An archive sensitive information automatic desensitization processing method and system based on dynamic semantic recognition

Abstract

The application provides a kind of based on dynamic semantic identification's archive sensitive information automatic desensitization processing method and system, it is related to data desensitization technical field.The method includes: parsing original archive text constructs dynamic local semantic topology graph, and mapping to privacy knowledge graph to obtain global privacy attribute and propagation path;Track local context correlation path, jointly dynamically calculate the sensitivity index of candidate entity node;Accordingly, the target sensitive entity is screened and its semantic role feature is extracted;Finally, the semantic retention type desensitization strategy is matched to carry out placeholder replacement and topology connectivity repair, and desensitization archive text is generated.The application overcomes the defects that static rules lead to sensitive entity context misjudgment and mask destroys syntactic dependency structure, while ensuring privacy security compliance, the machine reading availability of archive data is maximized.

Description

Technical Field

[0001] This invention relates to the field of data desensitization technology, and in particular to a method and system for automatic desensitization of sensitive information in archives based on dynamic semantic recognition. Background Technology

[0002] Existing de-identification technologies for archival texts primarily rely on static rule matching and sequence labeling models. Researchers typically employ Conditional Random Field (CRF) models or bidirectional Long Short-Term Memory (LSTM) networks to construct named entity recognition (NIR) systems. These systems perform hard-localization of sensitive entities within the archival text based on a pre-defined data dictionary and regular expressions. The localization targets mainly include names, identification numbers, and home addresses. After localization, the system directly performs masking or pseudonym replacement. These methods have been widely adopted in electronic medical record de-identification literature published in journals such as the *Journal of Chinese Information Processing*.

[0003] Knowledge graph technology and pre-trained language models are gradually being introduced into archival management. The utilization of archival data has shifted from simple keyword retrieval to deep semantic-level machine reading comprehension. This shift requires that archival texts, after anonymization, must retain their original grammatical dependency structure and contextual coherence. Research on anonymization techniques has now shifted towards context-aware, semantically preservative anonymization.

[0004] Current static anonymization schemes exhibit significant technical flaws when handling complex archival files. First, existing models cannot dynamically identify the sensitivity of entities based on context. The system relies excessively on static tag libraries for entity identification. The same entity can play different semantic roles in different contexts. In such multi-semantic role scenarios, static tag libraries are highly prone to false positives or false negatives of sensitive entities. Second, direct masking forcibly severs the subject-verb-object dependency chain of the original archival text. This operation disrupts the syntactic topological connectivity of the archival text. The lack of topological connectivity renders the anonymized archival data unusable for downstream knowledge extraction and event mapping tasks. Ultimately, while meeting security and compliance requirements, the archival data completely loses its usability for machine reading comprehension. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide an automatic desensitization method and system for sensitive information in archives based on dynamic semantic recognition. This invention solves the problems of static desensitization methods in the prior art, which lead to misjudgment of sensitivity due to ignoring contextual semantic roles, and the loss of the usability of secondary mining of archive data due to forced masking destroying the topological connectivity of text.

[0006] To achieve the above objectives, the present invention provides the following solution: An automatic desensitization method for sensitive information in archives based on dynamic semantic recognition includes: Obtain the original archive text to be processed, parse the original archive text to extract candidate entity nodes and the dependency association edges between the candidate entity nodes, and construct a dynamic local semantic topology graph based on the candidate entity nodes and the dependency association edges; The dynamic local semantic topology graph is mapped to a preset domain privacy knowledge graph to perform structural alignment, so as to obtain the global privacy attribute corresponding to the candidate entity node in the domain privacy knowledge graph, and determine the privacy propagation path associated with the global privacy attribute; In the dynamic local semantic topology graph, local contextual association paths surrounding the candidate entity nodes are traced and extracted; Based on the global privacy attribute, the privacy propagation path, and the local context association path, the sensitivity index of the candidate entity node is dynamically calculated. Based on the sensitivity index, target sensitive entities are selected from the candidate entity nodes, and semantic role features corresponding to the target sensitive entities are extracted from the dynamic local semantic topology graph. Based on the semantic role features, a semantically preserved desensitization strategy is matched for the target sensitive entity, and the original archive text is processed according to the semantically preserved desensitization strategy to obtain the target desensitized archive text.

[0007] An automatic desensitization system for sensitive information in archives based on dynamic semantic recognition includes: The topology graph construction module is used to obtain the original archive text to be processed, parse the original archive text to extract candidate entity nodes and the dependency association edges between the candidate entity nodes, and construct a dynamic local semantic topology graph based on the candidate entity nodes and the dependency association edges. The graph mapping and path determination module is used to map the dynamic local semantic topology graph to a preset domain privacy knowledge graph to perform structural alignment, so as to obtain the global privacy attribute corresponding to the candidate entity node in the domain privacy knowledge graph, and determine the privacy propagation path associated with the global privacy attribute; The local context tracking module is used to track and extract the local context association paths surrounding the candidate entity nodes in the dynamic local semantic topology graph. The sensitivity index dynamic calculation module is used to dynamically calculate the sensitivity index of the candidate entity node based on the global privacy attribute, the privacy propagation path, and the local context association path. The target entity and feature extraction module is used to filter out target sensitive entities from the candidate entity nodes according to the sensitivity index, and extract the semantic role features corresponding to the target sensitive entities from the dynamic local semantic topology graph. The strategy matching and desensitization processing module is used to match a semantically preserved desensitization strategy for the target sensitive entity based on the semantic role features, and process the original file text according to the semantically preserved desensitization strategy to obtain the target desensitized file text.

