Table data intelligent identification analysis and articulation system and method based on large model

Abstract

The invention relates to the technical field of data processing, in particular to an intelligent table data recognition, analysis and articulation system and method based on a large model, and the system comprises an image preprocessing module, a table analysis and recognition module, a key basic information extraction module, a multi-column cross-page complex table semantic splicing module and a rule-based articulation verification module. The image preprocessing module is used for performing red stamp removal, tilt correction and direction calibration operation on an input table image; the table analysis and identification module is used for identifying and analyzing a table image and carrying out subject standardization and target data extraction; the key basic information extraction module is used for extracting key metadata of the table; the multi-column cross-page complex table semantic splicing module is used for performing semantic fusion and recombination on complex tables and outputting standardized table data; and the rule articulation-based verification module is used for performing consistency verification and abnormality diagnosis on the analyzed table data. According to the invention, the accuracy, integrity and verification efficiency of table data processing are improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a system and method for intelligent identification, parsing and cross-referencing of tabular data based on a large model. Background Technology

[0002] In the field of data processing, the accurate recognition and parsing of tabular image data is a core aspect. Current mainstream processing methods mainly rely on traditional OCR mini-models or direct manual processing. Traditional OCR mini-models have significant technical bottlenecks. For example, they suffer from high text recognition error rates when faced with issues such as red stamp obscuring the text, page tilt, or interference lines. Furthermore, they lack effective semantic association capabilities for abnormal structures such as multi-page tables or tables without headers, often resulting in data breaks or misalignments, which prevents subsequent verification from proceeding properly. Manually entering and parsing tabular data is inefficient, has a high error rate, and consumes a significant amount of manpower.

[0003] While open-source multimodal large models (such as general visual language models) have improved numerical recognition accuracy, they suffer from limitations in scene adaptation: their built-in preprocessing modules cannot effectively address issues such as red stamp removal and skew correction; the output of some mainstream models (such as PaddleOCR-VL) cannot perform cross-page table merging and cross-checking of table data, requiring extensive manual review to ensure accuracy, resulting in high labor costs. Therefore, there is an urgent need for an end-to-end table data processing system that can integrate the advantages of large and small models, incorporate cross-checking rules, and adapt to complex scenarios. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent identification, parsing, and cross-referencing system for tabular data based on a large model, which solves the problems of poor adaptability to complex scenarios, inaccurate cross-page table splicing, low data cross-referencing efficiency, and high labor costs in existing technologies for tabular data processing.

[0005] To achieve the above objectives, the present invention provides an intelligent recognition, parsing and verification system for tabular data based on a large model, comprising an image preprocessing module, a table parsing and recognition module, a key basic information extraction module, a semantic splicing module for complex multi-column cross-page tables, and a rule-based verification module connected in sequence. The image preprocessing module is used to perform red stamp removal, tilt correction, and orientation calibration operations on the input table image; The table parsing and recognition module is used to perform high-precision text recognition and structured parsing on the preprocessed table image, and to perform subject standardization and target year data extraction on the parsing results to obtain the original table data; The key basic information extraction module is used to extract key metadata information such as enterprise name, table type, table body, reporting period, and monetary unit from the output of the table parsing and recognition module. The semantic splicing module for complex multi-column cross-page tables is used to perform semantic-level fusion and data reorganization on complex tables with multiple columns or content split across pages, and output standardized table data with a unified structure. The rule-based verification module is used to perform real-time consistency verification and anomaly diagnosis on the standardized table data based on predefined verification rules.

[0006] The image preprocessing module includes: The stamp-aware block processing unit is used to divide the input image into multiple sub-regions and use a dual-threshold collaborative stamp removal algorithm to process the sub-regions containing red stamps, thereby removing red stamps while protecting black text. The non-rigid tilt correction unit is used to determine the optimal correction angle by calculating the sum of squares of the first-order differences between adjacent columns of the vertical projection of the rotated binary image, and to perform rotation correction on the image. The table semantic orientation calibration unit is used to employ a dual-modal orientation decision mechanism, including OCR semantic verification and rule verification based on aspect ratio and keyword distribution, to ensure that the table orientation is positive.

