps. Step S1 outputs the following structured data: the cleaned text content tree for each document. and its text fragment collection The collection of image objects for each document and its location information Structured tabular data for each document Metadata for each document, including document type tag (electronic or scanned) and total number of pages. .

[0029] Step S2 involves extracting text content fingerprints and quickly filtering candidate sets based on the fusion of Burrows-Wheeler transform and Nilsimsa hashing. For example... Figure 2 As shown, step S2 receives the cleaned text content tree of each file output in step S1. and its text fragment collection The system first obtains the total number of all files. Assign a unique file identifier to each file ,in .

[0030] The system processes each document After text cleaning, the text content is segmented into sentences based on sentence boundaries. Sentence boundary recognition is based on period characters, question mark characters, exclamation mark characters, and newline characters. The resulting sentence set is then processed. ,in Indicates the first One sentence. , This represents the total number of sentences in the file. For each sentence... The system performs Chinese word segmentation to obtain a word sequence. ,in is the total number of words in the sentence. During the word segmentation process, the system loads a pre-set professional vocabulary, which contains more than 100,000 entries including terms in the field of bidding and tendering, industry standard vocabulary, and proprietary terms in laws and regulations. The bidirectional maximum matching algorithm is adopted to ensure the correct segmentation of professional vocabulary.

[0031] The system processes each sentence of the word sequence to perform stop word filtering. The stop word list contains about 500 common stop words, such as "of", "already", "in", "is", "I", "have", "and", "just", "not", "people", "all", "one", "a", "on", "also", "very", "to", "say", "want", "go", "you", "will", "with", "without", "look", "good", "oneself", "this", etc. After filtering, the effective word sequence is obtained, where is the number of effective words in the sentence.

[0032] The system simultaneously extracts high-weight key index words from the effective word sequence. The extraction rules include the following four matching methods. The rule for matching the unified social credit code is: match a string consisting of two uppercase letters or numeric characters, six numeric characters, and ten uppercase letters or numeric characters, where the letter part of the said uppercase letters or numeric characters is all uppercase letters excluded from the letter set I, O, S, V, Z. The rule for matching the resident identity card number is: match a string consisting of seventeen numeric characters and the last numeric character or uppercase letter X or lowercase letter x. The rule for matching the mobile phone number is: match a string starting with the numeric character 1 followed by ten numeric characters. The rule for matching the business license number is: match a string consisting of fifteen numeric characters. The set of extracted key index words includes four types: unified social credit code, resident identity card number, mobile phone number, and business license number. The set of extracted key index words is denoted as where is the number of key index words in the sentence.

[0033] The system constructs a set for all the effective words of each document , that is, the union of the effective words of all sentences of the document. The system uses the Burrows-Wheeler transform to perform transform processing on the effective word sequence of each document. For the effective word sequence of the document , the system regards the effective word sequence as a sequence composed of words where is the document The total number of valid words in the sequence. The system constructs all the sequences for this sequence. Each cyclic shift is performed, and all cyclic shifts are sorted in ascending lexicographical order. The last column of the sorted matrix is ​​taken as the Burrows-Wheeler transformation result. The system is derived from the Burrows-Wheeler transform. Extract the following feature: the longest run length of consecutive identical words. This refers to the maximum number of consecutive occurrences of the same word in the transformation result; run-length encoding compression ratio. ,in This represents the number of runs in the Burrows-Wheeler transform result. The total number of valid words; and the frequency vector of each word in the transformation result. ,in The total number of distinct words in the valid word set. For the first The frequency of different words appearing in the transformation result.

[0034] The system also uses the Nilsimsa hash algorithm for each file. The system generates a Nilsimsa hash digest from a valid word sequence. The system initializes an integer array of length 256. All elements are initialized to 0. For each word in the valid word sequence... The system calculates the SHA-256 hash value of the word, and takes the first three bytes of the SHA-256 hash value as three index values. , and ,in , and All are integers between 0 and 255. The system will store the array... In position , and The value increases by 1 for each word. The system calculates the context features of the window: It takes the word and the two words before and after it, totaling five words, to form a window. For words near the beginning or end of the file, the window only contains the existing words; that is, the window for the first word of the file consists of the first word and the four words after it, totaling five words; the window for the last word of the file consists of the last word and the four words before it, totaling five words. It calculates the XOR result of the SHA-256 hash values ​​of all words within the window, and takes the first three bytes of this XOR result as three additional index values. , and , array In position , and The value of each word is incremented by 1. After processing all words, the system iterates through the array. The 256 positions, for each position , ,like Greater than the average value of all positions in the array If the Nilsimsa hash digest is true, the corresponding bit is set to 1; otherwise, it is set to 0. This results in a 256-bit Nilsimsa hash digest. ,in For the first The binary value of a bit.

[0035] The system will process each document The Burrows-Wheeler transform features are fused with the Nilsimsa hash digest to construct the file. The fingerprint vector is obtained. The fusion process is as follows: The system extracts the first 64 run length values ​​and the first 64 word frequency values ​​from the Burrows-Wheeler transform result to form a 128-dimensional Burrows-Wheeler feature vector. The first 64 components to The run length values ​​are normalized values, obtained by dividing each run length value by the maximum value among all run length values; the last 64 components. to This is the normalized value of the word frequency, obtained by dividing each word frequency by the maximum value among all word frequencies. The system then generates a 256-bit Nilsimsa hash digest. Converting each group of 4 digits into a hexadecimal number yields a 64-dimensional Nilsimsa feature vector. ,in For the first The value corresponding to the four-bit binary number after conversion to hexadecimal. The system will use the Burrows-Wheeler feature vectors With Nilsimsa eigenvectors By concatenating the components, a 192-dimensional fused fingerprint vector is obtained: ;in For the file Fingerprint vector based on Burrows-Wheeler transform and Nilsimsa hash fusion to These are the 128 components of the Burrows-Wheeler eigenvector. to These are the 64 components of the Nilsimsa feature vector. The fused fingerprint vector... The first 128 components constitute the Burrows-Wheeler feature vector, where the first 64 components correspond to run length values ​​and the last 64 components correspond to word frequency values; the fused fingerprint vector The last 64 components are Nilsimsa feature vectors. The system is based on the fused fingerprint vectors of all files. Build a fast candidate set filtering index. The specific construction process is as follows.

