Multi-channel complex document parsing method and system for electronic information manufacturing industry

Abstract

The application discloses a kind of multi-channel complex document analysis method and system for electronic information manufacturing industry, solve the problem that existing scheme is prone to structural disorder when processing complex industrial document, model illusion, poor multi-format adaptation, high cost of artificial review.The application first discriminates document features and adaptively routes to corresponding analysis channel, respectively using MinerU, Docling to complete analysis, then generate Tree.json format chapter tree uniformly;Subsequently, confidence evaluation is carried out from three aspects and is shunted, local structure repair is carried out with the help of AI Agent, and finally, an optimized closed loop is formed through artificial review.The application can adapt to PDF, scan, various Office and other document formats, effectively avoid model illusion problem, completely retain document structure and semantic relationship, improve analysis accuracy, significantly reduce the workload of artificial review, and is suitable for electronic information manufacturing industry industrial document analysis, knowledge extraction and other scenes.

Description

Technical Field

[0001] This invention relates to a document parsing system, and more particularly to a method and system for parsing multi-channel complex documents for the electronics and information manufacturing industry. Background Technology

[0002] Document parsing is a crucial foundation for industrial knowledge extraction, knowledge base construction, and large-scale model corpus building. Process documents, equipment manuals, testing specifications, training materials, and standard documents in the electronics and information manufacturing industry typically feature diverse formats, complex layouts, dense charts and images, concentrated technical terminology, and frequent cross-page links. Compared to ordinary office documents, document parsing in the electronics and information manufacturing industry demands higher standards for structural integrity, text-image correspondence, chapter hierarchy, and fidelity of professional semantics. Existing document parsing solutions mainly fall into three categories: one category targets formatted documents such as PDFs, scanned documents, and images, typically relying on OCR, layout detection, table recognition, and reading order restoration; another category targets editable or semi-structured documents such as Word, Excel, PPT, HTML, and Markdown, typically relying on document object models, XML structures, style tags, or application interfaces to extract content; and the third category comprises generative OCR, visual language models, or multimodal large-scale model parsing solutions that have emerged in recent years, capable of comprehensively understanding page images, text regions, and contextual semantics, and generating natural language or structured parsing results. The above solutions work well in scenarios with a single format or regular layout, but they still have significant limitations in complex documents in the manufacturing industry: (1) Complex layouts in documents, such as cross-page tables, merged cells, and tables with incomplete borders, often lead to parsing failures or structural errors. For example, traditional OCR systems are prone to problems such as inconsistent column numbers and broken rows when processing nested tables. Even newer models such as HunyuanOCR still perform poorly in table structure recognition, resulting in duplicate outputs and reading order errors. Multimodal data fusion is insufficient. Multimodal documents contain various elements such as text, images, charts, formulas, audio, and video. How to effectively associate these data at the semantic level remains a challenge. Existing systems often can only process various modalities separately, lacking a unified cross-modal understanding capability, leading to information fragmentation. (2) End-to-end OCR or visual language models such as DeepSeek-OCR have strong image and text understanding capabilities, but their output is often presented as natural language generation results. When there are handwritten materials, low-resolution scanned documents, complex images, incomplete table boundaries, or missing context, the model may generate titles, fields, paragraphs, or table content that do not exist in the original text but seem semantically reasonable; it may also omit local areas, merge adjacent content, rewrite professional terms, or fill in missing information. For documents in the electronic information manufacturing industry, the above illusions will destroy the traceability of parameters, process steps, equipment part numbers, and testing standards, making it difficult to directly use them as highly reliable structured corpora; (3) For editable documents such as Word, Excel, and PPT, although existing methods can extract native objects such as text, tables, and images, they are difficult to fully preserve the hierarchical, spatial, referential, and text-image correspondence relationships between complex objects such as text boxes, graphics, connectors, combined objects, multi-level headers, formulas, comments, footnotes, and revision marks, resulting in inconsistent parsing results with the original text. (4) The document formats in the electronics and information manufacturing industry are complex, and a single parsing engine cannot adapt to scanned documents, PDFs, images and various Office documents at the same time. Using OCR uniformly is prone to losing the original structure, while using streaming parsing uniformly is difficult to restore the page space relationship. Existing systems usually lack automatic diversion mechanisms based on document type, page quality, structural complexity and parsing confidence, and also lack the ability to locally verify and repair disordered tables, cross-page content, mismatched text and images and suspected hallucination content, resulting in a large amount of manual review still required. Summary of the Invention

[0003] This invention addresses the aforementioned technical problems by providing a multi-channel complex document parsing method and system for the electronics and information manufacturing industry. It selects an appropriate parsing channel based on characteristics such as document format, whether it has been scanned, layout complexity, object nesting degree, and the density of images and tables. The results from different channels are uniformly mapped to a chapter tree (Tree.json). Based on Tree.json, confidence assessment, AI agent small-window structure repair, post-repair review, and manual verification are then performed. This improves the structural accuracy, traceability, and engineering usability of the parsed results for complex industrial documents, while reducing the workload of manual verification. It significantly enhances the ability to recognize and restore complex layout structures, avoids the illusion problem during parsing, fully preserves streaming document objects and semantic relationships, and provides a unified diversion, verification, and repair mechanism to ensure parsing quality.

[0004] Therefore, the technical solution of the present invention is a method for parsing multi-channel complex documents for the electronic information manufacturing industry, comprising the following steps: S1: Document input and type identification; Extract the file extension, text selectivity, image density, table density, object nesting degree, and whether there are slide objects or native table features of the document to be processed; S2: Multi-resolution channel adaptive routing; Based on the document type discrimination features extracted in S1, and combined with the parsed channel capability matrix, routing decisions are made to complete the following operations, including: Routing PDF documents, scanned documents, and image documents to the MinerU (an open-source layout document parsing tool) layout parsing channel; Routing editable streaming documents such as Word, PPT, and Excel to the Docling (an open-source native document parsing tool) streaming parsing channel; Depending on the document quality, parsing target, or project configuration, you can choose to route it to the Docling streaming parsing channel for native parsing, or first convert it to PDF format and then route it to the MinerU layout parsing channel. S3: Format-specific parsing; The document is parsed using the corresponding parsing channel determined by the routing, and the intermediate results corresponding to the channel are obtained. S4: Unified chapter tree construction; The intermediate results output from each parsing channel are uniformly mapped to a chapter tree in Tree.json (a unified chapter tree data format). Each node in the chapter tree contains at least the fields of number, title, body text, child node, and page number. Child nodes may contain titles, body text, tables, images, and formula objects. S5: Confidence assessment; Using the chapter tree as the calculation object, the overall confidence level is calculated around three dimensions: chapter number continuity, table duplication risk, and title rationality. The data is then processed and distributed according to the confidence level and preset rules. S6: AI Agent (Intelligent Repair Agent) for local structural repair; For chapter trees that are redirected to AI for repair, the goal is to correct errors in the chapter tree structure. The entire document is not re-parsed or the original business content is not rewritten. Only key information about abnormal nodes, adjacent nodes before and after abnormal nodes, and parent candidate nodes is provided to the Agent. Sending the complete chapter tree at once is prohibited. Candidate repair suggestions are generated based on deterministic rules. The repair agent determines the final repair scheme based on the local context semantics. The structure of the chapter tree is repaired through structured patching. After the repair is completed, the confidence of the chapter tree is recalculated. The repaired chapter tree is saved only if the confidence is improved after the repair and the original field information is not destroyed. Otherwise, the repair result is discarded and the original chapter tree is retained. S7: Manual review; For chapter trees that are routed to manual review, a split-screen review interface is used. The left side displays the repaired chapter tree structure, and the right side displays the corresponding page of the original document. Manual adjustments are made, and the differences before and after the manual adjustments are recorded to form a structured correction log. The difference information is used for subsequent optimization of parsing channel routing rules, spatial attribution algorithms, Agent prompts, and quality assessment thresholds, forming a closed loop consisting of parsing, assessment, repair, review, and optimization.

