Visual intelligent digital archive management method and system

Abstract

The invention relates to the technical field of archive management, in particular to a visual intelligent digital archive management method and system. The method comprises the following steps: acquiring a user retrieval request; analyzing the user retrieval request, and determining retrieval information; determining an associated node of the retrieval information according to a preset knowledge graph; determining a topological structure of the associated node according to the preset knowledge graph; determining a paternal document set and a sub-document set according to the topological structure; and determining a visual evolution structure according to the association node, the paternal document set and the subsystem document set. By establishing an encrypted communication link with a user terminal, a user retrieval request is captured in real time, and the semantic ambiguity rate can be reduced when the retrieval information is determined to process terminologies in an engineering document; and according to the topological structure, the paternal document set and the sub-document set are determined, so that the fuzzy retrieval accuracy can be improved.

Description

Technical Field

[0001] The present application relates to the technical field of archive management, and in particular to a visual intelligent digital archive management method and system. Background Art

[0002] With the accelerated progress of digitization, traditional archive management systems have gradually achieved the digital conversion of paper archives, forming a digital archive system with data perspective capabilities. Existing technologies can initially display the inheritance relationships and citation networks between archives by establishing association maps between documents.

[0003] In the prior art, when a user submits a fuzzy search request, the traditional keyword matching mechanism is difficult to accurately capture the user's true intention, resulting in a decrease in the accuracy of semantic understanding and association judgment. Summary of the Invention

[0004] This application provides a visual intelligent digital archive management method and system to solve the above problems.

[0005] In a first aspect, the present application provides a visual intelligent digital archive management method, which includes: obtaining a user search request; analyzing the user search request to determine search information; determining associated nodes of the search information based on a preset knowledge graph; determining a topological structure of the associated nodes based on the preset knowledge graph; determining a parent document set and a child document set based on the topological structure; and determining a visual evolution structure based on the associated nodes, the parent document set, and the child document set.

[0006] This solution establishes an encrypted communication link with the user terminal to capture original search requests in real time, helping to address the problem of intent drift caused by polysemous terms in the semantic understanding dimension and improving the accuracy of fuzzy searches. It also analyzes user search requests and determines search information, helping to reduce semantic ambiguity when processing specialized terminology in engineering documents. It also identifies associated nodes of the search information based on a pre-set knowledge graph, helping to reduce the mismatch rate of outdated documents while meeting millisecond-level response requirements. It also determines the topological structure of associated nodes based on a pre-set knowledge graph, helping to shorten topological update latency and adapt to the real-time evolution of the archive network. It also reduces logically contradictory edge relationships and ensures the logical consistency of the topological structure. It also determines parent and child document sets based on the topological structure, helping to improve the efficiency of filtering incorrectly associated documents and accelerating document relationship tracing. It also determines a visual evolutionary structure based on associated nodes, parent and child document sets, helping to reduce the overlap of three-dimensional nodes, improve the visual focus of associated nodes, and enhance user recognition efficiency.

[0007] Optionally, determining the associated nodes of the retrieval information based on a preset knowledge graph includes: parsing the retrieval information to determine entity keywords and contextual information; generating a grammatical structure tree based on the entity keywords and the contextual information; obtaining historical retrieval records; determining a personalized semantic mapping table based on the historical retrieval records based on the preset knowledge graph; generating an enhanced semantic vector by fusing the grammatical structure tree, the contextual information, and the personalized semantic mapping table based on an attention mechanism; and determining the associated nodes of the retrieval information based on the enhanced semantic vector based on the preset knowledge graph.

[0008] Optionally, determining the parent document set and the child document set based on the topological structure includes: determining a semantic feature vector based on the archive content of the associated node; determining a retrieval archive node based on the retrieval information; determining a semantic relevance threshold of the retrieval archive node based on the semantic feature vector; determining a branch relevance threshold of the retrieval archive node based on the topological structure; generating a dynamic relevance evaluation function based on the semantic relevance threshold and the branch relevance threshold; and determining the parent document set and the child document set of the retrieval archive node based on the dynamic relevance evaluation function.

[0009] Optionally, determining the semantic feature vector based on the archival content of the associated node includes: constructing a word frequency-inverse document frequency matrix based on the archival content of the associated node; obtaining existing archival data, analyzing the existing archival data, and determining the reference relationship and version change record between documents; determining the document creation time and the document modification timestamp based on the version change record; generating a timeliness weight coefficient based on the document creation time and the document modification timestamp; and determining the semantic feature vector by fusing the word frequency-inverse document frequency matrix, the reference relationship between the documents, the version change record, and the timeliness weight coefficient through a graph neural network model.

[0010] Optionally, determining the branch correlation threshold of the retrieval archive node based on the topological structure includes: retrieving the historical retrieval record based on the retrieval information, and determining the historical retrieval result that is the same as the retrieval information; parsing the historical retrieval result to determine the historical node path; analyzing the historical node path to determine the path node changes; based on the path node changes, counting the node transfer probability in the historical retrieval result; calculating the semantic similarity and time correlation between the retrieval archive node and the associated node; determining the user portrait based on the user retrieval request; adjusting the path weight allocation strategy based on the user portrait; inputting the node transfer probability, the semantic similarity and the time correlation into the LSTM network for time series modeling to generate the branch correlation threshold of the retrieval archive node.

[0011] Optionally, determining the visual evolution structure based on the associated nodes, the parent document set and the child document set includes: generating a three-dimensional force-directed layout based on the associated nodes; allocating morphological features of the visual icons based on the parent document set and the child document set; and generating a visual evolution structure based on the three-dimensional force-directed layout and the morphological features.

[0012] Optionally, generating a three-dimensional force-directed layout based on the associated nodes includes: calculating a first semantic attraction and a first topological repulsion between the associated nodes in the parent document set; calculating a second semantic attraction and a second topological repulsion between the associated nodes in the child document set; determining user attention based on the user portrait; determining a damping adjustment coefficient based on the user attention; setting edge weight constraints based on the first semantic attraction, the first topological repulsion, the second semantic attraction and the second topological repulsion; and generating a three-dimensional force-directed layout based on the damping adjustment coefficient and the edge weight constraints based on a preset physical simulation engine.

[0013] Optionally, the grammatical structure tree, the contextual information, and the personalized semantic mapping table are integrated according to the attention mechanism to generate an enhanced semantic vector, including: performing dependency analysis on the grammatical structure tree, extracting the core predicate-argument structure, and generating a grammatical role vector; encoding the contextual information through a pre-trained language model to generate a context embedding vector; based on the personalized semantic mapping table, calculating the cosine similarity between the user preference and the associated node to generate a personalized weight matrix; using the attention mechanism to cross-modally fuse the grammatical role vector, the context embedding vector, and the personalized weight matrix to output an enhanced semantic vector.

[0014] Optionally, the construction of the preset knowledge graph includes: analyzing the existing archival data to determine the archival spatiotemporal attributes, archival semantic labels and archival topological structure; establishing a three-dimensional spatiotemporal coordinate system of archival metadata based on the archival spatiotemporal attributes, the archival semantic labels and the archival topological structure; marking the domain ontology features of the existing archival data based on the archival spatiotemporal attributes; determining the evolutionary relationship graph of the existing archives based on the archival semantic labels and the archival topological structure; and constructing a preset knowledge graph based on the three-dimensional spatiotemporal coordinate system, the domain ontology features and the evolutionary relationship graph.

