Engineering document compliance intelligent review system based on multi-modal document visual analysis

Abstract

This invention pertains to visual inspection, specifically to an intelligent compliance review system for engineering documents based on multimodal document visual analysis. The system includes a page visual inspection component, a layout semantic graph construction component, a compliance constraint matching component, and an evidence location component. By performing multi-scale visual inspection on layout areas, table boundaries, figure frames, signature areas, and page number areas within engineering document pages, the system constructs a layout semantic graph containing file nodes, page nodes, layout area nodes, field nodes, and field value nodes, establishing spatial inclusion, adjacency, header mapping, and cross-page continuation relationships. Compliance rules are then converted into graph constraint templates, and the layout semantic graph performs field consistency, temporal order, and necessary node existence checks, outputting the page coordinates of abnormal nodes. This system maintains consistent expression of field sources, layout hierarchy, and cross-page relationships, reducing the difficulty of mismatched field matching and abnormal location in complex engineering documents.

Description

Technical Field

[0001] This invention pertains to visual inspection, specifically to an intelligent review system for the compliance of engineering documents based on multimodal document visual analysis. Background Technology

[0002] Compliance review of engineering documents typically targets documents such as construction application materials, acceptance records, drawing frame information, catalog attachments, approval forms, and archived files. In existing technologies, review systems often use page images or electronic document pages as input. They first perform preprocessing such as noise reduction, skew correction, binarization, and page segmentation, then use text recognition technology to extract text content such as titles, numbers, project names, dates, unit names, and approval opinions, storing the results in a field table. For documents with relatively fixed formats, the system presets the location ranges of the cover area, table area, signature area, and page number area. It then uses template coordinates or keyword anchors to find target fields and compares the field values ​​with preset compliance rules, such as checking if required fields are empty, if project names are consistent, if date formats meet requirements, and if attachment numbers are consecutive. This type of solution relies on text content and fixed template positions, and can handle engineering documents with stable layouts and good image quality, but it is insufficient for expressing the spatial allocation of complex tables, multi-page forms, and drawing frame information.

[0003] In further existing solutions, some systems introduce document layout analysis and visual inspection models to detect objects in title blocks, paragraph blocks, table boundaries, stamp areas, signature areas, and drawing blocks on the page, and then classify text blocks into the corresponding areas based on the location of the detection boxes. This type of solution typically divides the page into several visual areas, performs text recognition and field extraction within each area, and then applies the corresponding review rules according to the area category. For tabular documents, a common approach is to identify table lines, row and column boundaries, and cell positions, and infer the correspondence between table headers and field values ​​using row and column coordinates. For drawing documents, fields such as project name, drawing number, version number, design unit, and review section are located through the drawing frame area. This type of solution adds visual positioning capabilities compared to simple text recognition, but its output is mostly area boxes, text boxes, and field lists, lacking a unified data structure that can describe hierarchy, adjacency, table header mapping, reading order, and cross-page continuation.

[0004] Current compliance assessment processes often employ rule bases. Rule bases typically organize engineering document review requirements into rules for field completeness, field consistency, chronological order, attachment completeness, and format validation. During review, the appropriate rule is selected based on the document type, and the extracted fields are substituted into the rule conditions one by one. For cases where the same field appears repeatedly on different pages or in different documents, the system generally merges them based on identical field names, page types, or directory numbers. If multiple candidate fields exist in the identification results, a common practice is to use the candidate value with the highest confidence level, closest to the table header, or closest to the template position for judgment. This approach has low implementation costs, but lacks constraint propagation between the source of candidate fields, visual location, page structure, table hierarchy, and cross-page relationships. During rule execution, it is difficult to confirm whether the field value comes from the correct field, page, or attachment range. Therefore, when there are misaligned fields, broken table lines, scan skew, parallel appearance of fields with the same name, or cross-page continuation of attachments, the review results are easily affected by localized identification biases.

[0005] Therefore, existing compliance review technologies for engineering documents based on text recognition, template localization, or ordinary visual inspection lack semantic expression that maintains the hierarchical relationship of document layout, field spatial belonging, and cross-page association between page visual inspection results and compliance rule matching processes. This leads to field values ​​in complex engineering documents easily being detached from their respective layout areas, table cells, figure frames, or continuation pages and incorrectly included in the review process. The technical reason for this problem is that existing systems typically flatten visual inspection results into box coordinates and text fields, and express compliance rules as independent field conditions. There is a lack of structured connections between the two, including inclusion relationships, adjacency relationships, table header mapping relationships, cross-page continuation relationships, and rule hit evidence. When engineering documents contain multiple fields with the same name, table continuations across pages, figure frame fields coexisting with text references, or table of contents attachments being separated, the system cannot reliably maintain the correspondence between field sources and rule conditions, and it is difficult to trace abnormal conclusions back to complete visual evidence. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent review system for the compliance of engineering documents based on multimodal document visual analysis, which can solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The intelligent review system for engineering data compliance based on multimodal document visual parsing includes a page visual inspection component, a layout semantic graph construction component, a compliance constraint matching component, and an evidence location component. The page visual inspection component performs multi-scale visual inspection of the layout area, text block, table boundary, figure frame, signature column and page number area in the engineering data page image; The layout semantic graph construction component organizes the detection results into file nodes, page nodes, layout area nodes, field nodes, and field value nodes, and establishes spatial inclusion relationships, adjacency relationships, header mapping relationships, and cross-page continuation relationships to form a layout semantic graph. The compliance constraint matching component matches the compliance rules of the engineering data with the layout semantic graph; The evidence location component is a system that outputs the coordinate information of abnormal nodes on the page.

[0008] Preferably, the page visual detection component includes a page type determination process, which generates page type labels based on title visual blocks, table density, figure frame distribution, signature column position and page number style, and selects the corresponding detection scale set according to the page type labels; The detection scale set includes different receptive field detection paths for full page layout, partial table, column boundary and field text block, and the candidate boxes of each detection path are merged by reading order constraint and region mutual exclusion constraint. The merged candidate boxes retain the detection scale identifier, page coordinates and low confidence re-examination identifier.

[0009] Preferably, the layout semantic graph construction component assigns node attributes to the file nodes, page nodes, layout area nodes, field nodes, and field value nodes. The node attributes include page type, area category, text candidate, coordinate range, confidence level, and source detection path. The layout semantic graph construction component establishes edge attributes based on the vertical and horizontal projections, table line intersections, field title similarity, and visual block reading sequence. The edge attributes include at least containment edges, adjacent edges, same column edges, synonymous field edges, and cross-page continuation edges, and record the corresponding visual detection sources and edge generation basis.

