A Method for Classifying and Controlling the Exchange Access of Classified Documents Based on Natural Language Processing

Abstract

This invention discloses a method for classifying and managing the exchange of classified documents based on natural language processing, comprising the following steps: performing structured parsing of the input document, segmenting it into multiple original fragments, and extracting semantic features of the text; using a natural language processing model composed of a sparsely gated expert hybrid network and a neural causal reasoning network to classify the fragments by security level, outputting security level labels, sensitive entity labels, and permission vectors; inputting the above results into a reversible generative adversarial network to generate a content version that meets the permission and security level requirements, and using a discriminator to jointly judge the effectiveness of content desensitization, the accuracy of security level classification, and the consistency of permission allocation, thereby optimizing content generation. This method supports automatic recording and traceability of the entire process of content flow and permission adjustment, realizing intelligent classification, dynamic permission control, and secure traceable flow of classified documents. This invention achieves intelligent classification, dynamic permission control, and secure traceable flow of classified documents throughout the entire process.

Description

Technical Field

[0001] This invention relates to the fields of information security and natural language processing technology, and in particular to a method for classifying and controlling the exchange permissions of classified documents based on natural language processing. Background Technology

[0002] With the accelerated digitization of data in government, enterprise, finance, scientific research, and infrastructure sectors, the need for security management of classified documents in scenarios such as internal circulation, external exchange, cloud storage, and multi-platform collaboration is becoming increasingly prominent. Traditional methods for managing classified documents typically rely on static rules, manual classification, coarse-grained permission allocation, and manual approval processes. These methods suffer from the following main problems:

[0003] Most systems only support overall document security classification, failing to perform fine-grained security classification of document fragments, making it difficult to prevent partial content leakage. Security classification and sensitive content identification mainly rely on human experience, resulting in low efficiency, high subjectivity, and difficulty in adapting to large-scale dynamic data flow. Existing access control is mostly based on pre-set templates and fixed rules, making it difficult to dynamically adjust according to user identity, exchange context, and content sensitivity, leading to permissions that are either too lenient or too strict, easily causing security risks or hindering business operations. Content desensitization and multi-view generation for different permissions or external objects still require manual operation, with limited automation and intelligence levels. There is a lack of end-to-end traceability auditing and incident liability determination mechanisms for content versioning, access control adjustments, and data transfer operations.

[0004] In recent years, with the development of AI technologies such as deep learning, natural language processing, and graph neural networks, some studies have attempted to introduce NLP models, entity recognition algorithms, and intelligent rule engines to achieve automatic document classification and sensitive entity detection. However, existing methods still generally suffer from problems such as difficulty in accurately identifying implicit sensitive content in complex confidential scenarios, inability to achieve dynamic permission adjustment based on exchange contexts and automatic generation of multiple versions of content, and lack of end-to-end generative adversarial models and causal reasoning mechanisms.

[0005] Therefore, how to provide a method for classifying and controlling the exchange of classified documents based on natural language processing is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a method for classifying and managing the exchange of classified documents based on natural language processing. This invention can achieve intelligent discrimination of classified document fragment-level security classification and identification of sensitive entities based on multi-expert subnetworks and neural causal reasoning. It also dynamically generates content versions that meet different permission requirements by combining permission decision subnetworks and reversible flow neural networks, achieving triple consistency and automated control of content, permissions, and security classification. At the same time, it supports security auditing and traceability of content generation, permission allocation, and content flow throughout the entire process, greatly improving the intelligence, security, and compliance of classified document management.

[0007] The method for classifying and controlling the exchange of classified documents based on natural language processing according to embodiments of the present invention includes the following steps:

[0008] The input document is parsed in a structured manner, and it is divided into multiple original segments according to paragraphs, sentences, tables and images. The textual semantic features of each original segment are extracted as feature vectors.

[0009] Input the original fragment and feature vector, and use a natural language processing model composed of a sparse gating expert hybrid network and a neural causal reasoning network to perform security level discrimination, and output security level label, sensitive entity label and permission vector;

[0010] The original fragment, security level label, sensitive entity label, and permission vector are input together into a reversible generative adversarial network, which includes a generator and a discriminator.

[0011] The generator is a reversible neural network structure that generates a content version that meets the permission constraints and security requirements based on the original fragment, security level label, sensitive entity label and permission vector.

[0012] The discriminator directly receives the content version, security level label, permission vector, sensitive entity label and original fragment output by the generator as input, and performs joint discrimination on the effectiveness of content desensitization, the accuracy of security level classification and the consistency of permission allocation, and outputs the discrimination results to guide the optimization of the generator.

[0013] During document exchange, circulation, and permission adjustment operations, all content generation, security classification, permission allocation, and reverse mapping processes are automatically recorded.

[0014] Furthermore, the extraction of textual semantic features from the original fragment includes the following steps:

[0015] Each original segment is segmented and labeled with part-of-speech tags to extract word vector sequences;

[0016] The pre-trained Word2Vec model is used to encode fragment text into semantic embedding vectors;

[0017] Analyze entities, proper nouns, and technical terms in the fragment, and generate entity embedding representations;

[0018] Dependency parsing is performed on the fragments to obtain the grammatical structural features of subject-verb-object and modification relations;

[0019] The semantic embedding vector, entity embedding representation, and grammatical structure features of the original fragment are combined with structural information representing paragraph number, document level, and the position of the table or image to which it belongs to form a complete feature vector.

