Automatic reuse method of bid qualification performance based on state space model

Abstract

The application discloses a bid qualification performance automatic reuse method based on a state space model and relates to the technical field of artificial intelligence, including the following steps: step 1: performing text analysis on a bid document, performing entity boundary marking on a qualification performance reference paragraph to obtain a qualification performance entity set; step 2: performing associated assembly on the entities based on the geodesic distance between hyperbolic embedding vectors in the same paragraph to obtain a qualification performance associated group; step 3: calculating a non-consistency score for each qualification performance associated group and determining a calibration critical value, and automatically filling the entity information in the reusable associated group into a qualification performance reporting form of this bid. The application realizes long document unified modeling, hierarchical perception entity association matching and automatic reuse with statistical reliability guarantee, and effectively improves the efficiency and accuracy of bid qualification performance reporting.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method for automatically reusing bidding qualification and performance records based on a state-space model. Background Technology

[0002] In bidding processes, suppliers are required to submit qualification certificates, project performance records, and other supporting documentation for each bid, and to fill in key information such as qualification name, certificate number, project name, and project completion time in the corresponding forms of the bid documents. Because the same supplier repeatedly bids for different projects, a large amount of qualification and performance information is highly repetitive across bids. However, existing bidding systems generally lack automatic reuse mechanisms, forcing suppliers to manually fill in the information each time, resulting in a large workload, low efficiency, and a high error rate.

[0003] Currently, information extraction from tender documents primarily relies on a combination of optical character recognition (OCR) and named entity recognition (NAME) technologies. OCR converts images or scanned copies of the tender document into editable text, while NAME locates and extracts key entities such as qualification names and certificate numbers. However, tender documents are typically lengthy, often dozens or even hundreds of pages long, containing a significant amount of content unrelated to qualifications and performance. Existing sequence coding models based on self-attention mechanisms face a bottleneck when processing such long documents, as computational complexity increases quadratically with sequence length. This makes end-to-end unified modeling of the entire document difficult, often requiring the document to be truncated into multiple short segments for separate processing. This results in the loss of contextual information across pages and paragraphs, impacting the accuracy of entity extraction.

[0004] In entity association matching, existing technologies typically employ a matching strategy based on Euclidean space vector similarity. This involves encoding extracted entities into fixed-dimensional vectors and then calculating cosine or Euclidean distance for pairing. However, qualification names in industry practice naturally possess a hierarchical structure. For example, management system certifications may have specific certification categories such as ISO 9001 and ISO 14001, with each category further subdivided into different versions and levels. The flat geometry of Euclidean space cannot effectively represent this hierarchical relationship, making it difficult to correctly distinguish semantically similar qualification names at different levels within the embedding space, thus affecting the accuracy of association assembly.

[0005] Furthermore, existing automatic form filling schemes lack an effective confidence calibration mechanism before filling the matching results back into the form. They usually rely solely on fixed thresholds for truncation and screening, which cannot statistically guarantee that the proportion of correct matches among the filtered results reaches the preset reliability level. In high-risk government procurement and corporate procurement scenarios, it is difficult to meet the strict requirements for the accuracy of the form filling. Summary of the Invention

[0006] In view of this, the present invention provides an automatic reuse method for bidding qualification and performance based on a state-space model. The method marks the calibrated association groups as reusable association groups and automatically fills them into the bidding form. This achieves unified modeling of long documents, hierarchical entity association matching, and automatic reuse with statistical reliability assurance, effectively improving the efficiency and accuracy of filling in bidding qualification and performance information.

[0007] The technical solution adopted in this invention is as follows:

[0008] The method for automatically reusing bidding qualification and performance data based on state-space models includes the following steps:

[0009] Step 1: Perform text parsing on the tender document, concatenate the text content into a continuous tag sequence and divide it into several paragraph-level subsequences, input the continuous tag sequence into a sequence encoder built based on a selective state space mechanism to obtain the context encoding vector corresponding to each tag, filter out the qualification and performance reference paragraphs based on the context encoding vectors, and perform entity boundary annotation on the qualification and performance reference paragraphs to obtain the qualification and performance entity set.

[0010] Step 2: Project the Euclidean eigenvector of each entity in the qualification and performance entity set onto the hyperbolic embedding space defined by the Poincaré sphere model through exponential mapping to obtain a hyperbolic embedding vector. Perform association assembly on the entities within the same paragraph based on the geodesic distance between the hyperbolic embedding vectors to obtain the qualification and performance association group.

[0011] Step 3: Calculate the inconsistency score for each qualification and performance association group and determine the calibration threshold. Mark the qualification and performance association groups whose inconsistency scores do not exceed the calibration threshold as reusable association groups. Automatically fill the entity information in the reusable association groups into the qualification and performance filling form for this bid.

[0012] Furthermore, in step 1, the text content is concatenated into a continuous mark sequence according to the page number order. During the concatenation process, the continuous mark sequence is divided into several paragraph-level subsequences based on line break marks and page spacing. Each paragraph-level subsequence corresponds to an independent paragraph and is assigned a paragraph number. Each mark in the continuous mark sequence is assigned an absolute position number, which starts from 1 and increases by natural numbers.

[0013] Furthermore, in step 1, the sequence encoder is a Mamba sequence encoder, which consists of three components: an input projection layer, a selective state space computation layer, and an output projection layer. The input projection layer expands the embedding vector of each tag along the channel dimension to the internal dimension. The selective state space computation layer uses the expanded internal dimension vector as input to perform selective state recursion. The output projection layer compresses the internal dimension vector output by the selective state space computation layer back to the original embedding dimension to obtain the context encoding vector corresponding to each tag. The context encoding vector is in the Euclidean representation space.

[0014] Furthermore, the execution process of the selective state-space computation layer is as follows: For the first label in a continuous label sequence, the internal dimension vector of the first label is transformed into a discretization step size value, a state transition vector, and an input modulation vector specific to the first label through a linear mapping. Zero-order preserved discretization is performed on the state transition vector using the discretization step size value to obtain a discrete state transition vector. The discrete state transition vector is then multiplied element-wise with the state vector initialized to all zeros, and the element-wise product of the input modulation vector and the internal dimension vector of the first label is added to obtain the updated state vector corresponding to the first label. For each subsequent label starting from the second label in a continuous label sequence, the discretization step size value, state transition vector, and input modulation vector specific to each subsequent label are generated in the same manner. A dedicated discretization step size is used to perform zero-order preserved discretization on the state transition vector of each subsequent label to obtain a discrete state transition vector for each subsequent label. The discrete state transition vector of each subsequent label is then multiplied element-wise with the updated state vector corresponding to the previous label, and the element-wise product of the input modulation vector of each subsequent label and the internal dimension vector of each subsequent label is added to obtain the updated state vector corresponding to each subsequent label. The updated state vector corresponding to each label is then multiplied element-wise through a fixed output mapping vector and summed along the state dimension to obtain the single-channel output of each label in the selective state space computation layer. The single-channel outputs of all channels are then concatenated along the channel dimension to obtain the complete output of each label in the selective state space computation layer.

[0015] Furthermore, in step 1, the screening process for qualification and performance reference paragraphs is as follows: the average value of the context encoding vectors of all marked subsequences in each paragraph is taken along the sequence length dimension to obtain the paragraph representation vector. The paragraph representation vector is sent to the paragraph classification head. The paragraph classification head outputs the judgment score of each paragraph subsequence belonging to the qualification and performance reference paragraph. The paragraph subsequence with the judgment score exceeding the preset positioning threshold is marked as the qualification and performance reference paragraph.

[0016] Furthermore, in step 1, the entity boundary annotation process is as follows: the context encoding vector of each tag in the qualification and performance reference paragraph is sent to the entity annotation header. The entity annotation header outputs the score of each tag belonging to each of the nine categories: qualification name start position, qualification name internal position, certificate number start position, certificate number internal position, project name start position, project name internal position, project completion time start position, project completion time location and non-entity position. The category with the highest score is taken as the entity annotation result of each tag. Tags belonging to the same entity category and satisfying the adjacent position constraint are merged into one entity. Each entity contains entity category, entity text and paragraph number.

[0017] Furthermore, in step 2, the execution process of the exponential mapping is as follows: taking the origin of the Poincaré sphere model as the base point, calculate the Euclidean norm of the Euclidean eigenvectors, divide the Euclidean eigenvectors by the Euclidean norm to obtain the unit direction vector, input the Euclidean norm into the hyperbolic tangent function to obtain the hyperbolic radial component, multiply the hyperbolic radial component by the unit direction vector element by element to obtain the hyperbolic embedding vector of each entity inside the Poincaré sphere, and the hyperbolic embedding vector is in the hyperbolic embedding space.

