A multi-field adaptive document digitization semantic topology reconstruction method and system

Abstract

The application discloses a multi-field adaptive document digitization semantic topology reconstruction method and system. The method comprises the following steps: obtaining a material document image of a target field, performing full-image geometric space scanning on the material document image, and outputting a spatial layout matrix; performing flow distribution feature extraction on the material document image, and performing dynamic non-uniform slicing to generate a plurality of local image grid slices carrying physical space topology features; inputting various local image grid slices and flat character text strings in the spatial layout matrix into a preset multimodal semantic reasoning large model to output standard heterogeneous text data streams carrying dependency logic; and analyzing the standard heterogeneous text data streams, and dynamically mapping and assembling a knowledge topology dependency network model corresponding to the target field. The method can accurately identify the text and its position, and avoids problems such as reading order disorder, information series or broken line caused by the lack of semantic understanding ability of a traditional optical character recognition algorithm.

Description

Technical Field

[0001] This application relates to the fields of computer vision and intelligent document processing, and in particular to a method and system for digital semantic topology reconstruction of documents that is adaptable to multiple fields. Background Technology

[0002] Currently, when enterprises process regulations, manuals, technical standards, and legal documents for various vertical industries (such as coal mine safety production, power grid operation and maintenance, and legal compliance management), they usually need to convert these unstructured documents into structured industry knowledge topology networks in order to achieve digital flow of business data and intelligent decision-making.

[0003] Traditional technical approaches mainly employ the following two methods: 1. Traditional Optical Character Recognition (OCR) pipeline architecture: This directly uses text detection and recognition algorithms, employing image processing and convolutional neural network techniques to convert pixel-based images into computer-readable text characters and output their corresponding absolute planar coordinates. 2. General multimodal end-to-end large model architecture: This leverages the multi-domain generalization capabilities of a large model, downsampling and scaling the document image before inputting it as a whole. It attempts to directly acquire formatting knowledge through prompts.

[0004] However, the aforementioned existing technologies have significant limitations, specifically: Traditional optical character recognition (OCR) algorithms "recognize characters but not their meanings," completely lacking semantic understanding capabilities. Essentially, OCR is a static mapping algorithm between geometric space and pure symbols. It can only identify a "character" or a "number" at a certain coordinate position, but it completely fails to understand the physical meaning of that character within its industry context. When faced with complex layouts in vertical domains (such as multi-column layouts, cross-page tables, lineless tables, or causal logic diagrams), traditional algorithms, lacking any semantic context awareness, can only rigidly splice text according to physical coordinates from left to right and top to bottom. This easily leads to disordered reading order, disastrous information loss due to misalignment or line breaks, and it completely fails to capture the deep topological relationships such as causality, membership, or constraints implicit between text blocks.

[0005] High-density long documents suffer from recognition distortion under a pure large model architecture: Although the pure multimodal end-to-end large model has semantic generalization ability, it lacks the pixel perception and localization ability at the original image level. Direct scaling of large images into the input can easily lead to the loss of local details, resulting in serious numerical distortion or missed detection of critical data such as coal mine technical parameters and power grid technical parameters during reconstruction. This fails to meet the hard requirements of data accuracy and precision in industries such as energy and law. Summary of the Invention

[0006] The main purpose of this application is to provide a method and system for digital semantic topology reconstruction of documents that is adaptable to multiple fields, so as to solve the technical problems in the prior art.

[0007] To achieve the above objectives, this application provides a method for digital semantic topology reconstruction of documents adapted to multiple domains, the method comprising: S10, acquire the document image of the target field, and use the preset page layout geometric feature localization network to perform full-image geometric space scanning on the document image, and output a spatial layout matrix containing flat character text strings and their two-dimensional pixel absolute bounding boxes; S20: Based on the spatial layout matrix, the streaming distribution features of the document images are extracted, and dynamic non-uniform slicing is performed through the dynamic window slicing algorithm of information entropy statistics to generate multiple local image grid slices carrying physical spatial topological features. S30, input various local image grid slices and flat character text strings in the spatial layout matrix into a preset multimodal semantic reasoning large model, realize deep feature-level alignment and joint reasoning of spatial absolute coordinates, character semantics and multi-domain context through multimodal cross attention mechanism, and output a standard heterogeneous text data stream carrying dependency logic; S40, parse the standard heterogeneous text data stream, and execute the graph database writing instruction by adaptively calling the corresponding domain's schema mapping script through the rule decision layer, dynamically mapping and assembling to construct a knowledge topology dependency network model corresponding to the target domain.

[0008] Optionally, step S10 includes the following steps: S101, the document image is scanned using the physical layout tree probe module, structure probe module, text probe module, and coordinate probe module embedded in the page layout geometric feature localization network. The structure probe module collects physical boundary indicators for column and paragraph layouts. The coordinate probe module extracts the two-dimensional pixel absolute bounding box coordinates of all single-line text in the document image without discrimination using multi-layer convolution and anchor-free bounding box regression algorithms. The text probe module identifies non-semantic flat character text strings within the two-dimensional pixel absolute bounding boxes. The physical layout tree probe module collects page block type identifiers. S102, the text probe is invoked to perform coarse character recognition, and the non-semantic flat character text string within the two-dimensional pixel absolute bounding box is standardized and regularized, and uniformly converted into a spatial layout matrix.

[0009] Optionally, step S20 includes: S201, Read the two-dimensional coordinate distribution data of the spatial layout matrix in the preset page space based on the dynamic slice layer; S202, using a dynamic window slicing algorithm based on information entropy statistics, the document image is traversed by sliding windows, the pixel distribution density of characters in each sliding window is statistically analyzed, the mean and standard deviation of character density in the local layout area are calculated, and the mean and standard deviation of information entropy of character pixel distribution in each sliding window are calculated. S203, calculate the adaptive density threshold and the adaptive information entropy threshold. When the entity character flow density of a certain local typesetting area exceeds the adaptive density threshold and the information entropy exceeds the adaptive information entropy threshold, dynamically expand the geometric boundary of the current sliding window, generate a dynamic cropping bounding box, and perform dynamic non-uniform slicing on the document image to obtain local image grid slices.