[0015] In the second aspect, the present application provides a visual intelligent digital archive management system, which includes: a request analysis module for obtaining user retrieval requests; analyzing the user retrieval requests to determine retrieval information; a node determination module for determining the associated nodes of the retrieval information based on a preset knowledge graph; a topology structure determination module for determining the topological structure of the associated nodes based on the preset knowledge graph; a set determination module for determining the parent document set and the child document set based on the topological structure; and an evolutionary structure determination module for determining a visual evolutionary structure based on the associated nodes, the parent document set and the child document set. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0017] Figure 1 A schematic diagram of an application scenario provided in one embodiment of the present application; Figure 2 A flowchart of a visual intelligent digital archive management method provided in one embodiment of the present application; Figure 3 A schematic diagram of the structure of a visual intelligent digital archive management system provided in one embodiment of the present application. DETAILED DESCRIPTION

[0018] To make the purpose, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0019] In this document, the term "and / or" simply describes a relationship between related objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, or B exists alone. Furthermore, the character " / " in this document, unless otherwise specified, generally indicates an "or" relationship between the related objects.

[0020] The embodiments of the present application are described in further detail below with reference to the accompanying drawings.

[0021] With the accelerated progress of digitization, traditional archive management systems have gradually achieved the digital conversion of paper archives, forming a digital archive system with data perspective capabilities. Existing technologies can initially display the inheritance relationships and citation networks between archives by establishing association maps between documents.

[0022] In the prior art, when a user submits a fuzzy search request, the traditional keyword matching mechanism is difficult to accurately capture the user's true intention, resulting in a decrease in the accuracy of semantic understanding and association judgment.

[0023] Based on this, the present application provides a visual intelligent digital archive management method and system, which obtains user search requests; analyzes user search requests and determines search information; determines the associated nodes of the search information according to a preset knowledge graph; determines the topological structure of the associated nodes according to the preset knowledge graph; determines the parent document set and the child document set according to the topological structure; determines the visual evolution structure according to the associated nodes, the parent document set and the child document set. By establishing an encrypted communication link with the user terminal, the original search request is captured in real time, which helps to solve the problem of intention drift caused by polysemous terms in the semantic understanding dimension and improve the accuracy of fuzzy retrieval; analyzes user search requests and determines search information, which helps to reduce the semantic ambiguity rate when processing professional terms in engineering documents. Determining the associated nodes of the search information according to the preset knowledge graph helps to reduce the mismatch rate of outdated documents and meet the millisecond-level response requirements. Determining the topological structure of the associated nodes according to the preset knowledge graph helps to shorten the update delay of the topological structure, adapt to the real-time evolution of the archive network, and reduce logically contradictory edge relationships to ensure the logical consistency of the topological structure. Determining parent and child document sets based on the topological structure helps improve the efficiency of filtering out incorrectly associated documents and accelerates document relationship tracing. Determining the visual evolutionary structure based on associated nodes, parent and child document sets helps reduce the overlap of three-dimensional nodes, improve the visual focus of associated nodes, and enhance user recognition efficiency.

[0024] Figure 1This application provides a schematic diagram of an application scenario. The method provided herein is applied when a user performs an archive search. Specifically, the method provided herein is applied to any server, interacting with a user terminal. By establishing an encrypted communication link with the user terminal, the server captures user search requests in real time, helping to address the problem of intent drift caused by polysemous terms in the semantic understanding dimension and improving the accuracy of fuzzy searches. The method analyzes user search requests and determines search information, helping to reduce semantic ambiguity when processing specialized terminology in engineering documents. Based on a preset knowledge graph, the associated nodes of the search information are determined, helping to reduce the mismatch rate of outdated documents while meeting millisecond-level response requirements. Based on a preset knowledge graph, the topology of the associated nodes is determined, helping to shorten topology update latency, adapt to the real-time evolution of the archive network, reduce logically contradictory edge relationships, and ensure logical consistency of the topology. Based on the topology, parent and child document sets are determined, helping to improve the efficiency of filtering misassociated documents and accelerating document relationship tracing. Based on the associated nodes, parent and child document sets, a visual evolutionary structure is determined, helping to reduce the overlap rate of three-dimensional nodes, improve the visual focus of associated nodes, and enhance user recognition efficiency.

[0025] For specific implementation methods, please refer to the following embodiments.

[0026] Figure 2 This is a flow chart of a visual intelligent digital archive management method provided by an embodiment of the present application. The method of this embodiment can be applied to the server in the above scenario. Figure 2 As shown, the method includes: S201, obtaining a user search request; analyzing the user search request and determining search information; A user search request can be an archive query instruction containing unstructured sentences input by the user through natural language. The search information can be a multidimensional feature set containing core analytical elements such as enhanced semantic vectors, spatiotemporal tags, and user personalized preference weights.

[0027] Specifically, when users search for archives, they may have a vague understanding of the target content, making it difficult to accurately match the search results. By visualizing the knowledge graph and the topology of the archive structure, a more accurate target archive can be matched. By establishing an encrypted communication link with the user terminal, user search requests can be captured in real time. Using a cross-modal attention mechanism determined by the correlation weights between text feature vectors and user behavior characteristics, the search request undergoes a triple operation of word segmentation, syntactic parsing, and intent classification, outputting structured search information.

[0028] S202: Determine the associated nodes of the search information according to the preset knowledge graph; The preset knowledge graph can be a pre-built archive relationship network. It is pre-stored on the server and called upon when needed. The associated nodes can be document nodes in the preset knowledge graph that have semantic similarity with the search information and meet spatiotemporal constraints.

[0029] Specifically, the three-dimensional feature matrix of the spatiotemporal coordinates, semantic labels and topological connection strength of the preset knowledge graph is called; then, the enhanced retrieval intention vector is projected into the semantic subspace of the knowledge graph, and the semantic similarity threshold with each node is calculated; then, a dynamic sliding window mechanism with adaptive adjustment of the exponential decay factor in the time dimension and the spatial radius constraint is established to screen the node sets that meet the spatiotemporal constraints of the timestamp range and geographic coordinate range and whose semantic similarity exceeds the semantic similarity threshold, and mark them as associated nodes.

[0030] S203. Determine the topological structure of associated nodes according to a preset knowledge graph; The topology structure can be a network connection model that reflects the dynamic evolution relationship between associated nodes.

[0031] Specifically, by extracting the fully connected edge set of the associated nodes, the spatiotemporal evolution parameters of the corresponding edges are loaded; then, by extracting the timestamp sequence and geographic coordinate difference from the preset knowledge graph as the basic input of the time attenuation factor and the spatial attenuation coefficient, a dynamic association evaluation function is constructed to comprehensively calculate the edge weight; subsequently, the incremental graph calculation algorithm is determined by associating the nodes, edge weights, and current snapshot status, and then the incremental graph calculation algorithm is used to dynamically update the topological connection relationship between nodes to generate a topological structure with spatiotemporal labels.

[0032] S204: Determine the parent document set and the child document set based on the topological structure; The parent document set can be an upstream document set in the topology that meets the requirements of version generation time priority, semantic inheritance coherence, and reference frequency. The child document set can be a downstream document set in the topology that is derived from the parent document and passes the reverse semantic conflict detection.

[0033] Specifically, based on the topological structure diagram, the version evolution path of each associated node is traced back along the time axis in reverse, and several upper-level documents that directly or indirectly reference the historical version of the current node are marked to form a parent document set; at the same time, the derivative path of each associated node is traced forward along the time axis, and several lower-level documents that reference the latest version of the current node are captured to form a child document set.

[0034] S205: Determine a visualized evolution structure according to the associated nodes, the parent document set, and the child document set.

[0035] Visualizing the evolutionary structure can be to present the spatiotemporal distribution state of document nodes through a semantic attraction and topological repulsion balance model.