[0008] The present invention discloses the following technical effects: This invention provides a method and system for automatically de-identifying sensitive information in archives based on dynamic semantic recognition. By constructing a dynamic local semantic topology graph and mapping it to a domain privacy knowledge graph, this invention dynamically calculates the sensitivity index of candidate entity nodes by combining global privacy attributes, privacy propagation paths, and local contextual association paths. This overcomes the problem of false positives or false negatives of sensitive entities caused by neglecting context in existing static de-identification methods, achieving accurate location of sensitive information in multiple contexts. Furthermore, this invention matches and executes a semantically preserving de-identification strategy based on the semantic role features of the target sensitive entity in the topology graph. While effectively eliminating the entity's true privacy referential attributes, it fully maintains the contextual dependencies of the target de-identified archive text and the structural connectivity of the topology graph. This effectively solves the defect of traditional forced masking operations that destroy the syntactic topological connectivity of archive text. While ensuring the security and compliance of archive data, it maximizes the machine readability and usability of the de-identified archive in downstream knowledge extraction and graph construction tasks. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart of an automatic desensitization method for sensitive information in archives based on dynamic semantic recognition, provided as an embodiment of the present invention. Detailed Implementation

[0011] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0012] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0013] like Figure 1 As shown, this invention provides an automatic desensitization method for sensitive information in archives based on dynamic semantic recognition, comprising: Step 100: Obtain the original archive text to be processed, parse the original archive text to extract candidate entity nodes and the dependency association edges between the candidate entity nodes, and construct a dynamic local semantic topology graph based on the candidate entity nodes and the dependency association edges; Step 200: Map the dynamic local semantic topology graph to a preset domain privacy knowledge graph and perform structural alignment to obtain the global privacy attribute corresponding to the candidate entity node in the domain privacy knowledge graph, and determine the privacy propagation path associated with the global privacy attribute; Step 300: In the dynamic local semantic topology graph, trace and extract the local context association paths surrounding the candidate entity nodes; Step 400: Based on the global privacy attribute, the privacy propagation path, and the local context association path, dynamically calculate the sensitivity index of the candidate entity node; Step 500: Based on the sensitivity index, select target sensitive entities from the candidate entity nodes, and extract the semantic role features corresponding to the target sensitive entities from the dynamic local semantic topology graph; Step 600: Based on the semantic role features, match a semantically preserved desensitization strategy for the target sensitive entity, and process the original archive text according to the semantically preserved desensitization strategy to obtain the target desensitized archive text.

[0014] Furthermore, the specific implementation process of step 100 is as follows: This embodiment acquires the original archival text to be processed. Here, the original archival text refers to natural language business files that have not undergone any data anonymization processing and retain complete contextual coherence. This embodiment converts the original archival text to be processed into a digitized character stream sequence. Since judicial or medical files are often lengthy, to ensure the alignment of the underlying computational tensor dimensions and prevent long sentence contexts from being rigidly blocked, this embodiment uses a moving window slicing method to segment the original archival text. The maximum text span length of a single processing window is set to 512 character units, and the overlap buffer length between two adjacent processing windows is set to 64 character units. This embodiment uses the sliced ​​character stream sequence as the standard input data for a preset sequence labeling model. The preset sequence labeling model is a deep perception model composed of a bidirectional long short-term memory network layer and a conditional random field layer stacked and finely tuned with historical corpus from the vertical archival domain. The historical corpus from the vertical archival domain comes from a collection of publicly available litigation documents and declassified personnel files that have been manually verified by experts and pre-exposed to remove real identity markers. Its function is to provide the sequence labeling model with initial entity mapping features that conform to rigorous archival writing logic.

[0015] This embodiment utilizes a preset sequence labeling model to extract multiple business entities from the original archive text, obtaining candidate entity nodes. Specifically, the bidirectional long short-term memory network layer first maps the character stream sequence into a high-dimensional word embedding dense feature matrix and calculates the forward and backward semantic dependency states of each character. Then, the conditional random field layer decodes and outputs the state transition probability of each character belonging to each type of business entity. The state transition probability is a normalized value calculated by summing the emission score of the current character as a specific entity label and the transition score from the previous character to the current character. This embodiment extracts character segments with state transition probabilities greater than a preset confidence threshold globally as extraction results. This preset confidence threshold is strictly set to 0.85. Conversely, for character segments with state transition probabilities not greater than 0.85, this embodiment forcibly classifies them as background noise text and discards them, thereby absolutely filtering out semantically ambiguous noise data at the feature extraction source. This embodiment instantiates the multiple business entities selected through threshold filtering into independent graph node data structures and confirms them as candidate entity nodes.

[0016] This embodiment performs syntactic dependency analysis on the original archive text to extract the syntactic modification relations and logical pointing relations between each candidate entity node, thereby obtaining the dependency association edges. This embodiment employs a transition-based probabilistic context-free grammar parsing mechanism to construct a global dependency syntax tree for the original archive text from the bottom up. Breadth-first node pathfinding is then performed within this global dependency syntax tree to trace the shortest tree-like topological path between any two candidate entity nodes. The syntactic modification relations refer to the attribute definition and restriction structure generated by attributive or adverbial clauses on candidate entity nodes, and the logical pointing relations refer to the dynamic effect flow and transmission structure between the action initiator and the action receiver in a subject-verb-object grammatical structure. To prevent structural misjudgments caused by invalid grammatical noise across long sentences, this embodiment sets the maximum effective dependency search distance limit to 12 lexical spans. When the path length between any two candidate entity nodes in the global dependency syntax tree is not greater than this limit and there is a valid dependency connection, this embodiment confirms that there is a syntactic modification relationship or a logical pointing relationship between the two and extracts it as the dependency association edge. If the path length is greater than 12 lexical spans, this embodiment hard cuts off the dependency determination between the two and confirms that the two are not related at the topological level.

