Unstructured table parsing method and device, electronic equipment and storage medium

Abstract

The application relates to the technical field of data processing, can be applied to the field of financial technology / digital medical treatment, and discloses a non-structured table analysis method and device, an electronic device and a storage medium. The method comprises the following steps: obtaining a non-structured table file to be analyzed, extracting cell content data, row and column position data and cell attribute data, and generating table structure data; identifying a slice boundary according to the table structure data, and splitting the non-structured table file to generate data slices; reading a current row and an adjacent reference row in each data slice according to a row sequence to generate an inter-row structure continuity score; screening a candidate header row set according to the inter-row structure continuity score; performing semantic discrimination on the candidate header row and context row data to generate a target header row recognition result; identifying a data row range according to the target header row recognition result, and merging corresponding data rows to generate a structured data block and a structured analysis result. The method improves the accuracy and efficiency of table analysis.

Description

Technical Field

[0001] This application relates to the field of data processing technology and can be applied to the fields of fintech / digital healthcare, and in particular to a method, apparatus, electronic device and storage medium for parsing unstructured tables. Background Technology

[0002] With the widespread application of various business systems, a large amount of business data is stored and transferred in the form of spreadsheets. Spreadsheets are typically maintained long-term by different departments, personnel, or systems. Influenced by manual editing habits, display aesthetic requirements, and differences in business terminology, spreadsheets are often not strictly compiled according to database table structures, but rather exhibit features such as merged cells, multi-level headings, cross-row and cross-column descriptions, hidden rows and columns, intermediate calculation rows and columns, irregular empty rows and columns, and mixed data areas. For example, business statistics tables, fund detail tables, and risk analysis tables in the fintech field, as well as diagnosis and treatment statistics tables, expense summary tables, and departmental statistical reports in the healthcare field, may all have unstructured features such as merged cells, multi-level headings, hidden rows and columns, and mixed data areas. Traditional spreadsheet data extraction tools typically rely on preset templates, fixed row and column positions, or manual configuration of field relationships to process spreadsheet data. When table layouts change, header levels are inconsistent, data area boundaries are unclear, or the same file contains multiple business data blocks, traditional table data extraction tools are prone to problems such as header identification errors, inconsistent data row attribution, loss of field correspondences, and the miscollection of invalid data. In batch business scenarios, header identification errors, inconsistent data row attribution, and loss of field correspondences can also lead to inaccurate results in subsequent data retrieval, statistical analysis, and report generation. This requires a significant amount of manual verification, cleaning, and reconstruction, resulting in low data processing efficiency, high processing costs, and difficulty in adapting to the large-scale automated processing needs of different business domains and table formats. Summary of the Invention

[0003] The main technical problem addressed by the implementation method of this application is that traditional table parsing techniques suffer from inaccurate header identification, unclear data boundary delineation, and low batch processing efficiency when processing unstructured tables.

[0004] To solve the above-mentioned technical problems, the first technical solution adopted in the embodiments of this application is: providing a method for parsing unstructured tables, including: acquiring an unstructured table file to be parsed; extracting cell content data, row and column position data, and cell attribute data from the unstructured table file to generate table structure data; identifying slice boundaries in the unstructured table file based on the table structure data; segmenting the unstructured table file based on the slice boundaries to generate at least one data slice; reading the current row and adjacent reference rows in each data slice according to row order, and acquiring the cell content data, row and column position data, and cell attribute data corresponding to the current row and the adjacent reference rows, respectively; based on the current row and adjacent reference rows, the method further analyzes the data structure data to generate table structure data. The cell content data, row and column position data, and cell attribute data corresponding to the current row and the adjacent reference row are used to generate an inter-row structural continuity score between the current row and the adjacent reference row. Based on the inter-row structural continuity score, candidate header rows are selected from the data slices, and a candidate header row set is generated. Semantic discrimination is performed on the candidate header rows in the candidate header row set and the context row data corresponding to the candidate header rows to generate a target header row identification result. Based on the target header row identification result, the data row range corresponding to each target header row is identified, and the data rows within the data row range are merged into the corresponding target header row to generate a structured data block, thus obtaining the structured parsing result corresponding to the unstructured table file.

[0005] To solve the above-mentioned technical problems, the second technical solution adopted in the embodiments of this application is: providing an unstructured table parsing device, comprising: a table structure generation module, used to acquire an unstructured table file to be parsed, extract cell content data, row and column position data, and cell attribute data from the unstructured table file to generate table structure data; a data slice generation module, used to identify slice boundaries in the unstructured table file according to the table structure data, and to segment the unstructured table file based on the slice boundaries to generate at least one data slice; a row-level data reading module, used to read the current row and adjacent reference rows in each data slice according to row order, and acquire the cell content data, row and column position data, and cell attribute data corresponding to the current row and the adjacent reference rows, respectively; and a continuous scoring generation module, used to... Based on the cell content data, row and column position data, and cell attribute data corresponding to the current row and the adjacent reference row, respectively, an inter-row structural continuity score is generated between the current row and the adjacent reference row. A candidate header filtering module is used to filter candidate header rows from the data slice according to the inter-row structural continuity score and generate a candidate header row set. A target header recognition module is used to perform semantic discrimination on the candidate header rows in the candidate header row set and the context row data corresponding to the candidate header rows, generating a target header row recognition result. A data merging generation module is used to identify the data row range corresponding to each target header row according to the target header row recognition result, merge the data rows within the data row range into the corresponding target header row, generate a structured data block, and obtain the structured parsing result corresponding to the unstructured table file.

[0006] To solve the above-mentioned technical problems, the third technical solution adopted in the embodiments of this application is: to provide an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the unstructured table parsing method as described above.

[0007] To solve the above-mentioned technical problems, the fourth technical solution adopted in the embodiments of this application is: to provide a non-volatile computer-readable storage medium, wherein the non-volatile computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are executed by an electronic device, the electronic device performs the unstructured table parsing method as described above.

[0008] Unlike related technologies, this application first converts the unstructured table into table structure data containing cell content, row and column positions, and cell attributes. Then, it divides different logical regions into data slices based on slice boundaries, so that subsequent parsing does not directly deal with the entire complex table. Furthermore, by scoring the inter-row structural continuity between the current row and adjacent reference rows, candidate header rows that may undergo structural changes are first screened out. Then, semantic discrimination is performed in conjunction with context row data, thereby avoiding high-cost judgment of all rows and reducing the risk of misidentification caused by relying solely on fixed rules. Finally, based on the target header row identification results, the corresponding data rows are merged to generate structured data blocks, so that the header relationship and data ownership relationship in the unstructured table are clearly established, improving the accuracy of header identification, the stability of data block division, and the efficiency of batch parsing. Attached Figure Description

[0009] One or more embodiments are illustrated by way of example with reference to the accompanying drawings. These illustrations do not constitute a limitation on the embodiments. Elements having the same reference numerals in the drawings are denoted as similar elements. Unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0010] Figure 1 This is a schematic diagram of the operating environment of the unstructured table parsing method provided in the embodiments of this application.

[0011] Figure 2 This is a schematic diagram of the execution flow of the unstructured table parsing method provided in the embodiments of this application.

[0012] Figure 3 This is a schematic diagram of the execution flow for generating inter-row structure continuity scores in the unstructured table parsing method provided in this application embodiment.

[0013] Figure 4 This is a schematic diagram of the execution flow of weighted fusion in the unstructured table parsing method provided in the embodiments of this application.

[0014] Figure 5 This is a schematic diagram of the system structure of the unstructured table parsing device provided in the embodiments of this application.

[0015] Figure 6 This is a schematic diagram of the hardware structure of an electronic device that performs an unstructured table parsing method according to an embodiment of this application. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. Software tools, components, or servers not belonging to this company that appear in the embodiments of this application are merely illustrative examples and do not represent actual use.

[0017] It should be noted that, unless otherwise specified, the various features in the embodiments of this application can be combined with each other, all of which are within the protection scope of this application. Furthermore, although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device schematic diagram or the order in the flowchart.

[0018] Unless otherwise defined, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The term "and / or" as used in this specification includes any and all combinations of one or more of the associated listed items.

[0019] To facilitate understanding of this embodiment, a detailed description of an unstructured table parsing method disclosed in this application will be provided first. Please refer to [link to relevant documentation]. Figure 1 , Figure 1 This is a schematic diagram of the operating environment for the unstructured table parsing method provided in the embodiments of this application, such as... Figure 1 As shown, the unstructured table parsing method provided in this application is generally executed by an electronic device with a certain computing capability, such as a computer. In some possible implementations, this unstructured table parsing method can be implemented by a processor calling computer-readable instructions stored in memory. Figure 1 The computer equipment mentioned can be a server. A server can be a standalone server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. This can be understood as... Figure 1 The number of computer devices shown is merely illustrative and can be expanded in any number according to actual needs.

[0020] Please continue reading. Figure 2 , Figure 2 This is a schematic diagram illustrating the execution flow of the unstructured table parsing method provided in the embodiments of this application, such as... Figure 2 As shown, it includes the following steps: S1. Obtain the unstructured table file to be parsed, and extract the cell content data, row and column position data, and cell attribute data from the unstructured table file to generate table structure data.

