Document analysis inspection-free method and system, electronic equipment and storage medium

Abstract

The invention relates to the technical field of document processing, and discloses a document analysis inspection-free method and system, electronic equipment and a storage medium, and the method comprises the following steps: inputting a document, identifying layout elements, and adding element-level inspection-free marks to the layout elements meeting a preset credible condition; extracting text, image and path information, and recombining the text, image and path information into structured memory data according to a reading sequence; executing multi-source information fusion and rule matching, and adding a structure-level inspection-free mark to a chapter structure division result meeting a preset credible condition; attribute labeling: adding an attribute-level inspection-free mark to the text block meeting a preset credible condition; and executing post-processing operation, adding a processing-level inspection-free mark to a post-processing result meeting a preset credible condition, and storing the post-processing result as a standard structured document with the inspection-free mark. According to the method, identification errors caused by prejudice or limitation of a single model are avoided, the identification accuracy and the processing efficiency can be improved, manual intervention is reduced, and the automation level is improved.

Description

Technical Field

[0001] This invention belongs to the field of document processing technology, and in particular relates to a document parsing method, system, electronic device and storage medium that are exempt from inspection. Background Technology

[0002] Currently, with the rapid development of information technology, the number of digital documents is growing explosively. How to efficiently and accurately extract structured information from these massive amounts of documents has become a key problem that urgently needs to be solved in the field of information processing.

[0003] Traditional document processing techniques primarily rely on rule-based template matching or single artificial intelligence models. For example, in the layout analysis stage, early methods often employed heuristic rules, identifying elements such as headings, paragraphs, and charts by analyzing the document's geometric features (such as the position, size, and spacing of text blocks). However, this approach suffers from poor generalization ability and robustness when dealing with documents with complex layouts and diverse formats.

[0004] Because the recognition accuracy of a single model is limited, existing technologies often require significant manual intervention and verification of the results. For example, after document indexing is completed, the accuracy of each tag needs to be manually checked one by one, and elements with incorrect recognition need to be corrected, which is a huge workload. In addition, the manual verification process itself is subjective and inconsistent; different verifiers may have different standards, resulting in inconsistent quality of the final data.

[0005] Therefore, this application provides a document parsing exemption method to solve the above-mentioned technical problems. Summary of the Invention

[0006] The purpose of this invention is to provide a document parsing exemption method, system, electronic device and storage medium to solve the technical problem in the prior art of poor recognition accuracy and processing efficiency caused by recognition errors due to the bias or limitations of a single model.

[0007] To address the aforementioned technical problems, this invention provides a document parsing exemption method, comprising:

[0008] In response to the input document, the document page is analyzed in parallel using at least two models to identify layout elements and add element-level exemption marks to layout elements that meet preset confidence conditions.

[0009] Based on the layout elements with the aforementioned element-level inspection-free markers, text, images, and path information are extracted and reorganized into structured memory data according to the reading order;

[0010] In response to the structured memory data, multi-source information fusion and rule matching are performed, and a structure-level inspection-free mark is added to the chapter structure division results that meet the preset credibility conditions;

[0011] In response to the content tags of the chapter structure annotation with the structure-level exemption mark, attribute-level exemption mark is added to the text blocks that meet the preset credibility conditions;

[0012] In response to a text block with the attribute-level exemption mark, a post-processing operation is performed. Processing-level exemption marks are added to the post-processing results that meet the preset confidence conditions, and the results are stored as a standard structured document with multiple layers of exemption marks. The preset confidence conditions include that the consistency ratio of the output results of each model for the same object exceeds a preset consistency threshold, the confidence scores of each model are all higher than the dynamic confidence threshold, and the results are verified by the corresponding preset anomaly rule base.

[0013] In some specific embodiments, in response to the input document, the document page is analyzed in parallel using at least two models to identify layout elements, and element-level exemption marks are added to layout elements that meet preset confidence conditions, further including:

[0014] The multimodal pre-trained model and the object detection model are run in parallel, and the semantic type recognition results and spatial location recognition results of the layout elements are output respectively.

[0015] Heuristic rules are introduced to supplement and verify the semantic type recognition results and spatial location recognition results. The heuristic rules include header and footer position features, table of contents entry alignment features, and title font size features.

[0016] The semantic type recognition results and spatial location recognition results of the elements on the same page after supplementary verification are calculated. When the calculation results meet the preset credibility conditions, an element-level inspection exemption mark is added to the corresponding page element.

[0017] In some specific embodiments, in response to the structured memory data, multi-source information fusion and rule matching are performed, and a structure-level exemption mark is added to the chapter structure segmentation results that meet preset credibility conditions, further including:

[0018] Based on the structured memory data, the table of contents title metadata and the page title metadata are parsed.

[0019] In response to the metadata of table of contents titles and page titles, based on the table of contents parsing model and title association model, the text similarity between the table of contents titles and page titles is calculated, and the title formatting rule set is applied to verify the title layout features.

[0020] When the text similarity is greater than the preset similarity threshold, the layout feature verification is passed, and the boundaries are clear, a structure-level exemption mark is added.

[0021] In some specific embodiments, the preset trust condition further includes:

[0022] The table of contents contains records of article titles and starting page numbers;

[0023] The page number recognition results contain title text blocks with a similarity greater than the preset title threshold, wherein the title text blocks conform to the application title format rule set;

[0024] The end boundary of a chapter structure is determined by the starting position of the adjacent chapter structure.

[0025] In some specific embodiments, in response to the text structure annotation content tag with the structure-level exemption mark, an attribute-level exemption mark is added to the text block that meets the preset credibility condition, further including:

[0026] Content tagging is performed in parallel using rule-based models, natural language processing models, and large language models;

[0027] When the annotation result meets the preset credibility conditions, an attribute-level exemption mark is added to the corresponding text block, wherein the preset credibility conditions include verification through the text anomaly rule library.

