Tender procurement large file parsing method based on semantic misdefinition and segment reconstruction

Abstract

The present application belongs to the technical field of bidding procurement, and particularly relates to a bidding procurement large file analysis method based on semantic misdefinition and fragment reconstruction, comprising: extracting semantic vectors, boundary box coordinates and visual style attributes of each text fragment in the bidding procurement file; constructing fragment space embedding based on the projection difference of the boundary box corner points between fragments, and aggregating to form fragment space embedding, and mapping the visual style attributes to style hierarchical embedding; inputting the three types of features of semantics, space and style into a multi-modal fusion network to obtain context-aware features, and combining the style difference condition to predict the fragment successor relationship and logical relationship through a graph decoder to construct a fragment relationship graph; starting cross-page association for the page tail source node, aggregating the context of the adjacent page head and tail candidate fragments, analyzing the cross-page logical link, and outputting the structured analysis data based on the relationship graph. The present application can improve the coherence of the fragment relationship graph and the accuracy of the structured analysis.

Description

Technical Field

[0001] This invention relates to the field of bidding and procurement technology. More specifically, this invention relates to a method for parsing large bidding and procurement documents based on semantic misidentification and fragment reconstruction. Background Technology

[0002] In modern corporate bidding processes, tender documents are the core information carriers that clarify procurement needs, evaluation criteria, and performance terms. As business scale and procurement scenarios become increasingly complex, tender documents are typically characterized by their length, varied layouts, and deep logical hierarchy. They not only contain a large amount of text content but also visual style information such as font, font size, indentation, and bolding, as well as spatial layout information such as paragraph positions, table areas, headers, and footers.

[0003] Currently, in order to support the extraction of bidding business information, intelligent review, and compliance investigation, it is necessary to parse the above-mentioned unstructured large documents into structured data that can be recognized, correlated, and analyzed by computers, thereby reducing the cost of manual review, reducing the risk of human oversight, and improving the processing efficiency and decision-making quality of the bidding and procurement process.

[0004] However, in actual parsing processes, existing solutions often struggle to simultaneously address the variability of local page layouts and the coherence of the overall logical relationships. On one hand, the text density, alignment, and visual style vary significantly across different areas of large bidding and procurement documents. Relying solely on fixed layout rules or single semantic features can easily lead to errors in identifying the sequential, subordinate, or parallel relationships between text fragments. On the other hand, long clauses, paragraphs, and complete business modules are often truncated by pagination. Conventional single-page parsing methods struggle to aggregate the contextual information between the footer of one page and the beginning of the next, resulting in broken cross-page logical chains. These issues lead to fragmented parsing results, making it difficult to accurately reconstruct the original hierarchical structure and business element relationships of bidding and procurement documents, thus failing to meet the demands of high-level bidding and procurement operations for full-length, high-precision structured parsing data. Summary of the Invention

[0005] To address the technical challenges of misclassifying relationships, breaking cross-page associations, and fragmenting results when parsing large bidding documents, this invention proposes a method for parsing large bidding and procurement documents based on semantic misclassification and fragment reconstruction, which can improve the coherence of fragment relationship graphs and the accuracy of structured parsing.

[0006] In a first aspect, the present invention provides a method for parsing large bidding and procurement documents based on semantic misidentification and fragment reconstruction, comprising: S1, obtaining the semantic feature vectors, bounding box coordinates, and visual style attributes of each text fragment in the large document to be parsed; adjusting the spatial bucketing granularity according to the text density in the local neighborhood of the current text fragment, constructing fragment pair spatial embedding vectors based on the differences in the projection of bounding box corners between text fragments, and aggregating them to generate the spatial embedding vector of the current text fragment; mapping the visual style attributes to style-level embedding vectors through a pre-trained encoder network; S2, inputting the semantic feature vectors, spatial embedding vectors, and style-level embedding vectors into a multimodal fusion network for feature fusion processing. The process involves: 1) Obtaining context-aware features of each text fragment; 2) Inputting these features into a graph decoder model, using the difference in style hierarchy embedding vectors between the source node and candidate target nodes as conditional information, jointly predicting the subsequent fragment nodes of the source node and the logical relationship type between the source node and its successor fragment nodes, thus constructing a fragment relationship graph; 3) Activating the cross-page association module for the source node at the bottom of the page, aggregating and generating cross-page context vectors from candidate fragments in the current page's footer area and the next page's header area, and adding these cross-page context vectors to the graph decoder model to parse cross-page logical link relationships; and finally, generating structured bidding and procurement large document parsing results data based on the fragment relationship graph.

[0007] By employing the aforementioned technical solutions, semantic, spatial, and visual style multimodal features of text fragments are extracted. Spatial bucketing granularity is adaptively adjusted based on local text density to construct spatial embeddings. Simultaneously, visual styles are mapped to style-level embeddings, enhancing the expressive ability of fragment positional relationships across different page layouts. Context-aware features are obtained through a multimodal fusion network, and combined with style difference conditions to predict subsequent and logical relationships between fragments, reducing the risk of semantic misclassification and improving the quality of fragment relationship graph construction. Furthermore, the cross-page association module aggregates contextual information from candidate fragments at the footer and beginning of the page, correcting logical breaks caused by pagination and improving the coherence of structured parsing results and the quality of data extraction.

[0008] Preferably, the aggregation to generate the spatial embedding vector of the current text fragment includes: obtaining the set of first corner pixel coordinates and the set of second corner pixel coordinates of the bounding boxes of the source text fragment and the target text fragment, and calculating the difference between the horizontal and vertical distances; taking the center point of the bounding box of the source text fragment as the center, setting a radius to calculate the area of ​​the local enclosed region, calculating the total area of ​​the intersection between the text fragment and the local enclosed region, and dividing it by the area of ​​the local enclosed region to obtain the relative text density ratio; when the relative text density ratio is greater than a preset baseline value, configuring a first bucketing scale parameter, otherwise configuring a second bucketing scale parameter; dividing the difference between the horizontal and vertical distances by the configured bucketing scale parameter and rounding down to obtain the discretized topological index position, concatenating the index position features and mapping them to the fragment pair spatial embedding vector; aggregating the fragment pair spatial embedding vectors corresponding to the source text fragment and each target text fragment in the preset neighborhood to obtain the spatial embedding vector of the source text fragment.

[0009] By employing the above technical solution, the horizontal and vertical distance differences between the corner coordinates of the bounding boxes of the source and target text segments are calculated. Based on the relative ratio of text density within the local enclosed area, the bucketing scale parameter is dynamically configured. This allows for the use of smaller bucketing granularity in dense text areas to improve spatial resolution, and larger bucketing granularity in sparse areas to reduce computational redundancy, thereby enhancing the positional discrimination capability of spatial embedding vectors under different page densities. Furthermore, by concatenating the discretized topological index positional features and mapping them to segment pair spatial embedding vectors and performing neighborhood aggregation, more accurate spatial topological information can be provided for determining the logical relationships between segments.