[0020] Furthermore, the natural language processing model composed of a sparsely gated expert hybrid network and a neural causal reasoning network includes:

[0021] The entity recognition expert subnetwork encodes fragmented text into a context-dependent word vector sequence using the BERT model. The word vector sequence is input into a bidirectional LSTM for temporal modeling to capture the dependencies between words. The LSTM output is processed by a CRF layer to predict the entity category and location label of each word. All entities are embedded into entity embedding vectors through embedding mapping.

[0022] The context reasoning expert subnetwork concatenates the current segment and the preceding and following segments into an input sequence via word embedding. Multiple Transformers perform self-attention calculations on the concatenated sequence to model the higher-order semantic dependencies between the segment and its neighboring context. The mean pooling vector of the last Transformer layer serves as the context feature vector of the segment.

[0023] The structural relationship expert subnetwork adopts a graph neural network structure to construct a document structure graph from document fragments, table and image nodes. Node attributes are fragment features, and edges represent structural, reference and order relationships. The structure graph is subjected to several layers of GNN message passing, the features of neighboring nodes are aggregated, structural context information is modeled, and the structural feature vector of the current fragment node is obtained.

[0024] The external knowledge matching expert subnetwork performs matching calculations between the entities identified in the fragment and the entities in the knowledge base through vector embedding. If the match is successful, it extracts the associated security level, attributes, or hierarchical relationships to generate a knowledge matching feature vector. For entities without precise matching, it obtains relevant knowledge completion information through fuzzy matching and outputs knowledge matching features that describe the degree of association between the fragment and external knowledge and security rules.

[0025] The sparse gated routing mechanism assigns weights to each expert based on the features of the original fragment, dynamically routes the input, and outputs weighted fused high-order semantic features.

[0026] The neural causal reasoning layer receives weighted and fused high-order semantic features, performs causal inference based on the causal graph structure, and outputs the inferred security level label and sensitive entity combination label.

[0027] The permission decision layer encodes the security level label, sensitive entity label, context feature vector, and structural feature vector output by the neural causal reasoning layer to form a joint feature vector. The security level label and sensitive entity label are encoded using one-hot encoding, and the context feature and structural feature are normalized. The permission decision sub-network with a multilayer perceptron structure takes the joint feature vector as input, processes it through decision rules, performs feature fusion and condition judgment, and outputs a permission vector. The permission vector includes readable, writable, forwardable, shareable, downloadable, printable, and archiveable permission fields. Each field is 0 or 1, representing the accessibility of the fragment under the corresponding permission.

[0028] The classification output layer outputs security level labels, sensitive entity labels, and permission vectors. The security level labels include top secret, confidential, internal, and public.

[0029] Furthermore, the inference process of the neural causal reasoning layer includes:

[0030] It receives high-order semantic features obtained by gated routing and weighted fusion of multiple expert subnetworks as input;

[0031] Using the DAG-GNN algorithm, causal graphs are automatically constructed based on high-order semantic features:

[0032] The nodes of the causal graph include fragment text features, identified sensitive entities, context features, structural features, and external knowledge features;

[0033] Edges represent causal dependencies between nodes, and weights are learned through minimum description length, mutual information, or attention mechanisms.

[0034] Perform differentiable reasoning on the constructed causal graph:

[0035] The graph neural network employs a differentiable inference module. Each node's initial features are tensor representations of text features, identified sensitive entities, context features, structural features, and external knowledge features. The graph neural network performs multi-step message passing and feature aggregation on the causal graph. After each round of propagation, the node features are updated to a weighted sum of its own features and neighbor features. The weights are determined by the edge weights and trainable parameters. After multiple propagations, each node outputs a causal activation value.

[0036] In the last layer, the output node is defined as a security level label node and a sensitive entity combination label node. The final activation value of the node is used as the "causal contribution". The contribution of each input node to the security level label and sensitive entity combination label is quantified by L1 normalization.

[0037] The node with the largest contribution is selected as the main basis for judging the security level label and the sensitive entity combination label. The contribution of all nodes and the causal path are recorded as the inference link, and the final security level label and sensitive entity combination label are output.

[0038] Furthermore, the sparse gating routing mechanism includes:

[0039] The input consists of the concatenated text semantic feature vector, context feature vector, structural feature vector, and knowledge matching feature vector of the original fragment. The output is a weight vector with a length equal to the number of expert subnetworks, produced by a single-layer fully connected neural network.

[0040] Apply the Softmax activation function to the weight vector to obtain normalized weights;

[0041] Only the two expert branches with the highest weights are activated, and the weights of other branches are reset to zero;

[0042] The input data is dynamically routed to the activated expert sub-networks, and the features output by each sub-network are weighted and summed according to the Softmax weights to obtain the fused high-order semantic features.

[0043] Furthermore, the generator of the reversible generative adversarial network includes:

[0044] The generator adopts a RealNVP reversible flow neural network structure, which consists of multiple reversible coupling layers connected in series. Each layer performs bidirectional transformation on the input features. The input is a joint input feature, which is encoded into a high-dimensional vector or tensor by the original fragment, security level label, sensitive entity label, and permission vector respectively, and then concatenated in a set order to generate the generator.

[0045] The combined input features undergo multiple transformations to output content versions corresponding to the permission constraints and security level requirements, with each content version corresponding one-to-one with the input features.

[0046] Furthermore, the discriminator of the reversible generative adversarial network includes:

[0047] The discriminator is a multi-task neural network structure, and its input consists of the content version, security level label, permission vector, sensitive entity label, and original fragment output by the generator.