[0010] Optionally, step S30 includes: S301, the visual feature vectorization of each local image grid slice is performed using the visual encoding layer of the multimodal semantic reasoning big model to obtain visual feature vectors, and the text feature vectorization of the non-semantic flat character text string in the spatial layout matrix is ​​performed using the text encoding layer to obtain text feature vectors. S302, at the input end of the multimodal semantic reasoning large model, the visual feature vector and the text feature vector are subjected to cross-modal linear projection and splicing to assemble a physical topology feature network with unified feature level; S303, input the natural language text containing the target domain industry entity classification and hierarchical logical constraint prompts as an external industry control instruction into the multimodal semantic reasoning big model, activate the corresponding logical semantic call chain, and trigger the matching of the vertical latent semantic space of the target domain within the multimodal semantic reasoning big model; S304. Using a multimodal cross-attention mechanism, the visual feature vector and the text feature vector are mapped to a query matrix, a key matrix, and a value matrix, respectively. The logical dependency score of the target industry entity is calculated, and the standard heterogeneous text data stream is decoded and output end-to-end in an integrated manner.

[0011] Optionally, the dynamic mapping assembly constructs a knowledge topology dependency network model corresponding to the target domain, including: S401, Based on the multi-domain knowledge mapping layer, the standard heterogeneous text data stream is received in real time, and the standard heterogeneous text data stream is parsed to obtain vertical industry entity attributes, entity dependency relationships and logical dependency scores. S402, based on the domain tags in the standard heterogeneous text data stream, adaptively call the corresponding domain schema mapping script to convert the vertical industry entities in the standard heterogeneous text into graph nodes in the graph database, and assign corresponding attribute fields to the graph nodes; S403, based on the dependency logic and logical dependency score carried in the standard heterogeneous text data stream, directed connection edges between the nodes are dynamically created in real time within the graph database, thereby dynamically assembling a knowledge topology dependency network model oriented towards the target vertical domain without a preset static physical model.

[0012] The logical dependency score refers to a numerical indicator calculated within the multimodal semantic reasoning model through a multimodal cross-attention mechanism that incorporates a global cross-modal spatial distance bias matrix. It is used to quantitatively measure the strength of logical dependency between entities in different industries in both semantic and physical dimensions. The logical dependency score integrates text semantic relevance and page pixel geometric distance penalty, serving as the basis for determining whether directed connections are created between nodes in the graph database and for determining the initial adjacency matrix weights. It is then input into a two-layer graph convolutional neural network to perform deep graph feature updates and topological alignment.

[0013] Furthermore, to achieve the above objectives, embodiments of this application also provide a document digitization semantic topology reconstruction system adapted to multiple domains, the system comprising: The precision physical perception layer is used to acquire images of documents in the target domain and to perform full-image geometric space scanning of the document images using a pre-set layout geometric feature localization network, outputting a spatial layout matrix containing flat character text strings and their two-dimensional pixel absolute bounding boxes. The dynamic slice alignment layer is used to extract streaming distribution features from document images based on spatial layout matrix, and to perform dynamic non-uniform slicing through a dynamic window slicing algorithm based on information entropy statistics, generating multiple local image grid slices carrying physical spatial topological features. The multimodal semantic reasoning layer is used to input various local image grid slices and flat character text strings in the spatial layout matrix into a pre-set multimodal semantic reasoning large model. Through the multimodal cross attention mechanism, it realizes deep feature-level alignment and joint reasoning of spatial absolute coordinates, character semantics and multi-domain context, and outputs a standard heterogeneous text data stream carrying dependency logic. A multi-domain knowledge mapping layer is used to parse the standard heterogeneous text data stream. Through the rule decision layer, it adaptively calls the schema mapping script of the corresponding domain to execute the graph database writing instruction, and dynamically maps and assembles to construct a knowledge topology dependency network model corresponding to the target domain.

[0014] Optionally, the precision physical sensing layer includes: Structural probes are used to collect physical boundary indicators of the column and paragraph layout in the physical layout of the document image. A text probe is used to connect to the text encoding layer through character recognition and collect non-semantic flat character text strings that are detached from the industry vertical context within the absolute bounding box of the two-dimensional pixels. Coordinate probes are used to collect the absolute pixel coordinate bounding boxes corresponding to each single-line text block without discrimination through multi-layer convolution and anchorless bounding box regression algorithms. The information recombination unit is used to standardize and recombine the information output by the probes and convert it into the spatial layout matrix.

[0015] Optionally, the dynamic slice alignment layer includes: A sliding feature window is used to perform gridded traversal of local planar coordinates and character distribution in the spatial layout matrix, and to continuously scan and statistically analyze the streaming distribution features extracted within the window. The statistical feature unit is used to stream-calculate the mean and standard deviation of character density in the local layout area, as well as the mean and standard deviation of information entropy of character pixel distribution, based on the scan data of the sliding feature window.

[0016] Optionally, the multimodal semantic reasoning layer includes: The feature vectorization unit is used to generate visual feature vectors for local slices and text feature vectors for flat characters, respectively. Physical topology feature network is used to perform cross-modal linear projection and splicing of various visual feature vectors and text feature vectors to form a unified multimodal spatial skeleton network; Logical semantic call chain, used to dynamically switch and activate the potential decoding feature space of the corresponding vertical field according to industry control instructions; The text output unit is used to achieve deep feature-level alignment and joint reasoning of spatial absolute coordinates, character semantics and multi-domain context through a multimodal cross-attention mechanism, and outputs a standard heterogeneous text data stream carrying dependency logic.

[0017] Optionally, the multi-domain knowledge mapping layer includes: The text parsing unit is used to parse the standard heterogeneous text data stream; The schema mapping script is used to execute instructions to write nodes and directed edges to the target graph database; The streaming graph alignment unit is used to perform feature updates and alignment on the initial graph network structure using a two-layer graph convolutional neural network. The model building unit is used to dynamically map and assemble a knowledge topology dependency network model corresponding to the target domain. The verifier is used to verify the closed-loop integrity and topological consistency of the final generated digital knowledge topology-dependent network model.