[0005] Preferably, in step S3, the parsing process of the MinerU format parsing channel includes: (1) Perform layout analysis, text recognition, table recognition, and image region detection on the input document; (2) Generate page-level and block-level intermediate results, including page parsing information, block-level structure information, a content list organized in reading order, Markdown text, and chapter tree candidate structure.

[0006] Preferably, in step S3, the parsing process of the Docling streaming parsing channel includes: Extract headings, paragraphs, tables, images, text boxes, grouped objects, multilevel lists, slide objects, and metadata native objects from a document, and read the bounding box coordinates, layer order, style properties, and document hierarchy of each object; Also includes: Enhance the positioning step by reading the bounding box of the floating object I to be assigned to. Calculate the coordinates of the center point of the floating object. ; Iterate through all cells in table T. For the cell in row i and column j, retrieve its bounding box. To determine whether the center point of a floating object falls within the bounded box of the cell, use the following formula: ; like If the center point falls within the bounding box of the corresponding cell, the floating object will be attached as a subordinate object of the corresponding cell, and the relationship between the floating object and the corresponding table row and column position will be recorded in the chapter tree; if the center point does not fall within the bounding box of any cell, the floating object will be attached to the page, slide, paragraph area or chapter node where it is located. Before determining whether the center point falls within the bounding box of the cell, the upper and lower boundaries of the floating object and the cell can be normalized to eliminate the difference in the vertical coordinate direction between different analytical coordinate systems.

[0007] Preferably, in step S5, the formula for calculating the overall confidence level is: ; in, , , The weights are assigned to three dimensions: continuity of chapter numbering, risk of table duplication, and reasonableness of title, with the sum of each weight being 1. Score based on the continuity of chapter numbers. To score the risk of duplicate entries in the table, Score for the reasonableness of the title.

[0008] The calculation process for the chapter number continuity score includes: identifying the starting point of the text number, marking the nodes before the starting point as the preceding nodes, and performing continuity checks based only on the numbering of the text nodes. The continuity checks include checking for abnormal numbering formats, duplicate numbering, skipped numbering, hierarchical rollback errors, and abnormal starting points of sub-level numbering. The table duplication risk score is used to detect the risk of duplicate cells, duplicate rows, duplicate columns, and abnormal duplication of long text in a document. The table duplication risk score is only used to indicate risks in the underlying parsing or OCR, and does not directly authorize AI to modify the content of table cells, nor can it automatically rewrite the content of table cells based on the score. The title reasonableness score is used to detect anomalies in a document, such as empty titles, excessively long titles, titles containing body text, titles that appear to contain table fields, or titles containing too many numbers.

[0009] Preferably, in step S5, the triage rules for confidence assessment include: (1) If the confidence level of the chapter tree passes and the mandatory review condition is not triggered, the chapter tree is marked as automatically passed; (2) If the confidence level of the chapter tree is insufficient, or the numbering or title is abnormal and meets the preset repair conditions, the chapter tree will be marked as AI repair status; (3) If the chapter tree only has structural problems such as the risk of duplicate tables, abnormal chapter numbers, and abnormal titles, it will not enter the AI ​​repair process; (4) If the problem of the chapter tree cannot be safely repaired at the structural layer, or if it still fails the evaluation after repair, or if the mandatory review conditions are triggered, the chapter tree will be marked as a manual review status. The conditions for mandatory review include: a concentrated occurrence of critical structural anomalies, failure to improve confidence after repair, and a structural score of the chapter tree that is lower than a preset threshold.

[0010] Preferably, in step S6, the Agent is only allowed to return structured patches, and is prohibited from returning natural language repair instructions or returning complete new trees; Structured patching includes at least one of the following: node merging, node splitting, title splitting, title merging, and no need to modify tags. At the same time, the Agent is prohibited from performing the following operations: rewriting table cell content, rewriting OCR-recognized content, rewriting the entire chapter tree, rearranging chapters extensively without local evidence, and discarding original text, images, tables, page numbers, or coordinate information.

[0011] A multi-channel complex document parsing system for the electronics and information manufacturing industry includes: (1) Type discrimination module; It is used to input and type-discriminate the document to be processed, and extract the document's file extension, text selectivity, image density, table density, object nesting degree, and whether there are slide objects or native table features; (2) Routing module; It is used to perform multi-parsing channel adaptive routing. Based on features and the parsing channel capability matrix, it makes routing decisions and routes PDF documents, scanned documents, and image documents to the MinerU (open-source format document parsing tool) format parsing channel. Routing editable streaming documents such as Word, PPT, and Excel to the Docling (an open-source native document parsing tool) streaming parsing channel; It is also used to route editable Office documents to the Docling streaming parsing channel for native parsing, or to convert them to PDF format first and then route them to the MinerU layout parsing channel, depending on the document quality, parsing target, or project configuration. (3) Parsing module;

[0012] It includes the MinerU layout parsing unit and the Docling streaming parsing unit, which are used to parse the document through the corresponding parsing channel determined by the route and obtain the intermediate results corresponding to the channel; (4) Chapter tree construction module; This is used to uniformly map the intermediate results output from each parsing channel to a chapter tree in Tree.json (a unified chapter tree data format). Each node in the chapter tree must contain at least the fields of number, title, body text, child node, and page number. Child nodes may contain titles, body text, tables, images, and formula objects. (5) Confidence assessment module; It is used to calculate the overall confidence level based on the chapter tree, focusing on three dimensions: chapter number continuity, table duplication risk, and title rationality, and then process and distribute the data according to the confidence level and preset rules. (6) Repair module; Used for chapter trees that are redirected to AI for repair, with the goal of correcting chapter tree structure errors, without re-parse the entire document or rewrite the original business content, only providing the Agent with partial information on abnormal nodes, adjacent nodes before and after abnormal nodes, and parent candidate nodes, and prohibiting the sending of the complete chapter tree at once; Candidate repair suggestions are generated based on deterministic rules. The repair_agent determines the final repair scheme based on the local context semantics. The structure of the chapter tree is repaired through structured patching. After the repair is completed, the confidence of the chapter tree is recalculated. The repaired chapter tree is saved only if the confidence is improved after the repair and the original field information is not destroyed. Otherwise, the repair result is discarded and the original chapter tree is retained. (7) Review module; This is used to display a review interface for chapter trees that have been diverted to manual review. The interface includes the original document, the repaired chapter tree, the abnormal nodes, the abnormal tables, the abnormal titles and their page numbers. It receives manual adjustment operations and records the differences before and after the manual adjustment operations. This difference information is used for subsequent optimization of parsing channel routing rules, spatial attribution algorithms, Agent prompt words and quality assessment thresholds, forming a closed loop consisting of parsing, evaluation, repair, review and optimization.