[0010] Preferably, the step of mapping visual style attributes into style-level embedding vectors through a pre-trained encoder network includes: extracting font name features from text fragments and converting them into corresponding one-hot encoding sequences; recording the font size of text fragments, subtracting a preset minimum font size from the font size, and dividing by the global font size difference range to obtain a normalized font size value; obtaining the indentation ratio value (first-line indentation distance divided by the baseline width) and extracting the font weight calibration value; sequentially concatenating the one-hot encoding sequence, normalized font size value, indentation ratio value, and font weight calibration value to generate a joint feature vector; and inputting the joint feature vector into a pre-trained multilayer perceptron network, multiplying it sequentially by a weight matrix and adding a bias vector to output a style-level embedding vector.

[0011] By adopting the above technical solutions, the font name is converted into a one-hot encoding sequence, the font size is normalized, the first-line indentation ratio and the weight are introduced, and the above multi-dimensional visual style features are concatenated into a joint feature vector. This can comprehensively capture the style differences of text fragments in terms of font, size, indentation, and thickness. The joint feature vector is input into a pre-trained multilayer perceptron network and mapped into a style hierarchy embedding vector. This can provide a continuous and comparable hierarchical style feature representation for the subsequent determination of the hierarchical relationship and logical type of the fragments, which is beneficial for identifying the style differences between different levels of content such as titles and body text.

[0012] Preferably, obtaining the context-aware features of each text segment includes: concatenating the semantic feature vector, spatial embedding vector, and style hierarchical embedding vector to generate a total concatenated feature matrix; multiplying the total concatenated feature matrix by three independent weight parameter matrices to generate a query matrix, a key matrix, and a value matrix; performing a dot product operation on the transpose of the query matrix and the key matrix, using the feature dimension of the key matrix as the dimension scaling factor, and dividing the dot product result by the square root of the dimension scaling factor to calculate the attention weight distribution using the Softmax function; multiplying the attention weight distribution by the value matrix to obtain a primary fusion feature matrix; and performing a residual concatenation and superposition of the primary fusion feature matrix and the total concatenated feature matrix, outputting the context-aware features after layer normalization.

[0013] By adopting the above technical solution, semantic feature vectors, spatial embedding vectors, and style hierarchical embedding vectors are concatenated into a total concatenated feature matrix. The attention weight distribution is calculated using a multi-head self-attention mechanism, which enables each text segment to adaptively aggregate multimodal information from other related segments, thereby obtaining richer context-aware features. Through residual connection stacking and layer normalization, the gradient vanishing problem in deep network training can be effectively alleviated, the stability and representation quality of multimodal feature fusion can be improved, and more accurate node feature inputs can be provided for subsequent graph decoder relation prediction.

[0014] Preferably, the construction of the fragment relationship graph includes: traversing candidate target nodes within a preset candidate range using the source node to obtain context-aware features of the source node and candidate target nodes; subtracting the style level embedding vector of the candidate target node from the style level embedding vector of the source node to obtain a style difference condition vector; concatenating the context-aware features of the source node, the context-aware features of the candidate target node, and the style difference condition vector, inputting them into a classifier network for feature dimensionality reduction and nonlinear mapping, and outputting a classification probability distribution of the relationship type; when the maximum probability value in the classification probability distribution is greater than a preset probability threshold, and the corresponding relationship type is not an unconnected category, establishing an edge, determining the candidate target node as the successor fragment node of the source node, and using the relationship type corresponding to the maximum probability value as the logical relationship type label of the edge, thereby generating the fragment relationship graph.

[0015] By adopting the above technical solution, the source node traverses the candidate target nodes within a preset candidate range, and calculates the difference between the style level embedding vectors of the source node and the candidate target node as the style difference condition vector. This can provide the classifier network with visual level change signals of adjacent nodes in terms of font, font size, and font weight. By concatenating the context-aware features of the source node, the context-aware features of the candidate target node, and the style difference condition vector, and inputting them into the classifier network, semantic, spatial, and style level information can be jointly used to predict the relationship type. By filtering low-confidence prediction results and excluding unconnected categories through a preset probability threshold, the risk of semantic misclassification can be effectively reduced, and the accuracy of edge and logical relationship type labels in the fragment relationship graph can be improved.

[0016] Preferably, the step of aggregating candidate fragments from the current page's footer region and the next page's header region to generate a cross-page context vector includes: calculating the absolute vertical distance between the lower edge of the source node's bounding box and the bottom edge of the current page; activating the cross-page association module when the absolute vertical distance is less than a preset distance threshold; obtaining all text fragments located after the source node in the footer region of the current page, and all text fragments within a preset range from the top of the page in the header region of the next page, as a set of cross-page candidate fragments; extracting the context-aware features of each text fragment in the cross-page candidate fragment set, performing mean pooling operation according to the candidate fragment dimension, and calculating the average feature vector as the cross-page context vector.

[0017] Preferably, the step of supplementing the cross-page context vector into the graph decoder model to parse the cross-page logical link relationship includes: traversing each candidate segment in the cross-page candidate segment set and determining the currently traversed candidate segment as the target node; concatenating the cross-page context vector, the context-aware features of the source node and the context-aware features of the target node, and the difference between the style hierarchy embedding vectors of the source node and the target node as supplementary input features of the graph decoder model to predict the cross-page logical relationship between the source node and the target node.

[0018] Preferably, the step of generating structured bidding and procurement large document parsing result data based on the fragment relationship graph includes: performing edge normalization on the fragment relationship graph to eliminate conflicting edges and determine the root level node; traversing the in-degree information of all nodes in the fragment relationship graph after edge normalization, and taking the nodes with an in-degree of zero and the root level node determined when there are no nodes with an in-degree of zero as the root level node of the document; starting from the root level node, performing a depth-first traversal based on the edges between each node and the logical relationship type label to establish the subordinate and parallel hierarchical relationships between nodes; mapping each node to a hierarchical tree structure according to the established hierarchical relationship, replacing each node in the tree structure with the original content of the corresponding text fragment, and adding the corresponding logical level label; serializing the tree structure with logical level label into a standard structured document format, and outputting the structured bidding and procurement large document parsing result data.

[0019] Preferably, the step of performing edge normalization on the fragment relationship graph to eliminate conflicting edges and determine the root level node includes: performing edge normalization on the fragment relationship graph, wherein the edge normalization includes retaining the highest confidence edge among multiple incoming edges, deleting the lowest confidence edge in loops, and determining the root level node according to the reading order and style level when there are no nodes with an in-degree of zero, and ignoring incoming edges pointing to the root level node during the tree structure generation process.