[0021] The table structure data forms the data foundation after unifying the expression of cell content, spatial location, and display attributes in unstructured tables. Cell content data includes text data, numerical data, and null values ​​converted from table cells. Row and column position data includes the cell's row and column, merged areas, and adjacent position relationships. Cell attribute data includes formatting information such as font, color, size, and cell fill. In fintech scenarios, business statistics tables, fund detail tables, or risk analysis tables may have multi-level headers, merged cells, and hidden calculated columns. The above processing can convert manually edited table content into a calculable data representation. In digital healthcare scenarios, treatment statistics tables, cost summary tables, or departmental operation reports may have cross-row headers, blank separator columns, and data areas with different formats. The above processing can preserve table content, positional relationships, and format differences. This processing provides a unified data foundation for subsequent slice boundary recognition, inter-row structural continuity scoring, and header identification, improving the data usability and stability of the unstructured table parsing process.

[0022] As an optional implementation, the process of generating table structure data in step S2 above may also specifically include the following steps S11 to S14.

[0023] S11. Split and populate merged cells in an unstructured table file, and delete hidden rows, hidden columns, and blank rows.

[0024] In step S11, splitting merged cells refers to dividing a merged cell into multiple cells with independent row and column positions according to the row and column ranges corresponding to the merged cell. Filling the split cells with content means retaining the original content of the merged cell and writing it into the corresponding split cells, allowing the content in the merged area to be read according to cell position. Hidden rows and columns are typically used to store intermediate calculation data, auxiliary data, or data not directly displayed. Deleting hidden rows and columns prevents hidden content from participating in subsequent data slicing and header recognition. Empty rows are rows where none of the cells contain valid content; deleting empty rows reduces the impact of invalid rows on comparisons of adjacent rows. Empty columns used to separate different data areas are retained so that slice boundaries can be identified later based on consecutive empty columns. Through splitting merged cells, filling with content, deleting hidden rows and columns, and deleting empty rows, manually edited table formats can be converted into table formats that are easy to read row by row and compare column by column, providing a processing basis for subsequent cell content conversion, row and column position acquisition, and cell attribute extraction.

[0025] S12. Convert the data format in each cell to string format or numeric format to generate cell content data.

[0026] In step S12, the data in the cells may be stored in text, numerical, or other display formats. Data format conversion refers to reading the original content of each cell and converting it into string or numerical format based on the data characteristics of the original content. For cell content that can be identified as numerical, the cell content is converted into numerical format, enabling subsequent processing to perform calculations based on numerical precision, order of magnitude, and numerical variation. For content such as field names, item names, category names, numbers, and descriptive text, the cell content is converted into string format, enabling subsequent processing to compare based on character composition and text patterns. For cells without valid content, the corresponding empty value state can be retained. Cell content data is used to record the converted string content, numerical content, and empty value state. By converting the original content in different cells to a unified data format, the impact of differences in the original storage format on data type comparison, numerical distribution comparison, and text pattern comparison can be reduced, and calculable content data can be provided for subsequent inter-row structure continuity scoring.

[0027] S13. Obtain the row and column positions of each cell and generate row and column position data.

[0028] In step S13, row position indicates the row order of a cell in the unstructured table file, and column position indicates the column order of a cell. After splitting merged cells, deleting hidden rows and columns, and deleting blank rows, the row and column positions of each cell are recorded according to their row and column. Row and column position data can be saved according to row and column order and associated with corresponding cells. Row position identifies the adjacency relationship between the current row, the previous row, and the next row; column position identifies cells in the same column position in different rows and provides a positional basis for cell alignment between adjacent rows. Adjacent rows in an unstructured table may contain different numbers of valid cells, and may also have blank cells in some column positions. Row and column position data enables subsequent processing to establish cell correspondences based on actual column positions, avoiding field misalignment caused by comparing only according to the cell reading order, and providing positional data for data slicing, reading adjacent rows, and calculating the continuity of row structure.

[0029] S14. Extract the font type, font color, font size, cell size, and cell color of each cell to generate cell attribute data.

[0030] In step S14, cell attribute data is used to express the display format of each cell in the table. Cell attribute data includes font type, font color, font size, cell size, and cell color. Font type indicates the font category used for the cell content, font color indicates the display color of the cell content, font size indicates the font size of the cell content, cell size indicates the display size of the cell, and cell color indicates the cell's own color attribute. Each display attribute is extracted according to the row and column positions of each cell, and each display attribute is associated with the corresponding cell to generate cell attribute data. Corresponding cells in consecutive data rows typically have the same or similar font type, font color, font size, cell size, and cell color. Header rows, group header rows, or data area boundary rows may be distinguished from consecutive data rows by variations in font, font size, color, or cell size. Therefore, cell attribute data can express the structural differences between different cells at the display format level and provide format data for subsequent calculations of cell attribute continuity.

[0031] S15. Link the cell content data, row and column position data, and cell attribute data to generate table structure data.

[0032] In step S15, associating cell content data, row and column position data, and cell attribute data means establishing a correspondence between the string or numerical content, row and column positions, font type, font color, font size, cell size, and cell color of the same cell, using the cell as the association object. After completing the cell-level association, the data corresponding to each cell is organized according to row and column order, so that each cell simultaneously has content information, position information, and attribute information, generating table structure data. The table structure data does not only record the text or values ​​in the cells, but also records the cell content, the cell's position, and the cell's display format. The row and column position data in the table structure data is used to identify consecutive empty columns and establish cell correspondences between adjacent rows; the cell content data is used to calculate the continuity of content type, the continuity of numerical distribution, and the continuity of text pattern; and the cell attribute data is used to calculate the continuity of cell attributes. By associating these three types of data, a unified data input can be provided for subsequent slice boundary identification, inter-row structural continuity scoring, and candidate header row filtering.

[0033] As an example, in fintech scenarios, business statistics tables, fund detail tables, or risk indicator statistics tables may contain merged headers, hidden calculated rows and columns, blank rows, and field names in different formats. First, split and populate the cells corresponding to the merged headers, delete the hidden calculated rows and columns and blank rows, and then convert the field names, dates, amounts, ratios, and indicator values ​​to string or numeric formats. Generate row and column position data according to the row and column of each cell, and extract the font type, font color, font size, cell size, and cell color. After associating the content, position, and attributes of each cell, a table structure data that simultaneously expresses the fund field, numerical content, field position, and display format can be generated. In digital healthcare scenarios, treatment statistics tables, cost summary tables, or departmental statistics reports may contain cross-column headers, hidden calculated columns, blank rows, and category cells in different colors. After splitting merged cells, deleting hidden rows and columns, and deleting blank rows, convert the department name, project name, quantity, amount, and ratio to string or numeric formats, and record the row position, column position, and display attributes of each cell. By linking cell content data, row and column position data, and cell attribute data, a table structure data can be generated that simultaneously expresses diagnostic and treatment items, statistical values, cell position, and format differences, providing a data foundation for the segmentation of different data areas and the comparison of adjacent row structures.

[0034] Through steps S11 to S15, invalid data interference caused by hidden rows and columns and empty rows can be removed, and merged cells can be converted into cells that can be read and compared independently. By unifying the data format of cell content, calculable data can be provided for data type, numerical distribution, and text pattern analysis; by recording the row and column positions of cells, positional basis can be provided for slice boundary identification and adjacent row cell alignment; by extracting cell attributes, the display format differences between header rows and data rows can be preserved. By associating cell content data, row and column position data, and cell attribute data to generate table structure data, a unified and stable data foundation can be provided for subsequent data slicing, inter-row structure continuity scoring, and candidate header row selection, improving the data accuracy and processing stability of the unstructured table parsing process.

[0035] S2. Identify the slice boundaries in the unstructured table file based on the table structure data, and segment the unstructured table file based on the slice boundaries to generate at least one data slice.

[0036] In unstructured tables, slice boundaries are column-oriented separators used to distinguish different logical data areas. They are typically represented by consecutive empty columns, variations in the distribution of valid cells, and differences in column occupancy between adjacent data areas. For manually edited tables, multiple business tables, statistical areas, or explanatory areas may be arranged horizontally within the same worksheet. Parsing the entire worksheet directly can easily lead to the mixing of different data areas. In fintech scenarios, operational statistics tables may place a fund summary area, risk indicator area, and business detail area side-by-side within the same worksheet, with different areas usually visually separated by empty columns. In digital healthcare scenarios, departmental operation reports may record outpatient data, inpatient data, and expense data side-by-side, and blank separator columns may also exist between these areas. By identifying slice boundaries and creating data slices, subsequent header identification and data row attribution analysis can be performed within relatively independent data areas, reducing cross-regional misjudgments and improving the accuracy of region positioning in unstructured table parsing.

[0037] As an optional implementation, the process of generating data slices in step S2 above may also include steps S21 to S24.

[0038] S21. Based on the table structure data, count the number of valid cells, the number of empty cells, and the row span of valid cells in each column according to the column order, and generate a column occupancy status sequence.

