Standard digital analysis method based on semantic layering and vector reconstruction
By employing semantic layering and vector reconstruction methods, the hierarchical logic, table, and formula structure of standard documents are automatically identified and reconstructed. This solves the problems of long processing cycles and error-proneness in existing technologies, enabling efficient digitization of standard documents and AI applications.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHENGDU INSTITUTE OF STANDARDIZATION (CHENGDU WTO TBT CONSULTING CENTER CHENGDU MARKET SUPERVISION & MANAGEMENT DATA APPLICATION CENTER)
- Filing Date
- 2026-04-16
- Publication Date
- 2026-05-26
Smart Images
Figure CN122090476A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of document parsing technology, and in particular to a standard digital parsing method based on semantic layering and vector reconstruction. Background Technology
[0002] In the field of national standardization strategy, the number of standard system documents is enormous and the format is complex, containing a large number of tables, formulas, images, footnotes, and multi-level clause structures. Traditional manual parsing methods rely on manual annotation and input, which has the following drawbacks: long processing time (a single complex standard may take several days to a week); errors are prone to occur due to hierarchical structure and nested symbols; and the lack of a unified data structure is not conducive to the reuse of standard knowledge and AI applications.
[0003] While current OCR and natural language processing technologies can extract text content, they struggle to maintain standard hierarchical logic, table relationships, and mathematical formula structures.
[0004] Therefore, there is an urgent need for a parsing method that can automatically identify the semantic hierarchy of standard documents, vectorize and reconstruct charts and formulas, and generate structured knowledge data to improve the digitization capabilities of standards. Summary of the Invention
[0005] The purpose of this invention is to provide a standard digitization parsing method based on semantic layering and vector reconstruction, thereby improving standard digitization capabilities.
[0006] To achieve the above objectives, this invention provides a standard digital parsing method based on semantic layering and vector reconstruction, comprising the following steps: Standard documents are acquired through high-precision scanning or electronic import, and then layout correction and noise removal are performed. An improved large language model combined with a regular pattern recognition algorithm is used to perform multi-level semantic hierarchical parsing on the standard file, extract key fields and generate a semantic hierarchy tree; The tables, formulas, and images in the standard file are vectorized and reconstructed to generate structured HTML table objects, vector formulas, and vector graphics, respectively. Based on the large language model, the structured data obtained after semantic hierarchical parsing and vector reconstruction is subjected to semantic consistency verification; The proofread structured data is output as a standard format XML file, and multiple knowledge graphs are automatically constructed based on the XML file.
[0007] After obtaining the standard file, the method further includes: Automatically detect and identify the text area, image area, formula area, and symbol area in the standard file.
[0008] Specifically, an improved large language model combined with a regular pattern recognition algorithm is used to perform multi-level semantic hierarchical parsing on the standard file, extracting key fields and generating a semantic hierarchy tree, including: Based on the self-attention mechanism of the improved large language model, the semantic dependencies between nodes in the standard file are identified, key fields are extracted, and a semantic hierarchy tree is generated. The structure set S of the semantic hierarchy tree consists of chapter units. Clause Unit and paragraph units The structure, and its hierarchical relationship are represented as follows: .
[0009] The specific steps for table vectorization reconstruction are as follows: The DOM mapping function f_DOM takes the column set C, row set R, and merge cell information of the table as input and generates a structured HTML table object T_html. The function relationship is represented as T_html = f_DOM(C, R, merge).
[0010] The specific steps for formula vectorization and reconstruction are as follows: The identified formula region is processed by the render function to generate a vector formula graphic F_SVG in SVG format. Its functional relationship is represented as F_SVG = render(L_LaTeX), where L_LaTeX is the LaTeX encoding of the formula.
[0011] The vectorized reconstruction of the formula also supports bidirectional conversion between MathType and LaTeX encoding formats.
[0012] The method further includes vector reconstruction of the symbols, specifically: Symbols are identified using a pattern matching algorithm, and a symbol library is automatically generated based on the symbol type and context, and then indexed in the document.
[0013] The XML file includes standard names, numbers, definitions, clauses, and chapter structures.
[0014] Among them, the automatically constructed knowledge graphs include standard relationship graphs, indicator graphs, organizational and personnel graphs, equipment and product graphs, and domain knowledge graphs.