[0036] The system uses a random projection method to transform each 192-dimensional fused fingerprint vector Projected onto a lower-dimensional space. System generation. random projection vectors ,in This represents the number of projection vectors, with a value of 32. Each random projection vector... Each component is independently and identically distributed according to the standard normal distribution. ,in For each file fused fingerprint vector The system calculates in the first... random projection vectors Projection values ​​on: ;in For the file In the Projection values ​​on random projection vectors For the first The first random projection vector of the nth random projection vector One portion, For the file fused fingerprint vector The One portion, .

[0037] The system will put all files of The projected values ​​are combined to form a projected signature vector: ; in For the file The projected signature vector, For the first One projection value, , .

[0038] The system uses a position-sensitive hashing method to index and construct the projected signature vectors. The system then assigns each projected signature vector... Divided into One frequency band, of which The value is 16. Each frequency band contains Each projection value. and Substituting into the calculation, we get For each frequency band ,in The system will cover this frequency band The projected values ​​are concatenated sequentially into a single key value: ; in For the file In the Key values ​​on each frequency band For the projected signature vector The first in One portion, .

[0039] For each tender document ,in The system will submit the tender documents. Projected signature vector Divided according to the same frequency band division method Each frequency band is calculated, and the key value for each frequency band is determined. For each frequency band The system checks if any tender documents with the same key value exist in that frequency band. If they do, the system will include all tender documents in that frequency band. Having the same key value Tender documents Add to candidate similar file set ,in After all After querying the frequency band, the system obtains the tender documents. A collection of candidate similar tender documents .

[0040] For each candidate file pair in The system calculates the bid documents. fused fingerprint vector With the tender documents fused fingerprint vector Manhattan distance between them: ;in For the tender documents fused fingerprint vector The One portion, For the tender documents fused fingerprint vector The One portion, This represents the Manhattan distance. A smaller Manhattan distance indicates greater similarity between the two files. The system sets a Manhattan distance threshold. ,in The value is 42.0. This threshold was determined by scanning the parameters of a test set containing 1000 real bidding documents, and is the value corresponding to the maximum F1 score on the recall-precision balance curve. The system will satisfy... Candidate file pairs are marked as high-similarity candidate file pairs. All high-similarity candidate file pairs constitute a set. Step S2 outputs the following data: a set of high-similarity candidate document pairs. ; Fused fingerprint vector for each document Keyword set for each document And its location information in the file, including page numbers and sentence indexes.

[0041] Step S3 involves image feature extraction and similarity calculation based on the fusion of random projection and Hellinger distance. For example... Figure 3 As shown, step S3 receives the image object set of each file output in step S1. and its location information and the set of high similarity candidate files output in step S2. .

[0042] For each document Image object collection Each image in The system first performs a validity pre-screening. The pre-screening includes two checks. The first check is a size check: if the image width... Less than 32 pixels or height Images with less than 32 pixels are considered invalid and skipped from further processing. The second check is a content entropy check: the system calculates the entropy value of the image's pixel grayscale distribution. ;in grayscale value The frequency of occurrence in the image and satisfying , Let be the content entropy of the image. If... Images deemed to have insufficient information content are skipped from further processing. For each valid image that passes the validity pre-screening, the system extracts multiple visual features and performs dimensionality reduction and fusion using a random projection method.

[0043] The system first extracts the color features of the image. It converts the image from a red-green-blue color space to a hue-saturation-lightness color space, where H represents the hue channel, S represents the saturation channel, and V represents the lightness channel. The system calculates histograms for the hue, saturation, and lightness channels separately, with each histogram containing 32 intervals. The three histograms are then concatenated to obtain a 96-dimensional color histogram feature vector. ,in The first feature vector of the color histogram One portion, 96 is the dimension of the feature vector of the color histogram.

[0044] The system extracts texture features from the image. A Local Binary Pattern (LBMM) operator is used to extract these features. For each pixel in the image, its 3-pixel multiplied neighborhood is extracted. For image boundary pixels, the system uses zero-padding to expand the image boundary, filling the surrounding area with pixels of grayscale value 0, ensuring that boundary pixels also form complete 3-pixel multiplied neighborhoods. The grayscale values ​​of the eight neighboring pixels within the neighborhood are compared with the grayscale value of the center pixel. If the grayscale value of a neighboring pixel is greater than or equal to that of the center pixel, it is marked as 1; otherwise, it is marked as 0. The eight bits are arranged clockwise to form an 8-bit binary number, which is the LBMM encoded value for that pixel. The LBMM encoded value is an integer between 0 and 255, therefore the histogram contains 256 intervals, each corresponding to a possible encoded value. The system statistically analyzes the histogram of LBMM encoded values ​​to obtain a 256-dimensional LBMM texture feature vector. ,in The first local binary pattern texture feature vector is the first... One portion, .

[0045] The system extracts edge features from the image. The Sobel operator is used to extract the edge response of the image. The Sobel operator contains horizontal convolution kernels. and vertical convolution kernel The system convolves the image with horizontally oriented convolutional kernels. Convolution is performed to obtain the horizontal gradient. Convolve the image with a vertical convolution kernel Convolution is performed to obtain the gradient in the vertical direction. Calculate the gradient magnitude The system statistically analyzes the histogram of gradient magnitudes, which contains 64 intervals, resulting in a 64-dimensional edge feature vector. ,in The first edge feature vector One portion, .

[0046] The system will use the color histogram feature vector Local binary pattern texture feature vector and edge feature vectors The concatenation results in a 416-dimensional original visual feature vector: ;in For image The original visual feature vector, to These are the 96 components of the color histogram feature vector. to These are the 256 components of the local binary pattern texture feature vector. to 416 represents the 64 components of the edge feature vector and the dimension of the original visual feature vector.

[0047] The system uses a random projection method to transform the original 416-dimensional visual feature vector Dimensionality reduced to 64. The system generates 64 random projection vectors. Each random projection vector Each component is independently and identically distributed according to the standard normal distribution. ,in For each image The system calculates the dimensionality-reduced visual feature vector: ;in For image Dimensionally reduced visual feature vectors The One portion, For the first The first random projection vector of the nth random projection vector One portion, Original visual feature vector The One portion, The reduced visual feature vector is .

[0048] For a set of highly similar candidate document pairs Any file pair in The system can process any two images in the file pair. and Calculate visual feature vectors and The Hellinger distance between the two visual feature vectors. The system first combines the two visual feature vectors... and Normalize to probability distributions respectively: ;in For image The normalized probability distribution is at the th Values ​​in each dimension For image The visual feature vector in the first Values ​​in each dimension , For summation index variables.