[0036] Specifically, after retrieving the archives, the parent and child documents can be automatically expanded according to the current reading content, and the content relevance of the parent and child documents to the target retrieval is understood and displayed; based on the construction of a three-dimensional visualization space coordinate system, a semantic attraction field in which the topological connection strength is positively correlated with the semantic similarity and a topological repulsion field in which the topological connection strength is positively correlated with the semantic difference between nodes are defined; then, dynamic particles are initialized, and each associated node is mapped into a particle entity with mass attributes, and the particle mass value is proportional to the document importance score; then, an adaptive physical simulation engine is run, and by coupling explicit Verlet integration with adaptive step size control, the numerical integration is determined to solve the particle motion equation, and the particle spatial position is adjusted in real time until the potential energy reaches a stable state; finally, the three-dimensional space-time evolution structure is rendered, and color coding is used to map the changes in the time dimension, and the transparency of the connecting line is used to map the association strength to generate a visualized evolution structure; visualization is the real-time evolution of the parent and child according to real-time retrieval needs. When the retrieval content is different, the parent and child are also different.

[0037] This solution establishes an encrypted communication link with the user terminal to capture original search requests in real time, helping to address the problem of intent drift caused by polysemous terms in the semantic understanding dimension and improving the accuracy of fuzzy searches. It also analyzes user search requests and determines search information, helping to reduce semantic ambiguity when processing specialized terminology in engineering documents. It also identifies associated nodes of the search information based on a pre-set knowledge graph, helping to reduce the mismatch rate of outdated documents while meeting millisecond-level response requirements. It also determines the topological structure of associated nodes based on a pre-set knowledge graph, helping to shorten topological update latency and adapt to the real-time evolution of the archive network. It also reduces logically contradictory edge relationships and ensures the logical consistency of the topological structure. It also determines parent and child document sets based on the topological structure, helping to improve the efficiency of filtering incorrectly associated documents and accelerating document relationship tracing. It also determines a visual evolutionary structure based on associated nodes, parent and child document sets, helping to reduce the overlap of three-dimensional nodes, improve the visual focus of associated nodes, and enhance user recognition efficiency.

[0038] In some embodiments, the retrieval information is parsed to determine entity keywords and contextual information; a grammatical structure tree is generated based on the entity keywords and contextual information; historical retrieval records are obtained; based on a preset knowledge graph, a personalized semantic mapping table is determined based on the historical retrieval records; based on the attention mechanism, the grammatical structure tree, contextual information, and personalized semantic mapping table are integrated to generate an enhanced semantic vector; based on the preset knowledge graph, the associated nodes of the retrieval information are determined based on the enhanced semantic vector.

[0039] Entity keywords can be a collection of named entities such as names of people, places, and timestamps, as well as core domain terms. Contextual information can be a vector generated by the co-occurrence relationship between the characters before and after the entity keywords. The syntax structure can be a hierarchical structure generated based on the abstract syntax tree (AST). Historical search records can be high-frequency search keywords stored in user behavior logs, manually corrected associated node records, and long-term document categories and knowledge graph subdomain data. The personalized semantic mapping table can be a mapping relationship table stored in the form of triples of user ID, entity URI, and preference weight. The attention mechanism can be a fusion algorithm that uses gated weights to dynamically assign contributions to the syntax structure, contextual information, and personalized semantic mapping table. The enhanced semantic vector can be a composite vector generated by the cross-modal attention mechanism that fuses the syntax structure, contextual information, and personalized semantic mapping table.

[0040] Specifically, a dependency parser is used to identify named entities and core domain terms such as names of people, places, and timestamps in the text to form entity keywords. Then, a sliding window is used to capture the co-occurrence relationship between the characters before and after the entity keywords, and a context vector is constructed whose dimensions are aligned with the preset knowledge graph embedding space. Subsequently, based on the abstract syntax tree generation framework, the search information is converted into a syntax structure tree according to the entity keywords and context information. By traversing the syntax tree child nodes and calling the entity linking service of the preset knowledge graph, ambiguous terms are mapped to unique entity URIs in the preset knowledge graph, and the disambiguated syntax structure tree is output. Frequently searched entity keywords and successfully matched entity URIs, user-manually corrected associated node records, and long-term document categories and knowledge graph subdomains are extracted from user behavior logs to generate historical search records. Then, the user's personalized weight matrix is calculated: first, the ratio of the user's private word frequency to the global word frequency is calculated based on the word frequency-inverse document frequency matrix. Second, based on the embedding dimension of the preset knowledge graph entity vector, the user-entity interaction matrix is decomposed into a user latent factor matrix and an entity latent factor matrix using the collaborative filtering algorithm (ALS). The implicit entity association patterns in the user's historical click data are mined by minimizing the square loss function with a regularization term. The user's personalized weight matrix is then encoded into a personalized semantic mapping table. Three attention heads are then defined to process the grammatical structure tree, the context vector, and the personalized semantic mapping table, respectively, to determine modal alignment. Furthermore, a gating mechanism initialized with the Xavier normal distribution is used to dynamically calculate the attention score for each modality through dynamic weighted fusion of features using differentiable gating. Subsequently, the feature vectors of the three modalities are weighted summed to generate an enhanced semantic vector. Finally, the enhanced semantic vector is input into an improved variant of the TransE model, and the cosine similarity with several entity vectors in the preset knowledge graph is calculated based on the knowledge graph embedding model based on the entity-relationship vector space mapping theory. Then, the matching threshold is adaptively adjusted according to the current graph density. Furthermore, the set of entities with similarity higher than the matching threshold is returned as associated nodes.

[0041] Through this solution, the retrieval information is parsed, and entity keywords and contextual information are determined, which helps to eliminate terminology ambiguity, ensure that the extracted entity keywords are consistent with the preset knowledge graph node naming specifications, and avoid matching failures due to naming differences. Based on the entity keywords and contextual information, a grammatical structure tree is generated to clarify the combination of query conditions and eliminate semantic ambiguity. Obtaining historical retrieval records helps to ensure that personalized analysis results are synchronized with the user's latest needs. Based on historical retrieval records, a personalized semantic mapping table is determined, which helps the semantic vector generation process incorporate user-specific cognitive patterns and improve the accuracy of fuzzy query intent inference. According to the attention mechanism, the grammatical structure tree, contextual information, and personalized semantic mapping table are integrated to generate an enhanced semantic vector, which has a significantly improved representation dimension coverage compared to the unimodal vector. Based on the enhanced semantic vector, the associated nodes of the retrieval information are determined, which helps to stabilize the harmonic mean of the recall rate and the precision rate and reduce the false association rate compared to static rules.

[0042] In some embodiments, a semantic feature vector is determined based on the archive content of the associated node; a retrieval archive node is determined based on the retrieval information; a semantic relevance threshold of the retrieval archive node is determined based on the semantic feature vector; a branch relevance threshold of the retrieval archive node is determined based on the topological structure; a dynamic relevance evaluation function is generated based on the semantic relevance threshold and the branch relevance threshold; and a parent document set and a child document set of the retrieval archive node are determined based on the dynamic relevance evaluation function.

[0043] Archive content can be digital information entities such as text data, creation time, version number, and structured tags stored in archive nodes. The semantic feature vector can be an enhanced semantic vector formed by jointly encoding the text semantic features of the archive content and the topological relationship features of the knowledge graph. The retrieval archive node can be the target node that directly matches the user's current search request. The semantic relevance threshold can be a dynamically calculated judgment threshold that reflects the semantic similarity baseline value between several current nodes and the retrieval archive node. The branch relevance threshold can be a judgment threshold generated by linearly weighting the three indicators of the number of directly associated branches of the retrieval node, the maximum path depth, and the node degree centrality. The dynamic relevance evaluation function can be a comprehensive evaluation function used to quantify the comprehensive correlation strength between the candidate node and the retrieval node.

