An education scheme intelligent generation method and system based on OCR table semantic analysis
By using OCR table semantic parsing technology, the logical parsing problem of complex educational program tables is solved, realizing the conversion from unstructured documents to structured data, ensuring the accuracy and consistency of the generated text, and improving the automated processing capability of educational programs.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHENGFANG SOFTWARE CO LTD
- Filing Date
- 2026-05-09
- Publication Date
- 2026-06-05
Smart Images
Figure CN122154664A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of computer information processing technology, and in particular to an intelligent generation method and system for educational programs based on OCR table semantic parsing. Background Technology
[0002] With the deepening of educational informatization and the increasing demand for "smart academic affairs" management, the digital, structured, and intelligent management of core educational documents such as university curriculum plans and course outlines has become an inevitable trend. These documents are usually in PDF or scanned image format, and their core content, such as curriculum structure and credit requirements, is largely organized in tables to achieve a structured presentation of information. Currently, existing general-purpose optical character recognition (OCR) technology can extract text information from document images, and some advanced tools also claim to have table recognition capabilities, providing a certain foundation for automated processing.
[0003] However, in practical applications, when faced with educational program tables with complex logical structures, conventional OCR table recognition often focuses on restoring the visual borders of cells and extracting their internal text. The output is only the physical arrangement of text in the table, rather than semantic association. It cannot understand the inherent logical correspondence between the header cells and the data cells below, and it is even more difficult to handle complex structures such as merged cells and multi-level headers. As a result, the extracted information is a bunch of discrete text fragments without semantic annotation, losing the structured semantics originally carried by the table.
[0004] Furthermore, even after obtaining the text content, automatically converting this unstructured text information into standardized educational program descriptions presents another key challenge. Directly utilizing existing natural language processing technologies or simple template filling makes it difficult to guarantee the accuracy and logical consistency of the generated text, particularly ensuring that key values in the generated text strictly match the original table data, thus hindering the improvement of the automation level of academic affairs management. Summary of the Invention
[0005] To address the aforementioned technical issues, this application provides a method and system for intelligently generating educational solutions based on OCR table semantic parsing.
[0006] Firstly, this application provides an intelligent generation method for educational programs based on OCR table semantic parsing, employing the following technical solution:
[0007] A method for intelligently generating educational programs based on OCR table semantic parsing, the method comprising:
[0008] The system obtains the original document of the educational program uploaded by the user and performs image preprocessing to generate standardized image data; wherein the original document includes PDF electronic document or scanned image format;
[0009] Optical character recognition processing is performed on the standardized image data to extract text blocks and line elements from the document, calculate the page coordinate information of each text block and generate a discrete text list, and simultaneously identify the boundary coordinates of table areas in the document.
[0010] Based on the page coordinate information and table area boundary coordinates in the discrete text list, a table physical structure reconstruction operation is performed to generate a table physical matrix.
[0011] Using a pre-built education scheme business rule base and keyword base, a semantic mapping operation is performed on the physical matrix of the table to anchor the table header cells and associate data entities to establish key-value pair relationships and generate a structured intermediate representation model.
[0012] According to the logical segmentation rules in the education solution business rule base, the structured intermediate representation model is dynamically segmented into multiple semantic segments, and each semantic segment is configured with a prompt word template containing role definition, task instructions and structured data to generate segmented structured data.
[0013] The segmented structured data is input into the large language model to perform inference and generation operations, and numerical consistency checks are performed on the generated text segments. The text segments that pass the checks are then output.
[0014] The validated text segments are integrated, concatenated according to the logical order defined in the education solution business rule base, and converted into a standard document format to generate the final education solution text.
[0015] By employing the aforementioned technical solution, highly robust image preprocessing and OCR technology fused with physical coordinates are used to accurately capture the original information of the document. Then, through virtual grid reconstruction and semantic mapping based on a deep business rule base, accurate parsing of complex table logic structures and business meanings is achieved, transforming visual documents into semantically rich structured data. Next, by combining dynamic segmentation based on business logic and customized prompt word engineering, the powerful generation capabilities of the large model are effectively guided and constrained to specific tasks. Finally, through rigorous numerical consistency verification and standardized integration, the absolute accuracy of the final output text in terms of data and its high degree of professional formatting are ensured. This technical solution improves the automation level, accuracy, and consistency of educational solution processing, significantly reduces manual processing costs, and possesses good business adaptability and scalability.
[0016] Secondly, this application provides an intelligent educational solution generation system based on OCR table semantic parsing, which adopts the following technical solution:
[0017] An intelligent educational solution generation system based on OCR table semantic parsing, the intelligent educational solution generation system comprising:
[0018] The document processing module is used to acquire the original document of the educational program uploaded by the user and perform image preprocessing operations to generate standardized image data; wherein, the original document includes PDF electronic document or scanned image format;
[0019] The OCR recognition and layout analysis module is used to perform optical character recognition processing on the standardized image data, extract text blocks and line elements in the document, calculate the page coordinate information of each text block and generate a discrete text list, and at the same time recognize the boundary coordinates of table areas in the document.
[0020] The table physical structure reconstruction module is used to perform table physical structure reconstruction operations based on the page coordinate information and table area boundary coordinates in the discrete text list, and generate a table physical matrix.
[0021] The structured modeling module is used to perform semantic mapping operations on the physical matrix of the table using a pre-built education scheme business rule library and keyword library, anchor the table header cells and associate data entities to establish key-value pair relationships, and generate a structured intermediate representation model.
[0022] The dynamic segmentation and prompt word assembly module is used to perform dynamic segmentation processing on the structured intermediate representation model according to the logical segmentation rules in the education solution business rule library, splitting it into multiple semantic segments, and configuring prompt word templates containing role definitions, task instructions and structured data for each semantic segment to generate segmented structured data.
[0023] The model inference and verification module is used to input the segmented structured data into the large language model to perform inference and generation operations, and to perform numerical consistency verification on the generated text segments and output the text segments that pass the verification.
[0024] The document integration and output module is used to integrate the verified text paragraphs, concatenate them according to the logical order defined in the education solution business rule base, convert them into a standard document format, and generate the final education solution text.
[0025] Thirdly, this application provides a computer-readable storage medium, which adopts the following technical solution:
[0026] A computer-readable storage medium storing a computer program that can be loaded by a processor and executed as in any of the methods in the first aspect.
[0027] In summary, this application includes at least one of the following beneficial technical effects: by using OCR and table structure parsing to transform unstructured educational program documents into machine-understandable structured data, and then driving a large language model to generate standardized text, it solves the key bottlenecks of low efficiency in traditional manual compilation, the difficulty of general large models to understand table logic, and the lack of business semantic parsing capabilities in existing OCR technology. This achieves a fundamental shift in educational program compilation from reliance on human experience to automation and intelligence. In practical applications, this technical solution improves the accuracy of understanding and parsing complex course credit tables, ensures the logical consistency and numerical correctness of key information such as courses and credits in the generated program, significantly reduces the cost of manual verification and revision, and enhances the system's ability to generalize and process educational program documents from different formats and universities. Attached Figure Description
[0028] Figure 1 This is a schematic diagram of the first process of an intelligent generation method for educational programs based on OCR table semantic parsing, which is one embodiment of this application.
[0029] Figure 2 This is a schematic diagram of the second process of an intelligent generation method for educational programs based on OCR table semantic parsing, which is one embodiment of this application.
[0030] Figure 3 This is a schematic diagram of the third process of an intelligent generation method for educational programs based on OCR table semantic parsing, which is one embodiment of this application.
[0031] Figure 4 This is a schematic diagram of the fourth process of an intelligent generation method for educational programs based on OCR table semantic parsing, which is one embodiment of this application.
[0032] Figure 5 This is a schematic diagram of the fifth process of an intelligent generation method for educational programs based on OCR table semantic parsing, which is one embodiment of this application.
[0033] Figure 6 This is a schematic diagram of the sixth process of an intelligent generation method for educational programs based on OCR table semantic parsing, which is one embodiment of this application.
[0034] Figure 7 This is a schematic diagram of the seventh process of an intelligent generation method for educational programs based on OCR table semantic parsing, according to one embodiment of this application.
[0035] Figure 8 This is a schematic diagram of the eighth process of an intelligent generation method for educational programs based on OCR table semantic parsing, which is one embodiment of this application. Detailed Implementation
[0036] To make the purpose, technical solution, and advantages of this application clearer, the following description is provided in conjunction with the appendix. Figures 1-8 The present application will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the application.
[0037] This application discloses an intelligent method for generating educational programs based on OCR table semantic parsing.
[0038] Reference Figure 1 A method for intelligently generating educational plans based on OCR table semantic parsing, specifically including:
[0039] Step S101: Obtain the original document of the educational program uploaded by the user and perform image preprocessing to generate standardized image data;
[0040] The original documents include PDF electronic documents or scanned images. During the digitization process, problems such as noise, geometric distortion, and uneven lighting are often introduced due to scanning equipment, paper conditions, or the quality of the original documents. These problems can seriously interfere with the accuracy of subsequent optical character recognition.
[0041] In the embodiments of this application, image preprocessing operations include performing Gaussian filtering for noise reduction, adaptive thresholding for binarization, and Hough transform for image tilt correction on scanned image format documents to generate geometrically corrected standardized image data.