[0048] The discriminator contains three branches, namely:

[0049] The content desensitization validity branch determines whether sensitive information in the generated content has been effectively desensitized or blocked in accordance with permission requirements.

[0050] The security classification accuracy branch determines whether the security classification label of the generated content is consistent with the input security classification label;

[0051] The permission allocation consistency branch determines whether the generated content logically matches each permission field in the permission vector;

[0052] The three branches output a joint discrimination result.

[0053] Furthermore, the reversible generative adversarial network supports the reconstruction of the original fragment, original security level, sensitive entities, and initial permissions for any content version through reverse mapping.

[0054] Furthermore, when document content exchange, circulation, or permission adjustment operations are triggered, the system calls the permission decision sub-network to regenerate the permission vector based on the target recipient's user identity, organization affiliation, historical access behavior, current access environment, operation type, and content sensitivity. The permission adjustment operations include automatically tightening or relaxing specific operation permission fields, temporarily upgrading security level labels, and adding or removing sensitive entity protection strategies.

[0055] The beneficial effects of this invention are:

[0056] This invention achieves intelligent classification and sensitive entity identification of classified document fragments through the collaborative design of multi-expert subnetworks, neural causal reasoning, and permission decision subnetworks. This significantly improves the accuracy and fineness of automated classification and reduces human intervention.

[0057] This invention dynamically generates and adjusts permission vectors based on multi-dimensional features such as content sensitivity, exchange scenarios, and user identity. Combined with a reversible flow neural network generator, it automatically outputs content versions that meet different permission requirements, thereby effectively preventing the unauthorized spread of sensitive information and achieving flexible and secure content flow and access control.

[0058] It supports automated auditing and traceability of the entire process of content generation, permission allocation and content circulation, ensuring consistency and accountability of content, permissions and security classification decisions, and significantly improving the intelligence, dynamism and compliance capabilities of classified document security management. Attached Figure Description

[0059] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0060] Figure 1 This is a flowchart of the method for classifying and controlling the exchange permissions of classified documents based on natural language processing proposed in this invention;

[0061] Figure 2 This is a schematic diagram of the natural language processing model structure of the method for classifying and exchanging access control of classified documents based on natural language processing proposed in this invention.

[0062] Figure 3This is a flowchart of the reversible generative adversarial network for the method of classifying and exchanging access control of classified documents based on natural language processing proposed in this invention. Detailed Implementation

[0063] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0064] refer to Figure 1-3 A method for classifying and controlling the exchange of classified documents based on natural language processing includes the following steps:

[0065] The input document is parsed in a structured manner, and it is divided into multiple original segments according to paragraphs, sentences, tables and images. The textual semantic features of each original segment are extracted as feature vectors.

[0066] Input the original fragment and feature vector, and use a natural language processing model composed of a sparse gating expert hybrid network and a neural causal reasoning network to perform security level discrimination, and output security level label, sensitive entity label and permission vector;

[0067] The original fragment, security level label, sensitive entity label, and permission vector are input together into a reversible generative adversarial network, which includes a generator and a discriminator.

[0068] The generator is a reversible neural network structure that generates a content version that meets the permission constraints and security requirements based on the original fragment, security level label, sensitive entity label and permission vector.

[0069] The discriminator directly receives the content version, security level label, permission vector, sensitive entity label and original fragment output by the generator as input, and performs joint discrimination on the effectiveness of content desensitization, the accuracy of security level classification and the consistency of permission allocation, and outputs the discrimination results to guide the generator optimization.

[0070] During document exchange, circulation, and permission adjustment operations, all content generation, security classification, permission allocation, and reverse mapping processes are automatically recorded.

[0071] This invention organically combines structured document parsing, deep feature extraction, multi-expert hybrid neural networks, neural causal reasoning, dynamic permission generation, and reversible content generation models to achieve intelligent fragment-level classification of classified documents, automatic identification of sensitive entities, and multi-dimensional dynamic permission control. The system not only automatically generates different versions of secure content based on content, scenario, and permission requirements to prevent unauthorized dissemination of sensitive information, but also ensures consistency in content desensitization, classification, and permission allocation through a discriminator, greatly improving the automation and intelligence of content security flow and access control. The fully automated log recording and reverse traceability mechanism ensures the compliance, transparency, and accountability of classified document management. Overall, this invention significantly improves the security, flexibility, and auditability of classified documents in actual circulation, exchange, and permission adjustments.

[0072] In this embodiment, extracting the textual semantic features of the original fragment includes the following steps:

[0073] Each original segment is segmented and labeled with part-of-speech tags to extract word vector sequences;

[0074] The pre-trained Word2Vec model is used to encode fragment text into semantic embedding vectors;

[0075] Analyze entities, proper nouns, and technical terms in the fragment, and generate entity embedding representations;

[0076] Dependency parsing is performed on the fragments to obtain the grammatical structural features of subject-verb-object and modification relations;

[0077] The semantic embedding vector, entity embedding representation, and grammatical structure features of the original fragment are combined with structural information representing paragraph number, document level, and the position of the table or image to which it belongs to form a complete feature vector.

[0078] In this embodiment, the natural language processing model composed of a sparsely gated expert hybrid network and a neural causal reasoning network includes:

[0079] The entity recognition expert subnetwork encodes fragmented text into a context-dependent word vector sequence using the BERT model. The word vector sequence is input into a bidirectional LSTM for temporal modeling to capture the dependencies between words. The LSTM output is processed by a CRF layer to predict the entity category and location label of each word. All entities are embedded into entity embedding vectors through embedding mapping.