[0044] Specifically, based on the archival content of the associated node, the knowledge graph embedding model is called, and the corresponding semantic feature vector is jointly encoded based on the textual semantic features of the archival content and the topological relationship features of the knowledge graph. Then, based on the search information entered by the user, the entity node with the highest cosine similarity to the semantic feature vector is matched in the preset knowledge graph, and then the entity node is marked as the retrieval archive node. The semantic feature vector of the retrieval archive node is extracted, and the cosine similarity distribution with several entity nodes in the preset knowledge graph is calculated, and the median of the distribution is taken as the semantic relevance threshold. By analyzing the topological structure, the number of directly associated branches, the maximum path depth and the node degree centrality of the retrieval archive node are counted, and the branch relevance threshold is generated through linear weighting. Then, the semantic relevance threshold and the branch relevance threshold are input into the gated fusion unit dynamically adjusted by the user preference weight in the personalized semantic mapping table to generate a dynamic relevance evaluation function. Subsequently, the entity set directly associated with the retrieval archive node in the preset knowledge graph is traversed, and the parent document set and child document set are judged for each candidate node: if the path from the candidate node to the retrieval archive node meets the reverse timing constraint, and the dynamic association evaluation function exceeds the preset parent threshold value obtained based on the statistics of successful cases in the historical retrieval records, then the parent document set is added; at the same time, if the path from the candidate node to the retrieval archive node meets the forward evolution constraint, and the dynamic association evaluation function exceeds the preset child threshold value calculated based on the adaptive calculation of the graph density, then the child document set is added.

[0045] This solution determines semantic feature vectors based on the archival content of associated nodes, helping to address the inability to adapt to the dynamic evolution of archival relationships. It also quantifies the dynamic strength of archival relationships and eliminates the false association rate caused by fixed time windows. Retrieval archival nodes are determined based on retrieval information, avoiding mismatches caused by polysemy in entity keyword searches, thereby improving the accuracy of retrieval node positioning. A semantic relevance threshold is determined for retrieval archival nodes based on the semantic feature vectors, reflecting the baseline semantic relevance strength of archival nodes in the current knowledge graph and overcoming the inability to adapt to semantic evolution. A branch relevance threshold is determined for retrieval archival nodes based on the topological structure, effectively capturing the dynamic evolution of archival relationships. A dynamic relevance evaluation function is generated based on the semantic relevance threshold and branch relevance threshold, helping to strengthen personalized relevance determination, preventing high-connectivity nodes from overly dominating relevance results, and addressing the lack of a collaborative analysis mechanism for spatiotemporal attributes and semantic features. The dynamic relevance evaluation function is used to determine the parent and child document sets for retrieval archival nodes, helping to reduce the false association rate and avoid misjudgments caused by temporary branch documents.

[0046] In some embodiments, a word frequency-inverse document frequency matrix is constructed based on the archival content of the associated nodes; existing archival data is obtained and analyzed to determine the reference relationship and version change records between documents; the document creation time and the document modification timestamp are determined based on the version change records; a timeliness weight coefficient is generated based on the document creation time and the document modification timestamp; through a graph neural network model, the word frequency-inverse document frequency matrix, the reference relationship between documents, the version change records and the timeliness weight coefficient are integrated to determine the semantic feature vector.

[0047] The term frequency-inverse document frequency matrix can be a numerical matrix that represents the weight distribution of terms in a document. Existing archival data can be a collection of documents and metadata stored in a preset knowledge graph. Documents can be independent stored data units. Reference relationships can be logical associations established between documents through explicit tagging. Version change records can be a sequence of timestamps that record several modification operations from the creation of a document to its current state. The document creation time can be the timestamp recorded when the document was first generated and stored. The document modification timestamp can be the timestamp recorded when the document content was updated. The timeliness weight coefficient can be a dynamic weight value that reflects the timeliness of the document content. The graph neural network model can be a deep learning model for processing graph-structured data.

[0048] Specifically, the method extracts text data from the archive content of the associated node, performs word segmentation, and counts the frequency of each word in a single document and its inverse document frequency across multiple documents, generating a word frequency-inverse document frequency matrix that represents the core semantics of the document content. Next, the method traverses the metadata of several nodes in the pre-set knowledge graph, parses explicit reference tags between documents, and constructs reference relationships between documents. Simultaneously, the version number change sequence is associated with branch merge events to generate a time-stamped version change record. Furthermore, the version history chain is extracted from the document metadata, recording the creation timestamp of each document. The modification timestamps of each document recorded in the version change record are parsed and sorted in ascending order by version number. Subsequently, a basic time decay factor is calculated based on the interval between the document creation time and the current time. The timeliness weight is then adjusted based on the density of version modification timestamps. Finally, a normalized timeliness weight coefficient is output using an exponential decay function. Finally, the word frequency-inverse document frequency matrix is input into the node feature layer of the graph neural network; during the message passing process of the graph neural network, the reference relationship weights of adjacent nodes are aggregated; then, the version change records and timeliness weight coefficients are integrated through the temporal attention mechanism; thus, a semantic feature vector that integrates text semantics, reference topology and timeliness features is output.

[0049] This solution constructs a term frequency-inverse document frequency matrix based on the archival content of associated nodes, helping to eliminate semantic disambiguation failures caused by ignoring context in the semantic recognition dimension. Existing archival data is acquired and analyzed to determine inter-document citation relationships and version change records. An explicit inter-document association topology is established, eliminating the inability to adapt to dynamic version evolution and enabling simultaneous modeling of semantic and structural features. Based on version change records, document creation times and modification timestamps are determined, capturing the spatiotemporal evolution of documents, eliminating the problem of ignoring dynamic archival properties and quantifying document activity along the timeline. Timeliness weight coefficients are generated based on document creation times and modification timestamps, helping to dynamically quantify the timeliness value of archival content, eliminating version association bias caused by fixed time windows and avoiding the lag associated with manual threshold setting. A graph neural network model is used to integrate the term frequency-inverse document frequency matrix, inter-document citation relationships, version change records, and timeliness weight coefficients to determine a semantic feature vector. This achieves a three-dimensional fusion of text semantics, citation topology, and timeliness features, breaking through the limitation of a single-dimensional semantic vector representation and improving fuzzy retrieval accuracy.

[0050] In some embodiments, based on the retrieval information, historical retrieval records are retrieved to determine the historical retrieval results that are the same as the retrieval information; the historical retrieval results are parsed to determine the historical node path; the historical node path is analyzed to determine the path node changes; based on the path node changes, the node transfer probability in the historical retrieval results is counted; the semantic similarity and time correlation between the retrieval archive node and the associated node are calculated; based on the user retrieval request, the user portrait is determined; according to the user portrait, the path weight distribution strategy is adjusted; the node transfer probability, semantic similarity and time correlation are input into the LSTM network for time series modeling to generate a branch correlation threshold for the retrieval archive node.

