A Method and System for Intelligent Authentication and Automated Processing of Electronic Archives Based on Semantic Matching

Abstract

This invention relates to the field of electronic records management technology, and more particularly to a method and system for intelligent identification and automatic disposal of electronic records based on semantic matching. The method involves semantically decomposing identification rules and encoding record features, calculating semantic distance to generate recommended retention periods and destruction trigger scores. A directed acyclic graph of associated records is constructed and bidirectional information propagation is performed. Constraint conflicts or circular dependencies are detected and decomposed. Destruction conditions are determined based on the scores and recommended values ​​of nodes within the subgraph. When the conditions are met, the records are automatically migrated to the destruction database. This achieves automation, intelligence, and conflict resolution in electronic records identification, improving disposal efficiency and accuracy.

Description

Technical Field

[0001] This invention relates to the field of electronic records management technology, and in particular to a method and system for intelligent identification and automatic processing of electronic records based on semantic matching. Background Technology

[0002] The authentication and disposal of electronic records is a core aspect of archival management. Its goal is to assess the value of massive amounts of electronic records based on established rules and retention periods, ultimately deciding whether to preserve or destroy them. Traditional automated or semi-automated methods typically rely on keyword-based rule matching or simple metadata comparison to make authentication decisions.

[0003] Conventional practices typically treat pre-defined identification rules as a series of static logical judgments, such as directly matching specific keywords in the document title, document number, or classification number. The document content itself is represented by extracting keywords or through simple vectorization. The system triggers corresponding retention period determinations or disposal actions by comparing whether the document characteristics match the keywords in the rules. This method is feasible to a certain extent when the rules are clear and the document format is standardized.

[0004] With the increasing diversification of sources and the growing complexity of content formats in electronic archives, the conventional method based on surface keyword matching has gradually revealed its significant limitations. The true value of electronic archives often lies in their deeper semantic content, their historical context, and their relationships with other archives. Simple keyword matching is insufficient to accurately understand the complete semantics of the archive content and is prone to misjudgment or omission due to the diversity of vocabulary.

[0005] Conventional methods typically handle individual files in isolation, lacking a systematic consideration of the complex relationships between them. In actual management, files are generally closely linked through citations, attachments, and workflow continuity, forming an organic network. Decisions regarding the disposal of any file, especially its destruction, must consider the status of its related files to avoid disrupting the integrity of the file chain or causing inconsistencies in business logic. Existing methods struggle to model and reason about such complex dependencies, easily leading to disposal conflicts or errors, such as destroying critical documents still referenced by other important files. Summary of the Invention

[0006] The embodiments of the present invention provide a method and system for intelligent identification and automatic processing of electronic archives based on semantic matching, which can solve the problems in the prior art.

[0007] A first aspect of this invention provides a method for intelligent identification and automatic processing of electronic records based on semantic matching, comprising:

[0008] Obtain the electronic files to be authenticated, along with the corresponding authentication rules and retention period constraints;

[0009] The identification rules are semantically decomposed to extract conditional predicates. The content of the electronic archive to be identified is encoded with multi-level features and the time decay factor of the archive formation time and business context are introduced to obtain the archive features of the electronic archive to be identified.

[0010] The semantic distance between the archive features and the conditional predicates is calculated in the feature space, and the matching results of different conditional predicates are fused to generate a recommended retention period value and a destruction trigger score.

[0011] Construct a directed acyclic graph containing the electronic files to be identified and the corresponding multi-level related files. Perform bidirectional information propagation based on message passing on the directed acyclic graph. During forward propagation, the retention period constraint and disposal status of the parent node are passed along the edge to the child node and the constraint strength is accumulated. During reverse propagation, the unfinished disposal mark and blocking information of the child node are sent back to the parent node.

[0012] When a conflict or circular dependency in the retention period constraint between nodes is detected through the bidirectional information propagation, the directed acyclic graph is decomposed into multiple subgraphs that can be disposed of independently. The destruction trigger score and the recommended retention period value of each node in the subgraph are used to determine whether the destruction conditions are met. When the destruction conditions are met, the electronic file to be identified corresponding to the current node is migrated to the destruction database.

[0013] The identification rules are semantically decomposed to extract conditional predicates. The content of the electronic archive to be identified is then multi-level feature encoded, and a temporal decay factor of the archive's formation time and business context are introduced to obtain the archive features of the electronic archive to be identified, including:

[0014] Dependency parsing is performed on the text description in the identification rules to identify the subject-verb-object structure and modification relations in the identification rules. The predicate part in the subject-verb-object structure is used as the core semantic unit of the condition judgment, and the qualifiers and constraints in the modification relations are used as the boundary constraints of the condition judgment. The core semantic unit and the boundary constraints are extracted to form the structured representation of the condition predicate.

[0015] The structured fields and unstructured text of the electronic archive to be identified are encoded and feature extracted to obtain basic features. The interaction weight between the structured fields and the unstructured text is calculated through a self-attention mechanism. The basic features are then weighted and fused according to the interaction weight to obtain a fused feature representation.

[0016] The time difference between the archive creation time and the current time is calculated and mapped to an attenuation coefficient. The attenuation coefficient is inversely proportional to the time difference. The attenuation coefficient is multiplied element-wise with the fused feature representation to obtain the time-attenuated feature.

[0017] The identification information of the preceding and subsequent business nodes is extracted from the business process to which the electronic file to be identified belongs and encoded into a context vector. The context vector is then concatenated with the time-decayed features to obtain the file features.

[0018] The semantic distance between the archive features and the conditional predicates is calculated in the feature space, and the matching results of different conditional predicates are fused to generate a recommended retention period value and a destruction trigger score, including:

[0019] The structured representations of the archive features and the conditional predicates are mapped to a unified semantic feature space, and the distance metric between the vector representations of the archive features and the vector representations of the structured representations of the conditional predicates is calculated in the semantic feature space as the semantic distance.

[0020] The semantic distance between the file features and the multiple conditional predicates contained in the identification rule is calculated respectively to obtain multiple semantic distance values. The weight coefficient of each conditional predicate is determined according to the logical relationship of each conditional predicate in the identification rule. The multiple semantic distance values ​​and the corresponding weight coefficients are weighted and summed to obtain the comprehensive matching score.

[0021] The retention period labeling information associated with each conditional predicate is extracted from the identification rules. The retention period labeling information is weighted and averaged according to the semantic distance between each conditional predicate and the archival features to obtain the recommended retention period value. The weight of the retention period labeling information corresponding to each conditional predicate is inversely proportional to the semantic distance.

[0022] The recommended retention period is compared with the actual retention time of the electronic file to be identified. When the actual retention time exceeds the recommended retention period, the excess ratio is calculated. The excess ratio is weighted and combined with the comprehensive matching score to obtain the destruction trigger score.

[0023] Constructing a directed acyclic graph that includes the electronic file to be identified and its corresponding multi-level related files includes:

[0024] Extract a set of identifiers for associated files from the metadata and business attributes of the electronic file to be identified, parse the association type attribute between each associated file in the identifier set and the electronic file to be identified, and determine a business dependency relationship when the association type attribute is identified as a business process prerequisite, and a reference relationship when the association type attribute is identified as a content reference source.

[0025] The hierarchical structure between files is determined based on the business dependency relationship and the reference relationship. The electronic file to be identified is taken as the target node, the files that the electronic file to be identified depends on and the files that are referenced by the electronic file to be identified are taken as parent nodes, and the files that depend on the electronic file to be identified and the files that reference the electronic file to be identified are taken as child nodes. A directed edge of dependency is established from the dependent file node to the dependent file node based on the business dependency relationship, and a directed edge of reference is established from the referenced file node to the referencing file node based on the reference relationship.

[0026] The association strength value is calculated and normalized based on the access frequency and time span between the associated archive and the electronic archive to be identified, and used as the weight of the corresponding directed edge to obtain the directed acyclic graph.

[0027] On the directed acyclic graph, bidirectional message-passing-based information propagation is performed. During forward propagation, the retention period constraint and disposal status of the parent node are passed along the edges to the child nodes, accumulating constraint strength. During backward propagation, the incomplete disposal flag and blocking information of the child nodes are sent back to the parent node, including:

[0028] Starting from the root node of the directed acyclic graph, forward propagation is performed. The retention period constraint carried by the current parent node is read as the constraint base value. The constraint base value is multiplied by the weight of the edge between the current parent node and the child node to obtain the transit constraint value.

[0029] The cumulative constraint strength value of the child node is obtained by summing the transmitted constraint value with the historical constraint value received by the child node. At the same time, the current parent node's disposal status identifier is transmitted to the child node and recorded as the parent node's disposal status. The disposal status identifier includes pending disposal status, disposal status, completed disposal status, and destroyed status.

[0030] Backpropagation is performed starting from the leaf node of the directed acyclic graph. When an incomplete disposal task is detected in the current child node, an incomplete disposal identifier is generated. The cumulative constraint strength value of the current child node is compared with the recommended retention period value to determine whether there is a constraint conflict. The disposal status of the parent node of the current child node is matched with the preset dependency conditions to determine whether the dependency conditions are met. The comparison result of the constraint conflict and the matching result of the dependency conditions are encapsulated into the blocking information.

[0031] The incomplete processing flag and the blocking information are passed to the parent node along the reverse edge of the directed acyclic graph. After receiving the incomplete processing flag, the parent node suspends its own processing operation and waits for the processing completion signal from the child node.