[0028] In some specific embodiments, in response to a text block with the attribute-level exemption mark, a post-processing operation is performed, adding a processing-level exemption mark to the post-processing result that meets the preset trust conditions, and storing it as a standard structured document with multi-level exemption marks, further including:

[0029] The post-processing operations include performing metadata standardization, constructing chapter hierarchy trees, and chart captioning operations in parallel using multiple processing models.

[0030] By comparing the post-processing results output by different processing models, when the preset trust conditions are met, a processing-level inspection exemption mark is added and stored as a standard structured document with multiple layers of inspection exemption marks. The preset trust conditions include verification through a structural anomaly rule base.

[0031] In some specific embodiments, the construction of the chapter hierarchy tree further includes:

[0032] Parse text blocks with attribute-level exemption tags and parse the title number format to determine the explicit hierarchy;

[0033] Analyze the semantic content of the title to generate implicit logical relationships;

[0034] A chapter hierarchy tree structure is constructed by integrating explicit hierarchy and implicit logical relationships.

[0035] Based on the same concept, the present invention also provides a document parsing inspection-free system, comprising:

[0036] The element-level exemption tagging module is configured to respond to the input document by analyzing the document page in parallel using at least two models, identifying page elements, and adding element-level exemption tags to page elements that meet preset confidence conditions.

[0037] The structured memory data reorganization module is configured to extract text, images, and path information from page elements with the aforementioned element-level inspection-free markers, and reorganize them into structured memory data according to the reading order.

[0038] The structure-level inspection-free marking module is configured to respond to the structured memory data, perform multi-source information fusion and rule matching, and add structure-level inspection-free markings to the chapter structure division results that meet preset credibility conditions;

[0039] The attribute-level exemption tag module is configured to add attribute-level exemption tags to text blocks that meet preset credibility conditions in response to the text structure annotation content tags with the structure-level exemption tags.

[0040] A standard structured document storage module is configured to perform post-processing operations in response to text blocks with attribute-level exemption tags, add processing-level exemption tags to post-processing results that meet preset confidence conditions, and store them as standard structured documents with multi-layer exemption tags. The preset confidence conditions include that the consistency ratio of the output results of each model for the same object exceeds a preset consistency threshold, the confidence scores of each model are all higher than the dynamic confidence threshold, and they pass the verification of the corresponding preset anomaly rule base.

[0041] Based on the same concept, the present invention also provides an electronic device, including: a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of a document parsing exemption method.

[0042] Based on the same concept, the present invention also provides a computer-readable storage medium, characterized in that it stores a computer program executable by an electronic device, which, when run on the electronic device, causes the electronic device to perform the steps of a document parsing exemption method.

[0043] Compared with existing technologies, its advantages are as follows:

[0044] This invention discloses a document parsing exemption method, system, electronic device and storage medium, which avoids recognition errors caused by the bias or limitations of a single model, improves recognition accuracy and processing efficiency, reduces manual intervention and improves the level of automation. Attached Figure Description

[0045] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0046] Figure 1 This is a flowchart illustrating some specific embodiments of the document parsing exemption method of the present invention;

[0047] Figure 2 This is one of the flowcharts of a document parsing exemption method of the present invention in another specific embodiment;

[0048] Figure 3 This is a second flowchart illustrating another specific embodiment of the document parsing exemption method of the present invention;

[0049] Figure 4 This is a schematic diagram of the structure of a document parsing and inspection-free system according to some specific embodiments of the present invention;

[0050] Figure 5 This is a schematic diagram of the structure of an electronic device according to some specific embodiments of the present invention;

[0051] In the diagram, 710 is the processor; 720 is the memory; 730 is the input device; and 740 is the output device. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0053] The terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to limit the application. The singular forms “a,” “said,” and “the” used in the embodiments of this application and the appended claims are also intended to include the plural forms, and “multiple” generally includes at least two unless the context clearly indicates otherwise.

[0054] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0055] It should be understood that although the terms first, second, third, etc., may be used in the embodiments of this application, these descriptions should not be limited to these terms. These terms are only used to distinguish the descriptions. For example, first may also be referred to as second without departing from the scope of the embodiments of this application, and similarly, second may also be referred to as first.

[0056] Depending on the context, the words “if” or “suppose” as used here can be interpreted as “when” or “in response to determination” or “in response to detection.” Similarly, depending on the context, the phrases “if determination” or “if detection (of the stated condition or event)” can be interpreted as “when determination” or “in response to determination” or “when detection (of the stated condition or event)” or “in response to detection (of the stated condition or event).”

[0057] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that an article or device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such an article or device. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or device that includes said element.

[0058] It should be noted that any symbols and / or numbers present in the specification that are not marked in the accompanying drawings are not reference numerals.

[0059] Reference Figure 1 A document parsing exemption method includes:

[0060] S101, in response to the input document, analyze the document page in parallel using at least two models, identify page elements, and add element-level exemption marks to page elements that meet preset confidence conditions;

[0061] S102, Based on the layout elements with the aforementioned element-level inspection-free markers, extract text, images, and path information, and reorganize them into structured memory data according to the reading order;

[0062] S103, in response to the structured memory data, perform multi-source information fusion and rule matching, and add a structure-level inspection-free mark to the chapter structure division results that meet the preset credibility conditions;

[0063] S104, in response to the chapter structure annotation content tag with the structure-level exemption mark, add attribute-level exemption mark to the text block that meets the preset credibility condition;

[0064] S105, in response to the text block with the attribute-level exemption mark, perform post-processing operation, add processing-level exemption mark to the post-processing result that meets the preset confidence conditions, and store it as a standard structured document with multi-layer exemption marks. The preset confidence conditions include that the consistency ratio of the output results of each model for the same object exceeds the preset consistency threshold, the confidence score of each model is higher than the dynamic confidence threshold, and the corresponding preset anomaly rule base is verified.

