An electronic document format intelligent proofreading system and method based on multi-dimension checking rules

Abstract

This invention relates to the field of data processing technology, specifically to an intelligent proofreading system and method for electronic document formats based on multi-dimensional verification rules. It includes: a document structure parsing unit; a dynamic verification rule generation unit; an intelligent verification execution unit; and a verification result output unit. This invention eliminates differences in electronic document parsing. After extracting semantic blocks such as titles and body text by a paragraph-level semantic block recognition module, the intelligent verification execution unit calls a multi-dimensional rule set to verify core format elements. The verification result output unit accurately locates the page numbers, paragraphs, and character positions of deviations, generating correction guidelines and multi-format reports containing problem descriptions and compliance examples. This invention solves the problem of the lack of a formal, dedicated proofreading system for core document formats in existing technologies. It can comprehensively identify deviations and clarify the direction of correction, ensuring compliant document circulation while reducing labor costs and improving proofreading efficiency and convenience.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically, to an intelligent proofreading system and method for electronic document formats based on multi-dimensional verification rules. Background Technology

[0002] In modern office settings, official documents, contracts, bidding documents, and other formal electronic documents must strictly adhere to multi-dimensional formatting specifications. These specifications cover national standards, industry requirements, and legally binding details. Any formatting deviation may prevent the document from circulating properly or trigger compliance risks. Currently, most scenarios still rely on manual adjustments to document formatting page by page using office software. This not only consumes a significant amount of time but also suffers from a high error rate, severely hindering the improvement of office efficiency. There is an urgent need for automated technical solutions to achieve accurate formatting proofreading.

[0003] In the existing technology, relevant patents have explored the field of document proofreading. For example, Chinese patent CN202411681874.8 discloses a document proofreading method, device, storage medium, and computer equipment. When proofreading a set of documents, the method first obtains the set of documents and their business forms, embeds the set of documents into a canvas to form a document canvas proofreading page to enter proofreading mode; when the user clicks on a business field in the business form, the method extracts the field information and retrieves the corresponding field block in the document canvas, highlights and parses the text information of the field block and displays it with wireframes, then performs field mapping, and completes form backfilling or document return updating according to the mapping result and user instructions, realizing automated proofreading of multiple documents to improve accuracy and efficiency. For example, Chinese patent CN202411368346.7 discloses a document proofreading method based on a large model, which includes the following steps: document preprocessing is identified as text and segmented into text blocks; after vectorizing the text blocks, the large model is used to polish them to obtain new text blocks; vectorizing them again and comparing the vector similarity; marking the differences in characters and then comparing and integrating the proofreading results for output; this method can adapt to dynamic language changes and diverse document needs, and improve the professionalism and credibility of documents.

[0004] Despite the design advantages of the aforementioned technical solutions, they also suffer from the following technical shortcomings: Firstly, the verification rules lack the flexibility to dynamically adapt to document content. For example, the proofreading logic of Chinese patent CN202411681874.8 revolves around business field mapping and backfilling, only processing content associations between multiple documents and failing to consider the impact of document content characteristics on the proofreading focus. Similarly, the proofreading process of Chinese patent CN202411368346.7 is based on fixed text polishing and similarity comparison steps, focusing only on text content optimization and neglecting to consider the specific content of the documents. The mechanism of adjusting proofreading dimensions and parameters based on semantic features is difficult to adapt to the differentiated proofreading needs of different document types. Secondly, it lacks a dedicated proofreading system for the core format elements of formal documents: CN202411681874.8 only focuses on the matching accuracy of business fields and does not cover the proofreading of key format elements of formal documents such as page settings, font size, paragraph spacing, and chart number association; CN202411368346.7 focuses on improving text quality and does not cover the verification needs related to format compliance, failing to identify and correct format deviations that may lead to documents not meeting circulation standards. Therefore, we propose an intelligent electronic document format proofreading system and method based on multi-dimensional verification rules. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent proofreading system and method for electronic document formats based on multi-dimensional verification rules, so as to solve the problems mentioned in the background art, such as the lack of flexibility in the verification rules to dynamically adapt to the document content and the lack of a dedicated proofreading system for the core format elements of formal documents.

[0006] To address the aforementioned technical problems, one objective of this invention is to provide an intelligent proofreading system for electronic document formats based on multi-dimensional verification rules, comprising: The document structuring parsing unit is used to receive electronic documents and perform structuring parsing on the electronic documents to extract paragraph-level semantic blocks including titles, body text, and charts. A dynamic validation rule generation unit, connected to the document structure parsing unit, employs a graph neural network model to model the paragraph-level semantic blocks, constructing a document semantic topology graph with paragraph-level semantic blocks as nodes and text semantic similarity and citation tag matching results as edges. Through graph neural network embedding learning, it aggregates the semantic information of adjacent nodes to generate a node embedding vector for each paragraph-level semantic block. Then, an attention network is used to detect whether time-period constraint-related semantic features exist in the document content based on the node embedding vectors. When semantic features containing descriptions of clinical research, project management, or financial cycles are detected, the execution priority of the date logic validation dimension in the validation task queue is dynamically increased, and a call to the calendar service is triggered to verify the validity of the field calendar. When semantic features containing the context of abbreviations are detected, the terminology usage compliance validation dimension is activated, and the terminology exemption validation threshold is automatically adjusted based on the context semantic confidence. Finally, a multi-dimensional validation rule set adapted to the current document is generated based on the content features of each paragraph-level semantic block, allowing the validation rules to dynamically evolve with the document content. The intelligent verification execution unit is connected to the dynamic verification rule generation unit. It calls the multi-dimensional verification rule set to match and verify the electronic document format metadata. When performing date logic verification, it calls the calendar service to verify the validity of the date. When performing terminology compliance verification, it refers to the terminology exemption verification threshold and generates difference data containing deviation position, type and rule reference. The verification result output unit is connected to the intelligent verification execution unit, receives the difference data, and generates a format correction suggestion report containing deviation location information and targeted correction guidance.

[0007] As a further improvement to this technical solution, the document structure parsing unit includes a format-compatible access module and a paragraph-level semantic block recognition module, wherein: The format-compatible access module is used to receive electronic documents of different formats, and completes the standardization preprocessing of document formats by calling the corresponding format parsing interface to eliminate the parsing differences of office electronic document formats; The paragraph-level semantic block recognition module is based on the standardized document output by the format-compatible access module. It uses a combination of regular expression matching and deep learning text classification to identify and extract paragraph-level semantic blocks, including titles, body text, and charts.

[0008] As a further improvement to this technical solution, the dynamic verification rule generation unit includes a semantic topology modeling module, a node embedding generation module, and a verification rule adjustment module connected in sequence, wherein: The semantic topology modeling module is used to model paragraph-level semantic blocks and construct a document semantic topology graph with paragraph-level semantic blocks as nodes. The node embedding generation module aggregates the semantic information of adjacent nodes through embedding learning of graph neural networks to generate node embedding vectors for each paragraph-level semantic block. The verification rule adjustment module detects semantic features based on node embedding vectors and dynamically adjusts the verification dimensions to generate a multi-dimensional verification rule set.

[0009] As a further improvement to this technical solution, the semantic topology modeling module, when constructing a document semantic topology graph with paragraph-level semantic blocks as nodes, determines the weights of the edges of the topology graph through the following steps: S21.1, Target paragraph-level semantic block Input each word into a pre-trained word vector model and output the corresponding word vectors; S21.2 Calculate the text semantic similarity between two paragraph-level semantic blocks using the cosine similarity algorithm. This serves as the core weighting criterion for edges; S21.3 Perform citation tag matching on paragraph-level semantic blocks to identify citation relationships between different paragraph-level semantic blocks; S21.4, Text semantic similarity The results of matching with the reference tags are weighted and fused to form the final weight basis for the corresponding edges in the document semantic topology graph.

[0010] As a further improvement to this technical solution, the node embedding generation module, when generating the node embedding vector for each paragraph-level semantic block, implements the embedding learning and semantic information aggregation process through a graph attention network, including the following steps: S22.1 Initial feature vector for each paragraph-level semantic block Perform a linear transformation to obtain the transformed eigenvectors; S22.2 Calculate paragraph-level semantic blocks With adjacent nodes Attention coefficient This characterizes the strength of the semantic association between two nodes; S22.3, Attention Coefficient Normalization is performed to obtain attention weights. This enables the differentiated aggregation of semantic information of adjacent nodes; S22.4, Attention-based weights By aggregating the semantic information of adjacent nodes, a paragraph-level semantic block is finally generated. Node embedding vector This is used for subsequent semantic feature detection.