[0080] The context reasoning expert subnetwork concatenates the current segment and the preceding and following segments into an input sequence via word embedding. Multiple Transformers perform self-attention calculations on the concatenated sequence to model the higher-order semantic dependencies between the segment and its neighboring context. The mean pooling vector of the last Transformer layer serves as the context feature vector of the segment.

[0081] The structural relationship expert subnetwork adopts a graph neural network structure to construct a document structure graph from document fragments, table and image nodes. Node attributes are fragment features, and edges represent structural, reference and order relationships. The structure graph is subjected to several layers of GNN message passing, the features of neighboring nodes are aggregated, structural context information is modeled, and the structural feature vector of the current fragment node is obtained.

[0082] The external knowledge matching expert subnetwork performs matching calculations between the entities identified in the fragment and the entities in the knowledge base through vector embedding. If the match is successful, it extracts the associated security level, attributes, or hierarchical relationships to generate a knowledge matching feature vector. For entities without precise matching, it obtains relevant knowledge completion information through fuzzy matching and outputs knowledge matching features that describe the degree of association between the fragment and external knowledge and security rules.

[0083] The sparse gated routing mechanism assigns weights to each expert based on the features of the original fragment, dynamically routes the input, and outputs weighted fused high-order semantic features.

[0084] The neural causal reasoning layer receives weighted and fused high-order semantic features, performs causal inference based on the causal graph structure, and outputs the inferred security level label and sensitive entity combination label.

[0085] The permission decision layer encodes the security level label, sensitive entity label, context feature vector, and structural feature vector output by the neural causal reasoning layer to form a joint feature vector. The security level label and sensitive entity label are encoded using one-hot encoding, and the context feature and structural feature are normalized. The permission decision sub-network with a multilayer perceptron structure takes the joint feature vector as input, processes it through decision rules, performs feature fusion and condition judgment, and outputs a permission vector. The permission vector includes readable, writable, forwardable, shareable, downloadable, printable, and archiveable permission fields. Each field is 0 or 1, representing the accessibility of the fragment under the corresponding permission.

[0086] The classification output layer outputs security level labels, sensitive entity labels, and permission vectors. The security level labels include top secret, confidential, internal, and public.

[0087] This implementation method targets original fragments such as contracts, technical solutions, and project materials from information technology companies. First, the fragmented text is BERT encoded to obtain a context-dependent word vector sequence. This word vector sequence is then input into a bidirectional LSTM network for temporal feature modeling. The LSTM output is processed by a Conditional Random Field (CRF) layer to achieve sequence labeling of entity category and location tags for each word. Entity categories include API names, client companies, patent numbers, key employees, etc. All identified entities are then embedded to generate fixed-length entity embedding vectors, which serve as input for subsequent feature fusion and inference.

[0088] To achieve a comprehensive understanding of the fragment's content and context, the current fragment is concatenated with its preceding and following adjacent fragments, and encoded into a unified input sequence using the Word2Vec word embedding method. This sequence input is a three-layer Transformer structure, with each layer employing a self-attention mechanism to model high-order semantic dependencies between the fragment and its neighboring text. The mean pooling output of the last Transformer layer serves as the context feature vector for the current fragment, providing context-aware capabilities for subsequent discrimination.

[0089] All document fragments, tables, and image description nodes collectively construct a structural graph. Each node contains the structural features of the fragment, specifically including hierarchical number, belonging table, and position number. The edges of the graph represent the structural, referential, or sequential relationships between nodes. A GraphSAGE graph neural network is employed to complete multi-level message passing and node feature aggregation within the structural graph, ultimately outputting the structural feature vector of the current fragment. This achieves multi-level structural relationship modeling and enhances security discrimination capabilities.

[0090] By utilizing the output of the entity recognition expert subnetwork, key entities in the fragment are vectorized and matched one by one with the company's proprietary technical knowledge base, security thesaurus, and customer information database. Upon successful matching, the confidentiality level, technical affiliation, and hierarchical categories of the knowledge base entries are extracted to generate knowledge matching feature vectors. Unmatched entities are supplemented with weakly related features from the knowledge base using a semantic similarity algorithm. This module enables the automated identification and archiving of confidential and sensitive information such as technical terms, product models, and partners.

[0091] The entity embedding vector, context feature vector, structural feature vector, and knowledge matching feature vector are concatenated to form a multi-source feature vector for the fragment. This feature vector is input into a single-layer fully connected neural network, which outputs a weight vector of length equal to the number of expert subnetworks. The weight vector is then normalized using the Softmax activation function, retaining only the two expert subnetworks with the largest weights as activation branches. Input features are routed only to these activation branches, and the outputs of each expert are weighted and fused according to the Softmax weights to form the high-order semantic features of the current fragment, which serve as the sole input for subsequent causal inference.

[0092] A high-order fusion feature input is used to construct a neural causal inference layer. This layer employs a DAG-GNN causal structure learning algorithm to automatically build a directed acyclic causal graph with textual semantic features, contextual features, structural features, and knowledge matching features as nodes. Based on pre-trained parameters, causal edge weights between nodes are inferred, enabling causal link modeling of relationships such as technology dependence, personnel association, and customer level. Through multi-step message passing and node feature induction, the causal contribution of each node to the security level label and sensitive entity combination label is obtained, and the final security level label and sensitive entity combination label discrimination results are output.