[0049] ;in For image The normalized probability distribution is at the th Values ​​in each dimension For image The visual feature vector in the first The values ​​in each dimension. The Hellinger distance is defined as: ;in For image and images The Hellinger distance between them. The Hellinger distance takes values ​​within a closed interval. The smaller the value, the more similar the two images are.

[0050] The system converts the Hellinger distance into an image similarity score: ;in For image and images The image similarity score between them takes values ​​within a closed interval. The larger the value, the more similar the two images are.

[0051] For the images in the file identified as scanned in step S1, the system performs additional scale-invariant feature transformation feature extraction and matching on top of the random projection and Hellinger distance fusion comparison. The system then processes the two images... and Extract the set of key points for scale-invariant feature transformation respectively and ,in For image The total number of key points of scale-invariant feature transformation. For image The total number of scale-invariant feature transformation keypoints, where each keypoint contains location coordinates. Scale parameters Direction parameters and 128-dimensional descriptor vectors The system uses a fast approximate nearest neighbor search library to match the descriptor vectors of two sets of key points. For each key point... The system is Find the nearest neighbor key point and next nearest neighbor key points If the nearest neighbor distance Distance to the next nearest neighbor The ratio satisfies If a match is found to be a match, then the match is accepted. All accepted match pairs constitute the set of match pairs. The system calculates the scale-invariant feature transformation matching score: ; in The number of matching pairs, For image The total number of key points of scale-invariant feature transformation. For image The total number of key points of scale-invariant feature transformation. The score is the matching score of the scale-invariant feature transform and its value range is a closed interval. .

[0052] For scanned images, the system fuses the image similarity score using random projection and Hellinger distance. Matching score with scale-invariant feature transform Perform secondary fusion: ; in System default settings , , This is the overall image similarity score for the scanned image.

[0053] For candidate file pairs The system constructs a bipartite graph. The left node is the tender document. Image collection The node on the right is the tender document. Image collection The edge weights are the image similarity scores of the corresponding image pairs. For scanned images Alternative The system employs the Hungarian algorithm to find the maximum weight perfect matching in this bipartite graph. When the number of nodes on both sides is unequal, the system adds virtual nodes to make the number of nodes on both sides equal, and the edge weight between the virtual node and all nodes on the opposite side is 0. The system then obtains the optimal matching set. File-level image similarity is defined as: ; in For the tender documents The total number of valid images in the middle. For the tender documents The total number of valid images in the middle. The file-level image similarity is defined as a closed interval. .

[0054] The system identifies certificates and licenses by checking if the text content of the images contains keywords such as "business license," "software copyright," "patent certificate," "trademark registration certificate," and "qualification certificate." For image pairs identified as certificates and licenses, if the image similarity score is... The system triggers optical character recognition (OCR) for re-examination. Certificate images include five categories: business license images, software copyright certificate images, patent certificate images, trademark registration certificate images, and qualification certificate images. The system performs OCR on both images separately, extracting the core certificate number text, including four number types: unified social credit code text, certificate registration number text, patent number text, and trademark registration number text. If the extracted number texts are completely identical, the system sets the similarity of the image pair to 1.0. If the number texts are inconsistent but the image similarity is still higher than 0.7, the system marks the image pair as "suspected tampering of certificates" and records an alarm. Step S3 outputs the following data: file-level image similarity of the candidate file pair. Image-image best matching pair list The similarity details; the optical character recognition re-examination results of certificate images and the list of "suspected tampering of certificates" alarm markers.

[0055] Step S4 involves cross-modal similarity comprehensive calculation and iterative refinement based on multi-modal feature collaborative optimization. For example... Figure 4 As shown, step S4 receives the set of highly similar candidate files output in step S2. and key indicator set The file-level image similarity of the candidate file pairs output in step S3 And the details of the best image matching, and the structured table data output in step S1. .

[0056] For candidate file pairs The system processes the tender documents. Each sentence With the tender documents Each sentence Calculate text similarity. The system uses a composite text similarity metric that combines edit distance and Jaro-Winkler distance.

[0057] Edit distance Define as a sentence Convert to sentences The minimum number of single-character insertion operations, single-character deletion operations, or single-character replacement operations required. Edit distance similarity is: ; in For sentences The length of the characters, For sentences The length of the characters, The edit distance similarity is defined as a closed interval. The subscript "ED" indicates the edit distance.

[0058] The system first calculates the Jaro distance. (Two strings) and The formula for calculating the Jaro distance between them is: ; in The number of characters to be matched is defined as the number of identical characters at the same position in the two strings. The size of the matching window is [value missing]. ; The number of transpositions is defined as half the number of matched but out-of-order character pairs. For sentences The length of the characters, For sentences The length of the characters. If Then the Jaro distance is 0. The Jaro-Winkler distance is the Jaro distance with an added prefix factor: ; in For sentences and The length of the common prefix, with a maximum value of 4. The Jao-Winkler distance is a closed interval. The subscript "JW" indicates the Jaro-Winkler distance.

[0059] The system will edit distance similarity Distance from Jaro-Winkler Weighted fusion into sentence-level text similarity: ;in System default settings , , This represents sentence-level text similarity.

[0060] The system pairs candidate files Comparison of tender documents Key indicator word set and tender documents Key indicator term set If at least one keyword is completely identical, the system directly sets the file-level text similarity to 1.0 for the current candidate file pair and records the keyword complete match event. Otherwise, the system calculates the file-level text similarity as follows: First, calculate the keyword improvement: ; in This is the keyword weight coefficient with a value of 0.15. The value represents the increase in text similarity and is within a closed interval. Then, by aggregating sentence-level similarity and adding the keyword boosting factor, we obtain document-level text similarity: ; in For the tender documents A collection of sentences, For the tender documents Total number of sentences The function truncates the result to the upper limit of 1.0.

[0061] For candidate file pairs The system processes the tender documents. Structured tabular data With the tender documents Structured tabular data Perform a structured comparison. For the tender documents... Each structured table data and tender documents Each structured table data The system first compares the number of rows and columns in the tables. If the two tables have the same number of rows and columns, then... and The system compares the text content cell by cell and calculates the table similarity: ; in The number of rows in the table. The number of columns in the table. For structured tabular data The Middle Line 1 The text content of the column cells, For structured tabular data The Middle Line 1 The text content of the column cells.