[0013] Preferably, the MinerU format parsing unit is used for: (1) Perform layout analysis, text recognition, table recognition, and image region detection on the input document; (2) Generate page-level and block-level intermediate results, including page parsing information, block-level structure information, content list organized in reading order, Markdown text, and chapter tree candidate structure.

[0014] Preferably, the Docling streaming parsing unit is used to extract headings, paragraphs, tables, images, text boxes, grouped objects, multilevel lists, slide objects, and metadata native objects from the document, and read the bounding box coordinates, layer order, style attributes, and document hierarchy of each object; The Docling streaming parsing unit is also used to perform enhanced localization operations, as follows: Read the bounding box of the floating object I to be assigned Calculate the coordinates of the center point of the floating object. ; Iterate through all cells in table T. For the cell in row i and column j, retrieve its bounding box. To determine whether the center point of a floating object falls within the bounded box of the cell, use the following formula: ; like If the center point falls within the bounding box of the corresponding cell, the floating object will be attached as a subordinate object of the corresponding cell, and the relationship between the floating object and the corresponding table row and column position will be recorded in the chapter tree; if the center point does not fall within the bounding box of any cell, the floating object will be attached to the page, slide, paragraph area or chapter node where it is located. Before determining whether the center point falls within the bounding box of the cell, the upper and lower boundaries of the floating object and the cell can be normalized to eliminate the difference in the vertical coordinate direction between different analytical coordinate systems.

[0015] Preferably, in the repair module, the Agent is only allowed to return structured patches, and is prohibited from returning natural language repair instructions or returning a complete new tree; Structured patching includes at least one of the following: node merging, node splitting, title splitting, title merging, and no need to modify tags. Additionally, the fix module also prevents the Agent from performing the following operations: Rewrite table cell content, rewrite OCR-recognized content, rewrite the entire chapter tree, rearrange chapters on a large scale when there is no local evidence, and discard the original text, images, tables, page numbers, or coordinate information; The review module records the differences before and after the adjustment, including the anomaly type, repair method, node adjustment relationship, manual confirmation result, and change in confidence level.

[0016] The beneficial effects of this invention are: (1) Through a multi-channel routing mechanism, combined with the parsing channel capability matrix, accurate routing is achieved. PDF, scanned documents, and image documents can enter the MinerU format parsing channel, while editable documents such as Word, PPT, and Excel can enter the Dodling streaming parsing channel. This allows different documents to be processed according to their original organization, reducing structural loss or parsing distortion caused by incompatibility with a single engine, reducing the risk of parsing errors from the source. Moreover, the routing mechanism is not limited to specific tools and has a wider range of adaptability and protection. (2) By unifying the Tree.json data structure, information such as page blocks, native objects, tables, images, text boxes, page numbers and coordinates are mounted into a unified chapter tree. The top layer adopts the form of a node list object to ensure structural consistency, so that subsequent knowledge extraction, semantic retrieval and manual review can be carried out around the same structural object, avoiding the problem that flat content lists are difficult to express chapter ownership, and solving the problem of heterogeneous output from different parsing channels. (3) Reduce the impact of generative OCR illusion on industrial data; the system will not directly take the natural language output of end-to-end generative OCR as the final credible result, but will reduce the impact of model supplementation, miswriting, and omission on process parameters, component numbers and testing standards through constraints such as structured field retention, confidence assessment, prohibition of table content modification and manual review, and ensure the traceability of industrial data. (4) Improve the security boundary of automated processing; AI Agent only handles local structural problems that can be located and verified, such as misplaced chapters, mixed titles into the main text, and missplitting or merging nodes, and must be reviewed after each repair; the repair results are only saved when the confidence level is improved and the original fields are not destroyed, otherwise the original parsing results are retained, thereby avoiding uncontrollable rewriting of the original business content by the large model, ensuring the effectiveness of the repair while avoiding the illusion risk of generative models; (5) Reduce the cost of manual review and improve the efficiency of review. The system first screens out high-confidence samples through confidence assessment, then repairs some structural problems through AI Agent, and finally only hands over low-confidence, high-risk or unrepairable samples to manual review, so that manual work is focused on the anomaly locations that really need to be judged, rather than checking all documents page by page; at the same time, the system can be continuously optimized through the difference information of manual review, continuously improve the automation processing capability, and further reduce the subsequent manual cost. Attached Figure Description

[0017] Figure 1 This is the overall architecture diagram of the multi-channel complex document parsing system of the present invention; Figure 2 This is a flowchart of the confidence assessment and triage process of this invention; Figure 3 This is a flowchart of the AI ​​Agent small window local structure repair process of this invention; Figure 4 This is a flowchart of the manual review and final confirmation process for this invention. Detailed Implementation

[0018] The present invention will be further described below with reference to embodiments.

[0019] pass Figures 1-4 As can be seen, this multi-channel complex document parsing method for the electronics and information manufacturing industry includes the following steps.

[0020] S1: Document input and type identification;

[0021] Extracting the file extension, text selectivity, image density, table density, object nesting degree, and presence of slide objects or native tables from the document to be processed; in this step, multi-dimensional feature extraction can comprehensively characterize the document's format attributes and content complexity, providing accurate decision-making basis for subsequent parsing channel routing, avoiding routing deviations caused by relying solely on file extensions, and improving the accuracy of parsing channel adaptation from the entry point.

[0022] S2: Multi-resolution channel adaptive routing;

[0023] Based on the document type discrimination features extracted in S1, and combined with the parsing channel capability matrix, routing is determined. This channel routing mechanism is not limited to a specific tool name; its core lies in adapting and matching documents with parsing channels based on the capabilities of different parsing links in layout restoration, native object extraction, table structure restoration, image attribution determination, and chapter tree construction. Specifically, it is used for:

[0024] This process routes PDF documents, scanned documents, and image documents to the MinerU (an open-source document parsing tool) layout parsing channel; it routes editable streaming documents such as Word, PPT, and Excel to the Docling (an open-source native document parsing tool) streaming parsing channel. It also allows for routing editable Office documents like Word, PPT, and Excel to the Docling streaming parsing channel for native parsing, or converting them to PDF before routing to the MinerU layout parsing channel, depending on document quality, parsing goals, or project configuration. This step uses a parsing channel capability matrix for adaptive routing, ensuring documents of different organizational forms are matched to the most suitable parsing path. Layout-based documents use the layout parsing channel to guarantee accurate layout structure restoration, while stream-based editable documents use the native parsing channel to preserve native object attributes, reducing structural loss and parsing distortion caused by incompatibility with a single parsing engine. Furthermore, it provides a dual-path option for Office documents, allowing for flexible selection of parsing strategies based on actual business needs, balancing native structure preservation with layout consistency requirements.

[0025] S3: Format-specific parsing;

[0026] The document is parsed using the corresponding parsing channel determined by the routing, and the intermediate result corresponding to the channel is obtained.