[0015] This invention discloses a standard digitization parsing method based on semantic layering and vector reconstruction. The standard digitization parsing system based on semantic layering and vector reconstruction includes a file access and preprocessing module, a semantic layering parsing module, a vector reconstruction module, an AI semantic proofreading module, and a data output and knowledge graph construction module. The method acquires standard files through high-precision scanning or electronic import, and performs layout correction and noise removal. Using an improved large language model combined with a regular pattern recognition algorithm, the standard files undergo multi-level semantic layering parsing, extracting key fields and generating a semantic hierarchy tree. Tables, formulas, and images in the standard files are vectorized and reconstructed to generate structured HTML table objects, vector formulas, and vector graphics, respectively. Based on the large language model, the structured data obtained after semantic layering parsing and vector reconstruction undergoes semantic consistency verification. The verified structured data is output as a standard-format XML file, and multiple types of knowledge graphs are automatically constructed based on this XML file to improve the standard digitization capability. Attached Figure Description
[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.
[0017] Figure 1 This is a flowchart illustrating a standard digital parsing method based on semantic layering and vector reconstruction provided by the present invention.
[0018] Figure 2 This is a schematic diagram of the semantic hierarchy parsing structure of the standard document provided by this invention.
[0019] Figure 3 This is a schematic diagram illustrating the principle of table vectorization reconstruction provided by the present invention.
[0020] Figure 4 This is a flowchart of formula recognition and encoding conversion provided by the present invention.
[0021] Figure 5 This invention provides a knowledge graph generation logic diagram (showing the association between standard relationships and domain knowledge).
[0022] Figure 6 This is a schematic diagram illustrating the correspondence between the semantic hierarchy structure and the structure mapping matrix provided by this invention.
[0023] Figure 7 This is a structured process diagram of table parsing and DOM mapping provided by the present invention.
[0024] Figure 8 This is a flowchart illustrating the conversion process from LaTeX source code to SVG vector graphics provided by the present invention.
[0025] Figure 9 This is a schematic diagram of the steps of the standard digital parsing method based on semantic layering and vector reconstruction provided by the present invention. Detailed Implementation
[0026] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.
[0027] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.
[0028] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."
[0029] Please see Figures 1-9 This invention provides a standard digital parsing method based on semantic hierarchical and vector reconstruction, comprising the following steps: S101. Obtain standard documents through high-precision scanning or electronic import, and perform layout correction and noise removal.
[0030] Specifically, high-resolution scanners (600 dpi or higher) are used to scan paper standard documents to ensure that detailed information is captured. Dual-layer scanning technology is employed during the scanning process to acquire both sides of the document, and intelligent noise reduction algorithms are used to remove interference during scanning, ensuring clear and distortion-free document images. Existing electronic standard documents (such as PDF and Word formats) are directly imported into the system, and the document content is extracted through a file parsing engine for further OCR (Optical Character Recognition) processing. OCR technology not only recognizes text but can also parse complex content such as graphics, formulas, and tables. The processed text blocks are then used to generate structured data according to the original document's layout.
[0031] By using image processing algorithms based on region segmentation, combined with text density detection techniques (such as connected component analysis), plain text regions in a document are automatically identified. At this point, the document is divided into multiple parts, each labeled as a "text region".
[0032] This method uses edge detection and morphological analysis combined with deep learning algorithms (such as convolutional neural networks, CNNs) to detect chart regions. By analyzing the edges, lines, and regions of an image, it can automatically identify charts (such as line charts, bar charts, pie charts, etc.) and their locations.
[0033] The formula area automatically identifies mathematical symbols, numbers, and formula expressions within the document by analyzing their positions. This process combines character recognition and symbol localization technologies, employing image registration methods to identify formula regions and ensure that no formula information is missed.
[0034] The symbol area mainly includes unit symbols, mathematical symbols, and engineering symbols. Through symbol matching and pattern recognition technology, the system can accurately distinguish and extract information from various types of symbols.
[0035] The scanned document undergoes geometric correction, using image correction methods such as affine transformation to correct image tilt or distortion. Noise generated during the scanning process is removed, and denoising algorithms are employed to ensure the clarity of the final text areas.
