Government affair material intelligent verification method and system based on machine learning

Abstract

The invention relates to the technical field of material verification management, and discloses a government affair material intelligent verification method and system based on machine learning, and the method comprises the steps: obtaining the text information data and verification rules of historical materials; multi-modal feature extraction is carried out; establishing a dynamic verification rule knowledge graph, and verifying the multi-modal features and the target feature points of the text segments to generate a primary verification strategy; performing deep verification processing on the material text fragment to generate a secondary verification strategy; performing third-time verification on the material text fragment subjected to the second-time verification by triggering cross-page feature matching to realize cross-page text fragment association and multi-modal feature fitting so as to generate a third-time verification strategy; performing iterative comparison on the results of the three rounds of verification to generate a credibility score; and the comparison result is adjusted according to the credibility score in combination with the verification risk abnormity assessment, so that a final verification decision strategy can be obtained, and efficient, accurate, dynamic and comprehensive intelligent verification can be realized.

Description

Technical Field

[0001] The present invention relates to the technical field of material verification management, and more specifically, to a method and system for intelligent verification of government materials based on machine learning. Background Art

[0002] With the deepening of the digital transformation of government services, some government materials are currently verified through intelligent verification. The existing government material verification methods have problems such as insufficient multimodal feature extraction, static verification rules, insufficient deep verification processing, missing cross-page feature matching, lack of multi-dimensional cross-validation, and imperfect risk assessment mechanism. It is impossible to achieve efficient, accurate, dynamic and comprehensive intelligent verification of government materials.

[0003] Therefore, how to achieve efficient, accurate, dynamic and comprehensive intelligent verification has become a technical problem that needs to be solved urgently. Summary of the Invention

[0004] The present invention provides a method and system for intelligent verification of government materials based on machine learning, which solves the technical problem of how to achieve efficient, accurate, dynamic and comprehensive intelligent verification in the existing technology.

[0005] The present invention provides a method and system for intelligent verification of government documents based on machine learning, comprising the following steps: First, a method for intelligent verification of government documents based on machine learning, including: Obtain text information data and verification rules for historical materials; Perform multimodal feature extraction on text information data; Establish a dynamic verification rule knowledge graph to verify the multimodal features of the structured text fragments and target feature points in the current material to be verified to generate an initial verification strategy; Conduct in-depth verification of the material text fragments after the initial verification, obtain the verification results of related feature words and attribute descriptions, optimize the initial verification strategy, and generate a secondary verification strategy; By triggering cross-page feature matching, the text fragments of the material that have been verified twice are verified for a third time, so as to achieve cross-page text fragment association and multimodal feature fitting, and thus generate a three-time verification strategy; The results of the three rounds of verification are iteratively compared, and a multi-dimensional cross-validation is constructed, combining the verification conditions of multiple data sources to generate a credibility score; the comparison results are adjusted based on the credibility score and the verification risk anomaly assessment to obtain the final verification decision strategy.

[0006] Furthermore, multimodal feature extraction is performed on the text information data, including: Preprocessing the material, including noise removal and layout analysis, to obtain a preprocessed data set; An improved ResNet-101 network is used to extract multi-level features from the preprocessed dataset, and a multimodal spatiotemporal alignment network is used to perform three-level feature fusion on the multi-level features. The multi-level features include image pixel level, text semantic level, and spatiotemporal rule level. A three-level feature fusion mechanism is used to identify and associate image and text position information using a spatial attention mechanism to obtain multimodal features. The text semantic level fusion establishes semantic association between simple semantic data and complex semantic data through graph convolutional network. The multimodal features include a comprehensive representation of image features, text features and time features; The image features are complemented by simulating the complex situations of seal overlap, watermark interference and blurred text that are unique to government documents.

[0007] Furthermore, the multimodal features are standardized and normalized to obtain a consistent and complete multimodal feature representation to obtain a feature set; Extract text feature vectors from the feature set through the BERT model; Dynamically convert verification rules into verification rule vectors; Construct a material type recognition model based on text feature vectors and verification rule vectors; Automatically identify and judge the material category through the material type recognition model; A dynamic verification rule knowledge graph is established based on various verification rules obtained historically, and corresponding government documents are verified according to different verification rules; The verification rules include logical coherence between multimodal features, context segment association information, and keyword and word attribute constraints; And obtain the corresponding material verification strategy based on the material category and the corresponding verification rules.

[0008] Furthermore, the initial verification strategy includes: Identify the document structure through layout analysis of the materials to be verified to obtain structured text fragments; Extract the feature set of structured text fragments, verify the rule relationships and constraints, and obtain the extraction results; Add validity and expiration timestamp attributes to the extraction results to detect changes in verification rules and update verification constraints in real time; Develop a semantic differential engine to identify the impact of changes to verification rules. Obtain and record the trajectory of policy changes to establish a verification rule regulation mechanism; in, is the impact range index, is the node influence weight of the k-th rule, is the change range of the kth rule, is the node connectivity, is the distance attenuation parameter, is the semantic distance between nodes; Based on the constraint conditions, the dynamic verification rule knowledge graph is used to detect the verification rule conflicts. Obtain conflict scores and perform operations to implement version control and conflict detection of verification rules, proactively identifying inconsistencies in cross-departmental verification rules; in, Score the conflict. is the interaction weight of rules i and j, is the rule intersection size, is the rule contradiction function; Detect conflicts between verification rules based on conflict scores to obtain conflicting points in the verification rules; By establishing a rule priority system for conflicting points, the rule priority system is Obtain the rule priority index and handle conflicts and dependencies between verification rules in descending order based on the different rule priority indices. in, is the priority index of the i-th rule, 、 、 and are weight coefficients, For authoritative rating, is the frequency of use, Score timeliness. Rate the importance.

[0009] Furthermore, the initial verification strategy further includes: Based on the feature vocabulary and position description in the feature set, the target feature points are located in the structured text fragment; Apply each feature point to the corresponding attribute constraint rules to check whether it meets the format, range and logic requirements; Analyze the spatial position relationship and semantic association between feature points, and generate an initial verification strategy including verification path, verification result and confidence level; the verification confidence level is determined by Get; in, To verify the confidence, include any natural number between 0 and 1. is the matching probability of the i-th feature point, is the corresponding weight, To verify the error, is the standard deviation of error, represents an exponential function used to calculate the confidence decay factor based on the validation error; The target feature points include the spatial position relationship of the text segment; the context relationship and the position in the table; the feature words and attribute constraints; Among them, feature words include key field names and identifiers; attribute constraints include text description information of numerical ranges, date formats, and required items.

[0010] Furthermore, the text fragments of the material after the initial verification are deeply verified, and the verification results of the associated feature words and attribute descriptions are obtained to optimize the initial verification strategy to generate a secondary verification strategy, including: Performing segmented verification on the text segments based on punctuation marks, wherein the segmented verification includes identifying different punctuation marks in the text; Segment the text segment into multiple semantic units according to the semantic functions of different punctuation marks; Assign a unique identifier to each semantic unit and establish a tag signal index; Analyze the nested relationship between punctuation marks based on the tag signal index and construct the hierarchical structure of semantic units; Conduct semantic understanding of the initially verified material text based on the hierarchical structure of semantic units to obtain the logical relationships and constraints of the text fragments; By identifying and extracting key feature words and their attribute relationships in the text, a feature association network is constructed; The feature association network is matched and verified with the verification rule knowledge graph. The matching verification is passed. Get semantic matching degree; Preset a matching threshold, compare the semantic matching with the matching threshold, and if the comparison result shows a high semantic matching value, obtain a semantic feature relationship of compliance and consistency; if the semantic matching value is low, return to the step of constructing a feature association network and reconstruct it; in, is the semantic matching degree, is the weight of the nth feature, is the eigenvector, is a regular vector, is the similarity function; Apply each semantic unit to the corresponding verification rules according to the semantic feature relationship to generate fine-grained verification; Based on fine-grained verification combined with the semantic relationships implied by punctuation, the system verifies the logical consistency between related semantic units. It also records the verification time, applicable verification rule version, and meta-information between text segment features to generate a traceable verification path. The positioning and verification process of target feature points is obtained through fine-grained verification, forming a traceable verification evidence chain and obtaining in-depth verification results; Assign a confidence score between 0 and 1 to each deep verification result; Based on the verification results with a confidence score ≤ 0.6, the verification evidence chain that does not meet the verification rules will be classified as abnormal. Abnormal markings include missing, format errors, and logical conflicts. Integrate verification paths, confidence scores, and anomaly flags to form a secondary verification strategy.