[0010] Preferably, the compliance constraint matching component expresses compliance rules as a graph constraint template that includes target node type, limited edge type, field consistency condition, time order condition, and necessary node existence condition, and performs candidate subgraph retrieval in the layout semantic graph; For a target field with multiple candidate field value nodes, the compliance constraint matching component determines the field value nodes to participate in the matching based on regional priority, adjacent title relevance, cross-page reference relationship and field value format consistency, and records the mapping relationship between template nodes and candidate subgraph nodes as well as the source of field values.

[0011] Preferably, when the page visual inspection component identifies the table area and the frame bar, it first generates table orientation priors and frame position priors based on the page type label, and then performs joint detection on table lines, cell closing boundaries, merged cell empty boundaries and frame bar separator lines. For pages with tilted or broken lines, the page visual inspection component reconstructs cell boundaries based on the endpoint clustering relationship in the same row and column, and aligns and binds the reconstructed cell boundaries with text block candidate boxes. The binding result is written to the source attribute and table level attribute of the corresponding field node.

[0012] Preferably, the layout semantic graph construction component establishes a page sequence index for cross-page data, and the page sequence index includes page number recognition results, table of contents reference results, title continuity results, and table beginning and end continuation results; When adjacent pages have the same field title, adjacent table structure, or consecutive attachment numbers, the layout semantic graph construction component establishes cross-page continuation edges between the corresponding page nodes, uses the continuation field nodes to share field identifiers, uses cross-page field value nodes to belong to the same compliance inspection object, and records the source page range and directory entry number of the continuation edge.

[0013] Preferably, the compliance constraint matching component configures a rule trigger index, and the rule trigger index uses document type, page type, and key layout area category as index items; When a new or updated field value node is added to the layout semantic graph, the compliance constraint matching component only calls the graph constraint template that matches the index item, and writes the hit abnormal nodes, dependency edges and unmet conditions into the review evidence set. The review evidence set forms an associated record with rule identifiers, node identifiers and page coordinates. Multiple rules triggered by the same field are merged into a matching task according to the node sharing relationship.

[0014] Preferably, when the page visual detection component binds the text block candidate box to the cell boundary, it introduces text baseline direction, character arrangement density, cell white space ratio and candidate box edge overlap rate to form binding features, and merges multiple lines of text in the same cell into field value candidates according to the baseline projection order. For title text that spans multiple cells, the page visual detection component determines the coverage of the title text based on the empty boundaries of merged cells and the distribution of adjacent column fields, binds features and page type tags together to the field value node attributes, and associates them with the corresponding cell boundary identifiers.

[0015] Preferably, when configuring shared field identifiers, the layout semantic graph construction component first performs field name normalization on cross-page field nodes. The field name normalization generates normalization candidates based on the similarity of field title characters, the layout area category where the title is located, the set of adjacent fields on the same page, and the directory reference text. When multiple normalized candidates correspond to the same engineering data entity, the layout semantic graph construction component establishes entity merging edges and retains the original page coordinates and original text candidates for field value nodes on different pages. The entity merging edges are configured with merging source, conflict retention flag and entity category identifier.

[0016] Preferably, the evidence location component generates an abnormal evidence subgraph based on the review evidence set, and the abnormal evidence subgraph includes rule nodes, abnormal field nodes, dependency field nodes, page nodes, and coordinate frame nodes; The evidence location component sorts the cover field, table field, graph frame field, and cross-page attachment field involved in the same compliance rule according to the dependency edge order in the abnormal evidence subgraph, and configures a page thumbnail index, region zoom index, original detection path identifier, and graph constraint template mapping identifier for each coordinate frame node, and keeps the sorting result consistent with the node identifier of the abnormal evidence subgraph.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By employing an invention technology that combines page visual inspection, layout semantic graph construction, compliance constraint matching, and evidence location, the layout areas, text blocks, table boundaries, drawing frames, signature areas, and page number areas in engineering data pages are transformed into file nodes, page nodes, layout area nodes, field nodes, and field value nodes. These nodes form a unified layout semantic graph through spatial inclusion relationships, adjacency relationships, table header mapping relationships, and cross-page continuation relationships. Compliance rules are no longer matched solely with isolated field values, but rather with graph constraint templates that include target node types, boundary edge types, field consistency conditions, time sequence conditions, and necessary node existence conditions. Project name, document number, date, responsible unit, drawing frame fields, and attachment fields retain their page source, region source, table level, and cross-page continuation relationships during review. This reduces the number of fields from complex tables, drawing frames, and cross-page documents being misplaced in the judgment process, ensuring that abnormal nodes and dependent nodes are located within the same structure. This processing method places visual inspection results, field attribution relationships, and rule conditions within the same graph structure for constraint matching, resolving the misattribution problem caused by the lack of structured connection between visual inspection results and compliance rules, and providing review conclusions with traceable page coordinates and node mapping basis.

[0018] 2. Beyond the main effects mentioned above, the coordination between page type tags and the detection scale set enables different visual detection paths for covers, approval forms, tabular documents, drawing pages, and attachment directories using different receptive fields. After merging the candidate detection boxes using reading order constraints and region mutual exclusion constraints, the detection scale identifier and low-confidence re-inspection identifier are retained, facilitating subsequent field source tracing. In the processing of table areas and figure frames, direction priors, position priors, endpoint clustering relationships, and cell boundary reconstruction are introduced to ensure that incomplete boundaries caused by scanning tilt, broken lines, or merged cells can be included in field binding. For cross-page documents, a page sequence index is established using page number recognition results, table of contents reference results, title continuity results, and table beginning / end continuation results. Related field nodes are organized using shared field identifiers and entity merging edges. Rule-triggered indexes call graph constraint templates according to document type, page type, and key layout area categories. Multiple rules triggered by the same field are merged according to node sharing relationships, reducing node conflicts caused by duplicate matching. The abnormal evidence subgraph saves the review results according to rule nodes, abnormal field nodes, dependent field nodes, page nodes, and coordinate frame nodes, so that the page coordinates, region magnification index, original detection path, and template mapping identifier corresponding to the same compliant abnormality are consistent. Attached Figure Description

[0019] Figure 1 This is a flowchart of the overall process of an intelligent review system for engineering data compliance based on multimodal document visual analysis. Figure 2 Flowchart for page type identification and multi-scale visual inspection; Figure 3 Flowchart for visual inspection and field binding of table area and chart frame; Figure 4 Flowchart for constructing a layout semantic graph and merging cross-page fields; Figure 5 Flowchart for compliance constraint matching and abnormal evidence location. Detailed Implementation

[0020] refer to Figure 1In one embodiment, the intelligent review system for engineering document compliance based on multimodal document visual analysis performs visual inspection, layout semantic organization, compliance constraint matching, and abnormal evidence location on engineering document page images. The engineering document page images can originate from electronic document page rendering results, scanned page images, or image-based archived files. During processing, the system represents the page as a page object containing pixel coordinates, page number, document type candidates, and document ownership identifier. The system includes a page visual inspection component, a layout semantic graph construction component, a compliance constraint matching component, and an evidence location component. The page visual inspection component performs multi-scale visual inspection of layout areas, text blocks, table boundaries, figure frames, signature areas, and page number areas in the engineering document page images. The layout area includes a cover title area, document number area, body description area, table area, table of contents area, attachment index area, and drawing information area; text blocks include single-line text, multi-line text, field titles, field values, table header text, and footer text; table boundaries include outer borders, row lines, column lines, cell boundaries, and local break lines; drawing frame areas include project name, drawing number, version, design, review, and approval sections; signature / seal areas include candidate areas for seals, signatures, dates, and approval comments; page number areas include page numbers, attachment page numbers, table of contents page numbers, and footer numbers. When outputting each candidate detection box, the page visual inspection component records the coordinates of the top-left and bottom-right corners, area category, source detection path, and confidence information of the candidate box to prevent subsequent field extraction from deviating from the visual source.