[0017] This embodiment further performs attribute quantification processing on the dependency association edges, converting the extracted modification and logical relations into specific weight values ​​for graph calculation. Based on the specific dependency label type output from the global dependency syntax tree, this embodiment assigns a numerical weight representing the tightness of syntactic association to each dependency association edge. When the dependency association edge carries the core subject-verb relationship or verb-object dominance relationship in the logical pointing relationship, this embodiment assigns it the highest-level first structural weight, which is fixed at 0.90; when the dependency association edge carries the attributive-head modification relationship or adverbial-head modification relationship in the syntactic modification relationship, this embodiment assigns it a second-level second structural weight, which is fixed at 0.60. Simultaneously, this embodiment strictly follows the order of modification direction or action transmission in the real context, instantiating all dependency association edges assigned structural weights into directed graph edges with clearly defined starting and ending endpoints.

[0018] This embodiment constructs a multi-dimensional association network using each candidate entity node as a graph node and each dependency association edge as a directed graph edge to obtain the dynamic local semantic topology graph. This embodiment allocates an initial adjacency matrix storage space in the memory system, uses all candidate entity nodes as horizontal and vertical index identifiers of the adjacency matrix, and fills the dependency association edges and the corresponding quantized structural weight values ​​into the corresponding cross coordinate positions of the adjacency matrix, thereby completing the network construction. For isolated candidate entity nodes that fail to form dependency association edges with any other node, this embodiment adds a self-loop edge whose starting and ending endpoints both point to itself, and sets the corresponding structural weight value to 0.10, using this as the self-attention decay feature of the isolated entity to ensure that the connected components of the entire adjacency matrix do not experience singular breaks. To control the memory overhead of downstream graph computation, this embodiment sorts the candidate entity nodes in reverse order based on their frequency of occurrence in the text, limiting the maximum number of optimal candidate entity nodes allowed in a single network to 256. If the total number of extracted nodes exceeds 256, this embodiment strictly truncates and discards redundant nodes according to the sorting order, thereby generating and outputting the final dynamic local semantic topology graph.

[0019] Furthermore, the specific implementation process of steps 200-300 is as follows: This embodiment maps the dynamic local semantic topology graph to a pre-defined domain privacy knowledge graph for structural alignment. The domain privacy knowledge graph is a pre-constructed multi-relation directed graph network, whose underlying data originates from a database of tens of millions of real historical anonymized case files. This multi-relation directed graph network internally hard-coded sensitive hierarchical features of file categories and cross-domain anonymized business identifiers. To achieve continuous spatial measurement of the discrete graph structure, this embodiment pre-learns the full graph representation of the domain privacy knowledge graph using a graph neural network, projecting all graph nodes onto a low-dimensional dense semantic space. This embodiment extracts local subgraph features from the dynamic local semantic topology graph. This extraction process achieves topological representation vectors by vectorizing and fusing the node degree distribution, edge weights, and centrality indices in the topological structure. Subsequently, this embodiment performs isomorphic matching calculations between the extracted local subgraph features and the dense node vectors in the domain privacy knowledge graph. In the isomorphic matching calculation, this embodiment sets the graph structure similarity judgment threshold to 0.88. When the spatial mapping inverse Euclidean distance between the feature vectors of two topological subgraphs is greater than 0.88, this embodiment determines that the isomorphic matching is successful, and then performs the structure alignment and locates the completely corresponding knowledge graph anchor point in the global graph.

[0020] After successful matching, this embodiment extracts the sensitive level features and de-identified business identifiers mounted on the knowledge graph anchor points to obtain the global privacy attribute corresponding to the candidate entity node in the domain privacy knowledge graph. The sensitive level features characterize the absolute confidentiality assessment level of the entity within the national secrecy and privacy regulations system. This embodiment divides the sensitive level features into four discrete integer risk baseline values, assigning a weight of 0.95 to the top-secret level features, a weight of 0.50 to the regular sensitive level features, and a baseline value of 0.10 to the publicly available level features. The de-identified business identifier characterizes the specific case file business line classification to which the entity belongs, used to distinguish vertical fields such as criminal litigation, civil disputes, or medical and health care. To eliminate the dimensional barriers of heterogeneous data, this embodiment uses a multilayer perceptron to perform tensor concatenation with the one-hot encoding results of the extracted integer risk baseline values ​​and the de-identified business identifiers, and performs a linear feature mapping transformation to output a fixed-length high-dimensional dense vector feature, which is then confirmed as the global privacy attribute corresponding to the candidate entity node in the domain privacy knowledge graph.

[0021] This embodiment uses the knowledge graph anchor point as the starting node and performs multi-level traversal of the domain privacy knowledge graph according to a preset spillover diffusion mechanism to extract the link structure with topological connections, thereby obtaining the privacy propagation path associated with the global privacy attribute. The spillover diffusion mechanism refers to an algorithm that simulates the indirect association leakage risk of sensitive information in complex business chains due to graph structure associations. This embodiment uses a graph structure breadth-first search algorithm with a damping attenuation factor to perform the multi-level traversal. To prevent system memory overflow caused by endless computation of the entire graph, this embodiment strictly limits the maximum diffusion hop count to 3 hops. In the outward diffusion calculation of each hop, this embodiment sets the basic damping attenuation coefficient to 0.75, indicating that the privacy propagation potential of the secondary edge node is only 0.75 times that of the previous core node. The privacy propagation potential is calculated by combining the risk baseline value of the current node with the cumulatively multiplied basic damping attenuation coefficient.