[0011] Furthermore, the third verification of the material text fragment after the second verification is carried out by triggering cross-page feature matching to achieve cross-page text fragment association and multimodal feature fitting, so as to generate a three-verification strategy, including: Extract associated feature points from multiple pages of materials, including header and footer information, chapter numbers, and table continuation marks, to obtain an associated feature point dataset; Convert the extracted feature points into vector representation and construct a feature vector array; Based on the eigenvector array via: ; Obtain the similarity between feature vectors of different pages to obtain the cross-page similarity threshold; in, is the cross-page similarity threshold, For the feature vector weights, is the vector angle, For the hash feature weights, is the Hamming distance between the features of the mth and m+1th adjacent pages, is the page distance penalty coefficient, is the distance between pages; Determine whether there is a relationship between pages based on the cross-page similarity threshold and determine the type of relationship; Logically combine the identified cross-page related text fragment association types to reconstruct the text information; Based on the reconstructed text information, the image features, text features, and time features of the cross-page text in the second-verified material text fragment are comprehensively fitted to obtain a comprehensive fitting result to eliminate the defects of text page information fragmentation; Verify the semantic and logical consistency of the reconstructed text information based on the comprehensive fitting results and obtain the cross-page matching results; Based on the cross-page matching results, the first two rounds of verification strategies are integrated to form a three-round verification strategy with cross-page association.

[0012] Furthermore, the results of the three rounds of verification are iteratively compared, and a multi-dimensional cross-validation system is constructed, combining verification conditions from multiple data sources to generate a credibility score, including: Compare the text information after three verification strategies and check whether there are any contradictions or inconsistencies in the material text information to obtain the inspection results, which include numerical consistency, logical consistency and expression consistency; When the inspection results are abnormal, the material text information after three rounds of verification is iteratively calibrated to obtain the calibration results. Based on the calibration results, verify again whether the time information in the material complies with the time logic, including date sequence, timeliness requirements and historical time point verification, and obtain a round of verification results; Based on the results of the first round of verification, the authenticity and authority of the source of the materials are verified by querying authoritative databases, blockchain evidence storage, and government information sharing platforms to obtain the results of the second round of verification; Based on the results of the second round of verification, the anti-counterfeiting marks, electronic signatures, watermarks and special printing features in the materials are identified and verified, and: To obtain the credibility of the anti-counterfeiting feature; judge whether the material is forged or tampered according to the anti-counterfeiting credibility, and obtain the results of three rounds of verification; in, For anti-counterfeiting credibility, For the The credibility of the anti-counterfeiting mark, is the corresponding weight of the credibility of the anti-counterfeiting mark, For the The credibility of an electronic signature, is the corresponding weight of the electronic signature credibility, For the The credibility of the watermark feature, is the corresponding weight of the watermark feature credibility; Integrate the results of the first three rounds of verification to form a multi-dimensional cross-validation, and combine the multi-dimensional cross-validation with the verification conditions of three multi-data sources: internal government data, third-party trusted data, and historical verification records, to generate a comprehensive credibility score for the material; The comprehensive credibility score organizes the verification results of each dimension into a complete chain of evidence according to chronological order and logical relationships, supporting the interpretability of the verification results. Based on the completeness of the chain of evidence, the verification results of each dimension, and the authority of the verification data source, a comprehensive credibility score of the material is obtained. pass Get a comprehensive credibility score; where, For the overall credibility score, is the weight of the i-th dimension, For the The verification results of the dimensions, For the The data source authority score of each dimension, For the The time weight of each dimension, is the variance penalty coefficient, is the mean of the verification results, is the integrity weight of the evidence chain, Score the completeness of the chain of evidence.

[0013] Furthermore, based on the credibility score requirements and risk level adaptive adjustment, the final verification decision strategy is comprehensively generated, including: Based on the material type, application scenario, and business impact, the comprehensive credibility score of the current verification material text is graded into three levels: critical, important, and general; Comprehensively consider the various influencing factors involved in the material and the combined matrix of business impact and risk level, through: Assess the potential level of risk; in, is the risk level assessment value, is the weight of the fth influencing factor, For the degree of impact, is the probability of occurrence, is the weight of the rth severity factor, For severity, is the frequency, is the uncertainty coefficient, is the uncertainty factor; Automatically adjust the credibility score threshold of the verification results based on the risk level assessment of the currently verified material text, setting higher credibility score requirements for important and high-risk businesses; Set risk-based manual review trigger conditions. When the credibility of the verification result is lower than the threshold but the business importance is high, it will be automatically marked as requiring manual review. Based on the adjusted credibility score requirements, decide whether to enable a more in-depth verification process that requires manual review; Integrate the three rounds of verification results, credibility scores, and adjusted threshold requirements to generate a final verification decision strategy that includes verification conclusions, confidence levels, anomaly flags, and handling recommendations; pass Perform threshold adjustment operations; in, is the adjusted threshold, is the basic threshold, is the risk sensitivity coefficient, is the risk level assessment value, is the business importance coefficient, Score the business importance, is the historical error rate coefficient, is the historical error rate; Based on the verification conclusions and verification rules obtained through four rounds of verification strategies, we provide execution recommendations of passing, passing with supplementary materials, or rejecting the application, and give specific reasons and re-verification operation instructions.

[0014] In a second aspect, a system for intelligently verifying government documents based on machine learning is provided, which is used to implement a method for intelligently verifying government documents based on machine learning, including: Data collection module: used to obtain text information data and verification rules of historical materials; Feature extraction module: It is used to extract multimodal features from text information data; Initial verification module: This module is used to establish a dynamic verification rule knowledge graph, verify the multimodal features of the structured text fragments and target feature points in the current material to be verified, and generate an initial verification strategy; Secondary Verification Module: It is used to conduct in-depth verification of the material text fragments after the initial verification, obtain the verification results of the associated feature words and attribute descriptions, optimize the initial verification strategy, and generate a secondary verification strategy; Three-time verification module: It is used to conduct a third verification on the material text fragments after the second verification by triggering cross-page feature matching, so as to achieve cross-page text fragment association and multimodal feature fitting, so as to generate a three-time verification strategy; Credibility Assessment Module: This module is used to iteratively compare the results of the three rounds of verification, build multi-dimensional cross-validation, and combine verification conditions from multiple data sources to generate a credibility score. Final verification module: It is used to adjust the comparison results based on the credibility score and the verification risk anomaly assessment to obtain the final verification decision strategy.