[0021] In this embodiment, the page visual inspection component uses a combination of multi-scale candidate generation and category verification to form visual inspection results. The full-page scale is used to identify the main page layout and large area boundaries; the local scale is used to identify internal table columns, figure frame separators, and signature / signature positions; and the fine-grained scale is used to identify field text blocks, broken table lines, and page number characters. The overall confidence score of the candidate boxes is calculated using the following formula: ; in, It represents the overall confidence level of the candidate detection box, and its physical meaning is the reliability of the candidate box being retained and participating in the construction of the layout semantic graph; Indicates the confidence level of the regional category classification; This indicates the degree to which the candidate box edges fit the visual boundary of the page; This indicates the degree of matching between the text density or graphic texture density inside the candidate box and the target region. This indicates the degree of matching between the candidate box and the prior page layout. , , , Represents the corresponding weight, and satisfies + + + =1. For example, for a candidate box in a drawing frame, take... =0.35、 =0.25、 =0.25、 =0.15, if =0.88、 =0.76、 =0.80、 =0.70, then The candidate box is written into the detection result set, carrying its category and coordinates, and participates in subsequent graph structure generation. This embodiment retains the category, coordinates, source path, and confidence information of the page visual detection results, enabling subsequent compliance reviews to trace the source of fields and reducing misjudgments caused by fields being out of place in the layout.

[0022] In one embodiment, the layout semantic graph construction component receives candidate bounding boxes, text candidates, table boundaries, and page objects output by the page visual detection component, and organizes the project data into file nodes, page nodes, layout region nodes, field nodes, and field value nodes. File nodes record the document file identifier, document type candidates, page range within the file, and directory association status; page nodes record the page sequence number, page type, page number recognition result, and page image coordinate system; layout region nodes record the region category, region coordinates, region level, and source detection path; field nodes record the field title, field specification name, field location region, field level, and field role; and field value nodes record field content candidates, original text candidates, coordinate range, recognition confidence, and source detection box. The layout semantic graph construction component establishes edge connections based on the spatial position and semantic relationships between nodes, including spatial inclusion relationships, adjacency relationships, table header mapping relationships, and cross-page continuation relationships. Spatial containment relationships are used to indicate that a page contains a layout area, a layout area contains a field, and a field contains a field value; adjacency relationships are used to indicate the visually adjacent positions of field headers and field values; table header mapping relationships are used to indicate the hierarchical relationship between table headers, column headers, row headers, and cell content; and cross-page continuation relationships are used to indicate the continuation relationship of the same data object between adjacent pages, table of contents pages, and attachment pages.

[0023] In this embodiment, the compliance constraint matching component does not directly perform rule checks on the flat text list, but instead converts the compliance rules of the engineering documents into an executable graph constraint template. The graph constraint template consists of target node type, limiting edge type, field consistency condition, time sequence condition, and necessary node existence condition. The target node type is used to limit the types of nodes to be matched, such as project name field nodes, preparation date field nodes, approval date field nodes, drawing number field nodes, or attachment number field nodes; the limiting edge type is used to limit the target field to be in a specified layout area or a specified table column; the field consistency condition is used to compare the values ​​of the same field in the cover, approval table, and drawing frame column; the time sequence condition is used to compare the preparation date, review date, approval date, and archiving date; the necessary node existence condition is used to determine whether a specified column or a specified signature area exists on the document page. After the compliance constraint matching component outputs the abnormal node, the evidence location component reads the coordinate range of the abnormal node and its dependent nodes in the layout semantic graph, and outputs the page coordinate information, node identifier, rule identifier, and source detection path. This coordinate information can correspond to a rectangle in the page image, or it can correspond to multiple rectangles in a cross-page anomaly. This embodiment processes visual inspection results, layout hierarchy, field attribution, and compliance rules within the same graph structure, enabling anomalies in complex engineering data to be mapped to specific page areas and field sources. See Table 1 for an example of the layout semantic graph node and edge attribute configuration.

[0024] Table 1. Examples of Layout Semantic Graph Node and Edge Attribute Configuration

[0025] refer to Figure 2 In a preferred embodiment, the page visual inspection component configures a page type determination process. This process uses title visual blocks, table density, figure frame distribution, signature bar location, and page number style as input features to generate page type labels. Title visual blocks are determined by high-weight text areas at the top of the page or the header; table density is determined by row lines, column lines, cell closing boundaries, and text block arrangement rules; figure frame distribution is determined by fixed dividing lines near the page edges and the clustering of column blocks; signature bar location is determined by a combination of seal candidates, signature candidates, and approval opinion columns; and page number style is determined by the text pattern, position, and reference relationship with the table of contents in the page number area. Page type labels include cover page, table of contents page, approval form page, record form page, drawing page, attachment page, and general instruction page. The page type determination result does not replace subsequent inspections but provides constraints for the selection of the inspection scale set, making drawing pages focus more on the figure frame and drawing number columns, approval form pages focus more on the signature and date columns, and table of contents pages focus more on the correspondence between item numbers and page numbers.

[0026] The page type determination process can use the following formula to calculate the matching value for each page type: ; in, This represents the matching value for a specific page type, and its physical meaning is the degree to which the current page image belongs to that page type. This indicates the degree of matching between the title visual block and the distribution of this type of title; This indicates the degree to which the table density matches the table structure of this type; This indicates the degree of matching between the frame layout and the edge column structure of this type; This indicates the degree of matching between the signature / signature field position and the structure of this type of approval field; Indicates the degree of matching between the page number style and the page numbering rules for this type; to Assign weights to page types, and the sum of the weights is 1. For example, for an approval form page, take... =0.20、 =0.30、 =0.10、 =0.30、 =0.10, if =0.74、 =0.86、 =0.20、 =0.92、 =0.60, then =0.762. When this value is higher than the matching value of other page types, the page type tag is determined to be an approval form page, and it is included in the detection scale set corresponding to the approval form.

