Electronic archive big data intelligent collection method and system

Abstract

This invention provides a method and system for intelligent acquisition of big data from electronic archives, relating to the field of data processing technology. The method includes deconstructing and separating multimodal information from the original data of electronic archives to generate hybrid feature vectors, mapping and locating candidate entity sets in a knowledge graph, constructing a multi-level relationship network to obtain associated entity clusters, forming an archive knowledge association system, realizing semantic association analysis and knowledge mining of archive information, improving the efficiency of intelligent acquisition of archive information, and enhancing the ability to discover implicit relationships between cross-archive entities.

Description

Technical Field

[0001] This invention relates to data processing technology, and more particularly to a method and system for intelligent acquisition of big data from electronic archives. Background Technology

[0002] With the rapid development of information technology, electronic archives have become an important form of modern archives management. Electronic archive data is experiencing explosive growth, containing multimodal information such as text, images, audio, and video, posing a significant challenge to the effective utilization of archival resources. Traditional archives management methods, primarily relying on manual classification and organization, are ill-suited to handling the processing demands of large-scale electronic archives. In recent years, knowledge graph technology, as an effective knowledge representation and management method, has provided new insights into the intelligent management of electronic archives. Knowledge graphs construct semantic networks through entities and relationships, capable of expressing complex knowledge structures and connections, providing technical support for the intelligent acquisition and semantic analysis of electronic archives.

[0003] Existing technologies struggle to effectively process multimodal information in electronic archives. For archive content that includes various forms such as text and images, there is a lack of unified feature extraction and representation methods, leading to insufficient information utilization and affecting the accuracy of entity recognition in archives.

[0004] Traditional methods for mapping archival entities often employ simple character matching or keyword indexing, failing to fully consider the semantic relationships and contextual information between entities and ignoring the rich structured information contained in knowledge graphs, resulting in inaccurate archival entity mapping results.

[0005] Existing technologies have limitations in handling implicit relationships between archives. They are difficult to discover potential connections between non-directly related archival entities, cannot construct a complete archival knowledge association system, limit the mining and utilization of deep semantic relationships in archival resources, and cannot meet the knowledge discovery needs in complex scenarios. Summary of the Invention

[0006] The present invention provides a method and system for intelligent collection of electronic archive big data, which can solve the problems in the prior art.

[0007] A first aspect of the present invention provides a method for intelligent collection of big data from electronic archives, comprising:

[0008] Receive the raw data of the electronic archives to be processed, perform data deconstruction on the raw data of the electronic archives to separate multimodal information, and generate a hybrid feature vector;

[0009] Based on the hybrid feature vector, a candidate entity set is mapped and located in the semantic embedding space of the knowledge graph. The candidate entity set is then filtered by depicting the local subgraph structure and fusing path connectivity and attribute consistency constraints between entities within the subgraph. The archive entity mapping result is then output.

[0010] Each entity in the archive entity mapping result is defined as a core node. A multi-level relationship network is constructed in the knowledge graph. The semantic path range is extended through the multi-level relationship network to obtain associated entity clusters. The implicit association strength is calculated based on the path span between each entity in the associated entity cluster and the core node.

[0011] Based on the implicit association strength, a dynamic connection structure is constructed across archival entities to form an archival knowledge association system. By describing the association hotspots and sparse areas through the structural changes of the archival knowledge association system, a hierarchical semantic expression structure of the knowledge graph is obtained.

[0012] Based on the hybrid feature vectors, candidate entity sets are mapped and located in the semantic embedding space of the knowledge graph. Then, by depicting the local subgraph structure of the candidate entity sets and fusing path connectivity and attribute consistency constraints between entities within the subgraphs, the results are filtered, and the output archive entity mapping results include:

[0013] The hybrid feature vector is embedded into the semantic representation space to construct the semantic similarity distribution between the hybrid feature vector and the embedded representation of each entity in the knowledge graph. The density clustering characteristics of the semantic similarity distribution are analyzed to form candidate regions. Entities within the candidate regions are condensed to generate a candidate entity set.

[0014] Extend each candidate entity node in the candidate entity set, connect adjacent entities and connecting edges in the knowledge graph, and construct a local subgraph structure;

[0015] By integrating the relation types and directions in the local subgraph structure, the semantic path connectivity from each candidate entity to other candidate entities is quantified. At the same time, the structured attribute constraints in the original electronic archive data are separated, and the structured attribute constraints are compared with the attribute nodes of each candidate entity in the local subgraph structure to form an attribute consistency verification matrix.

[0016] By integrating the semantic path connectivity and the attribute consistency verification matrix, the comprehensive scores of each candidate entity are aggregated, and the candidate entity with the best score is locked to output the archive entity mapping result.

[0017] The hybrid feature vector is embedded into the semantic representation space to construct the semantic similarity distribution between the hybrid feature vector and the embedded representations of each entity in the knowledge graph. The density clustering characteristics of the semantic similarity distribution are analyzed to form candidate regions, including:

[0018] The hybrid feature vector is nonlinearly transformed and mapped to the semantic embedding space to obtain the vector representation of the hybrid feature vector in the semantic embedding space, and a reference vector set is constructed by generating the pre-trained embedding representation of each entity in the knowledge graph.

[0019] The semantic distance is calculated based on the vector representation and the reference vector set. The semantic similarity metric is constructed using the semantic distance. The semantic similarity metric is then projected into the semantic embedding space to form a semantic similarity distribution.

[0020] Based on the local density peak regions in the semantic similarity distribution, the spatial concentration and gradient change characteristics of similarity values ​​within the regions are analyzed to determine density clustering characteristics; according to the spatial boundaries and density threshold conditions of the density clustering characteristics, closed regions are located in the semantic embedding space to form candidate regions.

[0021] Each entity in the archive entity mapping result is defined as a core node. A multi-level relationship network is constructed in the knowledge graph. The semantic path range is extended through the multi-level relationship network to obtain associated entity clusters, including:

[0022] Each entity in the document entity mapping result is defined as a core anchor node. The first-level entities and relation edges that are directly related are extended in the knowledge graph through the progressive decomposition of entity elements, and the first-level relation network is dynamically constructed.

[0023] Based on the multidimensional connection features of the first-level relationship network, the second-level entities and relationship edges associated with each first-level entity are linked together, and the first-level relationship network and the second-level entities and relationship edges are hierarchically coupled and fused to generate a multi-level relationship network.

[0024] The semantic derivation structure of the multi-level relationship network is analyzed, the relationship transmission links from each core anchor node to each level entity are sorted out, the entity node and relationship edge sequence in each relationship transmission link are summarized, and the semantic penetration depth of each relationship transmission link is quantified.

[0025] Around the core anchor node in the multi-level relationship network, entities at various levels are integrated based on the semantic quantization strength of the relationship links to form a cluster of related entities.

[0026] Each entity in the document entity mapping result is defined as a core anchor node. Through progressive decomposition of entity elements, directly related first-level entities and relation edges are extended in the knowledge graph, dynamically constructing a first-level relation network including:

[0027] The structured elements of each entity in the archive entity mapping result are analyzed, and the structured elements are decomposed into attribute dimension set and semantic dimension set. Based on the combined features of the attribute dimension set and the semantic dimension set, the unique node corresponding to each entity is located in the knowledge graph, and the unique node is marked as the core anchor node.

[0028] In the knowledge graph, query the adjacent nodes that have a direct relationship with the core anchor node, extract the entity identifier and attribute information of the adjacent nodes, and define the adjacent nodes as first-level entities.

[0029] Extract the relationship edges connecting the core anchor node and each first-level entity, parse the relationship type and semantic direction of each relationship edge, construct relationship edge descriptors, and organize the core anchor node, the first-level entity and the relationship edge descriptors into a topological structure;

[0030] Based on the connection patterns between the core anchor nodes and the first-level entities in the topology, the semantic association strength of each first-level entity relative to the core anchor node is calculated. The topology is then weighted according to the semantic association strength to form a first-level relationship network.

[0031] Based on the implicit association strength, a dynamic connection structure is constructed across archival entities to form an archival knowledge association system. The structural changes of this system depict association hotspots and sparse regions, resulting in a hierarchical semantic representation structure of the knowledge graph, including:

[0032] Based on the implicit association weights of each entity in the associated entity cluster, the shared entity nodes between the associated entity clusters of different core anchor nodes are retrieved. The shared entity nodes and the implicit association weights are used to construct cross-cluster semantic connections. The weight propagation and aggregation calculation of the cross-cluster semantic connections form a dynamic connection structure across archival entities. The dynamic connection structure is integrated into a network topology to obtain the archival knowledge association system.