[0042] Specifically, Gaussian filtering denoising smooths the image through convolution operations, suppressing random noise and salt-and-pepper noise while preserving edge information. Adaptive threshold binarization dynamically calculates the binarization threshold based on the grayscale features of local image regions, effectively overcoming the problem of unclear text-background segmentation caused by uneven lighting or background contamination. Hough transform image tilt correction rotates the document by detecting the tilt angle of straight lines in the image, ensuring that text lines are level with image boundaries, which is crucial for coordinate-based text block aggregation and table structure analysis. Through these steps, a standardized image with geometrically correct alignment, a clean background, and clear text is obtained, laying the physical foundation for accurate recognition.
[0043] Step S102: Perform optical character recognition processing on the standardized image data, extract text blocks and line elements in the document, calculate the page coordinate information of each text block and generate a discrete text list, and at the same time identify the boundary coordinates of the table area in the document.
[0044] Traditional OCR focuses solely on character recognition, while this solution emphasizes the simultaneous parsing of the document's physical layout. Through optical character recognition (OCR), not only are individual characters identified, but more importantly, adjacent characters are aggregated into "text blocks," and the bounding box "page coordinate information" of each text block is precisely calculated. This page coordinate information (such as the coordinates of the top-left and bottom-right corners) together with the text content constitutes a "discrete text list," which records what all the text in the document "is" and "is where."
[0045] Simultaneously, the system identifies horizontal and vertical lines in the image in parallel, or locates "table regions" and records their "boundary coordinates" by detecting large areas of blank space and text alignment features. This step essentially deconstructs the image document into two types of machine-readable sets of elements with precise spatial location markings: unstructured text units and table region outlines with defined boundaries, providing accurate input data for the next step of reconstructing the logical relationships of the table from a physical perspective.
[0046] Step S103: Based on the page coordinate information and table area boundary coordinates in the discrete text list, perform a table physical structure reconstruction operation to generate a table physical matrix;
[0047] This step is a crucial link connecting visual information with logical structure, aiming to solve the problem that OCR output is only discrete text and loses the original table's row and column topological relationships.
[0048] Specifically, the system infers implicit table grid lines by analyzing the relative positional relationships between text blocks and calculating the vertical and horizontal alignment features of line intersections or text centers. For example, by clustering text block center points with similar vertical coordinates, row lines can be inferred; by clustering horizontal coordinates, column lines can be inferred.
[0049] For minor text misalignment caused by poor scanning quality, a positional tolerance mechanism based on OCR confidence is introduced. This means that for text blocks with low recognition confidence, a larger positional tolerance space is assigned when determining their grid affiliation, thus constructing a virtual grid with tolerance attributes. Text blocks are assigned to specific cells within this virtual grid. Merged cells spanning multiple rows or columns are identified and marked with "cross-row / column attributes" by analyzing the text block span against the blank spaces of adjacent cells.
[0050] Ultimately, the generated table physical matrix is a two-dimensional logical structure, in which each matrix element (i.e., cell) contains the original text content, row and column indexes, and merge status, fully restoring the physical skeleton of the table, but not yet giving it business meaning.
[0051] Step S104: Using the pre-built education solution business rule base and keyword base, perform semantic mapping operation on the physical matrix of the table, anchor the table header cells and associate data entities to establish key-value pair relationships, and generate a structured intermediate representation model.
[0052] The logic behind this step relies on a predefined, domain-knowledge-rich library of educational program business rules and keywords.
[0053] Specifically, the system first performs recursive matching in the header rows (or columns) of the matrix using a "course category keyword tree". For example, a three-level classification system (such as "General Education > Required > Humanities and Social Sciences") can more accurately anchor complex composite headers, avoiding ambiguity compared to simple keyword matching. Once semantic header cells such as "course category", "credit", and "course name" are anchored, the system establishes a mapping relationship with the data area cells below or to the right based on their row and column indices in the physical matrix.
[0054] For example, if the header "Course Name" is identified in the second row and first column, then the content of the cells in the third row and below, in the first column, is mapped to the "Course Name" entity. In this way, the discrete text in the table is systematically organized into a series of "key-value pair relationships," such as {"Category":"Core Major", "Course Name":"Data Structure", "Credits":"3.0"}. This collection of all relationships constitutes a structured intermediate representation model, a semantic network that is completely detached from the original visual layout and purely expresses business entities and their relationships—an ideal data form for subsequent intelligent processing.
[0055] Step S105: According to the logical segmentation rules in the education solution business rule base, perform dynamic segmentation processing on the structured intermediate representation model, split it into multiple semantic segments, and configure prompt word templates containing role definitions, task instructions and structured data for each semantic segment to generate segmented structured data.
[0056] Directly inputting the complete structured intermediate representation model into a large language model may exceed its context window or lead to chaotic generation due to its overly complex internal logic. Therefore, the logical principle of this step is to intelligently and adaptively segment the model based on the inherent business logic of the educational program.
[0057] Specifically, dynamic segmentation is based on logical segmentation rules in the education solution business rule base. The system monitors changes in the course category field. For example, when a course attribute changes from "General Education Required" to "Professional Core," this marks a natural knowledge module boundary, and the system automatically inserts a segmentation identifier at this point. Simultaneously, a credit accumulation threshold can be set. When the accumulated credits within the same category reach the module requirement (e.g., "Professional elective courses require 10 credits"), segmentation is also triggered. After segmentation, each semantic segment contains a logically consistent subset of data.
[0058] Next, a prompt word template is configured for each paragraph. The template clearly defines the "role" (e.g., "You are a curriculum design expert"), the "task instructions" (e.g., "Describe the following course list as professional training requirements"), and the corresponding "structured data" for that paragraph. This process transforms the massive amount of structured data into a series of targeted, contextually clear, and segmented structured data task packages that conform to the processing habits of large language models, greatly improving the feasibility and quality of the generated tasks.
[0059] Step S106: Input the segmented structured data into the large language model to perform inference generation operation, and perform numerical consistency verification on the generated text segments, and output the text segments that pass the verification.
[0060] The logical principle of this step is divided into two closely coupled stages: generation and verification.
[0061] First, segmented structured data (including roles, instructions, and data) is input into a large language model to perform inference and generation operations. Based on its vast knowledge and understanding of instructions, the model transforms structured course and credit information into fluent and standardized natural language descriptions. However, the large model may suffer from "illusions," leading to discrepancies between key values (such as credits and course numbers) in the generated text and the input data. Therefore, a numerical consistency verification mechanism is crucial.
[0062] Specifically, the system extracts key data entities such as credit values from the generated text segments using techniques like regular expression matching. It then compares these extracted data with the corresponding data in the original structured intermediate representation model. If an inconsistency is detected, the system doesn't simply accept the erroneous output; instead, it triggers a regeneration mechanism. This involves adjusting prompts, resampling, or switching decoding strategies to require the large model to regenerate until the values perfectly match. This creates a closed loop of generation, verification, and correction, ensuring the reliability of the automated output.
[0063] Step S107: Integrate the verified text paragraphs, concatenate them according to the logical order defined in the education solution business rule library, and convert them into a standard document format to generate the final education solution text.
[0064] This step aims to combine multiple independently generated and validated text modules into a professional, complete, and uniformly formatted final document. Its logical principle is to perform orderly synthesis based on domain specifications.
[0065] Specifically, the system concatenates all validated text paragraphs according to the logical order defined in the "educational solution business rule base" (e.g., "training objectives," "graduation requirements," "curriculum system," "practical components," etc.). This involves not only simple text connection but may also include generating transition sentences between paragraphs and adjusting formatting. Finally, the concatenated complete text is converted into a standard document format (such as DOCX, PDF, or a specific XML Schema) to generate a final, ready-to-use educational solution text. This process encapsulates all intelligent processing results into a standardized product that meets the end-user's needs.
[0066] In the above implementation, highly robust image preprocessing and OCR technology fused with physical coordinates are employed to accurately capture the original information of the document. Then, through virtual grid reconstruction and semantic mapping based on a deep business rule base, accurate parsing of complex table logic structures and business meanings is achieved, transforming the visual document into semantically rich structured data. Next, by combining dynamic segmentation based on business logic and customized prompt word engineering, the powerful generation capabilities of the large model are effectively guided and constrained to specific tasks. Finally, through rigorous numerical consistency verification and standardized integration, the absolute accuracy of the final output text in terms of data and its high degree of professional formatting are ensured. This technical solution improves the automation level, accuracy, and consistency of educational solution processing, significantly reduces manual processing costs, and possesses good business adaptability and scalability.
[0067] Reference Figure 2 As one implementation of step S102, the steps of performing optical character recognition processing on standardized image data, extracting text blocks and line elements from the document, calculating the page coordinate information of each text block and generating a discrete text list, and simultaneously recognizing the boundary coordinates of table areas in the document include:
[0068] Step S201: Perform document layout analysis on the standardized image data, identify text blocks and line elements in the document, and calculate the page coordinate information and corresponding text content of each text block.
[0069] The core logic of this step lies in deconstructing the visual layout of the document and identifying and locating the basic semantic units, namely text blocks.