[0007] Specifically, the table parsing and recognition module is used for: The locally deployed PaddleOCR-VL visual language model is used to recognize table images and output a Markdown-formatted structured text stream that preserves the topological relationship of the table rows and columns. For the subject column text in the structured text stream, subject standardization is performed by calculating its comprehensive similarity score with the subject names in the predefined standard subject library; Based on the timeliness requirements of table analysis, the data columns for the target year or target period are automatically extracted from the table header using either year-aware mode or relative period mode.

[0008] The overall similarity score is calculated using the following formula: ; in: , These represent the target words to be compared identified by the large model and the standard subject names in the standard subject library, respectively. Matching keywords; For fuzzy matching; For semantic similarity matching; , , These are the weighting coefficients.

[0009] Specifically, the key basic information extraction module is used for: By matching preset prefixes or recognizing high-frequency enterprise identifiers, enterprise names can be extracted from OCR-recognized text or file names; A two-level strategy, consisting of direct keyword matching and semantic feature analysis based on feature keyword sets, is used to identify table types and whether the table body is a single entity or a merged entity. A multi-regular expression joint matching strategy is adopted to parse reporting period information in various formats from the header area and select the latest year-month combination; OCR identifies monetary unit keywords in text by using regular expressions to iterate and match.

[0010] The semantic splicing module for multi-column, cross-page complex tables includes: The multi-column table semantic splicing unit is used to identify the pattern of repeated column names by column name frequency statistics. After determining that it is a multi-column table, it performs data column adaptation and splicing based on an adaptive column index matching algorithm. The cross-page table semantic splicing unit is used to locate cross-page data in the same table based on multi-dimensional related indexes, perform order calibration by using the page number identifier in the file name, and verify the consistency of the number of columns in each page before merging.

[0011] Specifically, the rule-based verification module is used for: Dynamically load the corresponding list of audit indicators and industry-standardized formulas based on the table metadata; Retrieve the required standard subject values ​​for the formula from standardized table data and calculate the theoretical values; The theoretical value is compared with the value identified in the table. Anomaly diagnosis is initiated for inconsistent results that exceed the threshold, and feedback is provided through a visual interface.

[0012] On the other hand, the present invention also includes a method for intelligent identification, parsing, and cross-referencing of tabular data based on a large model, the method comprising: The input table image is preprocessed, including removing red stamps, tilt correction, and orientation calibration. The preprocessed table image is subjected to high-precision character recognition and structured parsing. The parsing results are then standardized by subject and the target year data is extracted to obtain the original table data. Extract key metadata information such as company name, table type, table body, reporting period, and monetary unit from the original table data; Perform semantic-level fusion and data reorganization on complex tables with multiple columns or content split across pages to obtain standardized table data with a unified structure; Based on predefined checklist rules, the standardized table data is subjected to real-time consistency verification and anomaly diagnosis.

[0013] This invention discloses a large-model-based intelligent recognition, parsing, and cross-referencing system for tabular data. The system effectively addresses issues such as seal occlusion and misalignment in original images through an image preprocessing module. It utilizes a large model in the table parsing and recognition module to achieve high-precision structure and content recognition, and improves data standardization through post-processing. A key information extraction and complex table splicing module ensures the integrity and semantic coherence of fragmented, cross-page, and other multi-source heterogeneous tabular data. Finally, a rule-based cross-referencing verification module performs automated logical verification on the integrated standard data, achieving end-to-end intelligent processing from images to trusted structured data.

[0014] The beneficial effects of this invention are: significantly improving the accuracy and structuring of table data recognition; realizing automated splicing and semantic fusion of complex layouts and cross-page tables; greatly improving the efficiency and reliability of data verification through the built-in checklist rule engine; reducing the cost of manual review; and supporting large-scale, high-concurrency enterprise-level table processing needs. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0016] Figure 1 This is a schematic diagram of the structure of the intelligent identification, parsing, and cross-referencing system for tabular data based on a large model according to the present invention.

[0017] Figure 2 This is a schematic diagram of the image preprocessing module of the present invention.

[0018] Figure 3 This is a flowchart of the intelligent identification, parsing, and cross-referencing method for tabular data based on a large model, as described in this invention.