[0020] Preferably, the construction of the fragment relationship graph further includes: during the node pair relationship classification process, generating auxiliary relationship determination results for node pairs based on semantic similarity threshold rules, spatial proximity threshold rules, and style hierarchy difference threshold rules, respectively; if each auxiliary relationship determination result is inconsistent with the main determination result corresponding to the classification probability distribution output by the classifier network, or if the maximum probability value in the classification probability distribution is lower than a preset confidence threshold, then the corresponding node pair is marked as a semantically misclassified candidate node pair; the source node context-aware features, candidate target node context-aware features, fragment pair spatial embedding vector, and style difference conditional vector are reintroduced into the semantically misclassified candidate node pair, and input into a preset secondary classifier for secondary relationship determination; and the fragment relationship graph is updated according to the relationship type and confidence obtained from the secondary determination.

[0021] The present invention has the following beneficial effects: This invention, by integrating multimodal features of semantics, adaptive space, and style hierarchy, and combining them with a cross-page association correction mechanism, can effectively improve the construction quality of fragment relationship graphs in large bidding and procurement documents, and enhance the coherence and accuracy of structured analysis results.

[0022] Furthermore, this invention improves the accuracy of expressing the positional relationship between segments in different page areas by adaptively adjusting the spatial bucket granularity based on local text density and constructing spatial embedding; it provides a multi-dimensional information basis for segment relationship determination by mapping visual styles to style hierarchy features and integrating them with semantic and spatial features; it reduces the risk of semantic misclassification by introducing style difference conditions for relationship determination; and it can correct logical break defects caused by pagination by aggregating the contextual information of candidate segments at the footer and footer of the page through the cross-page association module. Attached Figure Description

[0023] Figure 1 This is a flowchart of a method for parsing large bidding and procurement documents based on semantic misclassification and fragment reconstruction; Figure 2 This is a schematic diagram of the mapping curve between the relative text density ratio and the bucket scale parameter; Figure 3 It is a bar chart showing the probability distribution of logical relationships between text fragments. Detailed Implementation

[0024] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0025] This invention discloses a method for parsing large bidding and procurement documents based on semantic misidentification and fragment reconstruction, referring to... Figure 1 This includes steps S1-S3: S1. Obtain the semantic feature vectors, bounding box coordinates, and visual style attributes of each text segment in the large file to be parsed; adjust the spatial bucketing granularity according to the text density in the local neighborhood of the current text segment; construct the spatial embedding vector of the segment pair based on the difference in the projection of the bounding box corners between text segments; and aggregate to generate the spatial embedding vector of the current text segment; map the visual style attributes into style hierarchical embedding vectors through a pre-trained encoder network.

[0026] Using optical character recognition engines such as PaddleOCR, or PDF-based parsing libraries such as PyMuPDF, the large tender and procurement documents to be parsed are analyzed to extract the text content, bounding box coordinate set, and visual style attributes of each text fragment. For electronic PDF files, the font name, font size, and font weight parameters are obtained by reading the PDF parsing library; for scanned image files, the font name, font size, and font weight parameters are obtained by a pre-trained font classification model based on the image region of the text fragment.

[0027] The pre-trained font classification model comprises a convolutional feature extraction layer, a pooling layer, a fully connected layer, and a multi-task output layer connected sequentially. The convolutional feature extraction layer extracts text stroke thickness, character outline, and local texture features; the pooling layer compresses feature sizes; the fully connected layer generates style discrimination features; and the multi-task output layer outputs the font name category, font size estimate or font size level, and font weight level, respectively. The model is trained on historical text segment image samples labeled with font name, font size, and font weight. The bounding box coordinate set includes the horizontal and vertical coordinates of the top-left and bottom-right corners; visual style attributes include font name, font size, whether it is bold, whether it is italic, first-line indent distance, and font weight parameter. The first-line indent distance is determined based on the horizontal distance between the left boundary of the text segment bounding box and the baseline left boundary of the paragraph or page body text; the font weight parameter is mapped to a font weight rating value based on the bold attribute or font weight parameter. The text content is lexicalized using a pre-trained language model such as BERT, generating a sequence of lexical numbers, position numbers, and attention masks. These are then input into the Transformer encoding layer of the BERT model for context encoding. The output vector at the [CLS] position is taken, or the output vectors of each lexical are averaged and pooled to obtain a fixed-dimensional semantic feature vector.

[0028] Using the center coordinates of the current text segment as the center, a circular region with a preset radius is defined as the local neighborhood. The sum of the intersection areas of the bounding boxes of each text segment within this neighborhood with the circular region is calculated. This sum of intersection areas is then divided by the area of ​​the circular region to obtain the text density. The binning granularity is adjusted according to a preset mapping rule; when the text density exceeds a density threshold, the binning granularity is reduced to improve spatial resolution. For any two text segments, the corner coordinates of their bounding boxes are extracted, and the difference between the horizontal and vertical coordinate projections is calculated. Discretization and binning are then performed using the adjusted binning granularity to obtain the bin index value. Simultaneously, the overlap ratio of the two bounding boxes in the horizontal and vertical directions, as well as the ratio of the intersection area to the joint area, are calculated to obtain the overlap feature value. Discretized bucket indices are input into the learnable embedding layer to obtain discrete spatial index embeddings; overlap ratios and overlap feature values ​​are input into the linear mapping layer to obtain continuous spatial feature embeddings; the two are then concatenated and input into a multilayer perceptron network to generate fragment pair spatial embedding vectors. The fragment pair spatial embedding vectors corresponding to the current text fragment and each target text fragment in the preset neighborhood are then aggregated to obtain the spatial embedding vector of the current text fragment.

[0029] The font name is input as a category feature into a one-hot encoder for discretization. Font size, first-line indentation distance, and weight level are normalized as continuous numerical features. The bold attribute is used to generate the weight level value, and the slant attribute is used as an optional additional style feature. The discretized encoding result and the normalized result are concatenated to form a comprehensive visual feature vector. A pre-trained multilayer perceptron encoder network containing linear layers and ReLU activation function layers is invoked to perform forward propagation on the comprehensive visual feature vector, outputting a continuous real-valued vector of a specified dimension as a style-level embedding vector.

[0030] In an optional embodiment, the spatial bucketing granularity is adjusted according to the text density in the local neighborhood of the current text segment, a segment pair spatial embedding vector is constructed based on the difference in the projection of the bounding box corners between text segments, and the spatial embedding vector of the current text segment is generated by aggregation, including: Obtain the set of pixel coordinates of the first corner point of the bounding box of the source text fragment and the set of pixel coordinates of the second corner point of the bounding box of the target text fragment; Calculate the difference in horizontal distance and the difference in vertical distance between the set of pixel coordinates of the first corner point and the set of pixel coordinates of the second corner point; Using the center point of the bounding box of the source text fragment as the center, calculate the area of ​​the local enclosed region with a set radius, calculate the total area of ​​the intersection between the text fragment and the local enclosed region, divide the total area of ​​the intersection by the area of ​​the local enclosed region, and obtain the relative text density ratio. When the relative text density ratio is greater than the preset baseline value, configure the first bucket scale parameter; when the relative text density ratio is not greater than the preset baseline value, configure the second bucket scale parameter which is greater than the first bucket scale parameter. The horizontal and vertical distance differences are divided by the configured bucket scale parameter and rounded down to extract the discretized topological index positions. The corresponding index positions are then concatenated with features and mapped to fragment pair spatial embedding vectors. The fragment pair spatial embedding vectors corresponding to the source text fragment and each target text fragment in the preset neighborhood are then aggregated to generate the spatial embedding vector of the source text fragment.