[0065] Specifically, in this embodiment of the invention, in response to the input PDF document, the page is processed in parallel by a visual positioning model and a multimodal understanding model: the visual positioning model outputs the bounding box coordinates of images / formulas / tables, and the multimodal understanding model outputs semantic tags for paragraphs / titles / tables; when the two models have completely consistent type judgments for the same element, the confidence levels are both higher than the dynamic threshold, and the layout anomaly rule base is verified (verifying element overlap rate and the rationality of header position), an element-level exemption mark is added. Based on the page elements with this mark, text content, image pixel data, and vector paths are extracted and reorganized into structured memory data in a reading order from left to right and from top to bottom. In response to this data, the title-page number mapping and position metadata of the title text block are extracted from the table of contents text block; the table of contents structure is analyzed by the table of contents parsing model, the title association model calculates the semantic similarity between the table of contents title and the page title, and the title format rule base is used to verify the title font size and centering feature; when the similarity exceeds a preset ratio, the format conforms to the rules, and the page numbers are consecutive, a structure-level exemption mark is added to the chapter division. In response to the marked text structure, text fragments are input into a pre-trained language model and a rule inference engine: the language model outputs semantic type labels (e.g., first-level heading / body text), and the rule engine generates labels based on a heading prefix dictionary; when the two models' annotations are consistent, the confidence score is higher than the threshold, and the semantic conflict rule base verification is passed, an attribute-level exemption tag is added. In response to the marked text block, the metadata conversion model standardizes the author / date format, and the structure generation model parses the heading number and semantic relationship to construct a chapter tree; when the chapter levels output by the two models are completely consistent, the confidence score meets the standard, and the structure rule base verification is passed, a processing-level exemption tag is added, and the document is stored as an XML document with four layers of exemption tags.

[0066] For example, given an A4-sized PDF document, the visual positioning model identifies the area on page 1 [coordinates (120,80)-(400,600)] as a table (confidence 0.92), while the multimodal model also identifies this area as a table (confidence 0.95). Both models are of the same type, with confidence levels > 0.9. The layout rule base verifies that the distance between this area and the header is > 50mm (meeting the ≥10mm rule), and an element-level exemption mark is added. Based on the marked table, the text "Experimental Data Comparison" is extracted from the cells and reorganized according to the reading order. ...Memory structure. The table of contents text block "Chapter 1 Introduction......1" is parsed. The table of contents model extracts the title "Introduction" and page number "1". The association model calculates the similarity between the table of contents title "Introduction" and the title on page 1 (cosine similarity 0.89 > 0.85). The format rule base verifies that the title body is 24pt (body text 20pt, meeting the ≥22pt requirement), and the page numbers are continuous without breaks, adding a structural-level exemption mark. For the tagged "Introduction" text block, the language model labels it as "Level 1 Heading" (confidence 0.88), and the rule engine labels it as "Level 1 Heading" based on the "Chapter X" prefix (confidence 0.90). The results of the two models are consistent and the confidence is > 0.85. The semantic rule base verification shows no conflict, adding an attribute-level exemption mark. For the tagged text block "Author: Zhang San", the metadata model normalizes it to... <author> Zhang San< / author> (Confidence level 0.96); The structural model identifies "1.1 Experimental Methods" as a child node of "1. Introduction" (confidence level 0.97); The outputs of the two models are consistent and the confidence level is >0.95, the structural rule base validation level depth is 2 (compliant with the ≤3 rule), a processing-level exemption mark is added, and XML tags are generated.<sectionlevel="1"> <title> introduction< / title> ...

[0067] In some applications, in response to an input document, the document page is analyzed in parallel using at least two models to identify layout elements. Element-level exemption tags are added to layout elements that meet preset credibility conditions. This includes running a multimodal pre-trained model and an object detection model in parallel, outputting semantic type identification results and spatial location identification results for layout elements, respectively. Heuristic rules are introduced to supplement and verify the semantic type identification results and spatial location identification results. These heuristic rules include header and footer position features, table of contents alignment features, and title font size features. The semantic type identification results and spatial location identification results for each layout element after supplementary verification are calculated. When the calculation results meet the preset credibility conditions, element-level exemption tags are added to the corresponding layout elements.

[0068] Understandably, in response to the input PDF document, the document page is processed in parallel by a multimodal pre-trained model and an object detection model: the multimodal pre-trained model outputs semantic type recognition results of page elements (such as paragraphs / headings / tables) based on text semantics and layout relationships, while the object detection model outputs spatial location recognition results (bounding box coordinates) of elements based on visual features; heuristic rules are introduced to supplement and verify the two types of results, including header and footer position features to verify whether the elements are located in the preset areas at the top / bottom of the page, table of contents item alignment features to verify whether the items meet the rules of left alignment and right alignment of page numbers, and title font size features to verify whether the font size is significantly larger than the body text; a confidence calculation is performed on the semantic type and spatial location results after supplementary verification. When the proportion of consistent type judgments of the same element by multiple models exceeds a threshold, the confidence of each model is higher than the dynamic standard, and the heuristic rule verification is passed, an element-level exemption mark is added to the corresponding page element.

[0069] For example, input an A4-sized (210mm×297mm) academic paper PDF; the multimodal model identifies the region on page 2 [coordinates (30,40)-(180,60)] as "chapter title" (confidence 0.93); the object detection model outputs the bounding box coordinates of this region as (32,42)-(178,58) (confidence 0.91);

[0070] Heuristic rule supplementary verification: Header position features: Y coordinate of the region 40-60mm (top region ≤80mm, which meets the requirements); Title font features: detected font size 16pt (body text 12pt, which meets the ≥14pt rule); Calculation results: the two models have consistent type judgment (both are identified as titles), and the average confidence score is 0.92 > 0.9 threshold; add element-level exemption mark.

[0071] In some applications, in response to the structured memory data, multi-source information fusion and rule matching are performed, and a structure-level exemption mark is added to the chapter structure division results that meet the preset credibility conditions. This includes parsing the table of contents title metadata and page title metadata based on the structured memory data; in response to the table of contents title metadata and page title metadata, calculating the text similarity between the table of contents title and page title based on the table of contents parsing model and the title association model, and applying the title format rule set to verify the title layout features; when the text similarity is greater than the preset similarity threshold, the layout feature verification is passed and the boundary is clear, a structure-level exemption mark is added.