[0062] If the two tables have different numbers of rows or columns, the system uses a table matching method based on optimal transmission. The system will then process the structured table data. Flattened to a length of A sequence of cells that will structure tabular data. Flattened to a length of The sequence of cells. The system solves for the transmission plan matrix. This matrix minimizes the total transmission cost: ; in For the transmission plan matrix, For the transmission plan matrix The elements represent data from structured tables. The From individual cells to structured table data The The transmission probability of each cell. For structured tabular data The flattened first One cell, For structured tabular data The flattened first One cell, This represents the total transmission cost. The constraints are as follows: For all Establishment, and For all Established, among which For structured tabular data The total number of cells, For structured tabular data The total number of cells. The sum of the transport plan matrix under this constraint setting. This satisfies the normalization constraints of the transmission plan. The table similarity is taken as the complement of the optimal transmission cost: ;in This is the minimum total transmission cost among all transmission plan matrices that satisfy the constraints. (File-level table similarity) Take the maximum similarity score across all tables.

[0063] The system introduces a multimodal feature collaborative optimization mechanism, which iteratively optimizes the weights of text, image, and table modalities. The system initializes with the number of iterations. The initial weights are set to , and The initial weights mentioned above are set based on the statistical distribution of information density in bidding documents under different modalities. Statistics show that text content accounts for approximately 50% of the total content in bidding documents, image content approximately 25%, and table content approximately 25%. In the... In this iteration, the system calculates the overall similarity: ; superscript Indicates the first iteration For the first The weight coefficients of the text modality in the next iteration For the first The weighting coefficients of the image modality in the next iteration For the first The weighting coefficients of the table modality in the next iteration For the first The comprehensive similarity of the nth iteration is completed. After one iteration, the system calculates the modal average similarity of all candidate file pairs: ; in For the first The average text modal similarity of all candidate file pairs in the next iteration. For the candidate document pair set The number of elements in the summation symbol This indicates traversing the candidate file pair set. All file pairs in the.

[0064] ;in For the first Average image modal similarity of all candidate file pairs in each iteration.

[0065] ;in For the first The average similarity of the tabular modalities of all candidate document pairs in each iteration.

[0066] The system calculates the overall average similarity. ,in For the first The average similarity of the previous iteration is calculated. Then, the weights for the next iteration are updated according to the following rules: ;in The learning rate is 0.05. For the first The weight coefficients of the text modality in the next iteration.

[0067] ;in For the first The weighting coefficients of the image modality in the next iteration.

[0068] ;in For the first The weight coefficients of the table modality in this iteration. The updated weights are normalized to ensure... The convergence condition for the iteration is: ;in This is the convergence accuracy threshold, and its value is 0.01. Indicates the set of candidate files Take the maximum value from the middle.

[0069] When the convergence condition is met, the iteration stops, and the system uses the comprehensive similarity of the current iteration as the final output. If the iteration count exceeds 100 and convergence is still not achieved, the system forcibly stops the iteration and uses the current result as the output. Step S4 outputs the following data: multimodal comprehensive similarity of candidate file pairs. Independent similarity scores for each modality , , List of events where key indicators are perfectly matched; detailed sentence-level similarity data and detailed table-level similarity data.

[0070] Step S5 involves identifying and comprehensively scoring bid-rigging and collusion risks based on a multi-level risk indicator system. Step S5 receives the multimodal comprehensive similarity of candidate document pairs output from Step S4. And modal-independent similarity scores, receiving the cleaned text content tree of each file output from step S1. Image object collection and metadata.

[0071] The system sets a comprehensive similarity threshold. ,in The value is 0.65. For each candidate file pair... ,like The system marks this file pair as a highly similar file pair. If The system marks these as moderately similar file pairs. If The system marks these as low-similar file pairs.

[0072] For file pairs marked as highly similar, the system performs the following four deep analyses.

[0073] The first in-depth analysis is the detection of duplicates in continuous long texts. The system analyzes the tender documents... and tender documents The entire text is compared using a sliding window, with a window size of [size missing]. A series of consecutive characters, among which The value is 50. The system slides the window from left to right in single-character increments, meaning the window moves one character position to the right each time, and the overlap length between adjacent windows is... Each time a character sequence within the window is identical, the duplicate text segment is recorded and its length is incremented. After duplicate detection is complete, the system calculates the total character length of the duplicate text as a percentage of the total length of the tender document. The proportion of total character length: ; in The sum of the character lengths of all consecutive repeating text segments. For the tender documents The total character length of the text after cleaning and removing headers, footers, and table of contents in step S1. If... The system determined that there was a large section of text suspected of being plagiarized and recorded the incident.

[0074] The second in-depth analysis is typo consistency detection. The system performs sentence-level text similarity analysis on highly similar file pairs. The system performs sentence-by-sentence typo analysis on text segments. It uses a BERT-based-chinese pre-trained language model to process each sentence. Each sentence is taken as the input sequence, and the BERT pre-trained language model outputs a probability vector of length equal to the dictionary size for each position. After softmax normalization, each dimension of this probability vector corresponds to the probability value of the corresponding character in the dictionary. If the BERT pre-trained language model predicts a character with a probability lower than 0.3 and the edit distance between this character and the character with the highest predicted probability by the BERT pre-trained language model is 1, then this character is identified as a typo. The edit distance here applies to the individual character level, i.e., the number of insertion, deletion, or replacement operations between individual characters. If two files contain the same typo or the same type of incorrect punctuation at the same page number and sentence sequence position, the system marks this phenomenon as a "common error feature" and records the error type and its location. Error types include four types: typo, missing character, extra character, and punctuation error.

[0075] The third in-depth analysis is a specific comparison of seal similarity. The system extracts the image region dominated by the red channel from the file as a candidate seal region. The extraction method is as follows: for each pixel in the image, if the red channel value is greater than the green channel value, the red channel value is greater than the blue channel value, and the red channel value is greater than the threshold of 200, then the pixel is marked as a candidate seal pixel. The values ​​of the red, green, and blue channels are all closed intervals. The system performs connected component analysis on the detected seal candidate regions to extract the complete seal image. The neighborhood type for connected component analysis is 8-connected, and the minimum connected component area threshold is 100 pixels. For the detected seal image, the system uses the single-image similarity score defined in step S3. Calculate the similarity between the seals. If the similarity between the seal images of two documents exceeds 0.8, the system marks it as "highly similar seals" and records the event.

[0076] The fourth in-depth analysis involves comparing file attribute metadata. The system extracts the following four metadata items from the metadata stream of the Portable Document Format file: creation time field. Modify the time field Author field Last saver field If the tender documents With the tender documents If the author field information or the last saver field information is completely identical, the system records it as "metadata author anomaly". If the tender document... The time interval between creation time and modification time and the tender document If the absolute value of the time difference between the creation time and modification time is less than 10 seconds, the system records it as a "metadata time anomaly". When any of the above anomalies is detected, the system records a "metadata anomaly" event.