[0027] S4: Unified chapter tree construction;

[0028] The intermediate results output from each parsing channel are uniformly mapped to a chapter tree in Tree.json (a unified chapter tree data format). The top level of Tree.json adopts a node list object form. Each node in the chapter tree contains at least the fields of number, title, body text, child nodes, and page number. Child nodes can contain titles, body text, tables, images, and formula objects. This step achieves format unification of multi-channel parsing results through a unified Tree.json data structure, eliminating the problem of heterogeneous output structures from different parsing engines and providing a unified operation object for subsequent quality assessment, AI repair, and manual review. The hierarchical structure based on the chapter tree can fully express the chapter membership and content nesting logic of the document, solving the problem that flat content lists cannot reflect the semantic hierarchy of the document, and facilitating direct calls by downstream applications such as knowledge extraction and semantic retrieval. The table below is the chapter tree field table.

[0029] num Chapter Number Used for hierarchical recovery and continuity checks. title Chapter Title Used to determine the appropriateness of the title and the content's attribution. text Chapter Text Save the main text content directly related to this chapter. children List of child objects It can contain objects such as title, text, table, and image. page_index Page number or page index Used for tracing the location of the original document and for manual verification.

[0030] S5: Confidence assessment;

[0031] Using the chapter tree as the calculation object, the overall confidence level is calculated around three dimensions: chapter number continuity, table duplication risk, and title rationality. The data is then processed and distributed according to the confidence level and preset rules.

[0032] S6: AI Agent (Intelligent Repair Agent) for local structural repair;

[0033] For chapter trees that are redirected to AI for repair, the goal is to correct errors in the chapter tree structure. The entire document will not be re-parsed or the original business content will not be rewritten. Only key information about abnormal nodes, adjacent nodes before and after abnormal nodes, and parent candidate nodes will be provided to the Agent. Sending the complete chapter tree at once is prohibited.

[0034] Candidate repair suggestions are generated based on deterministic rules. The repair agent determines the final repair scheme based on the local context semantics. The structure of the chapter tree is repaired through structured patching. After the repair is completed, the confidence of the chapter tree is recalculated. The repaired chapter tree is saved only if the confidence is improved after the repair and the original field information is not destroyed; otherwise, the repair result is discarded and the original chapter tree is retained.

[0035] S7: Manual review;

[0036] For chapter trees that are routed to manual review, the review interface adopts a split-screen layout. The left side displays the repaired chapter tree structure, and the right side displays the corresponding pages of the original document. Manual adjustments are made to chapter levels, parent-child relationships, title splitting, content mounting, image attribution, and table attribution. The interface supports one-click confirmation of no abnormal content and online editing to correct structural errors. Difference information before and after adjustments is recorded to form a structured correction log. The difference information can be used for subsequent optimization of parsing channel routing rules, spatial attribution algorithms, Agent prompt words, and quality assessment thresholds, forming a closed loop consisting of parsing, evaluation, repair, review, and optimization.

[0037] In a specific embodiment, step S3, the parsing process of the MinerU format parsing channel includes:

[0038] (1) Perform layout analysis, text recognition, table recognition and image region detection on the input document.

[0039] (2) Generate page-level and block-level intermediate results, including page parsing information, block-level structure information, a content list organized in reading order, Markdown text, and chapter tree candidate structure.

[0040] This parsing channel, through a full-process processing of layout analysis, text recognition, table recognition, and image region detection, can completely extract page elements and layout information from non-editable documents such as scanned documents and image-based PDFs; generating multi-dimensional intermediate results can provide sufficient structural and content materials for subsequent chapter tree construction, ensuring the parsing granularity and content integrity of layout documents.

[0041] In a specific embodiment, step S3, the parsing process of the Docling streaming parsing channel includes: extracting titles, paragraphs, tables, images, text boxes, grouped objects, multilevel lists, slide objects, and metadata native objects from the document, and reading the bounding box coordinates, layer order, style attributes, and document hierarchy relationships of each object.

[0042] This parsing method extracts various elements and their attributes based on the document's native object model. It can completely preserve the original hierarchical relationship, style information, and spatial attributes of the editable document, avoiding the loss of the original structure and the disorder of object relationships caused by OCR parsing, and restoring the original editing logic and semantic association of the document to the greatest extent.

[0043] In a specific embodiment, step S3, the parsing process of the Docling streaming parsing channel further includes:

[0044] Enhance the positioning step by reading the bounding box of the floating object I to be assigned to. Calculate the coordinates of the center point of the floating object. .

[0045] Iterate through all cells in table T. For the cell in row i and column j, retrieve its bounding box. To determine whether the center point of a floating object falls within the bounded box of the cell, use the following formula: .

[0046] like If the center point does not fall within the bounding box of any cell, the floating object will be attached as a subordinate object of the corresponding cell, and the affiliation relationship between the floating object and the corresponding table row and column position will be recorded in the chapter tree. If the center point does not fall within the bounding box of any cell, the floating object will be attached to the page, slide, paragraph area, or chapter node where it is located. This enhanced positioning step can accurately establish the affiliation relationship between the floating object and the table cell by determining whether the center point falls within the cell, which solves the problem of floating objects and tables being stored at the same level and the lack of image-text correspondence in the original parsing, and improves the semantic relevance and spatial affiliation accuracy of the parsing results. For objects that do not fall into cells, they can be attached by page and chapter, which can ensure that all floating objects have a clear chapter affiliation and prevent content from being outside the chapter system.

[0047] In a specific embodiment, before determining whether the center point falls within the cell bounding box, the upper and lower boundaries of the floating object and the cell can be normalized to eliminate the difference in vertical coordinate direction between different parsing coordinate systems. By normalizing the coordinates, the differences in coordinate systems of different parsing tools can be accommodated, improving the universality and robustness of the attribution determination algorithm and ensuring the consistency of attribution determination in scenarios where multiple tools are used for mixed parsing.

[0048] In a specific embodiment, in step S5, the formula for calculating the overall confidence level of the confidence assessment is as follows:

[0049] in, , , The weights are assigned to three dimensions: continuity of chapter numbering, risk of table duplication, and reasonableness of title, with the sum of each weight being 1. Score based on the continuity of chapter numbers. To score the risk of duplicate entries in the table, The system scores the reasonableness of the title; it constructs a quantitative evaluation system from three core dimensions: chapter number, table, and title, which can objectively and quantitatively evaluate the structural quality of the parsed results, replace subjective human judgment, realize the automated grading of parsing quality, and provide a clear decision basis for subsequent processing and diversion; the evaluation ratio of different dimensions can be flexibly adjusted through weight configuration to adapt to the quality emphasis needs of different industries and document types.

[0050] In a specific embodiment, the calculation process for the chapter numbering continuity score includes: identifying the starting point of the text numbering, marking the nodes before the starting point as preceding nodes, and performing continuity checks based solely on the numbering of the text nodes. The continuity checks include checking for abnormal numbering formats, duplicate numbers, skipped numbers, hierarchical rollback errors, and abnormal starting points for sub-level numbers. By identifying the starting point of the text numbering and eliminating interference from preceding pages such as the cover and table of contents, the text chapter system can be accurately located, improving the accuracy of the numbering continuity check. Comprehensive detection of multiple types of numbering anomalies can cover common parsing structure errors such as disordered chapter hierarchy and missing numbers, providing accurate anomaly location for subsequent structure repair.