[0032] When a conflict or circular dependency in the retention period constraint between nodes is detected through the bidirectional information propagation, the directed acyclic graph is decomposed into multiple independently manageable subgraphs, including:

[0033] During the bidirectional information propagation process, when the difference between the cumulative constraint strength value of the child node and the constraint reference value transmitted by the parent node exceeds a preset conflict threshold, it is determined to be a storage period constraint conflict; when a duplicate node identifier is detected in the reverse edge propagation path of the node, it is determined to be a circular dependency.

[0034] Identify the node pairs in the directed acyclic graph that conflict with the retention period constraint and the set of nodes that have cyclic dependencies. Mark the directed edges between the node pairs as conflict edges and the directed edges in the set of nodes that form closed loops as dependent edges. Disconnect the conflict edges and the dependent edges to obtain multiple disconnected subgraph structures.

[0035] Connectivity detection is performed on each subgraph structure, and the node identifiers and corresponding directed edge information in each subgraph structure are reorganized into an independent graph data structure. An independent processing task queue is assigned to each subgraph structure, and the bidirectional information propagation and processing operation is executed independently on each subgraph structure to obtain multiple subgraphs that can be processed independently.

[0036] Based on the destruction trigger score and the recommended retention period value of each node in the subgraph, it is determined whether the destruction conditions are met. When the destruction conditions are met, the electronic file to be identified corresponding to the current node is migrated to the destruction database, including:

[0037] Traverse all nodes in each subgraph, compare the destruction trigger score of each node with the preset destruction score threshold, obtain the actual storage time of each node according to the current system time, and determine that the current node meets the destruction conditions when the destruction trigger score is greater than the destruction score threshold and the actual storage time exceeds the recommended storage period value.

[0038] For nodes that meet the destruction conditions, perform dependency checks, read the set of child nodes of the current node, query the disposal status identifier of all child nodes in the set of child nodes, and when the disposal status identifier of all child nodes is in the completed disposal status, determine that the current node has no incomplete dependencies and generate a destruction operation instruction.

[0039] According to the destruction operation instruction, the electronic file to be identified is read from the current storage location and packaged into a data packet to be migrated. A data transmission channel is established with the destruction library. Based on the data transmission channel, the data packet to be migrated is written into the designated storage area of ​​the destruction library. The disposal status identifier of the current node is updated to the destroyed status, and the operation of migrating the electronic file to be identified to the destruction library is completed.

[0040] A second aspect of this invention provides an intelligent authentication and automatic processing system for electronic archives based on semantic matching, comprising:

[0041] The rule acquisition unit is used to acquire the electronic files to be authenticated along with the corresponding authentication rules and retention period constraints.

[0042] The feature encoding unit is used to perform semantic decomposition on the identification rules to extract conditional predicates, perform multi-level feature encoding on the content of the electronic archive to be identified, and introduce the time decay factor of the archive formation time and business context to obtain the archive features of the electronic archive to be identified.

[0043] The semantic matching unit is used to calculate the semantic distance between the archive features and the conditional predicates in the feature space, and to fuse the matching results of different conditional predicates to generate a recommended retention period value and a destruction trigger score.

[0044] The graph construction unit is used to construct a directed acyclic graph containing the electronic file to be identified and the corresponding multi-level related files. It performs bidirectional information propagation based on message passing on the directed acyclic graph. During forward propagation, the retention period constraint and disposal status of the parent node are passed along the edge to the child node and the constraint strength is accumulated. During reverse propagation, the unfinished disposal mark and blocking information of the child node are sent back to the parent node.

[0045] The conflict detection unit is used to decompose the directed acyclic graph into multiple independently disposable subgraphs when a conflict or circular dependency in the retention period constraint between nodes is detected through the bidirectional information propagation. The unit determines whether the destruction conditions are met based on the destruction trigger score and the recommended retention period value of the node in each subgraph. When the destruction conditions are met, the electronic file to be identified corresponding to the current node is migrated to the destruction database.

[0046] A third aspect of the present invention provides an electronic device, comprising:

[0047] processor;

[0048] Memory used to store processor-executable instructions;

[0049] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0050] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0051] This method achieves intelligent authentication and disposal of electronic archives through semantic matching and graph structure analysis, significantly improving the automation level and decision-making accuracy of archive management. The authentication rules are semantically decomposed and conditional predicates are extracted. Simultaneously, multi-level feature encoding is performed on the content of the archives to be authenticated, and time-series decay factors and contextual enhancement vectors are introduced, making the expression of archive features more comprehensive and accurate. This effectively overcomes the limitations of traditional keyword matching, enabling a deeper understanding of the semantic relationship between archive content and authentication requirements, thereby greatly improving the applicability and matching accuracy of authentication rules.

[0052] By calculating the semantic distance between archival features and conditional predicates in the feature space and fusing multi-condition matching results, more reliable retention period recommendations and destruction trigger scores can be generated. This process quantifies the value of archives and the urgency of disposal, providing an objective and quantifiable basis for subsequent automated decision-making, reducing the subjectivity and inconsistency of human judgment, and ensuring the scientific nature of disposal recommendations.

[0053] A directed acyclic graph containing multi-level related archives was constructed, and a message-passing-based bidirectional information propagation mechanism was implemented on the graph to achieve systematic modeling and analysis of complex dependencies between archives. Forward propagation accumulates the constraint strength of parent nodes, while backward propagation sends back blocking information from child nodes, enabling the system to dynamically perceive the global processing status and constraint conflicts. This mechanism ensures that the contextual dependencies of related archives are fully considered when processing them, avoiding logical errors or compliance risks caused by isolated processing.

[0054] When a conflict in retention period constraints or a circular dependency is detected, the system can automatically decompose the directed acyclic graph into multiple independently manageable subgraphs. Based on this, a comprehensive judgment is made by combining the destruction trigger score of each node with the recommended retention period value, triggering archive destruction and migration operations only when the conditions are met. This strategy effectively solves the disposal deadlock problem in complex archive association networks, realizes the decomposition and parallelization of the disposal process, and significantly improves the efficiency and robustness of large-scale archive batch disposal while ensuring compliance. Attached Figure Description

[0055] Figure 1 This is a flowchart illustrating a semantic matching-based intelligent identification and automated processing method for electronic records.

[0056] Figure 2 This is a schematic diagram of the semantic distance calculation and score generation process. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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.

[0058] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0059] Figure 1 This is a flowchart illustrating the intelligent identification and automatic processing method for electronic archives based on semantic matching, as described in an embodiment of the present invention. Figure 1 As shown, the intelligent authentication and automatic processing method for electronic records based on semantic matching includes:

[0060] Obtain the electronic files to be authenticated, along with the corresponding authentication rules and retention period constraints;

[0061] The identification rules are semantically decomposed to extract conditional predicates. The content of the electronic archive to be identified is encoded with multi-level features and the time decay factor of the archive formation time and business context are introduced to obtain the archive features of the electronic archive to be identified.

[0062] The semantic distance between the archive features and the conditional predicates is calculated in the feature space, and the matching results of different conditional predicates are fused to generate a recommended retention period value and a destruction trigger score.

[0063] Construct a directed acyclic graph containing the electronic files to be identified and the corresponding multi-level related files. Perform bidirectional information propagation based on message passing on the directed acyclic graph. During forward propagation, the retention period constraint and disposal status of the parent node are passed along the edge to the child node and the constraint strength is accumulated. During reverse propagation, the unfinished disposal mark and blocking information of the child node are sent back to the parent node.

[0064] When a conflict or circular dependency in the retention period constraint between nodes is detected through the bidirectional information propagation, the directed acyclic graph is decomposed into multiple subgraphs that can be disposed of independently. The destruction trigger score and the recommended retention period value of each node in the subgraph are used to determine whether the destruction conditions are met. When the destruction conditions are met, the electronic file to be identified corresponding to the current node is migrated to the destruction database.

[0065] In one optional implementation, the identification rules are semantically decomposed to extract conditional predicates, and the content of the electronic archive to be identified is multi-level feature encoded, incorporating a temporal decay factor of the archive's formation time and business context, to obtain the archive features of the electronic archive to be identified, including:

[0066] Dependency parsing is performed on the text description in the identification rules to identify the subject-verb-object structure and modification relations in the identification rules. The predicate part in the subject-verb-object structure is used as the core semantic unit of the condition judgment, and the qualifiers and constraints in the modification relations are used as the boundary constraints of the condition judgment. The core semantic unit and the boundary constraints are extracted to form the structured representation of the condition predicate.

[0067] The structured fields and unstructured text of the electronic archive to be identified are encoded and feature extracted to obtain basic features. The interaction weight between the structured fields and the unstructured text is calculated through a self-attention mechanism. The basic features are then weighted and fused according to the interaction weight to obtain a fused feature representation.

[0068] The time difference between the archive creation time and the current time is calculated and mapped to an attenuation coefficient. The attenuation coefficient is inversely proportional to the time difference. The attenuation coefficient is multiplied element-wise with the fused feature representation to obtain the time-attenuated feature.

[0069] The identification information of the preceding and subsequent business nodes is extracted from the business process to which the electronic file to be identified belongs and encoded into a context vector. The context vector is then concatenated with the time-decayed features to obtain the file features.