[0019] In the diagram: 101-Image preprocessing module, 102-Table parsing and recognition module, 103-Key basic information extraction module, 104-Multi-column cross-page complex table semantic splicing module, 105-Rule-based verification module, 106-Seal perception block processing unit, 107-Non-rigid tilt correction unit, 108-Table semantic direction calibration unit, 109-Multi-column table semantic splicing unit, 110-Cross-page table semantic splicing unit. Detailed Implementation

[0020] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, but should not be construed as limiting the present invention.

[0021] Please see Figure 1 and Figure 2 ,in Figure 1 This is a schematic diagram of the structure of a large-scale model-based intelligent recognition, parsing, and verification system for tabular data. Figure 2 This is a schematic diagram of the image preprocessing module.

[0022] This invention provides a large-model-based intelligent recognition, parsing, and verification system for tabular data, comprising an image preprocessing module 101, a table parsing and recognition module 102, a key basic information extraction module 103, a multi-column, cross-page complex table semantic splicing module 104, and a rule-based verification module 105 connected in sequence. The image preprocessing module 101 is used to perform red stamp removal, tilt correction, and orientation calibration operations on the input table image.

[0023] Specifically, the image preprocessing module 101 includes: The stamp-aware segmentation processing unit 106 is used to divide the input image into multiple sub-regions and process the sub-regions containing red stamps using a dual-threshold collaborative stamp removal algorithm, thereby removing the red stamps while protecting the black text.

[0024] In this embodiment, the seal perception and segmentation processing unit 106 performs the red seal removal operation, which adopts a dynamic segmentation strategy: considering that the current red seal removal operation is a pixel-level operation, in order to reduce the amount of computation and reduce the impact of the red seal removal operation on non-red seal areas, the input image is divided into R rows and C columns according to the table structure features, with a default R=16 and C=12, where the number of rows and columns is determined by table layout analysis: ; Indicates the height of the current image; Indicates the width of the current image; This represents the average height of the historical table cells, and is a fixed value. This represents the average width of the historical table cells, and is a fixed value.

[0025] Dual-threshold collaborative stamp removal: For each sub-region, extended HSV red domain detection is first applied, and a dual-threshold mask of light red ([0,40,40]~[15,255,255]) and dark red ([160,40,40]~[180,255,255]) is fused. Processing is triggered when the proportion of red pixels is >0.5%. For sub-images containing red stamps, adaptive OTSU fusion is performed to generate two stamp removal images: the high-threshold path retains dark text, and the low-threshold path eliminates light stamps. The two outputs are fused as the final output, thus removing red stamps while protecting black text.

[0026] Furthermore, the non-rigid tilt correction unit 107 is used to determine the optimal correction angle by calculating the sum of squares of the first-order differences between adjacent columns of the vertical projection of the rotated binary image, and to perform rotation correction on the image.

[0027] In this embodiment, the non-rigid tilt correction unit 107 projects the rotated binary image vertically and calculates the sum of squares of the first-order differences between adjacent positions in the horizontal direction. This sum is used as the tilt angle to evaluate the score, and the highest score corresponds to the state where the text line or table line is closest to horizontal alignment. The calculation formula is as follows: ; ; It is a binary image obtained by rotating the original image by an angle θ. It is a two-dimensional matrix with a size of H×W. x and y represent rows and columns, with a value of 0 representing the background and 255 representing the foreground, such as text or table lines.

[0028] The summation of all rows in the x-th column of the image represents the "total foreground pixels" in the x-th column horizontally at a rotation angle θ. If text or table lines cross this column, then... The value is relatively large; if it is a blank column, the value is close to 0.

[0029] The first-order difference represents the projection values ​​of two adjacent columns and is used to measure the degree of change in the projection signal in the horizontal direction. When text or table rows are horizontally aligned, the character edges will produce sharp abrupt changes on the projection curve (such as from blank space to text, or gaps between characters), resulting in a larger difference value; if the image is tilted, the text is distributed diagonally, the projection change tends to be gentle, and the difference value is smaller.

[0030] The square of the difference is used to eliminate the influence of the sign. Whether the transition is rising or falling, it is an effective feature. Squaring it can enhance significant transitions, and squaring large transitions contributes more.

[0031] Summing over all adjacent column pairs (W-1 pairs in total) yields a scalar score. This reflects the horizontal structural clarity of the entire image at that rotation angle.