[0018] Furthermore, in step 2, before performing the associated assembly, the hierarchical number of each qualification name entity in the qualification performance entity set is determined according to the preset qualification classification hierarchical knowledge base. The qualification classification hierarchical knowledge base records the hierarchical relationship between qualification names at all levels in the industry qualification system. A qualification hierarchical tree is constructed according to the hierarchical number. The root node of the qualification hierarchical tree is the virtual top-level node. Each qualification name entity is attached to the corresponding level position in the qualification hierarchical tree according to the hierarchical number. For each edge from the parent node to the child node in the qualification hierarchical tree, the geodesic distance between the hyperbolic embedding vector of the parent node and the hyperbolic embedding vector of the child node on the Poincaré sphere is calculated as the hierarchical distance value.

[0019] Furthermore, in step 2, the association assembly is divided into three stages: The first stage is qualification certificate pairing. Under the same paragraph number, the geodesic distance on the Poincaré sphere is calculated for each qualification name entity and each certificate number entity within the same paragraph. The certificate number entity with the smallest geodesic distance is paired with the qualification name entity to form a qualification certificate pair. The second stage is performance record pairing. Under the same paragraph number, the geodesic distance on the Poincaré sphere is calculated for each project name entity and each project completion time entity within the same paragraph. The project completion time entity with the smallest geodesic distance is paired with the project name entity to form a performance record pair. The third stage is cross-category association assembly, calculating qualification certificates... The geodesic distance on the Poincaré sphere between the hyperbolic embedding vector of the qualification name entity in the certificate pair and the hyperbolic embedding vector of the project name entity in the performance record pair is used as the first sub-item distance. The geodesic distance on the Poincaré sphere between the hyperbolic embedding vector of the certificate number entity in the qualification certificate pair and the hyperbolic embedding vector of the project completion time entity in the performance record pair is used as the second sub-item distance. The sum of the first sub-item distance and the second sub-item distance is used as the cross-category association score. Qualification certificate pairs and performance record pairs with cross-category association scores lower than the preset association threshold are merged into one qualification-performance association group. Each qualification-performance association group contains one qualification name, one certificate number, one project name, and one project completion time.

[0020] Furthermore, in step 3, the calculation process of the inconsistency score is as follows: the Euclidean feature vectors of the four entities in the qualification and performance association group are concatenated along the vector dimension to form a combined vector. The combined vector is in the Euclidean representation space. The combined vector is fed into the calibration scoring function. The calibration scoring function is a linear transformation pre-trained based on the labeled correct association group and incorrect association group. The calibration scoring function outputs the inconsistency score, which represents the degree of deviation between the qualification and performance association group and the confirmed correct association pattern. The process of determining the calibration threshold is as follows: all calibration samples are taken from the pre-constructed calibration sample pool. The calibration sample pool consists of qualification and performance association groups that have been manually verified and confirmed to be correctly or incorrectly associated, and their corresponding inconsistency scores. The inconsistency scores of all calibration samples are arranged in ascending order to form an ordered sequence. The process is traversed in ascending order along the ordered sequence. The inconsistency score at the position where the proportion of the number of samples with inconsistency scores not exceeding the current position in the ordered sequence is not less than the preset target coverage rate for the first time is taken as the calibration threshold.

[0021] By adopting the above technical solutions, the present invention achieves the following beneficial effects: The automatic reuse method for bidding qualification and performance based on the state-space model provided by the present invention introduces a selective state-space mechanism into the long sequence modeling of bidding documents, enabling the model to perform end-to-end unified encoding of the entire bidding document with linear computational complexity. This avoids the defect of traditional self-attention mechanisms in processing long documents where the computational load increases quadratically with the sequence length, forcing document truncation. It effectively preserves the long-distance contextual dependencies of qualification and performance information across pages and paragraphs, improving the accuracy of qualification and performance reference paragraph location and entity boundary labeling. By projecting the extracted entities from the Euclidean representation space to the hyperbolic embedding space defined by the Poincaré sphere model through exponential mapping, the geometric characteristics of the exponential volume growth of hyperbolic space are used to naturally adapt to the hierarchical structure of qualification names. This ensures that the distribution of qualification name entities at different levels in the embedding space is consistent with their subordinate relationships in the qualification system. Compared with the flat geometric structure of Euclidean space, this provides clearer entity distinction boundaries and improves the accuracy of entity association assembly within paragraphs. By introducing a compliance prediction calibration mechanism before automatic data entry, and determining the calibration threshold based on a pre-built calibration sample pool, the system performs statistical reliability screening on the inconsistency scores of each qualification and performance association group. This ensures that the coverage of correctly associated groups in the reusable association groups that have passed calibration is not lower than the preset target level under limited samples. This overcomes the lack of probability guarantee in the traditional fixed threshold truncation method. In high-risk procurement scenarios, it provides a quantifiable reliability basis for automatic data entry results. At the same time, by referencing calibration records, it provides a traceable verification reference for reviewers, balancing bidding efficiency and information accuracy. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the recursive principle of the Mamba selective state space computation layer in this embodiment of the invention;

[0023] Figure 2 This is a schematic diagram of the assembly of entities within a segment of the Poincaré disk in an embodiment of the present invention;

[0024] Figure 3 This is a schematic diagram comparing the F1 values ​​of associated assembly under different embedding spaces in an embodiment of the present invention;

[0025] Figure 4 This is a schematic diagram showing the relationship between the correlation accuracy and the automatic reporting ratio under different preset target coverage rates in embodiments of the present invention. Detailed Implementation

[0026] The method for automatically reusing bidding qualification and performance data based on state-space models includes the following steps:

[0027] Step 1: Perform text parsing on the tender document, concatenate the text content into a continuous tag sequence and divide it into several paragraph-level subsequences, input the continuous tag sequence into a sequence encoder built based on a selective state space mechanism to obtain the context encoding vector corresponding to each tag, filter out the qualification and performance reference paragraphs based on the context encoding vectors, and perform entity boundary annotation on the qualification and performance reference paragraphs to obtain the qualification and performance entity set.

[0028] Step 2: Project the Euclidean eigenvector of each entity in the qualification and performance entity set onto the hyperbolic embedding space defined by the Poincaré sphere model through exponential mapping to obtain a hyperbolic embedding vector. Perform association assembly on the entities within the same paragraph based on the geodesic distance between the hyperbolic embedding vectors to obtain the qualification and performance association group.

[0029] Step 3: Calculate the inconsistency score for each qualification and performance association group and determine the calibration threshold. Mark the qualification and performance association groups whose inconsistency scores do not exceed the calibration threshold as reusable association groups. Automatically fill the entity information in the reusable association groups into the qualification and performance filling form for this bid.

[0030] In actual bidding processes, tender documents are typically in PDF or Word format, ranging from dozens to hundreds of pages in length. They contain a significant amount of content unrelated to qualifications and performance records, such as bid prices, commercial terms, and technical specifications. Directly extracting entities from the entire tender document introduces substantial noise from this irrelevant text, leading to a decrease in entity recognition accuracy. Therefore, this embodiment employs a hierarchical processing strategy of first locating and then extracting entities. First, paragraphs relevant to qualifications and performance records are selected from the entire tender document. Then, only these paragraphs undergo refined entity boundary annotation, thereby reducing computational overhead while improving extraction accuracy.

[0031] During the text parsing phase, for PDF format bid documents, a PDF text extraction tool is used to read the text content of each page one by one, retaining page number information for subsequent sequential concatenation. For Word format bid documents, the main text paragraphs and table cell content are extracted segment by segment through a document parsing interface. During parsing, non-text information such as headers, footers, watermarks, and page numbers are filtered out, retaining only the valid text in the main text area. The extracted text content from all pages is then concatenated sequentially according to the page numbers in ascending order, forming a continuous text stream covering the entire bid document.

[0032] After concatenation, the continuous text stream needs to be segmented into paragraph-level subsequences. Paragraph segmentation is based on two signals: line break marks and page spacing. When consecutive line breaks appear in the text stream, their positions are used as the boundaries between adjacent paragraphs. Similarly, when the text source is a PDF file and page layout analysis detects that the vertical spacing between adjacent text blocks exceeds 1.5 times the line spacing of the main text, this spacing position is also used as the paragraph boundary. After this segmentation, the continuous text stream is divided into several paragraph-level subsequences, each corresponding to one independent paragraph. Each paragraph-level subsequence is assigned a paragraph number, starting from 1 and incrementing by natural numbers. These numbers are used to limit the scope of entity pairing during the subsequent association and assembly stage.

[0033] Subsequently, tokenization processing is performed on the continuous text stream. Tokenization segments each character or word in the continuous text stream into independent tokens, with the specific segmentation method depending on the tokenizer used. In this embodiment, a character-level tokenization strategy is adopted, treating each Chinese character and each English letter and number as an independent token. All tokens are assigned absolute position numbers according to their order of appearance in the continuous text stream, starting from 1 and incrementing by natural numbers, continuing throughout the entire tender document without being reset by paragraph boundaries. Each token also inherits the paragraph number of its corresponding paragraph-level subsequence, allowing subsequent processing to locate and group tokens based on both absolute position numbers and paragraph numbers.