[0072] Understandably, in response to structured in-memory data, the table of contents text blocks with element-level exemption tags are parsed to obtain table of contents title metadata (including title text and starting page number), while the metadata of title text blocks with element-level exemption tags in the layout (including text content and printed page number) is extracted. The table of contents structure is analyzed through the table of contents parsing model to generate a title-page number mapping table, and the title association model calculates the text similarity between the table of contents title and the layout title. The title formatting rule set is applied to verify the title layout features, including whether the font size exceeds the preset proportion threshold of the body text, whether it meets the centering tolerance range, and whether it is located in the preset top area of ​​the page. When the text similarity exceeds the preset similarity threshold, all layout features are verified, and the chapter boundaries are clearly defined by the positions of adjacent titles, a structure-level exemption tag is added to the chapter division result.

[0073] For example, when processing a conference proceedings PDF document, the table of contents metadata obtained from parsing the structured in-memory data contains "Distributed System Optimization......5", and the page title metadata is the text block "Distributed System Optimization Design" on page 5. The table of contents parsing model extracts the table of contents title as "Distributed System Optimization" and the starting page number as 5. The title association model calculates the cosine similarity between the two texts to be 0.92. The title formatting rule set is verified as follows: font size 18pt (body text 12pt, meeting the ≥15pt requirement), horizontal center offset 2% (meeting the ≤5% tolerance), and vertical position 10mm from the top of the page (meeting the ≤30mm area). Boundary clarity is verified: the next chapter starts on page 8. Because the similarity of 0.92 > 0.85 threshold, the layout verification is all passed, and the boundaries are clear, a structure-level exemption mark is added to this chapter. The mark information includes the table of contents title, page title, starting and ending page numbers, and verification timestamp.

[0074] In some applications, the preset credibility conditions include the existence of article titles and starting page number records in the table of contents; the existence of article title text blocks with a similarity greater than a preset article title threshold in the page recognition results of the corresponding page number, wherein the article title text blocks conform to the application title format rule set; and the end boundary of the chapter structure is determined by the starting position of the adjacent chapter structure.

[0075] Understandably, the process involves three steps: First, verifying the existence of valid article titles and starting page numbers in the table of contents data; second, verifying that in the layout recognition results for the specified page number in the table of contents, there exists an article title text block with a similarity exceeding a preset article title threshold, and that this text block passes the title formatting rule set validation, including a font size significantly larger than the body text baseline, a horizontal position meeting the centering tolerance range, and a vertical position located in the preset top area of ​​the page; finally, verifying that the end boundary of the chapter structure is accurately determined by the starting position of adjacent chapters, requiring that the end page number of the current chapter equals the starting page number of the next chapter minus one, and that the page numbers are continuous without any breaks. All three conditions must be met simultaneously for the passage to be considered credible.

[0076] For example, when processing a technical manual PDF, the table of contents contains a valid record with the title "Dynamic Load Balancing Algorithm" and the starting page number "17". In the layout recognition results of the corresponding page 17, the title text block "Research on Dynamic Load Balancing Algorithm" is detected, and the similarity with the table of contents title is calculated to be 0.91 (preset threshold 0.85). Title format rule set verification: font size 22pt (body text 16pt, meeting the ≥20pt requirement), horizontal center offset 3% (tolerance ≤5%), vertical position 15mm from the top of the page (≤30mm area); boundary determination: the next chapter "Performance Test Results" starts on page 23, so the end page number of the current chapter is confirmed to be 22 (23-1), and the page number sequence 17-22 is continuous and without missing pages. Since all three conditions are met, a structure-level exemption mark is added to this chapter, and the mark field includes the starting page number 17, the ending page number 22, the similarity value 0.91, and the verification identifier.

[0077] In some applications, in response to the text structure annotation content tags with the structure-level exemption tags, attribute-level exemption tags are added to text blocks that meet preset credibility conditions, including parallel annotation of content tags based on rule models, natural language processing models, and large language models; when the annotation results meet the preset credibility conditions, attribute-level exemption tags are added to the corresponding text blocks, wherein the preset credibility conditions include verification through a text anomaly rule base.

[0078] Understandably, in response to the text structure with structural-level exemption tags, text block content tags are annotated in parallel using rule models, natural language processing models, and language models: the rule model generates tags based on a predefined title prefix dictionary and caption keyword library, the natural language processing model outputs tags through semantic sequence analysis, and the language model predicts tags based on contextual semantic understanding; when at least two models have completely consistent annotation results for the same text block, the confidence of each model is higher than the dynamic threshold, and the text anomaly rule library verification is passed, an attribute-level exemption tag is added; the text anomaly rule library verification includes title-level conflict detection (such as a first-level title appearing in the body text area), caption semantic integrity verification (such as missing numbers in figure captions), and font logic contradiction checking (such as the title font size being smaller than the body text).

[0079] For example, when processing Chapter 3 of a research paper, the text block "3.2 Experimental Apparatus" with structural-level exemption tags was annotated in parallel: the rule model labeled it as a second-level heading based on the "XX" prefix (confidence 0.87), the natural language processing model identified it as a second-level heading through chapter numbering patterns (confidence 0.89), and the language model parsed it as a second-level heading based on contextual semantics (confidence 0.92); the rule model and the language model results were consistent (second-level heading), with a mean confidence of 0.90, which was higher than the threshold of 0.85; text anomaly rule base verification: font size 18pt (body text 12pt, meeting the requirement of heading > 15pt), position 45mm from the top of the page (non-body text area), no heading level conflict; because it met the consistency requirements, the confidence level was up to standard, and the anomaly verification passed, an attribute-level exemption tag was added, and the tag type, mean confidence level, and verification identifier were recorded.

[0080] In some applications, in response to a text block with the attribute-level exemption mark, a post-processing operation is performed. A processing-level exemption mark is added to the post-processing result that meets preset trust conditions, and the result is stored as a standard structured document with multi-level exemption marks. The post-processing operation includes using multiple processing models to perform metadata standardization, constructing a chapter hierarchy tree, and adding chart captions in parallel. The post-processing results output by different processing models are compared, and when the preset trust conditions are met, a processing-level exemption mark is added, and the result is stored as a standard structured document with multi-level exemption marks. The preset trust conditions include verification through a structural anomaly rule base.