[0011] As a further improvement to this technical solution, the verification rule adjustment module includes a feature input preprocessing submodule, a semantic feature classification and detection submodule, a verification dimension dynamic configuration submodule, and a multi-dimensional rule set integration and generation submodule, wherein: The feature input preprocessing submodule is used to receive the node embedding vector output by the node embedding generation module. Standardize the vector data to eliminate dimensionality differences; The semantic feature classification and detection submodule is connected to the feature input preprocessing submodule. It uses an attention network to classify the preprocessed node embedding vectors, identify the time period constraint semantic features and the contextual semantic features of professional term abbreviations in the document content, and output the corresponding feature classification probability values. The dynamic configuration submodule for the verification dimension is connected to the semantic feature classification and detection submodule. When the probability value of the semantic feature constrained by the time period is greater than the preset feature judgment threshold, an external calendar service is called through the HTTP or HTTPS interface protocol to verify the calendar validity of the date field. At the same time, the task priority parameter of the date logical verification dimension is increased to the preset high priority parameter. The multi-dimensional rule set integration and generation submodule is connected to the dynamic configuration submodule for verification dimensions. It is used to temporarily store the adjustment parameters of the verification dimensions and provide data support for the subsequent generation of multi-dimensional verification rule sets.

[0012] As a further improvement to this technical solution, the process by which the dynamic configuration submodule for verification dimensions generates a multi-dimensional verification rule set after dynamically adjusting the verification dimensions includes the following steps: S232.1 If the probability value of the contextual semantic feature of a technical term abbreviation is greater than the preset feature judgment threshold, then the probability value shall be used as the contextual semantic confidence level. ; S232.2, Context-based semantic confidence Dynamically determine the term exemption verification threshold This enables adaptive adjustments to terminology compliance verification. S232.3, Integrated Calendar Service Call Parameters, Date Validation Priority Parameters, and Terminology Exemption Validation Thresholds Generate a multi-dimensional set of validation rules in JSON format; S232.4 Add configuration fields related to interface calls to the multi-dimensional verification rule set to enable direct data interaction with the intelligent verification execution unit.

[0013] As a further improvement to this technical solution, the intelligent verification execution unit includes a rule set receiving and parsing module, a multi-dimensional verification execution module, and a difference data integration and generation module connected in sequence, wherein: The rule set receiving and parsing module is used to establish a data connection with the dynamic verification rule generation unit, receive multi-dimensional verification rule sets, parse the verification dimension identifiers, parameter data types and interface call fields in the multi-dimensional verification rule sets, and generate standardized verification instructions. The multi-dimensional verification execution module matches and verifies the metadata of electronic document formats based on standardized verification instructions. When performing date logic verification, it calls an external calendar service to complete date validity verification. When performing terminology compliance verification, it refers to the terminology exemption verification threshold in the rule set to determine compliance. The difference data integration and generation module is used to collect the verification results of the multi-dimensional verification execution module, extract deviation location information, deviation type and corresponding rule reference identifier, integrate and generate difference data according to a preset data structure, and provide data input for the verification result output unit.

[0014] As a further improvement to this technical solution, the verification result output unit includes a difference data receiving and processing module, a deviation information parsing and annotation module, a correction guidance generation module, and a report integration and output module connected in sequence, wherein: The difference data receiving and processing module establishes a data connection with the intelligent verification execution unit, receives difference data, and performs format standardization processing on the deviation location information, deviation type and rule reference identifier in the difference data to eliminate data format differences. The deviation information parsing and annotation module analyzes the page number, paragraph number, and character position of the electronic document corresponding to the deviation based on the standardized difference data, and generates accurate deviation positioning annotation information. The correction guidance generation module matches the preset correction strategy library according to the deviation type and the corresponding rule reference identifier, and generates a targeted correction guidance that is adapted to the deviation type. The correction guidance includes a description of the deviation problem and a compliance format example. The report integration and output module is used to integrate deviation positioning and annotation information with targeted correction guidelines according to a preset template, generate a formatted correction suggestion report with standardized format, and supports multiple output formats for users to choose from.

[0015] The second objective of this invention is to provide an intelligent proofreading method for electronic document formats based on multi-dimensional verification rules. The intelligent proofreading system for electronic document formats based on multi-dimensional verification rules, as described above, includes the following steps: S1. Receive electronic documents in different formats, perform standardized preprocessing on the document formats to eliminate parsing differences, and then use a combination of regular expression matching and deep learning text classification to identify and extract paragraph-level semantic blocks corresponding to titles, body text, and charts. S2. Perform word vector conversion and semantic similarity calculation on paragraph-level semantic blocks, construct a document semantic topology graph by combining the citation tag matching results, generate node embedding vectors for each paragraph-level semantic block, dynamically adjust the verification dimension parameters according to the document semantic features, and integrate and generate a multi-dimensional verification rule set in JSON format containing interface call fields. S3. Parse the multi-dimensional verification rule set to generate standardized verification instructions. Perform multi-dimensional matching verification on the electronic document format metadata according to the verification instructions. Among them, the date logic verification calls the external calendar service to verify the validity, and the terminology compliance verification refers to the preset exemption threshold for judgment. Collect the verification results and extract the deviation information, and integrate them to generate difference data. S4. Standardize the format of the discrepancy data, accurately mark the document location information corresponding to the deviation, generate a correction guide with problem description and compliance examples according to the deviation type and the correction strategy library, integrate the deviation marking information and correction guide according to the preset template, and generate a format correction suggestion report with multiple formats available.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention uses a graph neural network model to model paragraph-level semantic blocks through a dynamic verification rule generation unit, constructing a document semantic topology graph with semantic blocks as nodes and text semantic similarity and citation tag matching results as edges. Then, it aggregates the semantic information of adjacent nodes through graph neural network embedding learning to generate node embedding vectors. Finally, it uses an attention network to detect semantic features in the document such as time period constraints and the context of professional term abbreviations, dynamically adjusts the date logic verification priority and term exemption verification threshold, and generates a multi-dimensional verification rule set adapted to the current document. This allows the verification rules to evolve dynamically with the document content, effectively solving the problem that existing technology verification rules lack the flexibility to dynamically adapt to document content. It can accurately match the differentiated format proofreading needs of different scenarios such as official documents, contracts, and bidding documents. 2. This invention eliminates parsing differences between electronic documents of different formats through a format compatibility access module, extracts semantic blocks such as titles, body text, and charts using a paragraph-level semantic block recognition module, and then uses an intelligent verification execution unit to call a multi-dimensional verification rule set to perform matching verification on core format elements of formal documents, such as date logic and terminology compliance. Finally, the verification result output unit accurately locates the page number, paragraph number, and character position of the deviation, and generates targeted correction guidance and a report with multiple format options, including explanations of the deviation problem and examples of compliant formats. This effectively solves the problem that existing technologies do not build a dedicated proofreading system for the core format elements of formal documents, can comprehensively identify format deviations and provide clear correction directions, ensure that documents meet the circulation standards and compliance requirements of the corresponding scenarios, and at the same time reduce the cost of manual intervention and improve proofreading efficiency and convenience. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the system framework of the present invention; Figure 2 This is a schematic diagram of the process for generating dynamic verification rules in this invention; Figure 3 This is a schematic diagram of the method steps of the present invention; The meanings of the labels in the diagram are as follows: 1. Document structure parsing unit; 11. Format compatibility access module; 12. Paragraph-level semantic block recognition module; 2. Dynamic validation rule generation unit; 21. Semantic topology modeling module; 22. Node embedding generation module; 23. Validation rule adjustment module; 3. Intelligent verification execution unit; 31. Rule set receiving and parsing module; 32. Multi-dimensional verification execution module; 33. Difference data integration and generation module; 4. Verification result output unit; 41. Difference data receiving and processing module; 42. Deviation information parsing and annotation module; 43. Correction guidance generation module; 44. Report integration and output module. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] like Figures 1-2 As shown, this embodiment provides an intelligent proofreading system for electronic document formats based on multi-dimensional verification rules, including: Document structure parsing unit 1 is used to receive electronic documents and perform structure parsing on the electronic documents to extract paragraph-level semantic blocks including titles, body text, and charts. In this embodiment, the document structure parsing unit 1 includes a format-compatible access module 11 and a paragraph-level semantic block recognition module 12, wherein: The format-compatible access module 11 is used to receive electronic documents of different formats, and completes the standardization preprocessing of document formats by calling the corresponding format parsing interface to eliminate the parsing differences of office electronic document formats; Specifically, the format compatibility access module 11 targets mainstream electronic document formats in office scenarios (including but not limited to Word (.doc / .docx), PDF (.pdf), WPS (.wps), Excel (.xls / .xlsx), etc.) and achieves multi-format compatibility and standardized preprocessing through mature technologies in this field. The specific operation is as follows: Automatic format recognition: The format-compatible access module 11 has built-in file format recognition logic, which automatically determines the document format type by parsing the file header feature identifiers of electronic documents (such as the PK signature of .docx files and the %PDF- prefix identifier of .pdf files); Parsing Interface Adaptation: Based on the identified document format, the format compatibility access module 11 calls the corresponding format parsing tool interface commonly used in this field—for example, Word documents call the Apache POI interface, PDF documents call the iText interface, and WPS documents call the Kingsoft Office SDK interface, to ensure that the core content such as text and charts of different document formats is completely extracted; Standardized preprocessing: To eliminate structural differences between documents of different formats, the format compatibility access module 11 performs a unified preprocessing operation: Unified encoding: Convert the text encoding of all documents to UTF-8 format to avoid character garbled text or missing parsing due to inconsistent encoding; Structural regularity: PDF streaming text is split into independent paragraphs according to line breaks and page breaks, text in Excel spreadsheets is integrated into paragraph segments according to cell affiliation, and section breaks and page breaks of various documents are uniformly converted into standard paragraph separator marks; Redundancy stripping: Filter out format control characters and unnecessary metadata (such as author, modification time, and other information unrelated to format validation) in the document, and retain only the core text and chart-related information.