[0034] After the text parsing and tokenization processes described above, tender documents typically generate thousands to tens of thousands of tokens. Traditional encoding models based on self-attention mechanisms face significant computational bottlenecks when processing such long sequences. This is because the computational cost of self-attention mechanisms increases quadratically with sequence length. When the sequence length exceeds several thousand tokens, memory usage and computation time both rise sharply, making deployment in practical engineering scenarios difficult. To address this issue, this embodiment employs a Mamba sequence encoder based on a selective state-space mechanism to encode continuous token sequences. The computational cost of the Mamba sequence encoder increases linearly with sequence length, enabling it to maintain the ability to model long-distance contextual dependencies while keeping computational complexity within an engineering-acceptable range.

[0035] The Mamba sequence encoder consists of three components connected in series: an input projection layer, a selective state-space computation layer, and an output projection layer.

[0036] The input projection layer receives the embedding vector of each label as input. The embedding vector is obtained by adding the word embedding and the position embedding of the label. The word embedding is obtained from the pre-trained embedding matrix through a lookup table, and the position embedding is obtained from the learnable position embedding matrix based on the absolute position number. In this embodiment, the dimension of the embedding vector is set to 768. The input projection layer expands the embedding vector of each label from 768 dimensions to an inner dimension of 1536 through a linear transformation, i.e., an expansion factor of 2. The purpose of expanding to a higher dimension is to provide a richer feature representation space for the subsequent selective state space computation layer, enabling each channel to capture semantic information of different granularities.

[0037] The selective state-space computation layer is a core component of the Mamba sequence encoder. Unlike traditional state-space models that use fixed state transition parameters, the selective state-space computation layer dynamically generates state transition parameters based on the input content of each tag. This input-dependent parameter generation method allows the model to adaptively determine how much contextual information to retain and how much current input information to incorporate, based on the semantic content of the current tag. Specifically, when processing tags closely related to qualifications and performance (such as "certification," "qualification," "project," etc.), the model tends to generate larger discretization step sizes, thus more fully incorporating the information of the current tag into the state vector; while when processing tags unrelated to qualifications and performance (such as formatting punctuation, irrelevant terminology, etc.), the model tends to generate smaller discretization step sizes, causing the state vector to retain more existing contextual information and absorb less current input. This selective gating characteristic allows contextual information related to qualifications and performance to continuously accumulate and propagate in the state vector, even if this information is scattered in different locations in the tender document.

[0038] The selective state-space computation layer operates independently along the channel dimension, with a total of 1536 channels, each maintaining an independent state recursion chain. The state vector dimension of each channel is set to 16, meaning the order of the state space is 16. For the first label in a continuous label sequence, the internal dimension vector obtained after expanding the first label through the input projection layer is used to generate three specific parameters through a linear mapping: a discretization step size, a state transition vector, and an input modulation vector. The discretization step size is a positive real scalar, obtained by applying a softplus activation function to the output of the linear mapping to ensure that the step size is always positive; the dimension of the state transition vector is the same as the order of the state space, 16-dimensional; the input modulation vector is also 16-dimensional. Subsequently, zero-order hold discretization is performed on the state transition vector using the discretization step size. Zero-order hold discretization is a standard method for converting parameters of a continuous-time dynamical system into discrete-time parameters. For each element in the state transition vector... The corresponding elements after discretization The calculation method is as follows ,in This is the discretization step size value. The state transition vector is the first One element, It is a natural exponential function. Discretization yields a discrete state transition vector. Since there is no historical information before the first marker, the state vector is initialized as a 16-dimensional vector with all zeros. The discrete state transition vector... Multiply element-wise with the all-zero state vector (the result is still all zeros), then add the input modulation vector. The component values ​​of the inner dimension vector of the first label in the current channel The product of, i.e. ,in The initial state vector is all zero. This is the updated state vector corresponding to the first label. This represents element-wise product.

[0039] For each subsequent label in a continuous label sequence, starting from the second label, the same parameter generation and state recursion process is performed. The internal dimension vector of each subsequent label generates a unique discretization step size, state transition vector, and input modulation vector through the same linear mapping. After zero-order preserved discretization, a unique discrete state transition vector is obtained. This unique discrete state transition vector is then element-wise multiplied with the updated state vector corresponding to the previous label, and the result is added to the product of the unique input modulation vector and the component values ​​of the current label's internal dimension vector in the current channel, yielding the updated state vector corresponding to the current label. This can be expressed as: subscript This indicates the absolute position number of the current tag within a consecutive tag sequence. This is the updated state vector corresponding to the previous label. This is the discrete state transition vector specific to the current label. For the current label-specific input modulation vector, This represents the component value of the current marker in the current channel. Through the above recursive process, the state vector of each marker accumulates all contextual information from the start position of the sequence to the current position, and the degree of information retention is dynamically adjusted by the selective gating parameters specific to each marker.

[0040] In each channel, the updated state vector corresponding to each tag is mapped to a fixed output vector. Convert to scalar output. Specifically, convert the updated state vector... Mapping vector with output After performing element-wise multiplication, summation is performed along the state dimension, i.e. ,in is the order of the state space, with a value of 16, is the -th element of the output mapping vector, is the -th element of the updated state vector, is the single-channel output of the current token on the current channel. The single-channel outputs of all 1536 channels are concatenated along the channel dimension to form the complete output of each token at the selective state space computation layer, with a dimension of 1536.

[0041] The output projection layer compresses the 1536-dimensional complete output of the selective state space computation layer back to the original embedding dimension of 768 through a linear transformation, obtaining the context encoding vector corresponding to each token. The context encoding vector is in the Euclidean representation space and carries the semantic representation of the current token in the context of the full text. In actual implementation, multiple layers of the above structure can be stacked to enhance the expression capability of the model. In this embodiment, 12 layers are stacked, that is, the context encoding vector of the previous layer is used as the input embedding vector of the next layer, and encoding is performed progressively layer by layer. Finally, the output of the last layer is taken as the context encoding vector of each token.

[0042] Reference Figure 1 , Figure 1 the bottom of is a token sequence, which contains 8 tokens. From left to right, they are "tou", "biao", "zi", "zhi", "ISO", "9001", "ren", "zheng", representing a token sequence fragment formed by tokenizing a continuous text parsed from a bidding document. Each token occupies an independent rectangular box, and the corresponding token text is marked in the rectangular box. Above the token sequence is the discretized step size region. In this region, vertical bars represent the discretized step size generated by each token after linear mapping and softplus activation function . The discretized step sizes corresponding to the 8 tokens are 0.05, 0.08, 0.72, 0.68, 0.91, 0.85, 0.65 and 0.70 from left to right. It can be observed that the discretized step sizes of the first two tokens "tou" and "biao" are obviously smaller, both lower than 0.10. Starting from the third token "zi", the discretized step sizes increase significantly, concentrated in the range from 0.65 to 0.91. This difference intuitively reflects the core characteristic of the selective state space mechanism: tokens irrelevant to the semantics of qualification and performance will generate smaller discretized step sizes, so that the state vector mainly retains the information accumulated in the previous text and less integrates the current input when processing these tokens; tokens closely related to the semantics of qualification and performance will generate larger discretized step sizes, so that the state vector fully absorbs the semantic information carried by the current token. Figure 1The lower left and lower right corners respectively display textual descriptions of the two scenarios described above. The specific numerical value is marked above the discretization step size bar for each marker, with the bar height proportional to the value. Arrows extend upwards from the rectangle of each marker to the corresponding discretization step size bar, indicating that the discretization step size is dynamically generated from the input content of each marker. Above the discretization step size area is the state vector recursion chain. The state vector recursion chain consists of 8 circular nodes, labeled from left to right as follows: to This represents the updated state vector corresponding to each of the 1st to 8th tags. Adjacent circular nodes are connected by horizontal arrows, with the arrows pointing from left to right, indicating that the state vectors are recursively propagated according to the order of the tag sequence. A dashed arrow extends upwards from each discretization step size bar, pointing to the corresponding circular node of the state vector, representing the ratio between the amount of previous information retained and the amount of current input incorporated during the state update process. Above the state vector recursion chain is the output layer, containing 8 rectangles, labeled from left to right as follows: to , representing the context encoding vector output after each tag is processed by the selective state-space computation layer. An arrow pointing upwards from the circular node of each state vector to the corresponding output rectangle indicates that the context encoding vector is obtained by transforming the updated state vector through a fixed output mapping vector. Figure 1 The state recursion formula is marked at the top. ,in For the first The updated state vector corresponding to each label, For the first Each label-specific discrete state transition vector, This is the updated state vector corresponding to the previous label. For element-wise multiplication, For the first Each label-specific input modulation vector, For the first The component value of each marker on the current channel. The entire graph presents the complete data flow from bottom to top, including marker input, selective gating parameter generation, state recursion, and encoded output.

[0043] After encoding all tags, the next step is to filter the qualification and performance reference paragraphs. The goal of this filtering is to identify paragraphs from the entire tender document that contain descriptions of qualification and performance references, such as paragraphs containing statements like "Our company possesses ISO9001 quality management system certification" or "Completed the XX municipal engineering construction project in 2022." For each paragraph-level subsequence, the average of the context encoding vectors of all tags along the sequence length dimension is taken to obtain the paragraph representation vector for that paragraph. This averaging operation unifies paragraphs of different lengths into a fixed-length representation of the same dimension, facilitating subsequent classification. The paragraph representation vector has a dimension of 768, the same as the dimension of the context encoding vector.