[0051] In one specific embodiment, the table duplication risk score is used to detect the risk of duplicate cells, duplicate rows, duplicate columns, and abnormal duplication of long text in a document. The table duplication risk score is only used to indicate risks in the underlying parsing or OCR process, and does not directly authorize AI to modify the content of table cells, nor can it automatically rewrite the content of table cells based on the score. By detecting abnormal duplication of table content, it is possible to identify in advance the risks of recognition errors and duplicate outputs that may occur in the OCR or table recognition process. At the same time, limiting this indicator to only provide risk warnings and not triggering automatic rewriting can avoid AI automatically modifying table content, which may introduce illusions or data distortion, and ensure the traceability and accuracy of table data such as process parameters, testing standards, and part numbers in industrial documents.

[0052] In one specific embodiment, the title rationality score is used to detect anomalies in the document such as empty titles, excessively long titles, titles mixed with body text, titles that seem to contain table fields, or titles containing too many numbers. Through multi-dimensional title anomaly detection, it is possible to accurately identify title parsing error scenarios, provide targeted targets for subsequent title structure repair, avoid semantic confusion of chapters caused by title and body text confusion, and ensure the accuracy of the chapter tree title system.

[0053] In a specific embodiment, step S5 includes the following routing rules for confidence assessment:

[0054] (1) If the confidence level of the chapter tree passes and the mandatory review condition is not triggered, the chapter tree is marked as automatically passed.

[0055] (2) If the confidence level of the chapter tree is insufficient, or the numbering or title is abnormal and meets the preset repair conditions, the chapter tree will be marked as AI repair status.

[0056] (3) If the chapter tree only has structural problems such as risk of duplicate tables, abnormal chapter numbering, and abnormal titles, it will not enter the AI ​​repair process.

[0057] (4) If the problem of the chapter tree cannot be safely repaired at the structural layer, or if it still fails the evaluation after repair, or if the mandatory review conditions are triggered, the chapter tree will be marked as a manual review status.

[0058] The conditions for mandatory review include: a concentrated occurrence of critical structural anomalies, failure to improve confidence after repair, and a structural score of the chapter tree that is lower than a preset threshold.

[0059] The table below shows the reliability assessment indicators and processing rules.

[0060] Chapter numbering continuity Errors include: incorrect numbering format, empty number, duplicate number, skipped number, incorrect level rollback, incorrect starting point of sub-level numbering, and misjudgment of the starting point of text numbering. If the anomaly meets the preset repair conditions, it will proceed to the AI ​​Agent for local structural repair; if it cannot be safely repaired at the structural level, or if the confidence level still fails after repair, it will be transferred to manual review. Table duplication risk Duplicate cells, duplicate rows, duplicate columns, and abnormal repetition of long text. This is only used as a risk warning for underlying parsing / OCR; AI is prohibited from automatically rewriting table cell content. Documents with only this type of risk, no chapter numbering, and abnormal title structure will not be included in the AI ​​repair process. Title rationality Empty headings, excessively long headings, headings mixed into the body text, headings that appear to contain table fields, and headings containing too many numbers. If an anomaly meets the preset repair conditions, AI repair is performed through structured patches such as title splitting and title merging; after repair, the confidence level is recalculated, and if the confidence level does not improve, it is transferred to manual review.

[0061] The three-tiered triage mechanism described above matches the corresponding processing path based on the quality level of the parsing results. High-confidence results are automatically passed to improve processing efficiency, repairable structural anomalies are fixed by AI to reduce manual costs, and high-risk content is transferred to manual review to ensure parsing quality. For scenarios with only table-related risks, manual review is directly initiated, which avoids the security risks of AI modifying table data, thus improving automation while maintaining the accuracy of industrial data. The setting of mandatory review conditions can intercept low-quality parsing results, preventing documents with serious structural errors from flowing into downstream applications.

[0062] In a specific embodiment, in step S6, the Agent is only allowed to return structured patches, and is prohibited from returning natural language repair instructions or returning a complete new tree. The structured patch includes at least one of node merging, node splitting, title splitting, title merging, and no-modification markers. At the same time, the Agent is prohibited from performing the following operations: rewriting table cell content, rewriting OCR recognition content, rewriting the entire chapter tree, rearranging chapters on a large scale without local evidence, and discarding the original text, images, tables, page numbers, or coordinate information. By adopting a small-window local repair mode, only the local context surrounding the anomaly is provided to the Agent, which can provide sufficient semantic basis for repair and avoid the risk of leakage of the complete chapter tree and large-scale content tampering. Limiting the output to only structured patches and strictly prohibiting modification of the original business content can strictly limit the operation boundary of AI to the scope of structural adjustment, and eliminate the possibility of generative models rewriting business content and generating illusions from a mechanism perspective. The post-repair review and confidence improvement verification mechanism can ensure that the repair action effectively improves the structural quality, and ineffective repairs are automatically rolled back to avoid introducing new structural problems through erroneous repairs, thus ensuring the security and effectiveness of automated repair.

[0063] In a specific embodiment, in step S7, the difference information before and after the adjustment includes the anomaly type, repair method, node adjustment relationship, manual confirmation result, and confidence level change; the left-right split-screen review interface can help reviewers quickly locate the problem without having to check page by page, greatly improving the efficiency of manual review; the complete difference information recording and closed-loop optimization mechanism can transform the experience of manual review into input for system optimization, continuously iterate channel routing, attribution algorithm, evaluation threshold, and repair capability, achieve self-improvement of system parsing accuracy, and gradually reduce the proportion of manual review.

[0064] A multi-channel complex document parsing system for the electronics and information manufacturing industry, characterized by comprising:

[0065] (1) Type discrimination module;

[0066] This module is used to input and type-discriminate documents, extracting features such as file extensions, text selectivity, image density, table density, object nesting degree, and the presence of slide objects or native tables. It can comprehensively collect document format and content features, providing multi-dimensional input support for routing decisions, avoiding the one-sidedness of single feature judgment, and improving the accuracy of route matching.

[0067] (2) Routing module;

[0068] This mechanism is used to perform adaptive routing across multiple parsing channels. It determines routes based on features and the parsing channel capability matrix. This routing mechanism is not limited to a specific tool name; its core function is to adapt and match documents with parsing channels based on the capabilities of different parsing links in layout restoration, native object extraction, table structure reconstruction, image attribution determination, and chapter tree construction. Specifically, it is used for:

[0069] This module routes PDF documents, scanned documents, and image documents to the MinerU (an open-source document parsing tool) format parsing channel; it routes editable streaming documents such as Word, PPT, and Excel to the Docling (an open-source native document parsing tool) streaming parsing channel; and based on document quality, parsing target, or project configuration, it selects whether to route the document to the Docling streaming parsing channel for native parsing, or to first convert it to PDF format and then route it to the MinerU format parsing channel. This module achieves intelligent adaptation of documents and parsing channels based on a capability matrix, fully leveraging the technical advantages of different parsing links, reducing the risk of structural distortion from the parsing entry point, and supporting dual-path selection for Office documents to adapt to diverse business parsing needs.

[0070] (3) Parsing module;

[0071] It includes the MinerU layout parsing unit and the Docling streaming parsing unit, which are used to parse the document through the corresponding parsing channel determined by the route and obtain the intermediate result corresponding to the channel.