[0015] The beneficial effects of the present invention are as follows: the present invention obtains text information data and verification rules of historical materials; performs multimodal feature extraction on text information data; establishes a dynamic verification rule knowledge map, verifies the multimodal features and target feature points of structured text fragments in the current material to be verified, so as to generate an initial verification strategy; performs in-depth verification processing on the material text fragments after the initial verification, obtains the verification results of associated feature words and attribute descriptions, optimizes the initial verification strategy, and generates a secondary verification strategy; performs a third verification on the material text fragments after the secondary verification by triggering cross-page feature matching, so as to realize cross-page text fragment association and multimodal feature fitting, so as to generate a tertiary verification strategy; iteratively compares the three rounds of verification results, and constructs a multi-dimensional cross-validation, combined with multi-data source verification conditions The system generates a credibility score based on the credibility score; adjusts the comparison results based on the credibility score and the verification risk anomaly assessment to obtain the final verification decision strategy, thereby improving the accuracy of cross-modal consistency verification and effectively solving the problem of insufficient multi-modal feature extraction; the spatiotemporal knowledge modeling and version control mechanism based on the dynamic policy knowledge graph greatly shortens the adaptation time of new policies and solves the problem of static solidification of verification rules; the hierarchical update strategy reduces the model performance decay rate, improves stability and in-depth verification processing capabilities; ensures accuracy through credibility scoring and improves the risk assessment mechanism; reduces the material association error rate, improves the cross-page association accuracy, and solves the problem of missing cross-page feature matching through cross-page feature matching; realizes multi-dimensional cross-validation and improves the comprehensiveness and accuracy of government material verification. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flow chart of a method for intelligent verification of government documents based on machine learning provided in an embodiment of the present invention; Figure 2 It is a schematic diagram of a module of a government affairs material intelligent verification system based on machine learning provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0017] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0018] At least one embodiment of the present invention discloses a method and system for intelligent verification of government documents based on machine learning, including: like Figure 1As shown, a method for intelligent verification of government documents based on machine learning includes the following steps: Step 1: Obtain text information data and verification rules of historical materials.

[0019] First, the text data of historical government documents is obtained from the government service database, including but not limited to application forms, supporting documents, identity certificates, and other government documents. Simultaneously, the corresponding verification rules are obtained from the policy library. These rules contain the standards and requirements for verifying government documents. By establishing a historical materials library and a verification rules library, the basic data support for subsequent intelligent verification is provided.

[0020] Step 2: Perform multimodal feature extraction on text information data.

[0021] This step first preprocesses the material, including noise removal and layout analysis, to obtain a preprocessed dataset. Then, an improved ResNet-101 network is used to extract multi-level features from the preprocessed dataset, and a three-level feature fusion mechanism is implemented through the Multimodal Spatiotemporal Alignment Network (MSTA-Net), including image pixel-level fusion, text semantic-level fusion, and spatiotemporal rule-level fusion. MSTA-Net uses a spatial attention mechanism to identify and associate image and text position information to obtain multimodal features. Among them, text semantic-level fusion establishes semantic associations between simple semantic data and complex semantic data through a graph convolutional network. The extracted multimodal features include a comprehensive representation of image features, text features, and temporal features. In particular, image features effectively improve the accuracy of feature extraction by simulating and completing the complex situations of seal overlap, watermark interference, and blurred text that are unique to government materials.

[0022] Step 3: Establish a dynamic verification rule knowledge graph to verify the multimodal features of the structured text fragments and target feature points in the current material to be verified to generate an initial verification strategy.

[0023] This step first standardizes and normalizes the multimodal features to obtain a consistent and complete multimodal feature representation to obtain a feature set. Then, the feature set is used to extract the text feature vector through the BERT model, and the verification rules are dynamically converted into verification rule vectors. A material type recognition model is constructed based on the text feature vector and the verification rule vector, and the material category is automatically identified and judged through the model. A dynamic verification rule knowledge graph (DPKG) is established based on the various verification rules obtained historically, and different verification rules verify the corresponding government materials. The verification rules include logical coherence between multimodal features, context segment association information, keywords and word attribute constraints, and the corresponding material verification strategy is obtained according to the material category and the corresponding verification rules; the reinforcement learning algorithm is used to optimize the verification strategy selection process, and the reward function is used to guide the model to select the optimal verification path.

[0024] In the initial verification strategy, the document structure of the current material to be verified is identified through layout analysis to obtain structured text fragments; the feature set, verification rule relationship and constraint conditions of the structured text fragment are extracted to obtain the extraction results; the effective and abolishment timestamp attributes are added to the extraction results to perceive the changes in verification rules and update the verification constraints in real time; a semantic differential engine is developed to identify the scope of impact of verification rule changes and record the trajectory of policy changes to establish a verification rule regulation mechanism; verification rule conflict detection is performed on the dynamic verification rule knowledge graph based on constraint conditions to realize version control and verification conflict detection of verification rules, and actively discover contradictions in cross-departmental verification rules; by establishing a rule priority system for the contradictions, conflicts and dependencies between verification rules are handled; graph neural network technology is used to reason on the knowledge graph to realize automatic learning and inference of complex rule relationships.

[0025] In addition, the initial verification strategy also includes: locating target feature points in structured text fragments based on the feature word list and position descriptions in the feature set; applying each feature point to the corresponding attribute constraint rules to check whether it meets format, scope, and logic requirements; analyzing the spatial position relationship and semantic association between feature points to generate an initial verification strategy that includes a verification path, verification results, and confidence level. The target feature point includes the spatial position relationship of the text fragment, contextual relationship, position within the table, feature words, and attribute constraints. Feature words include key field names and identifiers; attribute constraints include textual descriptions of numerical ranges, date formats, and required items.

[0026] Step 4: Perform in-depth verification on the material text fragments after the initial verification, obtain the verification results of related feature words and attribute descriptions, optimize the initial verification strategy, and generate a secondary verification strategy.

[0027] This step verifies the text fragments by punctuation marks, and the segmented verification includes identifying different punctuation marks in the text; dividing the text fragments into multiple semantic units according to the semantic functions of different punctuation marks; assigning a unique identifier to each semantic unit and establishing a tag signal index; analyzing the nested relationship between punctuation marks based on the tag signal index and constructing a hierarchical structure of semantic units; performing semantic understanding on the material text after the initial verification based on the hierarchical structure of semantic units to obtain the logical relationship and constraint conditions of the text fragments; constructing a feature association network by identifying and extracting key feature words and their attribute relationships in the text; matching and verifying the feature association network with the verification rule knowledge graph to obtain semantic feature relationships of compliance and consistency; Among them, a matching threshold is preset, and the semantic matching degree is compared with the matching threshold. If the comparison result shows that the semantic matching degree value is high, the semantic feature relationship of compliance and consistency is obtained. If the semantic matching degree value is low, the step of constructing the feature association network is returned to reconstruct, specifically including: First, the system sets a semantic matching threshold based on historical verification data and expert experience. This threshold is typically set between 0.75 and 0.85 to balance verification accuracy and processing efficiency. The threshold is determined by taking into account the complexity of government documents, the strictness of verification rules, and the system's fault tolerance.

[0028] After the feature association network is constructed, the system calculates the semantic match score for each feature association pair. This score is calculated by using a pre-trained semantic similarity model to calculate the semantic similarity between the feature word and the verification rule; weighting the similarity based on contextual information and domain knowledge; and considering the importance weight of the feature word and its position in the document. The final result is a semantic match value between 0 and 1.

[0029] Next, the system compares the calculated semantic matching degree with the preset matching degree threshold: When the semantic match value exceeds the threshold, it indicates a good match between the feature association network and the verification rule knowledge graph, and the system confirms the compliance and consistency of the semantic feature relationship. At this point, the system marks the feature association as "verified" and incorporates it into the subsequent fine-grained verification process. The system also records information about successfully matched feature pairs to optimize subsequent feature extraction and matching processes.

[0030] When the semantic match value falls below the threshold, it indicates that the currently constructed feature association network is inaccurate or incomplete, possibly due to insufficient feature extraction, semantic understanding errors, or verification rule mismatches. At this point, the system triggers a reconstruction mechanism, returning to the step of building the feature association network for reconstruction. The reconstruction process includes: reanalyzing the semantic structure of the text fragment and adjusting the feature word extraction strategy; expanding the feature search scope to consider more contextual information; using different semantic analysis algorithms for feature association; and introducing artificial intelligence-assisted mechanisms for in-depth analysis of complex semantics.