[0033] Based on the time-series snapshot acquisition of the aforementioned archive knowledge association system, the topological state at different times is obtained. The nodes and edges with changing connection strength are determined by differential comparison. The clustered areas of nodes with increasing connection strength are defined as association hotspots, and the scattered areas of nodes with decreasing connection strength are defined as sparse areas.

[0034] Based on the in-degree and out-degree distribution of the associated hotspots, core semantic nodes are located. Hierarchical transmission paths are constructed according to the connection relationship between the core semantic nodes and the boundaries of the sparse regions. The semantic diffusion features of the core semantic nodes on the hierarchical transmission paths are used to generate a multi-level semantic structure of the knowledge graph.

[0035] By propagating and aggregating the weights of the cross-cluster semantic connections, a dynamic connection structure is formed across archival entities. Integrating this dynamic connection structure into a network topology yields an archival knowledge association system, including:

[0036] For each cross-cluster semantic connection, the implicit association weights of the nodes at both ends of the connection in their respective associated entity clusters are obtained. Bidirectional propagation operation is performed based on the propagation path of the cross-cluster semantic connection. The weight values ​​are attenuated and modulated according to the path length to obtain the propagation weight sequence.

[0037] Based on the aggregation operation of each weight component in the propagation weight sequence at the intermediate node of the cross-cluster semantic connection, the cross-cluster aggregation weight is calculated according to the propagation depth and attenuation degree of the weight component, and the cross-cluster aggregation weight is labeled with the corresponding cross-cluster semantic connection to construct a weighted dynamic connection structure.

[0038] The set of nodes in the dynamic connection structure is mapped to vertices of the network topology, the cross-cluster semantic connection is mapped to directed edges of the network topology, and the cross-cluster aggregation weight is mapped to the weight value of the directed edge. Based on the topological connection relationship between the vertices through the directed edges and their weight values, an archival knowledge association system is constructed.

[0039] A second aspect of the present invention provides an intelligent data acquisition system for electronic archives, comprising:

[0040] The receiving module is used to receive the original electronic archive data to be processed, perform data deconstruction on the original electronic archive data to separate multimodal information, and generate a hybrid feature vector;

[0041] The filtering module is used to map and locate a set of candidate entities in the semantic embedding space of the knowledge graph based on the hybrid feature vector, and to filter the entity by describing the local subgraph structure of the candidate entity set and integrating the path connectivity and attribute consistency constraints between entities within the subgraph, and output the file entity mapping result.

[0042] The definition module is used to define each entity in the archive entity mapping result as a core node, construct a multi-level relationship network in the knowledge graph, extend the semantic path range through the multi-level relationship network to obtain associated entity clusters, and calculate the implicit association strength based on the path span between each entity in the associated entity cluster and the core node.

[0043] The construction module is used to build a dynamic connection structure across archival entities based on the implicit association strength to form an archival knowledge association system. Through the structural changes of the archival knowledge association system, the association hotspots and sparse areas are depicted to obtain the hierarchical semantic expression structure of the knowledge graph.

[0044] A third aspect of the present invention provides an electronic device, comprising:

[0045] processor;

[0046] Memory used to store processor-executable instructions;

[0047] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0048] 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.

[0049] The beneficial effects of this application are as follows:

[0050] By deconstructing and separating multimodal information from the original data of electronic archives and generating hybrid feature vectors, the problem of insufficient processing of single-modal information in traditional archive data acquisition is solved, and the semantic features of the archive content can be fully captured.

[0051] Based on the mapping and local subgraph structure analysis of hybrid feature vectors in the semantic embedding space of knowledge graphs, and by integrating path connectivity and attribute consistency constraints for filtering, high-precision mapping of archive entities is achieved, effectively reducing entity ambiguity.

[0052] By constructing a multi-level relationship network with mapping entities as core nodes, the limitations of traditional archival classification boundaries are broken, enabling the automatic discovery of implicit relationships between archives and providing a new approach for in-depth mining of archival knowledge.

[0053] The archival knowledge association system built on implicit association strength can dynamically reflect the evolution characteristics of archival knowledge structure, help identify association hotspots and knowledge gaps, and provide decision support for the in-depth development and utilization of archival resources.

[0054] The overall approach enables the automated transformation of archival data into knowledge, significantly improving the intelligence level of electronic archival data acquisition and knowledge mining, and laying a technical foundation for the efficient management and application of large-scale archival resources. Attached Figure Description

[0055] Figure 1 This is a flowchart illustrating the intelligent data acquisition method for electronic archives according to an embodiment of the present invention.

[0056] Figure 2 This is a flowchart illustrating the dynamic weight propagation and topology construction process for cross-cluster semantic connections in an embodiment of the present invention. 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 data acquisition method for electronic archives according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0060] Receive the raw data of the electronic archives to be processed, perform data deconstruction on the raw data of the electronic archives to separate multimodal information, and generate a hybrid feature vector;

[0061] Based on the hybrid feature vector, a candidate entity set is mapped and located in the semantic embedding space of the knowledge graph. The candidate entity set is then filtered by depicting the local subgraph structure and fusing path connectivity and attribute consistency constraints between entities within the subgraph. The archive entity mapping result is then output.

[0062] Each entity in the archive entity mapping result is defined as a core node. A multi-level relationship network is constructed in the knowledge graph. The semantic path range is extended through the multi-level relationship network to obtain associated entity clusters. The implicit association strength is calculated based on the path span between each entity in the associated entity cluster and the core node.

[0063] Based on the implicit association strength, a dynamic connection structure is constructed across archival entities to form an archival knowledge association system. By describing the association hotspots and sparse areas through the structural changes of the archival knowledge association system, a hierarchical semantic expression structure of the knowledge graph is obtained.

[0064] In one optional implementation, candidate entity sets are mapped and located in the semantic embedding space of the knowledge graph based on the hybrid feature vectors. The candidate entity sets are then filtered by depicting the local subgraph structure and fusing path connectivity and attribute consistency constraints between entities within the subgraphs. The output archive entity mapping result includes:

[0065] The hybrid feature vector is embedded into the semantic representation space to construct the semantic similarity distribution between the hybrid feature vector and the embedded representation of each entity in the knowledge graph. The density clustering characteristics of the semantic similarity distribution are analyzed to form candidate regions. Entities within the candidate regions are condensed to generate a candidate entity set.

[0066] Extend each candidate entity node in the candidate entity set, connect adjacent entities and connecting edges in the knowledge graph, and construct a local subgraph structure;

[0067] By integrating the relation types and directions in the local subgraph structure, the semantic path connectivity from each candidate entity to other candidate entities is quantified. At the same time, the structured attribute constraints in the original electronic archive data are separated, and the structured attribute constraints are compared with the attribute nodes of each candidate entity in the local subgraph structure to form an attribute consistency verification matrix.

[0068] By integrating the semantic path connectivity and the attribute consistency verification matrix, the comprehensive scores of each candidate entity are aggregated, and the candidate entity with the best score is locked to output the archive entity mapping result.

[0069] The archival entity mapping process utilizes a pre-constructed knowledge graph semantic embedding space to represent the semantic features of each entity. This embedding space stores the semantic features of each entity in a dense vector format with a dimension of 512, and the vector elements take values ​​ranging from -1.0 to +1.0. The hybrid feature vector serves as a comprehensive semantic expression of the archival entity to be mapped. It consists of three parts: textual semantic features, structured attribute features, and contextual features. Each part occupies one-third of the total dimension of the vector, ensuring the balance and completeness of the feature representation.

[0070] Semantic similarity is calculated using cosine similarity, which is obtained by dividing the inner product of the mixed feature vector and the embedding representation of each entity in the knowledge graph by the product of the vector magnitudes. The similarity value ranges from -1.0 to +1.0, with values ​​closer to 1.0 indicating greater semantic similarity. During the similarity distribution formation process, all entities are sorted according to their similarity values ​​to construct a similarity density function. This function reflects the distribution characteristics of the number of entities within different similarity intervals. The peaks and troughs of the density function are observed to identify candidate region boundaries.