[0051] Historical search results can be a user's past search operation records and the corresponding document node set. A historical node path can be a chronological sequence of document nodes actually visited during a historical search. Path node changes can be sudden changes in the transition direction and path bifurcation between adjacent nodes in a historical node path. Node transition probability can be a quantitative indicator of inter-node transitions. Semantic similarity can be the degree of semantic matching between two nodes. Temporal relevance can be an indicator of the strength of temporal relevance. A user profile can be a labeling model generated by aggregating features of a user's historical search behavior. A path weight allocation strategy can be a weighted ratio rule used to adjust the contribution of semantics, time, and transition probability in relevance calculations. A node transition probability can be a distribution matrix for jumps from the current document node to other related nodes. An LSTM network can be a long-short-term memory neural network.

[0052] Specifically, historical search records matching the current search information are extracted from the user behavior database. The search information is feature-encoded using a hash fingerprint algorithm, and a Bloom filter is used to quickly match historical search records with the same search request. Furthermore, a topological traversal is performed on each historical search result to extract the sequence of document nodes actually browsed by the user, forming a historical node path. Then, historical node paths are compared across different time periods to identify changes in path nodes. First, the number and location distribution of newly added or missing document nodes are counted. Second, the path branch points and their frequency under the same search intent are detected. Finally, the standard deviation of the time intervals between visits to the same node in different paths is calculated. Subsequently, based on the memoryless assumption of Markov chains, a Markov chain state transition matrix is constructed. The transition frequencies between adjacent nodes in the historical node paths are counted, and the node transition probabilities are calculated. Using a pre-trained Doc2Vec model with a distributed memory and distributed bag-of-words framework, the text content of the search archive node and the associated node are mapped into the same vector space through unsupervised learning to calculate semantic similarity. Furthermore, the temporal correlation is calculated using an exponential decay function based on the difference in the last modification timestamps of the two nodes. Based on features such as the user's historical search frequency, commonly used node types, and average visit depth, a user profile containing cognitive preference weights is generated. The path weight allocation strategy is then personalized using the analytic hierarchy process (AHP), focusing on increasing the node transition probability weights for the node types that users focus on. Subsequently, a three-dimensional feature vector sequence is constructed by combining node transition probability, semantic similarity, and temporal correlation. A sliding time window is designed, and the feature sequence of continuous time segments is used as input. This is then fed into an LSTM network for time series modeling. The output layer uses a Sigmoid function that linearly transforms the weights of the LSTM network's hidden layer states. A gradient descent method is used to optimize the correlation distribution features in the fitting training data to generate a dynamic threshold. When the branch correlation exceeds the dynamic threshold, it is determined to be the branch correlation threshold.

[0053] This solution retrieves historical search records based on the search information and identifies historical search results that match the search information. The spatiotemporal correlation of historical search results is leveraged to enhance the ability to capture search intent. Historical search results are parsed to determine historical node paths, providing structured input for dynamic correlation analysis. Analyzing historical node paths and identifying changes in path nodes helps perceive topological changes brought about by version iterations. Based on changes in path nodes, node transition probabilities in historical search results are calculated to reflect the inherent strength of associations between document nodes, eliminating the flaw of static threshold mechanisms that fail to integrate spatiotemporal attributes. The semantic similarity and temporal correlation between search archive nodes and associated nodes are calculated, breaking through the dimensional limitations of the term frequency-inverse document frequency matrix and improving semantic understanding accuracy. Furthermore, the synchronization characteristics of document version iterations are quantified to address the lack of integration of spatiotemporal attributes with semantic evolution. Based on user search requests, user profiles are determined, helping to achieve the core goal of adaptively adjusting user cognitive preferences. Based on user profiles, the path weight allocation strategy is adjusted to address the issue of inefficient associated node identification and optimize the interactive experience in information overload scenarios. The node transfer probability, semantic similarity and time correlation are input into the LSTM network for time series modeling to generate the branch correlation threshold of the retrieval archive node. The threshold setting changes dynamically with the topological structure, replacing the defect of node overlap caused by fixed physical parameters.

[0054] In some embodiments, a three-dimensional force-directed layout is generated based on associated nodes; morphological features of visualization icons are assigned based on parent document sets and child document sets; and a visualization evolution structure is generated based on the three-dimensional force-directed layout and the morphological features.

[0055] The three-dimensional force-directed layout can be a dynamic equilibrium layout model. The visualization icon can be a graphical carrier representing a document node in the visualization evolution structure. The morphological feature can be a set of quantifiable visual attributes.

[0056] Specifically, a semantic attraction matrix is constructed based on the semantic similarity of associated nodes, a temporal repulsion coefficient is generated according to the temporal correlation, and the topological connection strength is calculated in combination with the node transfer probability. A three-dimensional spatial coordinate initialization model is established through the feature fusion theory of "spatiotemporal attributes, semantic labels, and topological structures" in the three-dimensional knowledge representation system. Then, the cognitive preference weights in the user portrait are input into the physical simulation engine, and the multi-dimensional force is used to dynamically adjust the weight ratios of semantic attraction, temporal repulsion, and topological connection force. Subsequently, a mass-spring model constructed through the interactive optimization principle of the force-directed layout algorithm and the archive evolution characteristics is used for spatial iterative calculation. In each round of iteration, the local repulsion strength is automatically adjusted according to the real-time node density distribution until the node overlap rate is reduced to a minimum, thereby generating a three-dimensional force-directed layout. The system traverses the parent document collection and assigns morphological features to the visualization icon: first, the size of a red gradient cube is generated based on the number of document iterations; second, based on the cognitive preference weights in the user profile, a texture map corresponding to the professional field is superimposed on the surface of the red gradient cube. Simultaneously, the system traverses the child document collection and assigns morphological features to the visualization icon: first, the glow intensity of the blue sphere is activated based on whether the relevance threshold is met; second, the transparency of the sphere is dynamically adjusted based on the node transition probability value. Subsequently, the output coordinates of the three-dimensional force-directed layout are spatially mapped to the icon morphological features, and the visual priority of each node is calculated based on the user's real-time viewing direction. Next, the system continuously monitors user interactions with the visualization structure, predicts the next operation intention through an LSTM network, and preloads the associated regions. Finally, when the branch relevance exceeds the dynamic threshold output by the Sigmoid function, the spatiotemporal evolution animation of the child document collection is automatically activated, displaying the version iteration path and key modification points, thereby generating a visualization evolution structure.

[0057] This solution generates a three-dimensional force-directed layout based on associated nodes, eliminating the problem of excessive node overlap in two-dimensional graphs and reducing spatiotemporal confusion errors caused by static layouts. Assigning morphological features to visualization icons based on parent and child document collections helps improve the visual recognition of the lack of integrity in parent document collections and resolves the high false association rate in static rule engines. It also enables visualization conversion of dynamic association evaluation functions and optimizes the interpretability of the document screening process. Generating a visual evolutionary structure based on the three-dimensional force-directed layout and morphological features helps eliminate visual jumps during refresh.

[0058] In some embodiments, a first semantic attraction and a first topological repulsion are calculated between associated nodes in a parent document set; a second semantic attraction and a second topological repulsion are calculated between associated nodes in a child document set; user attention is determined based on a user portrait; a damping adjustment coefficient is determined based on the user attention; edge weight constraints are set based on the first semantic attraction, the first topological repulsion, the second semantic attraction, and the second topological repulsion; based on a preset physical simulation engine, a three-dimensional force-guided layout is generated based on the damping adjustment coefficient and the edge weight constraints.

[0059] The first semantic attraction may be the semantic similarity in the parent document set. The first topological repulsion may be the repulsive force generated by the time difference between the version iterations of the parent document nodes. The second semantic attraction may be the transition probability in the child document set. The second topological repulsion may be the repulsive force generated by the Sigmoid function between the child document nodes. User attention may be the degree of cognitive focus generated by parsing the historical search records in the user portrait. The damping adjustment coefficient may be an adjustment parameter dynamically set based on the standard deviation of the quantified value of user attention. The edge weight constraint may be a set of rules for dynamically adjusting the weights of the semantic attraction and topological repulsion. The preset physical simulation engine may be a preset three-dimensional space layout optimization algorithm, which is pre-stored in the server and called when used.