[0093] The security level label, sensitive entity label, context feature vector, and structural feature vector output from the neural causal inference layer are concatenated. The security level label and sensitive entity label are encoded using one-hot encoding, and the feature vector is normalized. This joint feature vector is input into the multilayer perceptron (MLP) permission decision subnetwork. After decision rules are applied to fuse all feature information, a seven-dimensional permission vector for the fragment is output. Each permission field is strictly 0 or 1, corresponding to read, write, forward, share, download, print, and archive, respectively. This fully adapts to practical business scenarios such as read-only access to technical source code fragments, prohibition of sharing customer lists, and full-text access to public content.

[0094] The classification output layer uniformly outputs the security level tags (top secret, confidential, internal, public), sensitive entity tags, and permission vectors of the current segment. The above outputs serve as the basis for content generation, permission allocation, and secure content transfer, enabling fully automated, intelligent, and high-security management of business documents such as contracts, R&D materials, and customer data of information technology companies.

[0095] In this embodiment, the inference process of the neural causal reasoning layer includes:

[0096] It receives high-order semantic features obtained by gated routing and weighted fusion of multiple expert subnetworks as input;

[0097] Using the DAG-GNN algorithm, causal graphs are automatically constructed based on high-order semantic features:

[0098] The nodes of the causal graph include fragment text features, identified sensitive entities, context features, structural features, and external knowledge features;

[0099] Edges represent causal dependencies between nodes, and weights are learned through minimum description length, mutual information, or attention mechanisms.

[0100] Perform differentiable reasoning on the constructed causal graph:

[0101] The graph neural network employs a differentiable inference module. Each node's initial features are tensor representations of text features, identified sensitive entities, context features, structural features, and external knowledge features. The graph neural network performs multi-step message passing and feature aggregation on the causal graph. After each round of propagation, the node features are updated to a weighted sum of its own features and neighbor features. The weights are determined by the edge weights and trainable parameters. After multiple propagations, each node outputs a causal activation value.

[0102] In the last layer, the output node is defined as a security level label node and a sensitive entity combination label node. The final activation value of the node is used as the "causal contribution". The contribution of each input node to the security level label and sensitive entity combination label is quantified by L1 normalization.

[0103] The node with the largest contribution is selected as the main basis for judging the security level label and the sensitive entity combination label. The contribution of all nodes and the causal path are recorded as the inference link, and the final security level label and sensitive entity combination label are output.

[0104] In this embodiment, after receiving the high-order semantic features obtained by sparse gating routing and weighted fusion of the entity recognition expert subnetwork, context reasoning expert subnetwork, structural relationship expert subnetwork, and external knowledge matching expert subnetwork, the neural causal reasoning layer first automatically constructs a causal graph based on the DAG-GNN algorithm. The nodes of this causal graph strictly include fragment text feature nodes, sensitive entity feature nodes, context feature nodes, structural feature nodes, and external knowledge feature nodes, with each node corresponding to the fused output features of the aforementioned four types of expert networks.

[0105] Directed edges are configured between all node pairs in the causal graph. The edge weights are jointly optimized using the minimum description length criterion and mutual information index, and dynamically adjusted using an attention mechanism to ensure that the edge weights reflect the true causal dependencies between features. Subsequently, a differentiable inference module of a graph neural network is used to process the causal graph structure. The initial feature tensors of all nodes are input into the causal graph, and three rounds of message passing and feature aggregation operations are executed sequentially. During each round of propagation, the node features are updated according to the weighted sum of the node's features and the features of all its neighboring nodes. The weights are jointly determined by the aforementioned weighted sum neural network training parameters.

[0106] After the causal graph reasoning process is completed, the output nodes are designated as security level label nodes and sensitive entity combination label nodes. The final activation value of each input node is used as its causal contribution to the security level label and sensitive entity combination label. L1 normalization is applied to the causal contributions of all input nodes to quantify the impact of each type of input feature on the final reasoning result. Finally, based on the category of the node with the largest contribution, the security level label and sensitive entity combination label of the current segment are determined. The contribution results of all nodes and the complete causal path are stored as a reasoning link for subsequent permission decisions and content generation, achieving full interpretability and traceability of the discrimination process.

[0107] In this embodiment, the sparse gating routing mechanism includes:

[0108] The input consists of the concatenated text semantic feature vector, context feature vector, structural feature vector, and knowledge matching feature vector of the original fragment. The output is a weight vector with a length equal to the number of expert subnetworks, produced by a single-layer fully connected neural network.

[0109] Apply the Softmax activation function to the weight vector to obtain normalized weights;

[0110] Only the two expert branches with the highest weights are activated, and the weights of other branches are reset to zero;

[0111] The input data is dynamically routed to the activated expert sub-networks, and the features output by each sub-network are weighted and summed according to the Softmax weights to obtain the fused high-order semantic features.

[0112] In this embodiment, the generator of the reversible generative adversarial network includes:

[0113] The generator adopts a RealNVP reversible flow neural network structure, which consists of multiple reversible coupling layers connected in series. Each layer performs bidirectional transformation on the input features. The input is a joint input feature, which is encoded into a high-dimensional vector or tensor by the original fragment, security level label, sensitive entity label, and permission vector respectively, and then concatenated in a set order to generate the generator.