[0039] In step S21, the column occupancy status sequence is a structured statistical result formed along the column direction. This step takes the table structure data as input, traverses the reconstructed cell matrix according to column order, and counts the number of cells containing valid content, the number of empty cells, and the row range covered by valid cells in each column. The number of valid cells reflects the extent to which the column carries actual content, the number of empty cells reflects the blank distribution of the column, and the row span of valid cells reflects the vertical coverage of the column's content. The above statistical results, arranged in column order, form the column occupancy status sequence, allowing for continuous analysis of the content distribution of the table along the column direction.

[0040] S22. Based on the column occupancy status sequence, calculate the empty column discrimination value corresponding to each column, and mark the columns whose empty column discrimination values ​​meet the preset empty column conditions as empty columns.

[0041] In step S22, the empty column discrimination value is a parameter used to quantify whether each column belongs to a blank delimiter column. This step calculates the blankness and content occupancy of the corresponding column based on the number of valid cells, the number of empty cells, and the valid cell row span in the column occupancy status sequence. When the blankness of a column reaches the preset empty column condition, and the valid content occupancy is lower than the corresponding condition, the column is marked as an empty column. The preset empty column condition may include a valid cell number threshold, an empty cell ratio threshold, and a valid row span threshold, so that the empty column marking process does not rely on a single empty value judgment.

[0042] S23. Merge consecutive empty columns according to column order to generate consecutive empty column intervals.

[0043] In step S23, a continuous empty column interval is a column-oriented interval composed of multiple adjacent empty columns. This step scans the columns marked as empty columns in column order. When multiple empty columns are consecutively distributed in column order, these empty columns are merged into the same continuous empty column interval. When there is a non-empty column between two empty columns, the current interval is terminated and a new interval identification begins. Through interval merging, scattered column-level empty column markings can be converted into interval data with a start column position and an end column position.

[0044] S24. Based on the continuous length of the empty columns in the continuous empty column interval, the distribution status of the effective columns on both sides of the continuous empty column interval, and the effective column span, select the target continuous empty column interval from the continuous empty column interval.

[0045] In step S24, the target continuous empty column interval is a candidate slice separation region obtained by further filtering from the continuous empty column interval. This step not only determines the continuous length of the empty columns in the continuous empty column interval itself, but also combines the distribution of effective columns on both sides of the continuous empty column interval and the effective column span for filtering. The distribution of effective columns on both sides can reflect whether there are data regions with actual content on the left and right sides of the empty column interval, and the effective column span can reflect the coverage of adjacent data regions in the row direction. With the above information, it is possible to distinguish between continuous empty column intervals that truly serve the function of data region separation and ordinary blank column intervals.

[0046] S25. Use the target continuous empty column interval as the slice boundary, and generate the data slice range based on the effective column range on both sides of the slice boundary.

[0047] In step S25, the data slice range is a column-oriented parsing range formed based on the slice boundary and the effective column ranges on both sides of the boundary. This step uses the selected target continuous empty column intervals as the separating boundary between different logical data regions, and generates the corresponding data slice range based on the start and end positions of the continuous effective columns to the left and right of each slice boundary. The data slice range includes the start column, the end column, and the corresponding column-oriented coverage area, providing a clear column coordinate basis for subsequent horizontal segmentation.

[0048] S26. Divide the unstructured table file horizontally according to the data slice range to generate at least one data slice.

[0049] In step S26, horizontal segmentation is a process of dividing the unstructured table file into columns according to the data slice range. This step extracts cell content data, row and column position data, and cell attribute data from the original table structure within the corresponding column range based on each data slice range, while maintaining the row order and column relative position of each cell in the original table. After horizontal segmentation, multiple logical data areas arranged side-by-side in the same worksheet are divided into at least one data slice, and each data slice serves as the object for subsequent inter-row structure continuity scoring and header identification.

[0050] As an example, in fintech scenarios, business analysis tables, funding statistics tables, or risk indicator tables may horizontally arrange multiple data areas within the same worksheet, visually separated by consecutive empty columns. For instance, the left side might contain summary funding data, the middle a blank separator column, and the right side risk indicator data. By analyzing the content occupancy status of each column, the proportion of blank columns, and the distribution of valid columns on both sides, different logical areas can be divided into independent data slices. In digital healthcare scenarios, treatment statistics tables, departmental expense tables, or medical quality analysis tables may also set up outpatient statistics, inpatient statistics, and expense statistics areas side by side, distinguishing different data content through blank columns. By identifying consecutive empty column intervals and the range of valid columns on both sides, multiple medical data areas in the same table can be split into independent slices, facilitating subsequent header identification and data row attribution analysis.

[0051] Through steps S21 to S26, the column-oriented blank distribution in the unstructured table can be converted into a calculable slice boundary basis. Combined with the effective column distribution status on both sides of the continuous empty column interval, the target continuous empty column interval with actual separation function can be identified. Thus, multiple logical data areas that are horizontally parallel in the same table are divided into relatively independent data slices, reducing mutual interference between different data areas and improving the accuracy of subsequent table header identification, row structure continuity scoring and data attribution classification.

[0052] S3. Read the current row and adjacent reference rows in each data slice according to the row order, and obtain the cell content data, row and column position data and cell attribute data corresponding to the current row and adjacent reference rows respectively.

[0053] In this context, the current row and adjacent reference rows are the fundamental comparison objects for inter-row structural continuity analysis. Adjacent reference rows are typically the row preceding the current row in the row sequence or other adjacent rows. Reading the current row and adjacent reference rows requires not only reading the text and numerical content within the cells but also simultaneously reading the row and column position relationships and cell format attributes, enabling comparisons between adjacent rows across multiple dimensions: content, position, and format. In fintech scenarios, continuous data rows in fund detail tables are usually consistent in field types, numerical distribution, and formatting styles, while header rows and data rows often exhibit variations in content type and display format. In digital healthcare scenarios, continuous data rows in treatment statistics tables may display repetitive structures such as department names, quantities, amounts, or proportions in the same column, while header rows or grouping rows typically have different text patterns and formatting. By obtaining multiple types of data from adjacent rows sequentially, a comparable data foundation can be provided for subsequent continuity scoring.

[0054] S4. Based on the cell content data, row and column position data, and cell attribute data corresponding to the current row and the adjacent reference row, respectively, generate a score for the inter-row structure continuity between the current row and the adjacent reference row.

[0055] The row structure continuity score is a data metric used to quantify whether the current row and its adjacent reference rows are structurally continuous within the table. The row structure continuity score does not simply compare the text content within cells; it is calculated comprehensively based on cell content type, numerical distribution, text pattern, row-column correspondence, and cell attributes. For continuous data rows, the current row and its adjacent reference rows typically have similar data types and text patterns in the same column positions. Numerical cells usually have similar numerical precision, orders of magnitude, or data formats, and attributes such as font type, font color, font size, cell size, and cell fill color generally remain relatively stable. For header rows, group header rows, or data range boundary rows, the current row and its adjacent reference rows typically exhibit structural changes in terms of content type, numerical distribution, text pattern, or cell attributes.

[0056] This embodiment first calculates the inter-row structural continuity score between the current row and its adjacent reference rows locally, and then filters candidate header rows based on the inter-row structural continuity score. A high inter-row structural continuity score indicates that the current row and its adjacent reference rows have a high degree of structural continuity, and the current row can be processed as a continuous data row; a low inter-row structural continuity score indicates that there may be a structural discontinuity between the current row and its adjacent reference rows, and the current row can be included in the candidate header row set. The subsequent semantic discrimination process only inputs the candidate header rows and their corresponding context row data into the large language model, without needing to input all data rows in the data slice into the large language model.

[0057] In fintech scenarios, continuous data rows in business statistics tables, fund detail tables, or risk analysis tables typically maintain a stable data structure in terms of date, amount, percentage, and indicator name columns, while field name rows, category header rows, or summary header rows often vary in text content and cell formatting. In digital healthcare scenarios, continuous data rows in treatment statistics tables, cost summary tables, or departmental statistics reports typically maintain a stable data structure in terms of department name, project name, quantity, amount, and percentage columns, while project header rows, cost category header rows, or statistical description rows often exhibit significant content and format differences from adjacent data rows. By calculating the inter-row structural continuity score, continuous data rows and rows that may undergo structural changes can be converted into comparable quantitative results, providing a data foundation for candidate header row selection and subsequent semantic judgment using large-scale language models.

[0058] As an alternative implementation method, please continue reading. Figure 3 , Figure 3This is a schematic diagram of the execution flow for generating inter-row structure continuity scores in the unstructured table parsing method provided in this application embodiment, as shown below. Figure 3 As shown, the process includes steps S41 to S44.

[0059] S41. Based on the row and column position data of the current row and the adjacent reference row, align the cells with corresponding column positions in the current row and the adjacent reference row.

[0060] In step S41, cell alignment refers to establishing a comparison relationship between cells in the current row and cells in adjacent reference rows within the same data slice, based on column position. Unstructured tables may contain merged cells, blank cells, missing column positions due to hidden row and column processing, and differences in the number of cells between different data rows. Directly comparing cells according to the order of cell reading may incorrectly match cells located in different column positions with different field meanings, thus affecting the inter-row structural continuity score.