[0071] The density clustering algorithm uses a variant of the DBSCAN algorithm, setting a density radius threshold of 0.15 and a minimum number of sample points of 5 entities. It scans high-density regions in the similarity distribution that meet the conditions to form candidate regions. Entities within the candidate regions are filtered based on similarity values ​​and density connectivity, retaining entities with a similarity greater than 0.6 and density connectivity with neighboring entities to form a candidate entity set. The size of the set is controlled within the range of 20 to 50 entities to avoid excessive computational complexity in subsequent steps.

[0072] The local subgraph construction process begins with each candidate entity node in the candidate entity set and expands outward two hops depth along the connection edges in the knowledge graph. It collects the one-hop neighbor entities directly connected to the candidate entity, as well as the direct neighbor entities of those one-hop neighbor entities, forming a local subgraph structure centered on the candidate entity. Connection edge types include four types: semantic relationship edges, hierarchical relationship edges, attribute relationship edges, and instance relationship edges. Each edge type carries a directional identifier to distinguish between forward and reverse relationships. The number of edges is limited during subgraph construction; each entity node retains a maximum of 20 connection edges, prioritizing relationship edges with higher weights.

[0073] Semantic path connectivity quantification employs a multi-path aggregation mechanism, calculating the connected path between any two candidate entities through a subgraph structure. The path length is limited to three hops. The weight of each path is determined by the product of the weights of all connecting edges along the path. Edge weights are initialized based on the relation type: semantic relation edges have a weight of 0.9, hierarchical relation edges have a weight of 0.8, attribute relation edges have a weight of 0.7, and instance relation edges have a weight of 0.6. These weight values ​​reflect the strength and reliability of the relation. Multiple paths are aggregated using a weighted average to obtain the final semantic path connectivity. The connectivity value ranges from 0.0 to 1.0, with higher values ​​indicating stronger semantic associations between the two entities.

[0074] Structured attribute constraints are extracted from the structured fields of the original electronic archive data, including five categories of constraints: entity type constraints, time range constraints, geographic location constraints, numerical range constraints, and enumeration value constraints. Each type of constraint is converted into a standardized constraint expression: entity type constraints are represented as a set of type identifiers; time range constraints are represented as the interval between start and end times; geographic location constraints are represented as a coordinate range or administrative division code; numerical range constraints are represented as the interval between minimum and maximum values; and enumeration value constraints are represented as the set of allowed values.

[0075] Attribute consistency verification calculates the matching degree by comparing the attribute nodes of candidate entities in the local subgraph structure with the structured attribute constraints. Attribute nodes include the entity's type, time, location, numerical, and label attributes. The matching degree calculation uses an item-by-item comparison method. For each constraint, it checks whether the corresponding attribute of the candidate entity meets the constraint requirements. If it meets the requirements, it is scored as 1.0 points; if it partially meets the requirements, it is scored from 0.3 to 0.8 depending on the degree of deviation; and if it does not meet the requirements at all, it is scored as 0.0 points. The attribute consistency verification matrix stores the matching degree scores of each item in the form of a two-dimensional matrix with candidate entities as rows and constraints as columns.

[0076] The comprehensive scoring fusion process weights and combines semantic path connectivity and attribute consistency verification results. The semantic path connectivity weight is set to 0.4, and the average attribute consistency score weight is set to 0.6. This weighting reflects the importance of attribute matching in the mapping of archival entities. For each candidate entity, the average semantic path connectivity with other entities in the candidate entity set is calculated as the connectivity score, and the row average of its position in the attribute consistency verification matrix is ​​calculated as the consistency score. The final comprehensive score is the result of multiplying the connectivity score by 0.4 and the consistency score by 0.6.

[0077] The entity mapping output selects the candidate entity with the highest comprehensive score as the mapping target. When there is a tie for the highest score, the candidate entity with higher semantic similarity is selected first. If the semantic similarity is still the same, the candidate entity with higher connectivity in the knowledge graph is selected. The output includes fields such as the identifier of the selected entity, entity name, comprehensive score, semantic similarity score, connectivity score, consistency score, and mapping confidence. The mapping confidence is determined based on the difference between the comprehensive score and the second highest score. A difference greater than 0.2 indicates high confidence, a difference between 0.1 and 0.2 indicates medium confidence, and a difference less than 0.1 indicates low confidence.

[0078] In one optional implementation, the hybrid feature vector is embedded into a semantic representation space to construct a semantic similarity distribution between the hybrid feature vector and the embedded representations of each entity in the knowledge graph. The density clustering characteristics of the semantic similarity distribution are analyzed to form candidate regions, including:

[0079] The hybrid feature vector is nonlinearly transformed and mapped to the semantic embedding space to obtain the vector representation of the hybrid feature vector in the semantic embedding space, and a reference vector set is constructed by generating the pre-trained embedding representation of each entity in the knowledge graph.

[0080] The semantic distance is calculated based on the vector representation and the reference vector set. The semantic similarity metric is constructed using the semantic distance. The semantic similarity metric is then projected into the semantic embedding space to form a semantic similarity distribution.

[0081] Based on the local density peak regions in the semantic similarity distribution, the spatial concentration and gradient change characteristics of similarity values ​​within the regions are analyzed to determine density clustering characteristics; according to the spatial boundaries and density threshold conditions of the density clustering characteristics, closed regions are located in the semantic embedding space to form candidate regions.

[0082] In the process of embedding the hybrid feature vector into the semantic representation space, a nonlinear transformation is performed on the hybrid feature vector to map it to the semantic embedding space. Specifically, a deep neural network model is used as the carrier of the nonlinear transformation. This model contains multiple fully connected layers and activation functions. After the input hybrid feature vector undergoes a linear transformation with weight matrix W and bias vector b, it is then nonlinearly transformed by the ReLU activation function, expressed as f(x) = max(0, Wx + b). By stacking multiple such transformation structures, the hybrid feature vector is gradually mapped to the semantic embedding space, ultimately obtaining the vector representation of the hybrid feature vector in the semantic embedding space.

[0083] After obtaining the semantic space representation of the hybrid feature vectors, a reference vector set is constructed by extracting pre-trained entity embedding representations from the knowledge graph. The entity embedding representations in the knowledge graph are pre-obtained using knowledge graph embedding methods such as TransE, TransR, or ComplEx. These methods learn the distribution patterns of entities and relations in the semantic space during training, ensuring that semantically similar entities are close in the embedding space. The reference vector set contains the embedding representations of all entities in the knowledge graph, providing a foundation for subsequent semantic similarity calculations.

[0084] The semantic distance between the semantic space representation of the hybrid feature vector and the reference vector set is calculated. The semantic distance is measured using cosine similarity. For the semantic representation v of the hybrid feature vector and the entity embedding representation e_i in the reference vector set, their dot product is calculated and divided by their respective norms to obtain the semantic similarity value. By calculating the semantic similarity for all entity embedding representations, a semantic similarity distribution is formed. This distribution reflects the degree of relevance between the hybrid feature vector and each entity in the knowledge graph in the semantic space.

[0085] Based on the semantic similarity distribution, the density clustering characteristics of the distribution are analyzed to identify local density peak regions within the semantic similarity distribution. These regions indicate that the mixed feature vectors have high similarity with multiple entities in the semantic space. For each density peak region, the variance and average gradient of the similarity values ​​within the region are calculated to assess the spatial concentration of the similarity values. The smaller the variance, the more concentrated the similarity values ​​within the region; the larger the average gradient, the clearer the region boundaries.

[0086] Based on the characteristics of density clustering, a density threshold condition is set to determine the boundary of candidate regions. The density threshold condition includes a minimum density requirement and a density gradient threshold; only regions that simultaneously meet both conditions are considered valid density clusters. In the semantic embedding space, closed regions that meet the requirements are selected through the density threshold condition to form candidate regions. The candidate regions contain a set of knowledge graph entities that are highly semantically related to the mixed feature vector.

[0087] In practical applications, consider a product recommendation system where user behavioral characteristics (such as browsing history and click behavior) are fused with product characteristics (such as product description and category) to form a hybrid feature vector. This hybrid feature vector is then mapped to a semantic embedding space, and its semantic similarity to product entities in a knowledge graph is calculated to form a semantic similarity distribution. Density clustering characteristics are identified within this distribution to determine candidate regions containing semantically similar products, ultimately recommending a set of semantically relevant products to the user.