[0022] The expression for the privacy propagation potential is: ; in, To spread to the first The privacy propagation potential during node jumps; The risk baseline value corresponding to the starting knowledge graph anchor point; This represents the maximum number of hops in the current diffusion path. This is the hop count index in the diffusion path; The set basic damping attenuation coefficient; It is a natural constant; For the first The number of in-degree connections of a jump node in the domain privacy knowledge graph is used to characterize the dilution effect of the node structure on privacy spillover.

[0023] In this embodiment, connected branch paths with accumulated attenuation potential energy of not less than 0.25 are tracked in real time. Link structures that satisfy the damping constraint and have continuous topological connections are merged and recombined to obtain the privacy propagation path associated with the global privacy attribute.

[0024] In this embodiment, considering the current text context, the local context association paths surrounding the candidate entity nodes are tracked and extracted in the dynamic local semantic topology graph. Specifically, this embodiment first identifies first-order neighbor nodes and corresponding connecting edges that have direct connections with the candidate entity nodes in the dynamic local semantic topology graph. Subsequently, this embodiment performs connectivity topology diffusion along the connecting edges to capture predicate nodes that indicate contextual action semantics and modifying nodes that indicate contextual business characteristics. The predicate nodes carry the specific action behavior that triggers a change in the sensitive state of the candidate entity, while the modifying nodes define the spatiotemporal or logical scope of the action behavior. This embodiment generates a multi-order reachability matrix by performing Boolean algebra iterative multiplication operations in the bottom-level adjacency matrix of the topology graph, searches outward to a maximum expansion of 2nd-order nodes, and uses a preset grammar paradigm matching library to perform part-of-speech verification on the edge nodes connected by the reachability matrix. The preset grammar paradigm matching library contains regular expressions for standardized legal document action pointing logic, accurately capturing and locking all the predicate nodes and modifying nodes that meet the requirements of the regular grammar.

[0025] This embodiment obtains the local contextual association path around the candidate entity node from the connected subgraph topology composed of the candidate entity node, the predicate node, the modifier node, and the dependency association edges connecting the predicate node and the modifier node. After capturing all associated nodes, this embodiment performs subgraph structure pruning based on dependency relationships, forcibly removing isolated side branches without syntactic logical association and redundant connection edges without attributes. This embodiment forcibly verifies the path connectivity attributes of the pruned subgraph to ensure that starting from any candidate entity node, there exists at least one continuous dependency association edge that can reach all its predicate nodes and modifier nodes in a forward direction; if the pruning process causes the path to break and lose connectivity, this embodiment reverts and restores the connectivity state before the previous pruning. To ensure the computational focus of the subsequent local context feature calculation process, this embodiment imposes an absolute size constraint on the extracted connected subgraph topology, limiting the maximum number of nodes contained in a single local contextual association path to 8. This embodiment encapsulates and stores the topology of the connected subgraph that satisfies the above-mentioned node quantity constraints and has a complete dependency loop, and outputs the local context association path around the candidate entity node.

[0026] Furthermore, the specific implementation process of steps 400-500 is as follows: This embodiment dynamically calculates the sensitivity index of the candidate entity nodes based on the global privacy attribute, the privacy propagation path, and the local context association path. Specifically, this embodiment extracts the sensitivity level label features from the global privacy attribute. These sensitivity level label features represent the absolute confidentiality threshold value of the entity in the original case file database. To achieve the conversion from discrete values ​​to a continuous high-dimensional space, this embodiment pre-initializes a 128-dimensional embedding weight matrix in the video memory, performs vector multiplication between the sensitivity level label features and this embedding weight matrix, and then converts it into the static basic risk vector of the candidate entity node through a preset nonlinear mapping function. The nonlinear mapping function uses a tangent activation transformation mechanism to project linear features into a nonlinear continuous vector space, eliminating the gradient discontinuity phenomenon caused by direct calculation of discrete values. Subsequently, this embodiment extracts serialized features from the multi-hop node relationships contained in the privacy propagation path. The multi-hop node relationships refer to the indirect connection structure state radiating outward from the graph anchor point. In this embodiment, the node attributes and connection weights in the path are concatenated into a one-dimensional feature sequence according to the diffusion level. For links with a path length of less than 3 hops, this embodiment performs zero-padding at the end of the sequence to ensure the consistency of the tensor length, thereby obtaining a fixed-dimensional propagation link risk tensor used to represent the long-range privacy diffusion probability.

[0027] This embodiment further extracts the lexical features and dependency syntax features of each node in the local context association path. The lexical features reflect the node's native part-of-speech distribution structure, and the dependency syntax features reflect the node's grammatical position in the overall sentence components. This embodiment transforms the above two types of features into high-dimensional real-number features by querying a preset static lexical feature vector table, and uses a preset bidirectional long short-term memory network structure to encode the sequence features of these high-dimensional real-number features. The bidirectional long short-term memory network structure has a gating mechanism that can simultaneously capture the deep semantic dependencies of forward inference and backward backtracking of the text. This embodiment sets the number of hidden layer neurons in this network structure to 256, and generates and outputs a local context state vector containing dynamic context information through the fusion calculation of forward and backward propagation features. In order to accurately capture the key business actions that trigger the desensitization state mutation, this embodiment uses the static basic risk vector as the query feature and performs multi-head attention mechanism calculation on the local context state vector. Specifically, this embodiment inputs the local context state vector into two parallel linear fully connected layers, and generates the corresponding key matrix and value matrix through weight mapping. In the calculation process, this embodiment obtains the attention weight matrix by calculating the inner product similarity between the query features and the key matrix, and then performs a weighted summation of the attention weight matrix and the value matrix to extract the dynamic context adjustment vector with specific action indication features.