[0032] Specifically as follows: Adaptive binarization based on the maximum variance method: The Otsu thresholding algorithm is used to automatically determine the optimal binarization threshold without the need for manual threshold setting. Under the premise of adapting to images under different lighting conditions, it accurately separates the table content from the background and provides a clear binary image for subsequent angle detection.

[0033] Multi-angle rotation search strategy: A sequence of candidate rotation angles is generated within a defined angle range with a specified step size. Batch processing is achieved through image stacking, improving computational efficiency. The project code uses the `delta` parameter to control precision and the `limit` parameter to limit the search range, balancing precision and computational complexity.

[0034] An angle scoring function based on projection contour variance calculates the sum of squared differences between adjacent pixels in each row as the scoring standard. The gradient change of the row projection contour is most dramatic when the table is horizontal. In other words, variance quantifies the horizontality of the image; a larger variance indicates better alignment.

[0035] High-quality image rotation correction: The optimal correction angle is determined by maximizing the scoring function, using data-driven angle selection to avoid the limitations of heuristic rules. Image rotation is based on affine transformation, using cubic interpolation to maintain image quality, and a boundary replication mode to avoid black border effects.

[0036] Furthermore, the table semantic orientation calibration unit 108 is used to employ a dual-modal orientation decision mechanism, including OCR semantic verification and rule verification based on aspect ratio and keyword distribution, to ensure that the table orientation is positive.

[0037] In this embodiment, the table semantic orientation calibration unit 108 adopts a dual-modal orientation decision mechanism. First, it uses OCR semantic verification and calls a lightweight OCR engine to detect the text orientation. If the confidence level is >90%, it rotates the text directly.

[0038] When OCR fails, the table orientation is calibrated using the aspect ratio-field distribution joint rule. If H / W < 1 and keywords such as "assets" or "row number" are detected in the lower left quadrant, the table is rotated 90° clockwise. If the same keywords are detected in the upper right quadrant, the table is rotated 90° counterclockwise.

[0039] If H / W >= 1, and keywords such as "unit head" are detected at the bottom, it is considered a positive result.

[0040] Furthermore, the table parsing and recognition module 102 is used to perform high-precision text recognition and structured parsing on the preprocessed table image, and to perform subject standardization and target year data extraction on the parsing results to obtain the original table data.

[0041] Specifically, the table parsing and recognition module 102 is used for: The locally deployed PaddleOCR-VL visual language model is used to recognize table images and output a Markdown-formatted structured text stream that preserves the topological relationship of the table rows and columns. For the subject column text in the structured text stream, subject standardization is performed by calculating its comprehensive similarity score with the subject names in the predefined standard subject library; Based on the timeliness requirements of table analysis, the data columns for the target year or target period are automatically extracted from the table header using either year-aware mode or relative period mode.

[0042] In this embodiment, the table parsing and recognition module 102 employs a locally deployed open-source PaddleOCR-VL model to achieve high-precision text recognition and structured parsing of table content in table images. The core component of this model is PaddleOCR-VL-0.9B, a compact and powerful visual language model (VLM) composed of a NaViT-style dynamic resolution visual encoder and an ERNIE-4.5-0.3B language model, enabling accurate element recognition. The model input is a preprocessed table image, and the output is a structured text stream in Markdown format, fully preserving the table's row and column topological relationships and original semantic content.

[0043] Based on the Markdown format output of PaddleOCR-VL, this module further performs a two-stage post-processing workflow to ensure that the data conforms to the analysis specifications. Subject Standardization: User-uploaded forms contain diverse and non-standardized subject names, which may also lead to OCR errors, such as misrecognizing standard subject names as similar-looking typos, like misrecognizing "project" as "item day". This system employs a multi-level mapping mechanism to convert the target subject to be matched into a standard subject.

[0044] First, a predefined standard subject library ; This standard subject library covers all standard subjects that customers commonly use and that fit the current form category.