[0044] The paragraph representation vector is fed into the paragraph classification head. The paragraph classification head consists of a linear transformation and a sigmoid activation function concatenated. The linear transformation maps the 768-dimensional paragraph representation vector to a single scalar value, and the sigmoid activation function compresses the scalar value to the range of 0 to 1. The output is the judgment score for whether the paragraph belongs to the qualification / performance reference paragraph. The closer the judgment score is to 1, the higher the probability that the paragraph contains qualification / performance reference descriptions; the closer the judgment score is to 0, the higher the probability that the paragraph is unrelated to qualification / performance references. The judgment score is compared with a preset positioning threshold. Paragraphs with judgment scores exceeding the preset positioning threshold are marked as qualification / performance reference paragraphs. In this embodiment, the preset positioning threshold is set to 0.5, which can be adjusted according to the actual content distribution of the tender document: when the proportion of qualification / performance related paragraphs in the tender document is low and a high recall rate is required, the preset positioning threshold can be reduced to the range of 0.3 to 0.4; when high accuracy is required and a small number of marginal paragraphs can be missed, the preset positioning threshold can be increased to the range of 0.6 to 0.7.

[0045] In an optional implementation, the paragraph classification head can also be replaced with a two-layer fully connected network, that is, first mapping the 768-dimensional paragraph representation vector to a 256-dimensional intermediate representation, then processing it with the ReLU activation function and mapping it to a scalar value, and then outputting a judgment score through the sigmoid activation function, so as to improve the ability to discriminate complex paragraph semantics.

[0046] After filtering the paragraphs to obtain the qualification and performance references, entity boundary annotation is only performed on these paragraphs, and other paragraphs in the tender document are not processed. Entity boundary annotation adopts an extended form of the BIO sequence annotation paradigm, classifying each tag into one of nine categories. These nine categories are: qualification name start position, qualification name internal position, certificate number start position, certificate number internal position, project name start position, project name internal position, project completion time start position, project completion time internal position, and non-entity position. The start position tag indicates the first tag of a certain type of entity, the internal position tag indicates subsequent tags within that type of entity, and the non-entity position tag indicates a tag that does not belong to any entity. This annotation method is used because multiple entities of the same type may appear simultaneously in the same paragraph of the tender document (e.g., one paragraph describes two different qualification certifications). Distinguishing between start and internal positions accurately defines the start and end boundaries of each entity, preventing adjacent entities of the same type from being incorrectly merged.

[0047] The context encoding vector of each tag in the qualification and performance citation paragraph is sent to the entity annotation header. The entity annotation header is a linear transformation that maps the 768-dimensional context encoding vector to a 9-dimensional score vector. Each component of the 9-dimensional score vector corresponds to the score of one of the above 9 categories. The category with the highest score is taken as the entity annotation result for that tag. After all tags are annotated, tags belonging to the same entity category and satisfying the adjacent position constraint are merged into one entity. The specific method for determining the adjacent position constraint is: a series of consecutive internal bit tags of the same entity category following a start bit tag, together with the start bit tag, constitute a complete entity. When the continuous sequence of internal bit tags is interrupted by a non-entity bit tag or the start bit tag of another category, the merging process terminates. In actual tender documents, there may be OCR parsing errors or special formats that cause a small number of non-entity bit tags to appear within the same entity. To improve robustness, in an optional implementation, a tolerance window can be set. When the number of non-entity bit tags between internal bit tags of the same category does not exceed 2, the tags on both sides of the interval are still merged into the same entity.

[0048] After merging, a set of qualification and performance entities is obtained. Each entity in the qualification and performance entity set contains three attributes: entity category (one of qualification name, certificate number, project name, or project completion time), entity text (composed of the original text corresponding to all merged tags), and paragraph number (inherited from the paragraph number of the paragraph-level subsequence containing the first tag of this entity). The paragraph number will be used in the subsequent association assembly stage to limit the pairing scope, ensuring that only entities from the same paragraph are associated, and avoiding incorrect association of different qualification and performance information described in different paragraphs.

[0049] Let's take a real tender document as an example. Assume a tender document contains 86 pages. After text parsing, it is concatenated into a continuous token sequence containing 42,000 tokens, divided into 327 paragraph-level subsequences. After encoding by a Mamba sequence encoder, 42,000 context encoding vectors are obtained. Paragraph representation vectors are calculated for each of the 327 paragraphs and fed into the paragraph classification header. Assuming that 18 paragraphs have a judgment score exceeding a preset positioning threshold of 0.5, these 18 paragraphs are marked as qualification / performance reference paragraphs. Entity boundary annotation is performed on all tokens in these 18 paragraphs, ultimately resulting in a qualification / performance entity set, for example containing 12 qualification name entities, 10 certificate number entities, 8 project name entities, and 8 project completion time entities. This qualification / performance entity set will serve as input for subsequent hyperbolic space embedding and association assembly stages.

[0050] For training the paragraph classification head and entity annotation head, the training data consists of manually annotated tender document samples. Annotators mark each paragraph in the tender documents to indicate whether it is a qualification / performance reference paragraph, and also mark the entity category of each tag within the qualification / performance reference paragraph. In this embodiment, the training dataset contains 500 tender documents from different industries and formats, covering three typical bidding scenarios: engineering procurement, goods procurement, and service procurement. The paragraph classification head and entity annotation head are jointly trained with the Mamba sequence encoder, using the cross-entropy loss function as the optimization objective. The AdamW optimizer is used during training, with an initial learning rate of 0.00005, a batch size of 4, and 20 training epochs. In an optional implementation, the Mamba sequence encoder can be pre-trained on a general text corpus to obtain basic language representation capabilities, and then fine-tuned on the tender document annotation data to reduce the dependence on the amount of annotation data.

[0051] After step 1, a set of qualification and performance entities has been extracted from the tender documents. Each entity in the set contains an entity category, entity text, and its corresponding paragraph number. However, these entities are still isolated at this point, and no hierarchical relationship has been established between them. In actual tender documents, a complete qualification and performance record is usually composed of multiple entities of different categories—for example, "ISO9001 Quality Management System Certification" (qualification name) and "XXX-2023-001" (certificate number) jointly describe one qualification certificate, and "XX Municipal Engineering Construction Project" (project name) and "2022" (project completion time) jointly describe one performance record. Furthermore, there may be further hierarchical relationships between qualification certificates and performance records. If these entities are not correctly assembled into structured association groups, the subsequent automatic data entry will be unable to determine which certificate number corresponds to which qualification name, or which completion time corresponds to which project.

[0052] In industry practice, qualification names naturally exhibit a hierarchical structure. Taking quality management as an example, "Management System Certification" is the higher-level category, "ISO 9001 Quality Management System Certification" is its lower-level specific certification, and "ISO 9001:2015 Quality Management System Certification" is a more granular sub-level. Similar hierarchical relationships are widely present in various systems such as construction qualifications, environmental certification, and information security certification. Euclidean space has inherent limitations in expressing this hierarchical structure: the distance between any two points in Euclidean space increases uniformly with dimension, failing to naturally reflect the characteristic that "the closer to a leaf node in a hierarchy, the denser the nodes of the same level should be." Hyperbolic space, on the other hand, has the geometric characteristic of exponential volume growth; that is, as the distance from the origin increases, the number of points it can accommodate increases exponentially, which highly matches the characteristic of the number of nodes increasing layer by layer from the root to the leaf in a tree-like hierarchical structure. Therefore, this embodiment chooses to embed entities into hyperbolic space, utilizing the natural adaptability of hyperbolic space to hierarchical structures to assist in subsequent associative assembly.

[0053] Specifically, this embodiment uses the Poincaré sphere model as a concrete implementation of hyperbolic space. The Poincaré sphere model represents hyperbolic space as the interior of an open sphere, with the curvature set to -1 in this embodiment. The advantage of the Poincaré sphere model lies in its boundary effect—the closer to the sphere's boundary, the greater the geodesic distance corresponding to a unit change in coordinates. This means that deeper (more concrete) entities in the hierarchy tree are naturally pushed towards the edge of the sphere, while shallower (more abstract) entities cluster in the region near the origin, thus intuitively reflecting the hierarchical depth of the qualification system in the embedded space.

[0054] Mapping entities from the Euclidean representation space to the hyperbolic embedding space within the Poincaré sphere requires exponential mapping. For each entity in the qualification and performance entity set, the mean of the context encoding vectors of all corresponding tags is first taken along the sequence length dimension to obtain the entity's Euclidean feature vector. This averaging operation unifies entity texts of different lengths into a fixed-dimensional representation; the dimension of the Euclidean feature vector is the same as the dimension of the context encoding vector, which is 768 dimensions in this embodiment. The Euclidean feature vector resides in the Euclidean representation space and carries the semantic information of the entity within the full-text context.