[0020] The paragraph-level semantic block recognition module 12, based on the standardized document output by the format-compatible access module 11, uses a combination of regular expression matching and deep learning text classification to identify and extract paragraph-level semantic blocks, including titles, body text, and charts.

[0021] Specifically, the paragraph-level semantic block recognition module 12, based on the standardized document output by the format-compatible access module 11, adopts a hybrid scheme of "initial screening by regular expression matching + fine screening by deep learning text classification" to accurately extract three types of semantic blocks: title, body text, and charts. The specific implementation is as follows: Preliminary screening using regular expressions: The paragraph-level semantic block recognition module 12 pre-sets regular expression rules to adapt to the format features of office document titles and chart descriptions. By traversing all paragraphs of the standardized document, it initially filters out candidate title blocks and candidate chart descriptions. The remaining unmatched paragraphs are temporarily designated as candidate body text blocks. Deep learning text classification and screening: Model selection and training: The paragraph-level semantic block recognition module 12 adopts a lightweight BERT model, which is finely tuned based on a public corpus of office documents in multiple scenarios, including official documents, contracts, and bidding documents. The text fragments in the corpus have been manually labeled with three categories of tags: "title", "body text" and "figure description". Classification execution: The paragraph-level semantic block recognition module 12 divides all candidate blocks after initial screening into text segments of fixed length (padding is used to fill in the missing length), inputs them into the trained lightweight BERT model, and outputs the probability value of the text segment belonging to each category. When the probability value of a certain category is greater than the preset threshold (a common setting in this field), the semantic block type of the text segment is determined, and the misjudgment results of the regular initial screening are corrected. Chart semantic block integration: The paragraph-level semantic block recognition module 12 identifies the chart description text and associates it with the chart objects (images, tables) in the document through the chart number. It integrates the chart objects and the corresponding description text into a complete "chart semantic block" and records the position information of the chart objects in the document, such as page number and paragraph number.

[0022] Dynamic validation rule generation unit 2, connected to document structure parsing unit 1, employs a graph neural network model to model paragraph-level semantic blocks, constructing a document semantic topology graph with paragraph-level semantic blocks as nodes and text semantic similarity and citation tag matching results as edges. Through graph neural network embedding learning, it aggregates the semantic information of adjacent nodes to generate node embedding vectors for each paragraph-level semantic block. Then, an attention network is used to detect the presence of time-cycle constraint-related semantic features in the document content based on the node embedding vectors. When semantic features containing descriptions of clinical research, project management, or financial cycles are detected... During the verification process, the execution priority of the date logic verification dimension in the verification task queue is dynamically increased, and a call to the calendar service is triggered to verify the validity of the field calendar. When semantic features containing the context of abbreviations of technical terms are detected, the term usage compliance verification dimension is activated, and the term exemption verification threshold is automatically adjusted according to the semantic confidence of the context. Finally, a multi-dimensional verification rule set adapted to the current document is generated based on the content features of each paragraph-level semantic block, so that the verification rules evolve dynamically with the document content. The dynamic verification rule generation unit 2 includes a semantic topology modeling module 21, a node embedding generation module 22, and a verification rule adjustment module 23 connected in sequence.

[0023] The semantic topology modeling module 21 is used to model paragraph-level semantic blocks and construct a document semantic topology graph with paragraph-level semantic blocks as nodes. When constructing the document semantic topology graph with paragraph-level semantic blocks as nodes, the semantic topology modeling module 21 determines the weights of the edges of the topology graph through the following steps: S21.1, Target paragraph-level semantic block Input each word into a pre-trained word vector model and output the corresponding word vectors; Specifically, in order to transform paragraph-level semantic blocks in text form into computable feature vectors, we first select any two target paragraph-level semantic blocks. ( Both are independent semantic blocks output by the document structuring parsing unit 1, covering types such as title, body text, and chart semantic blocks. These are then input into commonly used pre-trained word vector models (such as Word2Vec and GloVe models). Through model computation, the output... The corresponding word vector sequence and The corresponding word vector sequence ,in represent word count, represent The word count is set, and the dimension of all word vectors is uniformly set to 300 (this dimension is a common choice in the field of natural language processing, which can achieve a balance between feature representation completeness and computational efficiency).

[0024] S21.2 Calculate the text semantic similarity between two paragraph-level semantic blocks using the cosine similarity algorithm. This serves as the core weighting criterion for edges; Specifically, for quantification To determine the semantic relevance, the word vector sequence must first be aggregated to obtain a paragraph-level overall feature vector. Definition for The average pooling vector of the word vector sequence, for The average pooled vector of the word vector sequence is calculated as follows: ; ; Based on the aforementioned average pooling vectors, the cosine similarity algorithm is used to calculate the textual semantic similarity between the two. (The value range is [0,1], and the closer the value is to 1, the stronger the semantic association between the two semantic blocks), the formula is as follows: ; in" " represents the vector dot product operation. The L2 norm of the vector is used to determine the similarity result, which will serve as the core basis for the edge weights of the topological graph.

[0025] S21.3 Perform citation tag matching on paragraph-level semantic blocks to identify citation relationships between different paragraph-level semantic blocks; Specifically, besides implicit semantic relationships, explicit references often exist between paragraphs in office documents, and these relationships are equally important for semantic modeling. Therefore, it is necessary to traverse... The text content is identified by recognizing text fragments containing citations such as "see", "as shown in Figure X", "see Part X", "quoted from", and "based on Clause X", and then the target paragraph identifiers (such as paragraph numbers, title keywords, figure numbers, etc.) pointed to in the citations are extracted.

[0026] like The citation mark in the middle clearly points to ,or The citation mark in the middle clearly points to If the two semantic blocks are directly referenced, the reference matching result is determined to be... If no valid identifier for mutual references is detected, the reference matching result is marked as... (R is a binary identifier, used only to distinguish whether a reference association exists).

[0027] Furthermore, in practical applications of office documents, in addition to the aforementioned "explicitly targeting paragraph" citations, there are also common vague citation types without specific target identifiers, such as "see related sections," "refer to previous text," "see related chapters," and "refer to the above content." While these vague citations do not explicitly point to a specific paragraph, they may imply semantic relationships between paragraphs. Without targeted processing, the semantic topology graph may fail to fully reflect the actual relationships between paragraphs, thus affecting the accuracy of subsequent node embedding generation and validation rule adjustments. Therefore, the following rules for handling vague citations are supplemented to complement the processing logic for explicit citations: Identification of fuzzy references: When traversing the text, if a reference expression without a clear target identifier is detected, it is determined to be a fuzzy reference.

[0028] Selection of candidate related paragraphs: Based on the paragraph where the fuzzy reference is located, extract the two adjacent paragraphs before and after it as candidate related paragraphs (if there are fewer than two adjacent paragraphs, then take all the actual adjacent paragraphs), thus forming a set of candidate targets.

[0029] Determination and labeling of association relationships: Following the cosine similarity calculation method disclosed in S21.2, the semantic similarity between the fuzzy reference paragraph and each candidate target paragraph is calculated; if the highest similarity is ≥0.7, a weak reference association is determined, and the corresponding reference matching result is labeled as... If the highest similarity is <0.7, then it is determined that there is no substantial reference association, and the reference matching result is marked as... .