[0045] For each row of the subject column in the md table Calculate its overall similarity score with the standard subjects in the standard subject library S. The calculation formula is as follows: ; in: , These represent the target words to be compared identified by the large model and the standard subject names in the standard subject library, respectively. For keyword matching, this system calculates the inclusion relationship between target words and standard subject names. For fuzzy matching, this system normalizes similarity using Levenshtein distance, and considers the target word and standard subject to be consistent when the similarity exceeds a set threshold; For semantic similarity matching, the cosine similarity between the target word and the standard subject is calculated based on the semantic embedding of the FinBERT-wwm fine-tuning model. , , These are weighting coefficients. This system sets them as fixed hyperparameters based on actual conditions.

[0046] Target year data extraction: Based on the timeliness requirements of table analysis, the latest reporting period values ​​are automatically extracted. Different extraction rules are designed for target year columns that contain years and relative periods in the table header.

[0047] Year-aware mode: For example, if the headers of a numerical list are "December 2022" and "December 2023", regular expressions are used to extract the year and month information from the header rows. The latest date column is selected as the target column by comparison, which supports subsequent numerical mapping with standard subject columns.

[0048] Relative periodic mode: For example, if the headers of the numerical list are "beginning balance" and "ending balance", the column related to "ending" is selected using predefined semantic rules. Specifically, the multiple possible expressions of the ending balance column are predefined, and the target column is selected by keyword matching to support subsequent numerical mapping with the standard account column.

[0049] Furthermore, the key basic information extraction module 103 is used to extract key metadata information such as enterprise name, table type, table body, reporting period, and monetary unit from the output of the table parsing and recognition module 102.

[0050] Specifically, the key basic information extraction module 103 is used for: By matching preset prefixes or identifying high-frequency enterprise identifiers, enterprise names can be extracted from OCR-recognized text or file names.

[0051] In this implementation, the company name is typically located at the top of the table page and is often preceded by prefixes such as "Compilation Unit:", "Accounting Unit:", or "Unit:". This module first matches a predefined prefix list in the first N lines (N=10) of the OCR-recognized text, extracting the subsequent content as candidate names. Simultaneously, to avoid misidentifying non-company fields such as "Unit: Yuan" as company names, a monetary unit keyword filtering mechanism is introduced. If no explicit prefix is ​​detected, further semantic heuristics are applied: if the text contains high-frequency company identifiers such as "company," "limited," or "technology," the text is directly adopted as the company name; otherwise, it reverts to extracting potential company information from the image filename (e.g., extracting "XX Company" when the filename is "XX Company_Balance Sheet.jpg"). Finally, punctuation cleaning and special symbol normalization are performed on the company name to improve the consistency of subsequent structured output.

[0052] Furthermore, a two-tiered strategy of direct keyword matching and semantic feature analysis based on feature keyword sets is adopted to identify the table type and whether the table body is a single entity or a merged entity.

[0053] In this implementation, a two-level classification strategy is used to identify the table type and the subject (single / merged): Level 1: Direct Keyword Matching Construct a mapping dictionary containing explicit report titles, covering standard names (such as "balance sheet") and their common misspellings or missing word variations (such as "consolidated profit statement" or "consolidated cash flow statement"). If any keyword exists in the OCR text, the report type and whether it is a consolidated report are directly determined.

[0054] Level 2: Semantic Feature Analysis If a direct match fails, a semantic discrimination mechanism based on technical terms is initiated. Three sets of feature keywords are predefined for the three core tables; some examples are shown below: Balance Sheet: Includes "Cash and Cash Equivalents", "Total Assets", "Total Liabilities", "Contract Liabilities", etc. Income statement: includes "operating profit" and "net profit", etc. Cash Flow Statement: Includes "Subtotal of Cash Inflows from Operating Activities" and "Cash," etc.

[0055] Meanwhile, to identify the "consolidation" attribute, consolidated financial statement-specific fields are introduced (such as "total comprehensive income attributable to owners of the parent company," "minority interests," and "dividends paid by subsidiaries to minority shareholders"). When a consolidation-specific keyword is detected, it is determined to be a consolidated financial statement; otherwise, it is a standalone financial statement. All text is cleaned before matching to improve the keyword accuracy rate.

[0056] This two-layer discrimination mechanism significantly improves robustness in scenarios involving missing titles, OCR recognition errors, or layout variations.

[0057] Furthermore, a multi-regular expression joint matching strategy is adopted to parse reporting period information in various formats from the header area and select the latest year-month combination.