[0031] In the stage of obtaining coordinates and distance differences, the pixel coordinates of the top-left corner of the bounding box of the source text fragment are extracted. and bottom right pixel coordinates The coordinates of the center point are calculated, and the coordinate parameters of the target text fragment are obtained similarly. The specific distance difference between the horizontal and vertical directions is obtained by subtracting the corresponding coordinates. For example, the horizontal difference is calculated to be 120 pixels and the vertical difference is 45 pixels. Taking the center point of the source text fragment as the center, the radius of the local enclosing region is set to be between 150 and 200 pixels, with a preferred value of 150 pixels. The total area of ​​the enclosing region is calculated, and the total intersection area of ​​all surrounding text fragments falling within the enclosing region is calculated using the geometric intersection area calculation method. The total area is divided by the total area of ​​the region to obtain the relative text density ratio. For example, the calculated intersection area is 28260 square pixels, which, when divided by the area of ​​a circle with a radius of 150 pixels (70650 square pixels), yields a text density ratio of 0.4.

[0032] After calculating the relative text density ratio, the ratio is compared with a preset baseline value, which is usually set between 0.35 and 0.45, preferably 0.40. If the calculated ratio of 0.4 is not greater than the baseline value of 0.40, the second binning scale parameter is configured to 20 pixels. Conversely, if the ratio is greater than 0.40 in other high-density areas, the first binning scale parameter is configured to 10 pixels. The previously obtained horizontal difference of 120 pixels and vertical difference of 45 pixels are divided by the currently configured binning scale parameter of 20 and rounded down to obtain the horizontal index position 6 and the vertical index position 2. The generated horizontal discretization index 6 and vertical discretization index 2 are input into a learnable embedding network layer with a feature dimension of 64 or 128, respectively. The learnable embedding network layer contains an embedding matrix corresponding to the position of the discretized topological index. The horizontal dense index features and vertical dense index features are obtained by looking up the table, and the two are concatenated to form a discrete spatial index feature. At the same time, the horizontal overlap ratio, the vertical overlap ratio, and the ratio of the intersection area of ​​the bounding boxes to the joint area are used to form a continuous overlap feature, which is then input into a linear mapping layer to obtain continuous spatial features.

[0033] Dense index features and continuous spatial features are concatenated and input into a multilayer perceptron network to generate a 1×128 segment pair spatial embedding vector. Then, mean pooling is performed on the segment pair spatial embedding vectors corresponding to the source text segment and multiple target text segments within a preset neighborhood to generate a 1×128 spatial embedding vector for the source text segment. The mapping curve between the relative text density ratio and the bucketing scale parameter is shown below. Figure 2 As shown.

[0034] In an optional embodiment, visual style attributes are mapped to style-level embedding vectors via a pre-trained encoder network, including: Extract the font name features of the text fragment and convert them into the corresponding one-hot encoding sequence; Record the font size of the text fragment, subtract the preset minimum font size from the font size, and divide by the global font size difference range to obtain the normalized font size value; Obtain the indentation ratio value, which is the first-line indentation distance divided by the baseline width, and extract the character weight calibration value; The one-hot encoding sequence, normalized font size value, indentation ratio value, and character weight calibration value are concatenated sequentially to generate a joint feature vector. The joint feature vector is input into a pre-trained multilayer perceptron network, multiplied by the weight matrix and then added with the bias vector to output the style hierarchical embedding vector.

[0035] In the visual feature extraction and representation stage, the underlying Extensible Markup Language (XML) tags of the large tender document are parsed to obtain font names. For example, SimSun, Heiti, and FangSong are mapped to one-hot encoded sequences with a dimension of 10. For instance, the encoding for Heiti is 0,1,0,0,0,0,0,0,0,0. The actual font size of the text fragment is extracted. For example, if the currently extracted font size is 18 points, and the minimum font size baseline for large document parsing is preset to 5 points, and the difference range between the maximum and minimum global font sizes is set to 40 points, then the dimensionless normalized font size value of 0.325 is calculated using the formula (18-5) / 40. The absolute indentation distance of the first line of the text fragment is extracted to be 24 points. This is divided by the current document's default standard character baseline width of 12 points to obtain the first line indentation ratio value of 2. At the same time, the font weight values ​​obtained by the system, such as bold 700 or regular 400, are divided by 1000 to obtain font weight calibration values ​​such as 0.7 or 0.4.

[0036] A 10-dimensional one-hot encoded sequence, a 1-dimensional normalized font size value, a 1-dimensional indentation ratio value, and a 1-dimensional weight rating value are concatenated along the feature channel dimension to construct a 13-dimensional joint feature vector. This 13-dimensional joint feature vector is then input into a pre-trained multilayer perceptron network containing alternating stacked fully connected layers and non-linear activation layers. The network has two fully connected layers. The 13-dimensional input feature vector is multiplied by a 13×64 first weight matrix and a 64-dimensional bias vector is added. After a non-linear transformation using a linear rectified activation function, it is multiplied by a 64×128 second weight matrix and the bias vector is added. Each layer performs the operation Y=WX+B, outputting a 1×128 high-dimensional style hierarchy embedding vector used to identify the hierarchical style feature differences between text segments. The pre-trained multilayer perceptron encoder network can be trained based on historical bidding and procurement document samples. These samples include visual style attributes of text fragments and their corresponding hierarchical category labels or adjacent fragment relationship labels. During training, the goal is to predict the hierarchical category of text fragments or the relationship between adjacent fragments, and the network parameters are updated using the cross-entropy loss function.

[0037] S2, the semantic feature vector, spatial embedding vector and style hierarchy embedding vector are input into the multimodal fusion network for feature fusion processing to obtain the context-aware features of each text segment; the context-aware features are input into the graph decoder model, and the difference between the style hierarchy embedding vectors of the source node and the candidate target node is used as conditional information to jointly predict the successor segment node of the source node and the logical relationship type between the source node and the successor segment node, thus constructing a segment relationship graph.