[0027] In this embodiment, the detection scale set includes full-page layout detection paths, partial table detection paths, column boundary detection paths, and field text block detection paths. The full-page layout detection path processes the main page area, header and footer, table of contents area, and drawing area; the partial table detection path processes the table outline, row and column boundaries, merged cells, and header area; the column boundary detection path processes the signature bar, drawing frame bar, and approval bar; and the field text block detection path processes the field title, field value, and page number characters. Candidate boxes output by each detection path are written into a candidate pool. Candidate boxes in the candidate pool are sorted according to reading order constraints, which are determined based on the top-to-bottom and left-to-right arrangement of page coordinates and the row and column order of the table. Candidate boxes in the candidate pool are also merged according to region mutual exclusion constraints, which prevent the same visual area from being repeatedly marked as both a text area and a table area, or the same column from being simultaneously marked as both a signature bar and a plain text area. The merged candidate boxes retain their respective detection scale identifier, page coordinates, and low-confidence re-inspection identifier. Low-confidence re-inspection flags are generated when there is a conflict between category confidence and layout priors. For example, a candidate box may be identified as a title block but not located at the page edge, or a candidate box may be identified as a signature block but does not contain a signature or date candidate. This preferred embodiment maintains a uniform data format for visual inspection results under different page formats by using page type-driven scale selection and candidate box merging, while preserving the sources of differentiated detections.

[0028] refer to Figure 4 In a preferred embodiment, the layout semantic graph construction component assigns node attributes to file nodes, page nodes, layout area nodes, field nodes, and field value nodes. Page node attributes include page type, page number candidates, page coordinate system, page rotation correction status, and image source identifier. Layout area node attributes include area category, coordinate range, area hierarchy, page to which the area belongs, and source detection path. Field node attributes include field title text, field specification name, field role, table hierarchy, layout area, and candidate compliance category. Field value node attributes include text candidate, coordinate range, recognition confidence, source detection box, original text fragment, and value format. The writing order of node attributes is consistent with the layout hierarchy: page nodes are the parent of layout area nodes, layout area nodes are the parent of field nodes, and field nodes are the parent of field value nodes. For fields within a frame, the field node also records the name of the frame sub-column; for fields within a table, the field node also records the row number, column number, table header path, and merged cell identifier.

[0029] The layout semantic graph construction component establishes edge attributes based on horizontal and vertical projections, table line intersections, field title similarity, and visual block reading sequence. Horizontal and vertical projections determine whether two visual blocks are in the same row or column; table line intersections determine whether a text block falls into a specific cell; field title similarity determines name variations of the same field on different pages or in different areas; and visual block reading sequence determines the reading order of field titles and values. Edge attributes include at least containing edges, adjacent edges, same-column edges, synonymous-field edges, and cross-page continuation edges. Containing edges record the parent node, child node, and containing area ratio; adjacent edges record the adjacent direction and spacing; same-column edges record column alignment and shared headers; synonymous-field edges record the source of field title similarity; and cross-page continuation edges record the page range and continuation basis. The reliability of edge generation can be calculated using the following formula: ; in, It indicates the reliability of edge connections, and its physical meaning is the degree to which the relationship between two nodes can be used for compliance constraint matching; This indicates the degree of overlap or alignment of two nodes in the longitudinal and transverse projections; Indicates the degree of matching between the node and the intersection of table lines or cell boundaries; Indicates the similarity of field title text; This indicates the degree of consistency in the order of two nodes within a visual reading sequence; to Weights are generated for the edges, and the sum of the weights is 1. For example, if the field title "Project Name" and the text value on the right are on the same line and there is no table separator in between, then... =0.30、 =0.25、 =0.25、 =0.20, if =0.90、 =0.72、 =0.96、 =0.88, then =0.866. This edge is marked as the adjacent edge from the field title to the field value, and the edge generation basis is written. This embodiment expresses the layout semantics through node attributes and edge attributes, so that the field value no longer exists only as text content, but carries page, area, table and reading order information.

[0030] refer to Figure 5In a preferred embodiment, the compliance constraint matching component expresses the compliance rules of engineering data as a graph constraint template. The graph constraint template includes a template identifier, applicable data type, target node type, limiting edge type, field consistency condition, time sequence condition, necessary node existence condition, abnormal output fields, and a set of evidence nodes. Taking project name consistency review as an example, the template applies to the cover page, approval form page, and drawing page. The target node type includes project name field nodes and project name field value nodes. The limiting edge type includes page containment edges, layout area containment edges, field adjacent edges, and synonymous field edges. The field consistency condition is used to constrain the consistency of standardized text for multiple project name field value nodes. Taking approval date sequence review as an example, the template applies to the approval form page. The target node type includes preparation date field value nodes, review date field value nodes, and approval date field value nodes. The limiting edge type includes containment edges and adjacent edges within the same approval form area. The time sequence condition is used to constrain the chronological relationship of date nodes. Taking the requirement field existence review as an example, the template is applicable to table data pages or drawing pages. The target node types include specified field nodes and field value nodes. The requirement node existence condition is used to determine whether a field is missing.

[0031] In this embodiment, the compliance constraint matching component performs candidate subgraph retrieval in the layout semantic graph. During candidate subgraph retrieval, the target node type is used to filter candidate nodes, the limiting edge type is used to filter the relationship paths between nodes, and field consistency conditions, time sequence conditions, and necessary node existence conditions are used to determine whether the candidate subgraph meets the review requirements. For a target field with multiple candidate field value nodes, the compliance constraint matching component determines the field value nodes participating in the matching based on regional priority, adjacent title relevance, cross-page reference relationships, and field value format consistency. Regional priority is jointly determined by the document type and page type; for example, the approval date in an approval form page preferentially comes from the approval column, and the drawing number in a drawing page preferentially comes from the drawing frame column. Adjacent title relevance is determined by the reliability of adjacent edges between field title nodes and field value nodes. Cross-page reference relationships are determined by table of contents entries, page numbers, and attachment numbers. Field value format consistency is used to determine whether the text value conforms to the format of the project number, date, drawing number, or attachment number.

[0032] The matching value of the candidate subgraph can be calculated using the following formula: ; in, This represents the matching value between the candidate subgraph and the graph constraint template, and its physical meaning is the degree to which the candidate subgraph is suitable for performing a certain compliance review; This indicates the degree of matching between the candidate node type and the template target node type; This indicates the degree of matching between the edge types of candidate nodes and the edge types specified by the template; Indicates the degree to which field value format, field consistency, or time sequence conditions are met; This indicates the degree of matching between the source region of the candidate field value and the region specified in the template; to Weights are matched to the graph template, and the sum of the weights is 1. For example, if the project name consistency template retrieves a candidate subgraph, then... =0.25、 =0.30、 =0.25、 =0.20, if =1.00、 =0.84、 =0.92、 =0.88, then =0.908. This candidate subgraph records the mapping relationship between template nodes and actual nodes, and saves the source of field values. This embodiment uses graph constraint template retrieval and candidate field value filtering to provide a clear structural basis for compliant matching in scenarios such as parallel fields with the same name, field position offsets, and cross-page references.

