A digital asset retrieval method based on big data processing

Abstract

The application discloses a kind of digital asset retrieval methods based on big data processing, it is related to digital asset management technical field, including the following steps: S1, obtain digital asset and associated data, obtain digital asset multi-source fusion dataset;S2, construct digital asset time sequence knowledge graph;S3, input the graph and the improved BLIP-2 model of search request, obtain asset semantic matching result;S4, calculate ownership trust score, obtain trusted candidate asset set;S5, disturb search request, calculate search stability evaluation value;S6, the above result is input DPP algorithm, and target digital asset set is filtered by evidence coverage constraint;S7, form digital asset retrieval result.The application is improved by ownership evidence guidance, time sequence semantic interaction and evidence coverage constraint, and the accuracy of search, ownership trustworthiness and result diversity are improved.

Description

Technical Field

[0001] This invention relates to the field of digital asset management technology, and in particular to a digital asset retrieval method based on big data processing. Background Technology

[0002] With the continuous expansion of digital content production, copyright transactions, and digital operations, the demand for unified management and rapid retrieval of digital assets such as images, videos, text, and design files is constantly growing. Existing digital asset retrieval methods mainly rely on file names, manual tags, keyword matching, or single content vector similarity for result ranking. However, these methods generally suffer from the following problems in practical applications:

[0003] Digital assets and related data from different sources differ in data format, field names, time records, and asset identifiers, which can easily lead to duplicate entries of the same digital asset, broken relationships, and uncontinuous tracking of ownership records. Existing multimodal retrieval models mainly focus on the semantic relevance between the retrieval request and the content of the digital asset, making it difficult to simultaneously utilize information on ownership subject, authorization status, transfer relationship, and time evolution. This results in digital assets with similar content but invalid authorization, incomplete ownership, or inconsistent subject records still being ranked higher. In the process of screening large-scale candidate results, traditional ranking methods usually select one by one based on a single similarity, without fully considering the content duplication and overlapping evidence paths between candidate assets. This can easily return multiple highly similar digital assets with duplicate ownership evidence, increasing the cost of manual review and reducing the credibility and diversity of retrieval results.

[0004] Therefore, how to provide a digital asset retrieval method based on big data processing is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a digital asset retrieval method based on big data processing. This invention employs a combination of a digital asset temporal knowledge graph, an improved BLIP-2 model, and the DPP algorithm to standardize digital assets and related data and construct data associations. It guides the mapping of graph relationship weights through ownership evidence, and integrates request features, asset features, and temporal features through a temporal semantic interaction structure to obtain asset semantic matching results. It combines ownership credibility scores and retrieval stability evaluation values ​​to screen a set of credible candidate assets, and corrects the candidate asset kernel matrix through evidence coverage constraints to obtain the target digital asset set. This reduces the number of digital assets with expired authorizations, incomplete ownership records, and duplicate content entering the retrieval results, offering advantages such as high retrieval accuracy, high ownership credibility, and good result diversity.

[0006] A digital asset retrieval method based on big data processing according to an embodiment of the present invention includes the following steps:

[0007] S1. Acquire digital assets and related data, and obtain a multi-source fusion dataset of digital assets through standardized processing;

[0008] S2. Construct a digital asset time-series knowledge graph based on the data association relationships in the multi-source fusion dataset of the digital assets;

[0009] S3. Input the retrieval request and the digital asset temporal knowledge graph into the improved BLIP-2 model, map the graph relationship weights through the ownership evidence guidance layer, and fuse request features, asset features and temporal features through the temporal semantic interaction structure to obtain the asset semantic matching result;

[0010] S4. Calculate the ownership credibility score based on the asset semantic matching result and the ownership association data in the digital asset time-series knowledge graph to obtain a set of credible candidate assets;

[0011] S5. Perform conditional perturbation on the search request, and calculate the search stability evaluation value based on the change in the ranking of the set of credible candidate assets before and after the perturbation;

[0012] S6. Input the asset semantic matching result, the ownership credibility score and the retrieval stability evaluation value into the DPP algorithm, construct the candidate asset kernel matrix through evidence coverage constraints, and jointly screen the credible candidate asset set to obtain the target digital asset set.

[0013] S7. Based on the target digital asset set, generate and output digital asset retrieval results.

[0014] Optionally, S1 specifically includes:

[0015] Acquire digital assets and related data, and extract digital asset content, asset attribute records, and related records respectively;

[0016] The digital asset content, the asset attribute records, and the associated records are format-converted, field-mapped, and time-unified to obtain standardized asset records.