[0077] The system calculates each bid document based on the results of in-depth analysis. Comprehensive risk assessment of bid rigging and collusion: ;in For the first The indicator function for each risk indicator has a value of 1 indicating that the risk indicator has been triggered, and a value of 0 indicating that it has not been triggered. The corresponding risk weights and the sum of all weights .

[0078] The first risk indicator is the overall similarity exceeding the standard. If any tender document exists Make ,but ,otherwise Weight .

[0079] The second risk indicator is the continuous long text duplication indicator. If any tender document exists Make ,but ,otherwise Weight .

[0080] The third risk indicator is the common error characteristic indicator. If at least one common error feature is detected, then ,otherwise Weight .

[0081] The fourth risk indicator is the seal high similarity indicator. If any seal image has a similarity exceeding 0.8, then... ,otherwise Weight .

[0082] The fifth risk indicator is the metadata anomaly indicator. If an abnormal metadata author or metadata time is detected, then ,otherwise Weight .

[0083] The system sets risk warning thresholds. ,in The value is 0.50. If The system will submit the tender documents. Marked as high-risk and generate a risk warning for bid rigging and collusion. If The system marks it as a medium-risk level and generates a warning alert. If The system marks these as low-risk. Step S5 outputs the following data: each tender document Comprehensive risk score The corresponding risk level; in-depth analysis details of highly similar document pairs, including a list of consecutive repeating text segments, a list of common error features, a detailed list of seal comparison results, and a detailed list of metadata comparison results; and a list of specific risk indicators that trigger warnings.

[0084] Step S6 refines the plagiarism detection results and corrects the risk level based on blacklist / whitelist collaborative filtering. Step S6 receives the lists of high-risk and medium-risk bid documents and their corresponding risk details output in Step S5. The system simultaneously reads a pre-set blacklist database. and whitelist database Blacklist database Includes sensitive word database and disabled word lists There are two sub-databases. The whitelist database. Includes a library of legal and regulatory provisions. Industry Standards Database and self-built whitelist database There are three sub-libraries in total.

[0085] The system first performs a blacklist word search on the text segments marked as duplicates in step S5. The specific process of the blacklist word search is as follows: The system compares the text segments to be searched with the blacklist database. Each blacklisted word in the text segment is matched. Matching employs both exact and fuzzy matching methods. Exact matching requires the text segment to contain a continuous sequence of characters identical to the blacklisted word. Fuzzy matching uses the edit distance method, calculated in the same way as defined in step S4, which is the sum of the minimum number of single-character insertion, deletion, or substitution operations required to convert one string to another. If the edit distance between a continuous character sequence in the text segment and a blacklisted word is less than or equal to 2, it is considered a fuzzy match. The system records all matched blacklisted words and their positions in the text segment, including page numbers and sentence numbers. If in a certain tender document... The system detected blacklisted words in the tender document and removed it. The overall risk score is increased by a blacklist penalty value, which is calculated using the following formula: ;in For the tender documents The total number of blacklisted words matched in the text. For the first The weight coefficient of each matched blacklisted word is set to 1.0. The base weight coefficient for blacklisted words is set to 0.05. This value ensures that the maximum contribution of a single blacklisted word to the final risk score is 0.05, and the total contribution of 5 blacklisted words is 0.25, which is consistent with the order of magnitude of the weights of various risk indicators in step S5, thus preventing the blacklisted word detection results from excessively dominating the final risk score. This is the penalty value for being blacklisted. The system will record the increased overall risk score as follows: ;in This is an intermediate risk score after the blacklist penalty is imposed. This intermediate score is only used to record intermediate results and is not included in the calculation of the subsequent refined risk score. The risk score will be corrected in subsequent steps. It is based on the original The recalculated value does not include blacklist penalty values. The original risk comprehensive score is output from step S5.

[0086] The system then compares text segments, images, and table content marked as duplicates with the whitelist. The system performs a full match on all content within the specified range. For each text segment, the system calculates the Jaccard similarity between the text segment to be checked and each item in the whitelist. The system then segments the text segment to be checked into a word set. After segmenting the whitelist content into words, a word set is obtained. ,calculate: ; in The similarity is based on Jaccard scores. If the Jaccard similarity exceeds the whitelist matching threshold... ,in A value of 0.85 indicates that the text segment belongs to the whitelist. For image content, the system uses the single-image similarity score defined in step S3. Calculate the similarity between the image to be checked and each image in the whitelist. If The image was identified as a whitelisted image. For the table content, the system uses the table similarity score defined in step S4. Calculate the similarity between the table to be checked and each table in the whitelist. It was determined to be a whitelist table.

[0087] The system removes all content identified as belonging to the whitelist from the plagiarism check results. For the file pair The text modal similarity contribution value corresponding to all text segments identified as whitelisted is obtained by summing the sentence-level text similarity of all text segments identified as whitelisted. For the file pair The image modal similarity contribution value corresponding to all images identified as whitelisted is obtained by summing the single image similarity scores of all images identified as whitelisted and dividing by the total number of valid images; For the file pair The table modal similarity contribution value is calculated for all tables identified as whitelisted. This contribution value is obtained by summing the table similarity scores of all whitelisted tables and dividing by the total number of tables. For each file pair... The system will calculate the original comprehensive similarity. After deducting contributions from content that is on the whitelist, the refined similarity is obtained: ;in These are the text modality weight coefficients after iterative convergence in step S4. These are the image modality weight coefficients after iterative convergence in step S4. These are the table modal weight coefficients after iterative convergence in step S4. The refined similarity... The range of values ​​is truncated to a closed interval. .

[0088] The system is based on refined similarity. Recalculate the overall risk score. The revised risk score is shown below. The calculation formula is: ;in To correct the overall similarity exceeding the standard based on the refined similarity: if any tender document... Make ,but ,otherwise ; The weighting of the similarity exceeding the standard; Assign the remaining four risk indicators the weights defined in step S5. , , , Weighted summation.

[0089] The system will correct the risk score. Blacklist penalty value The scores are combined to obtain the final risk score: ;in The system calculates the final risk score. Reassess the risk level: If Marked as high-risk level; if Marked as medium risk level; if If the risk level decreases after correction (i.e., from high risk to medium risk, from high risk to low risk, or from medium risk to low risk), the system will mark "Risk level downgraded after whitelist filtering" in the risk report and record details of the filtered whitelist content, including the sub-library category, content summary, and matching similarity score. If the risk level increases after correction (i.e., from medium risk to high risk, from low risk to medium risk, or from low risk to high risk), the system will mark "Risk level upgraded after blacklist retrieval" in the risk report and record details of the matched blacklist words, including the blacklist word content, page number, and sentence. Step S6 outputs the following data: refined comprehensive similarity. Final risk score And the final risk level; whitelist matching details; blacklist word matching details.