[0038] The semantic feature vectors, spatial embedding vectors, and style hierarchy embedding vectors of each text segment are concatenated along the feature dimension to obtain an initial multimodal feature matrix. A multi-layer self-attention mechanism network architecture based on Transformer is then constructed as the multimodal fusion network. This network includes a multi-head self-attention layer and a feedforward neural network layer. Layer normalization and residual connection mechanisms are used to prevent gradient vanishing. The initial multimodal feature matrix is ​​input into the multi-head self-attention layer. The dot product of the query matrix and the key matrix is ​​calculated using matrix multiplication, and then the attention weight distribution is calculated and multiplied by the value matrix to obtain the attention-weighted feature representation. This representation is then nonlinearly transformed by the feedforward neural network layer to output the fused context-aware features.

[0039] The graph decoder model consists of a candidate node relationship classifier. It iterates through each text segment in the file as a source node and extracts other text segments within a preset candidate range as candidate target nodes. Vector subtraction is used to calculate the difference vector between the source node and the candidate target node in their style-level embedding vectors, serving as conditional information. The context-aware features of the source node and candidate target node, along with the difference vector, are input into the candidate node relationship classifier, which outputs a probability distribution of the relationship type between the source node and the candidate target node. If the maximum probability value is greater than a preset probability threshold and the corresponding category is not an unconnected category, the candidate target node is determined as the successor segment node of the source node. Directed edges and relationship labels are established between each text segment node to construct a complete segment relationship graph. During the training phase, supervised samples are constructed based on the labeled node pairs and relationship labels. Unconnected node pairs are labeled as unconnected categories, and related node pairs are labeled as corresponding logical relationship categories. A multi-class cross-entropy loss function is used to train the candidate node relationship classifier.

[0040] In an optional embodiment, the semantic feature vector, spatial embedding vector, and style-level embedding vector are jointly input into a multimodal fusion network for feature fusion processing to obtain context-aware features for each text segment, including: The semantic feature vector, spatial embedding vector, and style hierarchical embedding vector are concatenated to generate a total concatenated feature matrix. Multiply the total concatenated feature matrix by three independent weight parameter matrices to generate the query matrix, key matrix, and value matrix respectively. The query matrix and the transpose of the key matrix are multiplied by a dot product. The feature dimension of the key matrix is ​​used as the dimension scaling factor. The result of the dot product is divided by the square root of the dimension scaling factor. The attention weight distribution is then calculated using the Softmax function. Multiply the attention weight distribution by the value matrix to obtain the initial fusion feature matrix; The primary fused feature matrix and the total spliced ​​feature matrix are superimposed by residual connection, and after layer normalization, the context-aware features are output.

[0041] The 768-dimensional semantic feature vector extracted by the pre-trained language model, the constructed 128-dimensional spatial embedding vector, and the 128-dimensional style hierarchical embedding vector are concatenated along the last dimension of the feature sequence to generate a total concatenated feature matrix with a dimension of 1024 and a shape of N×1024, where N is the number of text segment nodes.

[0042] In the self-attention module of the multimodal fusion network, three independent, learnable weight parameter matrices, each with a size of 1024×1024, are initialized and used. The total concatenated feature matrix is ​​then multiplied by these three weight matrices using dot products, sequentially generating the query matrix Q, key matrix K, and value matrix V of corresponding dimensions. The data dimension of all three remains N×1024. During the attention calculation and residual normalization stages, the extracted query matrix Q is multiplied by the key matrix after matrix transpose processing. Perform matrix dot product calculations to generate an N×N similarity matrix containing the original relevance scores between nodes. To avoid the dot product result becoming too large with increasing feature dimension and causing excessive concentration in the softmax distribution, an attention dimension coefficient is set. The value is 1024. The square root of the scaling factor for this dimension is calculated to be 32. The dot product result is then divided by 32 for scaling. The last dimension is normalized using a normalized exponential function to obtain an attention weight distribution matrix with values ​​between 0 and 1 and a sum of 1 for each row. This weight distribution matrix is ​​multiplied by the value matrix V to complete feature aggregation, resulting in a preliminary fusion feature matrix that integrates multimodal interaction information. A residual network structure is used to perform element-wise matrix addition on the preliminary fusion feature matrix and the N×1024 total concatenated feature matrix before the input module. This result is then passed to a layer normalization function layer for processing, unifying its variance and mean, and outputting an N×1024 context-aware feature containing cross-modal logic.

[0043] In an optional embodiment, context-aware features are input into the graph decoder model, using the difference between the style-level embedding vectors of the source node and the candidate target node as conditional information. This jointly predicts the successor fragment nodes of the source node and the logical relationship type between the source node and the successor fragment nodes, constructing a fragment relationship graph, including: Traverse the candidate target nodes within the preset candidate range using the source node to obtain the context-aware features of the source node and the context-aware features of the candidate target nodes. The style hierarchy embedding vector of the source node is subtracted from the style hierarchy embedding vector of the candidate target node to obtain the style difference conditional vector. The context-aware features of the source node, the context-aware features of the candidate target node, and the style difference conditional vector are concatenated and input into the classifier network for feature dimensionality reduction and nonlinear mapping, and the classification probability distribution of the relation type is output. When the maximum probability value in the classification probability distribution is greater than the preset probability threshold, and the relationship type corresponding to the maximum probability value is not an unconnected category, an edge is established between the source node and the candidate target node. The candidate target node is determined as the successor fragment node of the source node, and the relationship type corresponding to the maximum probability value is used as the logical relationship type label of the edge to generate a fragment relationship graph.

[0044] Set a preset candidate range size, such as the 20 most recent candidate text segments in both spatial and reading sequence dimensions after the source node, and iterate through each candidate target node within the sliding window. For each pair of source and target nodes, extract the previously obtained 1×1024-dimensional context-aware features of the source node and the target node in parallel. Simultaneously, subtract the 128-dimensional style hierarchy embedding vector of the candidate target node from the 128-dimensional style hierarchy embedding vector of the source node to obtain a 1×128-dimensional style difference conditional vector. For example, the difference reflects the hierarchical degradation relationship of font size from large to small or from bold to regular font.

[0045] These three vectors are concatenated along the feature dimension to generate a densely concatenated feature vector with a total dimension of 1×2176. In the feature classification and graph edge relationship generation stage, the generated 1×2176 concatenated feature is input into a classifier network consisting of three linear layers stacked with the LeakyReLU activation function. The classifier network contains multiple linear fully connected layers and non-linear activation layers stacked in series, as well as an output classification layer. The classifier network takes the concatenated feature vector as input and reduces the feature dimension to 512. This is then mapped to a 5-dimensional log-probability output containing 5 predefined relationship categories, such as title belonging to body text, parallel paragraphs at the same level, table title data, body text continuation, and no connection. The normalized exponential function layer then transforms these 5 predefined relationship categories into a classification probability distribution array ranging from 0 to 1, such as 0.85, 0.05, 0.02, 0.05, 0.03.