[0070] Specifically, document layout analysis refers to scanning standardized image data using computer vision algorithms (such as connected component analysis and deep learning segmentation models) to distinguish between text regions and non-text regions (such as graphics and background) in the image. For the identified text regions, they are further segmented into independent text blocks. Each text block usually corresponds to a continuous, visually distinguishable text unit, such as a word, a number, or a short phrase.
[0071] At the same time, the system identifies line elements through edge detection or line extraction algorithms, which are usually represented as table borders in educational program tables and are the key visual basis for distinguishing different cells.
[0072] Next, the page coordinates of each text block are calculated, which can be achieved by obtaining its smallest bounding rectangle (i.e., the smallest rectangle that completely encloses the text pixel area, with sides parallel to the image coordinate axes). Recording the coordinates of the top-left vertex, width, and height of this rectangle accurately depicts the absolute position and spatial extent of the text block on the page. Simultaneously, the OCR engine recognizes the text content within the text block.
[0073] Ultimately, the image document is transformed into a collection of text blocks carrying geometric attributes (coordinates, width and height) and semantic content (text), providing atomic data for subsequent logical reorganization based on spatial relationships.
[0074] Step S202: Based on the page coordinate information of the text blocks, aggregate the text blocks that meet the preset line alignment conditions to generate a text line set, and determine the line position range of each text line in the text line set;
[0075] The text blocks output by direct OCR are discrete and unordered, especially for multi-column layouts or table content, where the contextual row structure is lost. Therefore, the logical principle of this step is to reconstruct the document's horizontal logical structure, i.e., text lines, based on the vertical alignment of the text blocks.
[0076] Specifically, this can be achieved by calculating the vertical centerline distance between adjacent text blocks, where the vertical centerline is the center line of the text block rectangle in the vertical direction. Text blocks belonging to the same physical text line should have very close Y-coordinates (or vertical positions) of their vertical centerlines. The system sets a preset line merging threshold; when the vertical centerline distance between two text blocks is less than the preset threshold, they are considered to meet the preset line alignment conditions and should be merged into the same text line. By traversing and clustering all text blocks, a final set of text lines is generated.
[0077] Next, for each aggregated text line, the system marks the line position range of each text line, typically taking the minimum and maximum Y values of all text blocks in the vertical direction as the upper and lower boundaries of the line. This step reassembles the one-dimensional scattered points (text blocks) into a two-dimensional structure (text lines) with a linear order, restoring the basic reading order and logical grouping of the document in the line direction.
[0078] Step S203: Based on the row position range of the text line set, detect text line clusters with column alignment relationships, and calculate the bounding box coordinates of the text line clusters as the table area boundary coordinates.
[0079] Not all text lines belong to tables. The purpose of this step is to filter out areas from ordinary text paragraphs that have specific spatial distribution characteristics and are suspected to be tables.
[0080] In this embodiment of the application, detecting a cluster of text lines with column alignment includes: analyzing the periodicity features of the horizontal projection histogram of the text block and identifying regions with stable projection peak intervals as candidate areas for the table.
[0081] Specifically, a typical characteristic of tables is the "column alignment" of their content. That is, text describing similar attributes (such as course names or credits) in different rows is aligned horizontally, forming vertical "columns." To detect clusters of text rows with column alignment, the periodicity of the horizontal projection histogram of text blocks can be analyzed. This involves projecting and statistically analyzing the horizontal (X-axis) distribution of text blocks in all suspected areas to generate a histogram. In regular table areas, due to column alignment, this histogram will exhibit distinct, relatively evenly spaced peaks and troughs. Peaks correspond to columns containing text, and troughs correspond to the spaces separating columns. This periodicity is key to distinguishing tables from ordinary paragraphs (whose text projection is usually continuous or irregular).
[0082] Next, after identifying the table candidate region with this stable peak interval, it is necessary to accurately define its range. The bounding box coordinates are the minimum outer bounding matrix that encloses the cluster of text lines. In this embodiment, identifying the table region boundary coordinates specifically includes: calculating the convex hull of the geometric center points of all text blocks within the table candidate region (the convex hull is the smallest convex polygon containing all these points), and using the coordinates of the minimum bounding rectangle of the convex hull as the table region boundary coordinates.
[0083] In practical applications, this step does not rely on recognizing explicit table lines, but rather infers the table structure by analyzing the spatial arrangement patterns of the text itself, making it more robust to tables without lines or with incomplete lines.
[0084] Step S204: Associate and integrate the text content, page coordinate information and table area boundary coordinates of all text blocks to generate the discrete text list.
[0085] The system has grasped the details of "text blocks" at the micro level and the scope of "table areas" at the macro level. Through correlation and integration, the system marks the area to which each text block belongs (whether it is located within a table or a regular text paragraph) and binds its text content with coordinate information. The final generated discrete text list is actually a structured intermediate dataset. It retains the text results recognized by OCR, carries precise spatial location attributes through coordinate information, and clarifies which data belongs to the table objects to be parsed through the table area boundary coordinates.
[0086] The above embodiments can robustly identify and locate table areas in a document without relying on complete table borders, solely by analyzing the spatial distribution patterns of text elements themselves, and efficiently associate text content with its precise page layout and the information of the table area it belongs to. This technical solution improves the ability to recognize tables in documents with no borders, poor printing quality, or complex formats, laying a solid and reliable foundation for the subsequent accurate extraction and reconstruction of structured table data from unstructured document images.
[0087] Reference Figure 3 As one implementation of step S103, the step of performing a table physical structure reconstruction operation based on the page coordinate information and table region boundary coordinates in the discrete text list to generate a table physical matrix includes:
[0088] Step S301: Calculate the geometric features of the grid rows and columns based on the page coordinate information and the table area boundary coordinates, and construct a virtual grid model;
[0089] In this context, a well-organized table, regardless of whether it has visual borders, will exhibit an implicit row and column alignment pattern in its internal content elements (text blocks) within the page coordinate system. The virtual grid construction operation in this step is the process of discovering and explicitly defining this pattern.
[0090] In some embodiments, the system first determines the analysis range based on the table boundaries, and then infers grid lines within this range by calculating the geometric features of grid rows and columns, such as detecting text block alignment features or line intersections in page coordinate information.
[0091] Specifically, by analyzing the distribution of the left X-coordinate and top Y-coordinate of all text block bounding boxes, a clustering algorithm can be used to group left X-coordinates with very similar values together. The center line of each group is then inferred as a vertical virtual grid line (column line). Similarly, the top Y-coordinates are clustered to infer horizontal virtual grid lines (row lines). This method does not rely on recognizing actual drawn lines but constructs the grid based on the alignment features of the text blocks themselves. Another method is to directly detect line intersections. If physical table lines exist in the document, their intersections naturally constitute grid nodes.
[0092] It should be noted that, regardless of the method used, the final virtual grid model is a logical coordinate system composed of a series of horizontal and vertical lines. It divides the table area into several rows and columns. However, this grid is the result of computational abstraction, not a visual entity. It provides a reference system for subsequently placing discrete text blocks into specific logical cells.
[0093] Step S302: Based on the virtual grid model, calculate the geometric center coordinates of each text block and map them to the row and column index positions of the virtual grid model to generate cell position mapping data;
[0094] After the virtual grid coordinate system is established, each independent text block needs to be precisely placed in a specific position in this logical grid, that is, to determine which cell each text block belongs to. The logical principle is based on the mapping of spatial inclusion relationships.
[0095] In this embodiment of the application, calculating the geometric center coordinates of each text block includes: applying a preset fault tolerance threshold to process coordinate offsets, and forcibly aligning text blocks with offsets less than the threshold to the grid lines.
[0096] Specifically, the geometric center coordinates of a text block are the (X,Y) coordinates of the center point of its boundary rectangle, representing the average position of the text block on the page. The system compares these coordinates with the virtual grid model to determine which two adjacent horizontal virtual lines and which two adjacent vertical virtual lines it falls into within the rectangular grid. The row and column numbers corresponding to these grids are the "row and column indices" of the text block.
[0097] However, due to document scanning distortion, OCR recognition errors, or minor adjustments to the original layout, the actual geometric center of a text block may slightly deviate from the ideal grid line. To address this, the system introduces a preset error tolerance threshold. During mapping, the coordinates are not required to fall strictly within the range defined by the grid line; instead, a small deviation range is allowed. When the distance between the center point of a text block and a certain grid line is "less than the threshold," the system will "force-align" it to that grid line, thereby correcting the minor error and ensuring the text block is correctly positioned.
[0098] Finally, this step iterates through all text blocks and assigns a logical position to each block, generating cell position mapping data. This data essentially establishes a mapping table from text blocks to grid coordinates (row i, column j). However, it has not yet dealt with the complex cases where a cell contains multiple text blocks (such as a cell containing multiple lines of text) or a text block spans multiple grid cells (merged cells).
[0099] Step S303: Based on the cell position mapping data, analyze the proportional relationship between the geometric size of the text block and the virtual mesh model, determine the number of mesh cells covered by the text block, mark the number of rows spanned and the number of columns spanned, and generate cell row and column span attribute data.
[0100] In real tables, there are often merged cells that span multiple rows or columns. This is the difficult and key point in table structure analysis. The logical principle of this step is to analyze the proportional relationship between the physical size of the text block and the theoretical size of the basic grid unit to infer its row and column spanning attributes. In this way, those text blocks that physically occupy multiple grid spaces are correctly identified as logically single merged cells, and the complex layout of the table is accurately restored.