[0036] S102. Using an improved large language model combined with a regular pattern recognition algorithm, perform multi-level semantic hierarchical parsing on the standard file, extract key fields, and generate a semantic hierarchy tree.
[0037] Specifically, a modified Large Language Model (LLM) combined with a regular pattern recognition algorithm is used for semantic hierarchical parsing. This model, trained based on deep learning (e.g., the Transformer architecture), is capable of understanding and generating context-aware text information. In standard document parsing, LLM is used to identify chapter structure and semantic relationships in the text, performing syntactic analysis and semantic disambiguation. Regular pattern recognition, by defining a series of regular expression patterns, allows the system to match content with fixed formats (such as numbering, headings, and units) in standard documents. This helps to efficiently identify document structure and extract information such as chapters, clauses, and appendices. Simultaneously, for document images that have undergone layout analysis (already completed in S101), a lightweight convolutional neural network (CNN) encoder is used to extract visual features of each text block (such as a heading or a paragraph). These features include, but are not limited to: Spatial location characteristics: the bounding box coordinates (x, y, width, height) of the block and its relative position on the page.
[0038] Text style features: Attributes such as font size, bold, italics, and underline are recognized through OCR and image processing.
[0039] Layout relationship characteristics: relative indentation distance between blocks, alignment (such as left alignment, center alignment), and visual spacing with predecessor / successor blocks.
[0040] The improved LLM does not only process plain text; it integrates a visual-semantic cross-attention module. In this module, textual semantic feature vectors serve as "queries," while visual typography feature vectors serve as "keys" and "values." The model dynamically assigns attention weights to each semantic unit in the text (e.g., "Chapter 3") to "attention" to associated visual features (such as the "large font, centered, bold" characteristics typically found in chapter titles). This process enables the model to learn the implicit correspondence between specific visual styles and specific semantic levels. For example, the model learns that in most standard documents, text that occupies the top of the page, is large, bold, and centered, has a very high probability of being a "standard name" or a "first-level chapter title"; while text with indentation, numbering such as "5.2.1," and a slightly smaller font corresponds to a "third-level clause."
[0041] When document formatting is non-standard (e.g., missing numbering, unbold headings), a single feature stream can lead to ambiguity. The innovation of this system lies in introducing a structural ambiguity resolver. This resolver receives dual evidence from both the textual semantic stream and the visual-semantic fusion stream (e.g., a text block semantically resembles a "definition," but visually resembles a "heading"). Based on experience trained on a large number of multi-domain standard documents, the resolver performs a weighted evaluation of conflicting evidence. It may contain a series of lightweight rules and a small neural network classifier, ultimately outputting a joint confidence score and a hierarchical label prediction (e.g., "chapter title," "clause text," "figure caption"). The resolver also utilizes the initially parsed hierarchical relationships to perform logical coherence checks. For example, if the model identifies a text block as "3.1.2," but "3.1" is not detected before it, the system triggers a backtracking analysis to review visual features and contextual semantics to correct possible recognition errors or confirm that this is an irregular numbering jump.
[0042] like Figure 2As shown, this diagram illustrates the hierarchical relationship of a standard document from chapters to clauses to paragraphs, as well as the association between normative references, terminology definitions, and chart indexes, serving to explain how the semantic hierarchy tree is constructed. The system automatically identifies the hierarchical structure of the document (such as chapters, clauses, paragraphs, and items) and divides the document into multiple levels. The text of each level is extracted and labeled as the corresponding level, typically divided into a chapter-clause-paragraph-item hierarchy. For example, in standard documents, chapter titles are usually identified with "large font size" and "bold" formatting, while clauses and paragraphs usually have uniform numbering. Utilizing these formatting features, the model can automatically detect and identify the hierarchical relationships between chapters.