[0081] Understandably, in response to text blocks with attribute-level exemption tags, post-processing operations are performed in parallel through a metadata conversion model and a structure generation model: the metadata conversion model standardizes the author, institution, and date formats (e.g., unifying the rules for abbreviating institution names), and the structure generation model parses the title hierarchy relationships to construct a chapter tree and associates figures and captions; the output results of each model are compared, and when the processing results of the two models for the metadata format or chapter tree node relationships are completely consistent, the confidence level is higher than the dynamic threshold, and the structure anomaly rule base verification is passed, a processing-level exemption tag is added; the structure anomaly rule base verification includes chapter hierarchy depth conflicts (e.g., child nodes missing parent nodes), abnormal continuity of figure numbers (e.g., figure D directly following figure A), and metadata logical contradictions (e.g., publication date is earlier than writing date); finally, it is stored as an XML structured document with four levels of exemption tags: element-level, structure-level, attribute-level, and processing-level.

[0082] For example, when processing a technical report, for the author text block "Zhang Ming (Peking University)" with attribute-level exemption tags: the metadata transformation model is standardized as follows:<author affiliation="Peking University"> Zhang Ming (confidence level 0.96); For the title block "4.1 Performance Comparison": the structure generation model identifies it as a child node of "4.0" (confidence level 0.94); the chart association operation links "Figure X: Experimental Results" to the text reference "See Figure X" (confidence level 0.97); comparison of the output of each model: the metadata format is consistent, and the chapter tree node relationship is completely matched; structure anomaly rule base verification: the level depth = 2 (complies with the ≤3 rule), the chart sequence 2→3 is continuous, and the publication date March 2025 is later than the writing date January 2025; due to the consistency, confidence level (both > 0.95 threshold) and rule verification passing, a processing-level inspection exemption mark is added, stored as an XML document and the fourth-level inspection exemption mark and confidence level data are recorded.

[0083] In some of these applications, the construction of the chapter hierarchy tree includes parsing text blocks with attribute-level exemption tags, parsing the title number format to determine the explicit hierarchy, analyzing the semantic content of the titles to generate implicit logical relationships, and integrating the explicit hierarchy and implicit logical relationships to construct the chapter hierarchy tree structure.

[0084] Understandably, constructing a chapter hierarchy tree involves parsing title text blocks with attribute-level exemption tags: First, an explicit hierarchy identifier (such as numerical numbers like "1.2.3") in the title is identified through a numbering rule engine, and the hierarchy depth and parent node index are extracted; second, implicit logical relationships are generated by analyzing the semantic content of the title through a language model (such as the subordinate relationship between "experimental methods" and "results analysis"), and semantic similarity and logical dependency strength are calculated; finally, the explicit numbering hierarchy and implicit semantic relationships are integrated to construct a tree structure: when the explicit numbering is complete, the numbering hierarchy is used first; when the explicit numbering is missing, the hierarchy is inferred from the implicit relationship; if the two conflict, conflict resolution rules are activated (such as the explicit numbering priority principle).

[0085] For example, when processing a technical manual, parsing text blocks with attribute-level exemption tags: Explicit hierarchical parsing: Title "2.1.3 Temperature Calibration" → extracting three levels (2→1→3), with the parent node being "2.1"; Implicit relationship generation: Language model analysis of the title "Error Analysis": Semantic similarity with "Experimental Results" is 0.88 (strong correlation); similarity with "Experimental Apparatus" is 0.45 (weak correlation); inferred to be a child node of "Experimental Results"; Fusion construction: "2.1.3" is explicitly mounted under "2.1" (depth = 3); Unnumbered "Error Analysis" is implicitly mounted under "Experimental Results" (depth = 2); Conflict handling: The title "3.1 Data Processing" explicitly numbers the parent node as "3.0", but semantic analysis points to "2.4" (similarity 0.92), triggering the explicit priority principle, maintaining the child node relationship of "3.0". The spanning tree structure contains 4 layers and 26 nodes. After verification by the structure rule base, the layer depth is ≤4 and the parent-child distance of the nodes is compliant, and a processing-level inspection exemption mark is added.

[0086] The following is combined with Figure 2 and Figure 3 Another embodiment of the document parsing exemption method of the present invention is described below:

[0087] like Figure 2 and Figure 3 As shown, this embodiment includes:

[0088] 1. Input document: Receives the formatted document to be processed, such as PDF.

[0089] 2. Multi-engine Layout Recognition: Employing multiple deep learning models and heuristics, this function analyzes document pages to identify layout elements such as tables of contents, headings, paragraphs, charts, and formulas. At this stage, highly reliable layout element blocks are marked based on the consistency and confidence levels of the outputs from multiple models, as well as whether they pass the "abnormal pattern matching" check, thus achieving "exemption from inspection."

[0090] 3. Document parsing: Based on the layout recognition results, extract specific content such as characters, paths, and images from the document, and reorganize them according to the reading order to form structured data in memory.

[0091] 4. Multi-engine document segmentation: Combining the results of directory recognition and the article title information from page layout analysis, the document is divided into chapters. Through comparison verification and rule feature analysis, highly reliable chapter structures are identified, achieving "check-free" document segmentation.

[0092] 5. Multi-engine Attribute Annotation: This task transforms document content into attribute annotation, utilizing various pre-trained deep learning models for semantic understanding to annotate text blocks, assigning them labels such as metadata, chapters, and chart captions. Similarly, highly reliable text blocks are labeled based on the consistency and confidence of the multi-model results, as well as whether they pass the "abnormal pattern matching" check.

[0093] 6. Multi-engine post-processing: The annotation results are optimized and standardized, including metadata standardization, construction of chapter hierarchy trees, related diagrams and captions, etc. This stage also adopts multi-engine collaboration and inspection-free mechanisms to ensure the accuracy of the post-processing results.

[0094] 7. Storage: Save the final processed, highly structured content in standard formats such as XML for easy retrieval, analysis, and utilization.