[0030] Adaptation to weight calculation: For The weak reference associations will be directly substituted into the weighted fusion formula in S21.4 to participate in the edge weight calculation, ensuring that the semantic topology graph can reflect such weak associations.

[0031] S21.4, Text semantic similarity The results of matching with the reference tags are weighted and fused to form the final weight basis for the corresponding edges in the document semantic topology graph.

[0032] Specifically, to comprehensively reflect the dual impact of semantic association and reference association, the two results need to be weighted and fused to obtain the final weights of the topological graph edges. Definition for and The final weights of the corresponding edges are used to set the weight coefficients for text semantic similarity. The weighting coefficients of the reference tag matching results (This weighting is determined based on the characteristics of office documents, highlighting the core position of semantic content association while not ignoring the importance of explicit reference association.) The fusion formula is as follows: ; The final result The value ranges from [0,1]. The larger the value, the stronger the overall association between the two paragraph-level semantic blocks, providing accurate association weight support for the subsequent aggregation of semantic information in the node embedding vector.

[0033] Understandably, the above design breaks through the limitations of traditional semantic modeling, which relies solely on the single dimension of text similarity. This embodiment combines "textual semantic association" with "explicit citation association" to construct a document semantic topology graph, making the relationships between nodes more consistent with the actual logical structure of office documents such as official documents, contracts, and bidding documents. This solves the problem that single semantic modeling cannot fully reflect the citation associations between paragraphs, and provides more comprehensive and realistic basic data for subsequent deep semantic feature extraction.

[0034] Specifically, the node embedding generation module 22 aggregates the semantic information of neighboring nodes through embedding learning of a graph neural network to generate a node embedding vector for each paragraph-level semantic block. When generating the node embedding vector for each paragraph-level semantic block, the node embedding generation module 22 uses a graph attention network to achieve the embedding learning and semantic information aggregation process, which includes the following steps: S22.1 Initial feature vector for each paragraph-level semantic block Perform a linear transformation to obtain the transformed eigenvectors; Specifically, the paragraph average pooling vector output by the semantic topology modeling module 21 As a paragraph-level semantic block initial feature vector (300-dimensional) To map this initial feature vector to a unified feature space suitable for subsequent computation, a linear transformation is required. Definition The weight matrix of the fully connected layer (dimension 1) ,in For input dimensions, (This is the output dimension, which is a common dimension selection in this field). The bias term for the fully connected layer (dimension: ), The eigenvectors are the result of a linear transformation, and the transformation formula is as follows: ; in and Through training and optimization using a publicly available corpus of office documents, it was determined that the core purpose of linear transformation is to unify feature dimensions and enhance feature representation capabilities, thus laying a solid foundation for the subsequent calculation of attention coefficients.

[0035] S22.2 Calculate paragraph-level semantic blocks With adjacent nodes Attention coefficient This characterizes the strength of the semantic association between two nodes; Specifically, in order to distinguish the degree of influence of different adjacent nodes on the target semantic block, it is necessary to calculate the paragraph-level semantic block. With each adjacent node Attention coefficient This coefficient is used to characterize the strength of the semantic association between two nodes. Definition The attention weight vector (dimension 1) ), " " indicates a vector concatenation operation. The activation function (with the slope parameter set to 0.2, a standard setting within the domain) is used, and the attention coefficient is calculated using the following formula: ; This formula uses semantic blocks With adjacent nodes The transformed feature vectors are concatenated and then processed by the attention weight vector. Weighted summation and activation function operations yield attention coefficients that reflect the relative correlation between the two factors. .

[0036] S22.3, Attention Coefficient Normalization is performed to obtain attention weights. This enables the differentiated aggregation of semantic information of adjacent nodes; Specifically, in order to achieve differentiated aggregation of semantic information of adjacent nodes, the attention coefficients need to be normalized to obtain the attention weights. .definition Paragraph-level semantic blocks The set of all adjacent nodes is normalized using the Softmax function, as shown in the following formula: ; Normalized The value range is [0,1], and all adjacent nodes are semantic blocks. The sum of attention weights is 1, which allows neighboring nodes with stronger associations to focus on semantic blocks. The embedding vector contributes more.

[0037] S22.4, Attention-based weights By aggregating the semantic information of adjacent nodes, a paragraph-level semantic block is finally generated. Node embedding vector This is used for subsequent semantic feature detection.

[0038] Specifically, based on the normalized attention weights The transformed feature vectors of adjacent nodes Weighted summation is performed to aggregate semantic information. Then, a sigmoid activation function is used to perform a non-linear transformation on the aggregated feature vector to enhance its expressive power, ultimately generating paragraph-level semantic blocks. Node embedding vector (A vector symbol with a dot in the upper right corner).

[0039] definition Paragraph-level semantic blocks The set of all adjacent nodes, The aggregation formula for the Sigmoid activation function is as follows: ; generated It also covers paragraph-level semantic blocks. Its own core features, as well as the associated semantic features of its neighboring nodes, can accurately characterize the attributes and relationships of the semantic block in the document semantic topology graph, and can be directly used as the core input data for subsequent semantic feature detection.

[0040] Understandably, the above design abandons the average aggregation method used in traditional graph neural networks. Instead, it calculates and normalizes attention coefficients through graph attention networks, assigning differentiated attention weights to different neighboring nodes. This solves the problems of semantic information ambiguity and weakened influence of key related nodes caused by average aggregation, making the generated node embedding vectors more prominent in highlighting the core relationships and improving the accuracy of subsequent semantic feature detection.

[0041] The verification rule adjustment module 23 detects semantic features based on node embedding vectors and dynamically adjusts the verification dimensions to generate a multi-dimensional verification rule set. The verification rule adjustment module 23 includes a feature input preprocessing submodule, a semantic feature classification and detection submodule, a verification dimension dynamic configuration submodule, and a multi-dimensional rule set integration and generation submodule, wherein: The feature input preprocessing submodule is used to receive the node embedding vector output by the node embedding generation module. Standardize the vector data to eliminate dimensionality differences; Specifically, the feature input preprocessing submodule is used to process the node embedding vector output by the node embedding generation module 22. Standardization and noise reduction are performed to ensure the accuracy and stability of subsequent semantic feature detection. The specific steps are as follows: Standardization: Eliminates differences in the numerical ranges of features across different dimensions, preventing excessively large values ​​in some dimensions from interfering with classification results. Definition For the standardized vector, The mean of the embedding vectors for all nodes. To find the standard deviation, the Z-score standardization method is used, and the formula is as follows: ; After standardization, all vectors follow a standard normal distribution with a mean of 0 and a standard deviation of 1, ensuring fairness in the feature dimensions.

[0042] The semantic feature classification and detection submodule is connected to the feature input preprocessing submodule. It uses an attention network to classify the preprocessed node embedding vectors, identify the semantic features of time period constraints and the semantic features of the context of professional term abbreviations in the document content, and output the corresponding feature classification probability values. Specifically, the semantic feature classification and detection submodule uses an attention network to process the standardized vectors. Perform feature classification to accurately identify two types of core semantic features and output corresponding probability values. The specific steps are as follows: Attention Network Architecture: A BiLSTM-Attention (Bidirectional Long Short-Term Memory Network + Attention Mechanism) architecture is adopted, which has strong semantic capture capabilities in text semantic feature recognition. The network is fine-tuned and trained based on a publicly available corpus of semantic features of office documents (covering annotated samples from scenarios such as contracts, tenders, and financial reports) to ensure the accuracy of target feature recognition.

[0043] Feature detection target: Focus only on two types of core semantic features, namely time period constraint semantic features (such as project cycle, settlement deadline, etc.) and professional term abbreviation contextual semantic features (such as "API (Application Programming Interface)" and other term abbreviations with explanations).

[0044] Output format: The network independently outputs the classification probability values ​​of the two types of features, where The detection probability of semantic features corresponding to time period constraints. The detection probability of the semantic features of the context of the corresponding technical term abbreviation is [0,1]. The higher the probability value, the higher the credibility of the semantic block containing the corresponding feature.