[0018] Compared with the prior art, the beneficial effects of this application are as follows: The document digitization semantic topology reconstruction method adapted to multiple fields provided in this application uses a page geometric feature localization network to perform full-image geometric space scanning of the document image in the target field, which can accurately identify the text and its position, avoiding problems such as disordered reading order, information serialization or line breaks caused by the lack of semantic understanding ability in traditional optical character recognition algorithms.

[0019] Simultaneously, this method combines spatial layout matrix for streaming distribution feature extraction and dynamic non-uniform slicing, then concatenates the slices with the text and inputs them into a multimodal semantic reasoning model. This achieves deep feature-level alignment and joint reasoning of spatial absolute coordinates, character semantics, and multi-domain context. This process not only preserves the pixel-aware localization capability at the original image level, avoiding the loss of local details and data accuracy issues caused by scaling large images in a pure large model architecture, but also captures deep knowledge topological relationships between text blocks, constructing a knowledge topology dependency network model corresponding to the target domain. This meets the stringent requirements for data accuracy and precision in industries such as energy and law. Attached Figure Description

[0020] Figure 1 A flowchart illustrating a document digitization semantic topology reconstruction method adapted to multiple domains provided in this application embodiment; Figure 2 The structural block diagram of the document digitization semantic topology reconstruction system adapted to multiple fields provided in the embodiments of this application is shown.

[0021] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0022] It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of the application. Rather, these embodiments are provided to make the disclosure more thorough and complete, and to fully convey the scope of the disclosure to those skilled in the art.

[0023] Traditional optical character recognition algorithms, when processing complex typesetting documents, can only recognize characters but lack semantic understanding capabilities, leading to disordered reading order, information being strung together or broken lines, and failing to capture deep knowledge topological relationships. Meanwhile, general-purpose multimodal large models, when processing high-density, long documents, are prone to losing local details due to image scaling, resulting in numerical distortion or missed detection of high-precision key data, making it difficult to meet the industry's requirements for high data accuracy.

[0024] To address the aforementioned technical problems, embodiments of this application provide a document digital semantic topology reconstruction method adaptable to multiple domains. This method can be executed by a computer, such as... Figure 1 As shown, the method may include the following steps: S10, acquire the document image of the target domain, and use a preset layout geometric feature localization network to perform a full-image geometric space scan of the document image, outputting a spatial layout matrix M containing flat character text strings and their two-dimensional pixel absolute bounding boxes.

[0025] Specifically, the document images include unstructured vertical industry technical standards and procedures documents with multi-column nested layouts, barless tables, or logic diagrams. A pre-built page layout geometric feature localization network (based on a ResNet-50 backbone network combined with an FPN architecture and an anchor-free bounding box regression head) performs a full-image geometric spatial scan of the document images. This network is a high-precision geometric feature localization network pre-trained on a general document layout dataset. It is independent of the prior business logic of the target domain and does not require pre-constructing a business physical model for the target domain. It outputs a spatial layout matrix containing flat character text strings and their two-dimensional pixel absolute bounding boxes. For example, this model can use methods based on connected component analysis or sliding window detection to locate the text and output a simple list containing the string content of each text block and its coordinates of the top-left and bottom-right corners in the image. This forms a preliminary spatial layout matrix M, which records the text content and its approximate physical location.

[0026] Here, the spatial layout matrix M refers to a structured matrix used to store character text strings extracted from document images, along with their corresponding physical coordinates and layout attribute information. This matrix not only records non-semantic flat character text content... It also records the system processing timestamp. 2D pixel absolute bounding box coordinates and page layout type identifiers (Such as paragraph headings, body text, table cells, or figure text, providing precise spatial geometric context for subsequent multimodal semantic understanding and physical space alignment.)

[0027] In an exemplary embodiment, step S10 may include the following steps: S101, the document image is scanned using the physical layout tree probe module, structure probe module, text probe module, and coordinate probe module embedded in the page layout geometric feature localization network; wherein, the structure probe module is used to collect physical boundary indicators of column and paragraph layout; the coordinate probe module extracts the two-dimensional pixel absolute bounding box coordinates of all single-line text in the document image without discrimination through multi-layer convolution and anchor-free bounding box regression algorithms. The text probe module is used to identify non-semantic flat character text strings within the absolute bounding box of the two-dimensional pixels. The physical layout tree probe module is used to collect the layout block type identifier. ; S102, the text probe is invoked to perform coarse character recognition, and the non-semantic flat character text string within the two-dimensional pixel absolute bounding box is standardized and uniformly converted into a spatial layout matrix. The mathematical expression is as follows: in, , This represents the total number of single-line text blocks extracted from the image of the entire document. For the identified non-semantic flat character text string, For the system to process timestamps, These are the corresponding 2D pixel absolute bounding box coordinates. Use it to identify page block types (such as paragraph headings, body text, table cells, or legend text).

[0028] S20: Based on the spatial layout matrix M, the document image is subjected to streaming distribution feature extraction. A sliding window is used to traverse the document image to statistically analyze the character pixel distribution density. The mean and standard deviation of the character density in the local layout area are calculated to determine the adaptive density threshold, and the mean and standard deviation of the information entropy are calculated to determine the adaptive information entropy threshold. When the character streaming density of a local layout area exceeds the adaptive density threshold and the information entropy exceeds the adaptive information entropy threshold, the geometric boundary of the sliding window is expanded synchronously in four directions: up, down, left, and right. During the expansion process, the character density and information entropy of the expanded area are recalculated in real time. When the character density of the expanded area drops to less than or equal to the adaptive density threshold, the information entropy drops to less than or equal to the adaptive information entropy threshold, or the expansion boundary touches the image edge / outer blank separation zone, the expansion stops. The finally obtained expanded rectangular boundary is used as a dynamic cropping boundary box for dynamic non-uniform slicing, generating multiple local image grid slices carrying physical spatial topological features.