[0055] The execution process of the exponential mapping is as follows. Taking the origin of the Poincaré sphere model as the base point, the Euclidean norm of the Euclidean eigenvectors is first calculated. ,in These are Euclidean eigenvectors. for The sum of the squares of all components is taken as the arithmetic square root. The unit direction vector is obtained by dividing the Euclidean eigenvector by its Euclidean norm. The unit direction vector retains the directional information of the Euclidean eigenvectors and has a norm of 1. Then, the Euclidean norm is... Input the hyperbolic tangent function to obtain the hyperbolic radial component. ,in It is a hyperbolic tangent function, and its output range is an open interval. Since the Poincaré sphere model requires that the norm of interior points be strictly less than 1, the hyperbolic tangent function maps any positive real number to an open interval between 0 and 1, thus ensuring that the mapping result falls inside the Poincaré sphere without touching the boundary. Multiplying the hyperbolic radial component element-wise with the unit direction vector, i.e. ,in Let this be the hyperbolic embedding vector of the entity inside the Poincaré sphere. It lies in a hyperbolic embedding space, with the same dimension as the Euclidean eigenvector, which is 768-dimensional, but the geometric space it occupies changes from a flat Euclidean space to a hyperbolic space with a curvature of negative 1.

[0056] In an optional implementation, the base point for the exponential mapping is not limited to the origin of the Poincaré sphere; other base points can be chosen, such as the position after mapping the mean of the Euclidean eigenvectors of all entities in the competency performance entity set onto the Poincaré sphere, to make the distribution center of the embedding result more closely match the actual data. Furthermore, the curvature of the Poincaré sphere can also be set to other negative values, such as -0.5 or -2. The larger the absolute value of the curvature, the stronger the "bending" of the hyperbolic space, and the higher the resolution of deep hierarchical structures, but the stability of numerical computation may decrease. In this embodiment, a curvature of -1 is chosen as the default configuration to balance numerical stability and hierarchical expressive power.

[0057] refer to Figure 2 , Figure 2The main body is a circular area, representing the projection of a Poincaré sphere model with a curvature of -1 onto a two-dimensional plane. The boundary of the circle is the boundary of the Poincaré sphere, and a cross symbol is marked at the center to represent the origin. Inside the circular area are eight entity markers, divided into two groups. The first group contains four entity markers, concentrated in the upper right part of the circular area, representing the qualification name entity "ISO9001 certification", the certificate number entity "XXX-2023-001", the project name entity "XX Municipal Engineering", and the project completion time entity "2022" extracted from paragraph number 23. The second group also contains four entity markers, concentrated in the lower left part of the circular area, representing the qualification name entity "ISO14001 certification", the certificate number entity "YYY-2023-005", the project name entity "YY Environmental Engineering", and the project completion time entity "2023". The four entity categories are distinguished by different shaped markers: qualification name entities are marked with a circle, certificate number entities with a square, project name entities with a triangle, and project completion time entities with a diamond. Each entity marker is connected to a text label box outside the circular area by a leader line, with the label box indicating the specific text content of the entity. The associative assembly process is represented in the diagram by two types of lines. The first type is solid line connections, representing the results of the first stage qualification certificate pairing and the second stage performance record pairing. In the first group, "ISO9001 certification" and "XXX-2023-001" are connected by a solid line to form a qualification certificate pair, and "XX municipal engineering" and "2022" are connected by a solid line to form a performance record pair; in the second group, "ISO14001 certification" and "YYY-2023-005" are connected by a solid line to form a qualification certificate pair, and "YY environmental engineering" and "2023" are connected by a solid line to form a performance record pair. Solid lines indicate that the geodesic distance between the paired entities on the Poincaré sphere is the smallest, thus classifying them as the most likely pairing. The second type uses dashed lines, representing the association between qualification certificate pairs and performance record pairs in the third stage of cross-category association assembly. In the first group, "ISO9001 certification" and "XX municipal engineering" are connected by dashed lines, as are "XXX-2023-001" and "2022," corresponding to the first and second sub-item distances in the cross-category association score, respectively. The specific values ​​of the geodesic distances are marked near the midpoints of the four lines in the first group: geodesic distances for qualification certificate pairings. =3.2, geodesic distance of performance record pairing =2.8, distance of the first component =5.1, distance of the second component =4.6, cross-category association score is The value is below the preset association threshold of 15.0, therefore the qualification certificate pairs and performance record pairs in the first group are merged into qualification-performance association group A. The diagram uses dotted ellipses to circle all four entity markers in the first and second groups respectively, and labels "Association Group A" and "Association Group B" next to the ellipses to visually represent the scope of the two qualification-performance association groups. Each association group contains four entities: one qualification name, one certificate number, one project name, and one project completion time.

[0058] After completing the hyperbolic embedding of all entities, for all entities belonging to the qualification name category in the qualification performance entity set, the hierarchical relationships are further determined and a qualification hierarchy tree is constructed using a pre-defined qualification classification hierarchy knowledge base. The qualification classification hierarchy knowledge base is a pre-organized structured data table that records the hierarchical relationships between qualification names at various levels in the industry qualification system. Taking the construction industry as an example, the knowledge base records that "General Contracting Qualification" includes sub-qualifications such as "General Contracting of Building Construction" and "General Contracting of Municipal Public Works Construction," and each sub-qualification can be further subdivided into special grade, first grade, second grade, and third grade. The knowledge base assigns a unique hierarchy number to each qualification name. The value of the hierarchy number reflects the depth of the qualification in the system tree—the smaller the hierarchy number, the closer it is to the root node (more abstract), and the larger the hierarchy number, the closer it is to the leaf node (more specific).

[0059] The process of constructing a qualification hierarchy tree based on hierarchical numbers is as follows: A virtual top-level node is set as the root node of the qualification hierarchy tree. This virtual top-level node does not correspond to any actual qualification name entity; it serves only as the unified entry point for the tree structure. Each qualification name entity in the qualification performance entity set is attached to its corresponding level in the qualification hierarchy tree according to its hierarchical number. Entities with the same hierarchical number are treated as sibling nodes, and entities with smaller hierarchical numbers become the parent nodes of entities with larger hierarchical numbers. When a certain extracted qualification name entity is not included in the qualification classification hierarchy knowledge base, that entity is directly attached under the virtual top-level node as a first-level child node.

[0060] After the hierarchical tree is constructed, the hierarchical distance value is calculated for each edge from a parent node to a child node. The hierarchical distance is calculated by taking the hyperbolic embedding vectors of the parent and child nodes and calculating the geodesic distance between them on a Poincaré sphere. The geodesic distance between two points on a Poincaré sphere is the shortest path length connecting these two points in hyperbolic space. For two points inside the Poincaré sphere... and Geodesic distance The calculation method is as follows ,in and These are the hyperbolic embedding vectors for two points. It is the square of the Euclidean norm of the difference vector between the two vectors. and Let be the squares of the Euclidean norms of the hyperbolic embedding vectors at the two points. The hierarchical distance is an inverse hyperbolic cosine function. A smaller hierarchical distance value indicates that the parent and child nodes are closer in the hyperbolic embedding space, and their corresponding semantic levels in the qualification system are closer; a larger hierarchical distance value indicates a greater hierarchical span between the two in the qualification system. The hierarchical distance value plays an auxiliary filtering role in the subsequent association assembly stage—when the hierarchical distance value of a qualification name entity is too large relative to its parent node, it means that the entity may have been incorrectly classified into an inappropriate hierarchical position, and further verification can be performed based on this.

[0061] In an optional implementation, when the bidding scenario involves many industry types and the preset qualification classification hierarchy knowledge base is not comprehensive, the hierarchical relationship can be automatically inferred solely based on the geodesic distance between hyperbolic embedding vectors. Specifically, the Euclidean norm of the hyperbolic embedding vectors of all qualification name entities is used as an approximate indicator of hierarchy depth (a larger norm indicates proximity to the Poincaré sphere boundary, corresponding to a more specific hierarchy). Entities with smaller norms are considered potential parent nodes of entities with larger norms, and the closest parent-child relationships are then selected using geodesic distance. This method does not rely on an external knowledge base, but the accuracy of the inference depends on the quality of the hyperbolic embedding vectors.

[0062] After completing the hyperbolic embedding and qualification hierarchy tree construction, the association assembly stage begins. The goal of association assembly is to assemble isolated entities extracted from the same paragraph into structured qualification and performance association groups, with each group representing a complete qualification and performance record. Association assembly is limited to the same paragraph number range because, in the writing conventions of tender documents, the same qualification and performance information is usually described in its entirety within the same paragraph, and referencing different qualification and performance information across paragraphs is quite common. Without limiting the paragraph range, incorrect associations are easily generated.

[0063] The associated assembly is performed in three stages in sequence.