[0033] refer to Figure 3 In a preferred embodiment, when identifying table areas and title blocks, the page visual inspection component generates table orientation priors and title block position priors based on the page type label. The table orientation prior describes the main row and column directions, row height distribution, column width distribution, and header area position of the table on the current page. The title block position prior describes the position range of the title block near the page edge, the direction of column separators, and the structure of the drawing information area. For approval forms, the table orientation prior favors a uniform horizontal row and column distribution; for record forms, the table orientation prior allows for multi-layered headers and merged cells; for drawing pages, the title block position prior focuses on continuous columns near the page edge. Within the prior range, the page visual inspection component jointly detects table lines, cell closing boundaries, merged cell gap boundaries, and title block separators. The table line detection outputs horizontal and vertical line segments. The cell closing boundary is formed by the intersection of adjacent horizontal and vertical line segments. The empty boundary of merged cells is inferred from the area where line segments are missing but the text is continuous in the same row or column. The frame divider line is inferred from the local dense line segment area at the edge of the page.

[0034] In this embodiment, for pages with tilted or broken lines, the page visual inspection component reconstructs cell boundaries based on the endpoint clustering relationships within the same row and column. Endpoint clustering takes the line segment endpoint coordinates, line segment direction, and distance between adjacent endpoints as input, grouping broken endpoints belonging to the same row or column as the same candidate boundary. The endpoint distance can be calculated using the following formula: ; in, It represents the merging distance between the endpoints of two broken line segments, and its physical meaning is the quantified result of the distance between whether the two endpoints can be classified into the same table boundary; and These represent the horizontal coordinates of the endpoints to be compared in the page coordinate system; and These represent the vertical coordinates of the endpoints to be compared in the page coordinate system; and Indicates the horizontal and vertical normalized scales; and Indicates the direction angle of two line segments; This represents the conversion factor for directional differences. If the lateral coordinate difference between the endpoints of two transverse break segments is 4 and the longitudinal coordinate difference is 1, then take... =10、 =5、 =0.20, the angular difference between the two line segments is 0.05, then This value is used for endpoint clustering sorting; the smaller the value, the more suitable the two endpoints are to be reconstructed as the same boundary. The reconstructed cell boundaries are aligned and bound to the text block candidate boxes, and the binding result is written to the source attribute and table level attribute of the corresponding field node. This embodiment uses table direction prior, frame position prior, and endpoint clustering reconstruction to ensure that break lines, slanted lines, and merged cells can still form a table structure usable for field attribution. See Table 2 for an example of table boundary and field binding data configuration.

[0035] Table 2 Example of Data Configuration for Table Boundaries and Field Binding

[0036] In a preferred embodiment, when the page visual inspection component binds text block candidate boxes to cell boundaries, it introduces text baseline direction, character arrangement density, cell white space ratio, and candidate box edge overlap rate as binding features. The text baseline direction is obtained by fitting the center points of characters within the text block and is used to determine whether the text block and cell row direction are consistent. The character arrangement density is determined by the number of candidate characters per unit area and the character spacing, and is used to distinguish field titles, field values, and noise artifacts. The cell white space ratio is determined by the difference between the cell area and the area occupied by the text block, and is used to determine whether the text block falls within the main area of ​​the cell. The candidate box edge overlap rate is determined by the overlapping area between the text block candidate box and the cell boundary, and is used to determine whether the text block spans multiple cells. During binding, the system does not directly use the single condition of the text block center point falling into the cell, but rather comprehensively considers the above features to determine the relationship between the text block and the cell.

[0037] The binding value between a text block and a cell can be calculated using the following formula: ; in, This represents the binding value between the text block candidate box and the cell boundary, and its physical meaning is the degree to which the text block belongs to that cell; Indicates the degree of overlap between the text block candidate box and the cell boundary; Indicates the degree of consistency between the text baseline direction and the cell row direction; This indicates the degree of matching between character arrangement density and the field's text type; This indicates the degree of matching between the white space ratio within a cell and the text inclusion pattern; to To bind weights, the sum of the weights is 1. For example, if a text block is located inside a table cell, take... =0.40、 =0.20、 =0.25、 =0.15, if =0.94、 =0.88、 =0.82、 =0.70, then =0.862. When the same text block has bound values ​​to multiple cells, the system will take the cell with the highest bound value that is consistent with the reading order as the primary cell and record the other cells as cross-cell candidates.

[0038] In this embodiment, when multiple lines of text exist within the same cell, the page visual inspection component merges these lines into field value candidates according to the text baseline projection order. The baseline projection order is determined by the projection position of the baseline of each line of text in the vertical coordinate of the cell, with lines of text having smaller projection positions appearing first. If there is significant indentation between multiple lines of text, the system writes the indentation as hierarchical information into the field value node, but still retains the same cell identifier. For title text spanning multiple cells, the page visual inspection component determines the coverage of the title text based on the empty boundaries of the merged cells and the distribution of adjacent column fields. For example, when the header "Construction Unit's Opinion" spans multiple columns, and its adjacent columns below are "Responsible Person," "Date," and "Opinion Content," the system establishes a header mapping relationship between the header field node and the three column field nodes below. Binding features and page type tags are written together into the field value node attributes and associated with the corresponding cell boundary identifier. This embodiment, through multi-feature binding, baseline merging, and cross-cell title coverage determination, ensures that the cell affiliation and header hierarchy of field values ​​within the table can be stably recorded.

[0039] In a preferred embodiment, the layout semantic graph construction component establishes a page sequence index for cross-page documents. The page sequence index includes page number recognition results, table of contents reference results, title continuity results, and table beginning / end continuation results. Page number recognition results are formed by visual detection of page number areas and text candidate parsing, recording the correspondence between the displayed page number, attachment page number, table of contents page number, and page number within the document. Table of contents reference results are formed by the document name, attachment number, and page number range in the table of contents entries, recording the reference relationship between the table of contents entries and page nodes. Title continuity results are formed by the title text of adjacent pages, area category, and title position, used to determine whether the same document section continues on adjacent pages. Table beginning / end continuation results are formed by the column structure of tables on adjacent pages, table header name, last row boundary, and first row features on the first page, used to determine whether cross-page tables belong to the same document object. The page sequence index does not change the original page order, but rather adds structural relationships between page nodes for cross-page review.