[0029] In an exemplary embodiment, step S20 may include the following steps: S201, Read the spatial layout matrix based on the dynamic slice layer. Two-dimensional coordinate distribution data within a preset page space; S202 sets the initial sliding window using a dynamic window slicing algorithm based on information entropy statistics. Size is The document image is traversed using pixels and a sliding step of 64 pixels. The pixel distribution density of characters within each sliding window is counted, and the average character density of the local layout area is calculated. with standard deviation ; Meanwhile, in order to accurately measure the complexity of the local layout structure (distinguishing between pure continuous text paragraphs and high-density complex tables, multi-column nested areas, or logic diagrams), the current sliding window is calculated. Information entropy of internal character pixel distribution The mathematical formula is as follows: In the formula: Indicates the current sliding window The binary spatial information entropy based on the foreground and background distribution of a binary image is used to quantitatively measure the geometric complexity of a local layout. Indicates the current sliding window The proportion of foreground pixels of inner characters (i.e., non-whitespace text / lines / border pixels) to the total number of pixels in the window (i.e.) ,in This represents the total number of foreground pixels for the window characters. (Total number of pixels within the window). This represents the proportion of background pixels (blank area pixels) to the total number of pixels in the window, calculated using the following formula: ,in The total number of background pixels within the window; by convention, when or hour, ; When a sliding window traverses complex layouts (such as table borders, multi-column intersections, and areas with mixed symbols), the degree of overlap between characters and background pixels increases, leading to increased information entropy. The entropy increases significantly. During the process of traversing the entire image using a sliding window, the mean entropy of all sliding windows in the entire image is simultaneously statistically analyzed and calculated. with standard deviation .

[0030] S203, Calculate the adaptive density threshold With adaptive information entropy threshold When the entity character flow density in a certain local typesetting area exceeds the adaptive density threshold And the information entropy exceeds the adaptive information entropy threshold. At that time, with the current sliding window as the center, its geometric boundary is expanded synchronously and proportionally in four directions (up, down, left, and right) with a sliding step of 64 pixels; during the expansion process, the character density within the expanded area is recalculated in real time, and when the character density of the expanded area drops to less than or equal to the adaptive density threshold... The information entropy decreases to less than or equal to the adaptive information entropy threshold. The expansion stops when the boundary reaches the edge of the image / the blank dividing band outside the high-density layout area of ​​the image. The final expanded rectangle boundary is used as the dynamic cropping boundary box, and the document image is dynamically non-uniformly sliced ​​to obtain local image grid slices.

[0031] Specifically, adaptive density threshold and adaptive information entropy threshold The calculation formula is as follows: In the formula: Indicates an adaptive density threshold; Density weighting adjustment coefficient (range of values) (preferably 1.5). This represents the average character density across all sliding windows in the entire image. The standard deviation of character density; Indicates the adaptive information entropy threshold; This represents the mean of the information entropy of each sliding window in the entire graph; This represents the standard deviation of the information entropy of each sliding window in the entire graph; Information entropy weight adjustment coefficient (range of values) (Preferred to be 1.0). When the entity character flow density of a certain local typesetting area exceeds the adaptive density threshold... When, and the information entropy exceeds the adaptive information entropy threshold. At that time, the geometric boundary of the current sliding window is dynamically expanded to generate a dynamic cropping bounding box, and the document image is dynamically and non-uniformly sliced ​​to obtain a local image mesh slice carrying the physical space topology.

[0032] In one implementation, dynamic non-uniform slicing can be achieved through adaptive grid slicing rules. Specifically, based on the dynamic cropping bounding box, the entire document image is adaptively divided into several local rectangular regions of varying sizes, each of which is a local image grid slice. When these slices are generated, their boundaries can be set to align with the pixel grid of the original image, thereby ensuring that each slice carries the visual information of the area it covers and its relative position within the overall document. Thus, a series of image grid slices can be obtained, each slice carrying its corresponding physical spatial topological features.

[0033] S30, various local image grid slices and flat character text strings in the spatial layout matrix M are input into a preset multimodal semantic reasoning large model. Through the multimodal cross attention mechanism, deep feature-level alignment and joint reasoning of spatial absolute coordinates, character semantics and multi-domain context are realized, and a standard heterogeneous text data stream carrying dependency logic is output.

[0034] In an exemplary embodiment, step S30 may include the following steps: S301, the visual feature vectorization of each local image grid slice is performed using the visual encoding layer of the multimodal semantic reasoning big model to obtain visual feature vectors, and the text feature vectorization of the non-semantic flat character text string in the spatial layout matrix is ​​performed using the text encoding layer to obtain text feature vectors. S302, at the input end of the multimodal semantic reasoning large model, the visual feature vector and the text feature vector are subjected to cross-modal linear projection and splicing to assemble a physical topology feature network with unified feature level; S303, input the natural language text containing the target domain industry entity classification and hierarchical logical constraint prompts as an external industry control instruction into the multimodal semantic reasoning big model, activate the corresponding logical semantic call chain, and trigger the matching of the vertical latent semantic space of the target domain within the multimodal semantic reasoning big model; S304. Using a multimodal cross-attention mechanism, the visual feature vector and the text feature vector are mapped to a query matrix, a key matrix, and a value matrix, respectively. The logical dependency score of the target industry entity is calculated, and the standard heterogeneous text data stream is decoded and output end-to-end in an integrated manner.

[0035] Specifically, the large-scale multimodal semantic reasoning model is built upon a pre-trained VisionTransformer (VIT), specifically employing a large-size ViT (ViT-L / 14) architecture. The visual encoder includes an image patch embedding module and a multi-head self-attention network composed of multiple Transformer encoder blocks. The input image patch size is set to... The pixel and hidden dimension are set to 1024. Each local image grid slice is input into the visual coding layer, and after multi-head self-attention calculation and linear mapping, a corresponding fixed-dimensional local visual feature vector is output. Subsequently, tensor concatenation is performed on each local visual feature vector along the sequence axis. Specifically, if it includes... There are local image grid slices, and the shape of the visual feature vector of each slice is as follows: Then, in the sequence axis dimension, this The vectors are concatenated sequentially according to the spatial position order of the slices to form a shape of... Visual feature vectors (in The total length of the visual token sequence, which is the result of visual encoding of all local image mesh slices extracted from the whole image. This preserves the spatial sequence relationship between each local slice.