[0045] The dynamic configuration submodule for verification dimensions is connected to the semantic feature classification and detection submodule. When the probability value of the semantic feature constrained by the time period is greater than the preset feature judgment threshold, an external calendar service is called via HTTP or HTTPS interface protocol to verify the calendar validity of the date field. At the same time, the task priority parameter of the date logical verification dimension is increased to the preset high priority parameter. The process of generating a multi-dimensional verification rule set after dynamically adjusting the verification dimensions by the dynamic configuration submodule for verification dimensions includes the following steps: S232.1 If the probability value of the contextual semantic feature of a technical term abbreviation is greater than the preset feature judgment threshold, then the probability value shall be used as the contextual semantic confidence level. ; Specifically, preset feature determination threshold (A domain-standard value that balances accuracy and recall). The probability value of the semantic features of the context of technical term abbreviations. Then directly use the probability value Defined as contextual semantic confidence ,Right now .

[0046] S232.2, Context-based semantic confidence Dynamically determine the term exemption verification threshold This enables adaptive adjustments to terminology compliance verification. Specifically, based on contextual semantic confidence Determine the term exemption verification threshold This enables adaptive adjustments to terminology compliance verification. Definition The value range is [0, 0.2]. The calculation logic is that the higher the confidence level, the lower the exemption threshold (to reduce unnecessary exemptions). The formula is as follows: ; For example, if (High confidence level) then Only a small number of non-compliant terms are exempted; if ,but The exemption conditions should be appropriately relaxed.

[0047] If the probability value of semantic features is constrained by the time period If so, two configurations are enabled: First, external public calendar services (such as Microsoft Outlook Calendar API and Google Calendar API) are called via HTTP / HTTPS interface protocol. The interface address is preset to the official standard address of the service, and the request parameters include the date field to be verified. Second, the task priority parameter of the date logical verification dimension is increased to the preset high priority parameter (set to 1, and verification dimensions below this value are not high priority) to ensure that date verification is executed first.

[0048] S232.3, Integrated Calendar Service Call Parameters, Date Validation Priority Parameters, and Terminology Exemption Validation Thresholds Generate a multi-dimensional set of validation rules in JSON format; S232.4 Add configuration fields related to interface calls to the multi-dimensional verification rule set to enable direct data interaction with the intelligent verification execution unit 3.

[0049] The multi-dimensional rule set integration and generation submodule is connected to the dynamic configuration submodule for verification dimensions. This submodule is used to temporarily store the adjustment parameters of the verification dimensions, providing data support for the subsequent generation of multi-dimensional verification rule sets.

[0050] Specifically, the multi-dimensional rule set integration and generation submodule is used to temporarily store all adjustment parameters of the verification dimensions, integrate them according to a standardized format to generate a multi-dimensional verification rule set, and provide the intelligent verification execution unit 3 with a structured and directly parsable verification basis, specifically including: Parameter temporary storage mechanism: A dynamically adaptable temporary data cache pool is used to store adjustment parameters. The cache pool can automatically allocate memory block resources according to the amount of parameter data, avoiding memory waste or insufficient memory issues. The temporarily stored parameters include core parameters for calendar service calls, high-priority parameters for date logic validation, and terminology exemption verification thresholds. This ensures that parameters can be read in real time without loss during the integration process.

[0051] JSON format rule set generation—strictly follows the JSONSchema standardization specification to construct the rule set structure, balancing machine parsing efficiency and human readability. The core fields are as follows: Basic configuration fields: rule set version number and generation timestamp, used for rule set version management and traceability; Verify dimension configuration fields: dimension identifier, execution priority, and core threshold parameter. The core threshold parameter corresponds to the terminology exemption verification threshold. ; Service-related parameter fields: Only retain the necessary configuration information for calendar service calls, including the necessary configuration fields for the service interface call (such as interface address and request parameter format).

[0052] Output and Integrity Verification: The generated JSON-formatted rule set is directly stored in the system's preset encrypted storage path. Simultaneously, the rule set data undergoes integrity verification based on the MD5 algorithm. By extracting MD5 fingerprints in parallel from rule set data slices and synthesizing the final checksum, the checksum is stored synchronously with the rule set for integrity verification by the intelligent verification execution unit 3 when receiving the rule set, preventing tampering or damage during data transmission.

[0053] Furthermore, in actual deployment scenarios, encrypted storage paths may encounter access anomalies such as insufficient permissions or non-existent paths. Failure to handle these anomalies promptly will lead to rule set storage failure, thereby interrupting subsequent verification processes. Therefore, the following supplementary handling solutions are provided for encrypted path access anomalies: Anomaly detection and judgment: When writing the encrypted path into the rule set, monitor the error codes returned by system I / O in real time (such as permission error code "EACCES", path not found error code "ENOENT"), and judge it as an abnormal access to the encrypted path; Exception handling process: Prioritize switching to the system's preset backup encrypted storage path and automatically retry storage operations based on the rule set; If the backup path storage still fails, an error log will be triggered (the log content includes the exception type, timestamp, and unique identifier of the rule set), and an alarm message will be pushed to the system management node. Subsequent recovery mechanism: After the anomaly is resolved, the system will automatically read the information of the rule set that was not successfully stored from the error log, re-execute the storage process, and ensure that the rule set data is not lost.

[0054] Understandably, the above design breaks through the rigid configuration mode of traditional fixed rules, which is "one-size-fits-all". It takes semantic feature detection results as the core driver to realize the dynamic adaptation of verification rules. It ensures that key verification dimensions are executed first by adjusting the priority, and adapts to the terminology usage scenarios of different documents by dynamically calculating the terminology exemption threshold. This solves the problems of insufficient semantic adaptation of traditional rules to complex documents and low verification accuracy. At the same time, the standardized JSON rule set structure and integrity verification mechanism improve the stability and execution efficiency of rule transmission and avoid the defects of traditional rule set format disorder and easy loss.

[0055] The intelligent verification execution unit 3 is connected to the dynamic verification rule generation unit 2. It calls the multi-dimensional verification rule set to match and verify the electronic document format metadata. When performing date logic verification, it calls the calendar service to verify the validity of the date. When performing terminology compliance verification, it refers to the terminology exemption verification threshold and generates difference data containing deviation position, type and rule reference. In this embodiment, the intelligent verification execution unit 3 includes a rule set receiving and parsing module 31, a multi-dimensional verification execution module 32, and a difference data integration and generation module 33 connected in sequence, wherein: The rule set receiving and parsing module 31 is used to establish a data connection with the dynamic verification rule generation unit 2, receive multi-dimensional verification rule sets, parse the verification dimension identifiers, parameter data types and interface call fields in the multi-dimensional verification rule sets, and generate standardized verification instructions. Specifically, a data interaction bridge is established between the dynamic verification rule generation unit 2 and the intelligent verification execution unit 3. Through standardized parsing, the multi-dimensional verification rule set is converted into directly executable verification instructions, avoiding verification deviations caused by ambiguity in rule format. The specific process is as follows: First, a persistent connection is established using the HTTP / HTTPS interface protocol consistent with dynamic validation rule generation unit 2. Interface communication follows RESTful specifications, and a connection timeout of 5 seconds is set to balance connection stability and execution efficiency. The connection identifier is defined as... ,when This indicates a normal connection. An automatic reconnection mechanism is triggered (up to 3 reconnections, with a 1-second interval between each reconnection) to ensure stable data transmission channels.

[0056] Subsequently, the JSON-formatted multi-dimensional validation rule set output by the dynamic validation rule generation unit 2 is received. After receiving the data, the integrity is first verified based on the MD5 checksum (the checksum and rule set are transmitted synchronously); the checksum matching result is defined as... ,like (If the match is successful) then proceed to the next parsing process. If a match fails, a data retransmission request is immediately returned to prevent incomplete rule sets from entering the execution phase.

[0057] Then, a hierarchical parsing mechanism is used to... Accurately analyze key fields: extract the "Dimension Identifier" field and map it to an internally unified validation dimension code (date logical validation code). Terminology compliance verification code To avoid ambiguous identifiers; to identify date verification priority parameters and terminology exemption verification thresholds. The data type is checked and its validity is verified. If a type mismatch is found, it is marked as mismatched. It also returns a rule set correction request; extracts information such as the calendar service interface address and request parameter format from the "Interface Call Field" and stores it in the interface configuration cache pool. This provides fast query support for subsequent service calls.

[0058] Finally, all parsing results are integrated to generate standardized verification instructions. The instructions use a structured data format, with core fields including connection status identifiers. Verification dimension encoding ( (date / term), priority parameter Terminology exemption threshold Interface Configuration Index This ensures that the multi-dimensional verification execution module 32 can start verification directly without secondary parsing, thus improving execution efficiency.

[0059] Furthermore, in the above rule matching execution process, to ensure the orderliness and accuracy of the verification, the following logic is added: Rule matching priority: Date compliance rules are marked as high priority; the criterion for choosing conflicting rules is "high priority rules take precedence over low priority rules, and rules of the same priority are executed in reverse order of their generation time".