[0072] (4) Chapter tree construction module;

[0073] This module is used to uniformly map the intermediate results output from various parsing channels into a chapter tree in Tree.json (a unified chapter tree data format). The top level of Tree.json adopts a list of nodes object. Each node in the chapter tree contains at least the fields of number, title, body text, child nodes, and page number. Child nodes can contain titles, body text, tables, images, and formula objects. This module realizes the structural unification of multi-source parsing results, builds a unified hierarchical chapter data structure, provides a unified data foundation for subsequent end-to-end processing, eliminates the downstream adaptation costs caused by heterogeneous parsing outputs, and fully expresses the semantic hierarchy of the document through the hierarchical structure.

[0074] (5) Confidence assessment module;

[0075] This tool is used to calculate the overall confidence level based on the chapter tree, focusing on three dimensions: chapter number continuity, table duplication risk, and title rationality. It then processes and distributes data according to the confidence level and preset rules.

[0076] (6) Repair module;

[0077] This is used for chapter trees that are redirected to AI for repair. The goal is to correct errors in the chapter tree structure. It does not re-parse the entire document or rewrite the original business content. It only provides the Agent with partial information about abnormal nodes, adjacent nodes before and after abnormal nodes, and parent candidate nodes. Sending the complete chapter tree at once is prohibited.

[0078] Candidate repair suggestions are generated based on deterministic rules. The repair agent determines the final repair scheme based on the local context semantics. The structure of the chapter tree is repaired through structured patching. After the repair is completed, the confidence of the chapter tree is recalculated. The repaired chapter tree is saved only if the confidence is improved after the repair and the original field information is not destroyed. Otherwise, the repair result is discarded and the original chapter tree is retained.

[0079] (7) Review module;

[0080] This module is designed for chapter trees that have been routed to manual review. It displays a review interface including the original document, the repaired chapter tree, abnormal nodes, abnormal tables, abnormal titles, and their page numbers. It receives manual adjustments to chapter hierarchy, parent-child relationships, title splitting, content mounting, image attribution, and table attribution, recording the differences before and after adjustments. This difference information includes the abnormality type, repair method, node adjustment relationship, manual confirmation result, and change in confidence level. This difference information can be used to subsequently optimize parsing channel routing rules, spatial attribution algorithms, Agent prompts, and quality assessment thresholds, forming a closed loop of parsing, assessment, repair, review, and optimization. This module provides a manual fallback verification entry point for high-risk parsing results, and the split-screen display design effectively improves the efficiency of manual review. Simultaneously, by accumulating manual experience through difference logs, it enables continuous iterative optimization of the system's parsing capabilities, forming a positive closed loop of quality improvement.

[0081] In one specific embodiment, the MinerU layout parsing unit is used for:

[0082] (1) Perform layout analysis, text recognition, table recognition and image region detection on the input document; (2) Generate intermediate results at the page level and block level, including page parsing information, block-level structure information, content list organized in reading order, Markdown text and chapter tree candidate structure; This unit completes full-element layout parsing for non-editable layout documents, fully restores the page layout and content elements, and ensures the parsing granularity and content integrity of layout documents.

[0083] In one specific embodiment, the Docling streaming parsing unit is used to extract headings, paragraphs, tables, images, text boxes, grouped objects, multilevel lists, slide objects, and metadata native objects from a document, and reads the bounding box coordinates, layer order, style attributes, and document hierarchy relationships of each object. This unit extracts editable document content based on the native object model, fully preserving the original document hierarchy and object attributes, avoiding information loss and structural disorder caused by OCR, and restoring the original semantic relationships of the document to the greatest extent.

[0084] In one specific embodiment, the Docling streaming parsing unit is also used to perform enhanced localization operations, as follows:

[0085] Read the bounding box of the floating object I to be assigned Calculate the coordinates of the center point of the floating object. .

[0086] Iterate through all cells in table T. For the cell in row i and column j, retrieve its bounding box. To determine whether the center point of a floating object falls within the bounded box of the cell, use the following formula: .

[0087] like If the floating object is not within the bounding box of any cell, it will be attached as a subordinate object of the corresponding cell, and the relationship between the floating object and the corresponding table row and column position will be recorded in the chapter tree. If the center point does not fall within the bounding box of any cell, the floating object will be attached to the page, slide, paragraph area or chapter node where it is located. This enhanced positioning function can accurately reconstruct the subordinate relationship between floating objects and tables, complete the missing text and image correspondence logic in the original parsing, improve the semantic relevance and spatial belonging accuracy of the parsing results, and ensure that all content elements have a clear chapter belonging.

[0088] In one specific embodiment, the Docling streaming parsing unit is also used for:

[0089] Before determining whether the center point falls within the cell bounding box, the upper and lower boundaries of the floating object and the cell are normalized to eliminate the difference in vertical coordinate direction between different analytical coordinate systems. Coordinate normalization can be compatible with the coordinate system differences of different analytical tools, improve the universality and robustness of the attribution determination algorithm, and ensure the consistency of attribution determination in multi-analysis scenarios.

[0090] In a specific embodiment, the formula for calculating the overall confidence level by the confidence assessment module is as follows:

[0091]

[0092] in, , , The weights are assigned to three dimensions: continuity of chapter numbering, risk of table duplication, and reasonableness of title, with the sum of each weight being 1. Score based on the continuity of chapter numbers. To score the risk of duplicate entries in the table, The title is scored for its reasonableness; a multi-dimensional quantitative scoring system can achieve objective and automated evaluation of parsing quality, providing a clear decision-making basis for hierarchical and tiered management, while also supporting flexible weight configuration to adapt to the quality assessment needs of different types of documents.

[0093] In a specific embodiment, the confidence assessment module calculates the chapter number continuity score by identifying the starting point of the text number, marking the nodes before the starting point as preceding nodes, and performing continuity checks only based on the numbering of the text nodes. The continuity check includes checking for abnormal numbering formats, duplicate numbers, skipped numbers, hierarchical rollback errors, and abnormal starting points of child numbers. This eliminates interference from preceding nodes, improves the accuracy of the number continuity check, and the multi-dimensional anomaly detection can comprehensively cover chapter structure parsing errors, providing accurate anomaly location for subsequent repairs.

[0094] In a specific embodiment, the confidence assessment module uses a table duplication risk score to detect duplicate cells, duplicate rows, duplicate columns, and abnormal duplication of long text in a document. The table duplication risk score is mainly used to indicate risks in the underlying parsing or OCR recognition, and does not directly authorize AI to modify the content of table cells, nor can it automatically rewrite the content of table cells based on the score. By detecting table duplication risks, abnormalities in the underlying parsing can be identified in advance. At the same time, by prohibiting automatic rewriting, the accuracy and traceability of industrial document table data are ensured, and the impact of AI illusions on core business data is avoided.

[0095] In a specific embodiment, the confidence assessment module uses the title rationality score to detect anomalies in the document such as empty titles, excessively long titles, titles mixed with body text, titles that are suspected of containing table fields, or titles containing too many numbers; multi-dimensional title anomaly detection can accurately identify title parsing errors, provide targeted targets for structural repair, and ensure the accuracy and semantic rationality of the chapter tree title system.