[0064] The first stage is qualification certificate pairing. Under the same paragraph number, each qualification name entity needs to be paired with the most likely corresponding certificate number entity. The pairing is based on the geodesic distance in the hyperbolic embedding space. During the encoding process of the Mamba sequence encoder, tags that are close in position and semantically related in the text will generate similar context encoding vectors. After mean aggregation and exponential mapping, semantically related entities also tend to be distributed in close positions in the Poincaré sphere. Therefore, for each qualification name entity, the geodesic distance between the hyperbolic embedding vector of the qualification name entity and the hyperbolic embedding vectors of all certificate number entities in the same paragraph is calculated. The certificate number entity with the smallest geodesic distance is selected to pair with the qualification name entity, forming a qualification certificate pair. When there is only one certificate number entity in the same paragraph, the certificate number entity is directly paired with the qualification name entity in the paragraph. When there are multiple qualification name entities and multiple certificate number entities in the same paragraph, each qualification name entity independently selects the nearest certificate number entity. The same certificate number entity is allowed to be selected by multiple qualification name entities, but in actual bidding documents, the same certificate number usually corresponds to only one qualification name.

[0065] The second stage is performance record pairing. The pairing logic is the same as in the first stage. Under the same paragraph number, for each project name entity, the geodesic distance between the hyperbolic embedding vector of the project name entity and the hyperbolic embedding vector of all project completion time entities in the same paragraph is calculated. The project completion time entity with the smallest geodesic distance is selected to pair with the project name entity to form a performance record pair.

[0066] Phase 3 is cross-category association assembly, which aims to further merge the qualification certificate pairs generated in Phase 1 and the performance record pairs generated in Phase 2 into complete qualification-performance association groups. Cross-category association assembly requires determining which qualification certificate pair and which performance record pair describe the same qualification-performance record. This determination is based on the cross-category association score. The cross-category association score is calculated as follows: The geodesic distance on the Poincaré sphere is calculated between the hyperbolic embedding vector of the qualification name entity in the qualification certificate pair and the hyperbolic embedding vector of the project name entity in the performance record pair, respectively, as the first sub-item distance; the geodesic distance on the Poincaré sphere is calculated between the hyperbolic embedding vector of the certificate number entity in the qualification certificate pair and the hyperbolic embedding vector of the project completion time entity in the performance record pair, as the second sub-item distance. The first and second sub-item distances are summed to obtain the cross-category association score. The reason for using the sum of two sub-distances instead of just a single distance is that the semantic correlation between the qualification name and the project name, and the semantic correlation between the certificate number and the project completion time, respectively reflect different dimensions of attribution. Taking both into account can more comprehensively determine whether the qualification certificate pair and the performance record pair belong to the same qualification and performance record.

[0067] The cross-category association score is compared with a preset association threshold. Qualification certificate pairs and performance record pairs with cross-category association scores lower than the preset threshold are merged into one qualification-performance association group. In this embodiment, the preset association threshold is set to 15.0, a value determined based on the typical range of geodesic distances on a Poincaré sphere. When the curvature is negative 1 and the embedding dimension is 768, the geodesic distance between semantically highly associated entity pairs is typically in the range of 2.0 to 8.0, while the geodesic distance between semantically unrelated entity pairs is typically above 20.0. After summing the two sub-distances, highly associated combinations typically fall within the range of 4.0 to 16.0. A threshold setting of 15.0 can filter out most erroneous combinations while ensuring recall. In an optional implementation, the preset association threshold can be adaptively adjusted based on the precision-recall curve on the validation set. Each qualification-performance association group contains four entities: one qualification name, one certificate number, one project name, and one project completion time.

[0068] In an optional implementation, when only qualification certificate pairs are extracted from the same paragraph but no performance record pairs are extracted (i.e., the paragraph only describes qualification information but does not mention specific project performance), or only performance record pairs are extracted but no qualification certificate pairs are extracted, the incomplete pair can be independently treated as a qualification-performance association group. The missing entity fields in this association group are marked as null values, and only fields with values ​​are filled in during subsequent automatic filling.

[0069] After the association assembly is completed, the compliance prediction calibration stage begins. The qualification and performance association groups generated in the association assembly stage are automatically assembled based on hyperbolic geodesic distances, inevitably resulting in some erroneous associations. If erroneous association information is directly entered into the bidding form, the reported information will be inconsistent with the actual qualification and performance, potentially causing serious problems in the review stage. The traditional approach is to set a fixed distance threshold for truncation, but a fixed threshold cannot provide a probabilistic guarantee of reliability—that is, it cannot quantitatively answer the key question of "what is the minimum proportion of correct associations in the association groups filtered through this threshold?" The compliance prediction method precisely solves this pain point. Compliance prediction is a distribution-independent statistical inference framework. Its core property is that, under the assumption that only the calibration data and test data are independently and identically distributed, it can guarantee that the coverage of correct samples in the filtered samples is not lower than the preset target level. This guarantee holds strictly with a finite sample size and does not depend on the specific form of the underlying model or the data distribution assumptions.

[0070] The process for calculating the inconsistency score for each qualification and performance association group is as follows: The Euclidean eigenvectors of the four entities (qualification name, certificate number, project name, and project completion time) within the qualification and performance association group are concatenated along their vector dimensions to form a combined vector. Euclidean eigenvectors are used here instead of hyperbolic embedding vectors because vector concatenation in Euclidean space has a clear geometric meaning—the concatenated combined vector still exists in the Euclidean representation space, and its dimension is the sum of the dimensions of the four entities' Euclidean eigenvectors, i.e. If hyperbolic embedding vectors are used for concatenation, the concatenated vectors will no longer be valid points in the Poincaré sphere, making meaningful distance calculations impossible in hyperbolic space.

[0071] The combined vector is fed into the calibration scoring function. The calibration scoring function is a linear transformation pre-trained based on correctly labeled and incorrectly labeled association groups. This linear transformation maps the 3072-dimensional combined vector to a single scalar value. During training, correctly identified qualification / performance association groups are used as positive samples, and incorrectly identified qualification / performance association groups are used as negative samples. The parameters of the linear transformation are optimized by minimizing the binary cross-entropy loss function, making the output values ​​of positive samples tend to be smaller and the output values ​​of negative samples tend to be larger. After training, the scalar value output by the calibration scoring function is the inconsistency score of the qualification / performance association group. A larger inconsistency score indicates that the association group deviates further from the confirmed correct association pattern, and the higher the probability of incorrect association; a smaller inconsistency score indicates that the association group is closer to the correct association pattern, and the higher the probability of correct association.

[0072] In an optional implementation, the calibration scoring function can also use a two-layer fully connected network instead of a linear transformation. That is, the 3072-dimensional combined vector is first mapped to a 512-dimensional intermediate representation, processed by the ReLU activation function, and then mapped to a single scalar value to enhance the ability to discriminate complex association patterns. Furthermore, the calculation of the inconsistency score is not limited to the concatenation method based on Euclidean eigenvectors. In an optional implementation, the hyperbolic embedding vectors of the four entities can be transformed back to tangent vectors in Euclidean tangent space through logarithmic mapping before concatenation. Logarithmic mapping is the inverse operation of exponential mapping; the transformed tangent vectors remain in Euclidean space and retain the distance information in hyperbolic space.

[0073] After the inconsistency scores are calculated, a calibration threshold needs to be determined. The determination of the calibration threshold depends on a pre-built calibration sample pool. The calibration sample pool consists of qualification and performance association groups that have been manually verified as correctly or incorrectly associated, along with their corresponding inconsistency scores. The calibration sample pool is constructed as follows: before the system goes live or during periodic maintenance, several qualification and performance association groups are extracted from completed bidding projects. Manual verification personnel verify the correct attribution relationships of the four entities in each association group, and the verification results are stored along with the inconsistency scores of each association group. In this embodiment, the calibration sample pool contains 200 verified qualification and performance association groups, of which approximately 160 are confirmed as correctly associated and approximately 40 are confirmed as incorrectly associated.

[0074] The process for determining the calibration threshold is as follows: All 200 calibration samples in the calibration sample pool are arranged in ascending order of their inconsistency scores to form an ordered sequence. The ordered sequence is traversed from smallest to largest, and the inconsistency score at the position where the proportion of samples with inconsistency scores not exceeding the current position in the ordered sequence is first not lower than the preset target coverage rate is taken as the calibration threshold. In this embodiment, the preset target coverage rate is set to 0.90, meaning that at least 90% of the association groups that pass the screening are correctly associated. Taking 200 calibration samples as an example, when the preset target coverage rate is 0.90, it is necessary to find the 180th position in the ordered sequence (i.e.,... The non-consistency score of the 180th position in the ordered sequence is used as the calibration threshold. If the non-consistency score at the 180th position is exactly equal to the non-consistency score at the 181st position, then the value at the 181st position is taken to ensure that the coverage ratio is strictly not less than 0.90. According to the theoretical guarantee of compliance prediction, when the calibration sample and the actual qualification performance association group to be calibrated are independently and identically distributed, the calibration threshold determined in this way can ensure that the proportion of the correct association group that passes calibration in the new association group is not less than the preset target coverage rate in the expected sense.

[0075] In an optional implementation, the preset target coverage rate can be adjusted according to the risk tolerance of the business scenario. When applied to scenarios with extremely high accuracy requirements, such as government procurement, the preset target coverage rate can be set to 0.95 or even 0.99. In this case, the calibration threshold is lowered accordingly, the screening is more stringent, the number of related groups that pass calibration is reduced, but the accuracy rate is higher. When applied to scenarios with higher tolerance, such as internal procurement, the preset target coverage rate can be set to 0.80 to 0.85 in exchange for more related groups being automatically filled in.