[0031] To prevent infinite refactoring cycles, the system sets a maximum refactoring limit (usually three). If the matching threshold cannot be met after multiple refactoring attempts, the system marks the text segment as "requiring manual review" and generates a detailed exception report, including the specific reasons for the matching failure, the refactoring methods attempted, and the recommended solution.

[0032] In addition, the system also implements an adaptive threshold adjustment mechanism. By analyzing historical verification results and user feedback, the system can dynamically adjust the matching threshold to adapt to the characteristics of different types of government documents and changes in verification requirements. This adaptive mechanism significantly improves the system's flexibility and applicability, ensuring the quality and efficiency of in-depth verification processing. Based on the semantic feature relationship, each semantic unit is individually applied to the corresponding verification rule to generate fine-grained verification; based on the fine-grained verification combined with the semantic relationship implied by punctuation marks, the logical consistency between related semantic units is verified; and the verification time, applicable verification rule version and meta-information between text fragment features are recorded to generate a traceable verification path; based on the fine-grained verification, the positioning and verification process of the target feature points are obtained to form a traceable verification evidence chain and obtain the deep verification result; a confidence score between 0 and 1 is assigned to each deep verification result.

[0033] Step 5: Trigger cross-page feature matching to conduct a third verification on the material text fragments after the second verification, so as to achieve cross-page text fragment association and multimodal feature fitting, and generate a three-verification strategy.

[0034] This step utilizes a Fourier transform-based cross-page feature matching algorithm to implement association analysis of cross-page materials. First, the page number identifiers and continuity indicators in the materials are identified to establish a preliminary association between pages. Next, the feature vectors of each page are extracted, including header and footer information, document structure features, and content features. Fourier transforms are used to convert the feature vectors to the frequency domain to reduce noise interference. Similarity matrices between feature vectors of different pages are calculated to identify potentially related pages. Finally, content semantic analysis is used to verify the rationality of the association and ensure the accuracy of cross-page feature matching. This step effectively solves the problem of missing cross-page feature matching in traditional methods and significantly improves the integrity of material verification.

[0035] Step 6: Iteratively compare the three rounds of verification results and construct a multi-dimensional cross-validation, combining the verification conditions of multiple data sources to generate a credibility score; adjust the comparison results based on the credibility score and the verification risk anomaly assessment to obtain the final verification decision strategy.

[0036] This step achieves secure and efficient multi-dimensional cross-validation through the Federated Incremental Learning Framework (FIL-Gov). First, the results of the three rounds of verification are iteratively compared to identify points of consistency and discrepancies. Then, a multi-dimensional cross-validation is achieved through an image-text-policy triple alignment mechanism. External authoritative data sources are then introduced for supplementary verification to enhance the comprehensiveness of the verification process. A credibility score is calculated based on the verification results, taking into account factors such as verification consistency, data source reliability, and rule applicability. Finally, combined with a risk assessment model, high-risk materials are subject to additional review to ensure the accuracy and security of verification decisions.

[0037] Through the above steps, the present invention realizes efficient, accurate, dynamic and comprehensive intelligent verification of government materials, and effectively solves technical problems existing in the prior art, such as insufficient multimodal feature extraction, static solidification of verification rules, insufficient deep verification processing, lack of cross-page feature matching, lack of multi-dimensional cross-validation, and imperfect risk assessment mechanism.

[0038] Example 2 Based on Example 1, in step 2, more efficient feature extraction is achieved through an improved ResNet-101 network and a multimodal spatiotemporal alignment network (MSTA-Net).

[0039] Specifically, the document is first pre-processed, including noise removal and layout analysis. Noise removal uses an adaptive filtering algorithm to effectively remove common scanning noise, seal overlap, and watermark interference in government documents. Layout analysis uses an improved document segmentation algorithm to accurately identify the document's structural elements, such as the title, body text, tables, and signature area.

[0040] After preprocessing, a modified ResNet-101 network is used to extract multi-level features. By adding an attention mechanism and domain adaptation layers, this network improves its ability to extract unique features of government documents. In particular, the accuracy of extracting unique elements of government documents, such as blurred text, special seals, and table structures, is significantly improved.

[0041] The multimodal spatiotemporal alignment network implements a three-level feature fusion mechanism: 1. Image pixel-level fusion: Align image features with text position information through a spatial attention mechanism to accurately locate key areas in the document; 2. Text semantics fusion: Using graph convolutional networks to associate policy terms with text semantics, achieving complex semantic understanding and matching; 3. Spatial-temporal rule-level fusion: Filter valid policy versions based on the time gating mechanism to ensure the timeliness of verification rules.

[0042] This multi-level fusion mechanism improves the accuracy and comprehensiveness of feature extraction, providing high-quality feature input for subsequent verification steps. Multimodal feature extraction methods have achieved improvements in both accuracy and processing speed.

[0043] Example 3 On the basis of Example 1 and Example 2, step 3 realizes the dynamic update and intelligent adaptation of the verification rules based on the verification rule management method of the dynamic policy knowledge graph.

[0044] First, multimodal features are standardized and normalized to ensure that features from different sources and types can be compared and analyzed under a unified standard. Then, text feature vectors are extracted using the BERT model, which has been pre-trained on government data and has a deep understanding of government text.

[0045] Dynamic verification rule conversion involves converting text-based verification rules into structured rule vectors through semantic parsing, enabling machines to accurately understand and execute the rules. Based on text feature vectors and rule vectors, a material type recognition model was constructed. This model, employing a multi-layer perceptron architecture, can accurately identify different types of government documents.

[0046] The dynamic verification rule knowledge graph not only stores the content of the verification rules, but also records the relationship between the rules, applicable conditions, and timeliness information. It has the following characteristics: 1. Add effective / invalid timestamp attributes to each policy entity to ensure the timeliness of the rules; 2. Use differential coding technology to record policy revisions, making it easier to trace the history of rule changes; 3. Automatically identify the scope of impact of policy changes and accurately assess the impact of rule changes; 4. Based on the constraint satisfaction theory, identify cross-departmental policy contradictions and improve the consistency of rules.

[0047] The knowledge graph (DPKG) can automatically generate the most appropriate verification strategy based on material categories and currently valid verification rules, significantly improving verification accuracy and adaptability. Experiments have shown that compared to static rule bases, DPKG improves rule update response time and rule adaptation accuracy.

[0048] Example 4 On the basis of Examples 1-3, step 3 implements the structured text fragment processing method based on the initial verification strategy to achieve accurate structured analysis and preliminary verification of government materials.

[0049] First, layout analysis technology is used to identify the document structure and break down government documents into multiple structured text segments. Layout analysis utilizes a deep learning model to accurately identify different document elements, including titles, body text, tables, charts, and signatures. Compared to traditional OCR technology, this method improves recognition accuracy for complex layouts.

[0050] For identified structured text segments, feature sets, verification rule relationships, and constraints are extracted. Feature extraction uses named entity recognition technology to accurately identify key information in the text, such as names, dates, amounts, and ID numbers. Verification rule relationships and constraints are extracted from the rule base using semantic analysis technology, establishing a mapping between the text segments and verification rules.

[0051] The timestamp attribute adds validity and expiration timestamps to each extraction result, enabling timely management of verification rules. This time-sensitive processing mechanism enables real-time detection of verification rule changes and automatically updates verification constraints, avoiding the risk of using outdated rules for verification.

[0052] The semantic differential engine automatically identifies the impact of verification rule changes and records the trajectory of policy changes, providing a scientific basis for verification rule regulation. Through semantic differential analysis, it accurately assesses the impact of rule changes on different types of materials, enabling targeted rule adjustments.

[0053] Verification rule conflict detection performs conflict detection on the dynamic verification rule knowledge graph based on constraints, enabling version control and conflict detection of verification rules. This mechanism proactively identifies inconsistencies in cross-departmental verification rules, avoiding verification errors caused by conflicting rules.