[0036] At the same time, in order to determine the absolute coordinates of continuous pixels in the spatial layout matrix M To convert tokens into discrete tokens, this application introduces a coordinate quantization operator. The spatial coordinate normalization network is set to range [0, 1000], and the quantization calculation formula is as follows: in, , These represent the height and width of the input document image, respectively; the quantized discrete values ​​(integers ranging from 0 to 1000) are mapped to specific position tokens in a preset position vocabulary (e.g., the value 120 is mapped to a position identifier).<loc_120> ), forming a position token sequence: Token The location token sequence is then concatenated with the text tokens segmented by Byte Pair Encoder (BPE) to assemble a sequence containing... A discrete sequence of tokens, input to a text encoding layer (using Transformer Decoder, hidden layer dimension...). Generate text feature vectors. .

[0037] In generating the visual feature vector With the text feature vector Then, the visual feature vector With the text feature vector Each feature is mapped to a unified feature mapping dimension in a unified space through its corresponding linear transformation matrix. (wherein, the unified feature mapping dimension) (Set to 2048), resulting in a unified feature mapping dimension. Visual alignment features of the completed alignment Text alignment features .

[0038] Furthermore, the multimodal cross-attention mechanism refers to an attention mechanism used in large multimodal models, which allows information interaction and alignment between different modalities (such as image features and text features). Through this mechanism, the model can effectively fuse information from visual spatial coordinates, character semantics, and multi-domain context to achieve joint reasoning at the deep feature level.

[0039] Next, the specific tensor mapping, splicing, and attention solving process in step S30 will be described in more detail: S301, the visual encoding layer of the multimodal semantic reasoning big model is used to perform visual feature vectorization on each of the local image grid slices.

[0040] S302, at the input end of the multimodal semantic reasoning large model, a visual linear projection matrix is ​​used. With text linear projection matrix (Where the unified feature mapping dimension is set to) The visual feature vector and the text feature vector are mapped to a unified feature dimension. : In the formula: The visual feature vector obtained by extracting the local visual feature vectors of the visual coding layer (such as ViT-L / 14) in step S301 and concatenating them in spatial order is called the visual feature vector. The total length of the visual token sequence, which is the result of visual encoding of all local image mesh slices extracted from the whole image. The hidden layer dimension of the visual encoding layer ( ); This is a visual linear projection matrix used to project visual features from their original dimensions. Mapping to a unified feature dimension ; This is the visual feature matrix (i.e., visual alignment features) obtained after mapping. This refers to the text feature vector extracted from the spatial layout matrix M by the text encoding layer in step S301. The total length of the text token sequence generated by concatenating the position token sequence and the text token in the spatial layout matrix M. The hidden layer dimension of the text encoding layer (set to) ); This is a linear projection matrix for the text, used to transform text features from the original feature dimension. Linear mapping to a unified feature mapping dimension ; This is the text feature matrix (i.e., text alignment features) obtained after mapping.

[0041] Subsequently, the mapped visual feature matrix is ​​applied along the sequence axis (row dimension). With text feature matrix Tensor concatenation is performed to assemble and generate a physical topology feature network with a unified feature level. Its algebraic superposition formula is: In the formula: For the unified feature-level physical topology feature network (multimodal spatial skeleton network) generated by assembly; To map the visual feature matrix along the sequence axis dimension With text feature matrix The operation performed is a row-by-row tensor concatenation.

[0042] Here, the spliced ​​unified feature-level physical topology feature network The dimension on the sequence axis is the visual sequence length. With text sequence length sum The feature mapping dimension is uniformly maintained as This restores the operational rules of tensor algebra, thereby providing a standard high-dimensional input tensor for joint inference in subsequent multimodal cross-attention mechanisms.

[0043] S303, input the natural language text containing the target domain industry entity classification and hierarchical logical constraint prompts as an external industry control instruction into the multimodal semantic reasoning big model, activate the corresponding logical semantic call chain, and trigger the matching of the vertical latent semantic space of the target domain within the multimodal semantic reasoning big model; S304. Using a multimodal cross-attention mechanism, the visual feature vector and the text feature vector are mapped to a query matrix, a key matrix, and a value matrix, respectively. The logical dependency score of the target industry entity is calculated, and the standard heterogeneous text data stream is decoded and output end-to-end in an integrated manner.

[0044] Specifically, the formula for calculating the logical dependency score of the target industry entity is as follows: In the formula: (Q,K,V) represents the logical dependency score of the target industry entity, which incorporates spatial layout constraints. Its output dimension is consistent with the value matrix V; Q, K, and V are respectively the scores obtained by using the projection matrix. Mapping multimodal fusion features The resulting query matrix, key matrix, and value matrix; dk is the feature scalar scaling factor; This represents the cross-modal alignment adaptive tuning coefficient, used to dynamically scale the alignment coupling strength of multimodal features (value range is...). (preferably 1.0). Strict matching for dimensions of A global cross-modal spatial distance bias matrix of dimension is used to impose local penalty constraints on the spatial location of unstructured physical layouts.

[0045] Furthermore, the cross-modal spatial distance bias matrix In the middle, the first The first unstructured text entity and the second Penalty value for physical distance between structured table cells The global piecewise operator is defined as follows: If both the i-th text entity and the j-th cell have original pixel absolute geometric center coordinates ,but If it involves external industry control instructions or semantics without physical coordinates ,but .

[0046] In the formula: The attenuation scale factor representing the physical distance in space is calculated using the following formula: (in , (These are the height and width of the input document image, respectively), used to adjust the constraint weight of spatial physical proximity on the semantic association strength; This represents the geometric center coordinates of the i-th unstructured text entity in the absolute pixel coordinate system of the original image. This represents the absolute geometric center coordinates of the j-th structured table cell in the absolute pixel coordinate system of the original image. This bias matrix successfully applied a localized attenuation constraint to the spatial location of the unstructured physical layout, ultimately validated by the text parsing operator. The end-to-end integrated decoding output is a JSON-formatted standard heterogeneous text data stream containing entity attributes, dependencies, and dependency scores.