[0070] The process of generating archival features for electronic archives to be authenticated requires deep semantic analysis of the authentication rules. These rules are typically stored in the rule base of the archive management system in natural language form, such as "Contract documents involving major projects should be kept for 30 years" or "Routine office notices should be kept for 5 years after archiving." These rule texts contain complex semantic structures and need to be decomposed into computable semantic units using dependency parsing techniques. The dependency parser performs part-of-speech tagging and dependency relation annotation on the rule text, identifying the subject, predicate, object, and the dependency arcs between them. In the rule "Contract documents involving major projects should be kept for 30 years," dependency analysis identifies the subject as "contract documents," the predicate as "keep," and the object as "30 years." The modifier "involves major projects" is connected to the subject via a dependency arc. The predicate "keep" is used as the core semantic unit of the conditional judgment, carrying the core action semantics of the rule. The modifiers "involves major projects" and "contracts" serve as boundary constraints, defining the specific scope of the rule's application. These boundary constraints typically include limitations such as file type, business domain, and classification level. After extracting the core semantic units and boundary constraints, the conditional predicates are represented in triple form, with the structure (file type constraint, judgment action, retention period value), for example (contract documents AND involving major projects, should be retained, 30 years). For complex rules containing logical connectives such as "AND" or "unless", a logical combination tree of conditional predicates needs to be constructed to decompose complex conditions into logical operations of multiple basic conditional predicates.

[0071] The extraction of content features from electronic archives to be authenticated requires processing both structured fields and unstructured text. Structured fields include metadata fields such as archive title, document number, responsible party, security classification, and archive classification code. These fields are typically stored in the database tables of the archive management system. Encoding of structured fields employs a domain-specific embedding method, mapping each field value to a fixed-dimensional vector space. The archive classification code, as a hierarchical structure, uses path encoding, sequentially encoding each level of the classification code and then performing pooling to obtain the classification feature vector. If the responsible party field involves an organization, it is encoded using the hierarchical relationship of the organizational tree to obtain the organization feature vector. The security classification field uses sequential encoding, mapping levels such as "public," "internal," "secret," and "confidential" to incrementally increasing numerical codes. Unstructured text includes the main body of the archive, attachments, annotations, etc., which are encoded using a pre-trained language model. The unstructured text is input into an encoder based on a Transformer architecture. A multi-layer self-attention mechanism is used to obtain a contextualized representation of each text fragment. Average pooling or max pooling is then performed on the representation vectors of all text fragments to obtain the text feature vector. The feature vectors obtained from structured field encoding and the feature vectors obtained from unstructured text encoding constitute the basic feature set.

[0072] Implicit semantic relationships exist among basic features. For example, keywords in the document title are highly relevant to the theme of the main text, and there is a mapping relationship between the organization to which the responsible person belongs and the document classification code. These cross-modal feature interactions are captured through a self-attention mechanism. The feature vector sequence of structured fields is concatenated with the feature vector sequence of unstructured text to form a unified feature sequence, which is then input into the self-attention layer. The self-attention layer calculates the attention weight between each position in the sequence and all other positions, and the weight values ​​are... Indicates the first The feature position for the first The degree of attention given to each feature location. Attention weights are determined by the query vector. Key vector Sum value vector The calculation is as follows: ,in This represents the dimension of the key vector. After obtaining the attention weight matrix, the basic features are weighted and summed to fuse the feature representations. The Each position is calculated as This operation allows the representation of each feature location to incorporate information from other related feature locations. The multi-head self-attention mechanism captures feature interaction patterns in different subspaces by computing multiple sets of different query, key, and value matrices in parallel, and finally concatenates the outputs of the multiple heads to obtain a fused feature representation.

[0073] The impact of the time of archival creation on archival value exhibits a non-linear decay characteristic over time. Archival value is higher in its early stages, gradually decreasing over time while its historical value becomes more prominent. A time-decay factor is used to quantify this time-related impact. The time difference between the archival creation time and the current appraisal time is calculated. The unit is years. The time difference is mapped to an attenuation coefficient. Using an exponential decay function ,in This is the decay rate overparameter, set based on experience values ​​derived from the file type and business area. For important files requiring long-term preservation... Taking a smaller value results in a slower decay rate, which is beneficial for short-term transactional files. A larger value results in a faster attenuation rate. Attenuation coefficient The value of is between 0 and 1; the larger the time difference, the smaller the attenuation coefficient. The attenuation coefficient is then multiplied element-wise with the fused feature representation. ,in This represents element-wise multiplication, which applies the same decay weight to each dimension of the feature vector, thus numerically weakening the features of older archives.

[0074] Business context information reflects the position and relationship of an archive within a business process, and is of significant reference value for judging the archive's value. Electronic archives to be appraised typically belong to specific business processes, such as project approval processes, contract signing processes, and financial reimbursement processes. The process node identifier of the archive is retrieved from the business process table of the archive management system to obtain a list of its preceding and following nodes. Preceding nodes represent business steps completed before the archive's creation; for example, preceding nodes in a project initiation document include the project proposal and feasibility study report. Following nodes represent subsequent business steps after the archive's creation; for example, following nodes in a project initiation document include the project implementation plan and project acceptance report. The identifier information for preceding and following nodes includes node number, node name, and node type, which are mapped to fixed-dimensional vectors through an embedding layer. The preceding context vector is obtained by average pooling the vector sequence of preceding nodes. The subsequent context vector is obtained by using average pooling to transform the vector sequence of subsequent nodes. Context vector It is obtained by concatenating the preceding context vector and the following context vector, i.e. This context vector captures the location information and dependencies of the file in the business process. For files located at key nodes in the process, the context vector exhibits stronger semantic aggregation characteristics.

[0075] The time-decayed feature vector With business context vector By concatenating the features along the feature dimension, the archival features are obtained. The concatenation operation maintains the independence of features from different sources while integrating them into a unified feature representation space. The dimension of the archival features is equal to the sum of the dimension of the time decay features and the dimension of the context vector. This archival feature vector comprehensively integrates the content semantics, structural attributes, temporal characteristics, and business process location information of electronic archives, providing a rich feature representation foundation for subsequent semantic matching with conditional predicates. For batches of archives to be identified, the above feature extraction process can be parallelized, accelerating feature generation efficiency through batch encoding and matrix operations.

[0076] In one optional implementation, the semantic distance between the archive features and the conditional predicates is calculated in the feature space, and the matching results of different conditional predicates are fused to generate a recommended retention period value and a destruction trigger score, including:

[0077] The structured representations of the archive features and the conditional predicates are mapped to a unified semantic feature space, and the distance metric between the vector representations of the archive features and the vector representations of the structured representations of the conditional predicates is calculated in the semantic feature space as the semantic distance.

[0078] The semantic distance between the file features and the multiple conditional predicates contained in the identification rule is calculated respectively to obtain multiple semantic distance values. The weight coefficient of each conditional predicate is determined according to the logical relationship of each conditional predicate in the identification rule. The multiple semantic distance values ​​and the corresponding weight coefficients are weighted and summed to obtain the comprehensive matching score.

[0079] The retention period labeling information associated with each conditional predicate is extracted from the identification rules. The retention period labeling information is weighted and averaged according to the semantic distance between each conditional predicate and the archival features to obtain the recommended retention period value. The weight of the retention period labeling information corresponding to each conditional predicate is inversely proportional to the semantic distance.

[0080] The recommended retention period is compared with the actual retention time of the electronic file to be identified. When the actual retention time exceeds the recommended retention period, the excess ratio is calculated. The excess ratio is weighted and combined with the comprehensive matching score to obtain the destruction trigger score.

[0081] like Figure 2 As shown, the method includes:

[0082] To calculate the semantic distance between archival features and conditional predicates, both need to be mapped to the same feature space. The vector representation of archival features has already been obtained through multi-level feature encoding in the previous steps, and its dimension is typically a dense vector of 512 or 768 dimensions. The structured representation of the conditional predicate needs to be vectorized through semantic parsing. For the conditional predicate "document creation date earlier than 2010 and involves completed projects", it is split into two sub-predicates: time condition and state condition. The corresponding time feature encoder converts "earlier than 2010" into a time anchor vector, and the state feature encoder converts "completed projects" into a state attribute vector. The two sub-vectors are then fused into a unified representation vector for the conditional predicate through an attention mechanism. During the mapping process, a domain-pretrained archival terminology dictionary is introduced to ensure the semantic accuracy of professional terms such as "completed" and "retention period".

[0083] In a unified semantic feature space, the distance metric between the archival feature vector and the conditional predicate vector is calculated using a combination of cosine distance and Euclidean distance. Cosine distance reflects the directional difference between the two vectors; it is calculated by subtracting the cosine similarity of the two vectors from 1, with a value ranging from 0 to 2. A smaller value indicates closer directional similarity. Euclidean distance reflects the absolute positional difference between the vectors in the feature space, obtained by taking the square root of the sum of the squares of the differences in each dimension of the vectors. Since the two distances have different dimensions, they need to be normalized separately: the cosine distance is divided by 2, and the Euclidean distance is divided by the maximum possible distance in the feature space, normalizing both to the range of 0 to 1. The final semantic distance is calculated by adding 0.6 times the normalized cosine distance to 0.4 times the normalized Euclidean distance. This weighting configuration has been experimentally verified to effectively balance directional similarity and absolute positional difference in archival authentication scenarios.

[0084] When an authentication rule contains multiple conditional predicates, the semantic distance between the document features and each conditional predicate is calculated separately. For example, if an authentication rule contains five conditional predicates, relating to document creation time, business type, security classification, associated project status, and responsible department attribute, then five semantic distance calculations are required, yielding five distance values. These distance values ​​have different meanings: a smaller distance value for the time condition indicates a high degree of match between the document creation time and the time range required by the rule; a smaller distance value for the business type indicates that the document's business attributes are close to the type specified by the rule; and the distance value for the security classification reflects whether the document's security level meets the rule's requirements. The importance of each conditional predicate in the authentication rule varies, and its weight coefficient needs to be determined based on the rule's logical structure. For conditional predicates connected by "AND" logic, the failure to meet any condition will affect the overall matching, thus the weights of each condition are relatively balanced; for conditional predicates connected by "OR" logic, meeting any one condition is sufficient, resulting in a more dispersed weight distribution. By parsing the logical expression tree of the rule, the weight coefficient of each conditional predicate is automatically calculated, ensuring that the sum of all weights is 1.