[0058] In this implementation, the reporting period information is typically located in the table header area, and the format is highly heterogeneous. The date formats for the collected statistics tables include: Standard format: such as "December 2024" or "2024-12"; Range format: e.g., "January-December 2024"; Annual identifier: such as "2024" (default month is 12); Date strings: such as "20241231", "2024 / 12 / 31", "2024.12.01", etc.

[0059] This module employs a multi-regular expression matching strategy to parse the various formats mentioned above sequentially, extracting all valid year-month pairs to form a candidate set. Finally, the combination with the largest year and the largest corresponding month is selected as the reporting period (i.e., the latest reporting period). In actual business scenarios, the table typically displays the latest period-end data.

[0060] Furthermore, regular expressions are used to iterate and match OCR-recognized keywords related to monetary units in the text.

[0061] In this implementation, monetary units are generally displayed in the table header or footer in the form of "Unit: Yuan" or "(Thousand Yuan)". The module iterates through all OCR text, matching unit keywords using the regular expression `r'(Yuan|Thousand Yuan|Ten Thousand Yuan)'`, and prioritizes the first matching result. This design effectively avoids misjudgments caused by the mixing of numerical units within the table.

[0062] Furthermore, the semantic splicing module 104 for multi-column, multi-page complex tables is used to perform semantic-level fusion and data reorganization on complex tables with multiple columns or content split across pages, and output standardized table data with a unified structure.

[0063] Specifically, the semantic splicing module 104 for multi-column, multi-page complex tables includes: The multi-column table semantic splicing unit 109 is used to identify the repetitive column name pattern by column name frequency statistics. After determining that it is a multi-column table, it performs data column adaptation and splicing based on the adaptive column index matching algorithm.

[0064] In this implementation, the system first determines whether the current table is a multi-column table or a single-column table by parsing the table header data. This module only performs semantic concatenation for multi-column tables and outputs standard subject columns and their corresponding value columns.

[0065] A column name frequency statistical model is designed to count the frequency of occurrence of filtered core columns and determine whether there are duplicate column name patterns by using a preset threshold (frequency ≥ 2). Compared with traditional recognition methods based on character similarity, this technology directly focuses on the actual semantic repetition features of date columns, avoiding misjudgment problems such as "Amount in December 2024" and "Amount from January to December 2024" for similar column names.

[0066] A column number adaptation model is built based on an adaptive column index matching algorithm. For typical multi-column layouts such as 8 columns and 6 columns commonly found in actual table data, the column index selection rules are dynamically adjusted by matching core column keywords. For example, in an 8-column layout, if the 3rd column contains the core keyword, the index column [0,3,4,7] is selected; otherwise, the index column [0,2,4,6] is selected, thus achieving accurate adaptation of multi-column layouts.

[0067] When duplicate column names exist in the header row, it indicates that the current table is a multi-column table. The right column is appended to the left column to form the new table data, ensuring the uniqueness of the header row. Based on the selection of the target numerical columns, a binary structure of "subject column - numerical column" is formed. Simultaneously, standardized column names (such as "2024") are generated by combining the extracted table year information, achieving semantic fusion and format normalization of multi-column data.

[0068] Furthermore, the cross-page table semantic splicing unit 110 is used to locate cross-page data of the same table based on the multi-dimensional association index, perform order calibration by using the page number identifier in the file name, and verify the consistency of the number of columns of data on each page before merging.

[0069] In this implementation, to address the issue of splitting table content across pages, this module utilizes multi-dimensional relational indexes and data consistency verification technology to ensure the orderly concatenation and integrity of cross-page data. Core technologies include: Cross-page data association index construction: A data query index is built based on multi-dimensional data such as user_id, task_id, company_id, and slice_id to accurately locate cross-page split data in the same table and avoid confusion between data from different tables. The slice_id is pre-concatenated by the table type and report year to ensure the accuracy of the index.

[0070] Cross-page data order calibration mechanism: Extract page number identifiers (such as "page1" and "page2") from image file names, and achieve natural order calibration of cross-page table data through numerical sorting. This is because the current image data is obtained by splitting the original PDF document, and a "page" field will be automatically added to the original file name to indicate the page number of the current image in the original PDF. For abnormal scenarios without page number identifiers, log alarms will be output and the original data order will be preserved to ensure the orderliness of data splicing.