[0095] In this embodiment, multiple deep learning models are deployed simultaneously during the layout recognition stage, such as a multimodal model based on the Transformer architecture and a CV model based on object detection. Furthermore, heuristic methods are integrated to handle simple elements with obvious formatting characteristics. The same document page is processed in parallel, with each engine outputting its recognition results and confidence scores for the layout elements. The outputs of multiple engines are compared. If multiple engines achieve identical recognition results for the same text block (e.g., "title"), and their respective confidence scores are all above a preset threshold, then the text block is marked as a "highly reliable" or "inspection-free" element, requiring no manual verification and significantly saving processing time. After the layout recognition stage is completed, the precise location and type information of each element in the document are obtained. Based on this information, the corresponding content is extracted from the original layout document. For example, for a text block identified as a "paragraph," all characters within that area are extracted and reorganized according to the correct reading order. For elements identified as "figures" or "tables," the corresponding paths or image data are extracted. For formulas, both inline and cross-line formulas are extracted. The parsed content is organized into a temporary memory structure. This embodiment's article segmentation process also employs a multi-engine collaborative and inspection-free mechanism. First, using the table of contents information extracted during the layout recognition stage, metadata such as the article title, author, and page number of each article is obtained. Simultaneously, all text blocks identified as "article titles" in the layout are analyzed. Then, through multi-engine collaboration, the article titles in the table of contents are compared and verified with the actual article titles appearing in the layout. If the article title in the table of contents highly matches the content of a certain article title text block in the layout, and the position and format of that text block conform to preset article title rules (such as largest font size, centered, and at the top of the page), the starting position of that article is determined with high confidence. By analyzing the continuity of page numbers and the starting position of the next article, the end of each article can be accurately identified. For documents with missing or unclear table of contents information, article segmentation can be performed solely based on the rule characteristics of the article title elements in the layout. When the confidence score of a segment (determined by the matching score and rule compliance of multiple engines) exceeds a preset threshold, the segmentation result is marked as "exempt from inspection," thus avoiding the tedious work of manually checking the segmentation boundaries page by page. In the content indexing stage, the complex document understanding task is transformed into an attribute annotation task, segmenting the parsed document content (such as paragraphs, titles, figure captions, etc.) into text fragments. Then, multiple deep learning pre-trained models oriented towards semantic understanding are used to annotate these texts. Post-processing also employs a multi-engine collaboration and exemption mechanism to further optimize and standardize the attribute annotation results. Post-processing tasks include several aspects: First, the extracted metadata (such as author affiliation, keywords, etc.) is standardized, for example, by unifying the format of author names.Secondly, based on the hierarchical relationship of headings (such as first-level headings, second-level headings, etc.), the document's chapter hierarchy tree is automatically constructed, clearly reflecting the document's logical structure. Thirdly, the system analyzes the quoted words in the chart captions (such as "..."). Figure 1 The system automatically establishes associations between captions and corresponding figures / tables based on their numbers (Table 2, etc.). Finally, it precisely identifies and analyzes superscripts and subscripts in the text. These post-processing tasks are completed collaboratively by multiple specialized engines. For example, when building the chapter tree, a rule engine (to determine the numbering format of headings) and an LLM (to understand the semantic hierarchy of headings) can be combined. When the post-processing results from multiple engines are consistent, have high confidence, and pass the "abnormal pattern matching" check, the result is marked as "inspection-free" and directly output, thus ensuring the integrity and accuracy of the final structured data. After multi-engine collaborative processing and inspection-free filtering, the document content has been highly structured and refined. During the storage phase, these processing results are saved as standard XML documents according to a preset XML Schema definition.

[0096] This embodiment employs a modular design, primarily comprising an input interface, a multi-engine processing core, an exemption decision module, and a storage module. The input interface receives documents in various formats. The multi-engine processing core includes several specialized processing modules, such as a layout recognition module, a section segmentation module, an attribute annotation module, and a post-processing module, each integrating multiple heterogeneous engines. The exemption decision module runs throughout the entire processing flow, evaluating the multi-engine outputs at each key node and determining whether a data is exempt from inspection. Finally, non-exempt data undergoes manual correction. During the layout recognition stage, both a multimodal model and a CV model are deployed simultaneously. By fusing the outputs of these two models, a complete layout analysis result containing both semantic and precise location information can be obtained. For example, the multimodal model might identify a region as a "table," while the CV model provides the precise bounding box coordinates of that table; combining the two enables accurate table recognition and localization. In addition to deep learning models, heuristic rules are also included to assist in layout recognition. These rules are based on observations of common document layouts. For example, headers and footers are usually located in fixed positions on the page, table of contents entries have specific alignment and page number characteristics, and headings are usually in larger font than body text. These rules can serve as effective supplements and validations when the model is uncertain about the recognition results of certain elements. For example, if a text block is identified as "body text" by the model, but its position is at the top of the page and the font is small, heuristic rules might relabel it as a "header" based on its positional characteristics. Each deep learning model includes a confidence score with its output, indicating its degree of certainty about the result. The recognition results of multiple models for the same page element are compared. If multiple models completely agree on the type and position of the element, and their confidence scores are all higher than a preset threshold (e.g., 0.95), the element is considered "highly reliable," and the model output passes the "abnormal pattern matching" verification, thus being marked as "exempt from inspection." The document parsing module extracts specific content from the original document based on the page layout recognition results. For text areas, a character extraction engine extracts information such as text, font, and font size from the PDF. For image and vector graphics regions, pixel or path data is extracted directly. When processing multi-column formatted documents, the reading order of the text is recalculated based on the column information identified by the layout, ensuring that the extracted text flow conforms to human reading habits. All extracted content, along with its layout information (such as element type, chapter, etc.), is organized into a hierarchical memory object, providing structured data input for subsequent processing steps. In the multi-engine article segmentation stage, the document's table of contents is first identified and parsed to extract the title, author, and starting page number of each article. Then, all text blocks identified as "article titles" are searched in the entire document's layout recognition results. A comparison algorithm is used to match the article titles in the table of contents with the actual article titles appearing in the main text.The matching criteria include not only the similarity of text content but also auxiliary information such as author information and page numbers. If the information in the table of contents highly matches the information in the layout, the starting position of an article can be determined with high confidence. In addition to table of contents information, the regular features of article titles are also used to assist in article segmentation. The format of all text blocks in the layout is analyzed, and text blocks that conform to these regular features are identified and used as potential article titles. By combining table of contents comparison and regular feature analysis, a comprehensive understanding of the document's chapter structure can be constructed, and the confidence level of each chapter segmentation can be calculated. In the attribute annotation stage, three different types of annotation methods are used simultaneously: heuristic rules, NLP, and LLM. By annotating the same text fragment using these three methods, annotation results can be obtained from different perspectives, and the accuracy of the final decision can be improved by comparing these results. In the post-processing stage, the extracted metadata (such as author affiliation, date, keywords, etc.) is standardized. Utilizing the knowledge base and text generation capabilities of a large language model (LLM), this metadata is converted into a unified and standardized format. For example, keywords are separated by double semicolons, and different date representations are standardized to the format "YYYY-MM-DD". Based on the identified headings at each level, a hierarchical tree of the document's chapters is automatically constructed. The rule engine determines the level by analyzing the heading numbering format (e.g., "1", "1.1", "1.1.1"). The LLM, on the other hand, determines the logical hierarchy by understanding the semantic content of the headings. For example, LLM can determine that "Experimental Methods" is a parent chapter of "Results and Discussion". By combining these two methods, the system can construct a chapter tree that accurately reflects the document's logical structure. The post-processing module is also responsible for establishing the relationship between figures and their captions. This is achieved by analyzing quotations in the caption text (e.g., "..."). Figure 1 The system automatically matches the numbers of tables ("Table 2") and charts. It also performs precise analysis and recognition of superscripts and subscripts in the text. These post-processing tasks are also completed collaboratively by multiple engines, and their accuracy is ensured through a check-free mechanism. During the storage phase, the fully processed and validated structured data is saved as a standard XML document according to a predefined XML Schema (XSD).