[0060] Partial match threshold: When the match between a rule clause and a document semantic block is ≥0.6, a verification prompt is triggered.

[0061] The multi-dimensional verification execution module 32 uses standardized verification instructions to match and verify the metadata of electronic document formats. When performing date logic verification, it calls an external calendar service to complete date validity verification. When performing terminology compliance verification, it refers to the terminology exemption verification threshold in the rule set to determine compliance. Specifically, the standardized verification instructions are received and parsed based on the rule set. The system performs precise matching and verification of electronic document format metadata, calls external services as needed, and uses adaptive thresholds for judgment. The specific process is as follows: First, read the standardized verification instructions. The core parameters are encoded according to the verification dimension. Activate the corresponding verification logic ( Activate date logical validation, (Activate terminology compliance verification); simultaneously configure the cache pool from the interface. The corresponding interface configuration information is retrieved, and the verification execution environment is initialized (e.g., the field matching timeout is set to 3 seconds) to prepare for verification execution.

[0062] Subsequently, the metadata set of the electronic document format was extracted. ,in For all date type fields (such as contract signing date, project start date, etc.). A candidate set of terms is generated from all text type fields (fields containing technical terminology features are initially filtered through keyword matching); the original location information of each metadata is preserved during the extraction process. This provides a precise basis for subsequent deviation positioning.

[0063] Then, specific checks are performed in priority order: based on the highest priority parameters in the instruction. Prioritize date logic validation – based on the interface configuration index. from The system retrieves calendar service interface information and initiates a service call via HTTP / HTTPS protocol (defining the call status as...). , This indicates that the call was successful. (Indicates call failure, triggering two reconnections upon failure). Each date field The data is transmitted to an external calendar service to verify calendar validity (e.g., the existence of February 30th) and time period consistency (e.g., whether the end date is earlier than the start date). The date verification result is defined as follows. (1 indicates success, 0 indicates failure, and 2 indicates an exception).

[0064] Next, refer to the terminology exemption verification threshold in the instruction. Perform terminology compliance verification: Compare with industry terminology standard databases to obtain a candidate set of terms. Each term in Compliance score (Value range [0,1]); with To determine the compliance threshold, the verification result is determined by threshold comparison. (1 indicates compliance including exemption, 0 indicates non-compliance), enabling adaptive adjustment of terminology verification.

[0065] Finally, check all date verification results. Terminology verification results Location information of corresponding metadata Associated storage to the verification result cache pool The cache pool supports fast retrieval by verification dimension and result type, facilitating subsequent extraction of differential data.

[0066] The difference data integration and generation module 33 is used to collect the verification results of the multi-dimensional verification execution module 32, extract the deviation location information, deviation type and corresponding rule reference identifier, integrate and generate difference data according to the preset data structure, and provide data input for the verification result output unit 4.

[0067] Specifically, the difference data integration and generation module 33 extracts valid deviation information from the verification results and generates traceable difference data through structured integration, ensuring that the verification result output unit 4 can clearly present the verification issues. The specific process is as follows: First, iterate through the validation result cache pool. Filter out the results of failed verification ( or ) and verification anomaly results ( ), define the deviation dataset as (Corresponding to date deviation, terminology deviation, and verification anomaly data respectively), accurately pinpointing the core issues that need to be reported.

[0068] Subsequently, three types of core information were extracted from each record in the deviation dataset: the location information of the corresponding metadata was extracted. Batch field supplement field identifier Standardize the coding of deviation types (calendar invalid). Periodic logic error Non-compliant terminology Service call failed Extract the rule set version number. Verification dimension encoding The rules and reference identifiers for core parameters, etc., ensure that deviations can be traced back to specific verification rules.

[0069] Then, the extracted core information is structured and integrated according to the preset data structure to generate differential data in JSON format. Includes basic information fields (generation timestamp, electronic document identifier). Rule set version number), deviation details field (deviation location) Deviation type coding and description, rule reference identifier, supplementary explanation fields (reason for exception, compliance score) (etc.), taking into account both machine parsing and human readability.

[0070] Finally, for structured differential data Perform quality checks to verify field integrity and format validity (compliance with JSON Schema specifications); define the quality check result as... ,like (If verification passes) the difference data is output to verification result output unit 4. If the validation fails, mark the missing field or the format as incorrect, and re-execute the integration process to ensure high-quality output data.

[0071] Furthermore, for abnormal scenarios encountered during actual operation, the following handling solutions are adopted: Rule set version incompatibility: Automatically downgrade to the system's default compatible version, and prompt the user to update the rule set.

[0072] Verification timeout handling: Set the timeout threshold to 30 seconds; if the timeout is interrupted, the system saves the verification progress and resumes the unfinished verification process after restarting.

[0073] Verification result output unit 4 is connected to intelligent verification execution unit 3, receives difference data, and generates a formatted correction suggestion report containing deviation location information and targeted correction guidance.

[0074] In this embodiment, the verification result output unit 4 includes a difference data receiving and processing module 41, a deviation information parsing and annotation module 42, a correction guidance generation module 43, and a report integration and output module 44 connected in sequence, wherein: The difference data receiving and processing module 41 establishes a data connection with the intelligent verification execution unit 3, receives the difference data, and performs format standardization processing on the deviation location information, deviation type and rule reference identifier in the difference data to eliminate data format differences. Specifically, the difference data receiving and processing module 41 is used to establish a data interaction channel with the intelligent verification execution unit 3, receive difference data and complete standardization processing to eliminate data format differences, and provide basic data with a unified format for subsequent modules. The specific process is as follows: First, a stable data connection is established using the same HTTP / HTTPS interface protocol as the intelligent verification execution unit 3. The interface communication follows the RESTful specification, the connection timeout is set to 5 seconds, and an automatic reconnection mechanism is configured (up to 3 reconnections, with an interval of 1 second between each reconnection) to ensure that the data transmission channel remains uninterrupted and to avoid data loss due to connection interruption.

[0075] Subsequently, the JSON-formatted structured difference data output by the intelligent verification execution unit 3 is received. Upon receiving the data, an integrity check is immediately performed—by comparing the preset data length with the checksum transmitted synchronously, it is confirmed that the data has not been lost or tampered with. If the check fails, a data retransmission request is immediately returned to resolutely prevent incomplete data from entering the subsequent processing stage.

[0076] Then, the difference data that passed the integrity check. Standardize the format, focusing on unifying the specifications for three types of core information: integrate deviation position information into a standard expression of "page number-paragraph number-character start position-character end position"; convert deviation type codes into easy-to-understand text descriptions (such as "calendar invalid" or "terminology non-compliant"); and organize rule reference identifiers into a unified structure of "rule set version-validation dimension-core parameters" to completely eliminate data format differences between different modules and ensure that subsequent processing does not require secondary adaptation.

[0077] Finally, the standardized difference data is classified and stored in the result processing cache pool according to the deviation type (date deviation, terminology deviation, verification anomaly). At the same time, a data processing completion identifier is generated, and the deviation information parsing and annotation module 42 is triggered to start working synchronously, ensuring the continuity and efficiency of data flow.

[0078] The deviation information parsing and annotation module 42, based on the standardized difference data, parses the page number, paragraph number, and character position of the electronic document corresponding to the deviation, and generates accurate deviation location annotation information; Specifically, the deviation information analysis and annotation module 42, based on standardized difference data, accurately analyzes the specific location of deviations in electronic documents and generates intuitive location annotation information to help users quickly pinpoint the problem. The specific process is as follows: First, the standardized difference data is retrieved from the result processing cache pool. The core location fields such as page number, paragraph number, and character position are extracted from each deviation record. At the same time, the original structural data of the electronic document (such as paragraph division rules and character encoding format) is correlated to ensure that the parsing process is consistent with the actual structure of the document and to avoid positioning errors.

[0079] Subsequently, the location information is precisely analyzed in layers: page numbers directly map to the actual displayed page numbers of the electronic document, supporting cross-page positioning in multi-page documents; paragraph numbers are matched according to the natural paragraph order of the document (empty paragraphs are not included in the numbering to avoid positioning offset due to empty paragraphs); character positions are combined with row and column numbers to clarify the specific range of the deviation character (such as "3rd row 5th column - 3rd row 12th column"), achieving "precise positioning down to the character level".

[0080] Then, for deviations in batch fields such as tables and lists, supplement field identification information (such as "Table 1 - Column 2" and "List 3 - Item 4") to form a complete positioning chain of "Page Number - Paragraph Number - Field Identifier - Character Position". This ensures that even if there are batch repetitive deviations, the specific unit can be accurately located, avoiding user confusion.