[0046] The maximum probability value of 0.85 is extracted from the probability distribution array and compared with a preset probability threshold, which is typically set between 0.55 and 0.65, preferably 0.60. Since 0.85 is greater than 0.60, and the category corresponding to this maximum probability value is not a disconnected category, the current node pair is confirmed to have a relationship. A directed edge is established for the current node pair in the graph decoder, and the corresponding candidate target node is determined as the successor segment node of the source node. The edge corresponding to the title with the highest probability belonging to the body text is assigned as the logical relationship type label. If the maximum probability value corresponds to a disconnected category, no edge is established between the source node and the candidate target node. During the training phase, supervised samples are constructed based on the labeled node pair relationship labels. Disconnected node pairs are labeled as disconnected categories, and related node pairs are labeled as corresponding logical relationship categories. A multi-class cross-entropy loss function is used to train the classifier network.

[0047] During the node pair relationship classification process, auxiliary relationship determination results are generated based on semantic similarity threshold rules, spatial proximity threshold rules, and style hierarchy difference threshold rules, respectively. If the auxiliary relationship determination results are inconsistent with the main determination results corresponding to the classification probability distribution, or if the maximum probability value in the classification probability distribution is lower than the preset confidence threshold, the corresponding node pair is marked as a semantically misclassified candidate node pair. The source node context-aware features, candidate target node context-aware features, fragment pair spatial embedding vector, and style difference conditional vector are reintroduced into the semantically misclassified candidate node pair, and input into a preset secondary classifier for secondary relationship determination. The preset secondary classifier adopts the same linear classification structure as the aforementioned classifier network, and its input dimension matches the dimension of the composite vector obtained by concatenating the source node context-aware features, candidate target node context-aware features, fragment pair spatial embedding vector, and style difference conditional vector. The fragment relationship graph is updated according to the relationship type and confidence obtained from the secondary determination.

[0048] Iterative execution across the entire document constructs a relational graph of fragments within the large tender document. The probability distribution of logical relationships among text fragments is as follows: Figure 3 As shown.

[0049] S3 activates the cross-page association module for the source node at the bottom of the page, aggregates and generates a cross-page context vector from candidate fragments in the current page's footer area and the next page's header area, and adds the cross-page context vector to the graph decoder model to parse the cross-page logical link relationship; and generates structured bidding and procurement large document parsing result data based on the fragment relationship graph.

[0050] The system determines whether a text fragment is located at the bottom of the page based on the ratio of its maximum vertical coordinate to the total height of the current page. If this ratio is greater than a preset bottom-of-page threshold, the corresponding text fragment is marked as a source node at the bottom of the page, and the cross-page association module is activated. The cross-page association module extracts a preset number of candidate fragments from the current page's footer region and the next page's footer region. These candidate fragments are used as graph nodes, and undirected edges are established based on the reading order's adjacency and the bounding box's spatial proximity. A graph convolutional algorithm (GCN) is used to update the context-aware features of each candidate fragment, and average pooling is applied to the updated fragment features to generate a cross-page context vector. This cross-page context vector is concatenated with the source node's context-aware features, the target node's context-aware features, and the style difference conditional vector. This concatenation is then input into an extended classification layer that matches the dimensions of the composite input, performing node pair relationship classification and resolving the logical link relationships between cross-page text fragments.

[0051] Before generating the tree structure, the fragment relationship graph is processed by edge normalization: for cases where multiple incoming edges exist for the same target node, the incoming edge with the highest confidence is retained; for edge sets forming loops, the edge with the lowest confidence is deleted; if no node with zero in-degree exists after normalization, the node with the highest reading order and style level is determined as the root level node, and incoming edges pointing to this root level node are ignored in subsequent tree structure generation. The style level is determined comprehensively based on the font size, weight, indentation ratio, and the level score corresponding to the style level embedding vector of the text fragment. Subsequently, starting from the root level node, a depth-first search algorithm is used to traverse the fragment relationship graph, and the hierarchy is assembled according to the directed edge pointing and logical relationship type: hierarchical inclusion relationships are configured as parent-child structures, sibling relationships are configured as a list of sibling nodes, and the text content corresponding to paragraph continuation relationships is concatenated. Finally, the generated tree-like hierarchical structure is serialized into a JSON format file as the parsing result data of the large bidding and procurement file.

[0052] In an optional embodiment, the cross-page association module is activated for the source node at the bottom of the page, and a cross-page context vector is generated by aggregating candidate fragments from the current page's footer region and the next page's header region. This cross-page context vector is then added to the graph decoder model to parse cross-page logical link relationships, including: Calculate the absolute vertical distance between the bottom edge of the source node's bounding box and the bottom edge of the current page. When the absolute vertical distance is less than a preset distance threshold, activate the cross-page association module. Get all text fragments located after the source node in the footer area of ​​the current page, and all text fragments within a preset range from the top of the page in the footer area of ​​the next page, as a set of cross-page candidate fragments; Extract the context-aware features of each text segment in the cross-page candidate segment set, perform mean pooling operation according to the candidate segment dimension, and calculate the average feature vector as the cross-page context vector; Iterate through each candidate segment in the cross-page candidate segment set and determine the currently traversed candidate segment as the target node; The cross-page context vector, the context-aware features of the source node and the context-aware features of the target node, and the difference between the style hierarchy embedding vectors of the source node and the target node are concatenated as supplementary input features for the graph decoder model to predict the cross-page logical relationship between the source node and the target node.

[0053] During the cross-page feature boundary recognition and candidate set extraction process, the vertical coordinates of the source text nodes in the current document page are continuously tracked, and the vertical coordinates of the bottom edge of their bounding boxes are calculated as follows: Equal to 1050 pixels and the bottom reference coordinates of the entire page, such as the bottom height of a standard scanned document. A difference of 1169 pixels yields an absolute vertical distance of 119 pixels. When this distance is less than a preset distance threshold, preferably 15% of the total page height (e.g., 150 pixels), the cross-page association module is activated via an event trigger mechanism. The module's extraction rules are as follows: All remaining text fragments on the current page following the source node are selected. Then, all text fragments with a vertical coordinate between 0 and 250 pixels at the top of the next page are extracted. This preset range covers the height of the first line of a regular paragraph. All the selected text fragments are integrated to form a cross-page candidate fragment set containing approximately 5 to 15 fragment nodes.

[0054] The generated set of cross-page candidate segments is traversed, and the 1×1024 context-aware feature matrix of each node in the set is retrieved and extracted. Then, global mean pooling is performed on these feature matrices along the tensor dimension of the number of segments, and the average feature vector with a consistent dimension of 1×1024 is output as the global cross-page context vector. For the target node currently being traversed, a 1×128 style difference conditional vector is calculated between the source node and the target node. The 1×1024 source node context features representing the current segment, the 1×1024 target node context features representing the target segment, the 1×128 style difference conditional vector representing hierarchical style differences, and the 1×1024 cross-page context vector representing the global context are tightly concatenated in the channel dimension to form a composite input vector with a length of 3200. This vector is then imported into the extended classification layer in the graph decoder model that matches the 3200-dimensional composite input vector. This allows the classifier to predict whether there are cross-page logical connections such as cross-page sentence breaks or cross-page hierarchical subordination between the two segments, based on the global text environment at the connection between the two pages.