[0114] The combined input features undergo multiple transformations to output content versions corresponding to the permission constraints and security level requirements, with each content version corresponding one-to-one with the input features.

[0115] In this embodiment, the discriminator of the reversible generative adversarial network includes:

[0116] The discriminator is a multi-task neural network structure, and its input consists of the content version, security level label, permission vector, sensitive entity label, and original fragment output by the generator.

[0117] The discriminator contains three branches, namely:

[0118] The content desensitization validity branch determines whether sensitive information in the generated content has been effectively desensitized or blocked in accordance with permission requirements.

[0119] The security classification accuracy branch determines whether the security classification label of the generated content is consistent with the input security classification label;

[0120] The permission allocation consistency branch determines whether the generated content logically matches each permission field in the permission vector;

[0121] The three branches output a joint discrimination result.

[0122] The original fragment, security level label, sensitive entity label, and permission vector are encoded using high-dimensional vector or tensor encoding, and then concatenated in a preset order to form a joint input feature, which is then input into the generator module.

[0123] The generator employs a RealNVP reversible flow neural network structure, which consists of several reversible coupling layers connected in series. Each coupling layer performs a bidirectional affine transformation on the input features, enabling reversible information transfer between the joint input features and the content version. After multiple transformations, the generator output strictly corresponds to the content version under the constraints of the original input fragment, security level label, sensitive entity label, and permission vector. This achieves automatic generation of multiple versions of content under different permissions and security levels, ensuring a one-to-one correspondence between the content version and the joint input features.

[0124] The discriminator employs a multi-task neural network structure, taking the content version, security level label, permission vector, sensitive entity label, and original fragment output from the generator as input. Internally, the discriminator has three branches, each corresponding to a different discrimination task: the first branch judges the effectiveness of content desensitization, determining whether sensitive information in the generated content has been effectively desensitized or masked according to permission requirements; the second branch is responsible for judging the accuracy of security level classification, verifying whether the security level label of the generated content is completely consistent with the input security level label; the third branch performs permission allocation consistency judgment, verifying whether the actual performance of the generated content in each permission field logically matches the permission vector. The discrimination results output from the three branches are combined to comprehensively evaluate the effectiveness of the content version and provide feedback for optimizing generator parameters.

[0125] Each content generation, discrimination, and reverse mapping operation automatically generates a traceability log, recording input features, content version, discrimination results, and related operation information, enabling full-process traceability and auditing of content generation and permission transfer.

[0126] In this embodiment, the reversible generative adversarial network supports the reconstruction of the original fragment, original security level, sensitive entities, and initial permissions for any content version through reverse mapping.

[0127] In this embodiment, when document content exchange, circulation, or permission adjustment operations are triggered, the system calls the permission decision sub-network to regenerate the permission vector based on the target recipient's user identity, organization affiliation, historical access behavior, current access environment, operation type, and content sensitivity. The permission adjustment operations include automatically tightening or relaxing specific operation permission fields, temporarily upgrading security level labels, and adding or removing sensitive entity protection strategies.

[0128] Through the reverse mapping mechanism of reversible generative adversarial networks, the original fragments, original security levels, sensitive entities, and initial permissions can be accurately restored for any content version, thereby ensuring full-process traceability and content recovery capabilities in the multi-version content management process, and providing strong technical support for subsequent security audits, incident tracing, and responsibility allocation.

[0129] During content exchange, transfer, or permission adjustment, the permission decision subnetwork is dynamically invoked in real time to adaptively regenerate the permission vector by combining multi-dimensional information such as the recipient's identity, organizational affiliation, historical behavior, access environment, and content sensitivity. This enables flexible adjustments to permissions for different business scenarios, including automatic tightening, appropriate relaxation, temporary upgrade of security level, and protection strategies for sensitive entities. This significantly improves the intelligence, fine-grainedness, and dynamic response capabilities of access control for classified documents, effectively preventing the risk of unauthorized information dissemination and loss of control.

[0130] Example 1:

[0131] To verify the feasibility of this invention in practice, it was applied to the internal information management of an information technology company. The information involved various confidential documents, including contract management, technology research and development, customer data, and cooperation plans. Fifty actual contracts and internal R&D reports from the company over the past year were selected and divided into 400 segments, with content types including text, tables, and embedded image descriptions.

[0132] First, each document undergoes structured parsing, dividing the text, tables, and images into multiple original segments according to logical units. Then, semantic features are extracted from each segment: Chinese word segmentation tools are used for word segmentation and part-of-speech tagging, obtaining word sequences and parts of speech; the word sequences are fed into a Word2Vec model to extract the average word embedding vector for each segment; the BERT-CRF model is used for named entity recognition of the segment text, automatically labeling entities such as "API gateway source code," "customer name," and "employee number," and generating embedding features for each entity type; simultaneously, dependency parsing tools are used to extract structural grammatical features such as subject-verb-object and modifiers. These semantic features are then combined with the structural information of the segments to form the final feature vector used for subsequent AI inference.

[0133] Next, the feature vectors of each segment are sequentially input into four heterogeneous expert sub-networks: the entity recognition expert network outputs sensitive entities and their embedded representations based on the BERT-CRF structure; the context reasoning expert network concatenates the current and adjacent segment texts and feeds them into a three-layer Transformer to extract contextual dependency information; the structural relationship expert network, based on GraphSAGE, constructs a structural graph from document segments, tables, and image nodes, aggregates neighborhood features, and encodes structured relationships; and the external knowledge matching expert network, based on the company's knowledge base, aligns the identified entities with confidential terms, product lines, etc., to complete security attribute features. Through a sparse gating routing mechanism, these features are concatenated and fed into a single-layer fully connected neural network and normalized by Softmax, activating only the two expert branches with the largest weights. The features output by each sub-network are weighted and fused according to the assigned weights to obtain high-order semantic features.