[0081] Finally, visualized deviation location annotation information is generated, with two types of annotation formats: text labels are presented as "[Deviation Type] + Complete Location Description" (e.g., "[Invalid Calendar] Page 5, Paragraph 2, Table 1-Column 3, Characters 4-8"); graphic labels are highlighted marks at the corresponding locations in the electronic document (different colors are used for different deviation types: red for date deviations, yellow for terminology deviations, and orange for verification anomalies). The annotation information is then compared with the original deviation data. Associated storage provides precise location support for generating correction guidelines.

[0082] The correction guidance generation module 43 matches the preset correction strategy library according to the deviation type and the corresponding rule reference identifier, and generates a targeted correction guidance that is adapted to the deviation type. The correction guidance includes a description of the deviation problem and a compliance format example. Specifically, the correction guidance generation module 43 matches the preset correction strategy library based on the deviation type and rule reference identifier to generate targeted and immediately applicable correction guidance, allowing users to clearly understand "what the problem is and how to fix it." The specific process is as follows: First, initialize the preset correction strategy library. This library stores standardized correction solutions categorized by deviation type, covering all possible deviation scenarios. Each strategy in the library contains four core elements: deviation type, rule basis, problem description template, and compliance example template. It also supports dynamic optimization based on industry standards and user feedback to ensure the professionalism and practicality of the guidance.

[0083] Subsequently, the deviation type and rule reference identifier of each deviation record are extracted, and the corresponding verification rule is associated with the rule reference identifier (e.g., date deviation is associated with calendar service verification standards, terminology deviation is associated with verification standards based on...). The compliance determination rules ensure that the revision guidelines and verification rules are highly consistent, avoiding situations where "guidelines and rules conflict".

[0084] Then, based on the deviation type and the association rule matching correction strategy library, personalized correction guidance is generated: the deviation problem description section is generated by combining the specific content of the deviation with the template (e.g., "Date field 'February 30, 2024' does not conform to the Gregorian calendar rules and the date does not exist"), "The contextual explanation of the term 'API' is missing, and the compliance score is lower than the judgment threshold." The compliance format example section provides 1-2 directly replaceable compliance solutions (such as "Compliance Example 1: February 29, 2024; Compliance Example 2: March 1, 2024" and "Compliance Example: API (Application Programming Interface)"), ensuring that users can complete the modification without having to look up the standards separately.

[0085] Finally, the generated correction guidelines are validated for reasonableness, checking whether the problem descriptions are clear and unambiguous and whether the compliance examples are accurate and effective. At the same time, the correction guidelines are bound to the corresponding deviation location labeling information to form a complete set of related data of "location-problem-solution", providing a complete and orderly basic content for report integration.

[0086] The report integration and output module 44 is used to integrate deviation positioning and annotation information with targeted correction guidelines according to a preset template, generate a formatted correction suggestion report with standardized format, and supports multiple output formats for users to choose from.

[0087] Specifically, the report integration and output module 44 integrates the deviation location annotation information and targeted correction guidelines according to the standard template, generating a format correction suggestion report with a unified format and complete content, and supports multi-format output to meet the needs of different users' usage scenarios. The specific process is as follows: First, load the preset report template. The template structure is designed with reference to industry document verification report specifications and includes five core modules: Cover (including document name, report generation time, and rule set version), Table of Contents (supporting quick navigation to corresponding deviation items), Deviation Summary Table (statistics the number of deviations by type and page numbering), Detailed Deviations and Correction Guidelines (core content, presenting each deviation with location annotations, problem descriptions, and compliance examples), and Remarks (including report usage instructions, rule references, and clarifications). , (The value retrieval logic) ensures that the report structure is clear and logically coherent.

[0088] Subsequently, the associated data of "location annotation information + correction guidance" is sorted according to the severity of the deviation (compliance risk deviations take priority, followed by format deviations), and then populated into the corresponding modules of the report template. At the same time, a table of contents and page numbers are automatically generated to ensure that users can quickly locate the target deviation through the table of contents, thus improving the efficiency of report retrieval.

[0089] Next, the integrated report content is checked for format compliance: the font, font size, and line spacing are checked to see if they conform to the template standard (title in bold, size 2; body text in Song font, size 4; line spacing 1.5); the table structure is checked to see if it is complete and the data statistics are accurate; all deviation items are confirmed to be complete and the location information corresponds one-to-one with the correction guidelines to avoid errors of "misattribution".

[0090] Finally, the system offers several commonly used output formats for users to choose from, including PDF (suitable for archiving and cross-platform viewing, ensuring formatting integrity), Word (allowing users to directly reference compliance examples and modify documents within the report), and Excel (suitable for batch statistical analysis of deviation data and exporting data for analysis). It also supports online preview, local download, and printing. After the user selects an output format, the system quickly generates the corresponding report file, completing the output delivery.

[0091] like Figure 3 As shown, this embodiment also provides an intelligent proofreading method for electronic document formats based on multi-dimensional verification rules. The intelligent proofreading system for electronic document formats based on multi-dimensional verification rules, as described above, includes the following steps: S1. Receive electronic documents in different formats, perform standardized preprocessing on the document formats to eliminate parsing differences, and then use a combination of regular expression matching and deep learning text classification to identify and extract paragraph-level semantic blocks corresponding to titles, body text, and charts. S2. Perform word vector conversion and semantic similarity calculation on paragraph-level semantic blocks, construct a document semantic topology graph by combining the citation tag matching results, generate node embedding vectors for each paragraph-level semantic block, dynamically adjust the verification dimension parameters according to the document semantic features, and integrate and generate a multi-dimensional verification rule set in JSON format containing interface call fields. S3. Parse the multi-dimensional verification rule set to generate standardized verification instructions. Perform multi-dimensional matching verification on the electronic document format metadata according to the verification instructions. Among them, the date logic verification calls the external calendar service to verify the validity, and the terminology compliance verification refers to the preset exemption threshold for judgment. Collect the verification results and extract the deviation information, and integrate them to generate difference data. S4. Standardize the format of the discrepancy data, accurately mark the document location information corresponding to the deviation, generate a correction guide with problem description and compliance examples according to the deviation type and the correction strategy library, integrate the deviation marking information and correction guide according to the preset template, and generate a format correction suggestion report with multiple formats available.

[0092] Those skilled in the art will understand that the process of implementing all or part of the steps of the above embodiments can be carried out by hardware or by a program instructing the relevant hardware.

[0093] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent proofreading system for electronic document formats based on multi-dimensional verification rules, characterized in that, include: The document structure parsing unit (1) is used to receive electronic documents and perform structure parsing on the electronic documents to extract paragraph-level semantic blocks including titles, body text, and charts. The dynamic verification rule generation unit (2) is connected to the document structure parsing unit (1). It adopts a graph neural network model to model the paragraph-level semantic blocks and constructs a document semantic topology graph with paragraph-level semantic blocks as nodes and text semantic similarity and citation tag matching results as edges. Through the embedding learning of the graph neural network, the semantic information of adjacent nodes is aggregated to generate the node embedding vector of each paragraph-level semantic block. Then, an attention network is used to detect whether there are time period constraint-related semantic features in the document content based on the node embedding vector. When semantic features containing descriptions of clinical research, project management or financial cycles are detected, the execution priority of the date logic verification dimension in the verification task queue is dynamically increased, and the call to the calendar service is triggered to verify the validity of the field calendar. When semantic features containing the context of professional term abbreviations are detected, the term usage compliance verification dimension is activated, and the term exemption verification threshold is automatically adjusted according to the context semantic confidence. Finally, a multi-dimensional verification rule set adapted to the current document is generated according to the content features of each paragraph-level semantic block, so that the verification rules evolve dynamically with the document content. The intelligent verification execution unit (3) is connected to the dynamic verification rule generation unit (2). It calls the multi-dimensional verification rule set to match and verify the electronic document format metadata. When performing date logic verification, it calls the calendar service to verify the validity of the date. When performing term compliance verification, it refers to the term exemption verification threshold and generates difference data containing deviation position, type and rule reference. Verification result output unit (4) is connected to intelligent verification execution unit (3), receives the difference data, and generates a format correction suggestion report containing deviation location information and targeted correction guidance.

2. The intelligent proofreading system for electronic document formats based on multi-dimensional verification rules according to claim 1, characterized in that, The document structure parsing unit (1) includes a format-compatible access module (11) and a paragraph-level semantic block recognition module (12), wherein: The format-compatible access module (11) is used to receive electronic documents of different formats, and complete the standardization preprocessing of document formats by calling the parsing interface of the corresponding format, thereby eliminating the parsing differences of office electronic document formats; The paragraph-level semantic block recognition module (12) is based on the standardized document output by the format-compatible access module (11), and uses a combination of regular expression matching and deep learning text classification to identify and extract paragraph-level semantic blocks including titles, body text, and charts.