[0101] Specifically, the system first calculates the typical width and height of a single basic grid cell based on the virtual grid lines. Then, it analyzes the geometric dimensions of the text block in relation to the grid. If the width of a text block is much larger than (for example, close to 1.8 or more times) the width of the basic grid cell, and its position spans the column where the initial mapping is located and the adjacent column, then it is very likely a "spanning column" cell.
[0102] Similarly, the height ratio can be used to determine "line span," which can be precisely quantified by calculating the number of grid cells covered by the text block. The system can check the intersection of the boundary rectangle of the text block with the virtual grid lines, directly calculating how many columns it spans horizontally and how many lines it spans vertically.
[0103] The system then generates labeled data for the number of rows and columns spanned, i.e., cell row and column span attribute data. For example, a text block located in row 2 and column 3, if it covers rows 2 and 3 and columns 3 and 4, its cell attribute is labeled as {rowspan:2,colspan:2}.
[0104] Step S304: Integrate cell location mapping data and cell cross-row / column attribute data to generate a table physical matrix.
[0105] The location mapping data provides information about "which text is in which grid position," but may contain issues where multiple grid indices point to the same content due to merged cells. The attribute data, on the other hand, indicates which grid positions actually belong to the same merged cell.
[0106] In this embodiment, the integration process needs to resolve potential conflicts and organize information in an efficient and non-redundant manner. For example, a physical matrix of the table can be generated by serializing cell position mapping data and cell cross-row / column attribute data into a two-dimensional array format. This matrix is a two-dimensional array with the same number of rows and columns as the virtual grid model, and each element in the array corresponds to a logical cell. For ordinary cells, the element directly stores its text content. For merged cells, the text content and merge attributes (such as {content:“text”, rowspan:2, colspan:1}) are typically stored only in the “main cell” position defined by its cross-row and cross-column attributes (usually the first grid in the upper left corner), while other grid positions covered by the merged cell are marked as occupied or left empty to avoid data duplication.
[0107] Ultimately, the generated table physical matrix is a data structure that is completely independent of the original page coordinates, accessed purely by row and column indices, and contains all text content and cell merging relationships. It accurately restores the table's topology and provides clear and unambiguous input for subsequent semantic mapping steps.
[0108] In the above embodiments, the inherent logical row and column structure of a table can be robustly reconstructed from a set of discrete text coordinates lacking clear structural information, and merged cells can be accurately identified. This technical solution establishes the implicit geometric framework of the table by constructing a virtual grid, solves the inaccuracy problem in real-world documents through fault-tolerant mapping, and intelligently infers complex cell merging relationships through proportional analysis, ultimately generating a concise and accurate physical matrix of the table. This process is independent of the document's visual style (such as the presence of border lines), exhibiting strong versatility and adaptability.
[0109] Reference Figure 4 As one implementation of step S104, the steps of using a pre-built education program business rule base and keyword base to perform semantic mapping operations on the physical matrix of the table, anchoring the header cells and associating data entities to establish key-value pair relationships, and generating a structured intermediate representation model include:
[0110] Step S401: Based on the pre-set education solution business rule library and keyword library, scan the cell content in the physical matrix of the table, match the predefined education business keywords, identify and mark the table header cells, and generate anchor table header information;
[0111] The table physical matrix is a two-dimensional data structure containing the text content of all cells and their precise row and column coordinates, but its content has not yet been categorized (for example, it is impossible to distinguish whether a cell is a "course name" header or a specific course name). The core logic of this step is to assign "header" labels with clear business meanings to the parsed table physical skeleton, which only contains row and column position information.
[0112] In this embodiment, the first row or first column of the physical matrix of the table is traversed, and course category keywords from the keyword library are used for text matching. Specifically, this keyword library contains "educational business keywords," such as "course name," "credits," "course nature," "semester offered," "general education required courses," and "core professional courses."
[0113] Next, the system scans the text using text matching algorithms (such as exact matching, fuzzy matching, or regular expressions). Once the content of a cell successfully matches a term in the keyword database, that cell is identified and marked as a header cell. This marking not only records the cell's position (row and column index), but more importantly, it records the semantic tag it matched. For example, if the content of the cell located in row 1, column 2 is "credits," it will be anchored as a "credits" header cell.
[0114] It's worth noting that traversing the first row or column of the table's physical matrix is a common and efficient strategy, as the table headers for most tables are located at the top or left. The generated anchor header information is a crucial data structure; it essentially establishes a mapping dictionary from "physical location" (e.g., row 1, column 2) to "semantic role" (e.g., "credit"), creating a semantic "coordinate axis" for the entire table. This process heavily relies on the completeness of the education solution business rule base, which contains dynamically loaded course category rule tables. It receives external update commands through an interface, refreshing the keyword matching logic in real time, thus adapting to the diverse course classification systems and header naming conventions of different universities, enhancing the system's versatility.
[0115] Step S402: Based on the anchored header information and the row and column index data of the table physical matrix, establish the key-value pair mapping relationship between the data cells and the header cells, and generate key-value pair associated data;
[0116] After clarifying the semantic role and physical location of the table header, the specific content in the table data area (non-header cells) is correctly assigned to the corresponding header semantics based on the table's row and column topology, thereby forming meaningful business data units.
[0117] In this embodiment, the logic of this step is strictly based on the two-dimensional matrix characteristics of the table: In a regular table, the business attributes of a data cell (assuming it is located in row i and column j) are jointly determined by the row header corresponding to its row (possibly located in row i and column k) and the column header corresponding to its column (possibly located in row h and column j). By utilizing the precise coordinate information of the row and column index data of the table's physical matrix, combined with the anchor header information generated in the previous step, the data entity association operation can be performed.
[0118] Specifically, the system iterates through each non-header data cell, searching for the nearest anchored header cell in both the row and column directions based on its row and column indices. The semantic label of the found header cell becomes the "key" describing the attributes of that data cell, while the text content of the data cell itself is the "value," thus forming one or more "key-value pairs."
[0119] For example, if a data cell is located at row index 5 (corresponding to the row header "Core Courses") and column index 2 (corresponding to the column header "Credits"), and its content is "3", then a key-value pair similar to {"Course Category":"Core Courses", "Credits":"3"} will be generated.
[0120] Finally, by traversing the entire data area, the system generates a complete set of key-value pairs associated data. This step relies entirely on precise row and column index calculations to establish the association, avoiding the ambiguity that may be caused by fuzzy matching based on text content. This ensures that data under "Course Name" will not be incorrectly associated with the "Credits" header, thus guaranteeing the accuracy of semantic mapping.
[0121] Step S403: Based on the key-value pair association data, the mapping relationship is serialized into a computer-readable structured format to generate a structured intermediate representation model.
[0122] Specifically, integrating key-value pair data is not simply a matter of piling it up; it requires reorganization according to business logic. For example, all information belonging to the same course (such as course name, credits, nature, and semester) should be aggregated into a single logical object. The system typically groups and aggregates key-value pairs based on a core key (such as "course name"), combining multiple attribute key-value pairs describing the same entity into a composite structure.
[0123] Next, a serialization operation is performed, which converts these in-memory data structures into a byte stream or text conforming to a specific format standard, such as JSON object format or XML tree structure. JSON is a lightweight data-interchange format that uses a key-value pair and ordered list structure, which is very suitable for the data generated in this step; XML defines data structures through nested tags. Regardless of the format used, the serialization process produces a standardized, structured intermediate representation model.
[0124] For example, a course might be represented as a JSON object: {"course_name":"Data Structures","credit":"3","type":"Core Courses","semester":"Semester 3"}, while the entire course system can be represented as an array of such objects. This model completely departs from the visual layout and physical coordinates of the original table, and is a pure, semantically rich, structured data collection organized according to business logic.
[0125] In the above implementation, the inherent row and column topology of the table is utilized to establish mappings through coordinate calculations rather than text guessing, fundamentally avoiding semantic association errors. Simultaneously, relying on a configurable and dynamically updateable business rule base allows the system to flexibly adapt to tables with different formats and terminology systems, overcoming the limitations of fixed templates. The resulting structured intermediate representation model is a clear, unambiguous, and standardized semantic data pool, providing a solid, reliable, and high-quality data foundation for subsequent intelligent segmentation and text generation. It is a crucial link in the intelligent chain of educational solutions, connecting "document images" to "understanding data" and then to "generating text."
[0126] Reference Figure 5 As one implementation of step S105, the structured intermediate representation model is dynamically segmented into multiple semantic segments according to the logical segmentation rules in the education solution business rule base. Each semantic segment is then configured with a prompt word template containing role definitions, task instructions, and structured data. The steps to generate segmented structured data include:
[0127] Step S501: Obtain the logical segmentation rules from the education solution business rule base;
[0128] Specifically, the logical segmentation rule refers to the following: based on predefined course module separators (such as "General Education Required Module" and "Professional Core Course Module") in the education solution's business rule base, the rule automatically triggers semantic segmentation when a change in the course category attribute field of the structured intermediate representation model is detected in real time (e.g., switching from "Public Basic Course" to "Professional Core Course"). This generates a segmentation position index, thereby splitting the continuous data stream into independent semantic segments according to business logic. This rule ensures that the segmentation strictly follows the inherent modular structure of the education solution's curriculum system, avoiding logical breaks caused by mechanical segmentation.