[0088] To optimize candidate regions, an adaptive density threshold mechanism is introduced. This mechanism dynamically adjusts the density threshold based on the overall characteristics of the semantic similarity distribution, avoiding the limitations of a fixed threshold under different distribution conditions. When the overall distribution density is high, the density threshold is increased; when the distribution is sparse, the density threshold is appropriately decreased to ensure that the selected candidate regions have relatively stable quality.

[0089] To improve the accuracy of candidate regions, a region boundary refinement process is introduced. On the initially determined candidate region boundaries, gradient descent is used to iteratively optimize the boundary positions along the gradient direction of semantic similarity. In each iteration, the boundary is adjusted in the direction of the larger gradient until a local optimum or the upper limit of the number of iterations is reached, thereby obtaining more accurate candidate region boundaries.

[0090] Using the above method, the hybrid feature vector is embedded into the semantic representation space to construct the semantic similarity distribution between the entity and the knowledge graph. The density clustering characteristics of the distribution are analyzed to form candidate regions, providing effective support for subsequent knowledge reasoning and decision-making.

[0091] In one optional implementation, each entity in the archive entity mapping result is defined as a core node, and a multi-level relationship network is constructed in the knowledge graph. The semantic path range is extended through this multi-level relationship network to obtain associated entity clusters, including:

[0092] Each entity in the document entity mapping result is defined as a core anchor node. The first-level entities and relation edges that are directly related are extended in the knowledge graph through the progressive decomposition of entity elements, and the first-level relation network is dynamically constructed.

[0093] Based on the multidimensional connection features of the first-level relationship network, the second-level entities and relationship edges associated with each first-level entity are linked together, and the first-level relationship network and the second-level entities and relationship edges are hierarchically coupled and fused to generate a multi-level relationship network.

[0094] The semantic derivation structure of the multi-level relationship network is analyzed, the relationship transmission links from each core anchor node to each level entity are sorted out, the entity node and relationship edge sequence in each relationship transmission link are summarized, and the semantic penetration depth of each relationship transmission link is quantified.

[0095] Around the core anchor node in the multi-level relationship network, entities at various levels are integrated based on the semantic quantization strength of the relationship links to form a cluster of related entities.

[0096] Each entity in the archive entity mapping result is defined by a unique identifier. Each entity carries an entity type tag, a semantic embedding vector, a set of attribute fields, and association weight information, and is stored as a core anchor node in the graph data structure. The core anchor node is implemented in the form of a node object, which includes a node identifier field, a node type field, a semantic vector field, an attribute mapping table field, and a neighbor connection table field. The node identifier uses 64-bit integer encoding to ensure global uniqueness, the node type field uses enumerated values ​​to represent the category to which the entity belongs, and the semantic vector field stores a 512-dimensional floating-point array to express the semantic features of the entity.

[0097] The progressive decomposition process of entity elements obtains decomposition dimensions by parsing the attribute mapping table of core anchor nodes. These dimensions include five basic dimensions: time, space, theme, hierarchy, and association. The time dimension decomposition extracts the entity's temporal attribute information, including creation time, modification time, effective time, and expiration time. The spatial dimension extracts spatial attributes such as geographical location, administrative division, physical location, and logical location. The theme dimension extracts theme attributes such as domain classification, professional category, content tags, and keywords. The hierarchy dimension decomposition identifies the entity's hierarchical, inclusive, and subordinate relationships within the knowledge system. The association dimension decomposition discovers the entity's reference, dependency, and collaborative relationships with other entities.

[0098] The first-level relational network is constructed by traversing the relational edges directly connected to the core anchor nodes in the knowledge graph to obtain the first-level entity set. Relational edge types include four basic types: semantic relational edges, structural relational edges, temporal relational edges, and attribute relational edges. Semantic relational edges represent semantic similarity connections between entities, with edge weights ranging from 0.1 to 1.0, reflecting the strength of semantic relevance. Structural relational edges represent organizational structure connections between entities, including hierarchical relationships, part-whole relationships, and classification relationships; the edge weight is fixed at 0.8 to reflect the stability of structural relationships. Temporal relational edges represent the temporal sequence relationships between entities, including predecessor, successor, and parallel relationships; the edge weight is dynamically adjusted according to the time interval, with smaller time intervals resulting in higher weights, up to a maximum weight of 0.9. Attribute relational edges represent attribute-sharing relationships between entities; the edge weight is calculated based on the number and importance of shared attributes.

[0099] The first-level entity expansion process sets an expansion radius parameter for each core anchor node, with a default expansion radius of 2.0 and an expansion depth of 1 hop. Expansion considers the combined influence of relation edge weights and entity importance scores. Entity importance scores are calculated using three network topology metrics: degree centrality, proximity centrality, and betweenness centrality. Degree centrality reflects the number of connections an entity has, proximity centrality reflects the average distance from an entity to other entities, and betweenness centrality reflects the entity's bridging role in the network. The first-level entity selection retains entities with an importance score greater than 0.3 and relation edge weights with core anchor nodes greater than 0.4, avoiding the introduction of noisy entities that could negatively impact network quality.

[0100] Multidimensional connectivity feature extraction and analysis reveals the topological characteristics of the first-level relational network, including network density, clustering coefficient, average path length, and degree distribution. Network density is calculated as the ratio of the actual number of edges to the total number of edges, reflecting the tightness of network connections; the density value ranges from 0.0 to 1.0. The clustering coefficient measures the local clustering characteristics of nodes in the network, obtained by calculating the connection ratio between a node's neighbors. The average path length reflects the average shortest distance between any two nodes in the network, and the degree distribution describes the distribution of the number of nodes with different degrees in the network.

[0101] The second-level entity expansion is based on the connection pattern of the first-level relationship network. It extends outwards by one hop along the relationship edges of the first-level entities to obtain the candidate set of second-level entities. The second-level entity selection employs stricter filtering conditions, requiring candidate entities to have direct connections with at least two first-level entities, with each connection edge weight greater than 0.3, and the candidate entity's importance score greater than 0.2. During the second-level expansion process, each first-level entity is limited to expanding to a maximum of 15 second-level entities to prevent excessive network expansion and a sharp increase in computational complexity.

[0102] The hierarchical coupling and fusion process integrates the first-level relationship network with the second-level entities and relationship edges to construct a unified multi-level relationship network data structure. Network nodes are divided into three levels: core anchor nodes, first-level nodes, and second-level nodes. Each node is labeled with its level information and its sequence number within that level. Relationship edges are divided into four connection types: core-to-first-level edges, first-level internal edges, first-level-to-second-level edges, and second-level internal edges. Edge weights retain their original values, and edge type labels are used for subsequent path analysis and weight calculation.

[0103] Semantic derivation structure parsing traverses a multi-level relational network using a breadth-first search algorithm to identify all reachable paths from the core anchor node to entities at each level. The path search depth is limited to 3 hops, meaning the path length from the core anchor node to a second-level entity does not exceed 3 edges. Path records contain information such as the sequence of entity nodes on the path, the sequence of relation edges, the total path length, and the path weight. The path weight is calculated using the geometric mean of the weights of all relation edges on the path; this geometric mean avoids the excessive influence of a single low-weight edge on the overall path weight.

[0104] The relationship propagation link analysis groups the searched paths according to their start and end nodes, forming a set of relationship propagation links originating from the core anchor node. Each relationship propagation link includes fields such as link identifier, start node, end node, sequence of intermediate nodes, sequence of relationship types, link length, and link weight. The link weight calculation considers the combined effect of path weight and path length; the weight of long paths is attenuated with a attenuation coefficient set to 0.85, meaning that the weight is multiplied by 0.85 for every additional hop in the path length.

[0105] Semantic penetration depth quantifies the strength of semantic information transmission from core anchor nodes to outer entities in a relational transmission chain. The penetration depth calculation combines three factors: chain weight, path length, and node importance. The penetration depth value ranges from 0.0 to 1.0, with higher values ​​indicating deeper semantic penetration. Different attenuation weights are assigned to entities at different levels during the calculation: the first-level entity has an attenuation weight of 1.0, and the second-level entity has an attenuation weight of 0.7, reflecting the layer-by-layer attenuation characteristic of semantic transmission.