[0054] In response to the identified inconsistencies, a rule priority system was established to scientifically address conflicts and dependencies between verification rules. This priority system takes into account multiple factors, including the legal effect, publication date, and scope of application of the rules, ensuring that the most appropriate rule is selected for verification when conflicts arise.

[0055] The initial verification strategy of this embodiment has improved the accuracy of rule conflict detection and the efficiency of rule adaptation, effectively improving the accuracy and efficiency of government material verification.

[0056] Example 5 Based on Examples 1-4, step 3 uses a refined verification method based on target feature points to achieve precise positioning and verification of key information of government materials.

[0057] First, based on the feature word list and location descriptions in the feature set, the target feature points are precisely located within the structured text fragment. The feature word list contains key field names and identifiers, such as name, ID number, and application date; the location description contains information such as the spatial location of the text fragment, its contextual relationship, and its position within the table. This information enables the accurate location of key information points within government documents.

[0058] After locating the target feature points, each feature point is applied to the corresponding attribute constraint rules to check whether it meets format, range, and logic requirements. Attribute constraints include textual information such as numeric ranges, date formats, and required fields. For example, for ID card numbers, the system checks whether its length is 18 digits, whether its format complies with national standards, and whether the checksum is correct. For date fields, the system checks whether its format is standardized and whether it is within the valid range.

[0059] Through correlation analysis between feature points, the spatial and semantic relationships between them are analyzed and captured, generating an initial verification strategy that includes a verification path, verification results, and confidence levels. This correlation analysis can identify logical relationships between feature points, such as the application date must be earlier than the approval date and the applicant's age must be consistent with their ID number, thereby improving the accuracy and comprehensiveness of verification.

[0060] The multidimensional representation of target feature points represents them as multidimensional vectors that encompass the spatial positional relationships within text fragments, contextual relationships, position within tables, feature words, and attribute constraints. This allows for a comprehensive description and precise location of feature points. This multidimensional representation significantly improves the accuracy of feature point location, particularly for government documents with complex layouts and ambiguous expressions.

[0061] The target feature point refined verification method of this embodiment has improved the accuracy of feature point positioning and attribute constraint verification, effectively improving the accuracy and reliability of government material verification.

[0062] Example 6 Based on Examples 1-5, step 4 implements a safe and efficient deep verification process through a deep verification process method.

[0063] This example first verifies text segments based on punctuation. This fine-grained analysis method can capture subtle semantic differences in text. It identifies the semantic functions of different punctuation marks, segments the text into multiple semantic units, and assigns a unique identifier to each unit. By analyzing the nested relationships between punctuation marks, it constructs a hierarchical structure of semantic units, laying the foundation for deep semantic understanding.

[0064] Based on the hierarchical structure of semantic units, we conduct in-depth semantic understanding to capture the logical relationships and constraints of text fragments. By identifying key feature words and their attribute relationships, we construct a feature association network and match it with the verification rule knowledge graph to ensure the compliance and consistency of semantic feature relationships.

[0065] The deep verification processing method has the following characteristics: 1. Security update policy is updated through a hierarchical process including: Level 1 update: For non-substantive changes such as template adjustments, use online learning to update in real time Secondary update: When changes in approval authority are involved, the multi-party secure computing protocol is activated 2. Gradient protection mechanism: Gradient impact factor analyzer: prevents parameters from being inferred from original data and aggregation based on homomorphic encryption, ensuring cross-departmental data isolation.

[0066] Deep verification processing allows for continuous optimization of verification models to adapt to policy changes and new materials while protecting data privacy. This improves both model update efficiency and data privacy protection capabilities.

[0067] Example 7 On the basis of Examples 1-6, the cross-page feature matching method in step 5 effectively solves the problem of missing cross-page feature matching in traditional verification methods.

[0068] First, page numbers and continuity indicators, such as "continued from previous page" and "continued from next page," are identified in the document to establish a preliminary association between pages. Next, a feature vector is extracted for each page, including header and footer information, document structure features, and content features. Feature extraction utilizes a multi-scale feature fusion algorithm to ensure that page features of varying granularity are captured.

[0069] The core of cross-page feature matching is to convert feature vectors to the frequency domain, effectively reducing the impact of noise and document deformation. In the frequency domain, a similarity matrix is ​​calculated between feature vectors from different pages to identify potentially related pages. Finally, content semantic analysis verifies the rationality of the associations, ensuring the accuracy of cross-page feature matching.

[0070] In addition, a cross-page logical consistency verification mechanism has been developed to check the logical coherence and information consistency of cross-page content. For example, it verifies the data continuity of cross-page tables and the semantic coherence of cross-page text.

[0071] The cross-page feature matching method of the present invention improves the matching accuracy and processing speed, and effectively solves the cross-page association problem in the verification of government materials.

[0072] Example 8 Based on Examples 1-7, step 6 achieves comprehensive and accurate verification of government materials through a multi-dimensional cross-validation credibility scoring method.

[0073] First, the results of the three rounds of verification were iteratively compared to identify points of consistency and discrepancy. A multi-dimensional cross-validation mechanism, using an image-text-policy triplet alignment mechanism, ensured the comprehensiveness and accuracy of the verification results.

[0074] Multi-data source verification introduces external authoritative data sources for supplemental verification, such as identity information from the public security department, tax records from the tax department, and insurance information from the social security department. Through secure data exchange protocols, necessary verification information can be obtained while protecting privacy.

[0075] Based on the verification results, a credibility score is calculated. This score considers multiple factors, including verification consistency, data source reliability, rule applicability, and historical verification accuracy. The score uses a weighted fusion algorithm to ensure appropriate weighting of each factor.

[0076] Risk assessment evaluates verification risk based on credibility scores and material characteristics. High-risk materials trigger additional review processes, such as manual review and multi-party verification, to ensure the accuracy and security of verification decisions.

[0077] In another preferred embodiment, the authenticity of the material is determined by judging whether it is forged or tampered with based on the anti-counterfeiting credibility, and a multi-level judgment mechanism is used to determine the authenticity of the material: When the anti-counterfeiting credibility is ≥0.85$, the material is judged to be highly credible and authentic, with no signs of forgery or tampering; When 0.65≤anti-counterfeiting credibility<0.85, the material is judged to be of medium credibility, with slight anomalies but overall credibility; When 0.45≤anti-counterfeiting credibility<0.65, the material is judged to be of low credibility and has obvious abnormal characteristics, requiring further manual verification; When the anti-counterfeiting credibility is less than 0.45, the material is judged to be high-risk material and is very likely to be forged or tampered with; The system implements anti-counterfeiting judgment through the following specific detection mechanisms: 1. Anti-counterfeiting mark detection: A deep learning model is used to identify anti-counterfeiting marks in government documents, including special printed patterns, security threads, and color-changing inks. The matching degree is calculated by comparing them with a standard anti-counterfeiting feature library. 2. Electronic signature verification: Check the validity of the electronic signature through the digital certificate verification chain, including signature timestamp, certificate authority authority, and signature integrity verification; 3. Watermark feature analysis: Use frequency domain analysis technology to detect hidden watermarks, extract watermark information through wavelet transform and DCT transform, and verify the integrity and consistency of the watermark; 4. Printing quality assessment: Analyze the printing quality characteristics of materials, including font clarity, color saturation, paper texture and other physical characteristics, and identify low-quality counterfeit materials; Based on the above detection results, the system generates three rounds of verification results including authenticity judgment, abnormal feature description, credibility score and processing suggestions, providing key basis for subsequent comprehensive credibility score.

[0078] The multi-dimensional cross-validation method of the present invention has improved verification accuracy and risk identification capabilities, effectively improving the reliability and security of government document verification.