[0055] In an optional embodiment, structured bidding and procurement large document parsing result data is generated based on the fragment relationship graph, including: The fragment relationship graph is processed by edge regularization, which includes retaining the highest confidence edge in multiple edges, deleting the lowest confidence edge in loops, and determining the root level node according to the reading order and style level when there are no nodes with zero in-degree. In the process of generating the tree structure, the in-edges pointing to the root level node are ignored. Traverse the in-degree information of all nodes in the fragment relationship graph after edge normalization, and take the nodes with an in-degree of zero and the root level node determined when there are no nodes with an in-degree of zero as the root level node of the document. Starting from the root level node, perform a depth-first traversal based on the edges between nodes and the logical relationship type labels to establish the subordinate and parallel hierarchical relationships between nodes. Based on the established hierarchical relationship, each node is mapped to a hierarchical tree structure, each node in the tree structure is replaced with the original content of the corresponding text fragment, and corresponding logical hierarchy tags are added. The tree structure with logical hierarchy tags is serialized and converted into a standard structured document format, outputting structured bidding and procurement large document parsing results data.

[0056] In the graph dimensionality reduction and tree structure instantiation stages, the previously generated fragment relationship graph containing directed edges and classification weights is loaded, and a node degree counting algorithm is executed in the global topology network. Before performing depth-first search, the graph is normalized based on the edge confidence. For cases where multiple candidate parent nodes exist for the same target node, only the incoming edge with the highest confidence is retained. For detected loop paths, the edge with the lowest confidence in the loop path is deleted. If no node with an in-degree of 0 exists after normalization, the node with the highest reading order and style level is selected as the root-level node, and incoming edges pointing to this root-level node are no longer used as the basis for parent-child relationships during tree structure generation. The number of directed incoming edges for each text node in the graph is checked one by one. Discrete nodes, such as the document cover title or the top-level chapter title, that are not pointed to by any other text node (i.e., the directed edge in-degree is equal to 0) are extracted, uniformly identified, and selected as the root-level nodes of the structured document abstract syntax tree.

[0057] After establishing the root node, a depth-first search traversal is performed downstream from this node. During this process, a hierarchical relationship is established based on the attribute label constraints of the connected edges: for example, when a title is encountered as a subordinate text label, the child node pointer is pushed into the lower level as the subordinate level of the parent chapter; if a parallel paragraph label is encountered, a parallel linked list node is generated at the same level, decoupling and mapping the originally variable network graph into a standard hierarchical tree data structure with a unique predecessor path.

[0058] The constructed tree structure is integrated with the underlying document storage engine. Placeholders containing only index identifiers in each tree node are extracted and replaced with the corresponding original text string data identified by optical character recognition. Simultaneously, logical level label parameters are attached to the node metadata attributes, such as concrete category names like Heading_1, BodyText_Level2, and Table_Title. After content population and label binding are completed, the parser triggers the data serialization module. Following industry standards, the nested dictionary-like tree structure currently stored in memory is recursively parsed and output as a lightweight JSON standard format file or a more rigorous XML and HTML tag format. The persistent output is a structured parsing dataset of large bidding and procurement documents with an outline-based directory structure, free from image styles, which can then be used by the subsequent information extraction engine to read and call key elements.

[0059] The experiment utilized a large dataset containing 50,000 real bidding and procurement documents, divided into a 40,000-part training set and a 10,000-part test set. The hardware environment deployed four high-performance graphics processors, with an initial learning rate set to 0.05% and a cosine annealing decay strategy. The batch size was set to 16, and a total of 150 training epochs were performed. The baseline control group used a conventional graph attention network combined with a basic pre-trained language model for relation classification. The ablation experimental group, on the other hand, progressively added a spatial embedding module, a style hierarchy feature difference module, and a cross-page association fusion module. The core evaluation metrics for the model were the macro-average F1 score of relation edge prediction and the accuracy of reconstructing the tree structure of the entire document.

[0060] Test results show that the baseline control group achieved a macro-average F1 score of 79.4% for edge prediction and a tree structure reconstruction accuracy of 75.8%. After adding only the spatial embedding module, the macro-average F1 score increased to 84.2%, and the reconstruction accuracy improved to 81.6%. Further input of the style hierarchy feature difference module further increased the macro-average F1 score to 89.7%, and the reconstruction accuracy reached 88.3%. The complete solution using all innovative features achieved superior experimental results under these test conditions, with a macro-average F1 score of 94.5% for fragment edge prediction and a document tree structure reconstruction accuracy exceeding 93.1%.

[0061] The spatial embedding mechanism adjusts the spatial feature segmentation scale through text density, reducing the spatial topological mismatch rate in densely typed areas and enhancing the model's tolerance to paragraph structures. The style hierarchy feature difference module provides the network with font size and weight gradient signals between adjacent nodes, improving the classification confidence of hierarchical membership and parallel relationships. The cross-page association module integrates global contextual features from the page header and footer, correcting logical breaks caused by hard pagination, reducing the risk of broken hierarchical trees in long documents, and improving the high-fidelity output of parsed data.

[0062] In the description of this specification, "multiple" or "several" means at least two, such as two, three or more, unless otherwise expressly and specifically defined.

Claims

1. A method for parsing large bidding and procurement documents based on semantic misidentification and fragment reconstruction, characterized in that, include: S1. Obtain the semantic feature vectors, bounding box coordinates, and visual style attributes of each text segment in the large file to be parsed; adjust the spatial bucketing granularity according to the text density in the local neighborhood of the current text segment; construct the segment pair spatial embedding vector based on the difference in the projection of the bounding box corners between text segments, and aggregate to generate the spatial embedding vector of the current text segment; map the visual style attributes into style-level embedding vectors through a pre-trained encoder network; S2. Input the semantic feature vectors, spatial embedding vectors, and style-level embedding vectors into a multimodal fusion network for feature fusion processing to obtain the context-aware features of each text segment; Context-aware features are input into the graph decoder model. The difference between the style hierarchy embedding vectors of the source node and the candidate target node is used as conditional information to jointly predict the successor fragment nodes of the source node and the logical relationship type between the source node and the successor fragment nodes, thus constructing a fragment relationship graph. In S3, the cross-page association module is activated for the source node at the bottom of the page. Cross-page context vectors are aggregated from the candidate fragments in the current page footer area and the next page footer area, and the cross-page context vectors are added to the graph decoder model to parse the cross-page logical link relationship. The structured bidding and procurement large document parsing result data is generated based on the fragment relationship graph.