[0040] In this embodiment, when adjacent pages have the same field title, adjacent table structure, or consecutive attachment numbers, the layout semantic graph construction component establishes cross-page continuation edges between the corresponding page nodes. The same field title is determined by the normalized text of the field title and the area where the title is located, avoiding misjudgment of identical words appearing accidentally in the main text as continuation fields. Adjacent table structures are determined by the number of columns, column width ratio, table header text, and the form of the last row separator. Consecutive attachment numbers are determined by the attachment numbers in the table of contents and the page number range. The cross-page continuation value can be calculated using the following formula: ; in, This indicates the cross-page continuation value between adjacent pages, and its physical meaning is the degree of continuity between two pages on the same project data object; Indicates the degree of page number continuity; Indicates the degree of reference matching between directory entries and page ranges; Indicates the degree of continuity between adjacent page titles or field titles; Indicates the degree of structural continuity between the end of a table on the previous page and the beginning of a table on the following page; to For cross-page continuation weighting, the weight sum is 1. For example, if two consecutive pages of an acceptance record have consecutive page numbers, identical table of contents references, the same title, and the same table column structure, then... =0.25、 =0.25、 =0.20、 =0.30, if =1.00、 =0.90、 =0.86、 =0.92, then =0.923. Based on this, the system establishes a cross-page continuation edge and writes the source page range and directory entry number of the continuation edge.

[0041] Furthermore, the layout semantic graph construction component configures the continued field nodes as shared field identifiers and configures cross-page field value nodes as belonging to the same compliance inspection object. Shared field identifiers indicate that similar fields on multiple pages should be treated as the same set of fields under the same document object for review. Cross-page field value nodes retain their original page coordinates, original text candidates, and source detection paths to avoid losing location information after cross-page merging. For example, when an acceptance record continues across two pages, the project name field value node on the previous page and the project name field value node on the next page share the same field identifier, but the two field value nodes still record their respective pages and coordinates; if the next page has different project names, the system retains conflicting field values ​​within the same compliance inspection object. This embodiment, through page sequence indexing, cross-page continuation edges, and shared field identifiers, enables directories, attachments, and cross-page tables to be processed within the same review object, reducing review bias caused by splitting cross-page fields into irrelevant fields.

[0042] In a preferred embodiment, the layout semantic graph construction component first normalizes the field names of cross-page field nodes when configuring shared field identifiers. Field name normalization generates normalization candidates based on field title character similarity, the layout area category where the title is located, the set of adjacent fields on the same page, and the directory reference text. Field title character similarity is used to handle name variations such as project name, engineering name, and engineering project name; the layout area category where the title is located is used to distinguish between cover title areas, approval form fields, figure frame fields, and ordinary body text descriptions; the set of adjacent fields on the same page is used to determine the semantic environment of the field; for example, if the construction unit, construction company, and supervision unit are all present next to the project name, the field is more likely to belong to basic engineering information; the directory reference text is used to associate the document names in the directory entries with the title fields in the attachment pages. The normalization candidates retain the original field titles and do not directly overwrite the original identification text.

[0043] The entity merge value between field nodes can be calculated using the following formula: ; in, This indicates the degree to which two field nodes belong to the same project data entity. In physical terms, it represents the reliability of cross-page or cross-region fields being merged under the same entity identifier. Indicates the similarity of characters in the field headings; Indicates the degree of consistency in the layout area category where the field is located; Indicates the degree of consistency between adjacent field sets on the same page; This indicates the degree of match between the directory reference text and the page containing the field; to This is the entity merging weight, and the sum of the weights is 1. For example, if we need to determine whether the "Project Name" field on the cover page and the "Item Name" field in the drawing frame column should be merged, then... =0.35、 =0.20、 =0.25、 =0.20, if =0.78、 =0.70、 =0.88、 =0.92, then =0.817. When multiple normalization candidates correspond to the same engineering data entity, the layout semantic graph construction component establishes entity merging edges.

[0044] In this embodiment, entity merging edges do not delete any original field value nodes. Field value nodes on different pages retain their original page coordinates and original text candidates. Entity merging edges are configured with merging sources, conflict retention flags, and entity category identifiers. Merging sources include similar field titles, matching regional categories, matching adjacent field sets, and matching directory references. Conflict retention flags are used to record multiple different field values ​​under the same entity. Entity category identifiers are used to mark field categories such as project name, document number, drawing number, date, unit name, or attachment number. If the project name field exists in the cover page, approval form page, and drawing page, the system establishes the same entity category identifier and includes the three field value nodes under the same entity merging edge. If one of the field values ​​differs, that field value is not overwritten but participates in compliance constraint matching as a conflict candidate. This embodiment, through field name normalization and entity merging edges, enables the unified organization of field name variations, layout differences, and cross-page reference relationships while preserving original evidence.

[0045] In a preferred embodiment, the compliance constraint matching component configures a rule-triggered index. The rule-triggered index uses document type, page type, and key layout area categories as index items. Document types may include approval forms, acceptance records, drawings, catalog attachments, and archive covers; page types may include cover pages, table of contents pages, approval form pages, record form pages, drawing pages, and attachment pages; key layout area categories may include title areas, table areas, drawing frame columns, signature columns, date columns, and attachment index areas. The rule-triggered index records the set of drawing constraint templates corresponding to each type of index item. For example, the drawing frame column of a drawing page triggers the project name consistency template, drawing number integrity template, version number consistency template, and approval column date template; the signature column of an approval form page triggers the responsible unit field existence template, signature column corresponding template, and date sequence template; the attachment index area of ​​a table of contents page triggers the attachment number consecutive template and table of contents page number corresponding template. The rule-triggered index generates an initial matching set after page parsing is complete, and generates an incremental matching set when field value nodes are added or updated.

[0046] In this embodiment, when a field value node is added or updated in the layout semantic graph, the compliance constraint matching component only calls the graph constraint template that matches the document type, page type, and key layout area category. New field value nodes can originate from text updates after low-confidence review or from field binding updates after table boundary reconstruction; updated field value nodes can originate from changes in the field set after field name normalization, cross-page continuation, or entity merging. The compliance constraint matching component writes the matched abnormal nodes, dependency edges, and unmet conditions into the review evidence set. Abnormal nodes are used to mark field value nodes or necessary field nodes that do not meet the template conditions; dependency edges are used to record the included edges, adjacent edges, synonymous field edges, table header mapping edges, or cross-page continuation edges that the abnormal judgment depends on; unmet conditions are used to record inconsistencies in fields, non-compliance in time order, missing necessary nodes, or fields whose source area does not meet the template limitations. The review evidence set forms associated records with rule identifiers, node identifiers, and page coordinates. Multiple rules triggered by the same field are merged into a single matching task according to node sharing relationships. This embodiment reduces the participation of irrelevant templates in the judgment through rule-triggered indexing and incremental matching, enabling the review evidence set to be updated around the changed field nodes. For an example of the configuration of the rule triggering index and evidence set, please refer to Table 3.