[0043] Using regular expressions and text positioning algorithms, the system automatically identifies numbering (e.g., "Chapter 1" or "2.1") and headings (e.g., "Standard Name") in standard documents. It extracts sections from the text that contain standard definitions, unit descriptions, and terminology. For each definition, the system automatically extracts its keywords and descriptions and matches them with relevant domain knowledge. Citations in the document (e.g., mutual references between standards) are extracted using citation identifiers, and the system automatically identifies and labels the citation source and object. Based on the extracted key fields (number, heading, definition, unit, citation relationship, etc.), a semantic tree is built to ensure that chapter logic and contextual relationships are not lost. The expression for the constructed tree structure is: S is composed of chapter units. Clause Unit and paragraph units constitute, This represents the hierarchical relationship between chapters, i.e., chapters contain clauses, and clauses contain paragraphs. The model automatically builds this parent-child hierarchical tree. Here, S represents the semantic structure set of the entire standard document, containing all structural units such as chapters, clauses, and paragraphs. It is one of the outputs of the semantic parsing model. Chapter unit ( ), representing the i-th chapter in the document, such as "Chapter 1 General Provisions," which is the highest-level semantic node; clause unit ( ), indicating that it belongs to a chapter. The j-th item below, such as "1.1 Scope" or "1.2 Definition", corresponds to an intermediate layer node in the tree structure; paragraph unit ( (), indicating a specific paragraph or sub-item under a clause, such as "1.2.1 This standard applies to...". This is the finest-grained structural node, such as... Figure 6 As shown, the left side is a tree structure of chapters, clauses, and paragraphs, while the right side is a structure mapping matrix used to quantify semantic dependencies and illustrate the strength of semantic associations between nodes.
[0044] Each identified structural unit (such as chapter, clause, or paragraph) is treated as a node. Each node contains not only plain text content but also a visual style tag and spatial location fingerprint determined through fusion analysis. The hierarchical relationship (parent-child relationship) between nodes is determined collaboratively through: analyzing the contextual coherence of the text content using LLM's self-attention mechanism; judging physical containment relationships by comparing the spatial bounding boxes of nodes (e.g., a chapter title block visually "contains" all clause blocks below it); and combining this with a regular pattern recognition numbering system (e.g., "1.", "1.1", "a") and cross-validating with the visual hierarchy.
[0045] Based on all the above information, a semantic hierarchy tree is generated. Each node in this tree is a unified entity of visual and semantic information. Furthermore, based on the multiple relationships between nodes (semantic dependency, visual inclusion, and reference association), the system calculates a richer structural mapping matrix M_n×η. This matrix not only includes semantic dependency probabilities but can also be extended to include dimensions such as "visual layout proximity" and "numbering continuity strength," providing a multi-dimensional and computable mathematical model for the document's logical structure.
[0046] The model identifies clause dependencies through a self-attention mechanism, generating a structure mapping matrix M_n×η to quantify the dependencies between nodes in the document structure. The matrix rows (n) represent the number of clauses (or nodes) in the document; the matrix columns (η) represent the semantic attributes or reference relationships of each node (such as parent node index, reference index, etc.); the matrix element M(i,j) represents the semantic connection strength or dependency probability between node i and node j.
[0047] like Figure 6 As shown, firstly, based on LLM (Large Language Model), the document content is semantically segmented into chapters to identify... , , The nodes and their hierarchical relationships form a preliminary structure tree S. LLM calculates the dependency weights between chapters and clauses using a self-attention mechanism, outputting a dependency matrix M_n×η, where higher weights represent stronger semantic connections between nodes. Using matrix M_n×η, the system can achieve: Automatically determine the parent chapter of each clause; Identify cross-references between clauses; This provides a semantic mapping foundation for the node-edge structure in subsequent knowledge graph construction.
[0048] This model can be compared to the table of contents of a book: In this way, the system can fully express the logical hierarchy and semantic dependencies of a document in matrix form, ensuring that the parsing results can be understood and processed by a computer.
[0049] S103. The tables, formulas and images in the standard file are vectorized and reconstructed to generate structured HTML table objects, vector formulas and vector graphics respectively.
[0050] Specifically, such as Figure 3 As shown in the diagram, this illustration demonstrates the process of constructing an HTML table structure using DOM mapping functions through column set C, row set R, and merge cell information (merge), and shows the transformation relationship from the original table to a structured table. It analyzes the table's row and column structure, and reconstructs the table's hierarchical structure through row and column sets (C, R) and merge cell tags (merge). This is used in HTML. , , The `` tag ensures that the table hierarchy is consistent with the merged cells. Using the DOM mapping function `f_DOM`, the table's column set `C`, row set `R`, and merged cell information are taken as input to generate a structured HTML table object `T_html`. The functional relationship is represented as `T_html = f_DOM(C, R, merge)`, where `T_html` refers to the structured table result output by the system, i.e., the final generated HTML.