[0097] For the purpose of simplicity, the method steps disclosed in the above embodiments are described as a series of actions. However, those skilled in the art should understand that the embodiments of the present invention are not limited to the described order of actions, because according to the embodiments of the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily essential to the embodiments of the present invention.

[0098] like Figure 4As shown, the present invention also provides a document parsing inspection-free system, comprising:

[0099] The element-level inspection-free marking module 201 is configured to respond to the input document by analyzing the document page in parallel using at least two models, identifying page elements, and adding element-level inspection-free markings to page elements that meet preset credibility conditions.

[0100] The structured memory data reorganization module 202 is configured to extract text, images, and path information based on the page elements with the element-level inspection-free marks, and reorganize them into structured memory data according to the reading order;

[0101] The structure-level inspection-free marking module 203 is configured to respond to the structured memory data, perform multi-source information fusion and rule matching, and add structure-level inspection-free markings to the chapter structure division results that meet preset credibility conditions;

[0102] The attribute-level exemption tag module 204 is configured to add attribute-level exemption tags to text blocks that meet preset credibility conditions in response to the chapter structure annotation content tags with the structure-level exemption tags.

[0103] The standard structured document storage module 205 is configured to perform post-processing operations in response to text blocks with attribute-level exemption tags, add processing-level exemption tags to post-processing results that meet preset confidence conditions, and store them as standard structured documents with multi-layer exemption tags. The preset confidence conditions include that the consistency ratio of the output results of each model for the same object exceeds a preset consistency threshold, the confidence scores of each model are all higher than the dynamic confidence threshold, and they pass the corresponding preset anomaly rule base verification.

[0104] It is worth noting that although only some basic functional modules are disclosed in the embodiments of this invention, it does not mean that the composition of this system is limited to the above-mentioned basic functional modules. On the contrary, what this embodiment intends to express is that, based on the above-mentioned basic functional modules, those skilled in the art can arbitrarily add one or more functional modules in combination with existing technology to form an infinite number of embodiments or technical solutions. That is to say, this system is open rather than closed. The fact that this embodiment only discloses a few basic functional modules should not be considered as the scope of protection of the claims of this invention being limited to the disclosed basic functional modules. At the same time, for the convenience of description, the above device is described separately according to its functions as various units and modules. Of course, in implementing this invention, the functions of each unit and module can be implemented in one or more software and / or hardware.

[0105] like Figure 5As shown, the present invention also provides an electronic device, including: a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of a document parsing exemption method.

[0106] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. For example... Figure 5 The structure shown in this embodiment of the invention includes an electronic device comprising one or more processors 710 and a memory 720; the processors 710 in this electronic device may be one or more. Figure 5 Taking a processor 710 as an example; a memory 720 is used to store one or more programs; the one or more programs are executed by the one or more processors 710, so that the one or more processors 710 implement a document parsing exemption method as described in any one of the embodiments of the present invention.

[0107] The electronic device may also include an input device 730 and an output device 740.

[0108] The processor 710, memory 720, input device 730, and output device 740 in this electronic device can be connected via a bus or other means. Figure 5 Taking the example of a connection between China and Israel via a bus.

[0109] The memory 720 in this electronic device serves as a computer-readable storage medium, capable of storing one or more programs. These programs can be software programs, computer-executable programs, or modules, such as the program instructions / modules corresponding to the document parsing exemption method provided in this embodiment of the invention. The processor 710 executes various functional applications and data processing of the electronic device by running the software programs, instructions, and modules stored in the memory 720, thereby implementing the document parsing exemption method described in the above embodiment.

[0110] The memory 720 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the electronic device. Furthermore, the memory 720 may include high-speed random access memory and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some instances, the memory 720 may further include memory remotely located relative to the processor 710, which can be connected to the device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0111] Input device 730 can be used to receive input digital or character information, and to generate key signal inputs related to user settings and function control of the electronic device. Output device 740 may include display devices such as a display screen.