[0079] Example 9 Based on Examples 1-8, step 6 is a verification decision optimization method based on risk assessment, which achieves more accurate and reliable verification decisions by comprehensively analyzing verification results and risk factors.

[0080] First, a comprehensive analysis of the three rounds of verification results was conducted to identify points of agreement and discrepancy. Agreements indicate that different verification methods reached the same conclusion, which is highly reliable. Discrepancies indicate that different verification methods reached different conclusions, requiring further analysis and verification.

[0081] Risk factor analysis considers multiple risk factors, including document type risk, applicant risk, historical verification risk, and policy change risk. Document type risk is based on the historical error rates of different document types; applicant risk is based on the applicant's history and credit rating; historical verification risk is based on historical verification results for similar documents; and policy change risk is based on the impact of recent policy changes on verification.

[0082] By obtaining the risk level assessment value, an influencing factor assessment matrix is ​​constructed, which includes influencing factors in five dimensions: material sensitivity, amount involved, scope of personnel involved, timeliness requirements, and legal consequences; Each influencing factor is quantitatively scored according to three levels: low, medium, and high, corresponding to the scoring ranges of 1-3 points, 4-6 points, and 7-9 points respectively; Establish a business impact assessment system that includes four aspects: business continuity impact, public service impact, government image impact, and economic loss impact; Construct a risk level combination matrix, cross-analyze the impact factor assessment results and the business impact degree, and form a 9×4 risk assessment matrix; The comprehensive score of each dimension is calculated using the weighted average method, with a weight of 0.25 for material sensitivity, 0.20 for the amount involved, 0.15 for the scope of personnel involved, 0.20 for timeliness requirements, and 0.20 for legal consequences. Based on the comprehensive scoring results, the risk level is divided into four levels: low risk (score ≤ 3.0), medium risk (3.0 < score ≤ 6.0), high risk (6.0 < score ≤ 8.0) and very high risk (score > 8.0); For materials with extremely high risk levels, the system automatically triggers an expert review mechanism, requiring multiple professionals to conduct joint verification; The credibility score in Example 8 is calculated using a multi-factor weighted model that considers multiple dimensions, including verification consistency, data source reliability, rule applicability, and risk factors. The score uses a numerical value between 1 and 10 to indicate the credibility of the verification result.

[0083] Based on the credibility score, adaptive verification decisions are implemented. High-credibility verification results are directly adopted; medium-credibility results require additional verification measures; and low-credibility results trigger a manual review process. This adaptive decision-making mechanism significantly improves verification accuracy and efficiency.

[0084] The risk warning mechanism builds a risk warning model based on historical data and current verification results, enabling early identification of potential high-risk materials and anomalies. The warning model utilizes anomaly detection algorithms to identify verification results that are inconsistent with historical patterns, prompting relevant personnel to pay close attention.

[0085] The risk assessment and verification decision optimization method of the present invention improves both the decision-making accuracy and the risk warning accuracy, effectively improving the reliability and security of government material verification.

[0086] Through the detailed explanation of the above embodiments, the present invention effectively solves the technical problems existing in the prior art, such as insufficient multimodal feature extraction, static solidification of verification rules, insufficient deep verification processing, lack of cross-page feature matching, lack of multi-dimensional cross-validation, and imperfect risk assessment mechanism, and realizes efficient, accurate, dynamic and comprehensive intelligent verification of government materials.

[0087] like Figure 2 As shown, a government document intelligent verification system based on machine learning includes: Data collection module: used to obtain text information data and verification rules of historical materials; Feature extraction module: It is used to extract multimodal features from text information data; Initial verification module: This module is used to establish a dynamic verification rule knowledge graph, verify the multimodal features of the structured text fragments and target feature points in the current material to be verified, and generate an initial verification strategy; Secondary Verification Module: It is used to conduct in-depth verification of the material text fragments after the initial verification, obtain the verification results of the associated feature words and attribute descriptions, optimize the initial verification strategy, and generate a secondary verification strategy; Three-time verification module: It is used to conduct a third verification on the material text fragments after the second verification by triggering cross-page feature matching, so as to achieve cross-page text fragment association and multimodal feature fitting, so as to generate a three-time verification strategy; Credibility Assessment Module: This module is used to iteratively compare the results of the three rounds of verification, build multi-dimensional cross-validation, and combine verification conditions from multiple data sources to generate a credibility score. Final verification module: It is used to adjust the comparison results based on the credibility score and the verification risk anomaly assessment to obtain the final verification decision strategy.

[0088] Application Example 1 Based on Examples 1-9, the present invention provides a machine learning-based intelligent verification system for government documents, which is applied to the use scenario of medical insurance reimbursement review. This application example demonstrates the practical application effect of the present invention in a specific field.

[0089] First, the system receives medical insurance reimbursement application materials, including various documents such as medical expense invoices, diagnosis certificates, prescriptions, and medical record summaries. Through multimodal feature extraction, it simultaneously processes both image information (such as seals and watermarks) and text information (such as diagnosis results, drug names, and bill amounts) within these documents.

[0090] During the initial verification phase, the system uses a dynamic verification rule knowledge graph to perform structured analysis of medical documents. It accurately identifies key information on invoices, such as the hospital name, patient information, charges, and amounts, and correlates and verifies this information with the disease information in the diagnosis certificate. Through layout analysis and feature point positioning, it accurately identifies key fields in the document, such as diagnosis results and medication lists, and extracts the corresponding content for verification.

[0091] During the in-depth verification phase, the initial verification results are thoroughly analyzed. For example, the system checks whether the medications listed on the prescription match the diagnosis, whether the dosage is within a reasonable range, and whether the costs meet medical insurance reimbursement standards. Through segmented verification and semantic understanding, potential anomalies can be identified, such as multiple reimbursement requests for the same medication by the same patient in a short period of time, or significant discrepancies between the diagnosis and the prescribed medication.

[0092] During the cross-page feature matching phase, the system can effectively process multi-page medical documents. For example, for multi-page medical record summaries, Fourier transform frequency domain analysis can be used to ensure content coherence and consistency across different pages. Timeline information within medical record summaries can also be identified to ensure the logical rationality of consultation times, medication times, and reimbursement application times.

[0093] Through multi-dimensional cross-validation, the verification results are compared with the medical insurance policy database, drug catalog database, and historical reimbursement records to calculate a credibility score. Applications with low scores are automatically marked as requiring manual review and specific abnormalities are indicated.

[0094] Experimental data show that the application of this system in medical insurance reimbursement review has increased review efficiency by 60%-70% and reduced the error rate by 75%-80%, greatly reducing the workload of medical insurance agencies and improving the efficiency of medical insurance fund use.

[0095] Application Example 2 Based on Examples 1-9, the present invention provides a machine learning-based intelligent verification system for government documents, which is applied to the review of business registration documents. This application example demonstrates the advantages of the present invention in processing complex business documents.

[0096] First, the system receives business registration application materials, including the company name pre-approval notice, company charter, legal representative ID, business premises certificate, and other documents. Using an improved ResNet-101 network to extract multi-level features, it effectively processes seals, signatures, forms, and text content within these documents.

[0097] During the initial verification phase, a dynamic verification rule knowledge graph is used to perform a structured analysis of industrial and commercial registration materials. Key fields such as company name, registered capital, business scope, and shareholder information are accurately identified and verified against industrial and commercial registration rules. For complex fields like business scope, semantic analysis is used to determine whether they comply with national industrial policies and industry access requirements.

[0098] During the in-depth verification phase, the initial verification results are thoroughly analyzed. For example, the company's equity structure in the articles of association is consistent with the shareholders' capital contribution ratio, the registered capital meets the industry's minimum requirements, and any restrictions on the legal representative's position are examined. Through segmented verification and semantic understanding, potential issues can be identified, such as the inclusion of restricted or prohibited items in the business scope or discrepancies between shareholder information and identification documents.