[0129] Step S502: Scan the course category attribute field in the structured intermediate representation model, detect the preset segmentation trigger conditions, and generate semantic paragraph segmentation points based on logical segmentation rules;
[0130] The structured intermediate representation model is typically a list or array containing all course entries (each entry includes attributes such as course name, credits, category, and semester). Although the data is structured, it is still a linear sequence. The system scans the course category attribute field, that is, it sequentially reads the key classification attribute values such as "course category" or "module type" for each course entry.
[0131] In this embodiment, the preset segmentation trigger condition is: when the value of the course category attribute field changes, the semantic segmentation operation is triggered. Generating semantic segmentation points includes: traversing the data stream of the structured intermediate representation model, identifying the course module separator identifier defined in the education solution business rule base, and generating a segmentation position index list.
[0132] Specifically, the purpose of segmentation is to group courses belonging to the same knowledge module or with the same teaching attributes together to form an independent semantic unit. The system detects preset segmentation trigger conditions, which are rooted in logical segmentation rules. The most common trigger condition is when the value of the course category attribute field changes. For example, when the system sequentially scans the course list, if the course category of two adjacent entries changes from "General Education Required" to "Professional Core," this change point marks a natural boundary of the business module, and the system generates a semantic segmentation point at this location. The segmentation point can be a positional index, marking the end of the previous semantic segment and the beginning of the next semantic segment.
[0133] Step S503: Based on the semantic paragraph segmentation points, perform data slicing operation on the structured intermediate representation model, extract the key-value pair dataset corresponding to each semantic paragraph, and generate the segmented intermediate data set;
[0134] After obtaining a series of semantic segmentation points, the goal of this step is to perform actual data segmentation operations, splitting a single, large set of structured data into several smaller, logically cohesive subsets according to business logic. The logical principle is similar to dividing a book into several chapters based on bookmarks.
[0135] Specifically, the system segments the original linear data list using these semantic segmentation points as boundaries. For example, assuming the entire model contains 100 courses, and segmentation points are detected at courses 20 and 55 (i.e., the category changes), the system will segment the data into three segments: courses 1-20, courses 21-55, and courses 56-100.
[0136] Next, from each segment, a key-value pair dataset corresponding to each semantic segment is extracted. Each dataset contains complete structured information about all courses within that semantic segment, i.e., a collection of key-value pair objects. For example, the first dataset might contain detailed information about all “general education compulsory” courses, while the second dataset might contain information about all “professional core” courses. These independent subsets of data together constitute the “segmented intermediate dataset”.
[0137] Step S504: Based on the segmented intermediate data set, match the corresponding template in the preset prompt word template library, and inject role definition parameters, task instructions and structured data content into each data unit to generate a configured prompt word instance;
[0138] The matching pre-defined prompt word template library includes: dynamically selecting the corresponding large model inference task template based on the course module type field value of the segmented intermediate data. The injected role definition parameters include: loading predefined teaching expert role description text from the education solution business rule library and replacing the role placeholders in the template.
[0139] Specifically, the logical principle of this step is to assemble a clear and specific task description, i.e., a prompt word, for each subset of data to guide the large model in generating high-quality data. The system first matches each data unit in the segmented intermediate dataset with a corresponding template from a pre-built prompt word template library. This library stores pre-designed prompt word skeletons for different types of educational modules. The matching is typically based on the course module type field value of the data unit. For example, the "core professional courses" module matches a template emphasizing knowledge depth and professional description, while the "practical components" module matches a template emphasizing process, objectives, and outcomes.
[0140] Next, after matching the prompt word template, the system performs an injection operation, which is a parameter filling process. Specifically, this includes: injecting role definition parameters, such as loading descriptive text like "academic affairs expert" or "curriculum design consultant" from the rule base, giving the large model a clear identity perspective; injecting task instruction parameters, clearly specifying what the model should do, such as "Please convert the following course list into a coherent description of general education goals, including course objectives and value statements"; and finally, injecting the structured content of the current data unit itself as structured data content.
[0141] After filling in these three parameters, an abstract template becomes a configured prompt word instance, which is a complete and executable prompt text containing a specific role, a clear task, and data to be processed.
[0142] Step S505: Aggregate all prompt word instances and associated segmented intermediate data to generate segmented structured data.
[0143] In this process, aggregation is not a simple merging, but rather a process that maintains a strict correspondence between each prompt word instance (task instruction) and its source (segmented intermediate data), and may also include metadata such as segmentation order and module type.
[0144] Finally, segmented structured data is generated. This output is a structured collection or list, where each element represents an independent processing task unit. Each unit contains at least two core parts: first, a "cue word instance" specifically configured for the semantic segment and directly input into the large model; and second, a subset of the original "segmented intermediate data" pointed to by the cue word.
[0145] It's important to note that this encapsulation method ensures the indivisibility of data and instructions, providing extremely clear input for subsequent large-scale model inference and generation steps: the process processor can sequentially or in parallel extract each task unit, submitting the prompt word instances within it to the large language model. The model then generates the corresponding text paragraphs based on the roles, instructions, and data within the prompt words. The entire output structure clearly reflects the modular composition of the educational solution and achieves precise matching between task instructions and the data to be processed.
[0146] In the above implementation, a large, continuous structured course dataset is intelligently decomposed into multiple semantically cohesive and appropriately sized data modules through dynamic scanning and segmentation based on business rules. Then, by matching pre-set templates and injecting parameters, each data module is equipped with highly customized generation instructions and contextual roles, forming independent task packages. Finally, these task packages are systematically integrated and output as a clear, batch-processable segmented structured data format. This process cleverly solves the problems of context length pressure, ambiguous instructions, and logically chaotic or deviating-from-module focus issues caused by directly feeding large-scale structured data to a large language model.
[0147] In practical applications, this technical solution combines the "divide and conquer" strategy with the "precise prompting" technology, ensuring that the large model can process the most logically consistent subset of data with the most relevant context and the clearest instructions. This provides a crucial prerequisite for generating a well-structured and standardized educational solution text, and is one of the core innovative links in improving the overall system's generation quality, controllability, and efficiency.
[0148] Reference Figure 6 As one implementation of step S106, the steps of inputting segmented structured data into a large language model to perform inference generation operations, performing numerical consistency checks on the generated text segments, and outputting the text segments that pass the checks include:
[0149] Step S601: Input the segmented structured data into the pre-trained large language model, and generate an initial text paragraph corresponding to each semantic paragraph based on the prompt word examples and intermediate segment data contained in the segmented structured data.
[0150] Among them, the segmented structured data is a highly customized task unit after dynamic segmentation and prompt word assembly. Each unit contains two core parts: one is a prompt word example, which defines the roles, instructions and context required to generate the task; the other is the original "segmented intermediate data" subset pointed to by the prompt word, that is, the structured data content.
[0151] In the embodiments of this application, the mechanism by which the large language model performs reasoning and generation operations is to receive and understand this combined input. The model first parses the prompt words, establishes the "role" it should play (such as "course setting expert") and the "task instructions" it needs to complete (such as "please describe the following professional core course system in a professional and coherent manner, and summarize its credit composition").
[0152] Then, based on its vast pre-trained knowledge and understanding of instructions, the model creatively organizes and verbally expresses the subsequently provided structured data content as factual evidence. For example, it transforms key-value pairs like {“Course Name”:“Data Structures”, “Credits”:“3”, “Property”:“Required”} into natural language statements like “Required course 'Data Structures', worth 3 credits”, and integrates all course information into logically coherent and grammatically correct paragraphs. This process generates corresponding “initial text paragraphs” in parallel or sequentially for each independent semantic paragraph (such as “General Education Module”, “Professional Core Module”), completing the crucial transformation from machine-readable structured data to human-readable narrative text.
[0153] Step S602: Based on the initial text paragraph, identify and extract the predefined key field values in the education solution business rule base, and generate the extracted value results;
[0154] Because the generation of large language models is probabilistic and creative, the output text may be at risk in terms of factual accuracy, especially in the description of precise information such as numbers, which may be biased or omitted, a phenomenon known as the "illusion." Therefore, this step establishes an automated fact-checking mechanism for the generated text, the primary task of which is to accurately capture the key quantitative information mentioned in the text.
[0155] Specifically, the system performs key numerical extraction on the initial text paragraphs, locating and extracting numerical fields from the natural language text that are crucial to the business and must remain consistent with the original data. The most common example is the "credit value field." The system identifies and extracts these values using predefined pattern matching rules (such as targeted regular expressions) or a trained named entity recognition model. For example, from the text "This module contains 5 courses, totaling 15 credits.", the system needs to accurately extract "15" as the "credit value." This process is pure text parsing, and the output extracted numerical values represent the values that the model "believes" or "states."
[0156] Step S603: Based on the structured intermediate representation model, perform aggregation operations on the data entity values corresponding to the key fields to generate the original numerical sum;
[0157] In order to verify the accuracy of the values in the generated text, an absolutely reliable reference benchmark is required. The logical principle of this step is to bypass the potentially erroneous generated text and directly backtrack to the structured intermediate representation model generated in an earlier stage of the process, which served as the input source for the large model, for calculation.