[0017] Extract the asset identifier from the standardized asset record. When the asset identifier is missing, perform a hash calculation on the corresponding digital asset content and use the hash value as the asset identifier.

[0018] Standardized asset records belonging to the same digital asset are merged according to the asset identifier, and duplicate records are deleted to obtain a multi-source fusion dataset of digital assets.

[0019] Optionally, S2 specifically includes:

[0020] The digital asset multi-source fusion dataset is divided into asset record groups according to the asset identifier, and each asset record group is mapped to an asset node.

[0021] Extract the subject records and behavior records associated with the digital assets from the asset record group, and map the subject records and behavior records as associated nodes;

[0022] Establish node relationship edges based on the field correspondence between the asset record group, the subject record, and the behavior record, and convert the time field associated with the node relationship edge into a unified time sequence number;

[0023] Arrange the node relationship edges connected to the same asset node according to the unified time sequence number to form an asset relationship time sequence chain;

[0024] By aggregating the asset nodes, the associated nodes, the node relationship edges, and the asset relationship time-series chains, a digital asset time-series knowledge graph is constructed.

[0025] Optionally, the improved BLIP-2 model includes a multimodal input encoding module, a query bridging module, an ownership evidence guiding layer, a temporal semantic interaction structure, and a matching calculation module. The ownership evidence guiding layer and the temporal semantic interaction structure constitute the improved structure of the improved BLIP-2 model.

[0026] The retrieval request and the digital asset time-series knowledge graph are input into the multimodal input encoding module to extract the request content and retrieval conditions from the retrieval request. The request content is converted into a request word sequence, and the retrieval conditions are converted into a condition feature vector.

[0027] Extract asset nodes, associated nodes, node relationship edges, and asset relationship time-series chains from the digital asset time-series knowledge graph. Convert the content fields corresponding to the asset nodes and associated nodes into node feature sequences, and convert the node relationship edges and asset relationship time-series chains into relationship feature sequences to obtain multimodal input features.

[0028] The request term sequence, the conditional feature vector, and the multimodal input features are input into the query bridging module. Self-attention calculation is performed on the request term sequence, and cross-attention calculation is performed using the request term sequence as the query term and the multimodal input features as the key and value terms.

[0029] The cross-attention calculation result is concatenated with the conditional feature vector dimension by dimension, and linearly mapped to obtain the query bridging feature vector corresponding to each asset node. The vectors are then arranged in the order of the asset nodes to obtain the query bridging feature sequence.

[0030] The query bridging feature sequence and the digital asset time-series knowledge graph are input into the ownership evidence guidance layer. The evidence association value of each asset node is calculated according to the relationship type and connection path corresponding to the node relationship edge. The evidence association value is divided by the sum of the evidence association values ​​to obtain the graph relationship weight. The graph relationship weight is multiplied dimension by dimension by the query bridging feature vector of the corresponding asset node to obtain the ownership guidance feature sequence.

[0031] The ownership guidance feature sequence and the asset relationship time-series chain are input into the time-series semantic interaction structure. The time interval between adjacent node relationship edges is calculated according to the unified time sequence number. The ownership guidance feature sequence is subjected to time decay weighting. The weighting result is cross-attention calculation with the corresponding relationship feature vector to obtain the time-series interaction feature sequence.

[0032] The temporal interaction feature sequence is input into the matching calculation module, and the vector similarity between the retrieval request and each asset node is calculated respectively. The vector similarity corresponding to the same asset node is accumulated and normalized. The asset nodes are arranged according to the normalization result to obtain the asset semantic matching result.

[0033] Optionally, inputting the query bridging feature sequence and the digital asset time-series knowledge graph into the ownership evidence guiding layer specifically involves:

[0034] Receive the query bridging feature sequence and the digital asset time-series knowledge graph, extract the node relationship edges between each asset node and associated nodes, and determine the connection order of each node relationship edge according to the asset relationship time-series chain;

[0035] Starting with each asset node, the connection path is extracted along the node relationship edge, and the connection paths corresponding to the same asset node are aggregated into an evidence path set.

[0036] Count the number of node relationship edges and the number of relationship types contained in each connection path, and divide the number of relationship types by the number of node relationship edges to obtain the relationship coverage value;

[0037] Calculate the time interval between adjacent node relationship edges according to the unified time sequence number, divide each time interval by the preset time interval upper limit, adjust the ratio greater than one to one, and obtain the interval ratio; calculate the average value of all interval ratios corresponding to the same connection path, and subtract the average value from one to obtain the time continuity value.