[0085] The semantic distance values ​​of each conditional predicate are weighted and summed with their corresponding weight coefficients to obtain the overall matching score. During the weighted summation, a smaller distance value indicates a higher matching degree; therefore, the distance values ​​need to be converted into similarity values. The conversion formula is 1 minus the normalized semantic distance value. The converted similarity value ranges from 0 to 1, with a larger value indicating a higher matching degree. The overall matching score is obtained by multiplying the similarity value of each conditional predicate by its corresponding weight coefficient and summing the results. This score reflects the overall degree of conformity between the electronic file to be identified and the entire identification rule. A score close to 1 indicates that the file fully meets the rule requirements, while a score close to 0 indicates that the file seriously violates the rule requirements. In actual calculations, a satisfaction threshold for conditional predicates is also introduced. When the similarity value of a conditional predicate is lower than a preset threshold of 0.3, the condition is considered completely unmet. In this case, regardless of the matching degree of other conditions, the overall matching score will be forcibly reduced to below 0.5. This mechanism ensures strict matching of key conditions.

[0086] When generating recommended retention periods, retention period annotations associated with each conditional predicate are extracted from the identification rules. Each conditional predicate typically includes a suggested retention period value in the rule definition; for example, the retention period annotation for the condition "document creation date earlier than 2010" is 10 years, and the retention period annotation for the condition "involves completed projects" is 5 years. After extracting the retention period annotations for all conditional predicates, a weighted average is calculated based on the semantic distance between each conditional predicate and the archive features. The weighting method uses an inverse distance weighting, meaning that the smaller the semantic distance of a conditional predicate, the greater its weight in the average calculation. Specifically, the weight of each conditional predicate is equal to 1 divided by the semantic distance value of that condition plus a small constant of 0.01. This small constant is used to avoid division by zero when the distance is 0. After calculating the inverse distance weights for all conditional predicates, these weights are normalized so that their sum is 1. Then, the retention period annotation value of each conditional predicate is multiplied by its corresponding normalized weight and summed to obtain the recommended retention period value. This weighted averaging method ensures that the conditional predicate that best matches the archival characteristics has the greatest impact on the recommended retention period value, thereby improving the accuracy of the recommendation results.

[0087] The recommended retention period is compared with the actual retention time of the electronic records to be appraised. The actual retention time is obtained by subtracting the record creation time from the current system time, in years. When the actual retention time exceeds the recommended retention period, the excess percentage is calculated. The excess percentage is equal to the difference between the actual retention time and the recommended retention period, divided by the recommended retention period. For example, if the recommended retention period is 10 years and the record has actually been retained for 12 years, the excess percentage is 0.2, indicating that the record has been retained for 20% longer than recommended. The excess percentage is an important component of the destruction trigger score. The larger the excess percentage, the more the record has been retained beyond the recommended period, and the higher the urgency of destruction. When weighting the excess percentage with the comprehensive matching score, a weighting of 0.7 times the excess percentage plus 0.3 times the comprehensive matching score is used. This weighting reflects the dominant role of the time factor in the destruction decision; even if the record does not match the appraisal rules very well, as long as the retention time is significantly exceeded, the destruction trigger score will still be high. During the calculation process, an upper limit of 2.0 is set for the excess ratio, meaning that the maximum excess ratio is calculated as 200%, to avoid excessive distortion in the scoring of extremely overdue files.

[0088] The final value of the destruction trigger score ranges from 0 to 1, with a higher score indicating that the archive should be included in the destruction process. When the score is higher than 0.75, the archive is marked as a priority for destruction; when the score is between 0.5 and 0.75, the archive is marked as an archive awaiting review and requires manual verification; when the score is lower than 0.5, the archive is temporarily excluded from the destruction scope. For archives whose actual retention time has not exceeded the recommended retention period, the excess percentage is negative. In this case, the destruction trigger score is mainly determined by the comprehensive matching score, but the overall score usually does not exceed 0.4, ensuring that archives that have not yet expired are not mistakenly judged as eligible for destruction. In the weighted combination process, an archive importance correction coefficient is also introduced. For documents marked as important or involving significant matters, the destruction trigger score is multiplied by a correction coefficient of 0.8 to reduce the likelihood of them being destroyed. This mechanism reflects the principle of prudence in archive management.

[0089] To improve the accuracy of the recommended storage period, the historical verification accuracy of the conditional predicates is considered during the weighted average calculation. The performance of each conditional predicate in historical cases is recorded, and the deviation between the recommended storage period and the actual reasonable storage period is statistically analyzed to calculate the credibility coefficient of the conditional predicate. Conditional predicates with high credibility coefficients receive additional weight in the weighted average, while those with low credibility coefficients are weighted less. This dynamic adjustment mechanism allows the recommended storage period value to be continuously optimized as the system runs, gradually improving the recommendation accuracy. Simultaneously, for newly added conditional predicates, the initial credibility coefficient is set to a moderate level of 0.5 to avoid over-reliance on or complete disregard for the role of new conditions due to a lack of historical data.

[0090] In one optional implementation, constructing a directed acyclic graph including the electronic file to be identified and its corresponding multi-level associated files includes:

[0091] Extract a set of identifiers for associated files from the metadata and business attributes of the electronic file to be identified, parse the association type attribute between each associated file in the identifier set and the electronic file to be identified, and determine a business dependency relationship when the association type attribute is identified as a business process prerequisite, and a reference relationship when the association type attribute is identified as a content reference source.

[0092] The hierarchical structure between files is determined based on the business dependency relationship and the reference relationship. The electronic file to be identified is taken as the target node, the files that the electronic file to be identified depends on and the files that are referenced by the electronic file to be identified are taken as parent nodes, and the files that depend on the electronic file to be identified and the files that reference the electronic file to be identified are taken as child nodes. A directed edge of dependency is established from the dependent file node to the dependent file node based on the business dependency relationship, and a directed edge of reference is established from the referenced file node to the referencing file node based on the reference relationship.

[0093] The association strength value is calculated and normalized based on the access frequency and time span between the associated archive and the electronic archive to be identified, and used as the weight of the corresponding directed edge to obtain the directed acyclic graph.

[0094] The process iterates through the business attribute table of the electronic files to be authenticated, reading the marked preceding business process fields. These fields are typically stored in JSON format and contain a globally unique identifier for the preceding file, a business type code, and a creation timestamp. By parsing this JSON structure, all preceding file identifiers are obtained and organized into a set. Simultaneously, the system scans the reference markers in the main text of the archives and the metadata of the attachments. These markers may exist in the form of hyperlinks or specific tags. The system extracts the identifiers of all referenced archives to form a set. .Will and After merging and deduplication, a complete set of associated file identifiers is obtained. .

[0095] For sets For each identifier, the system accesses the archival relationship database and reads the association type attribute field between the corresponding archival document and the electronic archival document to be authenticated. This field is stored in the database as an enumeration type, including various types such as business prerequisite, content reference, attachment association, and version evolution. When the association type attribute value is "business prerequisite" or "process dependency," the association is determined to be a business dependency relationship, indicating that the formation of the electronic archival document to be authenticated must be based on the business results of the associated archival document. When the association type attribute value is "content reference" or "data source," the association is determined to be a reference relationship, indicating that some content of the electronic archival document to be authenticated is directly extracted or referenced from the associated archival document. Through this determination mechanism, all association relationships are divided into a business dependency set. With reference set .

[0096] The electronic file to be identified is used as the initial target node. Assign it a hierarchy number of 0. Traverse the business dependency set. For each relationship record in the graph, extract the identifier of the dependent file from the relationship record and create a corresponding node in the graph structure. If the dependent file also has its own business predecessor file, recursively trace upwards until a root node that no longer depends on any other file is found. Based on the depth of the dependency chain, assign a hierarchy number to each dependent file node; a larger hierarchy number indicates a greater distance from the target node. Similarly, traverse the set of reference relationships. A node is created for the referenced file and assigned a hierarchy number. For files that depend on or reference the target node, a corresponding node is also created, but its hierarchy number is negative, indicating that it is downstream of the target node.

[0097] For business dependencies, starting from the dependent file node, create a directed edge pointing to the dependent file node. This directed edge carries the edge type label "dependency" and is recorded in the edge set of the graph. In the process of creating a graph, for reference relationships, a directed edge is created pointing to the referencing archive node, with the edge type labeled "reference". During construction, loops are detected in real-time. Specifically, after each new edge is added, a depth-first search is performed starting from the endpoint of the edge to check if it is possible to backtrack to the origin of the edge. If a loop is detected, the edge is rejected, a warning message is recorded, and the relationship is marked as "weak," remaining only in the metadata and not included in the graph structure.

[0098] Retrieve joint access records between the related archive and the electronic archive to be authenticated from the archive access log database. Count the number of times both were accessed in the same business session within the past year, denoted as [missing information]. At the same time, the difference in the creation time of the two archives is calculated and denoted as . The unit is days. The initial value of the association strength is calculated using the formula... This formula reflects the characteristic that higher access frequency and shorter time span indicate a stronger association. Considering the significant differences in baseline access frequency across different business scenarios, the initial calculated values ​​need to be normalized. The initial calculated values ​​of all associated edges in the current batch are statistically analyzed, and the maximum value is found. and minimum value The min-max normalization method is used to convert the initial calculated value of each edge into a normalized weight. The normalized weight is limited to a value between 0 and 1. The normalized weight is stored as a weight attribute of the corresponding directed edge in the edge's attribute dictionary.