[0060] Specifically, based on the semantic similarity matrix between nodes in the parent document collection, the first semantic attraction is extracted for each parent node pair. A time decay function is applied based on the time difference between the parent document versions to generate the first topological repulsion. Then, in the child document collection, the second semantic attraction is calculated based on the node transition probability, with the value of the second semantic attraction being positively correlated with the transition probability. A second topological repulsion is generated based on whether the association threshold between child nodes has been met. Subsequently, historical search records in the user profile are parsed, and a cognitive focus vector is output through an LSTM network. This cognitive focus vector is mapped to the feature space of the parent document collection or the child document collection, and the cosine similarity between each node and the focus vector is calculated as the user's attention. Furthermore, the user attention is input into a sigmoid function to generate a dynamic damping adjustment coefficient. The first and second semantic attractions are then normalized to obtain a standardized semantic force. The first and second topological repulsion forces are mapped to the same mechanical dimension to generate a composite repulsion field. Edge weight constraints are established based on the linear combination relationship between the semantic force and the composite repulsion field. Initialize the three-dimensional space coordinate system in the physical simulation engine, and inject the damping adjustment coefficient into the energy conservation equation. Use the explicit Euler integration method for iterative calculations, and each round of iteration is executed: first, update the inter-node forces according to the edge weight constraints; second, detect the Euclidean distance between nodes in real time. When the Euclidean distance between nodes is less than the overlap rate, trigger the local repulsion enhancement mechanism; finally, terminate the iteration when the kinetic energy decays to a stable state and the node overlap rate reaches the minimum, and output the final three-dimensional force-guided layout.

[0061] This solution calculates the first semantic attraction and first topological repulsion between associated nodes in the parent document collection, which helps to break through the limitations of the static time window and effectively suppress the false association caused by outdated historical versions. Calculating the second semantic attraction and second topological repulsion between associated nodes in the child document collection helps to break through the fixed reference threshold defect and reduce the interference rate of related documents. Based on the user portrait, the user's attention is determined to achieve dynamic focusing driven by user intent. Based on the user's attention, the damping adjustment coefficient is determined to help eliminate the problem of visual layout oscillation and improve the iterative convergence speed. Based on the first semantic attraction, first topological repulsion, second semantic attraction, and second topological repulsion, edge weight constraints are set to break through the limitations of incomplete multi-dimensional feature fusion and realize the synergy between spatiotemporal attributes and semantic labels. Based on the preset physical simulation engine, a three-dimensional force-guided layout is generated according to the damping adjustment coefficient and edge weight constraints, avoiding the waste of computing resources caused by a fixed number of iterations and reducing the computing energy consumption of the layout optimization process.

[0062] In some embodiments, dependency analysis is performed on the grammatical structure tree to extract the core predicate-argument structure and generate a grammatical role vector; contextual information is encoded through a pre-trained language model to generate a context embedding vector; based on a personalized semantic mapping table, the cosine similarity between user preferences and associated nodes is calculated to generate a personalized weight matrix; using an attention mechanism, the grammatical role vector, context embedding vector and personalized weight matrix are cross-modally fused to output an enhanced semantic vector.

[0063] The core predicate-argument structure can be the core verb or predicate extracted from the sentence and its associated semantic roles. The grammatical role vector can be a numerical feature vector used to quantify the distribution pattern of grammatical roles in a document. The pretrained language model can be a deep neural network model pretrained on a large-scale corpus through self-supervised learning. The context embedding vector can be a global semantic representation vector. User preferences can be features of a user's personalized needs. Cosine similarity can be a metric that measures the directional consistency of the grammatical role vector and the context embedding vector in vector space.

[0064] The personalized weight matrix may be a probability distribution matrix.

[0065] Specifically, dependency parsing is performed on the input grammatical structure tree to identify the core predicate and its corresponding argument structure, such as agent, patient, time, and place. Each grammatical role is then mapped to a numerical feature vector based on a table of dependency relationship types and role weights. Furthermore, through vector concatenation and normalization, a grammatical role vector representing the document's grammatical structure is generated. A self-supervised pre-trained language model based on the Transformer architecture is used to encode document contextual information through masked language modeling and next sentence prediction tasks on a large corpus, capturing the semantic coherence between adjacent paragraphs. The CLS tag output vector of the pre-trained language model's hidden layer is then extracted as a global context representation. Dimensionality reduction is then performed through a fully connected layer to generate a fixed-dimensional context embedding vector. A personalized semantic mapping table, which records the mapping between terms in a user's historical search behavior and their preferred document categories, is loaded from the user profile. The cosine similarity between the current associated node's keyword set and the user's preferred terms in the personalized semantic mapping table is then calculated to generate a node-level preference matching matrix. This node-level preference matching matrix is then subjected to softmax normalization to form a personalized weight matrix. The grammatical role vector and context embedding vector are aligned by dimension and then concatenated into a joint feature matrix. A multi-head attention mechanism based on the design principles of the standard Transformer architecture is adopted, with a personalized weight matrix as the attention query vector, to calculate the weight distribution of different modal features in the joint feature matrix. Through weighted summation and residual connection operations, multimodal features are fused and an enhanced semantic vector is output.

[0066] Through this solution, dependency analysis is performed on the grammatical structure tree, the core predicate-argument structure is extracted, and a grammatical role vector is generated, which helps to eliminate the problem of insufficient dimensionality of grammatical feature representation caused by the grammatical tree analysis not taking into account the user's personalized semantic mapping. By encoding contextual information through a pre-trained language model and generating a context embedding vector, it helps to eliminate the problem of fuzzy retrieval intention understanding bias caused by ignoring the role of context in semantic disambiguation. Based on the personalized semantic mapping table, the cosine similarity between user preferences and associated nodes is calculated, and a personalized weight matrix is generated, which helps to carry out targeted optimization of the problem of decreased accuracy in semantic recognition and association judgment. Using the attention mechanism, the grammatical role vector, context embedding vector and personalized weight matrix are cross-modally fused to output an enhanced semantic vector, which helps to improve the accuracy of fuzzy retrieval.

[0067] In some embodiments, existing archival data is analyzed to determine the archival spatiotemporal attributes, archival semantic tags, and archival topological structure; a three-dimensional spatiotemporal coordinate system of archival metadata is established based on the archival spatiotemporal attributes, archival semantic tags, and archival topological structure; domain ontology features of existing archival data are annotated based on the archival spatiotemporal attributes; an evolutionary relationship map of existing archives is determined based on the archival semantic tags and archival topological structure; and a preset knowledge map is constructed based on the three-dimensional spatiotemporal coordinate system, domain ontology features, and evolutionary relationship map.

[0068] Archival spatiotemporal attributes can include spatial information such as document creation or modification timestamps, geographic coordinates, and institutional location. Archival semantic tags can be keywords or classification labels describing the subject matter of the archive, such as legal documents or medical reports. Archival topology can include logical relationships between archives, such as version inheritance, references, and affiliations. Archival metadata can be structured data representing the essential attributes of the archive. A three-dimensional spatiotemporal coordinate system can be a coordinate system of a three-dimensional mathematical space. Domain ontology features can be semantic path descriptions used to enhance the domain adaptability of semantic parsing. Existing archives can be historical documents that have been digitized and stored in the system. An evolutionary relationship graph can be a weighted directed graph composed of archives as nodes and associations as edges.