[0096] In one specific embodiment, the traffic splitting rules of the confidence assessment module include:

[0097] (1) If the confidence level of the chapter tree passes and the mandatory review condition is not triggered, the chapter tree is marked as automatically passed.

[0098] (2) If the confidence level of the chapter tree is insufficient, or the numbering or title is abnormal and meets the preset repair conditions, the chapter tree will be marked as AI repair status.

[0099] (3) If the chapter tree only has structural problems such as risk of duplicate tables, abnormal chapter numbering, and abnormal titles, it will not enter the AI ​​repair process.

[0100] (4) If the problem of the chapter tree cannot be safely repaired at the structural layer, or if it still fails the evaluation after repair, or if the mandatory review conditions are triggered, the chapter tree will be marked as a manual review status.

[0101] The mandatory review conditions include: a concentrated occurrence of critical structural anomalies, failure to improve confidence after repair, and a structural score of the chapter tree below a preset threshold.

[0102] The above-mentioned hierarchical and triage mechanism can achieve a balance between efficiency and quality. High-reliability documents are automatically processed to improve efficiency, and anomalies can be repaired by AI to reduce manual costs. High-risk content is manually guaranteed to ensure data quality, while low-quality parsing results are blocked through mandatory review conditions.

[0103] In one specific embodiment, the repair module prohibits sending the complete chapter tree to the Agent, and the Agent is only allowed to return structured patches, which include at least one of node merging, node splitting, title splitting, title merging, and no-modification tags. At the same time, the Agent is prohibited from returning natural language repair instructions and from returning a complete new tree.

[0104] The repair module also prohibits the Agent from performing the following operations: rewriting table cell content, rewriting OCR-recognized content, rewriting the entire chapter tree, rearranging chapters on a large scale without local evidence, and discarding original text, images, tables, page numbers, or coordinate information. This module completes the automated repair of the chapter structure under strict boundary constraints. The small window input and structured patching mechanism fundamentally limits the scope of AI operations, preventing generative models from tampering with the original business content and creating illusions. The post-repair review and invalid rollback mechanism can ensure the quality of repair, avoid introducing new problems through erroneous repairs, and effectively reduce the workload of manual review while ensuring data security.

[0105] However, the above description is merely a specific embodiment of the present invention and should not be construed as limiting the scope of the present invention. Therefore, any substitution of equivalent components or equivalent changes and modifications made in accordance with the scope of protection of the present invention should still fall within the scope of the claims of the present invention.

Claims

1. A method for parsing multi-channel complex documents for the electronics and information manufacturing industry, characterized by: Includes the following steps: S1: Document input and type identification; Extract the file extension, text selectivity, image density, table density, object nesting degree, and whether there are slide objects or native table features of the document to be processed; S2: Multi-resolution channel adaptive routing; Based on the document type discrimination features extracted in S1, and combined with the parsed channel capability matrix, routing decisions are made to complete the following operations, including: Routing PDF documents, scanned documents, and image documents to the MinerU (an open-source layout document parsing tool) layout parsing channel; Routing editable streaming documents such as Word, PPT, and Excel to the Docling (an open-source native document parsing tool) streaming parsing channel; Depending on the document quality, parsing target, or project configuration, you can choose to route it to the Docling streaming parsing channel for native parsing, or first convert it to PDF format and then route it to the MinerU layout parsing channel. S3: Format-specific parsing; The document is parsed using the corresponding parsing channel determined by the routing, and the intermediate results corresponding to the channel are obtained. S4: Unified chapter tree construction; The intermediate results output from each parsing channel are uniformly mapped to a chapter tree in Tree.json (a unified chapter tree data format). Each node in the chapter tree contains at least the fields of number, title, body text, child node, and page number. Child nodes may contain titles, body text, tables, images, and formula objects. S5: Confidence assessment; Using the chapter tree as the calculation object, the overall confidence level is calculated around three dimensions: chapter number continuity, table duplication risk, and title rationality. The data is then processed and distributed according to the confidence level and preset rules. S6: AI Agent (Intelligent Repair Agent) for local structural repair; For chapter trees that are redirected to AI for repair, the goal is to correct errors in the chapter tree structure. The entire document is not re-parsed or the original business content is not rewritten. Only key information about abnormal nodes, adjacent nodes before and after abnormal nodes, and parent candidate nodes is provided to the Agent. Sending the complete chapter tree at once is prohibited. Candidate repair suggestions are generated based on deterministic rules. The repair agent determines the final repair scheme based on the local context semantics. The structure of the chapter tree is repaired through structured patching. After the repair is completed, the confidence of the chapter tree is recalculated. The repaired chapter tree is saved only if the confidence is improved after the repair and the original field information is not destroyed. Otherwise, the repair result is discarded and the original chapter tree is retained. S7: Manual review; For chapter trees that are routed to manual review, a split-screen review interface is used. The left side displays the repaired chapter tree structure, and the right side displays the corresponding page of the original document. Manual adjustments are made, and the differences before and after the manual adjustments are recorded to form a structured correction log. The difference information is used for subsequent optimization of parsing channel routing rules, spatial attribution algorithms, Agent prompts, and quality assessment thresholds, forming a closed loop consisting of parsing, assessment, repair, review, and optimization.

2. The multi-channel complex document parsing method for the electronic information manufacturing industry according to claim 1, characterized in that: In step S3, the parsing process of the MinerU format parsing channel includes: (1) Perform layout analysis, text recognition, table recognition, and image region detection on the input document; (2) Generate page-level and block-level intermediate results, including page parsing information, block-level structure information, a content list organized in reading order, Markdown text, and chapter tree candidate structure.

3. The multi-channel complex document parsing method for the electronic information manufacturing industry according to claim 1, characterized in that: In step S3, the parsing process of the Docling streaming parsing channel includes: Extract headings, paragraphs, tables, images, text boxes, grouped objects, multilevel lists, slide objects, and metadata native objects from a document, and read the bounding box coordinates, layer order, style properties, and document hierarchy of each object; Also includes: Enhance the positioning step by reading the bounding box of the floating object I to be assigned to. Calculate the coordinates of the center point of the floating object. ; Iterate through all cells in table T. For the cell in row i and column j, retrieve its bounding box. To determine whether the center point of a floating object falls within the bounded box of the cell, use the following formula: ; like If the center point falls within the bounding box of the corresponding cell, the floating object will be attached as a subordinate object of the corresponding cell, and the relationship between the floating object and the corresponding table row and column position will be recorded in the chapter tree; if the center point does not fall within the bounding box of any cell, the floating object will be attached to the page, slide, paragraph area or chapter node where it is located. Before determining whether the center point falls within the bounding box of the cell, the upper and lower boundaries of the floating object and the cell can be normalized to eliminate the difference in the vertical coordinate direction between different analytical coordinate systems.