[0099] After the graph is constructed, perform a topological sort to verify its directed acyclic property. Starting with all nodes with an in-degree of zero, add them to the queue, dequeue them one by one, delete all their outgoing edges, and decrement the in-degree of the nodes pointed to by the outgoing edges. If the in-degree of a node becomes zero, add it to the queue. Repeat this process until the queue is empty. If there are still unvisited nodes at this point, it indicates the presence of a cycle in the graph. It is necessary to backtrack and investigate the cause of the cycle, which is usually due to incorrect business process configuration or incorrect metadata recording. For detected cycles, analyze the weights of each edge in the cycle, select the edge with the smallest weight to break it, downgrade the edge's type from forced dependency to reference association, and record this downgrade operation in the graph's additional information for subsequent manual review.

[0100] To improve graph construction efficiency, a layered construction strategy is adopted for complex business scenarios involving a large number of related archives. Only the first-level related nodes and edges directly adjacent to the target node are constructed, forming a local subgraph. When executing subsequent message passing, if incomplete constraint information of a boundary node is detected, the related subgraph of that node is dynamically expanded, gradually loading its parent or child nodes. This on-demand loading mechanism avoids the memory overhead of loading the entire related network at once, while ensuring that constraint information on the critical path can be obtained in a timely manner. For cross-archive associations, a cross-archive query connection is established through the archive identifier prefix in the archive identifier, retrieving the metadata summary of the related archives from the remote archive, rather than the complete content, to reduce network transmission overhead.

[0101] In the graph data structure implementation, an adjacency list is used for storage. Two lists are maintained for each node: one recording all incoming edges and their starting nodes, and the other recording all outgoing edges and their ending nodes. Each edge, in addition to storing its starting point, ending point, and weight, also records its type, formation time, and business context label. Node attributes, besides the file identifier, also store the recommended retention period, destruction trigger score, current disposal status, and the cumulative constraint strength propagated from the parent node. To support fast querying, a hash mapping from node identifiers to node objects is established, achieving node location with constant time complexity. The edge weight information is used to adjust the attenuation coefficient of information propagation during subsequent message passing; the higher the edge weight, the slower the decay of the propagated constraint information.

[0102] For archives with multiple version evolution relationships, different versions are treated as independent nodes, but version chain information is recorded in the node attributes. The latest version node in the version chain undertakes the main constraint propagation task, and the disposal decisions of historical version nodes are affected by the disposal status of the latest version node. Specifically, when establishing the version chain, a version evolution edge is created from the old version node pointing to the new version node. The weight of this edge is fixed at 1, representing a strong constraint relationship. In subsequent disposal processes, the destruction operation can only be performed uniformly when all nodes in the version chain meet the destruction conditions, avoiding the inconsistent state where some versions are destroyed while others are retained.

[0103] In one optional implementation, bidirectional message-passing-based information propagation is performed on the directed acyclic graph. During forward propagation, the retention period constraint and disposal status of the parent node are passed along the edges to the child nodes, and the constraint strength is accumulated. During backward propagation, the incomplete disposal flag and blocking information of the child node are sent back to the parent node, including:

[0104] Starting from the root node of the directed acyclic graph, forward propagation is performed. The retention period constraint carried by the current parent node is read as the constraint base value. The constraint base value is multiplied by the weight of the edge between the current parent node and the child node to obtain the transit constraint value.

[0105] The cumulative constraint strength value of the child node is obtained by summing the transmitted constraint value with the historical constraint value received by the child node. At the same time, the current parent node's disposal status identifier is transmitted to the child node and recorded as the parent node's disposal status. The disposal status identifier includes pending disposal status, disposal status, completed disposal status, and destroyed status.

[0106] Backpropagation is performed starting from the leaf node of the directed acyclic graph. When an incomplete disposal task is detected in the current child node, an incomplete disposal identifier is generated. The cumulative constraint strength value of the current child node is compared with the recommended retention period value to determine whether there is a constraint conflict. The disposal status of the parent node of the current child node is matched with the preset dependency conditions to determine whether the dependency conditions are met. The comparison result of the constraint conflict and the matching result of the dependency conditions are encapsulated into the blocking information.

[0107] The incomplete processing flag and the blocking information are passed to the parent node along the reverse edge of the directed acyclic graph. After receiving the incomplete processing flag, the parent node suspends its own processing operation and waits for the processing completion signal from the child node.

[0108] After constructing a directed acyclic graph containing the electronic archives to be authenticated and their multi-level related archives, a two-way information propagation mechanism is needed to transmit constraints and detect conflicts between the archives. This mechanism consists of two phases: forward propagation and backward propagation, and performs distributed computation on the graph structure through message passing.

[0109] Forward propagation begins at the root node of the directed acyclic graph (DAG). The root node typically represents the earliest created file or the highest-level file in the business process, and its retention period constraint serves as the benchmark for the entire associated file system. In each propagation step, the retention period constraint carried by the current parent node is read. This constraint, expressed in years, represents the minimum duration the file must be retained. For example, if the retention period constraint of a project initiation file is permanent, the read constraint benchmark value is recorded as positive infinity; if it's a 30-year retention period, the constraint benchmark value is 30. After obtaining the constraint benchmark value, the constraint attenuation or enhancement during the propagation process needs to be calculated based on the weights of the edges between the parent and child nodes. The edge weights reflect the strength of the relationship between the two files, with weights ranging from 0.5 to 1.5. When a child node file has a strong dependency on its parent node file, the edge weight is close to 1.5, indicating that the constraint needs to be enhanced during propagation; when the dependency is weak, the edge weight is close to 0.5, indicating that the constraint can be appropriately relaxed. The propagated constraint value is obtained by multiplying the constraint benchmark value by the edge weight. Taking the aforementioned project initiation file as an example, if the edge weight between it and the project implementation plan file is 1.2, then the transitive constraint value is 30 multiplied by 1.2, which equals 36. This means that the implementation plan file is subject to stricter retention period constraints.

[0110] When a child node receives transmitted constraint values, it needs to consider that it may receive different constraints from multiple parent nodes. An accumulation strategy is used to handle this multi-source constraint situation. The currently received transmitted constraint value is summed with the historical constraint values ​​already accumulated by the child node to obtain the child node's cumulative constraint strength value. This accumulation mechanism reflects the comprehensive constraint pressure that an archive experiences within the associated system. For example, if a meeting minutes archive is associated with three different project archives, receiving transmitted constraint values ​​of 25, 30, and 28 respectively, its cumulative constraint strength value is 83. Although this accumulation method may cause the constraint value to exceed the actual retention period of a single archive, it can quantitatively reflect the importance of the archive in the associated network. In actual judgment, the cumulative constraint strength value is compared with a preset threshold, rather than being directly used as the retention period.

[0111] In addition to passing constraint values, forward propagation also requires the synchronous transmission of the parent node's disposal status identifier. The disposal status identifier is divided into four types: pending disposal (meaning the file has not yet started any disposal process); in-process disposal (meaning the file is undergoing identification, approval, or technical processing); completed disposal (meaning the file has been identified and its final disposal method has been determined, but destruction has not yet been performed); and destroyed (meaning the file has been physically deleted from the system). During forward propagation, the parent node passes its current disposal status identifier along the edges to the child nodes, and the child nodes record the received identifier in a dedicated parent node disposal status field. Since a child node can have multiple parent nodes, this field is designed as a status set structure, capable of storing the disposal status from all parent nodes. This status transmission mechanism provides a data foundation for subsequent determination of whether dependencies between files are satisfied.

[0112] Forward propagation employs a breadth-first search strategy, maintaining a first-in-first-out (FIFO) queue structure. Initially, all root nodes are added to the queue. Each time, a node is retrieved from the head of the queue as the current parent node, its retention period constraint and disposal status are read, and then all outgoing edges of that node are traversed. For each outgoing edge, the target child node and edge weight are obtained, and the aforementioned constraint value calculation, accumulation, and state propagation operations are performed. After processing all child nodes of the current node, these child nodes are added to the tail of the queue in sequence. This process is repeated until the queue is empty, at which point all reachable nodes have completed forward propagation. During the traversal, the number of visits to each node is recorded. Since there are no cycles in a directed acyclic graph, the number of visits to each node should equal its in-degree, i.e., the number of parent nodes. If the number of visits to a node exceeds its in-degree, it indicates a potential cycle problem in the graph structure, requiring the triggering of an exception handling mechanism.

[0113] Backpropagation begins at a leaf node in the directed acyclic graph (DAG). A leaf node is a node with a degree of zero, representing the final or independent file of a business process. The primary task of backpropagation is to detect whether there are any incomplete processing tasks at the current child node. Specific detection methods include: querying the processing status identifier of the file corresponding to the node; if it is in a pending or in-processing state, an incomplete processing task is determined; checking whether the file contains any pending approval conclusions or pending technical processing tasks; and verifying whether the file's metadata integrity meets the processing requirements. When any condition is not met, an incomplete processing identifier is generated for the current child node. This identifier contains a specific blocking reason code, such as 101 for incomplete approval and 102 for missing metadata.

[0114] Constraint conflict determination is performed by comparing the cumulative constraint strength value of the child nodes with the recommended retention period value. The recommended retention period value is calculated through the feature space during the aforementioned semantic matching stage and represents a reasonable retention period inferred based on the characteristics of the archive content. The average constraint value is obtained by dividing the cumulative constraint strength value by the in-degree of the child nodes, and the difference is calculated with the recommended retention period value. If the absolute value of the difference exceeds a preset conflict threshold, such as 5 years, a constraint conflict is determined to exist. Conflicts are divided into two categories: when the average constraint value is significantly greater than the recommended value, it is an over-constraint conflict, leading to the unnecessary long-term retention of archives; when the recommended value is significantly greater than the average constraint value, it is an under-constraint conflict, leading to the premature destruction of important archives. The comparison result is encoded as a conflict type code and a conflict strength value.