[0038] Add the relationship coverage value and the time continuity value corresponding to the same connection path and divide by two to obtain the path evidence value. Add up all the path evidence values ​​corresponding to the same asset node to obtain the evidence association value.

[0039] Divide the evidence association value corresponding to each asset node by the sum of all evidence association values ​​to obtain the graph relationship weight;

[0040] Extract the query bridging feature vector corresponding to each asset node, multiply the graph relationship weight with the corresponding query bridging feature vector dimension by dimension, and arrange them according to the order of asset nodes to obtain the ownership guidance feature sequence.

[0041] Optionally, S4 specifically includes:

[0042] Read the asset semantic matching results, extract the semantic matching value corresponding to each asset node, and divide each semantic matching value by the sum of all semantic matching values ​​to obtain the normalized semantic matching value;

[0043] Extract ownership association data corresponding to each asset node from the digital asset time-series knowledge graph, and connect the corresponding node relationship edges according to the subject identifier and time sequence in the ownership association data. Form ownership association paths by connecting the node relationship edges that can be continuously connected.

[0044] The number of node relationship edges included in the ownership association path is counted, and the number of node relationship edges is divided by the total number of node relationship edges included in the ownership association data of the corresponding asset node to obtain the complete value of the ownership path.

[0045] The effective time value is obtained by dividing the number of records in the ownership association data that are within the effective time range by the total number of records in the ownership association data.

[0046] The number of records with the same ownership subject in the ownership association data is counted. The number of records corresponding to the ownership subject with the largest number is divided by the total number of records in the ownership association data to obtain the record consistency value.

[0047] The ownership credibility score is obtained by adding the normalized semantic matching value, the ownership path integrity value, the time validity value, and the record consistency value corresponding to the same asset node and then dividing by four.

[0048] Asset nodes whose ownership trust scores reach a preset trust threshold are retained and sorted in descending order according to the ownership trust scores to obtain a set of trustworthy candidate assets.

[0049] Optionally, S5 specifically includes:

[0050] Parse the retrieval request, extract the retrieval condition items that limit the retrieval scope in the retrieval request, and record the original sorting position of each asset node in the set of trusted candidate assets;

[0051] Each search condition item is either deleted or adjusted, with only one search condition item changed at a time, resulting in multiple perturbation search requests.

[0052] Each perturbation retrieval request and the digital asset temporal knowledge graph are input into the improved BLIP-2 model to obtain the asset semantic matching results corresponding to each perturbation retrieval request;

[0053] The ownership credibility score is recalculated based on the asset semantic matching results and ownership association data corresponding to each perturbation retrieval request, and the asset nodes are arranged according to the recalculated ownership credibility score to obtain the perturbation sorting sequence.

[0054] Extract the perturbation ranking position of each asset node in the set of trusted candidate assets in the perturbation ranking sequence; when an asset node does not appear in the perturbation ranking sequence, add one to the total number of asset nodes in the perturbation ranking sequence as the corresponding perturbation ranking position;

[0055] Calculate the absolute difference between the original sort position and each disturbed sort position of the same asset node;

[0056] The total number of asset nodes in the set of trusted candidate assets is compared with the total number of asset nodes in the corresponding perturbation sorting sequence plus one, and the larger value is taken as the sorting base; each absolute difference is divided by the corresponding sorting base to obtain the normalized sorting difference.

[0057] Calculate the average of all normalized sort differences corresponding to the same asset node, and subtract the average from one to obtain the retrieval stability evaluation value of the corresponding asset node.

[0058] Optionally, S6 specifically includes:

[0059] The asset semantic matching result, the ownership credibility score and the retrieval stability evaluation value are input into the DPP algorithm to extract the semantic matching value corresponding to each asset node in the set of credible candidate assets.

[0060] Divide the semantic matching value, the ownership credibility score, and the retrieval stability evaluation value by the sum of their corresponding values, and add the three ratios corresponding to the same asset node and divide by three to obtain the candidate asset quality value.

[0061] Extract the node feature sequence corresponding to each asset node from the digital asset time-series knowledge graph, and then sum the feature vectors in the node feature sequence dimension by dimension and divide by the number of feature vectors to obtain the candidate asset feature vector.

[0062] Multiply the values ​​of the corresponding dimensions of the feature vectors of the two candidate assets one dimension at a time and sum them up. Divide the summed result by the product of the magnitudes of the feature vectors of the two candidate assets to obtain the similarity of the candidate assets.

[0063] The evidence overlap value between two asset nodes is calculated based on the evidence coverage constraint. The evidence overlap value is added to the corresponding candidate asset similarity and then divided by two to obtain the initial corrected candidate asset similarity.