[0090] Step S7: Automatic generation of a multi-dimensional structured risk report and archiving of the evidence chain. Step S7 receives the refined deduplication results and final risk score output from Step S6. The output of step S4 includes the modal similarity details, the output of step S5 includes the depth analysis details, the output of step S1 includes the file metadata, and the output of step S3 includes the optical character recognition re-inspection alarm markers for certificate images.

[0091] The system organizes all processing results from steps S1 to S6 into a report data model according to the following five-level hierarchical structure. The project overview layer contains the text of the bidding project name and the numerical value of the number of bidding documents. Number of tender documents The document comparison layer contains four basic pieces of information, including the timestamp of the plagiarism detection task execution. It includes each bid document. With each tender document Refined overall similarity and modal independent similarity , , The risk analysis layer includes each tender document. Final risk score The final risk level and a list of triggered risk indicators are provided. The detailed evidence layer includes the specific content and location information of duplicate text segments, thumbnails and similarity values ​​of similar image pairs, preview content and similarity values ​​of similar table pairs, a list of common error features, a list of seal comparison results, and a list of metadata comparison results. The blacklist / whitelist filtering layer includes the filtered whitelist content and its corresponding sub-database category and matching similarity value, as well as the matched blacklist words and their locations. For low-risk bid documents that do not proceed to step S6, the corresponding filtering details in the blacklist / whitelist filtering layer are empty.

[0092] The system automatically generates a structured risk report based on a report data model. The risk report consists of six parts: an execution summary, a comprehensive similarity matrix, a risk ranking list, a detailed list of duplicate content, a risk evidence chain, and a blacklist / whitelist filtering detail. The execution summary provides a textual overview of the overall results of the plagiarism detection task, including the total number of duplicate files detected. The document provides four summary pieces of information: the number of high-risk documents, the number of medium-risk documents, and the distribution of major risk types. The comprehensive similarity matrix displays the refined overall similarity between all tender documents and all bidding documents in matrix form. The similarity is visually represented using a color gradient from green to red, with green indicating a similarity below 0.3, yellow indicating a similarity between 0.3 and 0.6, and red indicating a similarity above 0.6. The risk ranking list is based on the final risk score. All tender documents are arranged from highest to lowest risk, with each document's risk level and key risk indicator names clearly marked. The duplicate content details section displays all duplicate content for each high-risk document pair in list format, including the original text, page numbers, and coordinates of the duplicate text segments; thumbnails and similarity values ​​for similar image pairs; and previews and similarity values ​​for similar table pairs. The risk evidence chain section provides a complete evidence chain for each risk indicator that triggers a warning, including three pieces of evidence: screenshots of the original content, screenshots of the comparison results, and similarity values ​​output by the algorithm. The blacklist / whitelist filtering details section lists the details of content filtered by the whitelist and content matched by the blacklist, including a content summary, matching type, similarity value or edit distance value, and processing result.

[0093] The system exports the generated risk report as a portable document file conforming to the International Organization for Standardization 32000-2 standard for users to download and archive. At the same time, the system sends the report download link to authorized users through the system's message center.

[0094] Example 1: This example provides a specific implementation of a method for intelligent risk warning of procurement documents based on feature fusion and blacklists / whitelists. The method is implemented in two operating modes: a server-side software-as-a-service version and a standalone version. The server-side version runs on a Linux operating system, uses Docker container technology to build application images, supports horizontal scaling via Kubernetes container orchestration, and supports load balancing across multiple machines and GPUs. The deep learning inference module is based on the Nvidia TensorRT high-performance inference framework, supports Ampere, Ada Lovelace, and Blackwell computing power system acceleration cards, and its processing logic is designed asynchronously based on a unified computing device architecture. The CPU execution section supports advanced vector extension 512 single instruction multiple data instruction set acceleration. The standalone version supports Windows 10 and Windows 11 64-bit operating systems, including CPU and GPU versions. The CPU version is based on the Intel OpenVINO deep learning inference framework, and the GPU version is based on the Nvidia TensorRT deep learning inference framework.

[0095] In this embodiment, steps S1 to S7 are executed according to the above process. In step S2, the system uses the Burrows-Wheeler transform to transform the effective word sequence of each document and extracts the longest run length of consecutive identical words. Run-length encoding compression ratio Word frequency vector Simultaneously, the Nilsimsa hash algorithm is used to generate a 256-bit Nilsimsa hash digest. The two are fused into a 192-dimensional fused fingerprint vector. The system generates 32 random projection vectors, which will then be fused with the fingerprint vector. The projection is a 32-dimensional projected signature vector. It is divided into 16 frequency bands for index construction, and the Manhattan distance threshold is used. The value is 42.0.

[0096] In step S3, the system extracts the color histogram feature vector of each valid image. Local binary pattern texture feature vector and edge feature vectors The vectors are concatenated to form a 416-dimensional original visual feature vector. Then, the dimensionality was reduced to 64-dimensional visual feature vectors using 64 random projection vectors. For two images in a candidate file pair, the system calculates the Hellinger distance between their visual feature vectors. And convert it into an image similarity score. For scanned images, additional scale-invariant feature transformation (SMT) feature extraction and matching are performed. The image similarity score fused from random projection and Hellinger distance is then compared with the SMT matching score. , A secondary fusion is performed. The system uses the Hungarian algorithm to solve the maximum weight matching problem in the bipartite graph to obtain file-level image similarity. .

[0097] In step S4, the system uses a composite text similarity metric that combines edit distance similarity and Jaro-Winkler distance, with edit distance similarity weighted... Jaro-Winkler distance weights The system introduces a multimodal feature collaborative optimization mechanism, with initial weights set to... , and Learning rate Convergence accuracy threshold The maximum number of iterations is 100.

[0098] In step S5, the system sets a comprehensive similarity threshold. Size of window for detecting duplicate long texts The repetition rate threshold is 0.3, and the risk warning threshold is... The weights of the five risk indicators are as follows: , , , , .

[0099] In step S6, the system reads the preset blacklist database. and whitelist database Blacklisted word retrieval employs both exact matching and fuzzy matching with an edit distance less than or equal to 2. The base weight coefficient for blacklisted words... Whitelist matching threshold .