[0115] Dependency matching determines whether the handling operation of a child node is constrained by the state of its parent node. Preset dependency conditions include several rules: mandatory dependency rules require the parent node to be in a completed or destroyed state before the child node can initiate handling; suggested dependency rules recommend but do not mandate that the parent node complete handling; and time-series dependency rules require the parent node's handling time to be earlier than the child node's. During matching, the set of parent node handling states recorded by the child node is traversed, and each dependency rule is compared one by one. If a parent node's handling state does not meet the mandatory dependency rule, the dependency condition is determined to be unmet. The matching result is encoded as a dependency satisfaction score, ranging from 0 to 100, where 100 indicates all dependency conditions are fully met, and 0 indicates severe dependency blocking.

[0116] Blocking information is generated by encapsulating the comparison results of constraint conflicts and the matching results of dependency conditions. This information adopts a structured data format and includes fields such as conflict type code, conflict intensity value, dependency satisfaction score, and blocking reason code. For nodes with multiple conflicts, the blocking information will list all conflict items and their priority order so that the parent node can accurately identify the most critical blocking factor.

[0117] The incomplete handling flag and blocking information are passed to the parent node via reverse edges. A reverse edge refers to the edge direction in the directed acyclic graph, i.e., a virtual path from the child node to the parent node. At the implementation level, this is achieved by maintaining a list of parent nodes for each node to perform reverse traversal. The incomplete handling flag and blocking information are sent as the message body along each parent node in the list. Upon receiving the incomplete handling flag, the parent node immediately updates its handling scheduling state, suspending any planned destruction or migration operations. Simultaneously, a listening task waiting for the child node to complete its task is inserted into the parent node's pending task queue. This task continuously monitors changes in the child node's handling state. When the child node completes its handling and sends a handling completion signal, the parent node's listening task captures this signal, releases the suspension, reassesses its handling conditions, and decides whether to continue executing the handling operation.

[0118] Backpropagation also employs a breadth-first traversal strategy, but the traversal order is reversed compared to forward propagation. Initially, all leaf nodes are added to a queue, and each time a child node is processed, its parent node is added to the queue. To avoid duplicate processing, a set of visited nodes is maintained. A node is marked as fully processed and continues propagation upwards only after receiving reverse messages from all its child nodes. The entire backpropagation process ensures that inter-archive dependency and constraint conflict information is accurately transmitted from bottom to top, providing complete diagnostic data for subsequent conflict resolution and subgraph decomposition.

[0119] In an optional implementation, when a conflict or circular dependency in the retention period constraint between nodes is detected through the bidirectional information propagation, the directed acyclic graph is decomposed into multiple independently manageable subgraphs, including:

[0120] During the bidirectional information propagation process, when the difference between the cumulative constraint strength value of the child node and the constraint reference value transmitted by the parent node exceeds a preset conflict threshold, it is determined to be a storage period constraint conflict; when a duplicate node identifier is detected in the reverse edge propagation path of the node, it is determined to be a circular dependency.

[0121] Identify the node pairs in the directed acyclic graph that conflict with the retention period constraint and the set of nodes that have cyclic dependencies. Mark the directed edges between the node pairs as conflict edges and the directed edges in the set of nodes that form closed loops as dependent edges. Disconnect the conflict edges and the dependent edges to obtain multiple disconnected subgraph structures.

[0122] Connectivity detection is performed on each subgraph structure, and the node identifiers and corresponding directed edge information in each subgraph structure are reorganized into an independent graph data structure. An independent processing task queue is assigned to each subgraph structure, and the bidirectional information propagation and processing operation is executed independently on each subgraph structure to obtain multiple subgraphs that can be processed independently.

[0123] During bidirectional information propagation, it is necessary to continuously monitor the constraint information transmitted between nodes and the topological characteristics of the propagation path. Regarding the detection of retention period constraint conflicts, during the forward propagation phase, when the parent node's constraint baseline value... When passing the value to child nodes, the child nodes will combine this value with their own accumulated constraint strength values. Compare them. Specifically, calculate the difference. ,when At that time, among them If a pre-defined conflict threshold is met, a conflict in retention period constraints is determined between the parent and child node pairs. Such conflicts typically arise from conflicting retention period requirements set by different business rules for the same document. For example, a financial document might be required to be retained for thirty years under accounting regulations, but related project documents might be marked with a ten-year retention period; the constraints cannot be satisfied simultaneously. Conflict Threshold The setting needs to take into account the actual business tolerance. It can generally be set to 1 / 10 of the retention period in years, that is, when the difference exceeds three years, conflict identification is triggered.

[0124] To detect circular dependencies, during the backpropagation phase, it's necessary to record the identifiers of all nodes traversed along each message delivery path. Specifically, this is achieved by appending a path tracking list to each backpropagation message. Initially, this list only contains the identifier of the message's originating node. When a message originates from a node... Pass to its parent node First check Does the identifier already exist? If a duplicate is detected, i.e. This indicates that from After departing and passing through several nodes, it returned to This creates a closed-loop structure, which is then identified as a circular dependency. In document management, this circular dependency arises from cross-referencing of business processes. For example, a contract file might reference an approval file, which in turn might reference a contract attachment, forming a circular dependency chain. To prevent infinite loop propagation, once a circular dependency is detected, further propagation of this path is immediately terminated, and the identifiers of all nodes forming the closed loop are recorded.

[0125] After detecting conflicts and dependencies, the problematic nodes and edges in the directed acyclic graph need to be marked. For identified retention period constraint conflicts, the specific node pair where the conflict occurred needs to be located. ,in As the parent node, These are child nodes. The directed edge connecting these two nodes... Mark the edges as conflict edges and record the specific numerical information of the conflict, including the constraint baseline value of the parent node. Cumulative constraint strength value of child nodes and the calculated difference This information is of significant reference value for subsequent manual review or automated adjustment strategies. For handling circular dependencies, the complete closed-loop path is reconstructed from the path tracing list. Assuming the detected closed loop is... Then it is necessary to identify all the edges that constitute the closed loop. These edges are then marked as dependent edges. In practice, multiple independent closed loops exist, which need to be identified and marked separately.

[0126] After marking the edges, perform edge breaking operations to eliminate conflicts and dependencies. When breaking conflicting edges, directly remove the corresponding directed edge structure in the graph, releasing the constraint transitive relationship between parent and child nodes. When breaking dependent edges, the breaking position needs to be selected according to a certain priority strategy. A common strategy is to select the edge with the smallest edge weight in the closed loop for breaking. The edge weight can be defined based on the strength of the association, business importance, or the order in which the files were created. For example, the edge weight can be calculated... The weight is ,in Represents a node and The strength of the correlation between them Represents a node The corresponding time of the creation of the archives, For the current time, and represents the weighting coefficient. Disconnecting the edge with the smallest weight minimizes disruption to the overall file's relational structure.

[0127] After the edge-breaking operation, the originally connected directed acyclic graph will be decomposed into several disconnected subgraph structures. All connected components can be quickly identified using depth-first search or breadth-first search algorithms. Starting from any unvisited node, traverse all reachable nodes and group these nodes and their edges into a subgraph. Repeat this process until all nodes are assigned to a subgraph. Nodes within each subgraph retain their original parent-child relationships and constraint propagation paths, but different subgraphs are completely independent, with no cross-subgraph edge connections.

[0128] For each subgraph, a connectivity test is performed. This test verifies whether any two nodes within the subgraph are reachable. If two nodes can still be connected by a path after ignoring the edge direction, they are considered to belong to the same connected component. Simultaneously, the test checks for cycles that violate the properties of directed acyclic graphs. If a cycle is detected, it indicates that the edge-breaking operation has not completely eliminated circular dependencies, and edge-breaking needs to be performed again. Only subgraphs that pass the connectivity test can proceed to the subsequent reassembly stage.

[0129] During the reorganization phase, an independent graph data structure is constructed for each subgraph, and the identifiers of all nodes in the subgraph are extracted to form a node set. Extract the start and end points and related attribute information of all directed edges in the subgraph to form an edge set. .based on and Reconstruct graph objects The graph object contains complete topological structure information, node attributes (such as corresponding file identifiers, recommended retention periods, and destruction trigger scores), and edge attributes (such as constraint strength and association type). Each subgraph maintains an independent data structure, without interfering with each other, facilitating parallel processing.

[0130] To support parallel processing, an independent processing task queue is allocated to each subgraph. The task queue adopts a priority queue structure. Task items in the queue include information such as the identifier of the node to be processed, the current processing stage (e.g., pending identification, pending review, pending destruction), and priority score. The priority score is calculated based on a combination of factors, including the destruction trigger score, the urgency of the file's expiration date, and its business importance. All nodes requiring processing in the subgraph are inserted into the queue according to priority, and subsequent processing operations are executed sequentially according to the queue order, ensuring that high-priority files are processed first.

[0131] Bidirectional information propagation is performed independently for each subgraph, following the same mechanism as in the original graph, but with the propagation scope strictly limited to the subgraph itself. In the forward propagation phase, starting from the root node of the subgraph (the node with an in-degree of zero), the retention period constraint and disposal status are passed downwards along the directed edges. Since conflicting edges have been broken, there are no more constraint conflicts within the subgraph, and the calculation of the cumulative constraint strength value can be completed smoothly. In the backward propagation phase, starting from the leaf node of the subgraph (the node with an out-degree of zero), the incomplete disposal flag and blocking information are passed upwards. Since dependent edges have been broken, infinite propagation caused by circular dependencies will no longer occur. The propagation process for each subgraph is independent and can be executed in parallel using multi-threading or distributed methods, significantly improving disposal efficiency.