[0047] Wherein, the text parsing operator A pre-defined structured sequence decoding operator is used to receive the text token sequence output by the multimodal semantic reasoning big model, and extract the vertical industry entity attributes, entity dependency relationships and corresponding logical dependency scores from the text token sequence through syntax tree parsing and regular expression matching. This results in a JSON-formatted standard heterogeneous text data stream.

[0048] S40, parse the standard heterogeneous text data stream, and execute the graph database writing instruction by adaptively calling the corresponding domain's schema mapping script through the rule decision layer, dynamically mapping and assembling to construct a knowledge topology dependency network model corresponding to the target domain.

[0049] Specifically, the knowledge topology dependency network model refers to a knowledge model represented by a graph structure, where nodes represent industry entities or concepts, and edges represent dependencies, causal relationships, or membership relationships between these entities or concepts. This model aims to transform unstructured document information into a structured, reasonable industry knowledge graph.

[0050] In terms of specific model architecture, the knowledge topology dependency network model is composed of four core architectural layers working together: a feature input layer, a graph structure construction layer, a graph feature evolution layer (GCN encoder), and a topology verification and self-healing layer. Feature Input Layer: Responsible for receiving standard heterogeneous text data streams, extracting entity attributes, category labels and physical space coordinates, and constructing multimodal high-dimensional node feature matrices and initial feature vectors; Graph structure construction layer: Responsible for instantiating nodes and directed edges in the graph database by calling the schema mapping script based on entity dependency logic and dependency scores, forming an initial adjacency matrix containing explicit weights and implicit spatial constraints; Graph Feature Evolution Layer (GCN Encoder): Employs a two-layer graph convolutional neural network (GCN) that combines a self-loop adjacency matrix and a degree matrix to perform deep semantic and topological feature aggregation of multi-hop neighbor nodes.

[0051] Topology self-healing layer: Embedded isolated node detector, non-logic loop detector and constraint chain integrity checker, to perform structural integrity closed-loop verification and automatic repair of abnormal links on the evolved graph network.

[0052] In an exemplary embodiment, the dynamic mapping assembly constructing a knowledge topology dependency network model corresponding to the target domain includes: S401, Based on a multi-domain knowledge mapping layer, the standard heterogeneous text data stream is received in real time, and the text parsing operator is used to... Extract the attributes, dependencies, and logical dependency scores of entities in the vertical industry; S402, based on the domain tags in the standard heterogeneous text data stream, adaptively call the corresponding domain schema mapping script to convert the vertical industry entities in the standard heterogeneous text data stream into graph nodes in the graph database. The graph nodes are assigned corresponding attribute fields; and directed connection edges between nodes are created in the graph database based on entity dependency logic and logical dependency scores to form an initial graph network structure.

[0053] S403, In order to perform feature updates and topology alignment on the initial graph network, the initial node feature matrix is ​​first constructed. For each node in the graph database Extract the parsed entity text attribute strings With entity category labels The semantic vectors are extracted through a pre-trained language model encoder. and The data is then input into a multilayer perceptron (MLP) for feature fusion and dimensionality projection. Full map Feature vectors of each node Vertically stacked and spliced ​​to form the initial node feature matrix .

[0054] Subsequently, a two-layer graph convolutional neural network (GCN) was used based on the initial node feature matrix. Adjacency matrix containing dependency weights The initial graph network structure is updated using streaming graph features. The node feature update propagation formula is: in, To include the identity matrix The self-loop adjacency matrix, for The corresponding degree matrix (the i-th element on the diagonal) ), For the first Layer-trainable weight matrix, ReLU is a non-linear activation function. Through feature propagation in two layers of GCN, the node successfully aggregates the deep semantic and topological dependency features of multi-hop neighbor nodes; finally, the verifier performs isolated node detection, non-logical loop detection, and constraint chain integrity detection to verify and assemble a closed-loop complete and topologically consistent digital knowledge topological dependency network model.

[0055] In one implementation, parsing standard heterogeneous text data streams can be accomplished using a series of predefined rules and pattern matching algorithms. For example, rules can be set to identify specific keywords or phrases in the text and treat them as nodes in a knowledge graph. Dependencies between entities can be easily extracted using conjunctions appearing in the text, such as "belongs to," "contains," and "causes." Subsequently, these extracted nodes and relationships can be written into a basic graph database, thereby constructing a preliminary knowledge topology dependency network model. This model can represent explicitly mentioned entities in the text and their direct relationships.

[0056] The document digitization semantic topology reconstruction method adapted to multiple fields provided in this application uses a page geometric feature localization network to perform full-image geometric space scanning of the document image in the target field, which can accurately identify the text and its position, avoiding problems such as disordered reading order, information serialization or line breaks caused by the lack of semantic understanding ability in traditional optical character recognition algorithms.

[0057] Simultaneously, this method combines spatial layout matrix for streaming distribution feature extraction and dynamic non-uniform slicing, then concatenates the slices with the text and inputs them into a multimodal semantic reasoning model. This achieves deep feature-level alignment and joint reasoning of spatial absolute coordinates, character semantics, and multi-domain context. This process not only preserves the pixel-aware localization capability at the original image level, avoiding the loss of local details and data accuracy issues caused by scaling large images in a pure large model architecture, but also captures deep knowledge topological relationships between text blocks, constructing a knowledge topology dependency network model corresponding to the target domain. This meets the stringent requirements for data accuracy and precision in industries such as energy and law.