[0099] During the cross-page feature matching phase, we can effectively process multi-page corporate charter documents. Through Fourier transform frequency domain analysis, we ensure the coherence and consistency of content across different pages, especially for key information such as article numbers, page numbers, and signatures and seals in the charter.

[0100] Through multi-dimensional cross-validation, the verification results are compared with the enterprise credit information database, the legal person identity information database, and the industry access policy database to calculate a credibility score. Applications with low scores are automatically marked as requiring manual review and specific abnormalities are indicated.

[0101] The risk assessment constructs a business registration risk model based on historical data, capable of identifying high-risk registration behaviors such as potential zombie companies and shell companies. For example, it focuses on abnormal situations such as multiple companies registered at the same address in a short period of time, multiple companies with the same legal representative, and obvious mismatches between registered capital and business scope.

[0102] Experimental data shows that the application of this system in industrial and commercial registration review has increased review efficiency by 65%-75% and reduced the error rate by 80%-85%, effectively improving the work efficiency and supervision capabilities of market supervision departments.

[0103] Application Example 3 Based on Examples 1-9, the present invention also provides an application example of a machine learning-based intelligent verification system for government documents in the review of real estate registration documents. This application example demonstrates the security and accuracy of the present invention in processing high-value transaction documents.

[0104] First, the system receives real estate registration application materials, including real estate ownership certificates, sales contracts, identification documents, tax payment receipts, and other documents. Through multimodal feature extraction, it simultaneously processes both image information (such as seals, signatures, and property floor plans) and text information (such as owner information, property details, and transaction amounts) within these documents.

[0105] During the initial verification phase, a dynamic verification rule knowledge graph is used to perform structured analysis of real estate registration documents. This system accurately identifies key information on the property certificate, such as the owner, location, area, and purpose, and correlates it with relevant information in the sales contract. Through layout analysis and feature point positioning, it accurately identifies key fields in the document, such as the seller, buyer, and transaction price, and extracts the corresponding content for verification.

[0106] During the in-depth verification phase, the initial verification results are thoroughly analyzed. For example, the team will check whether the property information in the sales contract is consistent with the title deed, whether the transaction price is within a reasonable range, and whether tax payments comply with regulations. Through segmented verification and semantic understanding, potential risks can be identified, such as ambiguous terms in the contract, restrictions on rights, and special agreements.

[0107] During the cross-page feature matching phase, it can effectively process multi-page sales contracts. Through Fourier transform frequency domain analysis, it ensures the coherence and consistency of content across different pages, especially for key information such as clause numbers, page numbers, and signatures and seals.

[0108] Multi-dimensional cross-validation compares verification results with real estate registration databases, population identity databases, and tax databases to calculate a credibility score. For example, this process verifies the authenticity of the property owner's identity, whether the property is mortgaged or seized, and whether the two parties involved in the transaction are related. Applications with low scores are automatically flagged for manual review and are accompanied by specific anomalies.

[0109] Risk assessment mechanisms are particularly important in real estate registration. A real estate transaction risk model, built based on historical data, can identify potential fraud. For example, we focus on unusual situations such as properties frequently traded in a short period of time, transaction prices significantly deviating from market value, and the purchase of multiple properties by the same buyer within a short period of time.

[0110] Experimental data shows that the application of this system in real estate registration review has increased review efficiency by 55%-65% and reduced the error rate by 80%-90%, effectively improving the work efficiency and risk prevention and control capabilities of the real estate registration department and providing strong protection for the property safety of citizens. In summary, the results produced by the machine learning-based intelligent verification system for government materials provided by the present invention after use in different application scenarios are shown in Table 1.

[0111] Table 1: Comparison of performance indicators in three different application scenarios.

[0112] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make more forms of equivalent embodiments based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. A method for intelligent verification of government documents based on machine learning, characterized in that: The following steps are involved: Obtain text information data and verification rules for historical materials; Perform multimodal feature extraction on text information data; Establish a dynamic verification rule knowledge graph to verify the multimodal features of the structured text fragments and target feature points in the current material to be verified to generate an initial verification strategy; Conduct in-depth verification of the material text fragments after the initial verification, obtain the verification results of related feature words and attribute descriptions, optimize the initial verification strategy, and generate a secondary verification strategy; By triggering cross-page feature matching, the text fragments of the material that have been verified twice are verified for a third time, so as to achieve cross-page text fragment association and multimodal feature fitting, and thus generate a three-time verification strategy; The results of the three rounds of verification are iteratively compared, and a multi-dimensional cross-validation is constructed, combining the verification conditions of multiple data sources to generate a credibility score; the comparison results are adjusted based on the credibility score and the verification risk anomaly assessment to obtain the final verification decision strategy.

2. The method for intelligent verification of government documents based on machine learning according to claim 1 is characterized in that: Perform multimodal feature extraction on text information data, including: Preprocessing the material, including noise removal and layout analysis, to obtain a preprocessed data set; An improved ResNet-101 network is used to extract multi-level features from the preprocessed dataset, and a multimodal spatiotemporal alignment network is used to perform three-level feature fusion on the multi-level features. The multi-level features include image pixel level, text semantic level, and spatiotemporal rule level. A three-level feature fusion mechanism is used to identify and associate image and text position information using a spatial attention mechanism to obtain multimodal features. The text semantic level fusion establishes semantic association between simple semantic data and complex semantic data through graph convolutional network. The multimodal features include a comprehensive representation of image features, text features and time features; The image features are complemented by simulating the complex situations of seal overlap, watermark interference and blurred text that are unique to government documents.

3. The method for intelligent verification of government documents based on machine learning according to claim 2 is characterized in that: Standardize and normalize the multimodal features to obtain a consistent and complete multimodal feature representation to obtain a feature set; Extract text feature vectors from the feature set through the BERT model; Dynamically convert verification rules into verification rule vectors; Construct a material type recognition model based on text feature vectors and verification rule vectors; Automatically identify and judge the material category through the material type recognition model; A dynamic verification rule knowledge graph is established based on various verification rules obtained historically, and corresponding government documents are verified according to different verification rules; The verification rules include logical coherence between multimodal features, context segment association information, and keyword and word attribute constraints; And obtain the corresponding material verification strategy based on the material category and the corresponding verification rules.

4. The method for intelligent verification of government documents based on machine learning according to claim 3 is characterized in that: The initial verification strategy includes: Identify the document structure through layout analysis of the materials to be verified to obtain structured text fragments; Extract the feature set of structured text fragments, verify the rule relationships and constraints, and obtain the extraction results; Add validity and expiration timestamp attributes to the extraction results to detect changes in verification rules and update verification constraints in real time; Develop a semantic differential engine to identify the impact of changes to verification rules. Obtain and record the trajectory of policy changes to establish a verification rule regulation mechanism; in, is the impact range index, is the node influence weight of the k-th rule, is the change range of the kth rule, is the node connectivity, is the distance attenuation parameter, is the semantic distance between nodes; Based on the constraints, the dynamic verification rule knowledge graph is used to detect verification rule conflicts, so as to achieve version control and verification conflict detection of verification rules, and proactively discover contradictions in cross-departmental verification rules; In order to address the contradictions, a rule priority system is established to handle the conflicts and dependencies between verification rules.

5. The method for intelligent verification of government documents based on machine learning according to claim 4 is characterized in that: The initial verification strategy also includes: Based on the feature vocabulary and position description in the feature set, the target feature points are located in the structured text fragment; Apply each feature point to the corresponding attribute constraint rules to check whether it meets the format, range and logic requirements; Analyze the spatial position relationship and semantic association between feature points, and generate an initial verification strategy including verification path, verification result and confidence level; the verification confidence level is determined by Get; in, To verify the confidence, include any natural number between 0 and 1. is the matching probability of the i-th feature point, is the corresponding weight, To verify the error, is the standard deviation of error, represents an exponential function used to calculate the confidence decay factor based on the validation error; The target feature points include the spatial position relationship of the text segment, the context relationship and the position in the table, the feature words and the attribute constraints; Among them, feature words include key field names and identifiers; attribute constraints include text description information of numerical ranges, date formats, and required items.