[0100] In step S7, the system organizes all the processing results from steps S1 to S6 into a report data model according to five levels: project overview layer, document comparison layer, risk analysis layer, detailed evidence layer, and blacklist / whitelist filtering layer. This generates a structured risk report containing an execution summary, a comprehensive similarity matrix, a risk ranking list, details of duplicate content, a risk evidence chain, and details of blacklist / whitelist filtering. The report is then exported as a portable document file.

[0101] Example 2: Based on Example 1, Example 2 dynamically adjusts the Burrows-Wheeler transform and Nilsimsa hash fusion parameters in step S2 according to the statistical results of historical deduplication tasks.

[0102] The system records the recall rate of the most recent 100 plagiarism detection tasks. and accuracy Recall rate Precision is defined as the ratio of the number of correctly recalled similar file pairs to the total number of similar file pairs. Defined as the ratio of the number of correctly recalled similar file pairs to the total number of recalled file pairs. The total number of similar file pairs is the total number of manually labeled similar file pairs.

[0103] If recall rate Below 0.95, the system will reduce the number of frequency bands. The number of projected values ​​per frequency band increased from 16 to 24. This improves the sensitivity of the index. If the accuracy... If the value is below 0.80, the system will randomly project the number of vectors. Increasing the dimension of the projected signature vector from 32 to 48 increases the dimension of the vector, thereby reducing the probability of hash collisions. The thresholds and adjustment ranges mentioned above were determined through statistical analysis of 100 historical deduplication tasks. 0.95 and 0.80 represent the minimum acceptable values ​​for recall and precision in engineering applications, respectively. The adjustment range ensures that parameter changes have an observable impact on indexing performance without causing excessive oscillations. Before each deduplication task begins, the system decides whether to adjust the parameters based on the statistical results of the most recent 100 tasks. The adjusted parameters only apply to the current task. The other steps in Example 2 are exactly the same as in Example 1.

[0104] Example 3: Building upon Example 1, Example 3 introduces prior weight constraints based on historical data into the multimodal feature collaborative optimization mechanism in step S4. When initializing iterative weights, the system no longer uses fixed initial weights but calculates prior weights based on the average performance of each modality in historical plagiarism detection tasks. The system records the optimal weight combination at the final convergence of each of the last 500 plagiarism detection tasks and calculates the average weights for the text modality, image modality, and table modality respectively. , and The system sets the prior weights as follows: ;in For the prior weights of the text modality, The average text modal weights in historical tasks. ;in For the prior weights of the image modality, This represents the average image modal weights in historical tasks. ;in For the prior weights of the tabular modality, This represents the average table modal weights from historical tasks. The system incorporates prior constraints during each weight update iteration, resulting in the revised weight update formula: ; in This is the prior constraint strength coefficient, and its value is 0.1. For the first The weight coefficient of the text modality in the next iteration. This value was determined by cross-validation on 500 historical plagiarism detection tasks. Within the search range of 0.05 to 0.30, 0.1 achieved the optimal balance between convergence speed and final accuracy.

[0105] ; in For the first The weighting coefficients of the image modality in the next iteration.

[0106] ; in For the first The weight coefficients of the table modality in this iteration. The updated weights are normalized to ensure... The other steps in Example 3 are exactly the same as in Example 1.

Claims

1. A method for intelligent early warning of risks in procurement documents based on feature fusion and blacklists / whitelists, characterized in that: include: Step 1: Perform multimodal parsing on the tender documents and bid documents to extract text content, image content and table content. Clean the text content to form a structured text content tree and perform structured transformation on the table content to form structured table data. Step 2: Process the structured text content tree to extract a set of effective words and key indicator words, generate a fusion fingerprint vector based on the set of effective words, filter out a set of candidate similar tender documents based on the fusion fingerprint vector, and mark high similarity candidate document pairs according to the distance metric between the fusion fingerprint vectors; Step 3: Perform visual feature extraction and dimensionality reduction on the image content to obtain visual feature vectors; calculate the probability distribution distance metric between the visual feature vectors of the images in the high similarity candidate file pair and convert it into an image similarity score; calculate the file-level image similarity based on the image similarity score. Step 4: Calculate sentence-level text similarity for the text content in the high-similarity candidate file pairs and adjust the file-level text similarity using the key indicator words. Compare the structured table data to calculate file-level table similarity. Then, through multimodal feature collaborative optimization, fuse the file-level text similarity, the file-level image similarity, and the file-level table similarity into a comprehensive similarity. Step 5: Mark the file pairs whose comprehensive similarity exceeds the first threshold as highly similar file pairs, perform deep analysis on the highly similar file pairs, calculate the comprehensive risk score of bid rigging and collusion in the bid documents based on the deep analysis results, and mark the risk level. Step 6: Perform blacklist word search on the tender documents marked with risk level to calculate blacklist penalty value, perform whitelist content matching to remove similarity contribution from whitelist content, recalculate the comprehensive risk score based on the refined similarity after removal and superimpose it with the blacklist penalty value to obtain the final risk score, and correct the risk level according to the final risk score. Step 7: Organize all processing results into a report data model, and generate a structured risk report based on the report data model.

2. The intelligent early warning method for procurement document risks based on feature fusion and blacklists / whitelists as described in claim 1, characterized in that, In the step of generating a fused fingerprint vector based on the effective word set, the Burrows-Wheeler transform is used to transform the word sequence composed of the effective word set, constructing all cyclic shifts of the word sequence. After sorting all cyclic shifts in lexicographical order, the last column is taken as the Burrows-Wheeler transform result. From the Burrows-Wheeler transform result, the longest run length, run-length encoding compression ratio, and occurrence frequency vector of each word are extracted. The longest run length and the occurrence frequency vector are normalized and concatenated to form a Burrows-Wheeler feature vector. At the same time, the Nilsimsa hash algorithm is used to generate a Nilsimsa hash digest of the word sequence composed of the effective word set and convert it into a Nilsimsa feature vector. The Burrows-Wheeler feature vector and the Nilsimsa feature vector are concatenated to form the fused fingerprint vector.

3. The intelligent early warning method for procurement document risks based on feature fusion and blacklists / whitelists as described in claim 1, characterized in that, In the step of selecting a set of candidate similar tender documents based on the fused fingerprint vector, a preset number of random projection vectors are generated. The fused fingerprint vector is then multiplied by each random projection vector to obtain a corresponding number of projection values. All projection values ​​are combined into a projection signature vector. The projection signature vector is divided into a preset number of frequency bands, each containing an equal number of projection values. The projection values ​​within each frequency band are concatenated to form the key value for that frequency band. For each tender document, it is checked whether the key values ​​of each frequency band have the same key value in the tender document set. All tender documents with the same key value in any frequency band are added to the set of candidate similar tender documents.