[0058] Based on the above embodiments, refer to Figure 2 Another embodiment of this application also provides a document digitization semantic topology reconstruction system adapted to multiple domains. The document digitization semantic topology reconstruction system 200 adapted to multiple domains may include the following modules: The precision physical perception layer 210 is used to acquire the document image of the target field, and to perform a full-image geometric space scan of the document image using a preset page geometric feature localization network, and output a spatial layout matrix M containing flat character text strings and their two-dimensional pixel absolute bounding boxes. The dynamic slice alignment layer 220 is used to extract streaming distribution features of document images based on spatial layout matrix M, and to perform dynamic non-uniform slicing through a dynamic window slicing algorithm based on information entropy statistics, generating multiple local image grid slices carrying physical spatial topological features. The multimodal semantic reasoning layer 230 is used to input various local image grid slices and flat character text strings in the spatial layout matrix M into a preset multimodal semantic reasoning large model. Through the multimodal cross attention mechanism, it realizes deep feature-level alignment and joint reasoning of spatial absolute coordinates, character semantics and multi-domain context, and outputs a standard heterogeneous text data stream carrying dependency logic. The multi-domain knowledge mapping layer 240 is used to parse the standard heterogeneous text data stream. Through the rule decision layer, it adaptively calls the schema mapping script of the corresponding domain to execute the graph database writing instruction and construct the initial graph structure. The initial graph network structure is updated and aligned using a two-layer graph convolutional neural network (GCN) to dynamically map and assemble the knowledge topology dependency network model corresponding to the target domain.

[0059] Specifically, the precision physical sensing layer 210 includes: Structure probe 2101 is used to collect physical boundary indicators of the column and paragraph layout in the physical layout of the document image; Text probe 2102 is used to connect to the text encoding layer through character recognition and collect non-semantic flat character text strings that are detached from the industry vertical context within the two-dimensional pixel absolute bounding box; The coordinate probe 2103 is used to collect the absolute pixel coordinate bounding box corresponding to each single-line text block without discrimination through multi-layer convolution and anchorless bounding box regression algorithm. Information recombination unit 2104 is used to standardize and recombine the information output by the probe, and uniformly convert it into the spatial layout matrix. .

[0060] Furthermore, the dynamic slice alignment layer 220 includes: The sliding feature window 2201 is used to perform gridded traversal of the local planar coordinates and character distribution in the spatial layout matrix, and to continuously scan and statistically analyze the streaming distribution features extracted within the window. The statistical feature unit 2202 is used to dynamically calculate the mean and standard deviation of character density in the local layout area, as well as the mean and standard deviation of information entropy of character pixel distribution, based on the scan data of the sliding feature window.

[0061] Furthermore, the multimodal semantic reasoning layer 230 includes: The feature vectorization unit 2301 is used to generate visual feature vectors of local slices and text feature vectors of flat characters, respectively. The Physical Topology Feature Network 2302 is used to perform cross-modal linear projection and stitching of various visual feature vectors and text feature vectors to form a unified multimodal spatial skeleton network. ; Logical semantic call chain 2303 is used to dynamically switch and activate the potential decoding feature space of the corresponding vertical field according to industry control instructions; Text output unit 2304 is used to incorporate a global cross-modal spatial distance bias matrix. The multimodal cross-attention mechanism enables deep feature-level alignment and joint reasoning of spatial absolute coordinates, character semantics, and multi-domain context, outputting a standard heterogeneous text data stream carrying dependency logic.

[0062] In this embodiment, the multi-domain knowledge mapping layer 240 includes: Text parsing unit 2401 is used to parse the standard heterogeneous text data stream; Schema mapping script 2402 is used to execute instructions to write nodes and directed edges to the target graph database; The flow graph alignment unit 2403 is used to extract entity attributes and types to construct the initial node feature matrix. The initial graph structure is updated and topologically aligned using a two-layer graph convolutional neural network based on an adjacency matrix containing dependency scores. The model building unit 2404 is used to dynamically map and assemble a knowledge topology dependency network model corresponding to the target domain. Verifier 2405 is used to perform isolated node detection, non-logical loop detection, and constraint chain integrity detection to verify the closed-loop integrity and topological consistency of the final generated digital knowledge topology-dependent network model.

[0063] In the description of this application, it should be noted that the terms "first", "second", and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0064] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0065] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the coupling or direct coupling or communication connection shown or discussed may be through some communication interface; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0066] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0067] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0068] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0069] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the technical scope disclosed in this application. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be covered within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

[0070] Furthermore, although the operations of the method of this application are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

Claims

1. A method for reconstructing the semantic topology of digital documents adaptable to multiple domains, characterized in that, The method includes the following steps: S10, acquire the document image of the target field, and use the preset page layout geometric feature localization network to perform full-image geometric space scanning on the document image, and output a spatial layout matrix containing flat character text strings and their two-dimensional pixel absolute bounding boxes; S20: Based on the spatial layout matrix, the streaming distribution features of the document images are extracted, and dynamic non-uniform slicing is performed through the dynamic window slicing algorithm of information entropy statistics to generate multiple local image grid slices carrying physical spatial topological features. S30, input various local image grid slices and flat character text strings in the spatial layout matrix into a preset multimodal semantic reasoning large model, realize deep feature-level alignment and joint reasoning of spatial absolute coordinates, character semantics and multi-domain context through multimodal cross attention mechanism, and output a standard heterogeneous text data stream carrying dependency logic; S40, parse the standard heterogeneous text data stream, and execute the graph database writing instruction by adaptively calling the corresponding domain's schema mapping script through the rule decision layer, dynamically mapping and assembling to construct a knowledge topology dependency network model corresponding to the target domain.

2. The document digitization semantic topology reconstruction method adapted to multiple domains according to claim 1, characterized in that, Step S10 includes the following steps: S101, the document image is scanned using the physical layout tree probe module, structure probe module, text probe module, and coordinate probe module embedded in the page layout geometric feature localization network. The structure probe module collects physical boundary indicators for column and paragraph layouts. The coordinate probe module extracts the two-dimensional pixel absolute bounding box coordinates of all single-line text in the document image without discrimination using multi-layer convolution and anchor-free bounding box regression algorithms. The text probe module identifies non-semantic flat character text strings within the two-dimensional pixel absolute bounding boxes. The physical layout tree probe module collects page block type identifiers. S102, the text probe is invoked to perform coarse character recognition, and the non-semantic flat character text string within the two-dimensional pixel absolute bounding box is standardized and regularized, and uniformly converted into a spatial layout matrix.