3. The intelligent proofreading system for electronic document formats based on multi-dimensional verification rules according to claim 2, characterized in that, The dynamic verification rule generation unit (2) includes a semantic topology modeling module (21), a node embedding generation module (22), and a verification rule adjustment module (23) connected in sequence, wherein: The semantic topology modeling module (21) is used to model paragraph-level semantic blocks and construct a document semantic topology graph with paragraph-level semantic blocks as nodes. The node embedding generation module (22) aggregates the semantic information of adjacent nodes through the embedding learning of the graph neural network and generates the node embedding vector of each paragraph-level semantic block. The verification rule adjustment module (23) detects semantic features based on node embedding vectors and dynamically adjusts the verification dimension to generate a multi-dimensional verification rule set.

4. The intelligent proofreading system for electronic document formats based on multi-dimensional verification rules according to claim 3, characterized in that, The semantic topology modeling module (21) determines the weights of the edges of a document semantic topology graph by constructing a graph with paragraph-level semantic blocks as nodes, and the process includes the following steps: S21.1, Target paragraph-level semantic block Input each word into a pre-trained word vector model and output the corresponding word vectors; S21.2 Calculate the text semantic similarity between two paragraph-level semantic blocks using the cosine similarity algorithm. This serves as the core weighting criterion for edges; S21.3 Perform citation tag matching on paragraph-level semantic blocks to identify citation relationships between different paragraph-level semantic blocks; S21.4, Text semantic similarity The results of matching with the reference tags are weighted and fused to form the final weight basis for the corresponding edges in the document semantic topology graph.

5. The intelligent proofreading system for electronic document formats based on multi-dimensional verification rules according to claim 4, characterized in that, When generating the node embedding vector for each paragraph-level semantic block, the node embedding generation module (22) uses a graph attention network to achieve the process of embedding learning and semantic information aggregation, which includes the following steps: S22.1 Initial feature vector for each paragraph-level semantic block Perform a linear transformation to obtain the transformed eigenvectors; S22.2 Calculate paragraph-level semantic blocks With adjacent nodes Attention coefficient This characterizes the strength of the semantic association between two nodes; S22.3, Attention Coefficient Normalization is performed to obtain attention weights. This enables the differentiated aggregation of semantic information of adjacent nodes; S22.4, Attention-based weights By aggregating the semantic information of adjacent nodes, a paragraph-level semantic block is finally generated. Node embedding vector This is used for subsequent semantic feature detection.

6. The intelligent proofreading system for electronic document formats based on multi-dimensional verification rules according to claim 5, characterized in that, The verification rule adjustment module (23) includes a feature input preprocessing submodule, a semantic feature classification and detection submodule, a verification dimension dynamic configuration submodule, and a multi-dimensional rule set integration and generation submodule, wherein: The feature input preprocessing submodule is used to receive the node embedding vector output by the node embedding generation module. Standardize the vector data to eliminate dimensionality differences; The semantic feature classification and detection submodule is connected to the feature input preprocessing submodule. It uses an attention network to classify the preprocessed node embedding vectors, identify the time period constraint semantic features and the contextual semantic features of professional term abbreviations in the document content, and output the corresponding feature classification probability values. The dynamic configuration submodule for the verification dimension is connected to the semantic feature classification and detection submodule. When the probability value of the semantic feature constrained by the time period is greater than the preset feature judgment threshold, an external calendar service is called through the HTTP or HTTPS interface protocol to verify the calendar validity of the date field. At the same time, the task priority parameter of the date logical verification dimension is increased to the preset high priority parameter. The multi-dimensional rule set integration and generation submodule is connected to the dynamic configuration submodule for verification dimensions. It is used to temporarily store the adjustment parameters of the verification dimensions and provide data support for the subsequent generation of multi-dimensional verification rule sets.

7. The intelligent proofreading system for electronic document formats based on multi-dimensional verification rules according to claim 6, characterized in that, The process by which the dynamic configuration submodule for verification dimensions generates a multi-dimensional verification rule set after dynamically adjusting the verification dimensions includes the following steps: S232.1 If the probability value of the contextual semantic feature of a technical term abbreviation is greater than the preset feature judgment threshold, then the probability value shall be used as the contextual semantic confidence level. ; S232.2, Context-based semantic confidence Dynamically determine the term exemption verification threshold This enables adaptive adjustments to terminology compliance verification. S232.3, Integrated Calendar Service Call Parameters, Date Validation Priority Parameters, and Terminology Exemption Validation Thresholds Generate a multi-dimensional set of validation rules in JSON format; S232.4 Add configuration fields related to interface calls to the multi-dimensional verification rule set to realize direct data interaction with the intelligent verification execution unit (3).

8. The intelligent proofreading system for electronic document formats based on multi-dimensional verification rules according to claim 7, characterized in that, The intelligent verification execution unit (3) includes a rule set receiving and parsing module (31), a multi-dimensional verification execution module (32), and a difference data integration and generation module (33) connected in sequence, wherein: The rule set receiving and parsing module (31) is used to establish a data connection with the dynamic verification rule generation unit (2), receive the multi-dimensional verification rule set, parse the verification dimension identifier, parameter data type and interface call field in the multi-dimensional verification rule set, and generate standardized verification instructions. The multi-dimensional verification execution module (32) performs matching verification on the electronic document format metadata based on standardized verification instructions. When performing date logic verification, it calls an external calendar service to complete date validity verification. When performing terminology compliance verification, it refers to the terminology exemption verification threshold in the rule set to determine compliance. The difference data integration and generation module (33) is used to collect the verification results of the multi-dimensional verification execution module (32), extract the deviation location information, deviation type and corresponding rule reference identifier, integrate and generate difference data according to the preset data structure, and provide data input for the verification result output unit (4).

9. The intelligent proofreading system for electronic document formats based on multi-dimensional verification rules according to claim 8, characterized in that, The verification result output unit (4) includes a difference data receiving and processing module (41), a deviation information parsing and annotation module (42), a correction guidance generation module (43), and a report integration and output module (44) connected in sequence, wherein: The difference data receiving and processing module (41) establishes a data connection with the intelligent verification execution unit (3), receives difference data, and performs format standardization processing on the deviation location information, deviation type and rule reference identifier in the difference data to eliminate data format differences; The deviation information parsing and annotation module (42) analyzes the electronic document page number, paragraph number and character position corresponding to the deviation based on the standardized difference data, and generates accurate deviation positioning annotation information. The correction guidance generation module (43) matches the preset correction strategy library according to the deviation type and the corresponding rule reference identifier, and generates a targeted correction guidance that is adapted to the deviation type. The correction guidance includes a description of the deviation problem and a compliance format example. The report integration and output module (44) is used to integrate the deviation positioning annotation information and the targeted correction guidance according to the preset template, generate a format correction suggestion report with a standard format, and support multiple output formats for users to choose from.

10. A method for intelligent proofreading of electronic document formats based on multi-dimensional verification rules, based on the intelligent proofreading system for electronic document formats based on multi-dimensional verification rules as described in any one of claims 1-9, characterized in that, Includes the following steps: S1. Receive electronic documents in different formats, perform standardized preprocessing on the document formats to eliminate parsing differences, and then use a combination of regular expression matching and deep learning text classification to identify and extract paragraph-level semantic blocks corresponding to titles, body text, and charts. S2. Perform word vector conversion and semantic similarity calculation on paragraph-level semantic blocks, construct a document semantic topology graph by combining the citation tag matching results, generate node embedding vectors for each paragraph-level semantic block, dynamically adjust the verification dimension parameters according to the document semantic features, and integrate and generate a multi-dimensional verification rule set in JSON format containing interface call fields. S3. Parse the multi-dimensional verification rule set to generate standardized verification instructions. Perform multi-dimensional matching verification on the electronic document format metadata according to the verification instructions. Among them, the date logic verification calls the external calendar service to verify the validity, and the terminology compliance verification refers to the preset exemption threshold for judgment. Collect the verification results and extract the deviation information, and integrate them to generate difference data. S4. Standardize the format of the discrepancy data, accurately mark the document location information corresponding to the deviation, generate a correction guide with problem description and compliance examples according to the deviation type and the correction strategy library, integrate the deviation marking information and correction guide according to the preset template, and generate a format correction suggestion report with multiple formats available.

Images