[0134] The fused high-order features are input into the neural causal inference layer, automatically constructing a causal graph with textual semantic features, contextual features, structural features, and knowledge matching features as nodes. Edge weights are learned through an attention mechanism. A graph neural network is used for multi-step message passing and aggregation within the causal graph to obtain the causal contribution of each node to the security level label and the sensitive entity combination label. Finally, the node with the highest contribution is used as the decision criterion, outputting the security level label and the corresponding sensitive entity label for that segment. These outputs are then concatenated with contextual and structural features and fed into a three-layer multilayer perceptron permission decision subnetwork, automatically generating a permission vector containing seven fields: "readable, writable, forwardable, shareable, downloadable, printable, and archiveable," each field being either 0 or 1, achieving segment-level dynamic permission control.

[0135] The original fragment, security classification label, sensitive entity label, and permission vector are concatenated and input into the RealNVP reversible flow neural network generator to generate a content version that meets permission constraints and security classification requirements. For example, for a fragment containing "API gateway source code," if the permission vector only allows reading, the generator will automatically anonymize or mask the "source code" content. The discriminator simultaneously performs multi-task discrimination on the generated content version, including the effectiveness of content anonymization, the accuracy of security classification, and the consistency of permission allocation, and optimizes the generator parameters based on the discrimination feedback. All key operations, including content generation, security classification, permission allocation, content version flow, and reverse mapping, automatically generate traceability logs to ensure full traceability and security accountability throughout the process.

[0136] After extensive segment-level testing, the system can accurately identify and anonymize sensitive content from IT companies, such as "API gateway source code," "customer information," and "employee ID," and dynamically adjust permissions to achieve minimum authorization and secure delivery of multiple versions of content. Some experimental results are shown in the table below:

[0137] ;

[0138] Statistics show that the system's automatic identification and access control accuracy rate is over 96%, which greatly improves the company's automated security management capabilities for confidential documents and achieves a high degree of integration of content security, dynamic permissions, and full-process traceability.

[0139] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for classifying and controlling the exchange permissions of classified documents based on natural language processing, characterized in that, Includes the following steps: The input document is parsed in a structured manner, and it is divided into multiple original segments according to paragraphs, sentences, tables and images. The textual semantic features of each original segment are extracted as feature vectors. Input the original fragment and feature vector, and use a natural language processing model composed of a sparse gating expert hybrid network and a neural causal reasoning network to perform security level discrimination, and output security level label, sensitive entity label and permission vector; The original fragment, security level label, sensitive entity label, and permission vector are input together into a reversible generative adversarial network, which includes a generator and a discriminator. The generator is a reversible neural network structure that generates a content version that meets the permission constraints and security requirements based on the original fragment, security level label, sensitive entity label and permission vector. The discriminator directly receives the content version, security level label, permission vector, sensitive entity label and original fragment output by the generator as input, and performs joint discrimination on the effectiveness of content desensitization, the accuracy of security level classification and the consistency of permission allocation, and outputs the discrimination results to guide the generator optimization. In document exchange, circulation, and permission adjustment operations, all content generation, security classification, permission allocation, and reverse mapping processes are automatically recorded; The natural language processing model, composed of a sparsely gated expert hybrid network and a neural causal reasoning network, includes: The entity recognition expert subnetwork encodes fragmented text into a context-dependent word vector sequence using the BERT model. The word vector sequence is input into a bidirectional LSTM for temporal modeling to capture the dependencies between words. The LSTM output is processed by a CRF layer to predict the entity category and location label of each word. All entities are embedded into entity embedding vectors through embedding mapping. The context reasoning expert subnetwork concatenates the current segment and the preceding and following segments into an input sequence via word embedding. Multiple Transformers perform self-attention calculations on the concatenated sequence to model the higher-order semantic dependencies between the segment and its neighboring context. The mean pooling vector of the last Transformer layer serves as the context feature vector of the segment. The structural relationship expert subnetwork adopts a graph neural network structure to construct a document structure graph from document fragments, table and image nodes. Node attributes are fragment features, and edges represent structural, reference and order relationships. The structure graph is subjected to several layers of GNN message passing, the features of neighboring nodes are aggregated, structural context information is modeled, and the structural feature vector of the current fragment node is obtained. The external knowledge matching expert subnetwork performs matching calculations between the entities identified in the fragment and the entities in the knowledge base through vector embedding. If the match is successful, it extracts the associated security level, attributes or hierarchical relationships to generate a knowledge matching feature vector. If there is no precise matching entity, it obtains relevant knowledge completion information through fuzzy matching and outputs knowledge matching features describing the degree of association between the fragment and external knowledge and security rules. The sparse gated routing mechanism assigns weights to each expert based on the features of the original fragment, dynamically routes the input, and outputs weighted fused high-order semantic features. The neural causal reasoning layer receives weighted and fused high-order semantic features, performs causal inference based on the causal graph structure, and outputs the inferred security level label and sensitive entity combination label. The permission decision layer encodes the security level label, sensitive entity label, context feature vector, and structural feature vector output by the neural causal reasoning layer to form a joint feature vector. The security level label and sensitive entity label are encoded using one-hot encoding, and the context feature and structural feature are normalized. The permission decision sub-network with a multilayer perceptron structure takes the joint feature vector as input, processes it through decision rules, performs feature fusion and condition judgment, and outputs a permission vector. The permission vector includes readable, writable, forwardable, shareable, downloadable, printable, and archiveable permission fields. Each field is 0 or 1, representing the accessibility of the fragment under the corresponding permission. The classification output layer outputs security level labels, sensitive entity labels, and permission vectors. The security level labels include top secret, confidential, internal, and public.