[0071] Data integrity verification: Before merging data, the column count of each page of the table is checked for consistency to prevent incorrect merging of heterogeneous tables. If the column count does not match, a log is output and the abnormal data is skipped to avoid overall merging failure due to errors in a single page of data, thus improving the module's fault tolerance.

[0072] Furthermore, the rule-based verification module 105 is used to perform real-time consistency verification and anomaly diagnosis on the standardized table data based on predefined verification rules.

[0073] Specifically, the rule-based verification module 105 is used for: The corresponding list of audit indicators and industry-standard formulas are dynamically loaded based on the table metadata.

[0074] In this implementation, based on table metadata and semantic features, such as report type, the system automatically and dynamically loads the corresponding list of audit indicators. This list integrates industry-standard formulas (e.g., "Total Assets = Total Current Assets + Total Non-current Assets"), supports online expansion and version management of the rule base, and ensures synchronization with the latest accounting standards.

[0075] Furthermore, the standard subject values ​​required for the formula are retrieved from the standardized table data, and the theoretical values ​​are calculated.

[0076] In this implementation, the system accurately retrieves the subject values ​​required for the cross-referencing formula from a constructed standardized data source. The structured tabular data source has a two-column format: the first column is the semantically aligned standard subject name, and the second column is the corresponding numerical value for the standard subject. Multi-level formula parsing is performed using code logic to generate theoretical calculation values ​​in real time, and these values ​​are compared with the values ​​identified in the table.

[0077] Furthermore, the theoretical values ​​are compared with the values ​​identified in the table, and anomaly diagnosis is initiated for inconsistent results that exceed the threshold, with feedback provided through a visual interface.

[0078] In this implementation, for inconsistent results (such as "the deviation between the identified value and the calculated value in the total asset table exceeds the threshold"), the module initiates a multi-dimensional anomaly assessment mechanism. Through an interactive visual interface, the system automatically marks the problem areas, guiding manual review and significantly reducing the misjudgment rate and manual review costs.

[0079] Please see Figure 3 , Figure 3 This is a flowchart of a method for intelligent identification, parsing, and cross-referencing of tabular data based on a large model. The invention also includes a method for intelligent identification, parsing, and cross-referencing of tabular data based on a large model, the method comprising: S1: Preprocess the input table image, including removing red stamps, tilt correction, and orientation calibration; S2: Perform high-precision character recognition and structured parsing on the preprocessed table image, and perform subject standardization and target year data extraction on the parsing results to obtain the original table data; S3: Extract key metadata information such as company name, table type, table body, reporting period, and monetary unit from the original table data; S4: Perform semantic-level fusion and data reorganization on complex tables with multiple columns or content split across pages to obtain standardized table data with a unified structure; S5: Based on predefined checklist rules, perform real-time consistency verification and anomaly diagnosis on the standardized table data.

[0080] The above-disclosed embodiments are merely one or more preferred embodiments of this application and should not be construed as limiting the scope of this application. Those skilled in the art can understand that all or part of the processes for implementing the above embodiments and equivalent changes made in accordance with the claims of this application still fall within the scope of this application.

Claims

1. A system for intelligent recognition, parsing, and cross-referencing tabular data based on a large model, characterized in that: It includes an image preprocessing module, a table parsing and recognition module, a key basic information extraction module, a multi-column cross-page complex table semantic splicing module, and a rule-based verification module, which are connected in sequence. The image preprocessing module is used to perform red stamp removal, tilt correction, and orientation calibration operations on the input table image; The table parsing and recognition module is used to perform high-precision text recognition and structured parsing on the preprocessed table image, and to perform subject standardization and target year data extraction on the parsing results to obtain the original table data; The key basic information extraction module is used to extract key metadata information such as enterprise name, table type, table body, reporting period, and monetary unit from the output of the table parsing and recognition module. The semantic splicing module for complex multi-column cross-page tables is used to perform semantic-level fusion and data reorganization on complex tables with multiple columns or content split across pages, and output standardized table data with a unified structure. The rule-based verification module is used to perform real-time consistency verification and anomaly diagnosis on the standardized table data based on predefined verification rules.