[0158] Specifically, this structured intermediate representation model is the product of the preceding semantic mapping step. It is a clean, structured collection of business data (such as JSON objects), in which the attributes of each course (such as credits) are stored in explicit key-value pairs. The system performs raw numerical calculation operations, that is, iterates through the key-value pair data in the structured intermediate representation model, finds all relevant data entities for the key field that needs to be verified (such as credits), and performs aggregation operations (most commonly summation) on their values. For example, it calculates the sum of credits for all courses under the "Professional Core Module," thereby "generating the raw numerical sum."
[0159] Understandably, this summation comes from the original data that was initially parsed and structured from the table. Its accuracy has been guaranteed in the preceding OCR parsing, structural reconstruction, and semantic mapping steps. Therefore, it is regarded as the authoritative answer for verification, ensuring the independence and high reliability of the verification benchmark.
[0160] Step S604: Compare the extracted numerical results with the sum of the original numerical results, and generate a verification result identifier;
[0161] Among them, when the difference between the extracted numerical result and the sum of the original numerical values is less than a preset threshold, it is marked as consistent.
[0162] After obtaining the extracted numerical results from the generated text and the "true values" (the sum of the original numerical values) from the raw data, the core logic of this step is to perform automated comparison and judgment to objectively quantify the numerical accuracy of the generated content. The system directly compares these two values. Considering that the generated model may have reasonable rounding or unit conversions in its expression, the system usually sets a preset threshold (e.g., the difference is less than 0.5). When the difference is within the allowable range, they are judged to be consistent. The comparison result is formalized into a clear "verification result identifier," such as a Boolean value (True / False) or a status code (e.g., "PASS" / "FAIL").
[0163] In step S605, in response to the consistency indication of the verification result, the initial text paragraph is output as the text paragraph that has passed the verification.
[0164] When the indicator "Indication Consistency" is displayed, it means that the key values in the generated text match the original data and the text meets the accuracy standard. The system then "outputs the initial text paragraph as a verified text paragraph" and allows it to enter the final solution integration stage.
[0165] Conversely, if the indicators are inconsistent, a pre-defined error correction mechanism is triggered, such as regenerating the large language model. That is, the system can resubmit the paragraph containing the erroneous values, the original structured data, and the reinforcement instructions emphasizing numerical accuracy to the large model, requiring it to regenerate. This process can be iterative until the verification passes or the retry limit is reached.
[0166] In the above implementation, the powerful generation capabilities of the large language model are combined with the certainty of structured data processing. Through a "dual-track" numerical stream processing, one path extracts data from the generated text while the other path calculates it directly from the source structured data, thereby achieving automated and high-precision verification of key facts in the generated content.
[0167] In practical applications, this technical solution not only fully leverages the advantages of large models in language organization and creative description, but also effectively curbs the inherent "illusion" problem of large models in key numerical domains by introducing an independent calculation and comparison mechanism based on original structured data. The final decision output logic enables the entire system to have self-checking and correction capabilities, ensuring that the final generated educational program text is not only fluent and standardized in language, but also maintains strict consistency with the original document in core data (such as credits and the number of courses). This improves automation efficiency while ensuring the accuracy and credibility of the output results, achieving a balance between "quality" and "quantity" in the intelligent generation process.
[0168] Reference Figure 7As a further implementation of the intelligent generation method for educational programs, after the step of comparing the extracted numerical results with the sum of the original numerical results and generating a verification result identifier, the method further includes:
[0169] Step S701: In response to the inconsistency indicated by the verification result identifier, extract the corresponding key fields and generate text paragraph position information to generate an inconsistency analysis report;
[0170] Specifically, when the system detects a significant discrepancy between a key numerical value expressed in the generated text (such as the credit value "5.0" for a course) and the corresponding "data entity value" (the stored credit value "5.0") in the structured model, it generates an identifier instance to record this anomaly. This instance is a structured record containing metadata such as error type, location pointer, expected value, and actual value.
[0171] Next, after identifying specific instances of inconsistency, this step aims to perform in-depth analysis and contextualization of these discrete error points to generate a semantically rich diagnostic document that can be processed by the subsequent rule engine.
[0172] Specifically, the system first extracts the corresponding key fields based on the field paths or codes recorded in the instance, i.e., the core attribute names defined in the structured data model, such as "credit_hour" (credit hours) or "course_category" (course category). Next, the system extracts the paragraph location information of the generated text, which can be achieved by mapping it to the logical structure of the generated text (such as paragraph index, sentence number, or character offset), thereby accurately locating the language segment containing the erroneous statement in the final text output. This information, along with the specific inconsistencies (such as expected values, actual values, and deviations), is organized according to a predefined architecture to generate an inconsistency analysis report.
[0173] Step S702: Query the preset repair rule library and match the repair strategy instructions associated with the key fields; wherein, the repair rule library is used to store the field repair logic based on the business rules defined in the education solution.
[0174] Specifically, the pre-set repair rule base queried by the system is a knowledge base that encodes and stores the experience and standards of education business experts. The stored field repair logic defines the correction principles and steps that the system should follow when a specific type of inconsistency occurs in a specific field. For example, for inconsistencies in the "credit" field, one rule might be defined as "always overwrite with the source data value in the structured intermediate representation model"; for missing "course name", another rule might be defined as "select from the default name list of courses of the same category".
[0175] In this embodiment, the system uses the key fields extracted from the analysis report as primary keys to perform queries and matches in the repair rule base. This process may involve simple key-value lookups or complex reasoning based on rule priorities and conflict resolution strategies. The result of the matching is one or more repair strategy instructions, which are specific command codes that can be interpreted and executed by downstream components, such as "UPDATE_FIELD_WITH_SOURCE_VALUE" (update a field with the source value) or "RE_GENERATE_PARAGRAPH_WITH_TEMPLATE_X" (regenerate a paragraph using template X).
[0176] Step S703: According to the repair strategy instruction, update the data entity values in the structured intermediate representation model to generate the corrected structured intermediate representation model;
[0177] Among them, the structured intermediate representation model is the core data hub of the entire process. It carries the real values of all course entities and their attributes obtained from parsing the original tables in a structured form (such as JSON / XML).
[0178] In this embodiment, the system parses the operational semantics of the repair strategy instruction and locates the data entity value in the model that needs to be modified. For example, if the instruction requires resetting the credit field to the original value, the system will locate the specific credit attribute node in the model according to the field path in the analysis report, and then overwrite the current incorrect value with the correct value obtained from the original verification or extracted from other copies of the model.
[0179] It's important to note that the entire update operation is performed on the in-memory data model, requiring strict transactional compliance to prevent partial updates from causing inconsistencies in the model state. After the update is complete, the system generates a corrected structured intermediate representation model. This new model is a revised version of the original model, in which all identified numerical inconsistencies have been corrected according to business rules, thus ensuring the accuracy of core facts at the data source and laying a unique and accurate data foundation for regenerating correct text.
[0180] Step S704: Based on the corrected structured intermediate representation model, re-execute the dynamic segmentation processing, prompt word template configuration and large language model inference generation operations, and output the corrected text paragraphs;
[0181] Specifically, based on the original business logic segmentation rules (such as dividing by course modules), the corrected structured data containing the correct values is re-segmented into logical segments. Next, the prompt word template configuration is re-executed for each data segment, that is, the prompt word template is matched according to its module type, and the correct structured data is injected into it to form a new, data-accurate prompt word instance.
[0182] Finally, these new cue word instances are fed into a large language model to perform inference and generation operations. Based on the cue words containing accurate data, the large model performs a new round of text generation, thus outputting the corrected text paragraphs.
[0183] The core of this step lies in the fact that it doesn't simply patch up the existing erroneous text. Instead, it discards the old, flawed generated results and executes a standardized generation pipeline from scratch based on correct data. This ensures that the final output text not only matches the source data numerically, but its overall language organization and logical fluency are also reconstructed based on the correct context, thus guaranteeing high-quality output.
[0184] Step S705: The corrected text paragraph is taken as the text paragraph that has passed the verification.
[0185] The system marks the corrected text paragraph as "verification passed" to distinguish it from the initial version that failed verification. Then, in subsequent integration steps, the system uses this new, correct paragraph to replace the previously generated paragraph containing numerical errors, or inserts it into the corresponding position in the final document set. This operation ensures that all content in the final assembled complete educational solution document is reliable text that has undergone consistency verification (or has been corrected after verification).
[0186] The above implementation effectively solves the problem of numerical inconsistencies that may occur when large language models generate normative texts such as educational plans. It not only ensures the accuracy of facts by tracing back and correcting the structured intermediate representation model—the data root—but also guarantees the overall quality of the text by triggering a full-process regeneration based on correct data. This reduces the cost and time of manual review and modification of the generated results, improves the output reliability, business compliance, and automation level of the entire intelligent generation system, and ensures that the automatically generated documents meet the standard of direct usability in terms of the accuracy of key data.
[0187] Reference Figure 8 As a further implementation of the intelligent generation method for educational solutions, before step S107, which involves integrating and validating the text paragraphs, concatenating them according to the logical order defined in the educational solution business rule base, and converting them into a standard document format, the method further includes:
[0188] Step S801: Obtain data entity field information in the structured intermediate representation model, scan the course name attribute field and course category attribute field, and generate a course entity attribute list;
[0189] The structured intermediate representation model is a highly standardized data collection formed after deep semantic mapping. It usually exists in formats with clear hierarchy and key-value structure, such as JSON and XML. Each course entity has been assigned attribute tags with business significance, such as "course name" and "course category".