[0132] After bidirectional propagation is complete, each node in the subgraph is checked individually to see if it meets the destruction conditions. The determination is made by comprehensively considering the node's destruction trigger score. Recommended storage period .if And the current time With the time of the creation of the archives The difference satisfies If a node does not receive blocking information from its child nodes during backpropagation, it is determined that the node meets the destruction conditions. The electronic files corresponding to nodes that meet the conditions will be marked as destroyable, triggering a migration operation to move the file data from online storage to a dedicated destruction repository, awaiting final physical destruction or logical deletion. Nodes that do not meet the destruction conditions will remain in their original state, awaiting reassessment in the next authentication cycle.

[0133] Through the aforementioned decomposition and independent handling mechanism, the complex file association structure that was originally unable to be handled properly due to conflicts and circular dependencies has been effectively resolved. Each sub-graph can independently complete the identification and handling according to the standardized process, which not only ensures the correctness of the handling, but also improves the overall processing capacity of the system.

[0134] In one optional implementation, determining whether the destruction conditions are met based on the destruction trigger score and the recommended retention period value of each node in the subgraph, and migrating the electronic file to be identified corresponding to the current node to the destruction database when the destruction conditions are met, includes:

[0135] Traverse all nodes in each subgraph, compare the destruction trigger score of each node with the preset destruction score threshold, obtain the actual storage time of each node according to the current system time, and determine that the current node meets the destruction conditions when the destruction trigger score is greater than the destruction score threshold and the actual storage time exceeds the recommended storage period value.

[0136] For nodes that meet the destruction conditions, perform dependency checks, read the set of child nodes of the current node, query the disposal status identifier of all child nodes in the set of child nodes, and when the disposal status identifier of all child nodes is in the completed disposal status, determine that the current node has no incomplete dependencies and generate a destruction operation instruction.

[0137] According to the destruction operation instruction, the electronic file to be identified is read from the current storage location and packaged into a data packet to be migrated. A data transmission channel is established with the destruction library. Based on the data transmission channel, the data packet to be migrated is written into the designated storage area of ​​the destruction library. The disposal status identifier of the current node is updated to the destroyed status, and the operation of migrating the electronic file to be identified to the destruction library is completed.

[0138] After the archival system completes the decomposition and conflict resolution of the directed acyclic graph, it is necessary to perform precise destruction condition determination and actual migration operations on the archival nodes in each independent subgraph. For the multiple subgraphs obtained after decomposition, all nodes in each subgraph are traversed sequentially according to the topological sort order. For each node to be inspected, the previously calculated destruction trigger score is extracted from the node attributes. This score comprehensively reflects the semantic matching degree between the archival content and the identification rules, as well as the decay trend of the archival value over time. A preset destruction score threshold is maintained, which is dynamically adjusted according to the archival type and industry standards. For government archives, it is usually set to 0.75, and for business archives, it can be set to 0.68. The destruction trigger score of the current node is compared with this threshold. Only when the score exceeds the threshold is the node qualified to enter the destruction process.

[0139] After the scoring is passed, the current system timestamp is obtained, and the file creation time is read from the node's metadata. The actual retention period of the file is calculated using the time difference. This retention period is precisely calculated in days. For example, if the file was created on March 15, 2020, and the current time is September 10, 2023, the actual retention period is 1274 days. This actual retention period is compared with the recommended retention period value generated in the semantic matching stage. The recommended retention period value is usually in years, for example, 3 years corresponds to 1095 days, and 10 years corresponds to 3650 days. Only when the actual retention period is significantly greater than the number of days corresponding to the recommended retention period value is the file considered to meet the necessary conditions for destruction in the time dimension. This dual judgment mechanism ensures that only files that simultaneously meet the criteria of low semantic value and sufficient retention period can enter the destruction candidate set.

[0140] Once a node passes the two basic checks mentioned above, it is not immediately destroyed. Instead, a dependency check process is initiated. This involves reading the outgoing edges of the current node in the directed acyclic graph to obtain a list of identifiers for all direct child nodes. A disposal status table is maintained, recording the real-time disposal status of each archive node, including states such as undisposed, in process, archived, and destroyed. For each child node of the current node, a query request is sent to the disposal status table to retrieve the disposal status identifier field of that child node. Only when the returned result shows that all child nodes have a disposal status identifier of "completed," meaning the child nodes have completed archiving or destruction, is it determined that the current node has no incomplete dependencies. This mechanism ensures the sequential nature of archive disposal, preventing premature destruction of parent archives before child archives are properly processed, and avoiding compliance risks caused by broken archive chains.

[0141] If the disposal status of a child node is displayed as "not disposed of" or "in disposal," the destruction operation of the current node is suspended, the node is added to the pending queue, and the reason for the blockage is recorded in the system log. The dependency status of these suspended nodes is periodically rechecked, and the processing flow is automatically resumed when the blocking conditions are resolved. This asynchronous processing mechanism allows the system to advance disposal work in parallel across different archive nodes, improving overall disposal efficiency.

[0142] Once dependency checks confirm that the current node has no outstanding dependencies, a structured destruction operation instruction is generated. This instruction includes key fields such as file identifier, storage path, migration target, operation timestamp, and operator information. Based on this instruction, the physical or logical location of the electronic file to be authenticated in the current storage system is located, and the original file data is read into a memory buffer via a file system interface or object storage interface. During the reading process, the system synchronously calculates the hash value of the data for subsequent integrity verification.

[0143] After the archival data is read, it is packaged into a data packet to be migrated. This data packet uses a unified packaging format and includes three parts: the archival text, metadata, and disposal audit information. The archival text retains its original format, while the metadata section supplements information such as the reason for destruction, the basis for destruction, and the appraisal score. The disposal audit information records the complete disposal decision-making process, including semantic matching scores, dependency check results, and approval processes, ensuring the traceability of the destruction operation. During the packaging process, the system compresses the data packet to reduce subsequent transmission overhead and adds a version identifier and checksum to the header of the data packet. Subsequently, a dedicated data transmission channel is established with the destruction repository, an independent storage area with special access control and auditing capabilities, typically deployed on physically isolated storage nodes. When the transmission channel is established, authentication is performed to verify that the current operation has destruction permissions, and then encryption parameters are negotiated, using transport layer encryption protocols to protect the confidentiality of data during transmission. After the channel is established, a migration request is sent to the access layer of the destruction repository to request the allocation of storage space. The destruction repository returns the specified storage area identifier based on the archival type and retention period; for example, financial archives are assigned to area A of the destruction repository, and personnel archives are assigned to area B.

[0144] The data packets to be migrated are streamed into the designated storage area of ​​the destruction database via the established transmission channel. The writing process employs a block-based transmission mechanism; each data block contains a sequence number and verification information. The receiving end of the destruction database performs real-time verification on each data block, immediately requesting retransmission when a transmission error is detected to ensure data integrity. After all data blocks have been written, a write completion signal is sent, and the destruction database performs a final integrity check, comparing the hash value of the received data with the original hash value provided by the sending end. Upon successful verification, the destruction database returns a successful migration confirmation and generates a unique storage identifier for the file in the destruction database.

[0145] A second aspect of this invention provides an intelligent authentication and automatic processing system for electronic archives based on semantic matching, comprising:

[0146] The rule acquisition unit is used to acquire the electronic files to be authenticated along with the corresponding authentication rules and retention period constraints.

[0147] The feature encoding unit is used to perform semantic decomposition on the identification rules to extract conditional predicates, perform multi-level feature encoding on the content of the electronic archive to be identified, and introduce the time decay factor of the archive formation time and business context to obtain the archive features of the electronic archive to be identified.

[0148] The semantic matching unit is used to calculate the semantic distance between the archive features and the conditional predicates in the feature space, and to fuse the matching results of different conditional predicates to generate a recommended retention period value and a destruction trigger score.

[0149] The graph construction unit is used to construct a directed acyclic graph containing the electronic file to be identified and the corresponding multi-level related files. It performs bidirectional information propagation based on message passing on the directed acyclic graph. During forward propagation, the retention period constraint and disposal status of the parent node are passed along the edge to the child node and the constraint strength is accumulated. During reverse propagation, the unfinished disposal mark and blocking information of the child node are sent back to the parent node.

[0150] The conflict detection unit is used to decompose the directed acyclic graph into multiple independently disposable subgraphs when a conflict or circular dependency in the retention period constraint between nodes is detected through the bidirectional information propagation. The unit determines whether the destruction conditions are met based on the destruction trigger score and the recommended retention period value of the node in each subgraph. When the destruction conditions are met, the electronic file to be identified corresponding to the current node is migrated to the destruction database.

[0151] A third aspect of the present invention provides an electronic device, comprising:

[0152] processor;

[0153] Memory used to store processor-executable instructions;

[0154] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0155] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0156] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0157] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent identification and automatic processing of electronic archives based on semantic matching, characterized in that, include: Obtain the electronic files to be authenticated, along with the corresponding authentication rules and retention period constraints; The identification rules are semantically decomposed to extract conditional predicates. The content of the electronic archive to be identified is encoded with multi-level features and the time decay factor of the archive formation time and business context are introduced to obtain the archive features of the electronic archive to be identified. The semantic distance between the archive features and the conditional predicates is calculated in the feature space, and the matching results of different conditional predicates are fused to generate a recommended retention period value and a destruction trigger score. Construct a directed acyclic graph containing the electronic files to be identified and the corresponding multi-level related files. Perform bidirectional information propagation based on message passing on the directed acyclic graph. During forward propagation, the retention period constraint and disposal status of the parent node are passed along the edge to the child node and the constraint strength is accumulated. During reverse propagation, the unfinished disposal mark and blocking information of the child node are sent back to the parent node. When a conflict or circular dependency in the retention period constraint between nodes is detected through the bidirectional information propagation, the directed acyclic graph is decomposed into multiple subgraphs that can be disposed of independently. The destruction trigger score and the recommended retention period value of each node in the subgraph are used to determine whether the destruction conditions are met. When the destruction conditions are met, the electronic file to be identified corresponding to the current node is migrated to the destruction database.