[0055] The `` element preserves the original table's hierarchical structure, merged cells, and formatting information. It is the output of the entire table reconstruction module; `f_DOM(·)` represents the transformation function that maps the original table data to a DOM hierarchical structure. It generates the HTML structure by parsing the table's rows, columns, and merged information; in the system, it is responsible for converting the identified table geometry into semantic DOM tree nodes; `C` represents the collection of all columns in the table, used to describe the table's horizontal structure, such as column width, column position, column headers, etc. In HTML, this corresponds to... and The `` element; `R` represents all rows of data in a table, used to describe the table's vertical organization, such as the number of rows, row height, and row order. In HTML, this corresponds to... and The `merge` element indicates the merging information for cells in a table. The system identifies which cells in the table are merged horizontally or vertically, and uses this parameter to add the `rowspan` and `colspan` attributes when generating the HTML, thus maintaining the original layout logic of the table. The specific process is as follows: Row and column parsing phase: Parse the table data after OCR recognition to obtain the coordinates (x, y), size (width, height), borders and content of each cell.
[0051] Row set R: Stores the cell information and its sequence number for each row.
[0052] Column set C: Stores the width, header, and position index of each column.
[0053] Merged cell detection phase: The system detects the overlap of borders between adjacent cells. If multiple cells are found to share the same boundary, the system automatically generates a merge mapping table to record the rowspan and colspan of each cell.
[0054] DOM mapping construction phase: Generate DOM tree nodes based on C, R, and merge: Each node corresponds to a [link / structure] in HTML. element; Each row corresponds to one ; Write the attributes of each column If merged cells are detected, the merge attribute is automatically added to that node. Output stage: Generate the final HTML table structure: title Data 1 Data 2 The output maintains semantic and visual consistency with the table structure in the original standard file. For example... Figure 7 As shown, this further illustrates the roles and relationships of column sets, row sets, and merge mappings in generating HTML table structures.
[0056] For tables with multi-level headers, the system merges cell data layer by layer and reconstructs the order of each header level.
[0057] like Figure 4As shown, the process includes formula region detection, character recognition, syntax structure construction, format normalization, and LaTeX / MathType encoding generation and vectorization output. Formula regions are identified using image processing technology and converted into SVG format vector graphics. Each formula uses an algorithm to identify its symbols, operators, variables, and other elements. The identified formula regions are then processed by the `render` function to generate an SVG format vector formula graphic, F_SVG, whose functional relationship is represented as F_SVG = render(L_LaTeX). The layout of the formula is automatically adjusted, and its functional relationship conforms to the LaTeX or MathType encoding standard. It supports converting formulas from MathType to LaTeX format, or vice versa, ensuring compatibility across different application scenarios. Here, F usually represents "file" or "function." More likely, it represents the final output result, i.e., a graphic file or graphic object. SVG is an abbreviation for Scalable Vector Graphics. It is a vector-based XML image format. Unlike bitmap formats such as JPEG or PNG, SVG images do not lose quality when enlarged or reduced, and edges remain sharp. This makes it ideal for displaying mathematical formulas, icons, and charts in web pages and documents. Overall meaning: F_SVG represents the final output of a formula as a clear, infinitely scalable vector graphic in SVG format. `render`: This is a core term in computer graphics and web development, meaning "rendering" or "drawing." It refers to the process of transforming an abstract model or instruction described in code into an image that can be perceived by human vision. In the context of formulas, `render` is a function or operation. Its role is to receive a specific input (LaTeX code), perform calculations and processing (such as parsing syntax, calculating character positions, drawing lines and curves), and finally generate a visual image (SVG graphics). `L`: Represents "language" or "LaTeX source code." LaTeX: This is a powerful typesetting system based on TeX, particularly adept at handling complex mathematical formulas, academic papers, and scientific literature. Users describe the structure of formulas by writing specific markup commands (such as `\frac{a}{b}` to represent the fraction a / b). Overall meaning: L_LaTeX represents input, i.e., plain text code written in the LaTeX language describing a specific mathematical formula or symbol. For example, an SVG format vector graphics file (F_SVG) is obtained by converting the LaTeX source code (L_LaTeX) describing the graphic through a rendering process.