[0061] This embodiment maps the current row and adjacent reference rows to the same column coordinate system based on their respective row and column position data, and establishes cell correspondence based on column number or column position. When both the current row and adjacent reference rows have valid cells at the same column position, the two valid cells are marked as valid aligned cells; when neither the current row nor adjacent reference rows have valid cells at the corresponding column position, the corresponding column position is marked as a missing aligned position; when a cell covers multiple column positions, a correspondence is established based on the range of columns covered by the cell. After column alignment, a unified cell comparison benchmark is formed between the current row and adjacent reference rows, enabling subsequent processing to calculate content type continuity, numerical distribution continuity, text pattern continuity, and cell attribute continuity around the same field position. Column alignment also reduces the impact of cell misalignment on the inter-row structural continuity score, reducing the possibility of consecutive data rows being incorrectly selected as candidate header rows due to column misalignment, thereby reducing the amount of data input into large language models for semantic discrimination.

[0062] S42. Based on the aligned cell content data, calculate the content type continuity, numerical distribution continuity, and text pattern continuity between the current row and the adjacent reference row.

[0063] In step S42, content type continuity, numerical distribution continuity, and text pattern continuity quantify the structural continuity between the current row and adjacent reference rows from different content dimensions. Content type continuity expresses whether the data types of the current row and adjacent reference rows are consistent in the same column position. Content types can include text, numeric, date, and null values. For each pair of valid aligned cells, the data types of the current row cells and adjacent reference row cells can be compared to see if they are the same. The content type continuity is calculated based on the ratio between the number of valid aligned cells with the same data type and the number of cells participating in the content type comparison. Numerical distribution continuity expresses whether the numerical data in the current row and adjacent reference rows belong to the same data structure. Numerical distribution continuity can be calculated based on the numerical precision, order of magnitude, numerical format, and variation range of the numerical content in the valid aligned cells. For example, when both the current row and adjacent reference rows contain monetary data with two decimal places in the same column position, the corresponding numerical cells have a high degree of numerical distribution continuity; when the cell in the current row is a field name, while the corresponding cell in the adjacent reference row is numerical data, the corresponding column position has a low degree of numerical distribution continuity. Text pattern continuity is used to express whether the text content in the current row and the adjacent reference row have similar structural patterns. Text patterns can be divided according to text length, character composition, date expression format, number expression format, field description format, and text content type. For example, numbers, names, or dates in consecutive data rows usually have similar character compositions and format structures, while field names in header rows usually have different text patterns than specific field values ​​in data rows.

[0064] By calculating content type continuity, numerical distribution continuity, and text pattern continuity separately, structural changes between the current row and adjacent reference rows can be identified at the cell content level, avoiding judgments based solely on whether cell text is identical. Content type continuity, numerical distribution continuity, and text pattern continuity can also serve as content feature criteria for locally filtering candidate header rows, preventing rows with high structural continuity from entering the semantic discrimination scope of large language models, allowing large language models to focus on processing candidate header rows with changes in content structure.

[0065] S43. Based on the aligned cell attribute data, calculate the continuity of cell attributes between the current row and the adjacent reference row.

[0066] In step S43, cell attribute continuity is a data metric used to quantify whether the current row and its adjacent reference rows maintain consistency in cell display attributes. Cell attribute data can include font type, font color, font size, cell size, and cell fill color. Consecutive data rows typically use the same or similar cell attributes in the same column position. For example, consecutive numerical cells in the same amount column typically use the same font, font size, and cell fill color. Header rows, group header rows, or data range starting rows are usually distinguished from data rows by bolding the font, adjusting the font size, changing the font color, changing the cell fill color, or changing the cell size. Therefore, changes in cell attributes can serve as a formatting feature to identify changes in table structure.

[0067] This embodiment compares the attribute data of effectively aligned cells in the current row and adjacent reference rows column by column based on the cell correspondence generated in step S41. For each group of effectively aligned cells, the matching results for font type, font color, font size, cell size, and cell fill color can be calculated separately. The matching results of each attribute are then weighted according to their respective weights to generate a cell attribute continuity score. A high cell attribute continuity score indicates a high degree of consistency between the current row and adjacent reference rows in terms of table display structure; a low cell attribute continuity score indicates changes in font, color, size, or fill attributes between the current row and adjacent reference rows. Cell attribute continuity, along with content type continuity, numerical distribution continuity, and text pattern continuity, participates in the inter-row structural continuity scoring, ensuring that candidate header row selection considers both content structure changes and format structure changes. By incorporating cell attribute continuity, the problem of not being able to identify formatted header rows based solely on text content can be reduced, and format feature basis can be provided for selecting candidate header rows that require further semantic discrimination using a large language model.

[0068] S44. Weighted fusion of content type continuity, numerical distribution continuity, text pattern continuity, and cell attribute continuity is performed to generate an inter-line structure continuity score.

[0069] In step S44, weighted fusion refers to converting content type continuity, numerical distribution continuity, text pattern continuity, and cell attribute continuity into a unified inter-row structural continuity score. The content structure and format structure in different unstructured tables may have different degrees of importance; using a single continuity score cannot fully express the structural continuity between the current row and adjacent reference rows.

[0070] This embodiment allows configuring corresponding feature weights for content type continuity, numerical distribution continuity, text pattern continuity, and cell attribute continuity. Content type continuity primarily reflects whether the data types in the same column are consistent; numerical distribution continuity primarily reflects whether the precision, order of magnitude, and format of numerical data are continuous; text pattern continuity primarily reflects whether the text structure and character composition are similar; and cell attribute continuity primarily reflects whether the font, color, size, and fill attributes are consistent. Based on the feature weights corresponding to each continuity, each continuity is multiplied by its corresponding feature weight, and the product results are accumulated to generate an inter-row structural continuity score between the current row and adjacent reference rows. A higher inter-row structural continuity score indicates that the current row and adjacent reference rows are closer in terms of content type, numerical distribution, text pattern, and cell attributes, and are more likely to belong to the same continuous data region. A lower inter-row structural continuity score indicates that there are more likely structural changes such as table headers, group titles, or data region boundaries between the current row and adjacent reference rows.

[0071] This embodiment performs local prediction based on inter-row structural continuity scoring. When the inter-row structural continuity score meets the discrimination criteria for consecutive data rows, the current row is processed as a consecutive data row; when the inter-row structural continuity score does not meet the discrimination criteria for consecutive data rows, the current row is filtered as a candidate header row. After processing each row in the data slice, only the candidate header row and its corresponding context row data are input into a large language model. The large language model then performs semantic discrimination on the candidate header row. Based on the textual semantics of the candidate header row and the semantic relationship between the candidate header row and its context row data, the large language model determines whether the candidate header row belongs to the true header row. Local prediction using inter-row structural continuity scoring avoids the large language model processing all data rows in the data slice row by row and reduces the number of calls to the large language model.

[0072] As an example, in fintech scenarios, consecutive detail rows in funding statistics tables or risk indicator tables typically maintain similar data types in the same columns. For instance, date, amount, percentage, and indicator name columns present stable text, numerical, or date formats, respectively. The header row, however, may contain field names, category headings, or bold fill styles. Therefore, by aligning columns and comparing content types, numerical precision, text format, and cell styles, it can be determined whether the current row and the previous row belong to the same data area. In digital healthcare scenarios, consecutive data rows in treatment statistics tables or department expense tables typically have a stable structure in columns such as department name, item name, quantity, amount, and percentage. However, department category headings, expense category headings, or statistical description rows may differ from adjacent data rows in text patterns, numerical distribution, and format attributes. Therefore, multidimensional continuity calculations can be used to identify structural changes between adjacent rows, providing a basis for subsequent candidate header row selection.

[0073] Through steps S41 to S44, based on column alignment, the structural continuity between the current row and adjacent reference rows can be quantified from multiple dimensions, including content type, numerical distribution, text pattern, and cell attributes, avoiding reliance solely on single text matching or fixed formatting rules for judgment. Therefore, it can more accurately distinguish between continuous data rows and candidate header rows, group header rows, or data region boundary rows, providing calculable scoring criteria for local filtering of candidate header rows and subsequent semantic discrimination by large-scale language models. This reduces the amount of data input to large-scale language models and improves the accuracy and processing efficiency of identifying complex row structures in unstructured tables.

[0074] As another example, the interline structural continuity score can be obtained by first weighting and fusing the continuity scores of each dimension to obtain an initial continuity score, which can be expressed as: S=Σ(w_k s_k) / (Σ w_k), k∈{type,num,text,fmt}, where S represents the initial continuity score, w_k represents the feature weight corresponding to the k-th continuity category, s_k represents the calculation result of the k-th continuity category, type represents content type continuity, num represents numerical distribution continuity, text represents text pattern continuity, and fmt represents cell attribute continuity. Content type continuity is generated based on the ratio between the number of data type matches in the corresponding column positions of the current row and adjacent reference rows and the number of comparable cells. Numerical distribution continuity is generated based on the ratio between the number of numerical precision, order of magnitude, or numerical format matches in the corresponding column positions and the number of comparable numerical cells. Text pattern continuity is generated based on the ratio between the number of text structure, character composition, or format pattern matches in the corresponding column positions and the number of comparable text cells. Cell attribute continuity is calculated by weighting the matching results of attributes such as font type, font color, font size, cell size, and cell fill color. The above method can transform the content differences, numerical differences, text pattern differences, and format differences between the current row and adjacent reference rows into a unified continuous scoring basis, and provide a calculation basis for subsequent validity identifier generation, validity weight normalization, and coverage ratio correction.