[0069] Specifically, timestamp sequences and spatial information, including document creation and modification timestamps, are extracted from existing archival data to construct archival spatiotemporal attributes. Natural language processing is then used to identify entities and categorize themes within the archival content, generating archival semantic tags containing entities, manual annotations, or predefined classification systems. Subsequently, logical relationships between archives, such as version inheritance, references, and affiliations, are determined based on hyperlinks, database foreign keys, and internal document references. Based on these logical relationships and user operation logs, an archival topology is constructed, characterizing the document association strength and evolutionary paths of parent and child document collections. The timestamp sequence is converted into relative time axis coordinates, and the spatial information is converted into three-dimensional spatial coordinates to form a spatiotemporal attribute embedding vector. Furthermore, the archival semantic tags are converted into numerical vectors using one-hot encoding and concatenated with the spatiotemporal attribute embedding vectors by dimension to generate a three-dimensional spatiotemporal coordinate system. A predefined domain ontology model is loaded, and the concept hierarchical relationships and attribute constraints of the domain ontology model are analyzed. The time series features in the archival spatiotemporal attributes are matched with the temporal constraints of the domain ontology, and the domain ontology features are annotated. Based on the archival topology, the direct reference relationships between parent and child document collections are identified and initial association edges are generated. Dynamic evolution modeling is then performed. First, based on the Lancaster information aging curve, the distribution of historical version access intervals in user operation logs is statistically analyzed. A time decay function is introduced to calculate the weight decay coefficient of historical version association edges. Second, the semantic similarity weights of association edges are dynamically updated based on the co-occurrence frequency of semantic tags and the overlap of time windows. Finally, the initial association edges are integrated with the dynamic evolution modeling to generate an evolutionary relationship graph with dual spatiotemporal and semantic weights. The coordinate data in the three-dimensional spatiotemporal coordinate system, the domain ontology feature annotation results, and the edge weights of the evolutionary relationship graph are input into the feature fusion module. Subsequently, a graph neural network is used for multimodal feature alignment. First, the three-dimensional spatiotemporal coordinate system is encoded into initial node feature vectors. Second, the domain ontology feature category labels are mapped into node type embedding vectors. Finally, the edge weights of the evolutionary relationship graph are converted into edge features for the graph attention mechanism. Finally, by jointly encoding the graph structure data and feature vectors, a pre-defined knowledge graph containing spatiotemporal, semantic, and topological features is output.

[0070] This solution analyzes existing archival data to determine its spatiotemporal attributes, semantic labels, and topological structure. This helps eliminate the problem of incomplete feature representation caused by relying solely on a single temporal or semantic dimension, as well as the ambiguity of polysemous words in cross-domain scenarios. It also improves the accuracy of semantic parsing and eliminates the problem of static rule engines being unable to capture dynamic association networks, providing a data foundation for determining evolutionary relationships. A three-dimensional spatiotemporal coordinate system for archival metadata is established based on its spatiotemporal attributes, semantic labels, and topological structure, helping to achieve a unified mathematical representation of multi-source spatiotemporal attributes. Based on its spatiotemporal attributes, domain ontology features are annotated for existing archival data, helping to eliminate semantic mapping biases caused by a lack of domain knowledge in syntax tree analysis methods and providing a traceable semantic association basis for cross-modal attention mechanisms. Based on its semantic labels and topological structure, an evolutionary relationship map of existing archives is determined, reducing the rate of false associations and quantifying the evolutionary strength and direction between archives, eliminating the limitation of only being able to determine binary associations. Constructing a preset knowledge graph based on the three-dimensional space-time coordinate system, domain ontology characteristics and evolutionary relationship graph can help eliminate the problem of incomplete feature extraction caused by the lack of a three-dimensional knowledge representation system, while reducing the response time of fuzzy retrieval.

[0071] Figure 3 This is a structural diagram of a visual intelligent digital archive management system provided by an embodiment of the present application, such as Figure 3 As shown, the visual intelligent digital archive management system 300 of this embodiment includes: a request analysis module 301, a node determination module 302, a topology structure determination module 303, a set determination module 304, and an evolution structure determination module 305.

[0072] The request analysis module 301 is used to obtain a user search request; analyze the user search request and determine the search information; A node determination module 302 is configured to determine associated nodes of the search information based on a preset knowledge graph; A topology structure determination module 303 is configured to determine the topology structure of the associated nodes based on the preset knowledge graph; A collection determination module 304 is configured to determine a parent document collection and a child document collection according to the topological structure; The evolution structure determination module 305 is configured to determine a visual evolution structure according to the associated nodes, the parent document set, and the child document set.

[0073] Optionally, when the node determination module 302 determines the associated nodes of the search information according to a preset knowledge graph, it is configured to: Parsing the search information to determine entity keywords and contextual information; Generate a grammatical structure tree according to the entity keywords and the context information; Obtaining historical search records; determining a personalized semantic mapping table based on the preset knowledge graph and the historical search records; According to the attention mechanism, the syntax structure tree, the context information, and the personalized semantic mapping table are integrated to generate an enhanced semantic vector; Based on the preset knowledge graph and according to the enhanced semantic vector, the associated nodes of the retrieval information are determined.

[0074] Optionally, when the set determination module 304 determines the parent document set and the child document set according to the topological structure, it is configured to: Determining a semantic feature vector based on the archive content of the associated node; Determining a retrieval archive node based on the retrieval information; Determining a semantic relevance threshold of the retrieval archive node based on the semantic feature vector; Determining a branch relevance threshold of the retrieval archive node according to the topological structure; generating a dynamic relevance evaluation function according to the semantic relevance threshold and the branch relevance threshold; According to the dynamic relevance evaluation function, the parent document set and the child document set of the retrieval archive node are determined.

[0075] Optionally, when determining the semantic feature vector based on the archive content of the associated node, the set determination module 304 is configured to: Constructing a term frequency-inverse document frequency matrix based on the archive content of the associated node; Acquire existing archival data, analyze the existing archival data, and determine reference relationships and version change records between documents; Determine the document creation time and document modification timestamp based on the version change record; Generate a timeliness weight coefficient according to the document creation time and the document modification timestamp; Through a graph neural network model, the word frequency-inverse document frequency matrix, the reference relationship between the documents, the version change record and the timeliness weight coefficient are integrated to determine a semantic feature vector.

[0076] Optionally, when the set determination module 304 determines the branch relevance threshold of the retrieval archive node according to the topological structure, it is configured to: Retrieving the historical search records based on the search information to determine the historical search results that are identical to the search information; Analyze the historical search results to determine the historical node path; Analyze the historical node path and determine the path node changes; Based on the path node changes, counting the node transition probabilities in the historical search results; Calculating the semantic similarity and temporal correlation between the retrieved archive node and the associated node; Determining a user profile based on the user search request; Adjusting the path weight distribution strategy based on the user profile; The node transition probability, the semantic similarity and the temporal correlation are input into the LSTM network for time series modeling to generate a branch correlation threshold of the retrieval archive node.

[0077] Optionally, when the evolutionary structure determination module 305 determines the visualized evolutionary structure according to the associated nodes, the parent document set, and the child document set, it is configured to: generating a three-dimensional force-directed layout according to the associated nodes; Allocating morphological features of visualization icons according to the parent document set and the child document set; A visual evolution structure is generated according to the three-dimensional force-directed layout and the morphological characteristics.