[0058] For example, such as Figure 8As shown, the process by which the rendering function parses the formula encoding and generates the vector graphics F_SVG is clearly illustrated. When you see a beautifully crafted mathematical formula on a webpage (such as Wikipedia, Stack Overflow, or some LaTeX-enabled note-taking software), this process is very likely taking place behind the scenes. Input (L_LaTeX): You or your editor have written LaTeX code, such as E = mc^2.
[0059] Rendering: The rendering engine integrated into the webpage or software (such as MathJax or KaTeX) begins its work. It recognizes the commands (E, =, m, c^2) in the code and calculates the size, position, and spacing of each character according to mathematical typesetting rules.
[0060] Output (F_SVG): The rendering engine ultimately does not directly generate a PNG image, but instead generates a piece of SVG vector graphics code and sends it to your browser. Your browser then renders this code into a sharp and clear formula on the screen: E=mc 2 .
[0061] Image objects (such as drawings and charts) are generated in SVG format using image vectorization technology, preserving image details and supporting lossless scaling. This ensures that images can be displayed clearly at different resolutions.
[0062] Symbols (such as units of measurement and engineering symbols) are identified using pattern matching algorithms. The system automatically generates a symbol library based on the symbol type and context, and indexes it in the document. Footnote areas are automatically identified and reformatted, while ensuring that footnote markers are consistent with citations in the main text.
[0063] S104. Perform semantic consistency verification on the structured data obtained after semantic hierarchical parsing and vector reconstruction based on the large language model.
[0064] Specifically, LLM (Limited Language Modeling) is used for semantic proofreading to ensure consistency in hierarchical structure, citation relationships, and terminology. Semantic conflicts and logical errors are automatically identified through context analysis and text matching. Through self-attention, the LLM model can detect logical consistency between sections of the document and ensure accurate connections between terminology usage and definitions. A combination of human and AI proofreading is supported, providing a human review interface after initial automated proofreading. Cross-checking and expert review processes improve the accuracy of the final proofreading results.
[0065] S105. Output the proofread structured data as a standard format XML file, and automatically construct multiple types of knowledge graphs based on the XML file.
[0066] Specifically, following the "General Processing Indexing System Schema (2021).xsd" standard, the parsed data is transformed into a compliant XML format. This format includes information such as the standard name, number, definition, clauses, and chapter structure, generating XML data compliant with XSD. Five types of knowledge graphs are also automatically generated, such as... Figure 5 The diagram illustrates the entire process of entity extraction, relation extraction, graph mapping, and graph database import from structured data (XML / JSON / HTML): Standards Relationship Graph: Displays the references and dependencies between standards; Indicator Chart: Compiles all technical indicators involved in the standard and their interrelationships; Organizational personnel map: showing the drafting units of the standard, the participating personnel, and their roles; Equipment Product Catalog: Lists the equipment and products mentioned in the standard, along with their specifications and model information; Domain knowledge graph: Extracting knowledge nodes and their relationships related to specific domains (such as environmental protection and energy) from standards.
[0067] By employing a four-step technical approach of "document structure recognition + semantic hierarchical parsing + vector reconstruction + large model semantic verification", we can achieve high-precision structural and semantic processing of complex standard documents (including charts, formulas, symbols, and nested clauses).
[0068] This XML file is not a flat list of data, but a deeply nested document object model containing rich semantic and structural relationships. Its key elements include: document metadata, a structured representation of a semantic hierarchy tree, associative encoding of a structure mapping matrix, and embedded storage of vectorized reconstructed content: non-text elements such as tables and formulas are not stored as images, but are embedded in XML as structured data objects.
[0069] During initialization, a set of graph construction patterns are loaded. Each pattern is a template that is explicitly defined: Target graph type: such as "standard relationship graph" or "index graph".
[0070] Node extraction rules: Specify which paths in the XML, based on what conditions, and what type of entities to extract as graph nodes.
[0071] Relationship extraction rules: Specify how to establish what type of semantic edges between different nodes based on structural relationships, reference relationships, or content matching in XML.
[0072] Attribute mapping rules: Define how to map the attributes of entities in XML (such as name, number, value, unit) to the attributes of graph nodes.