[0075] As another alternative implementation method, please continue reading. Figure 4 , Figure 4 This is a schematic diagram of the execution flow of weighted fusion in the unstructured table parsing method provided in the embodiments of this application, as shown below. Figure 4 As shown, the process includes steps S441 to S445.

[0076] S441. Obtain the feature weight table corresponding to the data slice. The feature weight table includes content type weight, numerical distribution weight, text pattern weight, and cell attribute weight.

[0077] In step S441, the feature weight table is weight configuration data that matches the structural features of the data slices, used to limit the participation of different continuity indicators in the inter-row structural continuity scoring. Content type weight corresponds to content type continuity, numerical distribution weight corresponds to numerical distribution continuity, text pattern weight corresponds to text pattern continuity, and cell attribute weight corresponds to cell attribute continuity. Since the table content in different data slices may have different data organization methods—for example, some data slices are mainly numerical data, while others are mainly text fields and categorical content—and the format attributes of some data slices are more valuable for distinguishing header rows from data rows, this step obtains the feature weight table corresponding to the data slices. This allows the subsequent scoring process to allocate the influence ratio of different continuity indicators according to the structural features of the current data slice, rather than applying fixed weights to all data slices.

[0078] S442. Based on the number of cells involved in alignment in the current row and adjacent reference rows, generate validity indicators corresponding to content type continuity, numerical distribution continuity, text pattern continuity, and cell attribute continuity, respectively.

[0079] In step S442, the validity label is data used to mark whether each continuity indicator has a reliable calculation basis. Because unstructured tables may contain blank cells, merged cells, missing columns due to hidden rows and columns, local data sparsity, and inconsistent column counts across different data regions, even though the current row and adjacent reference rows have achieved column alignment, not every type of continuity has a sufficient number of comparable cells. This step determines whether content type continuity, numerical distribution continuity, text pattern continuity, and cell attribute continuity meet the valid calculation conditions by counting the number of cells involved in alignment and considering the data basis upon which each type of continuity depends. For example, numerical distribution continuity requires comparable numerical cells, text pattern continuity requires comparable text cells, and cell attribute continuity requires comparable formatted data. By generating validity labels, continuity indicators lacking data support can be avoided from participating in subsequent scoring.

[0080] S443. Based on the validity identifier, select valid feature weights from content type weight, numerical distribution weight, text pattern weight, and cell attribute weight, and normalize the valid feature weights to generate normalized feature weights.

[0081] In step S443, the effective feature weights are weight data selected from the feature weight table based on the validity identifier. If the validity identifier corresponding to a certain continuity index indicates that the continuity does not meet the calculation conditions, then the weight corresponding to that continuity will not participate in the current inter-row structure continuity score calculation. Since the sum of the remaining weights may no longer meet the original weight sum condition after some weights are removed, it is necessary to normalize the effective feature weights. The normalization process involves recalculating the retained effective feature weights to a uniform weight scale, so that the continuity indices participating in the calculation maintain a relative proportional relationship within the effective weight range. This process can avoid the dilution of the score caused by invalid feature dimensions, and can also avoid the score being too low or too high due to directly deleting a certain weight.

[0082] S444. Based on the normalized feature weights, the effective continuity of content type continuity, numerical distribution continuity, text pattern continuity and cell attribute continuity are weighted and calculated to generate an initial continuity score.

[0083] In step S444, effective continuity refers to the continuity index that matches the validity identifier and meets the participation conditions. This step performs a weighted calculation of effective continuity based on normalized feature weights. Specifically, the continuity values ​​retained from content type continuity, numerical distribution continuity, text pattern continuity, and cell attribute continuity are multiplied by their corresponding normalized weights, and the products are summed to generate an initial continuity score. This initial continuity score reflects the structural similarity between the current row and adjacent reference rows in terms of effective feature dimensions. Compared to directly averaging all continuity values, this method dynamically adjusts the feature dimensions involved in the scoring based on the actual availability of the current row alignment data, making the scoring results more consistent with situations where there are partial missing values, inconsistent formats, and mixed data types in unstructured tables.

[0084] S445. Based on the effective cell coverage ratio between the current row and adjacent reference rows, correct the initial continuity score and generate an inter-row structural continuity score.

[0085] In step S445, the effective cell coverage ratio is a parameter used to describe the degree of comparable cell coverage between the current row and adjacent reference rows. This ratio is generated based on the proportional relationship between the number of cells participating in alignment and having valid content or attributes and the number of cells in the current row and adjacent reference rows that should be compared. If the effective cell coverage ratio is high, it indicates that there is a sufficient basis for comparison between the current row and adjacent reference rows, and the reliability of the initial continuity score is high. If the effective cell coverage ratio is low, even if some effective continuity values ​​are high, it may only mean that a small number of cells are locally similar, which cannot fully represent the overall structural continuity between the two rows. Therefore, this step corrects the initial continuity score based on the effective cell coverage ratio, so that the final generated inter-row structural continuity score reflects both the multi-dimensional continuity calculation results and the degree of comparable cell coverage, thereby improving the reliability of inter-row structure judgment in unstructured tables.

[0086] As an example, in fintech scenarios, some data slices in funding statistics tables or risk indicator tables may be dominated by numerical columns, with relatively high weighting for numerical distribution. Text pattern weighting and cell attribute weighting can serve as auxiliary judgment criteria. When only a few amount or percentage columns can be effectively aligned between a row and its adjacent rows, text pattern continuity lacking a basis for comparison can be excluded through validity indicators, and the remaining weights can be renormalized. The score can then be corrected by combining this with the effective cell coverage ratio, preventing the entire row from being misjudged as continuous data due to a few similar columns. In digital healthcare scenarios, department expense tables or treatment statistics tables may simultaneously contain fields such as department name, project name, quantity, amount, and percentage. The proportion of text and numerical columns varies in different data slices. If a slice has many comparable text cells, text pattern weighting can have a higher participation rate. If a row has many blank cells or merged header areas, the impact of local similarity features on the overall score can be reduced through validity indicators and coverage ratio correction, making the inter-row structural continuity score more consistent with the actual structure of the current data slice.

[0087] Through steps S441 to S445, validity judgment and coverage ratio correction are introduced into the multidimensional continuity weighted calculation process. This ensures that feature dimensions lacking a valid basis for comparison are excluded from scoring, and that the feature weights involved in the calculation are redistributed within an effective range. Therefore, the inter-row structural continuity score can simultaneously reflect the feature importance of the current data slice, the effective alignment between adjacent rows, and the multidimensional structural similarity. This reduces the interference of blank cells, merged areas, or missing local columns on the scoring results, improving the accuracy and stability of candidate header row selection.

[0088] S5. Based on the inter-row structure continuity score, filter candidate header rows from the data slices and generate a set of candidate header rows.

[0089] Candidate header rows are selected from data slices based on inter-row structural continuity scores, identifying potential structural start rows or field description rows. Generally, when the structural continuity between the current row and adjacent reference rows is low, the current row may not be a regular continuous data row, but rather a new header row, group header row, or data area start row. This process does not directly identify the header row but first filters out a small number of candidate rows requiring further judgment from all rows, thus avoiding complex semantic analysis for each row. In fintech scenarios, new field titles or summary titles may appear between different business segments in a funds summary table, and these rows will show significant structural changes compared to the previous data row. In digital healthcare scenarios, new header rows or classification description rows may also appear between different department categories or different cost categories in a treatment statistics table. By generating a set of candidate header rows, the scope of subsequent semantic judgment can be narrowed, improving the processing efficiency of the header recognition process.

[0090] As an optional implementation, the process of generating the candidate header row set in step S5 above may also specifically include the following steps S51 to S54.

[0091] S51. Obtain the continuity threshold corresponding to the data slice.

[0092] In step S51, the continuity discrimination threshold is a criterion used to distinguish between continuous data rows and suspected structural discontinuities. This threshold can be set corresponding to a data slice, allowing different data slices to adopt appropriate discrimination scales based on their own table structure complexity, column occupancy status, and row-level structural characteristics. For data slices with relatively regular structures, the inter-row structural continuity score between continuous data rows is usually high, and the continuity discrimination threshold can be set to a higher level. For data slices with many blank cells, merged areas, or format differences, the continuity discrimination threshold can be adjusted appropriately to avoid normal data rows being over-screened as candidate header rows.

[0093] S52. Compare the line structure continuity score and the continuity discrimination threshold.