[0112] The present invention also provides a computer-readable storage medium storing a computer program executable by an electronic device, which, when run on the electronic device, causes the electronic device to perform the steps of a document parsing exemption method.

[0113] Specifically, the computer storage medium in this embodiment of the invention can be any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. For example, a computer-readable storage medium can be—but is not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this embodiment, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0114] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A document parsing exemption method, characterized in that, include: In response to the input document, the document page is analyzed in parallel using at least two models to identify layout elements and add element-level exemption marks to layout elements that meet preset confidence conditions. Based on the layout elements with the aforementioned element-level inspection-free markers, text, images, and path information are extracted and reorganized into structured memory data according to the reading order; In response to the structured memory data, multi-source information fusion and rule matching are performed, and a structure-level inspection-free mark is added to the chapter structure division results that meet the preset credibility conditions; In response to the content tags of the chapter structure annotation with the structure-level exemption mark, attribute-level exemption mark is added to the text blocks that meet the preset credibility conditions; In response to a text block with the attribute-level exemption mark, a post-processing operation is performed. Processing-level exemption marks are added to the post-processing results that meet the preset confidence conditions, and the results are stored as a standard structured document with multiple layers of exemption marks. The preset confidence conditions include that the consistency ratio of the output results of each model for the same object exceeds a preset consistency threshold, the confidence scores of each model are all higher than the dynamic confidence threshold, and the results are verified by the corresponding preset anomaly rule base.

2. The document parsing exemption method according to claim 1, characterized in that, In response to the input document, the document page is analyzed in parallel using at least two models to identify layout elements. Element-level exemption markers are added to layout elements that meet preset confidence criteria. Further steps include: The multimodal pre-trained model and the object detection model are run in parallel, and the semantic type recognition results and spatial location recognition results of the layout elements are output respectively. Heuristic rules are introduced to supplement and verify the semantic type recognition results and spatial location recognition results. The heuristic rules include header and footer position features, table of contents entry alignment features, and title font size features. The semantic type recognition results and spatial location recognition results of the elements on the same page after supplementary verification are calculated. When the calculation results meet the preset credibility conditions, an element-level inspection exemption mark is added to the corresponding page element.

3. The document parsing exemption method according to claim 1, characterized in that, In response to the structured memory data, multi-source information fusion and rule matching are performed, and a structure-level exemption mark is added to the chapter structure segmentation results that meet the preset credibility conditions, further including: Based on the structured memory data, the metadata of the table of contents titles and the metadata of the page titles are parsed. In response to the metadata of table of contents titles and page titles, based on the table of contents parsing model and title association model, the text similarity between the table of contents titles and page titles is calculated, and the title formatting rule set is applied to verify the title layout features. When the text similarity is greater than the preset similarity threshold, the layout feature verification is passed, and the boundaries are clear, a structure-level exemption mark is added.

4. The document parsing exemption method according to claim 3, characterized in that, The preset trust conditions further include: The table of contents contains records of article titles and starting page numbers; The page number recognition results contain title text blocks with a similarity greater than the preset title threshold, wherein the title text blocks conform to the application title format rule set; The end boundary of a chapter structure is determined by the starting position of the adjacent chapter structure.

5. The document parsing exemption method according to claim 1, characterized in that, In response to the content tags of the chapter structure annotation with the aforementioned structure-level exemption mark, attribute-level exemption marks are added to text blocks that meet preset credibility conditions, further including: Content tagging is performed in parallel using rule-based models, natural language processing models, and large language models; When the annotation result meets the preset credibility conditions, an attribute-level exemption mark is added to the corresponding text block, wherein the preset credibility conditions include verification through the text anomaly rule library.

6. The document parsing exemption method according to claim 1, characterized in that, In response to a text block with the attribute-level exemption mark, a post-processing operation is performed, adding a processing-level exemption mark to the post-processing result that meets the preset confidence conditions, and storing it as a standard structured document with multi-level exemption marks, further including: The post-processing operations include performing metadata standardization, constructing chapter hierarchy trees, and chart captioning operations in parallel using multiple processing models. By comparing the post-processing results output by different processing models, when the preset trust conditions are met, a processing-level inspection exemption mark is added and stored as a standard structured document with multiple layers of inspection exemption marks. The preset trust conditions include verification through a structural anomaly rule base.

7. The document parsing exemption method according to claim 6, characterized in that, The construction of the chapter hierarchy tree further includes: Parse text blocks with attribute-level exemption tags and parse the title number format to determine the explicit hierarchy; Analyze the semantic content of the title to generate implicit logical relationships; A chapter hierarchy tree structure is constructed by integrating explicit hierarchy and implicit logical relationships.

8. A document parsing inspection-free system, characterized in that, include: The element-level exemption tagging module is configured to respond to the input document by analyzing the document page in parallel using at least two models, identifying page elements, and adding element-level exemption tags to page elements that meet preset confidence conditions. The structured memory data reorganization module is configured to extract text, images, and path information from page elements with the aforementioned element-level inspection-free markers, and reorganize them into structured memory data according to the reading order. The structure-level inspection-free marking module is configured to respond to the structured memory data, perform multi-source information fusion and rule matching, and add structure-level inspection-free markings to the chapter structure division results that meet preset credibility conditions; The attribute-level exemption tag module is configured to add attribute-level exemption tags to text blocks that meet preset credibility conditions in response to the text structure annotation content tags with the structure-level exemption tags. A standard structured document storage module is configured to perform post-processing operations in response to text blocks with attribute-level exemption tags, add processing-level exemption tags to post-processing results that meet preset confidence conditions, and store them as standard structured documents with multi-layer exemption tags. The preset confidence conditions include that the consistency ratio of the output results of each model for the same object exceeds a preset consistency threshold, the confidence scores of each model are all higher than the dynamic confidence threshold, and they pass the verification of the corresponding preset anomaly rule base.

9. An electronic device, characterized in that, include: The system includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus; the memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores a computer program executable by an electronic device, which, when run on the electronic device, causes the electronic device to perform the steps of the method according to any one of claims 1 to 7.

Images