6. The method for intelligent verification of government documents based on machine learning according to claim 1 is characterized in that: Perform in-depth verification on the text fragments of the materials after the initial verification, obtain the verification results of the associated feature words and attribute descriptions, optimize the initial verification strategy, and generate a secondary verification strategy, including: Performing segmented verification on the text segments based on punctuation marks, wherein the segmented verification includes identifying different punctuation marks in the text; Segment the text segment into multiple semantic units according to the semantic functions of different punctuation marks; Assign a unique identifier to each semantic unit and establish a tag signal index; Analyze the nested relationship between punctuation marks based on the tag signal index and construct the hierarchical structure of semantic units; Conduct semantic understanding of the initially verified material text based on the hierarchical structure of semantic units to obtain the logical relationships and constraints of the text fragments; By identifying and extracting key feature words and their attribute relationships in the text, a feature association network is constructed; Match and verify the feature association network with the verification rule knowledge graph to obtain the semantic feature relationship of compliance and consistency; Apply each semantic unit to the corresponding verification rules according to the semantic feature relationship to generate fine-grained verification; Based on fine-grained verification combined with the semantic relationships implied by punctuation, the system verifies the logical consistency between related semantic units. It also records the verification time, applicable verification rule version, and meta-information between text segment features to generate a traceable verification path. The positioning and verification process of target feature points is obtained through fine-grained verification, forming a traceable verification evidence chain and obtaining in-depth verification results; Assign a confidence score between 0 and 1 to each deep verification result; Based on the verification results with a confidence score ≤ 0.6, the verification evidence chain that does not meet the verification rules will be classified as abnormal. Abnormal markings include missing, format errors, and logical conflicts. Integrate verification paths, confidence scores, and anomaly flags to form a secondary verification strategy.

7. The method for intelligent verification of government documents based on machine learning according to claim 1 is characterized in that: By triggering cross-page feature matching, the text fragments of the material that have been verified twice are verified for a third time to achieve cross-page text fragment association and multimodal feature fitting, thereby generating a three-level verification strategy, including: Extract associated feature points from multiple pages of materials, including header and footer information, chapter numbers, and table continuation marks, to obtain an associated feature point dataset; Convert the extracted feature points into vector representation and construct a feature vector array; Based on the feature vector array, the similarity between feature vectors of different pages is obtained to obtain the cross-page similarity threshold; Determine whether there is a relationship between pages based on the cross-page similarity threshold and determine the type of relationship; Logically combine the identified cross-page related text fragment association types to reconstruct the text information; Based on the reconstructed text information, the image features, text features, and time features of the cross-page text in the second-verified material text fragment are comprehensively fitted to obtain a comprehensive fitting result to eliminate the defects of text page information fragmentation; Verify the semantic and logical consistency of the reconstructed text information based on the comprehensive fitting results and obtain the cross-page matching results; Based on the cross-page matching results, the first two rounds of verification strategies are integrated to form a three-round verification strategy with cross-page association.

8. The method for intelligent verification of government documents based on machine learning according to claim 1 is characterized in that: The results of the three rounds of verification are iteratively compared, and a multi-dimensional cross-validation system is constructed, combining verification conditions from multiple data sources to generate a credibility score, including: Compare the text information after three verification strategies and check whether there are any contradictions or inconsistencies in the material text information to obtain the inspection results, which include numerical consistency, logical consistency and expression consistency; When the inspection results are abnormal, the material text information after three rounds of verification is iteratively calibrated to obtain the calibration results. Based on the calibration results, verify again whether the time information in the material complies with the time logic, including date sequence, timeliness requirements and historical time point verification, and obtain a round of verification results; Based on the results of the first round of verification, the authenticity and authority of the source of the materials are verified by querying authoritative databases, blockchain evidence storage, and government information sharing platforms to obtain the results of the second round of verification; Based on the results of the second round of verification, the anti-counterfeiting marks, electronic signatures, watermarks and special printing features in the materials are identified and verified, and: To obtain the credibility of the anti-counterfeiting feature; judge whether the material is forged or tampered according to the anti-counterfeiting credibility, and obtain the results of three rounds of verification; in, For anti-counterfeiting credibility, For the The credibility of the anti-counterfeiting mark, is the corresponding weight of the credibility of the anti-counterfeiting mark, For the The credibility of an electronic signature, is the corresponding weight of the electronic signature credibility, For the The credibility of the watermark feature, is the corresponding weight of the watermark feature credibility; Integrate the results of the first three rounds of verification to form a multi-dimensional cross-validation, and combine the multi-dimensional cross-validation with the verification conditions of three multi-data sources: internal government data, third-party trusted data, and historical verification records, to generate a comprehensive credibility score for the material; The comprehensive credibility score organizes the verification results of each dimension into a complete chain of evidence according to chronological order and logical relationships, supporting the interpretability of the verification results; and obtains the comprehensive credibility score of the material based on the integrity of the chain of evidence, the verification results of each dimension and the authority of the verification data source.

9. The method for intelligent verification of government documents based on machine learning according to claim 8 is characterized in that: Based on the credibility score requirements and risk level adaptive adjustments, the final verification decision strategy is comprehensively generated, including: Based on the material type, application scenario, and business impact, the comprehensive credibility score of the current verification material text is graded into three levels: critical, important, and general; Comprehensively consider the various influencing factors involved in the material and the combined matrix of business impact and risk level, through: Assess the potential level of risk; in, is the risk level assessment value, is the weight of the fth influencing factor, For the degree of impact, is the probability of occurrence, is the weight of the rth severity factor, For severity, is the frequency, is the uncertainty coefficient, is the uncertainty factor; Automatically adjust the credibility score threshold of the verification results based on the risk level assessment of the currently verified material text, setting higher credibility score requirements for important and high-risk businesses; Set risk-based manual review trigger conditions. When the credibility of the verification result is lower than the threshold but the business importance is high, it will be automatically marked as requiring manual review. Based on the adjusted credibility score requirements, decide whether to enable a more in-depth verification process that requires manual review; Integrate the three rounds of verification results, credibility scores, and adjusted threshold requirements to generate a final verification decision strategy that includes verification conclusions, confidence levels, anomaly flags, and handling recommendations; Based on the verification conclusions and verification rules obtained through four rounds of verification strategies, we provide execution recommendations of passing, passing with supplementary materials, or rejecting the application, and give specific reasons and re-verification operation instructions.

10. A system for intelligent verification of government documents based on machine learning, used to execute the method for intelligent verification of government documents based on machine learning according to any one of claims 1 to 9, characterized in that: include: Data collection module: used to obtain text information data and verification rules of historical materials; Feature extraction module: It is used to extract multimodal features from text information data; Initial verification module: This module is used to establish a dynamic verification rule knowledge graph, verify the multimodal features of the structured text fragments and target feature points in the current material to be verified, and generate an initial verification strategy; Secondary Verification Module: It is used to conduct in-depth verification of the material text fragments after the initial verification, obtain the verification results of the associated feature words and attribute descriptions, optimize the initial verification strategy, and generate a secondary verification strategy; Three-time verification module: It is used to conduct a third verification on the material text fragments after the second verification by triggering cross-page feature matching, so as to achieve cross-page text fragment association and multimodal feature fitting, so as to generate a three-time verification strategy; Credibility Assessment Module: This module is used to iteratively compare the results of the three rounds of verification, build multi-dimensional cross-validation, and combine verification conditions from multiple data sources to generate a credibility score. Final verification module: It is used to adjust the comparison results based on the credibility score and the verification risk anomaly assessment to obtain the final verification decision strategy.

Images