[0094] In step S52, comparing the inter-row structural continuity score and the continuity discrimination threshold maps the structural continuity between the current row and its adjacent reference rows to an executable row type filtering result. The inter-row structural continuity score reflects the comprehensive continuity between the current row and its adjacent reference rows in dimensions such as content type, numerical distribution, text pattern, and cell attributes, while the continuity discrimination threshold provides a judgment benchmark. By comparing the two, it is possible to identify whether there are significant structural changes between the current row and its adjacent reference rows, thus providing a direct basis for subsequent candidate header row filtering.

[0095] S53. When the inter-row structure continuity score is less than the continuity discrimination threshold, the current row is marked as a candidate header row.

[0096] In step S53, when the inter-row structural continuity score is less than the continuity discrimination threshold, it indicates that the structural continuity between the current row and the adjacent reference row is insufficient. The current row may correspond to a new field description row, a group header row, or the starting row of a data area. In this embodiment, the current row is not directly identified as the target header row. Instead, it is first marked as a candidate header row, so that subsequent semantic discrimination can further identify the true header within the candidate range. This process can transform the structural score result into a candidate filtering result, while avoiding misidentification of the header due to relying solely on threshold judgment.

[0097] S54. Generate a set of candidate header rows according to the row order position of each candidate header row in the data slice.

[0098] In step S54, the candidate header row set is an ordered set formed according to the row order position of each candidate header row in the data slice. By preserving the row order position of the candidate header rows, the contextual relationship and sequential distribution of the candidate header rows in the data slice can be reflected. This set not only contains the candidate header rows themselves, but also serves as the data basis for subsequent acquisition of context row data, performance of semantic discrimination, and identification of the data row range corresponding to the target header row.

[0099] As an example, in fintech scenarios, funding statistics tables or risk indicator tables may contain multiple consecutive detail rows and several field header rows. Consecutive detail rows typically have high inter-row structural continuity scores, while field header rows, category header rows, or the starting row of a new data area differ significantly from the previous row in terms of content type, numerical distribution, text pattern, and format attributes. Therefore, rows with scores below the continuity threshold can be marked as candidate header rows, and a set of candidate header rows can be formed according to row order. In digital healthcare scenarios, treatment statistics tables or department expense tables may contain department category titles, item field rows, and consecutive statistical data rows. When the structural continuity between a row and the previous row is significantly reduced, that row can be retained as a candidate header row. This allows subsequent semantic judgment to be performed only around the set of candidate header rows, without requiring complex judgments on each row in the data slice.

[0100] Through the above steps S51 to S54, the suspected structural fault lines in the data slice can be initially screened based on the inter-row structural continuity score. The current line that may belong to the header line or the starting line of the data area is included in the candidate header line set, thereby narrowing the processing scope of subsequent semantic discrimination, reducing invalid judgments on continuous data lines, and improving the efficiency of header recognition and the accuracy of candidate screening.

[0101] S6. Perform semantic discrimination on the candidate header rows and their corresponding context row data in the candidate header row set to generate the target header row recognition result.

[0102] Semantic discrimination is a semantic analysis process that further identifies whether candidate header rows belong to the true header rows. The discrimination object includes not only the candidate header rows themselves but also the context row data corresponding to them. Context row data can include the previous and next rows of the candidate header row and the local structural information of the data slice it belongs to, enabling the semantic discrimination process to combine field names, data styles, and adjacent row relationships for judgment. In fintech scenarios, candidate rows may contain field text such as "date, amount, account, indicator name," or they may simply be business descriptions or totals, requiring consideration of preceding and following row data to determine whether they constitute a true header. In digital healthcare scenarios, candidate rows may contain field text such as "department, project, quantity, cost," or they may be statistical descriptions, category names, or remarks, again requiring differentiation based on context. Through semantic discrimination, true header rows can be further identified within the candidate range, reducing misjudgments caused by relying solely on structural scoring.

[0103] S7. Based on the target header row identification results, identify the data row range corresponding to each target header row, merge the data rows within the data row range into the corresponding target header row, generate structured data blocks, and obtain the structured parsing results corresponding to the unstructured table file.

[0104] Structured data blocks are structured parsing units formed by merging the corresponding data row ranges, based on the target header rows as field descriptions. After identifying the target header rows, the data row range corresponding to each target header row needs to be identified according to its row order position in the data slice. Then, the data rows within that range are merged under the corresponding target header row, transforming the manually formatted table content into data blocks with field relationships and data ownership relationships. In fintech scenarios, different business indicator areas in operational statistics tables can be formed into corresponding structured data blocks, facilitating subsequent retrieval, statistics, and analysis. In digital healthcare scenarios, different departments, projects, or cost categories in treatment statistics tables or cost summary tables can also form corresponding data blocks, facilitating subsequent summarization and querying. By generating structured data blocks, the correspondence between headers, fields, and data rows in unstructured tables can be established, improving the usability of parsing results and the accuracy of subsequent data processing.

[0105] As an optional implementation, the process of generating structured data blocks in step S7 above may also specifically include the following steps S71 to S75.

[0106] S71. Obtain the previous row data, the next row data, and the slice structure information of the data slice to which each candidate header row is located, and generate the context row data corresponding to the candidate header row.

[0107] In step S71, the context row data serves as the basis for local semantic judgment constructed around the candidate header row. The candidate header row itself may only contain field names, category names, or descriptive text. Judging solely based on this row's content can easily lead to misjudgments. Therefore, it is necessary to simultaneously obtain the previous row, the next row, and the slice structure information of the data slice containing the candidate header row. The previous row reflects whether there were consecutive data rows or the end state of the previous data region before the candidate header row; the next row reflects whether data content corresponding to the field name appears after the candidate header row; and the slice structure information reflects the number of columns, column occupancy status, empty column separation status, and row order position of the region containing the candidate header row. Through the combination of the above data, a local contextual expression corresponding to the candidate header row can be formed.

[0108] S72. Input the candidate header row, context row data, and slice structure information into the semantic discrimination model to generate the header semantic discrimination result corresponding to the candidate header row.

[0109] In step S72, the semantic discrimination model jointly analyzes the candidate header row, context row data, and slice structure information to generate the semantic discrimination result for the candidate header row. This semantic discrimination does not simply determine whether a specific keyword exists in the candidate header row, but rather combines the text content of the candidate header row, the data type changes between adjacent rows, field description relationships, and the structural position of the data slice. For example, whether the text in the candidate header row is closer to the field name, whether the next row contains data values ​​corresponding to the field names, and whether there is a clear structural break between the previous row and the candidate header row can all serve as criteria for semantic discrimination. The semantic discrimination result may include a discrimination identifier indicating whether it is a header row, or it may include corresponding discrimination confidence or semantic explanation information.

[0110] S73. Based on the semantic discrimination result of the table header, select the target table header row from the candidate table header row set and generate the target table header row recognition result.

[0111] In step S73, the target header row is the actual header row selected from the candidate header row set. Since the candidate header row set originates from the inter-row structure continuity scoring, it may contain group header rows, description rows, total rows, or other structurally varied rows. Therefore, further filtering based on the header semantic discrimination results is necessary. This step can retain candidate header rows that meet the header semantic conditions from the candidate header row set, and filter out candidate header rows that do not meet the header semantic conditions, based on the discrimination identifier, discrimination confidence level, or preset semantic filtering conditions, generating the target header row identification result. The target header row identification result may include the row order position of the target header row, its data slice identifier, and the corresponding field content.

[0112] S74. Identify the range of data rows corresponding to each target header row according to the row order position of each target header row in the data slice.

[0113] In step S74, the data row range is the data belonging interval identified based on the row order position of the target header row in the data slice. When multiple target header rows are located in the same data slice, the target header rows can be sorted according to their row order, and the row order interval between two adjacent target header rows can be used as the data row range corresponding to the previous target header row. When a target header row is the last target header row in the data slice, the continuous data rows from that target header row to the end position of the data slice can be identified as the corresponding data row range. This step divides the row data in the data slice into the belonging interval corresponding to each target header row based on the positional relationship of the target header rows.

[0114] S75. Merge the data rows within the data row range into the corresponding target header row to generate a structured data block.

[0115] In step S75, a structured data block is a data parsing unit formed by establishing a relationship between the target header row and the corresponding data row range. This step merges the data rows within the data row range into the corresponding target header row, establishing a column-to-column correspondence between the field content in the target header row and the cell content in the corresponding data row. During the merging process, the row order, column position, field names, and cell content of the data rows can be preserved, transforming the table area that originally relied on manual typesetting into a data block with field relationships and data ownership relationships. Through this processing, the target header row is no longer just a single identified row, but participates in data organization as the field basis of the structured data block.

[0116] As an example, in fintech scenarios, financial statistics tables or risk indicator tables may contain a mix of field rows, category description rows, and continuous data rows. A candidate header row might contain field-based text such as "date, amount, indicator name, percentage," or it might simply be descriptive content for a certain type of business data. Therefore, semantic discrimination can be performed by combining the data in the previous and next rows of the candidate header row with the structural information of the data slice it belongs to, filtering out the true header row. Then, according to the row order of the true header row in the data slice, the subsequent corresponding data rows are merged under that true header row, forming data blocks with field correspondences. In digital healthcare scenarios, treatment statistics tables or departmental expense tables may contain field rows such as "department, item, quantity, cost, percentage," or they may also contain departmental category titles, statistical descriptions, or remarks rows. By combining the contextual data of the candidate header row and the slice structure information, the true header row and ordinary description rows can be distinguished. Then, the treatment data, expense data, or statistical data within the corresponding range are merged under the corresponding header row, forming structured data blocks.