[0106] The integration of associated entity clusters is performed independently around each core anchor node, aggregating peripheral entities based on the semantic quantization strength of the relationship links. The semantic quantization strength is calculated using a weighted average, comprehensively considering the relationship propagation link weight, semantic penetration depth, and semantic similarity between entities. An aggregation threshold of 0.4 is set; entities with a semantic quantization strength greater than 0.4 are included in their corresponding associated entity clusters. Each associated entity cluster contains a core anchor node and peripheral associated entities, with entities within the cluster sorted according to their semantic quantization strength.

[0107] In one optional implementation, each entity in the archive entity mapping result is defined as a core anchor node. The first-level relationship network is dynamically constructed by extending directly related first-level entities and relation edges in the knowledge graph through progressive decomposition of entity elements, including:

[0108] The structured elements of each entity in the archive entity mapping result are analyzed, and the structured elements are decomposed into attribute dimension set and semantic dimension set. Based on the combined features of the attribute dimension set and the semantic dimension set, the unique node corresponding to each entity is located in the knowledge graph, and the unique node is marked as the core anchor node.

[0109] In the knowledge graph, query the adjacent nodes that have a direct relationship with the core anchor node, extract the entity identifier and attribute information of the adjacent nodes, and define the adjacent nodes as first-level entities.

[0110] Extract the relationship edges connecting the core anchor node and each first-level entity, parse the relationship type and semantic direction of each relationship edge, construct relationship edge descriptors, and organize the core anchor node, the first-level entity and the relationship edge descriptors into a topological structure;

[0111] Based on the connection patterns between the core anchor nodes and the first-level entities in the topology, the semantic association strength of each first-level entity relative to the core anchor node is calculated. The topology is then weighted according to the semantic association strength to form a first-level relationship network.

[0112] The structured elements of each entity in the archive entity mapping results are analyzed and decomposed into attribute dimension sets and semantic dimension sets. The attribute dimension set includes the entity's basic features, such as entity identifier, name, timestamp, and category label; the semantic dimension set contains the entity's intrinsic features, such as topic description, key terms, and semantic vectors, representing the entity's semantic content. For example, for the entity "Archive File A", its attribute dimensions include "Number XC2023001", "Creation Date 2023-05-20", and "Security Level 2"; while the semantic dimensions include "Topic: System Design" and "Keywords: Data Processing, Knowledge Graph, Entity Mapping". Based on the combined features of these two dimension sets, a multi-dimensional index query is performed in the knowledge graph. A combination of attribute matching and semantic similarity calculation is used to locate the unique node corresponding to each entity. The location process employs a weighted matching algorithm, which assigns weight coefficients to the attribute and semantic dimensions, calculates a comprehensive matching score, and selects the knowledge graph node with the highest score exceeding a preset threshold as the unique corresponding node, marking it as the core anchor node.

[0113] When querying adjacent nodes in a knowledge graph that are directly linked to the core anchor node, a graph traversal algorithm is used. Specifically, starting from the core anchor node, all edges directly connected to it and their pointed-to nodes are explored. The traversal process is limited to a depth of 1 to ensure that only adjacent nodes reachable in one hop are retrieved. For each discovered adjacent node, its entity identifier (such as unique ID, URI, etc.) and attribute information (such as entity type, name, key attribute values, etc.) are extracted. For example, if "File A" is the core anchor node, the queried adjacent nodes include "Author B", "Project C", "Reference D", etc., and their identifiers and attribute information are extracted respectively. After extraction and standardization, the adjacent nodes are defined as first-level entities, forming the first-level entity set surrounding the core anchor node.

[0114] When extracting the relationship edges connecting the core anchor nodes and each first-level entity, the relationship type and direction are identified for each pair of connected nodes. Relationship type identification involves querying a predefined relationship type dictionary in the knowledge graph and matching the semantic labels of the edges; the relationship direction is determined based on the edge's orientation attribute in the knowledge graph, pointing from the source node to the target node. For each relationship edge, a relationship edge descriptor is constructed, containing elements such as relationship type, source entity ID, target entity ID, relationship strength, and time attribute. For example, the relationship edge descriptor between "file A" and "author B" can be represented as {type: "creator", source: file A_ID, target: author B_ID, strength: 0.92, time: "2023-05", etc.}. The core anchor nodes, first-level entities, and relationship edge descriptors are organized into a data structure representing the topological associations between entities according to their connection relationships in the knowledge graph.

[0115] Based on the connection patterns between core anchor nodes and first-level entities in the topology, the semantic association strength of each first-level entity relative to the core anchor node is calculated. The calculation process considers multiple factors: the type weight of relation edges (different types of relations have different semantic importance), attribute features on relation edges (such as time decay coefficients, citation frequency, etc.), and importance indicators of entity nodes (such as the centrality measure of an entity in the knowledge graph). These factors are combined to derive a numerical index representing the degree of semantic association. Based on the calculated semantic association strength, the relation edges in the topology are weighted; the higher the weight value, the stronger the association between the corresponding first-level entity and the core anchor node. Finally, a weighted first-level relation network is formed, which visually displays the most directly related entities around the core anchor node and their association strength distribution.

[0116] In practical applications, when a user queries a document entity (such as "Project Report P"), it is mapped to a core anchor node in the knowledge graph, and then a first-level relationship network is constructed using the method described above. Users can intuitively see which people, organizations, documents, and other entities the report is directly associated with, as well as the strength and nature of each association, thereby quickly grasping the context of the entity in the knowledge network and providing a foundation for further knowledge discovery and association analysis.

[0117] In one optional implementation, a dynamic connection structure across archival entities is constructed based on the implicit association strength to form an archival knowledge association system. The structural changes of this archival knowledge association system are used to depict association hotspots and sparse regions, resulting in a hierarchical semantic representation structure of the knowledge graph, including:

[0118] Based on the implicit association weights of each entity in the associated entity cluster, the shared entity nodes between the associated entity clusters of different core anchor nodes are retrieved. The shared entity nodes and the implicit association weights are used to construct cross-cluster semantic connections. The weight propagation and aggregation calculation of the cross-cluster semantic connections form a dynamic connection structure across archival entities. The dynamic connection structure is integrated into a network topology to obtain the archival knowledge association system.

[0119] Based on the time-series snapshot acquisition of the aforementioned archive knowledge association system, the topological state at different times is obtained. The nodes and edges with changing connection strength are determined by differential comparison. The clustered areas of nodes with increasing connection strength are defined as association hotspots, and the scattered areas of nodes with decreasing connection strength are defined as sparse areas.

[0120] Based on the in-degree and out-degree distribution of the associated hotspots, core semantic nodes are located. Hierarchical transmission paths are constructed according to the connection relationship between the core semantic nodes and the boundaries of the sparse regions. The semantic diffusion features of the core semantic nodes on the hierarchical transmission paths are used to generate a multi-level semantic structure of the knowledge graph.

[0121] When constructing an archival knowledge association system based on the implicit association weights of entities within associated entity clusters, starting from each archival entity, shared entity nodes between associated entity clusters with different core anchor nodes are retrieved. In practical implementation, implicit association weights can be determined by calculating the semantic similarity and co-occurrence frequency between entities; these weights reflect the strength of the potential semantic connections between entities.

[0122] For archival entities A and B, their implicit association weights can be calculated as follows: First, extract the semantic feature vectors of the entities and calculate the cosine similarity between the vectors; then, count the frequency of the entities appearing in the same context; finally, weight and fuse the semantic similarity and co-occurrence frequency to obtain the implicit association weight value. For example, in the analysis of historical archives, the implicit association weight between the entities "Opium War" and "Lin Zexu" will be relatively high because they are semantically closely related and often co-occur.

[0123] Once shared entity nodes are identified, these nodes are used as bridges between different associated entity clusters to construct cross-cluster semantic connections. Specifically, for each pair of associated entity clusters S1 and S2, all entity sets E shared by both clusters are identified. The implicit association weights of these shared entity nodes determine the strength of the cross-cluster connection. The weights of the cross-cluster semantic connections can be calculated through weight propagation and aggregation of shared entities, forming a weight matrix that represents the semantic association strength between different entity clusters.