[0064] The square of the candidate asset quality value is written into the main diagonal position of the candidate asset kernel matrix, and the product of the two candidate asset quality values ​​and the corresponding initial modified candidate asset similarity is written into the non-main diagonal position to construct the candidate asset kernel matrix.

[0065] The candidate asset kernel matrix is ​​decomposed by eigenvalue decomposition, and the eigenvalues ​​less than zero are adjusted to zero. The candidate asset kernel matrix is ​​then reconstructed based on the adjusted eigenvalues.

[0066] Add the asset node with the highest candidate asset quality value to the selected asset set, and add the remaining asset nodes to the principal sub-matrix of the candidate asset core matrix corresponding to the selected asset set. Calculate the difference in determinant between the principal sub-matrix of the candidate asset core matrix before and after adding them.

[0067] The asset node with the largest determinant difference is added to the selected asset set, and the candidate asset kernel matrix is ​​updated according to the evidence coverage constraint. The process of adding asset nodes and updating the candidate asset kernel matrix is ​​repeated until the selected asset set reaches the preset number of assets or all determinant differences are not greater than zero. The selected asset set is then determined as the target digital asset set.

[0068] Optionally, updating the candidate asset kernel matrix based on the evidence coverage constraint, and repeatedly performing asset node addition and candidate asset kernel matrix update, specifically involves:

[0069] Using each asset node in the set of trusted candidate assets as the starting point of the path, extract the set of evidence paths along the node relationship edges in the digital asset time-series knowledge graph.

[0070] The starting node identifier, ending node identifier, and relationship type of each node relationship edge are combined into an evidence edge identifier. The evidence edge identifiers corresponding to the same asset node are collected to obtain the evidence edge identifier set.

[0071] Perform intersection and union operations on the evidence edge identifier sets corresponding to two asset nodes, and divide the number of evidence edge identifiers in the intersection by the number of evidence edge identifiers in the union to obtain the evidence overlap value.

[0072] The initial corrected candidate asset similarity is obtained by adding the evidence overlap value to the corresponding candidate asset similarity and then dividing by two.

[0073] By aggregating the set of evidence edge identifiers corresponding to the selected asset set, a set of covered evidence edge identifiers is obtained.

[0074] Perform a difference operation between the set of evidence edge identifiers corresponding to the asset node to be added and the set of evidence edge identifiers that have been covered. Divide the number of evidence edge identifiers contained in the difference set by the total number of evidence edge identifiers corresponding to the asset node to be added to obtain the evidence increment value.

[0075] The evidence duplication coefficient is obtained by subtracting the evidence increment from one. The similarity of the initial corrected candidate assets is multiplied by the evidence duplication coefficient to obtain the similarity of the corrected candidate assets.

[0076] The two symmetrical non-main diagonal positions in the candidate asset kernel matrix corresponding to the asset node to be added are updated to the product of the quality values ​​of the two candidate assets and the similarity of the modified candidate assets.

[0077] Optionally, S7 specifically includes:

[0078] Arrange the corresponding digital assets according to the order in which each asset node in the target digital asset set is added;

[0079] Extract the digital asset content, associated data, and asset identifiers corresponding to each digital asset from the multi-source fusion dataset of digital assets;

[0080] The digital asset content, the associated data, and the asset identifier are associated with the corresponding asset semantic matching results, the ownership credibility score, and the retrieval stability evaluation value to form digital asset retrieval results;

[0081] Output the digital asset retrieval results.

[0082] The beneficial effects of this invention are:

[0083] (1) This invention performs format conversion, field mapping, time unification and asset identification merging on digital assets and related data to obtain a multi-source fusion dataset of digital assets. Based on asset nodes, related nodes, node relationship edges and asset relationship time sequence chains, a digital asset time sequence knowledge graph is constructed, so that digital assets from different sources and in different formats can form a unified association expression, reducing the problems of duplicate entry, relationship fragmentation and the inability to continuously track ownership records.

[0084] (2) By improving the ownership evidence guidance layer in the BLIP-2 model, the graph relationship weights are calculated based on the relationship type, connection path and time continuity. Time decay weighting and cross attention calculation are performed through the temporal semantic interaction structure, so that the semantic content of digital assets, ownership association and circulation time participate in asset semantic matching. This reduces the arrangement position of digital assets with invalid authorization, incomplete ownership path and inconsistent subject records, and improves the retrieval accuracy and ownership credibility.