[0047] Table 3. Example of rule triggering index and evidence set configuration

[0048] In a preferred embodiment, the evidence localization component generates an anomaly evidence subgraph based on the review evidence set. The anomaly evidence subgraph includes rule nodes, anomaly field nodes, dependent field nodes, page nodes, and coordinate frame nodes. Rule nodes record the triggered graph constraint template identifiers and unmet conditions; anomaly field nodes record field nodes or field value nodes that do not meet the conditions; dependent field nodes record related fields involved in the judgment, such as other page fields used for consistency comparison or adjacent date fields used for determining time order; page nodes record the pages where the anomaly field and dependent field are located; and coordinate frame nodes record the rectangular position of the anomaly field or dependent field in the page coordinate system. The evidence localization component reads the rule identifiers, node identifiers, and page coordinates from the review evidence set and reconstructs the anomaly judgment path according to the dependent edges in the layout semantic graph. If the anomaly involves cover fields, approval form fields, and graph frame fields, the anomaly evidence subgraph simultaneously includes all three types of page nodes and corresponding coordinate frame nodes to avoid outputting only a single anomaly text.

[0049] In this embodiment, the evidence location component sorts the cover field, table field, graph frame field, and cross-page attachment field involved in the same compliance rule according to the dependency edge order in the abnormal evidence subgraph. The sorting criteria include page sequence index, layout area hierarchy, the position of field nodes in the reading sequence, and the node order in the graph constraint template. The cover field is usually located at the beginning of the sort, the table of contents field follows its table of contents reference order, the table field is arranged in row and column order, the graph frame field is arranged in the reading order within the graph frame, and the cross-page attachment field is arranged according to the page range. Each coordinate frame node is configured with a page thumbnail index, an area magnification index, an original detection path identifier, and a graph constraint template mapping identifier. The page thumbnail index is used to locate the page where the anomaly is located, the area magnification index is used to locate the local area around the anomaly coordinate frame, the original detection path identifier is used to trace back the origin of the anomaly field from full-page layout detection, local table detection, column boundary detection, or field text block detection, and the graph constraint template mapping identifier is used to trace back the target node of the anomaly field in the template. This embodiment maintains consistency between the compliance judgment conclusion, the visual detection source, and the page location information through the abnormal evidence subgraph and coordinate frame nodes.

[0050] In a preferred embodiment, the system performs a re-examination and graph structure update on low-confidence fields. Low-confidence fields can be triggered by a low-confidence re-examination flag generated by the page visual inspection component, or by the compliance constraint matching component after discovering a field source conflict during graph constraint template matching. During the re-examination, the system reads the page node, layout area node, adjacent field nodes, and candidate cell boundaries where the low-confidence field is located. Instead of re-parseing the entire document, the system re-extracts text blocks, line segments, and column boundaries within a local area surrounding the abnormal or conflicting field. After the re-examination results generate new field value candidates, the layout semantic graph construction component does not directly overwrite the original field value node. Instead, it establishes candidate replacement edges, which record the original field value node, the new field value node, the re-examination source, and the replacement conditions. In subsequent matching, the compliance constraint matching component selects field value nodes to participate in the matching based on the field source area, the relevance of adjacent titles, and the consistency of field value format. If neither the new field value nor the original field value meets the template source conditions, both candidates are retained as evidence of conflict.

[0051] In this embodiment, the re-examination area is determined based on the coordinate frame nodes and dependent edges in the abnormal evidence subgraph. If the anomaly originates from a table field, the re-examination area extends to the same cell and its header path; if the anomaly originates from a frame field, the re-examination area extends to the same frame sub-column and adjacent separator lines; if the anomaly originates from a cross-page field, the re-examination area extends to the page range associated with the shared field identifier. The field value nodes generated after re-examination continue to carry the original page coordinates, source detection path, and re-examination identifier, and trigger the incremental matching set in the rule trigger index. This optional embodiment, through local re-examination, candidate alternative edges, and incremental matching, enables low-confidence identification results to be reorganized while preserving the original evidence, avoiding a one-time identification bias that directly determines the compliance conclusion.

[0052] In a preferred embodiment, the system employs different field value format consistency processing methods for the date, number, and name fields. The date field records the original date text, normalized date value, and date source field in its field value node. The normalized date value is obtained by parsing year, month, and day text fragments. If the original text is missing or has different delimiters, the original text is retained and normalization candidates are generated. The number field records the prefix, main number, version identifier, and attachment number in its field value node to avoid confusion between page numbers, document numbers, and drawing numbers. The name field records the original name text, name candidates after removing whitespace, unit name candidates, and project entity candidates in its field value node. Field value format consistency does not form a separate review result but serves as a condition for matching drawing constraint templates and selecting candidate field values. When multiple candidate values ​​exist for the same field, the system uses the format consistency result in conjunction with the field source area, allowing fields with correct sources but abnormal formats to be retained as evidence of anomalies, while fields with incorrect sources but normal formats will not replace the target field in the review.

[0053] In this embodiment, the processing of the time sequence condition relies on the template mapping relationship between date field value nodes. The preparation date, review date, approval date, and archiving date are compared by comparing the target node order of the graph constraint template. The relationships between date field value nodes come from the containing edges, adjacent edges, and header mapping edges within the same approval table area. If a date field is missing, the system marks the necessary node existence condition as unmet; if a date field exists but originates from an incorrect field, the system marks the field source region as unmet; if a date field exists and its source region is correct, but the normalized date value order does not conform to the template condition, the system marks the time sequence condition as unmet. This optional embodiment, by synchronously recording the field value format and field source region, ensures that dates, numbers, and names retain both textual content and visual origin and field attribution during compliance matching.

[0054] In one embodiment, the overall system workflow is completed sequentially by page object generation, visual inspection, semantic graph construction, graph constraint matching, and evidence localization. During page object generation, each page of engineering data image is assigned a file ownership identifier, page number, and page coordinate system. During visual inspection, the page visual inspection component selects a set of inspection scales based on the page type discrimination result and outputs layout areas, text blocks, table boundaries, figure frame borders, signature borders, and page number areas. During semantic graph construction, the layout semantic graph construction component converts the inspection results into file nodes, page nodes, layout area nodes, field nodes, and field value nodes, and establishes edge connections based on inclusion relationships, adjacency relationships, table header mapping relationships, and cross-page continuation relationships. During graph constraint matching, the compliance constraint matching component calls graph constraint templates according to data type, page type, and key layout area categories, retrieves candidate subgraphs, and judges node, edge, and field value conditions. During evidence localization, the evidence localization component reads the review evidence set and generates anomaly evidence subgraphs, keeping the abnormal fields, dependent fields, and coordinate frames within the same structure.