[0117] Through steps S71 to S75, semantic discrimination can be further performed by combining contextual data and slice structure information on the structural screening of candidate header rows. The true header rows are identified from the candidate header row set, and the corresponding data row range is divided according to the row order position of the true header rows. Therefore, a clear attribution relationship can be established between header rows and their corresponding data rows, generating structured data blocks with field correspondences. This reduces the possibility of description rows, category rows, or total rows being misidentified as header rows, improving the accuracy and usability of unstructured table parsing results.

[0118] The unstructured table parsing method provided in this application first converts the cell content, row and column positions, and cell attributes of the unstructured table file into table structure data, giving the table content, which originally relied on manual typesetting, a computable data foundation. Then, based on the table structure data, it identifies slice boundaries, dividing different logical data areas arranged horizontally or in a mixed manner within the same table into relatively independent data slices, avoiding interference between different data areas during subsequent parsing. Subsequently, it compares the current row and adjacent reference rows according to row order, considering content, position, and attributes, and generates an inter-row structural continuity score, enabling the structural differences between continuous data rows and suspected header rows to be quantified. Furthermore, this application does not directly perform semantic discrimination on all rows; instead, it first filters candidate header rows based on the inter-row structural continuity score, and then performs semantic discrimination on the candidate header rows and their corresponding context rows. Therefore, it can limit the semantic discrimination scope to a small number of suspected structurally variable rows, reducing invalid judgments on continuous data rows, improving the efficiency of batch table parsing, and reducing processing costs. Simultaneously, the semantic discrimination process, combined with the context row data and slice structure information of the candidate header rows, can further distinguish between real header rows, category description rows, total rows, and ordinary data rows, reducing misjudgments caused by relying solely on fixed rules or a single threshold. Finally, based on the target header row identification result, the data row range is identified, and the corresponding data rows are merged under the target header row to generate structured data blocks, thus clearly establishing the field relationships and data ownership relationships in unstructured tables. Therefore, the embodiments of this application can improve the accuracy of header identification, data area division, and structured data generation in unstructured tables while ensuring parsing efficiency, and are suitable for automated parsing and processing of multi-format, multi-region, and complex layout tables.

[0119] Please continue reading. Figure 5 , Figure 5 This is a schematic diagram of the system structure of the unstructured table parsing device provided in the embodiments of this application, such as... Figure 5 As shown, the unstructured table parsing device 50 includes: a table structure generation module 51, a data slice generation module 52, a row-level data reading module 53, a continuous score generation module 54, a candidate header filtering module 55, a target header identification module 56, and a data merging generation module 57.

[0120] The table structure generation module 51 is used to acquire the unstructured table file to be parsed, extract cell content data, row and column position data, and cell attribute data from the unstructured table file, and generate table structure data; the data slice generation module 52 is used to identify slice boundaries in the unstructured table file according to the table structure data, and segment the unstructured table file based on the slice boundaries to generate at least one data slice; the row-level data reading module 53 is used to read the current row and adjacent reference rows in each data slice according to row order, and acquire the cell content data, row and column position data, and cell attribute data corresponding to the current row and the adjacent reference rows, respectively; the continuous scoring generation module 54 is used to generate data based on the cells corresponding to the current row and the adjacent reference rows, respectively. The data slice is processed using cell content data, row and column position data, and cell attribute data to generate a row structure continuity score between the current row and the adjacent reference row. The candidate header filtering module 55 is used to filter candidate header rows from the data slice based on the row structure continuity score and generate a candidate header row set. The target header recognition module 56 is used to perform semantic discrimination on the candidate header rows in the candidate header row set and the context row data corresponding to the candidate header rows to generate a target header row recognition result. The data merging generation module 57 is used to identify the data row range corresponding to each target header row based on the target header row recognition result, and merge the data rows within the data row range into the corresponding target header row to generate a structured data block, thereby obtaining the structured parsing result corresponding to the unstructured table file.

[0121] As an optional implementation, the table structure generation module 51 is specifically used to split and fill the content of merged cells in the unstructured table file, and delete hidden rows, hidden columns and blank rows; convert the data format in each cell into string format or numeric format to generate the cell content data; obtain the row position and column position of each cell to generate the row and column position data; extract the font type, font color, font size, cell size and cell color of each cell to generate the cell attribute data; and associate the cell content data, the row and column position data and the cell attribute data to generate the table structure data.

[0122] As an optional implementation, the data slicing generation module 52 is specifically used to generate a column occupancy status sequence by counting the number of valid cells, the number of empty cells, and the row span of valid cells in each column according to the column structure data; calculating the empty column discrimination value corresponding to each column based on the column occupancy status sequence, and marking columns whose empty column discrimination values ​​meet the preset empty column conditions as empty columns; merging consecutive columns marked as empty columns according to the column order to generate consecutive empty column intervals; selecting target consecutive empty column intervals from the consecutive empty column intervals according to the consecutive length of empty columns in the consecutive empty column intervals, the distribution status of valid columns on both sides of the consecutive empty column intervals, and the effective column span; using the target consecutive empty column intervals as the slice boundary, and generating a data slice range according to the effective column range on both sides of the slice boundary; and horizontally slicing the unstructured table file according to the data slice range to generate at least one data slice.

[0123] As an optional implementation, the continuous score generation module 54 is specifically used to align cells with corresponding column positions in the current row and the adjacent reference row according to the row and column position data of the current row and the adjacent reference row, respectively; calculate the content type continuity, numerical distribution continuity, and text pattern continuity between the current row and the adjacent reference row according to the aligned cell content data; calculate the cell attribute continuity between the current row and the adjacent reference row according to the aligned cell attribute data; and perform weighted fusion of the content type continuity, the numerical distribution continuity, the text pattern continuity, and the cell attribute continuity to generate the inter-row structural continuity score.

[0124] As an optional implementation, the continuous score generation module 54 is further configured to obtain a feature weight table corresponding to the data slice, the feature weight table including content type weight, numerical distribution weight, text pattern weight, and cell attribute weight; generate validity identifiers corresponding to the content type continuity, numerical distribution continuity, text pattern continuity, and cell attribute continuity based on the number of cells participating in alignment in the current row and the adjacent reference rows; filter valid feature weights from the content type weight, numerical distribution weight, text pattern weight, and cell attribute weight based on the validity identifiers, and normalize the valid feature weights to generate normalized feature weights; perform weighted calculation on the valid continuity corresponding to the content type continuity, numerical distribution continuity, text pattern continuity, and cell attribute continuity based on the normalized feature weights to generate an initial continuity score; and correct the initial continuity score based on the effective cell coverage ratio between the current row and the adjacent reference rows to generate the inter-row structure continuity score.

[0125] As an optional implementation, the candidate header filtering module 55 is specifically used to obtain the continuity discrimination threshold corresponding to the data slice; compare the inter-row structure continuity score with the continuity discrimination threshold; when the inter-row structure continuity score is less than the continuity discrimination threshold, mark the current row as a candidate header row; and generate the candidate header row set according to the row order position of each candidate header row in the data slice.

[0126] As an optional implementation, the data merging and generation module 57 is specifically used to obtain the previous row data, the next row data, and the slice structure information of the data slice to which each candidate header row belongs, and generate the context row data corresponding to the candidate header row; input the candidate header row, the context row data, and the slice structure information into the semantic discrimination model to generate the header semantic discrimination result corresponding to the candidate header row; filter the target header row from the candidate header row set according to the header semantic discrimination result, and generate the target header row identification result; identify the data row range corresponding to each target header row according to the row order position of each target header row in the data slice; merge the data rows within the data row range to the corresponding target header row to generate the structured data block.

[0127] It should be noted that the above-described unstructured table parsing device can execute the unstructured table parsing method provided in the embodiments of this application, and has the corresponding functional modules and beneficial effects of the method. Technical details not described in detail in the embodiments of the unstructured table parsing device can be found in the unstructured table parsing method provided in the embodiments of this application.

[0128] Figure 6 This is a schematic diagram of the hardware structure of an electronic device that performs an unstructured table parsing method according to an embodiment of this application, such as... Figure 6 As shown, the electronic device 600 includes: One or more processors 610 and memory 620, Figure 6 Take the 610 processor as an example.

[0129] The processor 610 and the memory 620 can be connected via a bus or other means. Figure 6 Taking the example of a connection between China and Israel via a bus.

[0130] The memory 620, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules, such as the program instructions / modules corresponding to the unstructured table parsing method in the embodiments of this application. The processor 610 executes various functional applications and data processing of the server by running the non-volatile software programs, instructions, and modules stored in the memory 620, thereby implementing the unstructured table parsing method in the above-described embodiments.