[0073] The construction of a standard relational graph includes: Node instantiation: Based on the schema, each tag representing an independent semantic unit in the XML is instantiated as a "standard clause" node in the graph. Node attributes are derived from the unit's ID and title text.
[0074] Relation instantiation: Hierarchical relationships: Based directly on the nested structure of XML, establish "inclusion" relationships between parent and child clause nodes.
[0075] Reference relationships: Parsing relational encoding in XML. When clause A references clause B, a "reference" or "based on" relationship edge is established between the corresponding two "standard clause" nodes. This process automatically extracts the internal logical network of the document.
[0076] By using the dependency strength quantified by the structure mapping matrix M_n×η generated in S102, weight attributes are assigned to the "reference" edges to form a weighted reference network, which can support more refined knowledge tracing and impact analysis.
[0077] The construction of the indicator graph includes: Node instantiation: The construction of this graph relies heavily on the vectorized, reconstructed tabular data. The engine mode will locate all nodes in the XML.
[0078]
[0079]
[0080] <clause>(The tag contains) establishes a "defined in" relationship between the "Technical Specifications" node and the parent "Standard Clauses" node.
[0081] Relationships between indicators: By analyzing the logic of rows and columns in a table, or by using LLM to perform micro-analysis of the indicator description text, relationships such as "mutual exclusion", "substitution", and "hierarchy" between indicators can be established.
[0082] The construction of the organizational personnel map includes: Node instantiation: The engine pattern guides named entity recognition (NER) in specific fields of XML (such as "drafting unit", "participating unit", "main drafter"), and the identified organization name and personnel name are instantiated as "organization" node and "personnel" node respectively.
[0083] Relation instantiation: Establish a "subordinate" relationship between the "personnel" node and the "organization" node.
[0084] Establish a "Draft" or "Participate" relationship between the "Organization" / "Personnel" node and the "Standard Documents" entity in the root node.
[0085] This process can connect to authoritative external databases to normalize and enrich the identified organization names, and automatically supplement external knowledge attributes such as organization type and location.
[0086] The construction of the equipment product map includes: Node instantiation: This graph construction is a combination of text analysis and structured data mining.
[0087] The engine pattern first identifies entities such as equipment, products, and materials from the text terms of XML through NER and keyword matching, and instantiates them as "equipment / product" nodes.
[0088] Key enhancement: At the same time, the engine will perform deep parsing of the vectorized table, directly converting the data in columns such as "instrument name", "model" and "technical specifications" into structured "device" nodes and their attributes.
[0089] Relation instantiation: Establish an "Application Constraint" relationship between the "Equipment / Product" node and the "Standard Clause" node that specifies its usage requirements.
[0090] Establish a "parameterized" relationship between the equipment and its technical parameters (from the table).
[0091] Based on text descriptions (such as "the system consists of A, B, and C"), establish the "composition" relationship between devices.
[0092] The construction of a domain knowledge graph includes: Node and relation instantiation: Invokes a domain ontology mapper. This mapper has a built-in or dynamically linked top-level ontology framework for a specific domain (such as environmental protection or machinery).
[0093] It guides LLM to perform deep semantic analysis on key clauses in XML (especially the "Terms and Definitions" section and the core requirements section) and identify instances that conform to domain ontology concepts (such as "contaminants", "safety protection devices", and "test methods").
[0094] The instantiated node types are no longer "document units" but "domain concepts" (such as "heavy metal lead" and "compressive strength test"). The relationships between nodes also follow the domain ontology definition (such as "lead" is a "heavy metal"; "a certain test" is used to detect "compressive strength"). The essence of this process is to reorganize and associate the scattered knowledge points in the standard document according to the thinking of domain experts, generating a knowledge network that truly supports domain reasoning.
[0095] All five types of graphs undergo cross-graph association and fusion after generation. For example, a "Gas Chromatograph" node in the "Equipment Graph" might be linked to the "Pesticide Residue Limits" node it detects in the "Indicator Graph" through the "For Detection" relationship. Finally, this set of interconnected multidimensional knowledge graphs is serialized into standard knowledge graph formats (such as RDF, attribute graphs), stored in a graph database, and a query interface is provided.