2. The method according to claim 1, characterized in that, The identification rules are semantically decomposed to extract conditional predicates. The content of the electronic archive to be identified is then multi-level feature encoded, and a temporal decay factor of the archive's formation time and business context are introduced to obtain the archive features of the electronic archive to be identified, including: Dependency parsing is performed on the text description in the identification rules to identify the subject-verb-object structure and modification relations in the identification rules. The predicate part in the subject-verb-object structure is used as the core semantic unit of the condition judgment, and the qualifiers and constraints in the modification relations are used as the boundary constraints of the condition judgment. The core semantic unit and the boundary constraints are extracted to form the structured representation of the condition predicate. The structured fields and unstructured text of the electronic archive to be identified are encoded and feature extracted to obtain basic features. The interaction weight between the structured fields and the unstructured text is calculated through a self-attention mechanism. The basic features are then weighted and fused according to the interaction weight to obtain a fused feature representation. The time difference between the archive creation time and the current time is calculated and mapped to an attenuation coefficient. The attenuation coefficient is inversely proportional to the time difference. The attenuation coefficient is multiplied element-wise with the fused feature representation to obtain the time-attenuated feature. The identification information of the preceding and subsequent business nodes is extracted from the business process to which the electronic file to be identified belongs and encoded into a context vector. The context vector is then concatenated with the time-decayed features to obtain the file features.

3. The method according to claim 1, characterized in that, The semantic distance between the archive features and the conditional predicates is calculated in the feature space, and the matching results of different conditional predicates are fused to generate a recommended retention period value and a destruction trigger score, including: The structured representations of the archive features and the conditional predicates are mapped to a unified semantic feature space, and the distance metric between the vector representations of the archive features and the vector representations of the structured representations of the conditional predicates is calculated in the semantic feature space as the semantic distance. The semantic distance between the file features and the multiple conditional predicates contained in the identification rule is calculated respectively to obtain multiple semantic distance values. The weight coefficient of each conditional predicate is determined according to the logical relationship of each conditional predicate in the identification rule. The multiple semantic distance values ​​and the corresponding weight coefficients are weighted and summed to obtain the comprehensive matching score. The retention period labeling information associated with each conditional predicate is extracted from the identification rules. The retention period labeling information is weighted and averaged according to the semantic distance between each conditional predicate and the archival features to obtain the recommended retention period value. The weight of the retention period labeling information corresponding to each conditional predicate is inversely proportional to the semantic distance. The recommended retention period is compared with the actual retention time of the electronic file to be identified. When the actual retention time exceeds the recommended retention period, the excess ratio is calculated. The excess ratio is weighted and combined with the comprehensive matching score to obtain the destruction trigger score.

4. The method according to claim 1, characterized in that, Constructing a directed acyclic graph that includes the electronic file to be identified and its corresponding multi-level related files includes: Extract a set of identifiers for associated files from the metadata and business attributes of the electronic file to be identified, parse the association type attribute between each associated file in the identifier set and the electronic file to be identified, and determine a business dependency relationship when the association type attribute is identified as a business process prerequisite, and a reference relationship when the association type attribute is identified as a content reference source. The hierarchical structure between files is determined based on the business dependency relationship and the reference relationship. The electronic file to be identified is taken as the target node, the files that the electronic file to be identified depends on and the files that are referenced by the electronic file to be identified are taken as parent nodes, and the files that depend on the electronic file to be identified and the files that reference the electronic file to be identified are taken as child nodes. A directed edge of dependency is established from the dependent file node to the dependent file node based on the business dependency relationship, and a directed edge of reference is established from the referenced file node to the referencing file node based on the reference relationship. The association strength value is calculated and normalized based on the access frequency and time span between the associated archive and the electronic archive to be identified, and used as the weight of the corresponding directed edge to obtain the directed acyclic graph.

5. The method according to claim 1, characterized in that, On the directed acyclic graph, bidirectional message-passing-based information propagation is performed. During forward propagation, the retention period constraint and disposal status of the parent node are passed along the edges to the child nodes, accumulating constraint strength. During backward propagation, the incomplete disposal flag and blocking information of the child nodes are sent back to the parent node, including: Starting from the root node of the directed acyclic graph, forward propagation is performed. The retention period constraint carried by the current parent node is read as the constraint base value. The constraint base value is multiplied by the weight of the edge between the current parent node and the child node to obtain the transit constraint value. The cumulative constraint strength value of the child node is obtained by summing the transmitted constraint value with the historical constraint value received by the child node. At the same time, the current parent node's disposal status identifier is transmitted to the child node and recorded as the parent node's disposal status. The disposal status identifier includes pending disposal status, disposal status, completed disposal status, and destroyed status. Backpropagation is performed starting from the leaf node of the directed acyclic graph. When an incomplete disposal task is detected in the current child node, an incomplete disposal identifier is generated. The cumulative constraint strength value of the current child node is compared with the recommended retention period value to determine whether there is a constraint conflict. The disposal status of the parent node of the current child node is matched with the preset dependency conditions to determine whether the dependency conditions are met. The comparison result of the constraint conflict and the matching result of the dependency conditions are encapsulated into the blocking information. The incomplete processing flag and the blocking information are passed to the parent node along the reverse edge of the directed acyclic graph. After receiving the incomplete processing flag, the parent node suspends its own processing operation and waits for the processing completion signal from the child node.

6. The method according to claim 5, characterized in that, When a conflict or circular dependency in the retention period constraint between nodes is detected through the bidirectional information propagation, the directed acyclic graph is decomposed into multiple independently manageable subgraphs, including: During the bidirectional information propagation process, when the difference between the cumulative constraint strength value of the child node and the constraint reference value transmitted by the parent node exceeds a preset conflict threshold, it is determined to be a storage period constraint conflict; when a duplicate node identifier is detected in the reverse edge propagation path of the node, it is determined to be a circular dependency. Identify the node pairs in the directed acyclic graph that conflict with the retention period constraint and the set of nodes that have cyclic dependencies. Mark the directed edges between the node pairs as conflict edges and the directed edges in the set of nodes that form closed loops as dependent edges. Disconnect the conflict edges and the dependent edges to obtain multiple disconnected subgraph structures. Connectivity detection is performed on each subgraph structure, and the node identifiers and corresponding directed edge information in each subgraph structure are reorganized into an independent graph data structure. An independent processing task queue is assigned to each subgraph structure, and the bidirectional information propagation and processing operation is executed independently on each subgraph structure to obtain multiple subgraphs that can be processed independently.

7. The method according to claim 1, characterized in that, Based on the destruction trigger score and the recommended retention period value of each node in the subgraph, it is determined whether the destruction conditions are met. When the destruction conditions are met, the electronic file to be identified corresponding to the current node is migrated to the destruction database, including: Traverse all nodes in each subgraph, compare the destruction trigger score of each node with the preset destruction score threshold, obtain the actual storage time of each node according to the current system time, and determine that the current node meets the destruction conditions when the destruction trigger score is greater than the destruction score threshold and the actual storage time exceeds the recommended storage period value. For nodes that meet the destruction conditions, perform dependency checks, read the set of child nodes of the current node, query the disposal status identifier of all child nodes in the set of child nodes, and when the disposal status identifier of all child nodes is in the completed disposal status, determine that the current node has no incomplete dependencies and generate a destruction operation instruction. According to the destruction operation instruction, the electronic file to be identified is read from the current storage location and packaged into a data packet to be migrated. A data transmission channel is established with the destruction library. Based on the data transmission channel, the data packet to be migrated is written into the designated storage area of ​​the destruction library. The disposal status identifier of the current node is updated to the destroyed status, and the operation of migrating the electronic file to be identified to the destruction library is completed.

8. A semantic matching-based intelligent identification and automatic processing system for electronic archives, used to implement the method as described in any one of claims 1-7, characterized in that, include: The rule acquisition unit is used to acquire the electronic files to be authenticated along with the corresponding authentication rules and retention period constraints. The feature encoding unit is used to perform semantic decomposition on the identification rules to extract conditional predicates, perform multi-level feature encoding on the content of the electronic archive to be identified, and introduce the time decay factor of the archive formation time and business context to obtain the archive features of the electronic archive to be identified. The semantic matching unit is used to calculate the semantic distance between the archive features and the conditional predicates in the feature space, and to fuse the matching results of different conditional predicates to generate a recommended retention period value and a destruction trigger score. The graph construction unit is used to construct a directed acyclic graph containing the electronic file to be identified and the corresponding multi-level related files. It performs bidirectional information propagation based on message passing on the directed acyclic graph. During forward propagation, the retention period constraint and disposal status of the parent node are passed along the edge to the child node and the constraint strength is accumulated. During reverse propagation, the unfinished disposal mark and blocking information of the child node are sent back to the parent node. The conflict detection unit is used to decompose the directed acyclic graph into multiple independently disposable subgraphs when a conflict or circular dependency in the retention period constraint between nodes is detected through the bidirectional information propagation. The unit determines whether the destruction conditions are met based on the destruction trigger score and the recommended retention period value of the node in each subgraph. When the destruction conditions are met, the electronic file to be identified corresponding to the current node is migrated to the destruction database.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images