[0055] In this embodiment, the core of the system's compliance review of engineering documents lies in converting the page visual inspection results from flat candidate boxes into a layout semantic graph, and then converting compliance rules from isolated field conditions into graph constraint templates. Project names, document numbers, approval dates, drawing numbers, version numbers, responsible units, and attachment numbers in the engineering documents exist as field nodes and field value nodes, with their page source, region source, table level, field title, and cross-page relationships expressed through edge objects. The compliance constraint matching component reads node attributes and edge attributes during the review process, avoiding judgments based solely on identical text or similar coordinates. The page coordinates and original detection path output by the evidence location component allow anomaly conclusions to be mapped to specific page areas. This embodiment covers the collaborative processing of the page visual inspection component, the layout semantic graph construction component, the compliance constraint matching component, and the evidence location component. Its advantage lies in incorporating complex document layout relationships, field source relationships, and compliance rule matching relationships into a unified data structure, reducing review biases caused by misaligned fields, cross-page fields, and table fields in the engineering documents.

Claims

1. An intelligent review system for engineering data compliance based on multimodal document visual analysis, characterized in that, This includes a page visual inspection component, a layout semantic graph construction component, a compliance constraint matching component, and an evidence location component; The page visual inspection component performs multi-scale visual inspection of the layout area, text block, table boundary, figure frame, signature column and page number area in the engineering data page image; The layout semantic graph construction component organizes the detection results into file nodes, page nodes, layout area nodes, field nodes, and field value nodes, and establishes spatial inclusion relationships, adjacency relationships, header mapping relationships, and cross-page continuation relationships to form a layout semantic graph. The compliance constraint matching component matches the compliance rules of the engineering data with the layout semantic graph; The evidence location component is a system that outputs the coordinate information of abnormal nodes on the page.

2. The intelligent review system for engineering data compliance based on multimodal document visual parsing according to claim 1, characterized in that, The page visual inspection component includes a page type determination process, which generates page type labels based on title visual blocks, table density, figure frame distribution, signature column position, and page number style, and selects the corresponding set of detection scales according to the page type labels. The detection scale set includes different receptive field detection paths for full page layout, partial table, column boundary and field text block, and the candidate boxes of each detection path are merged by reading order constraint and region mutual exclusion constraint. The merged candidate boxes retain the detection scale identifier, page coordinates and low confidence re-examination identifier.

3. The intelligent review system for engineering data compliance based on multimodal document visual parsing according to claim 2, characterized in that, The layout semantic graph construction component assigns node attributes to the file nodes, page nodes, layout area nodes, field nodes, and field value nodes. The node attributes include page type, area category, text candidate, coordinate range, confidence level, and source detection path. The layout semantic graph construction component establishes edge attributes based on the vertical and horizontal projections, table line intersections, field title similarity, and visual block reading sequence. The edge attributes include at least containment edges, adjacent edges, same column edges, synonymous field edges, and cross-page continuation edges, and record the corresponding visual detection sources and edge generation basis.

4. The intelligent review system for engineering data compliance based on multimodal document visual parsing according to claim 3, characterized in that, The compliance constraint matching component expresses compliance rules as a graph constraint template that includes target node type, limited edge type, field consistency condition, time order condition, and necessary node existence condition, and performs candidate subgraph retrieval in the layout semantic graph; For a target field with multiple candidate field value nodes, the compliance constraint matching component determines the field value nodes to participate in the matching based on regional priority, adjacent title relevance, cross-page reference relationship and field value format consistency, and records the mapping relationship between template nodes and candidate subgraph nodes as well as the source of field values.

5. The intelligent review system for engineering data compliance based on multimodal document visual parsing according to claim 4, characterized in that, When identifying the table area and the frame bar, the page visual inspection component first generates table orientation priors and frame position priors based on the page type label, and then performs joint detection on table lines, cell closing boundaries, merged cell empty boundaries and frame bar separator lines. For pages with tilted or broken lines, the page visual inspection component reconstructs cell boundaries based on the endpoint clustering relationship in the same row and column, and aligns and binds the reconstructed cell boundaries with text block candidate boxes. The binding result is written to the source attribute and table level attribute of the corresponding field node.

6. The intelligent review system for engineering data compliance based on multimodal document visual parsing according to claim 5, characterized in that, The layout semantic graph construction component establishes a page sequence index for cross-page data. The page sequence index includes page number recognition results, table of contents reference results, title continuity results, and table beginning and end continuation results. When adjacent pages have the same field title, adjacent table structure, or consecutive attachment numbers, the layout semantic graph construction component establishes cross-page continuation edges between the corresponding page nodes, uses the continuation field nodes to share field identifiers, uses cross-page field value nodes to belong to the same compliance inspection object, and records the source page range and directory entry number of the continuation edge.

7. The intelligent review system for engineering data compliance based on multimodal document visual parsing according to claim 6, characterized in that, The compliance constraint matching component configures a rule trigger index, which uses document type, page type, and key layout area category as index items. When a new or updated field value node is added to the layout semantic graph, the compliance constraint matching component only calls the graph constraint template that matches the index item, and writes the hit abnormal nodes, dependency edges and unmet conditions into the review evidence set. The review evidence set forms an associated record with rule identifiers, node identifiers and page coordinates. Multiple rules triggered by the same field are merged into a matching task according to the node sharing relationship.

8. The intelligent review system for engineering data compliance based on multimodal document visual parsing according to claim 7, characterized in that, When the page visual detection component binds text block candidate boxes to cell boundaries, it introduces text baseline direction, character arrangement density, cell white space ratio, and candidate box edge overlap rate to form binding features, and merges multiple lines of text in the same cell into field value candidates according to the baseline projection order. For title text that spans multiple cells, the page visual detection component determines the coverage of the title text based on the empty boundaries of merged cells and the distribution of adjacent column fields, binds features and page type tags together to the field value node attributes, and associates them with the corresponding cell boundary identifiers.

9. The intelligent review system for engineering data compliance based on multimodal document visual parsing according to claim 8, characterized in that, When configuring shared field identifiers, the layout semantic graph construction component first performs field name normalization on cross-page field nodes. The field name normalization generates normalization candidates based on the similarity of field title characters, the layout area category where the title is located, the set of adjacent fields on the same page, and the directory reference text. When multiple normalized candidates correspond to the same engineering data entity, the layout semantic graph construction component establishes entity merging edges and retains the original page coordinates and original text candidates for field value nodes on different pages. The entity merging edges are configured with merging source, conflict retention flag and entity category identifier.

10. The intelligent review system for engineering data compliance based on multimodal document visual parsing according to claim 9, characterized in that, The evidence location component generates an abnormal evidence subgraph based on the review evidence set. The abnormal evidence subgraph includes rule nodes, abnormal field nodes, dependency field nodes, page nodes, and coordinate frame nodes. The evidence location component sorts the cover field, table field, graph frame field, and cross-page attachment field involved in the same compliance rule according to the dependency edge order in the abnormal evidence subgraph, and configures a page thumbnail index, region zoom index, original detection path identifier, and graph constraint template mapping identifier for each coordinate frame node, and keeps the sorting result consistent with the node identifier of the abnormal evidence subgraph.

Images