[0131] The memory 620 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the unstructured table parsing device. Furthermore, the memory 620 may include high-speed random access memory and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 620 may optionally include memory remotely located relative to the processor 610, and this remote memory may be connected to the unstructured table parsing device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0132] The one or more modules are stored in the memory 620. When executed by the one or more processors 610, they perform the unstructured table parsing method in any of the above method embodiments, for example, the method described above. Figure 2 Method steps S1 to S7, Figure 3 Method steps S41 to S44, Figure 4 Steps S441 to S445 in the method are implemented. Figure 5 The functions of modules 51-57 in the document.

[0133] The above-described product can perform the methods provided in the embodiments of this application, and has the corresponding functional modules and beneficial effects for performing the methods. Technical details not described in detail in this embodiment can be found in the methods provided in the embodiments of this application.

[0134] This application provides a non-volatile computer-readable storage medium storing computer-executable instructions that are executed by one or more processors, for example... Figure 6 One of the processors 610 can enable the one or more processors to execute the unstructured table parsing method in any of the above method embodiments, for example, to execute the method described above. Figure 2 Method steps S1 to S7, Figure 3 Method steps S41 to S44, Figure 4Steps S441 to S445 in the method are implemented. Figure 5 The functions of modules 51-57 in the document.

[0135] This application provides a computer program product, which includes a computer program stored on a non-volatile computer-readable storage medium. The computer program includes program instructions that, when executed by an electronic device, enable the electronic device to perform the unstructured table parsing method described in any of the above method embodiments, for example, to perform the method described above. Figure 2 Method steps S1 to S7, Figure 3 Method steps S41 to S44, Figure 4 Steps S441 to S445 in the method are implemented. Figure 5 The functions of modules 51-57 in the document.

[0136] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0137] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented using software and a general-purpose hardware platform, or of course, using hardware. Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0138] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0139] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and not to limit them; under the concept of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of this application as described above, which are not provided in detail for the sake of brevity; although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for parsing unstructured tables, characterized in that, include: Obtain the unstructured table file to be parsed, and extract the cell content data, row and column position data, and cell attribute data from the unstructured table file to generate table structure data; Identify the slice boundaries in the unstructured table file based on the table structure data, and segment the unstructured table file based on the slice boundaries to generate at least one data slice; Read the current row and adjacent reference rows in each data slice according to the row order, and obtain the cell content data, row and column position data and cell attribute data corresponding to the current row and the adjacent reference rows respectively; Based on the cell content data, row and column position data, and cell attribute data corresponding to the current row and the adjacent reference row, respectively, an inter-row structure continuity score is generated between the current row and the adjacent reference row; Based on the inter-row structure continuity score, candidate header rows are selected from the data slices, and a set of candidate header rows is generated. Semantic discrimination is performed on the candidate header rows in the candidate header row set and the context row data corresponding to the candidate header rows to generate the target header row identification result; Based on the target header row identification results, the data row range corresponding to each target header row is identified, and the data rows within the data row range are merged into the corresponding target header row to generate a structured data block, thereby obtaining the structured parsing result corresponding to the unstructured table file.

2. The unstructured table parsing method according to claim 1, characterized in that, The step of extracting cell content data, row and column position data, and cell attribute data from the unstructured table file to generate table structure data includes: The merged cells in the unstructured table file are split and populated with content, and hidden rows, hidden columns, and blank rows are deleted; Convert the data format in each cell to string or numeric format to generate the cell content data; Obtain the row and column positions of each cell, and generate the row and column position data; Extract the font type, font color, font size, cell size, and cell color of each cell to generate the cell attribute data; The table structure data is generated by associating the cell content data, the row and column position data, and the cell attribute data.

3. The unstructured table parsing method according to claim 1, characterized in that, The step of identifying slice boundaries in the unstructured table file based on the table structure data, and segmenting the unstructured table file based on the slice boundaries to generate at least one data slice includes: Based on the table structure data, the number of valid cells, the number of empty cells, and the row span of valid cells in each column are counted according to column order to generate a column occupancy status sequence. Based on the column occupancy status sequence, calculate the empty column discrimination value corresponding to each column, and mark the columns whose empty column discrimination values ​​meet the preset empty column conditions as empty columns; Merge consecutive empty columns according to column order to generate consecutive empty column intervals; Based on the continuous length of the empty columns in the continuous empty column interval, the distribution status of the effective columns on both sides of the continuous empty column interval, and the effective column span, a target continuous empty column interval is selected from the continuous empty column interval; The target continuous empty column interval is used as the slice boundary, and the data slice range is generated according to the effective column range on both sides of the slice boundary; The unstructured table file is horizontally divided according to the data slice range to generate at least one data slice.

4. The unstructured table parsing method according to claim 1, characterized in that, The step of generating a row structure continuity score between the current row and the adjacent reference row based on the cell content data, row and column position data, and cell attribute data corresponding to the current row and the adjacent reference row, respectively, includes: Based on the row and column position data corresponding to the current row and the adjacent reference row, the cells with corresponding column positions in the current row and the adjacent reference row are aligned; Based on the aligned cell content data, calculate the content type continuity, numerical distribution continuity, and text pattern continuity between the current row and the adjacent reference row; Calculate the cell attribute continuity between the current row and the adjacent reference row based on the aligned cell attribute data; The continuity of content type, the continuity of numerical distribution, the continuity of text pattern, and the continuity of cell attribute are weighted and fused to generate the interline structure continuity score.

5. The unstructured table parsing method according to claim 4, characterized in that, The step of weightedly fusing the content type continuity, the numerical distribution continuity, the text pattern continuity, and the cell attribute continuity to generate the interline structure continuity score includes: Obtain the feature weight table corresponding to the data slice, the feature weight table including content type weight, numerical distribution weight, text pattern weight and cell attribute weight; Based on the number of cells participating in alignment in the current row and the adjacent reference row, validity identifiers corresponding to the content type continuity, the numerical distribution continuity, the text pattern continuity, and the cell attribute continuity are generated respectively. Based on the validity identifier, valid feature weights are selected from the content type weight, the numerical distribution weight, the text pattern weight, and the cell attribute weight, and the valid feature weights are normalized to generate normalized feature weights. Based on the normalized feature weights, the effective continuity of the content type continuity, the numerical distribution continuity, the text pattern continuity, and the cell attribute continuity are weighted and calculated to generate an initial continuity score. The initial continuity score is corrected based on the effective cell coverage ratio between the current row and the adjacent reference row to generate the inter-row structural continuity score.

6. The unstructured table parsing method according to claim 1, characterized in that, The step of filtering candidate header rows from the data slice based on the inter-row structure continuity score and generating a candidate header row set includes: Obtain the continuity threshold corresponding to the data slice; Compare the line structure continuity score with the continuity discrimination threshold; When the inter-row structure continuity score is less than the continuity discrimination threshold, the current row is marked as a candidate header row; The candidate header row set is generated according to the row order position of each candidate header row in the data slice.

7. The unstructured table parsing method according to claim 1, characterized in that, The step of identifying the data row range corresponding to each target header row based on the target header row identification result, and merging the data rows within the data row range to the corresponding target header row to generate a structured data block includes: Obtain the previous row data, the next row data, and the slice structure information of the data slice to which each candidate header row is located, and generate the context row data corresponding to the candidate header row; Input the candidate header row, the context row data, and the slice structure information into the semantic discrimination model to generate the header semantic discrimination result corresponding to the candidate header row; Based on the semantic discrimination result of the header, the target header row is selected from the candidate header row set, and the target header row identification result is generated; Identify the range of data rows corresponding to each target header row according to the row order position of each target header row in the data slice; The data rows within the specified data row range are merged into the corresponding target header row to generate the structured data block.

8. An unstructured table parsing device, characterized in that, include: The table structure generation module is used to obtain the unstructured table file to be parsed, extract the cell content data, row and column position data and cell attribute data from the unstructured table file, and generate table structure data. The data slice generation module is used to identify slice boundaries in the unstructured table file based on the table structure data, and to slice the unstructured table file based on the slice boundaries to generate at least one data slice. The row-level data reading module is used to read the current row and adjacent reference rows in each data slice according to the row order, and to obtain the cell content data, row and column position data and cell attribute data corresponding to the current row and the adjacent reference rows, respectively; The continuous score generation module is used to generate an inter-row structural continuity score between the current row and the adjacent reference row based on the cell content data, row and column position data and cell attribute data corresponding to the current row and the adjacent reference row, respectively. The candidate header filtering module is used to filter candidate header rows from the data slice based on the inter-row structure continuity score and generate a set of candidate header rows. The target header recognition module is used to perform semantic discrimination on the candidate header rows in the candidate header row set and the context row data corresponding to the candidate header rows, and generate the target header row recognition result; The data merging and generation module is used to identify the data row range corresponding to each target header row based on the target header row identification result, merge the data rows within the data row range into the corresponding target header row, generate a structured data block, and obtain the structured parsing result corresponding to the unstructured table file.

9. An electronic device, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the unstructured table parsing method according to any one of claims 1-7.

10. A non-volatile computer-readable storage medium, characterized in that, The non-volatile computer-readable storage medium stores computer-executable instructions, which, when executed by an electronic device, cause the electronic device to perform the unstructured table parsing method according to any one of claims 1-7.

Images