4. The multi-channel complex document parsing method for the electronic information manufacturing industry according to claim 1, characterized in that: In step S5, the formula for calculating the overall confidence level is: ; Among them, the , , The weights are assigned to three dimensions: continuity of chapter numbering, risk of table duplication, and reasonableness of title, with the sum of each weight being 1. The The score is based on the continuity of chapter numbers. For the duplicate risk score of the table, the Score for the reasonableness of the title; The calculation process for the continuity score of the chapter number includes: identifying the starting point of the text number, marking the nodes before the starting point as the preceding nodes, and performing continuity checks based only on the numbering of the text nodes. The continuity checks include checking for abnormal numbering formats, duplicate numbers, skipped numbers, hierarchical rollback errors, and abnormal starting points of sub-level numbers. The table duplication risk score is used to detect the risk of duplicate cells, duplicate rows, duplicate columns and abnormal duplication of long text in the document. The table duplication risk score is only used to indicate the risk of underlying parsing or OCR, and does not directly authorize AI to modify the content of table cells, and the content of table cells must not be automatically rewritten based on the score. The title rationality score is used to detect anomalies in documents such as empty titles, excessively long titles, titles mixed with body text, titles that appear to contain table fields, or titles containing too many numbers.

5. The multi-channel complex document parsing method for the electronic information manufacturing industry according to claim 1, characterized in that: In step S5, the traffic splitting rules for confidence assessment include: (1) If the confidence level of the chapter tree passes and the mandatory review condition is not triggered, the chapter tree is marked as automatically passed; (2) If the confidence level of the chapter tree is insufficient, or the numbering or title is abnormal and meets the preset repair conditions, the chapter tree will be marked as AI repair status; (3) If the chapter tree only has structural problems such as the risk of duplicate tables, abnormal chapter numbers, and abnormal titles, it will not enter the AI ​​repair process; (4) If the problem of the chapter tree cannot be safely repaired at the structural layer, or if it still fails the evaluation after repair, or if the mandatory review conditions are triggered, the chapter tree will be marked as a manual review status. The mandatory review conditions include: a concentrated occurrence of critical structural anomalies, failure to improve confidence after repair, and a structural score of the chapter tree that is lower than a preset threshold.

6. The multi-channel complex document parsing method for the electronic information manufacturing industry according to claim 1, characterized in that: In step S6, the Agent is only allowed to return structured patches, and is prohibited from returning natural language repair instructions or returning complete new trees; The structured patch includes at least one of node merging, node splitting, title splitting, title merging, and no need to modify the tag; At the same time, the Agent is prohibited from performing the following operations: rewriting table cell content, rewriting OCR-recognized content, rewriting the entire chapter tree, rearranging chapters extensively without local evidence, and discarding original text, images, tables, page numbers, or coordinate information.

7. A multi-channel complex document parsing system for the electronics and information manufacturing industry, characterized in that, include: (1) Type discrimination module; It is used to input and type-discriminate the document to be processed, and extract the document's file extension, text selectivity, image density, table density, object nesting degree, and whether there are slide objects or native table features; （ 2) Routing module; It is used to perform multi-parsing channel adaptive routing. Based on features and the parsing channel capability matrix, it makes routing decisions and routes PDF documents, scanned documents, and image documents to the MinerU (open-source format document parsing tool) format parsing channel. Routing editable streaming documents such as Word, PPT, and Excel to the Docling (an open-source native document parsing tool) streaming parsing channel; It is also used to route editable Office documents to the Docling streaming parsing channel for native parsing, or to convert them to PDF format first and then route them to the MinerU layout parsing channel, depending on the document quality, parsing target, or project configuration. (3) Parsing module; It includes the MinerU layout parsing unit and the Docling streaming parsing unit, which are used to parse the document through the corresponding parsing channel determined by the route and obtain the intermediate results corresponding to the channel; (4) Chapter tree construction module; This is used to uniformly map the intermediate results output from each parsing channel to a chapter tree in Tree.json (a unified chapter tree data format). Each node in the chapter tree must contain at least the fields of number, title, body text, child node, and page number. Child nodes may contain titles, body text, tables, images, and formula objects. (5) Confidence assessment module; It is used to calculate the overall confidence level based on the chapter tree, focusing on three dimensions: chapter number continuity, table duplication risk, and title rationality, and then process and distribute the data according to the confidence level and preset rules. (6) Repair module; Used for chapter trees that are redirected to AI for repair, with the goal of correcting chapter tree structure errors, without re-parse the entire document or rewrite the original business content, only providing the Agent with partial information on abnormal nodes, adjacent nodes before and after abnormal nodes, and parent candidate nodes, and prohibiting the sending of the complete chapter tree at once; Candidate repair suggestions are generated based on deterministic rules. The repair_agent determines the final repair scheme based on the local context semantics. The structure of the chapter tree is repaired through structured patching. After the repair is completed, the confidence of the chapter tree is recalculated. The repaired chapter tree is saved only if the confidence is improved after the repair and the original field information is not destroyed. Otherwise, the repair result is discarded and the original chapter tree is retained. (7) Review module; This is used to display a review interface for chapter trees that have been diverted to manual review. The interface includes the original document, the repaired chapter tree, the abnormal nodes, the abnormal tables, the abnormal titles and their page numbers. It receives manual adjustment operations and records the differences before and after the manual adjustment operations. This difference information is used for subsequent optimization of parsing channel routing rules, spatial attribution algorithms, Agent prompt words and quality assessment thresholds, forming a closed loop consisting of parsing, evaluation, repair, review and optimization.

8. The multi-channel complex document parsing system for the electronic information manufacturing industry according to claim 7, characterized in that: The MinerU format parsing unit is used for: (1) Perform layout analysis, text recognition, table recognition, and image region detection on the input document; (2) Generate page-level and block-level intermediate results, including page parsing information, block-level structure information, content list organized in reading order, Markdown text, and chapter tree candidate structure.

9. The multi-channel complex document parsing system for the electronic information manufacturing industry according to claim 7, characterized in that: The Docling streaming parsing unit is used to extract headings, paragraphs, tables, images, text boxes, grouped objects, multilevel lists, slide objects, and metadata native objects from a document, and read the bounding box coordinates, layer order, style attributes, and document hierarchy of each object; The Docling streaming parsing unit is also used to perform enhanced localization operations, as follows: Read the bounding box of the floating object I to be assigned Calculate the coordinates of the center point of the floating object. ; Iterate through all cells in table T. For the cell in row i and column j, retrieve its bounding box. To determine whether the center point of a floating object falls within the bounded box of the cell, use the following formula: ; like If the center point falls within the bounding box of the corresponding cell, the floating object will be attached as a subordinate object of the corresponding cell, and the relationship between the floating object and the corresponding table row and column position will be recorded in the chapter tree; if the center point does not fall within the bounding box of any cell, the floating object will be attached to the page, slide, paragraph area or chapter node where it is located. Before determining whether the center point falls within the bounding box of the cell, the upper and lower boundaries of the floating object and the cell can be normalized to eliminate the difference in the vertical coordinate direction between different analytical coordinate systems.

10. The multi-channel complex document parsing system for the electronic information manufacturing industry according to claim 7, characterized in that: In the repair module, the Agent is only allowed to return structured patches, and is prohibited from returning natural language repair instructions or returning a complete new tree; The structured patch includes at least one of node merging, node splitting, title splitting, title merging, and no need to modify the tag; Additionally, the fix module also prevents the Agent from performing the following operations: Rewrite table cell content, rewrite OCR-recognized content, rewrite the entire chapter tree, rearrange chapters on a large scale when there is no local evidence, and discard the original text, images, tables, page numbers, or coordinate information; The review module records the differences before and after the adjustment, including the anomaly type, repair method, node adjustment relationship, manual confirmation result, and change in confidence level.

Images