[0076] The inconsistency score of each qualification and performance association group is compared with the calibration threshold. When the inconsistency score does not exceed the calibration threshold, the qualification and performance association group is determined to have passed the compliance prediction calibration and is marked as a reusable association group. When the inconsistency score exceeds the calibration threshold, the association group fails the calibration and is not entered into the automatic reporting process. Instead, it is marked as an association group requiring manual review and is sent back to the supplier for manual confirmation.

[0077] After compliance prediction calibration, the entity information in the reusable association groups enters the auto-fill stage. The qualification name, certificate number, project name, and project completion time from each reusable association group are automatically filled into the corresponding qualification name, certificate number, project name, and project completion time fields in the qualification and performance reporting form for this tender. The filling process is achieved through a preset mapping relationship between form field names and entity categories—entity text for the qualification name category is filled into the qualification name field, entity text for the certificate number category is filled into the certificate number field, entity text for the project name category is filled into the project name field, and entity text for the project completion time category is filled into the project completion time field.

[0078] Simultaneously, a reference calibration record is generated for each reusable association group in the form. The reference calibration record contains two pieces of information: the inconsistency score and the calibration threshold for that reusable association group. The purpose of setting up reference calibration records is to provide a traceable and reliable basis for subsequent review processes—reviewers can intuitively judge the reliability of the automatically generated information by viewing the difference between the inconsistency score and the calibration threshold in the reference calibration record: association groups with inconsistency scores far below the calibration threshold are highly reliable, while association groups with inconsistency scores close to the calibration threshold are relatively less reliable. Reviewers can then decide whether to focus on verifying specific association groups based on this information.

[0079] In an optional implementation, the autofill process can be further enhanced with a supplier confirmation step. After autofill is complete, the completed form is previewed and shown to the supplier, who then confirms or modifies the autofill results for each reusable group. When the supplier modifies the content of a field, the modified content is recorded along with the original autofilled content for subsequent updates to the calibration sample pool. The supplier's confirmation can also be fed back into the calibration sample pool as new annotation samples for the next update of calibration thresholds. Through this feedback mechanism, the calibration sample pool continuously expands with usage time, and the determination of calibration thresholds becomes increasingly accurate.

[0080] The following is a complete explanation using Supplier A's participation in a procurement project bidding as an example. In Step 1, a set of qualification and performance entities was extracted from Supplier A's 86-page PDF bid document. Assume the following entities were extracted from a single qualification and performance reference paragraph (paragraph number 23): qualification name entity "ISO9001 Quality Management System Certification", certificate number entity "XXX-2023-001", project name entity "XX Municipal Engineering Construction Project", and project completion time entity "2022". In Step 2, the four entities are respectively aggregated using the mean to obtain Euclidean eigenvectors, and then subjected to exponential mapping to obtain hyperbolic embedding vectors. In the first stage of association assembly, "ISO9001 Quality Management System Certification" and "XXX-2023-001" are in the same paragraph. The geodesic distance between their hyperbolic embedding vectors is calculated to be 3.2. Since this paragraph contains only this one certificate number entity, they are directly paired as a qualification certificate pair. In Phase 2, the geodesic distance between "XX Municipal Engineering Construction Project" and "2022" is 2.8, pairing them as performance record pairs. In Phase 3, the geodesic distance between the hyperbolic embedding vector of "ISO9001 Quality Management System Certification" in the qualification certificate pair and the hyperbolic embedding vector of "XX Municipal Engineering Construction Project" in the performance record pair is calculated to be 5.1 (first sub-item distance). The geodesic distance between the hyperbolic embedding vector of "XXX-2023-001" and the hyperbolic embedding vector of "2022" is calculated to be 4.6 (second sub-item distance). The cross-category association score is... The value is below the preset association threshold of 15.0, therefore the qualification certificate pair and the performance record pair are merged into one qualification-performance association group.

[0081] In step 3, the Euclidean eigenvectors of the four entities within the qualification and performance association group are concatenated into a 3072-dimensional combined vector, which is then fed into the calibration scoring function to obtain a non-consistency score of 0.12. The non-consistency score at the 180th position in the ordered sequence of 200 samples in the calibration sample pool is 0.35, meaning the calibration threshold is 0.35. Since 0.12 does not exceed 0.35, this association group passes the compliance prediction calibration and is marked as a reusable association group. Subsequently, "ISO9001 Quality Management System Certification" is entered into the qualification name field, "XXX-2023-001" into the certificate number field, "XX Municipal Engineering Construction Project" into the project name field, and "2022" into the project completion time field. Simultaneously, a reference calibration record is generated, containing the non-consistency score of 0.12 and the calibration threshold of 0.35, for review by the assessors.

[0082] refer to Figure 3 , Figure 3The horizontal axis represents the distance threshold, ranging from 2 to 20, which is the distance cutoff value used to determine whether entity pairs belong to the same association group during the association assembly stage; the vertical axis represents the F1 value, ranging from 0.40 to 1.00, which is the comprehensive evaluation index of the association assembly results. The higher the F1 value, the better the overall performance of the association assembly in terms of precision and recall. Figure 3 The graph contains two curves, corresponding to two different embedding space schemes. The first curve represents the scheme using the hyperbolic embedding space defined by the Poincaré sphere model. In this scheme, the Euclidean feature vectors of entities are projected onto the hyperbolic embedding space via an exponential mapping and then associated and assembled based on geodesic distance. The second curve represents the scheme using Euclidean space cosine distance. In this scheme, the Euclidean feature vectors of entities are directly associated and assembled in Euclidean space based on cosine distance without undergoing hyperbolic mapping, serving as a baseline for comparison. Both curves show a trend of first rising and then falling. This is because when the distance threshold is too small, a large number of correct entity pairs are missed due to distances exceeding the threshold, resulting in low recall. When the distance threshold is too large, a large number of incorrect entity pairs are misclassified due to distances below the threshold, resulting in low precision. Both extreme cases lead to low F1 scores. The hyperbolic space scheme reaches its optimal F1 score of 0.86 at a distance threshold of 8, as indicated by the text "Optimal F1 = 0.86, Threshold = 8" in the graph. The curve for the Euclidean space scheme reaches its optimal F1 value of 0.76 at a distance threshold of 10, as indicated by the text "Optimal F1 = 0.76, Threshold = 10" in the figure. A comparison of the two curves shows that, under the same experimental conditions, the optimal F1 value of the hyperbolic space scheme is 0.10 higher than that of the Euclidean space scheme. Furthermore, the hyperbolic space scheme maintains an F1 value above 0.80 across a wider range of distance thresholds from 6 to 12, while the Euclidean space scheme only reaches above 0.70 within a narrower range of distance thresholds from 8 to 12. This difference validates the effectiveness of embedding entities into the hyperbolic space of a Poincaré sphere: the exponential volume growth characteristic of hyperbolic space causes semantically closely related entities to cluster together more closely after embedding, while semantically unrelated entities become more distant, thus achieving clearer distinction boundaries when determining associations based on distance thresholds.

[0083] refer to Figure 4 , Figure 4 The horizontal axis represents the preset target coverage rate, ranging from 0.80 to 0.99, which represents the expected coverage level set during the compliance prediction calibration phase; the vertical axis represents the ratio, ranging from 0.55 to 1.00. Figure 4The data includes two curves. The first curve is the correlation accuracy curve, representing the proportion of correctly correlated associations among reusable association groups that pass compliance prediction calibration. The second curve is the automatic reporting rate curve, representing the proportion of reusable association groups that pass compliance prediction calibration out of all qualification performance association groups, reflecting the level of automation. As the preset target coverage gradually increases from 0.80 to 0.99, the correlation accuracy curve shows a downward trend, decreasing from 0.97 to 0.82. This is because the higher the preset target coverage, the larger the calibration threshold, allowing for a wider range of association groups to pass calibration, but also increasing the number of erroneous association groups, leading to a decrease in accuracy. The automatic reporting rate curve, on the other hand, shows an upward trend, rising from 0.65 to 0.97. This is because the higher the preset target coverage, the more lenient the calibration conditions, allowing more association groups with inconsistency scores below the calibration threshold to pass calibration, thus increasing the proportion of automatically reported association groups. The two curves show an inverse relationship, reflecting the inherent trade-off between automation and accuracy. Figure 4 A vertical reference line is marked at the preset target coverage rate of 0.90, with the note "Recommended working point, coverage = 0.90" next to it. At this working point, the correlation accuracy is 0.94, and the auto-fill ratio is 0.80, meaning that 80% of all qualification and performance correlation groups can be automatically filled into the bidding form without manual operation, and 94% of the automatically filled correlation groups are correctly correlated. When the application scenario has extremely high accuracy requirements, the preset target coverage rate can be reduced to the range of 0.80 to 0.85 to obtain a correlation accuracy rate of 0.96 to 0.97, but the auto-fill ratio will decrease to 0.65 to 0.72; when the application scenario prioritizes efficiency and can tolerate a certain error rate, the preset target coverage rate can be increased to the range of 0.95 to 0.99 to obtain an auto-fill ratio of 0.88 to 0.97, but the correlation accuracy rate will decrease to 0.82 to 0.91.