[0124] By propagating and aggregating weights across cluster semantic connections, a dynamic connection structure is formed across archival entities. The weight propagation process is iterative. Initially, each entity cluster has an initial weight value, and then the weight is passed to neighboring clusters through connections. In each iteration, the weight passed from cluster i to cluster j is equal to the product of cluster i's current weight and the connection strength between the two clusters. After multiple iterations, the weight distribution tends to stabilize, forming the final dynamic connection structure.

[0125] By integrating the dynamic connection structure into a network topology, an archival knowledge association system is obtained. In this network, nodes represent entity clusters, edges represent cross-cluster semantic connections, and edge weights reflect connection strength. For example, when analyzing the relationships between ancient documents, a network structure is formed with the "Four Books and Five Classics" as the center, extending outwards to the annotations of classics from various dynasties, clearly demonstrating the relationship of knowledge inheritance and evolution.

[0126] When conducting time-series analysis based on an archival knowledge association system, periodic snapshots are taken of the association system to obtain the topological state at different times. These snapshots record the complete structural information of the network at a specific point in time, including the set of nodes, the set of edges, and their weight values. By performing differential comparison on snapshots from different time points, nodes and edges whose connection strength has changed significantly can be identified. In the specific implementation of differential comparison, the rate of change of edge weights is calculated, and changes exceeding a preset threshold are marked.

[0127] Regions where node clusters exhibit an increasing trend in connection strength are defined as associated hotspots. Associated hotspots represent thematic areas that attract significant attention and form dense knowledge connections over a period of time. In practical applications, a connection strength growth rate threshold α can be set. When more than n% of nodes in a region exhibit a connection strength growth rate greater than α, and the clustering degree of these nodes exceeds a preset value, the region is marked as an associated hotspot.

[0128] A sparse region is defined as a scattered area of ​​nodes where the connection strength is decreasing. Similarly, a threshold β for the connection strength attenuation rate is set. When more than m% of the nodes in the region have a connection strength attenuation rate greater than β and the nodes are relatively scattered, the region is marked as a sparse region.

[0129] Core semantic nodes are located based on the in-degree and out-degree distribution of nodes within related hotspots. These core semantic nodes are entities within the related hotspot regions that play a crucial role in information transmission; they can be effectively identified by analyzing the in-degree and out-degree of nodes. Specifically, the weighted sum of the in-degree and out-degree of each node in the hotspot region is calculated, and the nodes with the highest ranking are selected as core semantic nodes. For example, in a scientific literature network, nodes that are heavily cited and simultaneously cite multiple important documents are typically the core concepts or theories in the field.

[0130] Based on the connections between core semantic nodes and sparse region boundaries, hierarchical transmission paths are constructed. These paths reflect the way and intensity of knowledge diffusion from the core region to the edge region. In the construction process, the set of core semantic nodes C and the set of sparse region boundary nodes B are first determined. Then, the shortest path from each node in C to a node in B is calculated, and weights are assigned based on the connection strength along the path.

[0131] By leveraging the semantic diffusion features of core semantic nodes along hierarchical transmission paths, a multi-layered semantic structure for the knowledge graph is generated. Semantic diffusion features manifest as the changing patterns of semantic information within nodes during transmission, which can be captured by analyzing changes in semantic similarity between adjacent nodes along the path. Based on a stepwise decrease in semantic similarity, the knowledge graph can be divided into multiple concentric layers: the core layer contains basic concepts and theories, while the outer layers contain application cases and extended knowledge. The resulting hierarchical semantic representation structure intuitively demonstrates the inherent organization of the knowledge domain, providing structured support for efficient retrieval and knowledge mining of archival data.

[0132] In one optional implementation, a dynamic connection structure across archival entities is formed by weight propagation and aggregation calculation of the cross-cluster semantic connections. Integrating this dynamic connection structure into a network topology yields an archival knowledge association system, including:

[0133] For each cross-cluster semantic connection, the implicit association weights of the nodes at both ends of the connection in their respective associated entity clusters are obtained. Bidirectional propagation operation is performed based on the propagation path of the cross-cluster semantic connection. The weight values ​​are attenuated and modulated according to the path length to obtain the propagation weight sequence.

[0134] Based on the aggregation operation of each weight component in the propagation weight sequence at the intermediate node of the cross-cluster semantic connection, the cross-cluster aggregation weight is calculated according to the propagation depth and attenuation degree of the weight component, and the cross-cluster aggregation weight is labeled with the corresponding cross-cluster semantic connection to construct a weighted dynamic connection structure.

[0135] The set of nodes in the dynamic connection structure is mapped to vertices of the network topology, the cross-cluster semantic connection is mapped to directed edges of the network topology, and the cross-cluster aggregation weight is mapped to the weight value of the directed edge. Based on the topological connection relationship between the vertices through the directed edges and their weight values, an archival knowledge association system is constructed.

[0136] For each cross-cluster semantic connection, the implicit association weights of the nodes at both ends of the connection in their respective associated entity clusters are obtained. These implicit association weights are typically calculated based on the centrality and importance of the nodes in the entity clusters. For example, when there is a cross-cluster semantic connection between node a in entity cluster A and node b in entity cluster B, the implicit association weight Wa of node a in entity cluster A and the implicit association weight Wb of node b in entity cluster B are calculated respectively. The implicit association weights can be determined by indicators such as the degree centrality, proximity centrality, or eigenvector centrality of the nodes.

[0137] Bidirectional propagation is performed based on the propagation path of cross-cluster semantic connections. Bidirectional propagation means that weights propagate both from node a to node b and from node b to node a. For propagation from a to b, the initial weight value is Wa; for propagation from b to a, the initial weight value is Wb. During propagation, the weight values ​​are attenuated and modulated according to the path length. The attenuation function can adopt an exponential attenuation form, that is, when the propagation traverses a path of length L, the propagated weight value is the original weight value multiplied by the attenuation factor α^L, where α is the attenuation coefficient, typically ranging from (0, 1). In this way, the propagation weight sequence {Wa_1, Wa_2, ..., Wa_n} and {Wb_1, Wb_2, ..., Wb_m} are obtained, where n and m represent the propagation depth from a to b and from b to a, respectively.

[0138] The aggregation operation is performed on the intermediate nodes of cross-cluster semantic connections based on the propagation weights in the propagation weight sequence. The aggregation operation considers the propagation depth and attenuation of the weights to calculate the cross-cluster aggregation weight. The aggregation function can be defined as a weighted sum of two propagation weight sequences, i.e., W_agg = Σ(Wa_i * γ_i) + Σ(Wb_j * γ_j), where γ_i and γ_j are the depth adjustment factors for the corresponding weights, reflecting the influence of propagation depth on the aggregation weight. For example, in the application scenario of archival resource management, when calculating the cross-cluster aggregation weight between "personal archives" and "project archives," if there are multiple semantic connection paths between the two archive entities, each path will generate a corresponding propagation weight component, and the final connection strength is obtained through the aggregation function.

[0139] Cross-cluster aggregation weights are labeled onto the corresponding cross-cluster semantic connections to construct a weighted dynamic connection structure. In this structure, nodes represent archival entities, connections represent semantic relationships between entities, and weights represent the strength of those relationships. The characteristic of this dynamic connection structure is that the connection weights are dynamically adjusted as archival data is updated and new semantic connections are discovered, thus reflecting the latest state of knowledge association.

[0140] This method maps the set of nodes in a dynamic connection structure to vertices in the network topology, cross-cluster semantic connections to directed edges in the network topology, and cross-cluster aggregation weights to weights of directed edges. Based on the topological connections between vertices formed by directed edges and their weights, a complete archival knowledge association system is constructed. This network topology representation method facilitates knowledge graph visualization and subsequent association analysis.

[0141] In practical applications, taking the association of historical archive knowledge as an example, when analyzing the association between the entity clusters "historical figures" and "historical events," the semantic connections connecting the figure nodes and event nodes are first identified, such as relationships like "participation," "initiation," or "witnessing." For each cross-cluster connection, the centrality weight of the figure node in the figure cluster and the importance weight of the event node in the event cluster are calculated. Then, through bidirectional propagation operations, the weights of the figure nodes are propagated to the event nodes, and vice versa. Assuming there is a direct connection between figure A and event X, the weight propagation will consider other events in which figure A participates and other figures associated with event X, forming a complex propagation network.