[0028] The expression for the dynamic context adjustment vector is: ; in, The generated dynamic context adjustment vector; It is a normalized exponential function; The static underlying risk vector serves as the query feature; The key matrix generated by mapping the local context state vector; Let be the transpose of the key matrix; is a scalar dimension of the latent feature space during feature alignment, used to prevent gradient vanishing caused by excessively large inner product values; The value matrix generated by mapping the local context state vector.

[0029] After obtaining the multidimensional features, this embodiment performs matrix-dot product fusion on the static basic risk vector, the propagation link risk tensor, and the dynamic context adjustment vector in a preset latent feature space to obtain a multidimensional risk feature matrix. The latent feature space refers to the parameter-shared vector space that has undergone orthogonal transformation to eliminate collinearity interference between different features. Since the initial dimensions of the three heterogeneous feature vectors differ, this embodiment first calls three independent feedforward neural network channels to project and map these three vectors to a unified 512-dimensional space to achieve feature alignment. Subsequently, this embodiment performs Hadamard product operations on the dimension-aligned three feature matrices in the latent feature space to preserve the nonlinear interaction properties between each feature component, outputting a final 512-dimensional multidimensional risk feature matrix. Then, this embodiment inputs the multidimensional risk feature matrix into a preset fully connected perceptron for normalization and dimensionality reduction processing. The fully connected perceptron consists of three cascaded layers of gradually converging neural networks, with the number of neurons decreasing layer by layer until a single numerical node is finally output. In this embodiment, an exponential normalization function is used at the output node to compress the numerical values ​​and dynamically calculate the sensitivity index of the candidate entity node in the current context. The output value of this sensitivity index is rigidly limited to a continuous real number range from 0 to 100.

[0030] After completing the above index calculation, this embodiment filters out target sensitive entities from the candidate entity nodes based on the sensitivity index, and extracts the semantic role features corresponding to the target sensitive entities from the dynamic local semantic topology graph. Specifically, this embodiment compares the calculated sensitivity index with a dynamic adjustment threshold. The dynamic adjustment threshold is a security baseline parameter calculated in reverse from the confidentiality permission level of the terminal user currently accessing the archive retrieval system. This embodiment pre-configures a permission level mapping table; when the terminal user's permission is Level 1 (senior administrator), the dynamic adjustment threshold is set to 90; when the permission is Level 3 (ordinary staff), the dynamic adjustment threshold is lowered to 75. This embodiment forcibly filters and confirms candidate entity nodes with a sensitivity index not less than the dynamic adjustment threshold as the target sensitive entities. For each outstanding target sensitive entity, this embodiment extracts its subgraph position in the dynamic local semantic topology graph and extracts the edge type of the dependent association edges directly connected to it. This embodiment identifies the syntactic logic role and business operation role of the target sensitive entity in the current context by reading the absolute centrality parameter of the subgraph position and the pointing relationship of the edge type. This embodiment sets a classification matching rule: when the absolute centrality parameter is greater than 0.6 and the dependency relationship shows an outward diverging dominant pointing relationship, this embodiment identifies the entity as a syntactic action initiator; otherwise, it is identified as a modifier.

[0031] Finally, this embodiment combines the grammatical logic roles and the business operation roles to extract the semantic role features corresponding to the target sensitive entity. This embodiment allocates a dedicated feature storage block in the system's underlying memory, concatenating the one-hot encoded vector of the grammatical logic roles with the attribute classification enumeration values ​​of the business operation roles using tensors. Specifically, this embodiment strictly connects the grammatical role encoding array of length 4 and the business role encoding array of length 6 according to a preset positional order, generating a composite role vector with a total feature length of 10. This embodiment converts this composite role vector containing complete state information into a binary data block and persists it in the attribute field of the corresponding node in the topology graph, serving as the final output semantic role feature corresponding to the target sensitive entity. This semantic role feature provides a unique and deterministic classification decision basis for accurately matching semantically preserving desensitization rules with anti-fragmentation capabilities in subsequent technical steps, fundamentally avoiding the desensitization strategy mismatch and grammatical destruction caused by relying solely on a single part-of-speech tag in traditional desensitization techniques.

[0032] Furthermore, the specific implementation process of step 600 is as follows: This embodiment, based on the grammatical logic role and business operation role indicated in the semantic role features, performs conditional retrieval in a preset desensitization strategy rule base to match and obtain the corresponding target placeholder generation rule, thus obtaining the semantically preserving desensitization strategy. Specifically, the preset desensitization strategy rule base is a non-relational configuration table storing pre-defined graph structure mapping conditions. This configuration table internally has a rigidly cascaded set of five-level lexical superordinate classification dictionaries. After parsing the semantic role features, this embodiment performs precise branch logic retrieval in the desensitization strategy rule base: If the semantic role features indicate that the target sensitive entity is the core subject, this embodiment retrieves matching similar concept forgery and replacement rules in the preset desensitization strategy rule base as the target placeholder generation rule; if the semantic role features indicate that the target sensitive entity is a modifying subordinate entity, this embodiment retrieves matching category superordinate concept generalization rules in the preset desensitization strategy rule base, and extracts parent node category words spanning two levels as desensitized content by querying the vocabulary superordinate classification dictionary, thus obtaining the target placeholder generation rule; if the semantic role features indicate that the target sensitive entity is a logical reasoning hub entity, indicating that the entity is connected to more than or equal to 3 predicate nodes in the topology graph, this embodiment retrieves matching homomorphic encryption anti-fragmentation desensitization rules in the preset desensitization strategy rule base, thus obtaining the target placeholder generation rule.