[0190] In this embodiment, a programmed scanning operation is used to accurately locate and extract the core entities and attributes necessary for subsequent logical reasoning from the complex nested structure of the model. Specifically, the system selects a path based on predefined fields (such as JSON Path $.courses[*].name) and traverses all course data nodes in the model. For each course node, the system not only extracts its "course name" as a unique identifier for the entity, but also simultaneously extracts its key classification attributes such as "course category". These extracted discrete information are repackaged and aggregated to form a new, flattened list of course entity attributes.
[0191] Understandably, the course entity attribute list is essentially a course metadata index that has removed the complex relationships and redundant information that may exist in the original model. It contains only the minimum necessary set of information for relational queries (such as ID, name, and category), providing a standardized and lightweight input interface for efficient and batch queries to external knowledge bases.
[0192] Step S802: Based on the course entity attribute list, query the pre-set educational knowledge relation database to obtain the set of preceding course relation data corresponding to the course name; wherein, the educational knowledge relation database stores the logical dependency rules between courses;
[0193] Specifically, the Educational Knowledge Relationship Database is a pre-built structured database that stores logical rules between courses within a domain. Its data model is designed to represent relationships such as "course A is a prerequisite course for course B".
[0194] In this embodiment, the system initiates a batch query to this knowledge base using the course name in the "Course Entity Attribute List" as the query key. The query operation is typically performed using a database query language (such as SQL) or a graph query language (such as Cypher), and its core is matching the association fields of the dependency rule table in the knowledge base. The query result is "retrieving the set of predecessor course relationship data corresponding to the course name," where each record clearly defines a typed dependency relationship from "predecessor course" to "current course."
[0195] Understandably, the OCR parsing process can only extract explicit text and numerical attributes from the physical layout of a table. However, the complex logical constraints between courses are deep-seated business rules that are not usually listed in an intuitive table format. This key semantic information must be supplemented by the introduction of external structured knowledge sources to lay the foundation for a complete and accurate understanding of the internal logic of the educational program.
[0196] Step S803: Based on the set of prior course relationship data, perform a graph construction operation to generate a course logical relationship graph model, which includes a graph structure with course entities as nodes and dependencies as edges.
[0197] Specifically, performing graph construction is a transformation process from relational data to graph-structured data. The system takes the set of prerequisite course relationship data as input and transforms each triple record (current course, dependency type, prerequisite courses) into a basic element in graph theory: each unique course is instantiated as a node, and each node is appended with attributes inherited from the course entity attribute list (such as course category); each dependency relationship is instantiated as a directed "edge," with the edge pointing from the knowledge to the current course, and the edge itself can carry "dependency type" as an attribute. By systematically processing all relationship records and merging identical nodes and edges, a course logical relationship graph model is finally generated.
[0198] In this embodiment, the course logical relationship graph model is a powerful mathematical abstraction that intuitively reveals the internal structure of a course group (e.g., which courses constitute a tightly connected prerequisite knowledge chain) and the boundaries between modules using a network of nodes and edges. Compared to linear lists, graph structures can more efficiently support the calculation and analysis of complex logical attributes such as accessibility between courses, loop detection, and critical paths. It is a core step in transforming implicit business rules into a computable and reasonable form.
[0199] Step S804: Inject the course logical relationship graph model into the task instruction parameter part of the prompt word template to generate a constraint-enhanced prompt word instance;
[0200] The underlying logic of this step is to bridge the modal gap between structured graph data and natural language model input. The raw data format of the graph (such as an adjacency matrix) directly input is something that large language models cannot directly understand.
[0201] Therefore, the system needs to perform an injection operation, which involves serializing the course logical relationship graph model into a human-readable, descriptive text summary (e.g., transforming it into a statement or structured list such as "Course dependencies include: 'Advanced Mathematics' must be taken before 'Probability Theory', and 'Probability Theory' must be taken before 'Machine Learning'"). This text description is then injected into the task instruction parameter section of the prompt word template. The prompt word template is a predefined, placeholder-based natural language instruction framework.
[0202] Then, using template engine technology, the textual description of the graph is filled into specific slots in the instructions regarding logical constraints, thereby generating constraint-enhanced prompt word instances. These newly generated prompt word instances explicitly include hard constraint descriptions of inter-course dependencies in their task instructions, upgrading the task from generating text based solely on static course attributes to a text generation task under conditional constraints that must simultaneously satisfy given logical rules.
[0203] Step S805: Replace the original prompt word template in the segmented structured data with constraint-enhanced prompt word instances, update the segmented structured data, and input the updated segmented structured data into the large language model to perform inference generation operations, generating and outputting text paragraphs with enhanced logical constraints.
[0204] The underlying logic of this step is to leverage the powerful adherence of large language models to natural language instructions, proactively avoiding logical contradictions during the generation process. The system first replaces the original prompt template with constraint-enhanced prompt word instances, thereby updating the core instruction portion of the entire segmented structured data.
[0205] In this embodiment, when the updated segmented structured data is input into the large language model to perform inference and generation operations, the large model, while interpreting the prompt words, simultaneously understands two requirements: first, to generate a basic description of a certain module's courses; and second, to adhere to the explicitly stated order of courses in the instructions. In each step of the autoregressive text generation process, the model's internal inference mechanism subconsciously references these constraints, tending to select words and expressions that conform to logical order, thereby generating and outputting text paragraphs with enhanced logical constraints.
[0206] For example, when describing the order of courses, prerequisite courses are naturally listed before subsequent courses, and conjunctions indicating dependencies, such as "only after mastering… can one proceed to…", are used. This process, by transforming external structured knowledge into internal contextual constraints during model generation, greatly reduces the probability of logical fallacies such as reversed course order or disordered dependencies in the generated text.
[0207] In the above implementation, the boundaries of OCR parsing are extended from the explicit physical structure and attributes of tables to the implicit logical dependencies inherent in educational programs. Information gaps in the original documents are filled by querying external educational knowledge relation databases, and graph modeling is innovatively used to transform complex course relationship networks into computable and injectable structured constraints. Finally, by dynamically generating and injecting constraint-enhancing prompts, these structured constraints are seamlessly integrated into the generation context of a large language model, thereby achieving intelligent guidance and logical correction of the automated text generation process.
[0208] In practical applications, this technical solution fundamentally improves the logical rigor and business compliance of generated educational program texts, ensuring the accurate expression of key business logic such as course offering order and prerequisite / advancement relationships. It makes the automatically generated program texts closer to the review standards of academic affairs experts in terms of depth, and is an effective and innovative path to solve the common problem of "accurate facts but chaotic logic" in the field of intelligent document generation.
[0209] This application also discloses an intelligent educational program generation system based on OCR table semantic parsing.
[0210] An intelligent educational solution generation system based on OCR table semantic parsing, specifically including:
[0211] The document processing module is used to acquire the original documents of the educational program uploaded by the user and perform image preprocessing operations to generate standardized image data; the original documents include PDF electronic documents or scanned image formats.
[0212] The OCR recognition and layout analysis module is used to perform optical character recognition processing on standardized image data, extract text blocks and line elements in the document, calculate the page coordinate information of each text block and generate a discrete text list, and at the same time recognize the boundary coordinates of table areas in the document.
[0213] The table physical structure reconstruction module is used to perform table physical structure reconstruction operations based on page coordinate information and table area boundary coordinates in the discrete text list, and generate a table physical matrix.
[0214] The structured modeling module is used to perform semantic mapping operations on the physical matrix of tables using a pre-built education scheme business rule library and keyword library, anchoring the table header cells and associating data entities to establish key-value pair relationships, and generating a structured intermediate representation model.
[0215] The dynamic segmentation and prompt word assembly module is used to perform dynamic segmentation processing on the structured intermediate representation model according to the logical segmentation rules in the education solution business rule library, splitting it into multiple semantic segments, and configuring prompt word templates containing role definitions, task instructions and structured data for each semantic segment to generate segmented structured data.
[0216] The model inference and verification module is used to input segmented structured data into the large language model to perform inference and generation operations, and to perform numerical consistency verification on the generated text segments, and output the text segments that pass the verification.
[0217] The document integration and output module is used to integrate validated text paragraphs, concatenate them according to the logical order defined in the education solution business rule library, convert them into a standard document format, and generate the final education solution text.
[0218] The intelligent educational scheme generation system based on OCR table semantic parsing in this application embodiment can implement any of the above methods, and the specific working process of each module in the system can refer to the corresponding process in the above method embodiments.
[0219] In the several embodiments provided in this application, it should be understood that the provided methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for example, the division of a certain module is merely a logical functional division, and in actual implementation there may be other division methods, such as multiple modules can be combined or integrated into another system, or some features can be ignored or not executed.
[0220] This application also discloses a computer-readable storage medium.
[0221] A computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described above in any of the methods for intelligent generation of educational programs based on OCR table semantic parsing.
[0222] The computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device; the program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.
[0223] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.