[0142] During propagation, increased path length leads to weight decay. For example, when person A is indirectly associated with event Y through person B, the weight of this second-order connection is lower than that of a direct connection. In the aggregation phase, the final aggregation weight between person A and event X is calculated based on the weight components of all propagation paths, and this weight is then labeled onto the corresponding semantic connection. Ultimately, all person nodes, event nodes, and the weighted semantic connections between them constitute a complete historical knowledge association network, supporting association analysis and knowledge discovery in historical research.

[0143] The archival knowledge association system constructed in this way can not only reflect explicit direct connections, but also discover implicit indirect connections, providing strong support for the in-depth utilization of archival resources and knowledge mining.

[0144] The method further includes:

[0145] The data receiving and task scheduling layer manages the registration and transfer records of archival items through the archive receiving module. This module uses structured data storage, including fields for item identifier, receiving time, number of archives, archive type, and transferring unit. The item identifier uses 32-bit string encoding to ensure global uniqueness, the receiving time field uses a timestamp format accurate to the second, the number of archives is limited to 1 to 100,000 items, and the archive type field uses enumerated values ​​to support 15 basic types, including document archives, scientific and technological archives, and accounting archives. The task generation module automatically creates corresponding digitization processing tasks based on the archive receiving results. The task creation process parses the archive type and quantity information and generates three basic tasks—scanning, cataloging, and quality inspection—according to a preset processing flow template. Each task includes attribute fields such as task identifier, task type, priority, estimated working hours, and dependencies. The task distribution module implements an intelligent allocation mechanism based on process type and personnel skills. The allocation algorithm comprehensively considers three evaluation dimensions: current workload, historical completion quality, and professional skill level. Workload is calculated by the number of unfinished tasks and estimated remaining working hours. Historical completion quality is evaluated by the weighted average of the task completion rate and rework rate over the past 30 days. Professional skill level is divided into three levels: primary, intermediate, and advanced, corresponding to different task complexity thresholds.

[0146] The multi-source scanning module of the image acquisition and processing layer supports both flatbed scanners and high-speed scanners. It establishes a communication connection with the scanning device via the standard TWAIN protocol. Scanning parameter configuration includes four core items: resolution settings, color mode settings, paper size settings, and scan area settings. Resolution supports an adjustable range of 300 DPI to 600 DPI; color modes support black and white, grayscale, and color; and paper sizes support 12 standard sizes including A3, A4, and A5, as well as custom sizes. The batch operation module provides image file moving, sorting, and insertion scanning functions. Moving operations support batch transfer of images across file numbers and automatic renaming, using a file number prefix followed by a four-digit incrementing serial number. Sorting operations support sorting by filename, creation time, and file size. The insertion scanning function allows for scanning new images at any position and automatically rearranging subsequent page numbers, using a continuous numbering method with an increment of 1.

[0147] The quality inspection and rework closed-loop layer uses an image quality inspection module to check the quality of scanned images and mark errors. The quality inspection interface offers two viewing modes: online browsing and local download. The online browsing mode supports image zooming, rotation, and brightness adjustment, while the local download mode supports batch downloading of compressed file formats. Preset error types include five common error types: image blurring, excessive tilt, missing pages, duplicate scanning, and blemish interference. Quality inspectors can select preset error types or enter custom error descriptions, with the error description field limited to 500 characters. The rework assignment function allows quality inspectors to specify specific rework personnel. Rework tasks are automatically pushed to the assigned personnel's task list, and the push message includes error descriptions, rework requirements, and deadlines. The cataloging quality inspection module verifies the accuracy of cataloging data. Quality inspection tasks support two allocation methods: by volume or by item. Assigning by volume assigns the entire archive file to a single quality inspector, while assigning by item distributes the archives to multiple quality inspectors for parallel processing. The quality inspection interface displays the file image on the left and the corresponding bibliographic field information on the right. It supports field-by-field comparison and verification. The field-level modification function records the operation time, modification content, and operator information for each modification. The rework process tracking module maintains a real-time update mechanism for task status. Task status includes five basic states: pending, processing, completed, reworking, and reworked. Status changes trigger automatic notifications to relevant personnel, supporting both in-site messages and email alerts.

[0148] The intelligent cataloging and recognition layer's image cataloging module adopts a left-right split layout. The left image display area supports adjustable image scaling from 25% to 400%, and image navigation supports thumbnail preview and quick location functions. The right cataloging field area dynamically loads corresponding cataloging templates according to the document type. The cataloging template includes basic fields such as title, document number, date of issuance, responsible person, and security classification, as well as extended fields. Basic fields are mandatory, while extended fields are configured according to the specific needs of the document type. The automatic extraction module uses optical character recognition (OCR) technology to intelligently recognize the text content of the documents. The OCR engine supports both printed and handwritten recognition modes, with a printed recognition accuracy of over 95% and a handwritten recognition accuracy of over 85%. The recognition results are intelligently matched with the cataloging fields. The matching algorithm combines keyword extraction and pattern recognition. Keyword extraction is based on a predefined dictionary for text analysis, while pattern recognition uses regular expressions to match formatted fields such as document number and date. The manual editing module provides functions for adding, deleting, modifying, and querying fields. Adding fields supports custom field names and data type configurations. Deleting fields requires administrator privileges. Modifying fields supports both batch editing and single-item editing modes. Querying fields supports both fuzzy matching and exact matching search methods.

[0149] The data upload and synchronization layer's task upload module is responsible for uploading completed scanned images, cataloging data, and quality inspection results to the server. Uploading uses a chunked transmission method to improve the efficiency of large file transfers, with each data chunk size set to 2MB. It supports breakpoint resumption to avoid transmission failures caused by network interruptions. Upload status includes four states: waiting to upload, uploading in progress, upload successful, and upload failed. Failed uploads are automatically added to the retry queue, with an exponential backoff strategy for the retry interval. The initial retry interval is 30 seconds, and the maximum number of retries is 5. The progress monitoring module displays real-time statistics for each process, including the number of pending tasks, tasks in progress, completed tasks, and rework tasks. The refresh frequency is set to automatically update every 30 seconds, and manual immediate refresh is also supported. The data synchronization module maintains consistency between the local database and the server database. The synchronization strategy uses incremental synchronization to reduce network transmission volume. Synchronization trigger conditions include scheduled synchronization and event-triggered synchronization. Scheduled synchronization is executed hourly, while event-triggered synchronization is executed immediately upon critical data changes.

[0150] A second aspect of the present invention provides an intelligent data acquisition system for electronic archives, comprising:

[0151] The receiving module is used to receive the original electronic archive data to be processed, perform data deconstruction on the original electronic archive data to separate multimodal information, and generate a hybrid feature vector;

[0152] The filtering module is used to map and locate a set of candidate entities in the semantic embedding space of the knowledge graph based on the hybrid feature vector, and to filter the entity by describing the local subgraph structure of the candidate entity set and integrating the path connectivity and attribute consistency constraints between entities within the subgraph, and output the file entity mapping result.

[0153] The definition module is used to define each entity in the archive entity mapping result as a core node, construct a multi-level relationship network in the knowledge graph, extend the semantic path range through the multi-level relationship network to obtain associated entity clusters, and calculate the implicit association strength based on the path span between each entity in the associated entity cluster and the core node.

[0154] The construction module is used to build a dynamic connection structure across archival entities based on the implicit association strength to form an archival knowledge association system. Through the structural changes of the archival knowledge association system, the association hotspots and sparse areas are depicted to obtain the hierarchical semantic expression structure of the knowledge graph.

[0155] A third aspect of the present invention provides an electronic device, comprising:

[0156] processor;

[0157] Memory used to store processor-executable instructions;

[0158] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0159] 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.

[0160] 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.