[0033] This embodiment executes the target placeholder generation rule to remove the target sensitive entities in the original file text and generates desensitized placeholder nodes that match the original part-of-speech tags of the target sensitive entities. Specifically, in the underlying character stream sequence, this embodiment locks the start and end byte offsets of the target sensitive entities and completely clears and overwrites the real character data within this span. When executing the homomorphic encryption anti-fragmentation desensitization rule, considering the uncertainty of the original character length, this embodiment uses the national commercial cryptography hash algorithm standard to perform irreversible hash mapping transformation on the real entity characters. This embodiment truncates the fixed-length digest data output after hash transformation, fixing the output length to 16 bytes as the desensitized placeholder node. This ciphertext sequence is forcibly bound to the part-of-speech tag corresponding to the original entity by the system to ensure that the part-of-speech recognition probability of this node in the downstream machine reading comprehension model remains absolutely consistent before and after desensitization, thereby preventing the collapse of the lexical parse tree.

[0034] This embodiment maps the desensitized placeholder nodes back to the original dynamic local semantic topology graph before processing according to their original character sequence index positions to obtain a post-topology graph containing the desensitized placeholder nodes. In the graph memory space, this embodiment replaces the original sensitive entity graph nodes with the newly generated desensitized placeholder nodes, inheriting all in-degree and out-degree edges attached to them, thus generating the post-topology graph. To quantify the connectivity changes of the graph structure in a rigorous mathematical space, this embodiment extracts the post-Laplace matrix of the post-topology graph and the initial Laplace matrix of the original dynamic local semantic topology graph. Given that the dependency network is a directed asymmetric structure, this embodiment first ignores the direction attributes of all directed graph edges, reducing its dimensionality to an undirected structural skeleton graph. Subsequently, this embodiment calculates the connectivity degree of each node in the two structural skeleton graphs and constructs the corresponding diagonal matrix, subtracting the adjacency matrix formed by the edge weights, outputting the initial Laplace matrix and the post-Laplace matrix with strictly aligned feature dimensions.

[0035] This embodiment verifies the consistency of the network structure by calculating the eigenvalue difference between two Laplace matrices. Eigenvalue difference is a core mathematical measure of whether two complex networks undergo a fundamental shift in the number of connected components and the efficiency of information flow. To eliminate complex interference from asymmetric references, this embodiment performs singular value decomposition on both the initial and subsequent Laplace matrices, extracting two sets of real singular value sequences arranged in ascending order of numerical value. This embodiment calculates the absolute difference of the corresponding rank values ​​in each of the two sets of singular value sequences and sums all the absolute differences to obtain the total eigenvalue difference, which characterizes the severity of structural drift.

[0036] The expression for the total eigenvalue difference is: ; in, The difference in the total eigenvalues ​​obtained through calculation; This represents the total number of non-zero singular values ​​extracted after singular value decomposition. The descending index number of the singular value; The first of the post-Laplace matrix One singular value; The first Laplace matrix is ​​the first... One singular value; The symbol for absolute value calculation; This is a maximum value function used to extract the larger values ​​in the sequence as the normalized denominator; To prevent extremely small positive numbers with a denominator of zero.

[0037] Subsequently, this embodiment compares the feature value difference with a preset graph structure tolerance threshold. The graph structure tolerance threshold is fixed at 0.15, used to define the critical safety line at which entity substitution operations cause damage to the context syntax structure. When the calculated feature value difference is not greater than 0.15, this embodiment determines that no topological distortion has occurred, i.e., the graph structure consistency verification passes; when the feature value difference is greater than 0.15, it indicates that the graph structure consistency verification fails.

[0038] When the difference in eigenvalues ​​exceeds a preset graph structure tolerance threshold, indicating that the graph structure consistency verification has failed, this embodiment locates the broken connected components in the subsequent topology graph where dependency relationships have been broken, and triggers the grammatical coherence repair rule. This embodiment determines the increment in the number of connected branches by comparing the vector dimensions corresponding to the null spaces of the Laplacian matrices before and after, and reversely calculates the set of isolated entity nodes that have lost their core syntactic dependencies, thus completing the precise location of the broken connected components. According to the grammatical coherence repair rule, this embodiment specifically inserts logical conjunctions or prepositions to bridge the broken connected components at the corresponding positions of the target sensitive entities in the original archive text. Specifically, this embodiment queries a preset repair compensation dictionary; when the broken edge type is an action-dominant association, it forcibly inserts a preposition character representing a passive or dispositional action; when the broken edge type is an attribute-limited association, it forcibly inserts a structural particle character representing a subordinate relationship, in order to reconstruct the logical pointing relationship between the two on the abstract syntax tree. After the insertion is completed, this embodiment uses the updated text character sequence to reconstruct the underlying network and re-executes the aforementioned Laplacian eigenvalue difference calculation to implement graph structure consistency verification. Through closed-loop iteration until the eigenvalue difference drops below 0.15, the graph structure consistency verification is passed. Finally, the target desensitized archive text that maintains the connectivity of the dynamic local semantic topology graph is generated and output.