Claims
1. A method for intelligently generating educational plans based on OCR table semantic parsing, characterized in that, The method includes: The system obtains the original document of the educational program uploaded by the user and performs image preprocessing to generate standardized image data; wherein the original document includes PDF electronic document or scanned image format; Optical character recognition processing is performed on the standardized image data to extract text blocks and line elements from the document, calculate the page coordinate information of each text block and generate a discrete text list, and simultaneously identify the boundary coordinates of table areas in the document. Based on the page coordinate information and table area boundary coordinates in the discrete text list, a table physical structure reconstruction operation is performed to generate a table physical matrix. Using a pre-built education scheme business rule base and keyword base, a semantic mapping operation is performed on the physical matrix of the table to anchor the table header cells and associate data entities to establish key-value pair relationships and generate a structured intermediate representation model. According to the logical segmentation rules in the education solution business rule base, the structured intermediate representation model is dynamically segmented into multiple semantic segments, and each semantic segment is configured with a prompt word template containing role definition, task instructions and structured data to generate segmented structured data. The segmented structured data is input into the large language model to perform inference and generation operations, and numerical consistency checks are performed on the generated text segments. The text segments that pass the checks are then output. The validated text segments are integrated, concatenated according to the logical order defined in the education solution business rule base, and converted into a standard document format to generate the final education solution text.
2. The method for intelligently generating educational plans based on OCR table semantic parsing according to claim 1, characterized in that, The steps of performing optical character recognition processing on the standardized image data, extracting text blocks and line elements from the document, calculating the page coordinate information of each text block and generating a discrete text list, and simultaneously recognizing the boundary coordinates of table areas in the document include: Perform document layout analysis on the standardized image data to identify text blocks and line elements in the document, and calculate the page coordinate information and corresponding text content of each text block; Based on the page coordinate information of the text blocks, text blocks that meet the preset line alignment conditions are aggregated to generate a text line set, and the line position range of each text line in the text line set is determined. Based on the row position range of the text line set, detect text line clusters with column alignment relationships, and calculate the bounding box coordinates of the text line clusters as the table area boundary coordinates; The text content and page coordinate information of all text blocks are associated and integrated with the boundary coordinates of the table area to generate the discrete text list.
3. The method for intelligently generating educational plans based on OCR table semantic parsing according to claim 2, characterized in that, Based on the page coordinate information and table region boundary coordinates in the discrete text list, the steps for performing a table physical structure reconstruction operation and generating a table physical matrix include: Based on the page coordinate information and the table area boundary coordinates, calculate the geometric features of the grid rows and columns, and construct a virtual grid model; Based on the virtual grid model, the geometric center coordinates of each text block are calculated and mapped to the row and column index positions of the virtual grid model to generate cell position mapping data; Based on the cell position mapping data, the geometric dimensions of the text block and the proportional relationship between them and the virtual mesh model are analyzed, and the number of mesh cells covered by the text block is determined. The row span and column span attributes are marked, and cell row span and column span attribute data are generated. The cell position mapping data and cell cross-row and column attribute data are integrated to generate the table physical matrix.
4. The method for intelligently generating educational plans based on OCR table semantic parsing according to claim 1, characterized in that, The steps for generating a structured intermediate representation model by using a pre-built educational solution business rule base and keyword base to perform semantic mapping operations on the physical matrix of the table, anchoring the table header cells and associating data entities to establish key-value pair relationships include: Based on the pre-set education solution business rule base and keyword base, scan the cell content in the physical matrix of the table, match predefined education business keywords, identify and mark the table header cells, and generate anchored table header information; Based on the anchored header information and the row and column index data of the table physical matrix, a key-value pair mapping relationship is established between data cells and header cells, and key-value pair associated data is generated; Based on the key-value pair association data, the serialization mapping relationship is converted into a computer-readable structured format, generating a structured intermediate representation model.
5. The method for intelligently generating educational programs based on OCR table semantic parsing according to claim 4, characterized in that, Based on the logical segmentation rules in the education solution business rule base, the structured intermediate representation model is dynamically segmented into multiple semantic segments. Each semantic segment is then configured with a prompt word template containing role definitions, task instructions, and structured data. The steps for generating segmented structured data include: Obtain the logical segmentation rules from the business rule base of the education solution; Scan the course category attribute field in the structured intermediate representation model, detect the preset segmentation trigger conditions, and generate semantic paragraph segmentation points based on the logical segmentation rules; Based on the semantic paragraph segmentation points, a data slicing operation is performed on the structured intermediate representation model to extract the key-value pair dataset corresponding to each semantic paragraph and generate a segmented intermediate data set. Based on the segmented intermediate data set, the corresponding template in the preset prompt word template library is matched, and role definition parameters, task instructions and structured data content are injected into each data unit to generate a configured prompt word instance; Aggregate all the aforementioned prompt word instances and associated segmented intermediate data to generate segmented structured data.
6. The method for intelligently generating educational programs based on OCR table semantic parsing according to claim 5, characterized in that, The steps of inputting the segmented structured data into a large language model to perform inference and generation operations, performing numerical consistency checks on the generated text segments, and outputting the text segments that pass the checks include: The segmented structured data is input into a pre-trained large language model, and based on the prompt word examples and intermediate segment data contained in the segmented structured data, an initial text paragraph corresponding to each semantic paragraph is generated. Based on the initial text paragraph, identify and extract the predefined key field values in the education program business rule base, and generate the extracted value results; Based on the structured intermediate representation model, aggregation operations are performed on the data entity values corresponding to the key fields to generate the original numerical sum; The extracted numerical results are compared with the sum of the original numerical results to generate a verification result identifier; In response to the consistency indicated by the verification result identifier, the initial text paragraph is output as the text paragraph that has passed the verification.
7. The method for intelligently generating educational programs based on OCR table semantic parsing according to claim 6, characterized in that, After the step of comparing the extracted numerical results with the sum of the original numerical results to generate a verification result identifier, the method further includes: In response to the inconsistency indicated by the verification result identifier, the corresponding key fields are extracted and text paragraph position information is generated to generate an inconsistency analysis report; The system queries a pre-defined repair rule base and matches repair strategy instructions associated with the key fields; wherein, the repair rule base is used to store field repair logic based on the business rules defined in the education solution. According to the repair strategy instruction, update the data entity values in the structured intermediate representation model to generate the corrected structured intermediate representation model; Based on the modified structured intermediate representation model, the dynamic segmentation process, prompt word template configuration, and large language model inference generation operations are re-executed to output the modified text paragraphs. The corrected text paragraph is taken as the text paragraph that passes the verification.
8. A method for intelligently generating educational programs based on OCR table semantic parsing according to any one of claims 1 to 7, characterized in that, Before the steps of integrating the validated text paragraphs, concatenating them according to the logical order defined in the education program business rule base, and converting them into a standard document format, the method further includes: Obtain the data entity field information in the structured intermediate representation model, scan the course name attribute field and course category attribute field, and generate a course entity attribute list; Based on the course entity attribute list, a pre-set educational knowledge relationship database is queried to obtain a set of preceding course relationship data corresponding to the course name; wherein, the educational knowledge relationship database stores logical dependency rules between courses; Based on the aforementioned set of prior course relationship data, a graph construction operation is performed to generate a course logical relationship graph model, which includes a graph structure with course entities as nodes and dependencies as edges. The course logical relationship graph model is injected into the task instruction parameter part of the prompt word template to generate a constraint-enhanced prompt word instance; The constraint-enhanced prompt word instance replaces the original prompt word template in the segmented structured data, updates the segmented structured data, and inputs the updated segmented structured data into the large language model to perform inference generation operations, generating and outputting text paragraphs with enhanced logical constraints.
9. An intelligent educational program generation system based on OCR table semantic parsing, characterized in that, The system includes: The document processing module is used to acquire the original document of the educational program uploaded by the user and perform image preprocessing operations to generate standardized image data; wherein, the original document includes PDF electronic document or scanned image format; The OCR recognition and layout analysis module is used to perform optical character recognition processing on the standardized image data, extract text blocks and line elements in the document, calculate the page coordinate information of each text block and generate a discrete text list, and at the same time recognize the boundary coordinates of table areas in the document. The table physical structure reconstruction module is used to perform table physical structure reconstruction operations based on the page coordinate information and table area boundary coordinates in the discrete text list, and generate a table physical matrix. The structured modeling module is used to perform semantic mapping operations on the physical matrix of the table using a pre-built education scheme business rule library and keyword library, anchor the table header cells and associate data entities to establish key-value pair relationships, and generate a structured intermediate representation model. The dynamic segmentation and prompt word assembly module is used to perform dynamic segmentation processing on the structured intermediate representation model according to the logical segmentation rules in the education solution business rule library, splitting it into multiple semantic segments, and configuring prompt word templates containing role definitions, task instructions and structured data for each semantic segment to generate segmented structured data. The model inference and verification module is used to input the segmented structured data into the large language model to perform inference and generation operations, and to perform numerical consistency verification on the generated text segments and output the text segments that pass the verification. The document integration and output module is used to integrate the verified text paragraphs, concatenate them according to the logical order defined in the education solution business rule base, convert them into a standard document format, and generate the final education solution text.
10. A computer-readable storage medium, characterized in that: The computer program is stored that can be loaded by a processor and executed as described in any one of claims 1 to 8.