[0096] Employing an improved LLM language model and a self-developed OCR vector recognition algorithm, this method can automatically identify and reconstruct the hierarchical structure, formula expressions, chart data, and symbol system of standard documents, generating machine-readable HTML / JSON / XML structured data and automatically building a domain knowledge graph. This method significantly improves the efficiency of standard digitization processing (by approximately 90%), saves 70% of manual labor, and can be widely applied to scenarios such as standardization research, policy system construction, AI training corpus generation, and enterprise standard compliance analysis.
[0097] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.
[0098] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.< / clause> Tags, especially those for tables containing technical parameters (such as "limits," "requirements," and "performance metrics"). It instantiates each row (or each key cell) of the table as a "Technical Metric" node. Node attributes are extracted from the table content, such as metric name, specified value, unit, test conditions, etc. This information has been structured during the table vectorization process in S103. Relationship instantiation: Metric-clause attribution relationship: by analyzing the context of the table in the XML (i.e., which clause it belongs to).
Claims
1. A standard digital parsing method based on semantic hierarchical structure and vector reconstruction, characterized in that, Includes the following steps: Standard documents are acquired through high-precision scanning or electronic import, and then layout correction and noise removal are performed. An improved large language model combined with a regular pattern recognition algorithm is used to perform multi-level semantic hierarchical parsing on the standard file, extract key fields and generate a semantic hierarchy tree; The tables, formulas, and images in the standard file are vectorized and reconstructed to generate structured HTML table objects, vector formulas, and vector graphics, respectively. Based on the large language model, the structured data obtained after semantic hierarchical parsing and vector reconstruction is subjected to semantic consistency verification; The proofread structured data is output as a standard format XML file, and multiple knowledge graphs are automatically constructed based on the XML file.
2. The standard digital parsing method based on semantic layering and vector reconstruction as described in claim 1, characterized in that, After obtaining the standard file, the method further includes: Automatically detect and identify the text area, image area, formula area, and symbol area in the standard file.
3. The standard digital parsing method based on semantic layering and vector reconstruction as described in claim 1, characterized in that, An improved large language model combined with a regular pattern recognition algorithm is used to perform multi-level semantic hierarchical parsing on the standard document, extract key fields, and generate a semantic hierarchy tree, including: Based on the self-attention mechanism of the improved large language model, the semantic dependencies between nodes in the standard file are identified, key fields are extracted, and a semantic hierarchy tree is generated. The structure set S of the semantic hierarchy tree consists of chapter units. Clause Unit and paragraph units The structure, and its hierarchical relationship are represented as follows: .
4. The standard digital parsing method based on semantic layering and vector reconstruction as described in claim 1, characterized in that, The specific steps for table vectorization reconstruction are as follows: The DOM mapping function f_DOM takes the column set C, row set R, and merge cell information of the table as input and generates a structured HTML table object T_html. The function relationship is represented as T_html = f_DOM(C, R, merge).
5. The standard digital parsing method based on semantic hierarchical and vector reconstruction as described in claim 1, characterized in that, The specific steps for formula vectorization and reconstruction are as follows: The identified formula region is processed by the render function to generate a vector formula graphic F_SVG in SVG format. Its functional relationship is represented as F_SVG = render(L_LaTeX), where L_LaTeX is the LaTeX encoding of the formula.
6. The standard digital parsing method based on semantic hierarchical and vector reconstruction as described in claim 5, characterized in that, The vectorized reconstruction of the formula also supports bidirectional conversion between MathType and LaTeX encoding formats.
7. The standard digital parsing method based on semantic layering and vector reconstruction as described in claim 2, characterized in that, The method further includes vector reconstruction of the symbols, specifically: Symbols are identified using a pattern matching algorithm, and a symbol library is automatically generated based on the symbol type and context, and then indexed in the document.
8. The standard digital parsing method based on semantic hierarchical and vector reconstruction as described in claim 1, characterized in that, The XML file includes standard names, numbers, definitions, clauses, and chapter structures.
9. The standard digital parsing method based on semantic hierarchical and vector reconstruction as described in claim 1, characterized in that, The automatically constructed knowledge graphs include standard relationship graphs, indicator graphs, organizational and personnel graphs, equipment and product graphs, and domain knowledge graphs.