2. The method according to claim 1, characterized in that, The aggregation to generate the spatial embedding vector of the current text fragment includes: obtaining the set of first corner pixel coordinates and the set of second corner pixel coordinates of the bounding boxes of the source text fragment and the target text fragment, and calculating the difference between the horizontal and vertical distances; taking the center point of the bounding box of the source text fragment as the center, setting a radius to calculate the area of ​​the local enclosed region, calculating the total area of ​​the intersection between the text fragment and the local enclosed region, and dividing it by the area of ​​the local enclosed region to obtain the relative text density ratio; when the relative text density ratio is greater than a preset baseline value, configuring a first bucketing scale parameter, otherwise configuring a second bucketing scale parameter; dividing the difference between the horizontal and vertical distances by the configured bucketing scale parameter and rounding down to obtain the discretized topological index position, concatenating the index position features and mapping them to the fragment pair spatial embedding vector; aggregating the fragment pair spatial embedding vectors corresponding to the source text fragment and each target text fragment in the preset neighborhood to obtain the spatial embedding vector of the source text fragment.

3. The method according to claim 2, characterized in that, The process of mapping visual style attributes into style-level embedding vectors through a pre-trained encoder network includes: extracting font name features from text fragments and converting them into corresponding one-hot encoding sequences; recording the font size of text fragments, subtracting a preset minimum font size from the font size, and dividing by the global font size difference range to obtain a normalized font size value; obtaining the indentation ratio value (first-line indentation distance divided by the baseline width) and extracting the font weight calibration value; sequentially concatenating the one-hot encoding sequence, normalized font size value, indentation ratio value, and font weight calibration value to generate a joint feature vector; and inputting the joint feature vector into a pre-trained multilayer perceptron network, multiplying it by a weight matrix and adding a bias vector to output a style-level embedding vector.

4. The method according to claim 1, characterized in that, The process of obtaining context-aware features for each text segment includes: concatenating semantic feature vectors, spatial embedding vectors, and style hierarchical embedding vectors to generate a total concatenated feature matrix; multiplying the total concatenated feature matrix by three independent weight parameter matrices to generate a query matrix, a key matrix, and a value matrix; performing a dot product operation on the transpose of the query matrix and the key matrix, using the feature dimension of the key matrix as a dimension scaling factor, and dividing the dot product result by the square root of the dimension scaling factor to calculate the attention weight distribution using the Softmax function; multiplying the attention weight distribution by the value matrix to obtain a primary fusion feature matrix; and performing a residual concatenation and superposition of the primary fusion feature matrix and the total concatenated feature matrix, followed by layer normalization processing to output the context-aware features.

5. The method according to claim 1, characterized in that, The construction of the fragment relationship graph includes: traversing candidate target nodes within a preset candidate range using the source node to obtain context-aware features of the source node and candidate target nodes; subtracting the style level embedding vector of the candidate target node from the style level embedding vector of the source node to obtain a style difference condition vector; concatenating the context-aware features of the source node, the context-aware features of the candidate target node, and the style difference condition vector, inputting them into a classifier network for feature dimensionality reduction and nonlinear mapping, and outputting a classification probability distribution of the relationship type; when the maximum probability value in the classification probability distribution is greater than a preset probability threshold, and the corresponding relationship type is not an unconnected category, establishing an edge, determining the candidate target node as the successor fragment node of the source node, and using the relationship type corresponding to the maximum probability value as the logical relationship type label of the edge, thereby generating the fragment relationship graph.

6. The method according to claim 4, characterized in that, The process of aggregating candidate fragments from the current page's footer region and the next page's header region to generate a cross-page context vector includes: Calculate the absolute vertical distance between the bottom edge of the source node's bounding box and the bottom edge of the current page. When the absolute vertical distance is less than a preset distance threshold, activate the cross-page association module. Get all text fragments located after the source node in the footer area of ​​the current page, and all text fragments within a preset range from the top of the page in the footer area of ​​the next page, as a set of cross-page candidate fragments; Extract context-aware features from each text segment in the cross-page candidate segment set, perform mean pooling operation according to the candidate segment dimension, and calculate the average feature vector as the cross-page context vector.

7. The method according to claim 6, characterized in that, The step of supplementing the cross-page context vector into the graph decoder model to parse cross-page logical link relationships includes: Iterate through each candidate segment in the cross-page candidate segment set and determine the currently traversed candidate segment as the target node; The cross-page context vector, the context-aware features of the source node and the context-aware features of the target node, and the difference between the style hierarchy embedding vectors of the source node and the target node are concatenated as supplementary input features for the graph decoder model to predict the cross-page logical relationship between the source node and the target node.

8. The method according to claim 1, characterized in that, The process of generating structured bidding and procurement large document parsing results data based on the fragment relationship graph includes: performing edge normalization on the fragment relationship graph to eliminate conflicting edges and determine the root level node; traversing the in-degree information of all nodes in the fragment relationship graph after edge normalization, and taking the nodes with an in-degree of zero and the root level node determined when there are no nodes with an in-degree of zero as the root level node of the document; starting from the root level node, performing a depth-first traversal based on the edges between each node and the logical relationship type labels to establish subordinate and parallel hierarchical relationships between nodes; mapping each node to a hierarchical tree structure based on the established hierarchical relationships, replacing each node in the tree structure with the original content of the corresponding text fragment, and adding the corresponding logical level label; serializing the tree structure with logical level labels into a standard structured document format, and outputting the structured bidding and procurement large document parsing results data.

9. The method according to claim 8, characterized in that, The step of performing edge normalization on the fragment relationship graph to eliminate conflicting edges and determine the root level node includes: performing edge normalization on the fragment relationship graph, wherein the edge normalization process includes retaining the highest confidence edge among multiple incoming edges, deleting the lowest confidence edge in loops, and determining the root level node according to the reading order and style level when there are no nodes with an in-degree of zero, and ignoring incoming edges pointing to the root level node during the tree structure generation process.

10. The method according to claim 5, characterized in that, The construction of the fragment relationship graph also includes: In the process of classifying node pairs, auxiliary relationship determination results are generated based on semantic similarity threshold rules, spatial proximity threshold rules, and style hierarchy difference threshold rules, respectively. If the auxiliary relation determination results are inconsistent with the main determination results corresponding to the classification probability distribution output by the classifier network, or if the maximum probability value in the classification probability distribution is lower than the preset confidence threshold, then the corresponding node pair is marked as a semantic misclassification candidate node pair; the source node context-aware features, candidate target node context-aware features, fragment pair spatial embedding vector, and style difference conditional vector are reintroduced into the semantic misclassification candidate node pair, and input into the preset secondary classifier for secondary relation determination; the fragment relation graph is updated according to the relation type and confidence obtained from the secondary determination.

Images