2. The method for classifying and controlling the exchange permissions of classified documents based on natural language processing according to claim 1, characterized in that, The extraction of text semantic features from the original fragment includes the following steps: Each original segment is segmented and labeled with part-of-speech tags to extract word vector sequences; The pre-trained Word2Vec model is used to encode fragment text into semantic embedding vectors; Analyze entities, proper nouns, and technical terms in the fragment, and generate entity embedding representations; Dependency parsing is performed on the fragments to obtain the grammatical structural features of subject-verb-object and modification relations; The semantic embedding vector, entity embedding representation, and grammatical structure features of the original fragment are combined with structural information representing paragraph number, document level, and the position of the table or image to which it belongs to form a complete feature vector.

3. The method for classifying and controlling the exchange permissions of classified documents based on natural language processing according to claim 1, characterized in that, The inference process of the neural causal reasoning layer includes: It receives high-order semantic features obtained by gated routing and weighted fusion of multiple expert subnetworks as input; Using the DAG-GNN algorithm, causal graphs are automatically constructed based on high-order semantic features: The nodes of the causal graph include fragment text features, identified sensitive entities, context features, structural features, and external knowledge features; Edges represent causal dependencies between nodes, and weights are learned through minimum description length, mutual information, or attention mechanisms. Perform differentiable reasoning on the constructed causal graph: The graph neural network employs a differentiable inference module. Each node's initial features are tensor representations of text features, identified sensitive entities, context features, structural features, and external knowledge features. The graph neural network performs multi-step message passing and feature aggregation on the causal graph. After each round of propagation, the node features are updated to a weighted sum of its own features and neighbor features. The weights are determined by the edge weights and trainable parameters. After multiple propagations, each node outputs a causal activation value. In the last layer, the output node is defined as a security level label node and a sensitive entity combined label node. The final activation value of the node is used as the "causal contribution". The contribution of each input node to the security level label and sensitive entity combined label is quantified by L1 normalization. The node with the largest contribution is selected as the main basis for judging the security level label and the sensitive entity combination label. The contribution of all nodes and the causal path are recorded as the inference link, and the final security level label and sensitive entity combination label are output.

4. The method for classifying and controlling the exchange permissions of classified documents based on natural language processing according to claim 1, characterized in that, The sparse gating routing mechanism includes: The input consists of the concatenated text semantic feature vector, context feature vector, structural feature vector, and knowledge matching feature vector of the original fragment. The output is a weight vector with a length equal to the number of expert subnetworks, produced by a single-layer fully connected neural network. Apply the Softmax activation function to the weight vector to obtain normalized weights; Only the two expert branches with the highest weights are activated, and the weights of other branches are reset to zero; The input data is dynamically routed to the activated expert sub-networks, and the features output by each sub-network are weighted and summed according to the Softmax weights to obtain the fused high-order semantic features.

5. The method for classifying and controlling the exchange permissions of classified documents based on natural language processing according to claim 1, characterized in that, The generator of the reversible generative adversarial network includes: The generator adopts a RealNVP reversible flow neural network structure, which consists of multiple reversible coupling layers connected in series. Each layer performs bidirectional transformation on the input features. The input is a joint input feature, which is encoded into a high-dimensional vector or tensor by the original fragment, security level label, sensitive entity label, and permission vector respectively, and then concatenated in a set order to generate the generator. The combined input features undergo multiple transformations to output content versions corresponding to the permission constraints and security level requirements, with each content version corresponding one-to-one with the input features.

6. The method for classifying and controlling the exchange permissions of classified documents based on natural language processing according to claim 1, characterized in that, The discriminator of the reversible generative adversarial network includes: The discriminator is a multi-task neural network structure, and its input consists of the content version, security level label, permission vector, sensitive entity label, and original fragment output by the generator. The discriminator contains three branches, namely: The content desensitization effectiveness branch determines whether sensitive information in the generated content has been effectively desensitized or masked according to permission requirements; the security classification accuracy branch determines whether the security classification label of the generated content is consistent with the input security classification label. The permission allocation consistency branch determines whether the generated content logically matches each permission field in the permission vector; The three branches output a joint discrimination result.

7. The method for classifying and controlling the exchange permissions of classified documents based on natural language processing according to claim 1, characterized in that, The reversible generative adversarial network supports the inverse mapping of any content version to restore the original fragment, original security level, sensitive entities, and initial permissions.

8. The method for classifying and controlling the exchange permissions of classified documents based on natural language processing according to claim 1, characterized in that, When document content exchange, circulation, or permission adjustment operations are triggered, the system calls the permission decision sub-network to regenerate the permission vector based on the target recipient's user identity, organization affiliation, historical access behavior, current access environment, operation type, and content sensitivity. The permission adjustment operations include automatically tightening or relaxing specific operation permission fields, temporarily upgrading security level labels, and adding or removing sensitive entity protection policies.

Images