3. The method according to claim 1, characterized in that, Step S20 includes: S201, Read the two-dimensional coordinate distribution data of the spatial layout matrix in the preset page space based on the dynamic slice layer; S202, using a dynamic window slicing algorithm based on information entropy statistics, the document image is traversed through a sliding window to count the pixel distribution density of characters within each sliding window, and the mean and standard deviation of character density in the local layout area are calculated. S203, calculate the adaptive density threshold. When the entity character flow density of a certain local typesetting area exceeds the adaptive density threshold, dynamically expand the geometric boundary of the current sliding window, generate a dynamic cropping bounding box, and perform dynamic non-uniform slicing on the document image to obtain local image grid slices.

4. The method according to claim 1, characterized in that, Step S30 includes: S301, the visual feature vectorization of each local image grid slice is performed using the visual encoding layer of the multimodal semantic reasoning big model to obtain visual feature vectors, and the text feature vectorization of the non-semantic flat character text string in the spatial layout matrix is ​​performed using the text encoding layer to obtain text feature vectors. S302, at the input end of the multimodal semantic reasoning large model, the visual feature vector and the text feature vector are subjected to cross-modal linear projection and splicing to assemble a physical topology feature network with unified feature level; S303, input the natural language text containing the target domain industry entity classification and hierarchical logical constraint prompts as an external industry control instruction into the multimodal semantic reasoning big model, activate the corresponding logical semantic call chain, and trigger the matching of the vertical latent semantic space of the target domain within the multimodal semantic reasoning big model; S304. Using a multimodal cross-attention mechanism, the visual feature vector and the text feature vector are mapped to a query matrix, a key matrix, and a value matrix, respectively. The logical dependency score of the target industry entity is calculated, and the standard heterogeneous text data stream is decoded and output end-to-end in an integrated manner.

5. The method according to claim 1, characterized in that, The dynamic mapping assembly constructs a knowledge topology dependency network model corresponding to the target domain, including: S401, Based on the multi-domain knowledge mapping layer, the standard heterogeneous text data stream is received in real time, and the standard heterogeneous text data stream is parsed to obtain vertical industry entity attributes, entity dependency relationships and logical dependency scores. S402, based on the domain tags in the standard heterogeneous text data stream, adaptively call the corresponding domain schema mapping script to convert the vertical industry entities in the standard heterogeneous text into graph nodes in the graph database, and assign corresponding attribute fields to the graph nodes; S403, based on the dependency logic and logical dependency score carried in the standard heterogeneous text data stream, directed connection edges between the nodes are dynamically created in real time within the graph database, thereby dynamically assembling a knowledge topology dependency network model oriented towards the target vertical domain without a preset static physical model.

6. A document digitization semantic topology reconstruction system adaptable to multiple domains, characterized in that, include: The precision physical perception layer is used to acquire images of documents in the target domain and to perform full-image geometric space scanning of the document images using a pre-set layout geometric feature localization network, outputting a spatial layout matrix containing flat character text strings and their two-dimensional pixel absolute bounding boxes. The dynamic slice alignment layer is used to extract streaming distribution features from document images based on spatial layout matrix, and to perform dynamic non-uniform slicing through a dynamic window slicing algorithm based on information entropy statistics, generating multiple local image grid slices carrying physical spatial topological features. The multimodal semantic reasoning layer is used to input various local image grid slices and flat character text strings in the spatial layout matrix into a pre-set multimodal semantic reasoning large model. Through the multimodal cross attention mechanism, it realizes deep feature-level alignment and joint reasoning of spatial absolute coordinates, character semantics and multi-domain context, and outputs a standard heterogeneous text data stream carrying dependency logic. A multi-domain knowledge mapping layer is used to parse the standard heterogeneous text data stream. Through the rule decision layer, it adaptively calls the schema mapping script of the corresponding domain to execute the graph database writing instruction, and dynamically maps and assembles to construct a knowledge topology dependency network model corresponding to the target domain.

7. The system according to claim 6, characterized in that, The precision physical sensing layer includes: Structural probes are used to collect physical boundary indicators of the column and paragraph layout in the physical layout of the document image. A text probe is used to connect to the text encoding layer through character recognition and collect non-semantic flat character text strings that are detached from the industry vertical context within the absolute bounding box of the two-dimensional pixels. Coordinate probes are used to collect the absolute pixel coordinate bounding boxes corresponding to each single-line text block without discrimination through multi-layer convolution and anchorless bounding box regression algorithms. The information recombination unit is used to standardize and recombine the information output by the probes and convert it into the spatial layout matrix.

8. The system according to claim 6, characterized in that, The dynamic slice alignment layer includes: A sliding feature window is used to perform gridded traversal of local planar coordinates and character distribution in the spatial layout matrix, and to continuously scan and statistically analyze the streaming distribution features extracted within the window. The statistical feature unit is used to stream-calculate the mean and standard deviation of character density in the local layout area, as well as the mean and standard deviation of information entropy of character pixel distribution, based on the scan data of the sliding feature window.

9. The system according to claim 6, characterized in that, The multimodal semantic reasoning layer includes: The feature vectorization unit is used to generate visual feature vectors for local slices and text feature vectors for flat characters, respectively. Physical topology feature network is used to perform cross-modal linear projection and splicing of various visual feature vectors and text feature vectors to form a unified multimodal spatial skeleton network; Logical semantic call chain, used to dynamically switch and activate the potential decoding feature space of the corresponding vertical field according to industry control instructions; The text output unit is used to achieve deep feature-level alignment and joint reasoning of spatial absolute coordinates, character semantics and multi-domain context through a multimodal cross-attention mechanism, and outputs a standard heterogeneous text data stream carrying dependency logic.

10. The system according to claim 6, characterized in that, The multi-domain knowledge mapping layer includes: The text parsing unit is used to parse the standard heterogeneous text data stream; The schema mapping script is used to execute instructions to write nodes and directed edges to the target graph database; The streaming graph alignment unit is used to perform feature updates and alignment on the initial graph network structure using a two-layer graph convolutional neural network. The model building unit is used to dynamically map and assemble a knowledge topology dependency network model corresponding to the target domain. The verifier is used to verify the closed-loop integrity and topological consistency of the final generated digital knowledge topology-dependent network model.

Images