[0039] This embodiment also provides an automatic desensitization system for sensitive information in archives based on dynamic semantic recognition, including: The topology graph construction module is used to obtain the original archive text to be processed, parse the original archive text to extract candidate entity nodes and the dependency association edges between the candidate entity nodes, and construct a dynamic local semantic topology graph based on the candidate entity nodes and the dependency association edges. The graph mapping and path determination module is used to map the dynamic local semantic topology graph to a preset domain privacy knowledge graph to perform structural alignment, so as to obtain the global privacy attribute corresponding to the candidate entity node in the domain privacy knowledge graph, and determine the privacy propagation path associated with the global privacy attribute; The local context tracking module is used to track and extract the local context association paths surrounding the candidate entity nodes in the dynamic local semantic topology graph. The sensitivity index dynamic calculation module is used to dynamically calculate the sensitivity index of the candidate entity node based on the global privacy attribute, the privacy propagation path, and the local context association path. The target entity and feature extraction module is used to filter out target sensitive entities from the candidate entity nodes according to the sensitivity index, and extract the semantic role features corresponding to the target sensitive entities from the dynamic local semantic topology graph. The strategy matching and desensitization processing module is used to match a semantically preserved desensitization strategy for the target sensitive entity based on the semantic role features, and process the original file text according to the semantically preserved desensitization strategy to obtain the target desensitized file text.

[0040] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0041] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for automatically de-identifying sensitive information in archives based on dynamic semantic recognition, characterized in that, include: Obtain the original archive text to be processed, parse the original archive text to extract candidate entity nodes and the dependency association edges between the candidate entity nodes, and construct a dynamic local semantic topology graph based on the candidate entity nodes and the dependency association edges; The dynamic local semantic topology graph is mapped to a preset domain privacy knowledge graph to perform structural alignment, so as to obtain the global privacy attribute corresponding to the candidate entity node in the domain privacy knowledge graph, and determine the privacy propagation path associated with the global privacy attribute; In the dynamic local semantic topology graph, local contextual association paths surrounding the candidate entity nodes are traced and extracted; Based on the global privacy attribute, the privacy propagation path, and the local context association path, the sensitivity index of the candidate entity node is dynamically calculated. Based on the sensitivity index, target sensitive entities are selected from the candidate entity nodes, and semantic role features corresponding to the target sensitive entities are extracted from the dynamic local semantic topology graph. Based on the semantic role features, a semantically preserved desensitization strategy is matched for the target sensitive entity, and the original archive text is processed according to the semantically preserved desensitization strategy to obtain the target desensitized archive text.

2. The method for automatic desensitization of sensitive information in archives based on dynamic semantic recognition according to claim 1, characterized in that, The process of acquiring the original archive text to be processed, parsing the original archive text to extract candidate entity nodes and dependency edges between the candidate entity nodes, and constructing a dynamic local semantic topology graph based on the candidate entity nodes and the dependency edges includes: The original archive text to be processed is obtained, and multiple business entities in the original archive text are extracted using a preset sequence labeling model to obtain the candidate entity nodes; Syntactic dependency analysis is performed on the original archive text to extract the syntactic modification relations and logical pointing relations between each candidate entity node, thereby obtaining the dependency association edges; A multidimensional association network is constructed using each candidate entity node as a graph node and each dependency association edge as a directed graph edge to obtain the dynamic local semantic topology graph.

3. The method for automatic desensitization of sensitive information in archives based on dynamic semantic recognition according to claim 1, characterized in that, The step of mapping the dynamic local semantic topology graph to a preset domain privacy knowledge graph to perform structural alignment, in order to obtain the global privacy attributes corresponding to the candidate entity nodes in the domain privacy knowledge graph, and determining the privacy propagation path associated with the global privacy attributes, includes: Extract the local subgraph features of the dynamic local semantic topology graph, and perform isomorphic matching with the preset domain privacy knowledge graph to perform the structure alignment and locate the corresponding knowledge graph anchor point; Extract the sensitive hierarchical features and desensitized business identifiers mounted on the anchor points of the knowledge graph to obtain the global privacy attributes corresponding to the candidate entity nodes in the domain privacy knowledge graph; Starting from the anchor point of the knowledge graph, a multi-level traversal is performed on the domain privacy knowledge graph according to a preset spillover diffusion mechanism to extract the link structure with topological connection relationship and obtain the privacy propagation path associated with the global privacy attribute.

4. The method for automatic desensitization of sensitive information in archives based on dynamic semantic recognition according to claim 3, characterized in that, The domain privacy knowledge graph is a pre-constructed multi-relation directed graph network, which includes: sensitive hierarchical features of file categories and de-identified business identifiers across domains.

5. The method for automatic desensitization of sensitive information in archives based on dynamic semantic recognition according to claim 1, characterized in that, The step of tracing and extracting local contextual association paths around the candidate entity nodes in the dynamic local semantic topology graph includes: In the dynamic local semantic topology graph, the first-order neighbor nodes that have a direct connection relationship with the candidate entity node and the corresponding connection edges are identified. Connectivity topology diffusion is performed along the connecting edges to capture predicate nodes that indicate contextual action semantics and modifier nodes that indicate contextual business features; The local context association path around the candidate entity node is obtained by taking the connected subgraph topology formed by the candidate entity node, the predicate node, the modifier node, and the dependency association edges connecting the predicate node and the modifier node.