[0161] 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 data collection of electronic archives, characterized in that: include: Receive the raw data of the electronic archives to be processed, perform data deconstruction on the raw data of the electronic archives to separate multimodal information, and generate a hybrid feature vector; Based on the hybrid feature vector, a candidate entity set is mapped and located in the semantic embedding space of the knowledge graph. The candidate entity set is then filtered by depicting the local subgraph structure and fusing path connectivity and attribute consistency constraints between entities within the subgraph. The archive entity mapping result is then output. Each entity in the document entity mapping result is defined as a core node. A multi-level relationship network is constructed in the knowledge graph. The semantic path range is extended through the multi-level relationship network to obtain associated entity clusters, including: Each entity in the document entity mapping result is defined as a core anchor node. The first-level entities and relation edges that are directly related are extended in the knowledge graph through the progressive decomposition of entity elements, and the first-level relation network is dynamically constructed. Based on the multidimensional connection features of the first-level relationship network, the second-level entities and relationship edges associated with each first-level entity are linked together, and the first-level relationship network and the second-level entities and relationship edges are hierarchically coupled and fused to generate a multi-level relationship network. The semantic derivation structure of the multi-level relationship network is analyzed, the relationship transmission links from each core anchor node to each level entity are sorted out, the entity node and relationship edge sequence in each relationship transmission link are summarized, and the semantic penetration depth of each relationship transmission link is quantified. Around the core anchor node in the multi-level relationship network, entities at all levels around the periphery are integrated to form a cluster of related entities based on the semantic quantification strength of the relationship links. The implicit association strength is calculated based on the path span between each entity in the cluster and the core node. Based on the implicit association strength, a dynamic connection structure is constructed across archival entities to form an archival knowledge association system. By describing the association hotspots and sparse areas through the structural changes of the archival knowledge association system, a hierarchical semantic expression structure of the knowledge graph is obtained.

2. The method according to claim 1, characterized in that, Based on the hybrid feature vectors, candidate entity sets are mapped and located in the semantic embedding space of the knowledge graph. Then, by depicting the local subgraph structure of the candidate entity sets and fusing path connectivity and attribute consistency constraints between entities within the subgraphs, the results are filtered, and the output archive entity mapping results include: The hybrid feature vector is embedded into the semantic representation space to construct the semantic similarity distribution between the hybrid feature vector and the embedded representation of each entity in the knowledge graph. The density clustering characteristics of the semantic similarity distribution are analyzed to form candidate regions. Entities within the candidate regions are condensed to generate a candidate entity set. Extend each candidate entity node in the candidate entity set, connect adjacent entities and connecting edges in the knowledge graph, and construct a local subgraph structure; By integrating the relation types and directions in the local subgraph structure, the semantic path connectivity from each candidate entity to other candidate entities is quantified. At the same time, the structured attribute constraints in the original electronic archive data are separated, and the structured attribute constraints are compared with the attribute nodes of each candidate entity in the local subgraph structure to form an attribute consistency verification matrix. By integrating the semantic path connectivity and the attribute consistency verification matrix, the comprehensive scores of each candidate entity are aggregated, and the candidate entity with the best score is locked to output the archive entity mapping result.

3. The method according to claim 2, characterized in that, The hybrid feature vector is embedded into the semantic representation space to construct the semantic similarity distribution between the hybrid feature vector and the embedded representations of each entity in the knowledge graph. The density clustering characteristics of the semantic similarity distribution are analyzed to form candidate regions, including: The hybrid feature vector is nonlinearly transformed and mapped to the semantic embedding space to obtain the vector representation of the hybrid feature vector in the semantic embedding space, and a reference vector set is constructed by generating the pre-trained embedding representation of each entity in the knowledge graph. The semantic distance is calculated based on the vector representation and the reference vector set. The semantic similarity metric is constructed using the semantic distance. The semantic similarity metric is then projected into the semantic embedding space to form a semantic similarity distribution. Based on the local density peak regions in the semantic similarity distribution, the spatial concentration and gradient change characteristics of similarity values ​​within the regions are analyzed to determine density clustering characteristics; according to the spatial boundaries and density threshold conditions of the density clustering characteristics, closed regions are located in the semantic embedding space to form candidate regions.

4. The method according to claim 1, characterized in that, Each entity in the document entity mapping result is defined as a core anchor node. Through progressive decomposition of entity elements, directly related first-level entities and relation edges are extended in the knowledge graph, dynamically constructing a first-level relation network including: The structured elements of each entity in the archive entity mapping result are analyzed, and the structured elements are decomposed into attribute dimension set and semantic dimension set. Based on the combined features of the attribute dimension set and the semantic dimension set, the unique node corresponding to each entity is located in the knowledge graph, and the unique node is marked as the core anchor node. In the knowledge graph, query the adjacent nodes that have a direct relationship with the core anchor node, extract the entity identifier and attribute information of the adjacent nodes, and define the adjacent nodes as first-level entities. Extract the relationship edges connecting the core anchor node and each first-level entity, parse the relationship type and semantic direction of each relationship edge, construct relationship edge descriptors, and organize the core anchor node, the first-level entity and the relationship edge descriptors into a topological structure; Based on the connection patterns between the core anchor nodes and the first-level entities in the topology, the semantic association strength of each first-level entity relative to the core anchor node is calculated. The topology is then weighted according to the semantic association strength to form a first-level relationship network.

5. The method according to claim 1, characterized in that, Based on the implicit association strength, a dynamic connection structure is constructed across archival entities to form an archival knowledge association system. The structural changes of this system depict association hotspots and sparse regions, resulting in a hierarchical semantic representation structure of the knowledge graph, including: Based on the implicit association weights of each entity in the associated entity cluster, the shared entity nodes between the associated entity clusters of different core anchor nodes are retrieved. The shared entity nodes and the implicit association weights are used to construct cross-cluster semantic connections. The weight propagation and aggregation calculation of the cross-cluster semantic connections form a dynamic connection structure across archival entities. The dynamic connection structure is integrated into a network topology to obtain the archival knowledge association system. Based on the time-series snapshot acquisition of the aforementioned archive knowledge association system, the topological state at different times is obtained. The nodes and edges with changing connection strength are determined by differential comparison. The clustered areas of nodes with increasing connection strength are defined as association hotspots, and the scattered areas of nodes with decreasing connection strength are defined as sparse areas. Based on the in-degree and out-degree distribution of the associated hotspots, core semantic nodes are located. A hierarchical transmission path is constructed according to the connection relationship between the core semantic nodes and the boundaries of the sparse region. The hierarchical semantic expression structure of the knowledge graph is generated by utilizing the semantic diffusion features of the core semantic nodes on the hierarchical transmission path.

6. The method according to claim 5, characterized in that, By propagating and aggregating the weights of the cross-cluster semantic connections, a dynamic connection structure is formed across archival entities. Integrating this dynamic connection structure into a network topology yields an archival knowledge association system, including: For each cross-cluster semantic connection, the implicit association weights of the nodes at both ends of the connection in their respective associated entity clusters are obtained. Bidirectional propagation operation is performed based on the propagation path of the cross-cluster semantic connection. The weight values ​​are attenuated and modulated according to the path length to obtain the propagation weight sequence. Based on the aggregation operation of each weight component in the propagation weight sequence at the intermediate node of the cross-cluster semantic connection, the cross-cluster aggregation weight is calculated according to the propagation depth and attenuation degree of the weight component, and the cross-cluster aggregation weight is labeled with the corresponding cross-cluster semantic connection to construct a weighted dynamic connection structure. The set of nodes in the dynamic connection structure is mapped to vertices of the network topology, the cross-cluster semantic connection is mapped to directed edges of the network topology, and the cross-cluster aggregation weight is mapped to the weight value of the directed edge. Based on the topological connection relationship between the vertices through the directed edges and their weight values, an archival knowledge association system is constructed.

7. An intelligent data acquisition system for electronic archives, used to implement the method described in any one of claims 1-6, characterized in that, include: The receiving module is used to receive the original electronic archive data to be processed, perform data deconstruction on the original electronic archive data to separate multimodal information, and generate a hybrid feature vector; The filtering module is used to map and locate a set of candidate entities in the semantic embedding space of the knowledge graph based on the hybrid feature vector, and to filter the entity by describing the local subgraph structure of the candidate entity set and integrating the path connectivity and attribute consistency constraints between entities within the subgraph, and output the file entity mapping result. The definition module is used to define each entity in the archive entity mapping result as a core node, construct a multi-level relationship network in the knowledge graph, extend the semantic path range through the multi-level relationship network to obtain associated entity clusters, and calculate the implicit association strength based on the path span between each entity in the associated entity cluster and the core node. The construction module is used to build a dynamic connection structure across archival entities based on the implicit association strength to form an archival knowledge association system. Through the structural changes of the archival knowledge association system, the association hotspots and sparse areas are depicted to obtain the hierarchical semantic expression structure of the knowledge graph.

8. 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 6.

9. 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 6.

Images