4. The intelligent early warning method for procurement document risks based on feature fusion and blacklists / whitelists as described in claim 1, characterized in that, In the step of extracting visual features and reducing dimensions of the image content to obtain a visual feature vector, the image is converted from the red-green-blue color space to the hue-saturation-brightness color space, and the histograms of the hue channel, saturation channel and brightness channel are calculated and concatenated into a color histogram feature vector. The texture features of the image are extracted using a local binary mode operator. For each pixel in the image, its neighborhood is taken, and the gray values ​​of the adjacent pixels in the neighborhood are compared with the gray value of the center pixel to generate local binary mode encoded values. The histogram of the local binary mode encoded values ​​of all pixels is used as the local binary mode texture feature vector. The horizontal and vertical gradients of the image are extracted using the Sobel operator, and the gradient magnitudes are calculated. The histogram of the gradient magnitudes is used as the edge feature vector. The color histogram feature vector, the local binary mode texture feature vector, and the edge feature vector are concatenated to form the original visual feature vector. A preset number of random projection vectors are generated to project and reduce the dimensionality of the original visual feature vector to the original visual feature vector.

5. The intelligent early warning method for procurement document risks based on feature fusion and blacklists / whitelists according to claim 1, characterized in that, In the step of calculating the probability distribution distance metric between the visual feature vectors of the images in the high similarity candidate file pair and converting it into an image similarity score, the visual feature vectors of the two images are normalized to probability distributions respectively, the Hellinger distance between the two probability distributions is calculated as the probability distribution distance metric, and the complement of the Hellinger distance is used as the image similarity score.

6. The intelligent early warning method for procurement document risks based on feature fusion and blacklists / whitelists according to claim 1, characterized in that, In the step of fusing the file-level text similarity, file-level image similarity, and file-level table similarity into a comprehensive similarity through multimodal feature collaborative optimization, the number of iterations is initialized and the initial weights of the text modality, image modality, and table modality are set. In the t-th iteration, the comprehensive similarity of each candidate file pair is calculated according to the current weights. The average similarity of all candidate file pairs in the text modality, image modality, and table modality in this round of iteration is calculated. The weights of each modality in the next round of iteration are updated according to the deviation between the average similarity of each modality and the comprehensive average similarity based on a preset learning rate. The iteration stops when the absolute value of the change in the comprehensive similarity of all candidate file pairs in two adjacent iterations is less than the convergence accuracy threshold.

7. The intelligent early warning method for procurement document risks based on feature fusion and blacklists / whitelists according to claim 1, characterized in that, In the deep analysis, continuous long text duplication detection uses a sliding window method to compare the full text of the tender documents and bidding documents. The window size is a preset number of consecutive characters. The window slides with a single character step size, recording the repeated text segments when the character sequences within the window are completely identical and accumulating the duplication length. The ratio of the total length of repeated text characters to the total length of the cleaned text of the tender documents is calculated. If the ratio is greater than or equal to the duplication ratio threshold, it is determined that there is a large section of text plagiarism. For misspelling consistency detection, a BERT pre-trained language model is used to analyze the text segments in highly similar documents whose sentence-level text similarity exceeds a preset value sentence by sentence. The sentence is used as the input sequence to obtain the correct probability of each character position. Characters with a predicted probability lower than the preset probability threshold and an edit distance of 1 from the highest probability character are judged as misspellings. If the two documents have the same misspellings or the same type of incorrect punctuation marks at the same page number and the same sentence number position, they are marked as common error features. Seal similarity comparison extracts the image region dominated by the red channel as the seal candidate region, performs connected component analysis on the seal candidate region to extract the complete seal image, and calculates the similarity between seals using single image similarity scores. If the seal image similarity exceeds the seal similarity threshold, it is marked as highly similar seals.

8. The intelligent early warning method for procurement document risks based on feature fusion and blacklists / whitelists according to claim 1, characterized in that, The comprehensive risk score is calculated by weighting and summing four risk indicators: comprehensive similarity exceeding the standard, continuous long text duplication, common error feature, and seal high similarity, according to preset weights. The comprehensive similarity exceeding the standard indicator is triggered when any tender document has a comprehensive similarity greater than or equal to the first threshold. The continuous long text duplication indicator is triggered when any tender document has a continuous duplication ratio greater than or equal to the duplication ratio threshold. The common error feature indicator is triggered when at least one common error feature is detected. The seal high similarity index is triggered when the similarity of any seal image exceeds the seal similarity threshold.

9. The intelligent early warning method for procurement document risks based on feature fusion and blacklists / whitelists according to claim 1, characterized in that, In the step of performing whitelist content matching to remove similarity contributions belonging to whitelist content, the whitelist content includes three sub-libraries: a legal and regulatory text library, an industry standard library, and a self-built whitelist library. For text content, the Jaccard similarity between the text segment to be checked and each entry in the whitelist library is calculated. If the Jaccard similarity exceeds the whitelist matching threshold, the text segment is determined to belong to the whitelist content. For image content, the similarity between the image to be checked and each image in the whitelist library is calculated using a single image similarity score. If the single image similarity score exceeds a preset image whitelist threshold, it is determined to be a whitelist image. For table content, the similarity between the table to be checked and each table in the whitelist library is calculated using table similarity. If the table similarity exceeds a preset table whitelist threshold, it is determined to be a whitelist table. The similarity contributions of all text content, whitelist images, and whitelist tables determined to be whitelisted are deducted from the comprehensive similarity.

10. The intelligent early warning method for procurement document risks based on feature fusion and blacklists / whitelists according to claim 5, characterized in that, In the step of calculating the probability distribution distance metric between the visual feature vectors of the images in the high similarity candidate file pair and converting it into an image similarity score, scale-invariant feature transformation (SMT) feature extraction and matching are additionally performed on the images. SMT keypoints are extracted from the two images respectively. The descriptor vectors of the two sets of keypoints are matched, and the ratio of the nearest neighbor distance to the second nearest neighbor distance is calculated. Matching pairs with a ratio less than a preset matching threshold are considered valid matching pairs. The ratio of the number of valid matching pairs to the average of the total number of keypoints in the two images is calculated as the SMT matching score. The image similarity score obtained by the Hellinger distance conversion and the SMT matching score are fused twice according to a preset weight to obtain a comprehensive image similarity score.

Images