[0084] This invention is not limited to the specific embodiments described above. The invention extends to any new feature or combination disclosed in this specification, as well as any new method or process step or combination disclosed herein.

Claims

1. A method for automatic reuse of bidding qualification and performance records based on a state-space model, characterized in that, Includes the following steps: Step 1: Perform text parsing on the tender document, concatenate the text content into a continuous tag sequence and divide it into several paragraph-level subsequences, input the continuous tag sequence into a sequence encoder built based on a selective state space mechanism to obtain the context encoding vector corresponding to each tag, filter out the qualification and performance reference paragraphs based on the context encoding vectors, and perform entity boundary annotation on the qualification and performance reference paragraphs to obtain the qualification and performance entity set. Step 2: Project the Euclidean eigenvector of each entity in the qualification and performance entity set onto the hyperbolic embedding space defined by the Poincaré sphere model through exponential mapping to obtain a hyperbolic embedding vector. Perform association assembly on the entities within the same paragraph based on the geodesic distance between the hyperbolic embedding vectors to obtain the qualification and performance association group. Step 3: Calculate the inconsistency score for each qualification and performance association group and determine the calibration threshold. Mark the qualification and performance association groups whose inconsistency scores do not exceed the calibration threshold as reusable association groups. Automatically fill the entity information in the reusable association groups into the qualification and performance filling form for this bid.

2. The method according to claim 1, characterized in that, In step 1, the text content is concatenated into a continuous mark sequence according to the page number order. During the concatenation process, the continuous mark sequence is divided into several paragraph-level subsequences based on line break marks and page spacing. Each paragraph-level subsequence corresponds to an independent paragraph and is assigned a paragraph number. Each mark in the continuous mark sequence is assigned an absolute position number, which starts from 1 and increases by natural numbers.

3. The method according to claim 1 or 2, characterized in that, In step 1, the sequence encoder is a Mamba sequence encoder, which consists of three components: an input projection layer, a selective state space computation layer, and an output projection layer. The input projection layer expands the embedding vector of each tag along the channel dimension to the internal dimension. The selective state space computation layer uses the expanded internal dimension vector as input to perform selective state recursion. The output projection layer compresses the internal dimension vector output by the selective state space computation layer back to the original embedding dimension to obtain the context encoding vector corresponding to each tag. The context encoding vector is in the Euclidean representation space.

4. The method according to claim 3, characterized in that, The execution process of the selective state space computation layer is as follows: For the first label in a continuous label sequence, the internal dimension vector of the first label is transformed into a discretization step size value, a state transition vector, and an input modulation vector specific to the first label through a linear mapping. Zero-order preserved discretization is performed on the state transition vector with the discretization step size value to obtain a discrete state transition vector. The discrete state transition vector is multiplied element-wise with the state vector initialized to all zeros, and then the element-wise product of the input modulation vector and the internal dimension vector of the first label is added to obtain the updated state vector corresponding to the first label. For each subsequent label starting from the second label in a continuous label sequence, a discretization step size, state transition vector, and input modulation vector specific to each subsequent label are generated in the same manner. Zero-order preserved discretization is performed on the state transition vector specific to each subsequent label using the discretization step size specific to each subsequent label to obtain a discrete state transition vector specific to each subsequent label. The discrete state transition vector specific to each subsequent label is multiplied element-wise with the updated state vector corresponding to the previous label, and then the element-wise product of the input modulation vector specific to each subsequent label and the internal dimension vector of each subsequent label is added to obtain the updated state vector corresponding to each subsequent label. The updated state vector corresponding to each label is multiplied element-wise through a fixed output mapping vector and then summed along the state dimension to obtain the single-channel output of each label in the selective state space computation layer. The single-channel outputs of all channels are concatenated along the channel dimension to obtain the complete output of each label in the selective state space computation layer.

5. The method according to claim 1, characterized in that, In step 1, the screening process for qualification and performance reference paragraphs is as follows: the average value of the context encoding vector of all marked subsequences in each paragraph is taken along the sequence length dimension to obtain the paragraph representation vector. The paragraph representation vector is sent to the paragraph classification head. The paragraph classification head outputs the judgment score of each paragraph subsequence belonging to the qualification and performance reference paragraph. The paragraph subsequence with the judgment score exceeding the preset positioning threshold is marked as the qualification and performance reference paragraph.

6. The method according to claim 1, characterized in that, In step 1, the entity boundary annotation process is as follows: The context encoding vector of each tag in the qualification and performance reference paragraph is sent to the entity annotation header. The entity annotation header outputs the score of each tag belonging to each of the following 9 categories: qualification name start position, qualification name internal position, certificate number start position, certificate number internal position, project name start position, project name internal position, project completion time start position, project completion time location and non-entity position. The category with the highest score is taken as the entity annotation result of each tag. Tags belonging to the same entity category and satisfying the adjacent position constraint are merged into 1 entity. Each entity contains entity category, entity text and paragraph number.

7. The method according to claim 1, characterized in that, In step 2, the exponential mapping process is as follows: using the origin of the Poincaré sphere model as the base point, calculate the Euclidean norm of the Euclidean eigenvectors, divide the Euclidean eigenvectors by the Euclidean norm to obtain the unit direction vector, input the Euclidean norm into the hyperbolic tangent function to obtain the hyperbolic radial component, multiply the hyperbolic radial component element-wise with the unit direction vector to obtain the hyperbolic embedding vector of each entity inside the Poincaré sphere, and the hyperbolic embedding vector is in the hyperbolic embedding space.

8. The method according to claim 1, characterized in that, In step 2, before performing the association assembly, the hierarchical number of each qualification name entity in the qualification performance entity set is determined according to the preset qualification classification hierarchical knowledge base. The qualification classification hierarchical knowledge base records the hierarchical relationship between qualification names at all levels in the industry qualification system. A qualification hierarchical tree is constructed according to the hierarchical number. The root node of the qualification hierarchical tree is the virtual top-level node. Each qualification name entity is attached to the corresponding level position in the qualification hierarchical tree according to the hierarchical number. For each edge from the parent node to the child node in the qualification hierarchical tree, the geodesic distance between the hyperbolic embedding vector of the parent node and the hyperbolic embedding vector of the child node on the Poincaré sphere is calculated as the hierarchical distance value.

9. The method according to claim 1, characterized in that, In step 2, the association assembly is divided into three stages: Stage 1 is qualification certificate pairing. Under the same paragraph number, the geodesic distance on the Poincaré sphere is calculated between each qualification name entity and each certificate number entity within the same paragraph. The certificate number entity with the smallest geodesic distance is paired with the qualification name entity to form a qualification certificate pair. Stage 2 is performance record pairing. Under the same paragraph number, the geodesic distance on the Poincaré sphere is calculated between each project name entity and each project completion time entity within the same paragraph. The project completion time entity with the smallest geodesic distance is paired with the project name entity to form a performance record pair. Stage 3 is cross-category association assembly, calculating the distance between qualification certificate pairs... The geodesic distance on the Poincaré sphere between the hyperbolic embedding vector of the qualification name entity and the hyperbolic embedding vector of the project name entity in the performance record pair is used as the first sub-item distance. The geodesic distance on the Poincaré sphere between the hyperbolic embedding vector of the certificate number entity in the qualification certificate pair and the hyperbolic embedding vector of the project completion time entity in the performance record pair is used as the second sub-item distance. The sum of the first sub-item distance and the second sub-item distance is used as the cross-category association score. Qualification certificate pairs and performance record pairs with cross-category association scores lower than the preset association threshold are merged into one qualification-performance association group. Each qualification-performance association group contains one qualification name, one certificate number, one project name, and one project completion time.

10. The method according to claim 1, characterized in that, In step 3, the calculation process of the inconsistency score is as follows: the Euclidean feature vectors of the four entities in the qualification and performance association group are concatenated along the vector dimension to form a combined vector. The combined vector is in the Euclidean representation space. The combined vector is fed into the calibration scoring function. The calibration scoring function is a linear transformation pre-trained based on the labeled correct association group and incorrect association group. The calibration scoring function outputs the inconsistency score, which represents the degree of deviation between the qualification and performance association group and the confirmed correct association pattern. The process of determining the calibration threshold is as follows: all calibration samples are taken from the pre-constructed calibration sample pool. The calibration sample pool consists of qualification and performance association groups that have been manually verified to be correctly or incorrectly associated and their corresponding inconsistency scores. The inconsistency scores of all calibration samples are arranged in ascending order to form an ordered sequence. The process is traversed in ascending order along the ordered sequence. The inconsistency score at the position where the proportion of the number of samples with inconsistency scores not exceeding the current position in the ordered sequence is not less than the preset target coverage rate for the first time is taken as the calibration threshold.

Images