[0078] Optionally, when the evolved structure determination module 305 generates a three-dimensional force-directed layout according to the associated nodes, it is configured to: Calculating a first semantic attraction and a first topological repulsion between associated nodes in the parent document set; Calculating a second semantic attraction and a second topological repulsion between associated nodes in the subset document set; Determining user attention based on the user portrait; Determining a damping adjustment coefficient according to the user attention; Setting an edge weight constraint condition according to the first semantic attraction, the first topological repulsion, the second semantic attraction, and the second topological repulsion; Based on a preset physical simulation engine, a three-dimensional force-directed layout is generated according to the damping adjustment coefficient and the edge weight constraint condition.

[0079] Optionally, when the node determination module 302 generates an enhanced semantic vector by fusing the syntax structure tree, the context information, and the personalized semantic mapping table according to the attention mechanism, it is used to: Performing dependency analysis on the grammatical structure tree, extracting the core predicate-argument structure, and generating a grammatical role vector; Encoding the context information through a pre-trained language model to generate a context embedding vector; Based on the personalized semantic mapping table, calculating the cosine similarity between the user preference and the associated node to generate a personalized weight matrix; An attention mechanism is used to cross-modally fuse the grammatical role vector, the context embedding vector, and the personalized weight matrix to output an enhanced semantic vector.

[0080] Optionally, the visual intelligent digital archive management system 300 further includes a graph construction module 306 for: Analyze the existing archival data to determine the temporal and spatial attributes, semantic tags, and topological structure of the archives; Establishing a three-dimensional spatiotemporal coordinate system of archive metadata according to the archive spatiotemporal attributes, the archive semantic tags, and the archive topological structure; According to the temporal and spatial attributes of the archives, marking the domain ontology features of the existing archive data; Determining an evolutionary relationship graph of existing archives based on the archive semantic tags and the archive topological structure; A preset knowledge graph is constructed based on the three-dimensional space-time coordinate system, the domain ontology characteristics and the evolutionary relationship graph.

[0081] The system of this embodiment can be used to execute the method of any of the above embodiments. Its implementation principles and technical effects are similar and will not be described in detail here.

Claims

1. A visual intelligent digital archive management method, characterized in that: include: Get user search request; Analyzing the user search request and determining search information; Determine the associated nodes of the search information according to a preset knowledge graph; Determining the topological structure of the associated nodes according to the preset knowledge graph; Determine a parent document set and a child document set according to the topological structure; A visual evolution structure is determined according to the associated nodes, the parent document set, and the child document set.

2. The method according to claim 1, characterized in that The determining of the associated nodes of the search information according to the preset knowledge graph includes: Parsing the search information to determine entity keywords and contextual information; Generate a grammatical structure tree according to the entity keywords and the context information; Obtaining historical search records; determining a personalized semantic mapping table based on the preset knowledge graph and the historical search records; According to the attention mechanism, the syntax structure tree, the context information, and the personalized semantic mapping table are integrated to generate an enhanced semantic vector; Based on the preset knowledge graph and according to the enhanced semantic vector, the associated nodes of the retrieval information are determined.

3. The method according to claim 2, characterized in that Determining the parent document set and the child document set according to the topological structure includes: Determining a semantic feature vector based on the archive content of the associated node; Determining a retrieval archive node based on the retrieval information; Determining a semantic relevance threshold of the retrieval archive node based on the semantic feature vector; Determining a branch relevance threshold of the retrieval archive node according to the topological structure; generating a dynamic relevance evaluation function according to the semantic relevance threshold and the branch relevance threshold; According to the dynamic relevance evaluation function, the parent document set and the child document set of the retrieval archive node are determined.

4. The method according to claim 3, characterized in that Determining the semantic feature vector according to the archive content of the associated node includes: Constructing a term frequency-inverse document frequency matrix based on the archive content of the associated node; Acquire existing archival data, analyze the existing archival data, and determine reference relationships and version change records between documents; Determine the document creation time and document modification timestamp based on the version change record; Generate a timeliness weight coefficient according to the document creation time and the document modification timestamp; Through a graph neural network model, the word frequency-inverse document frequency matrix, the reference relationship between the documents, the version change record and the timeliness weight coefficient are integrated to determine a semantic feature vector.

5. The method according to claim 3, characterized in that Determining the branch relevance threshold of the retrieval archive node according to the topological structure includes: Retrieving the historical search records based on the search information to determine the historical search results that are identical to the search information; Analyze the historical search results to determine the historical node path; Analyze the historical node path and determine the path node changes; Based on the path node changes, counting the node transition probabilities in the historical search results; Calculating the semantic similarity and temporal correlation between the retrieved archive node and the associated node; Determining a user profile based on the user search request; Adjusting the path weight distribution strategy based on the user profile; The node transition probability, the semantic similarity and the temporal correlation are input into the LSTM network for time series modeling to generate a branch correlation threshold of the retrieval archive node.

6. The method according to claim 1, characterized in that The determining of the visual evolution structure according to the associated node, the parent document set, and the child document set includes: generating a three-dimensional force-directed layout according to the associated nodes; Allocating morphological features of visualization icons according to the parent document set and the child document set; A visual evolution structure is generated according to the three-dimensional force-directed layout and the morphological characteristics.

7. The method according to any one of claims 5 or 6, characterized in that Generating a three-dimensional force-directed layout according to the associated nodes includes: Calculating a first semantic attraction and a first topological repulsion between associated nodes in the parent document set; Calculating a second semantic attraction and a second topological repulsion between associated nodes in the subset document set; Determining user attention based on the user portrait; Determining a damping adjustment coefficient according to the user attention; Setting an edge weight constraint condition according to the first semantic attraction, the first topological repulsion, the second semantic attraction, and the second topological repulsion; Based on a preset physical simulation engine, a three-dimensional force-directed layout is generated according to the damping adjustment coefficient and the edge weight constraint condition.

8. The method according to claim 2, characterized in that The step of fusing the syntax structure tree, the context information, and the personalized semantic mapping table according to the attention mechanism to generate an enhanced semantic vector includes: Performing dependency analysis on the grammatical structure tree, extracting the core predicate-argument structure, and generating a grammatical role vector; Encoding the context information through a pre-trained language model to generate a context embedding vector; Based on the personalized semantic mapping table, calculating the cosine similarity between the user preference and the associated node to generate a personalized weight matrix; An attention mechanism is used to cross-modally fuse the grammatical role vector, the context embedding vector, and the personalized weight matrix to output an enhanced semantic vector.

9. The method according to claim 4, characterized in that The construction of the preset knowledge graph includes: Analyze the existing archival data to determine the temporal and spatial attributes, semantic tags, and topological structure of the archives; Establishing a three-dimensional spatiotemporal coordinate system of archive metadata according to the archive spatiotemporal attributes, the archive semantic tags, and the archive topological structure; According to the temporal and spatial attributes of the archives, marking the domain ontology features of the existing archive data; Determining an evolutionary relationship graph of existing archives based on the archive semantic tags and the archive topological structure; A preset knowledge graph is constructed based on the three-dimensional space-time coordinate system, the domain ontology characteristics and the evolutionary relationship graph.

10. A visual intelligent digital archive management system, characterized in that: The method according to any one of claims 1 to 9, characterized in that it includes: A request analysis module is used to obtain a user search request; analyze the user search request and determine the search information; A node determination module, configured to determine the associated nodes of the search information according to a preset knowledge graph; A topology structure determination module, configured to determine the topology structure of the associated nodes based on the preset knowledge graph; A collection determination module, configured to determine a parent document collection and a child document collection according to the topological structure; The evolution structure determination module is used to determine the visual evolution structure according to the associated nodes, the parent document set and the child document set.

Images