2. The intelligent identification, parsing, and cross-referencing system for tabular data based on a large model as described in claim 1, characterized in that, The image preprocessing module includes: The stamp-aware block processing unit is used to divide the input image into multiple sub-regions and use a dual-threshold collaborative stamp removal algorithm to process the sub-regions containing red stamps, thereby removing red stamps while protecting black text. The non-rigid tilt correction unit is used to determine the optimal correction angle by calculating the sum of squares of the first-order differences between adjacent columns of the vertical projection of the rotated binary image, and to perform rotation correction on the image. The table semantic orientation calibration unit is used to employ a dual-modal orientation decision mechanism, including OCR semantic verification and rule verification based on aspect ratio and keyword distribution, to ensure that the table orientation is positive.

3. The intelligent identification, parsing, and cross-referencing system for tabular data based on a large model as described in claim 1, characterized in that, The table parsing and recognition module is specifically used for: The locally deployed PaddleOCR-VL visual language model is used to recognize table images and output a Markdown-formatted structured text stream that preserves the topological relationship of the table rows and columns. For the subject column text in the structured text stream, subject standardization is performed by calculating its comprehensive similarity score with the subject names in the predefined standard subject library; Based on the timeliness requirements of table analysis, the data columns for the target year or target period are automatically extracted from the table header using either year-aware mode or relative period mode.

4. The intelligent identification, parsing, and cross-referencing system for tabular data based on a large model as described in claim 3, characterized in that, The overall similarity score is calculated using the following formula: ； in: , These represent the target words to be compared identified by the large model and the standard subject names in the standard subject library, respectively. Matching keywords; For fuzzy matching; For semantic similarity matching; , , These are the weighting coefficients.

5. The intelligent identification, parsing, and cross-referencing system for tabular data based on a large model as described in claim 1, characterized in that, The key basic information extraction module is specifically used for: By matching preset prefixes or recognizing high-frequency enterprise identifiers, enterprise names can be extracted from OCR-recognized text or file names; A two-level strategy, consisting of direct keyword matching and semantic feature analysis based on feature keyword sets, is used to identify table types and whether the table body is a single entity or a merged entity. A multi-regular expression joint matching strategy is adopted to parse reporting period information in various formats from the header area and select the latest year-month combination; OCR identifies monetary unit keywords in text by using regular expressions to iterate and match.

6. The intelligent identification, parsing, and cross-referencing system for tabular data based on a large model as described in claim 1, characterized in that, The semantic concatenation module for complex multi-column, cross-page tables includes: The multi-column table semantic splicing unit is used to identify the pattern of repeated column names by column name frequency statistics. After determining that it is a multi-column table, it performs data column adaptation and splicing based on an adaptive column index matching algorithm. The cross-page table semantic splicing unit is used to locate cross-page data in the same table based on multi-dimensional related indexes, perform order calibration by using the page number identifier in the file name, and verify the consistency of the number of columns in each page before merging.

7. The intelligent identification, parsing, and cross-referencing system for tabular data based on a large model as described in claim 1, characterized in that, The rule-based verification module is specifically used for: Dynamically load the corresponding list of audit indicators and industry-standardized formulas based on the table metadata; Retrieve the required standard subject values ​​for the formula from standardized table data and calculate the theoretical values; The theoretical value is compared with the value identified in the table. Anomaly diagnosis is initiated for inconsistent results that exceed the threshold, and feedback is provided through a visual interface.

8. A method for intelligent identification, parsing, and cross-referencing of tabular data based on a large model, applied to the system described in any one of claims 1-7, characterized in that, The method includes: The input table image is preprocessed, including removing red stamps, tilt correction, and orientation calibration. The preprocessed table image is subjected to high-precision character recognition and structured parsing, and the parsing results are standardized by subject and extracted from the target year data to obtain the original table data. Extract key metadata information such as company name, table type, table body, reporting period, and monetary unit from the original table data; Perform semantic-level fusion and data reorganization on complex tables with multiple columns or content split across pages to obtain standardized table data with a unified structure; Based on predefined checklist rules, the standardized table data is subjected to real-time consistency verification and anomaly diagnosis.

Images