6. The method for automatic desensitization of sensitive information in archives based on dynamic semantic recognition according to claim 1, characterized in that, The dynamic calculation of the sensitivity index of the candidate entity node based on the global privacy attribute, the privacy propagation path, and the local context association path includes: The sensitive hierarchical label features in the global privacy attributes are extracted and converted into static basic risk vectors of the candidate entity nodes through a preset nonlinear mapping function. Serialization features are extracted from the multi-hop node relationships contained in the privacy propagation path to obtain a propagation link risk tensor representing the long-range privacy diffusion probability; Lexical features and dependency syntax features of each node in the local context association path are extracted, and sequence feature encoding is performed using a bidirectional long short-term memory network structure to obtain a local context state vector. Using the static basic risk vector as the query feature, a multi-head attention mechanism is performed on the local context state vector to extract a dynamic context adjustment vector with specific action indication features. The static basic risk vector, the propagation link risk tensor, and the dynamic context adjustment vector are fused by matrix multiplication in a preset latent feature space to obtain a multidimensional risk feature matrix. The multidimensional risk feature matrix is ​​input into a preset fully connected perception network for normalization and dimensionality reduction processing, and the sensitivity index of the candidate entity node in the current context is dynamically calculated and output.

7. The method for automatic desensitization of sensitive information in archives based on dynamic semantic recognition according to claim 1, characterized in that, The step of filtering target sensitive entities from candidate entity nodes based on the sensitivity index and extracting the semantic role features corresponding to the target sensitive entities from the dynamic local semantic topology graph includes: Candidate entity nodes whose sensitivity index is not less than the dynamic adjustment threshold are selected and confirmed as the target sensitive entities; Based on the subgraph position of the target sensitive entity in the dynamic local semantic topology graph and the edge type of the associated dependency association edge, the syntactic logic role and business operation role of the target sensitive entity in the current context are identified. The grammatical logic role and the business operation role are combined and represented to extract the semantic role features corresponding to the target sensitive entity.

8. The method for automatic desensitization of sensitive information in archives based on dynamic semantic recognition according to claim 1, characterized in that, The process of matching a semantically preserving desensitization strategy to the target sensitive entity based on the semantic role features, and processing the original document text according to the semantically preserving desensitization strategy to obtain the target desensitized document text, includes: Based on the syntactic logic role and business operation role indicated in the semantic role features, a conditional search is performed in the preset desensitization strategy rule base to match the corresponding target placeholder generation rule and obtain the semantic preservation desensitization strategy. The target placeholder generation rule is executed to remove the target sensitive entity in the original file text, and a desensitized placeholder node that matches the original part of speech of the target sensitive entity is generated. The desensitized placeholder nodes are mapped back to the original character sequence index position in the dynamic local semantic topology graph before processing to obtain a post-topology graph containing the desensitized placeholder nodes. Extract the post-Laplacian matrix of the post-topology graph and the initial Laplacian matrix of the dynamic local semantic topology graph before processing, and verify the consistency of the graph structure by calculating the difference in eigenvalues ​​of the two Laplacian matrices. When the difference in eigenvalues ​​exceeds a preset graph structure tolerance threshold, indicating that the graph structure consistency verification has failed, the broken connected domains in the subsequent topology graph where dependency relationships have been broken are located, and the syntax coherence repair rule is triggered. According to the grammatical coherence repair rules, logical connectors or prepositions are inserted at the corresponding positions of the target sensitive entities in the original archive text to bridge the broken connected domains, and the graph structure consistency verification is re-executed until the graph structure consistency verification passes, thus obtaining the target desensitized archive text that maintains the connectivity of the dynamic local semantic topology graph.

9. The method for automatic desensitization of sensitive information in archives based on dynamic semantic recognition according to claim 8, characterized in that, The syntactic logic role and business operation role indicated by the semantic role features are used to perform conditional retrieval in a preset desensitization strategy rule base to obtain the corresponding target placeholder generation rule, including: If the semantic role feature indicates that the target sensitive entity is the core subject, then a matching concept forgery and replacement rule is retrieved from the preset desensitization strategy rule base as the target placeholder generation rule; If the semantic role feature indicates that the target sensitive entity is a decorative subordinate entity, then the matching category superordinate concept generalization rule is retrieved from the preset desensitization strategy rule base to obtain the target placeholder generation rule; If the semantic role feature indicates that the target sensitive entity is a logical reasoning hub entity, then a matching homomorphic encryption anti-fragmentation desensitization rule is retrieved from the preset desensitization strategy rule base to obtain the target placeholder generation rule.

10. An automatic desensitization system for sensitive information in archives based on dynamic semantic recognition, characterized in that, include: The topology graph construction module is used to obtain the original archive text to be processed, parse the original archive text to extract candidate entity nodes and the dependency association edges between the candidate entity nodes, and construct a dynamic local semantic topology graph based on the candidate entity nodes and the dependency association edges. The graph mapping and path determination module is used to map the dynamic local semantic topology graph to a preset domain privacy knowledge graph to perform structural alignment, so as to obtain the global privacy attribute corresponding to the candidate entity node in the domain privacy knowledge graph, and determine the privacy propagation path associated with the global privacy attribute; The local context tracking module is used to track and extract the local context association paths surrounding the candidate entity nodes in the dynamic local semantic topology graph. The sensitivity index dynamic calculation module is used to dynamically calculate the sensitivity index of the candidate entity node based on the global privacy attribute, the privacy propagation path, and the local context association path. The target entity and feature extraction module is used to filter out target sensitive entities from the candidate entity nodes according to the sensitivity index, and extract the semantic role features corresponding to the target sensitive entities from the dynamic local semantic topology graph. The strategy matching and desensitization processing module is used to match a semantically preserved desensitization strategy for the target sensitive entity based on the semantic role features, and process the original file text according to the semantically preserved desensitization strategy to obtain the target desensitized file text.

Images