[0085] (3) Calculate the quality value of candidate assets based on the semantic matching results of assets, the ownership credibility score and the retrieval stability evaluation value, and calculate the evidence overlap value and evidence increment value through evidence coverage constraints. Dynamically correct the candidate asset kernel matrix to reduce the probability that digital assets with similar content and duplicate ownership evidence are selected at the same time, so that the target digital asset set takes into account relevance, stability and evidence coverage, and improves the diversity and usability of retrieval results. Attached Figure Description

[0086] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0087] Figure 1 This is an overall flowchart of a digital asset retrieval method based on big data processing proposed in this invention;

[0088] Figure 2 This is a structural diagram of the improved BLIP-2 model proposed in this invention;

[0089] Figure 3 This is a flowchart of the DPP algorithm and evidence coverage constraint screening proposed in this invention. Detailed Implementation

[0090] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0091] refer to Figures 1-3 A digital asset retrieval method based on big data processing includes the following steps:

[0092] S1. Acquire digital assets and related data, and obtain a multi-source fusion dataset of digital assets through standardized processing;

[0093] S2. Construct a time-series knowledge graph of digital assets based on the data association relationships in a multi-source fusion dataset of digital assets;

[0094] S3. Input the retrieval request and the digital asset temporal knowledge graph into the improved BLIP-2 model, map the graph relationship weights through the ownership evidence guidance layer, and fuse request features, asset features and temporal features through the temporal semantic interaction structure to obtain the asset semantic matching result;

[0095] S4. Calculate the ownership credibility score based on the asset semantic matching results and the ownership association data in the digital asset time-series knowledge graph to obtain a set of credible candidate assets.

[0096] S5. Perturb the search request conditions and calculate the search stability evaluation value based on the change in the ranking of the set of credible candidate assets before and after the perturbation.

[0097] S6. Input the asset semantic matching results, ownership credibility score and retrieval stability evaluation value into the DPP algorithm, construct the candidate asset kernel matrix through evidence coverage constraints, and jointly screen the set of credible candidate assets to obtain the target digital asset set.

[0098] S7. Generate and output digital asset retrieval results based on the target digital asset set.

[0099] In this embodiment, S1 specifically refers to:

[0100] Acquire digital assets and related data, and extract digital asset content, asset attribute records, and related records respectively;

[0101] Standardized asset records are obtained by converting the format of digital asset content, asset attribute records, and associated records, mapping fields, and unifying time.

[0102] Extract asset identifiers from standardized asset records. When an asset identifier is missing, perform a hash calculation on the corresponding digital asset content and use the hash value as the asset identifier.

[0103] Standardized asset records belonging to the same digital asset are merged according to asset identifiers, and duplicate records are deleted to obtain a multi-source fusion dataset of digital assets.

[0104] In the specific implementation process, digital assets and related data are obtained from the digital asset management platform, distributed storage nodes and business record database. Incremental reading is performed according to a 30-minute collection cycle, and the collection time and data source of each data item are recorded. The digital asset content includes the original file or file access address. The asset attribute record is stored in JSON format, and the association record records the data association relationship between the subject and the behavior according to the asset identifier.

[0105] The content of digital assets is uniformly converted into binary file indexes, the field names of asset attribute records and related records are mapped to a preset field table, and each time field is converted into a millisecond-level timestamp to obtain standardized asset records; the preset field table includes at least an asset identifier field, a content field, a subject field, a behavior field, and a time field;

[0106] The asset identifier is read from the standardized asset record. When the asset identifier is missing, the SHA-256 algorithm is used to perform a hash calculation on the corresponding digital asset content to obtain a hash value with a length of 256 bits, and the hash value is used as the asset identifier. The standardized asset records are merged according to the asset identifier. When the asset identifier, time field and association field of two standardized asset records are the same, the standardized asset record with the later collection time is retained to obtain the multi-source fusion dataset of digital assets.

[0107] In this embodiment, S2 specifically refers to:

[0108] The digital asset multi-source fusion dataset is divided into asset record groups according to asset identifiers, and each asset record group is mapped to an asset node.

[0109] Extract the subject records and behavior records associated with the digital assets from the asset record group, and map the subject records and behavior records to associated nodes;

[0110] Establish node relationship edges based on the field correspondence between asset record groups, entity records, and behavior records, and convert the time fields associated with the node relationship edges into a unified time sequence number;

[0111] The nodes connected to the same asset node are arranged according to a unified time sequence number to form an asset relationship time sequence chain;

[0112] By aggregating asset nodes, associated nodes, node relationship edges, and asset relationship time-series chains, a digital asset time-series knowledge graph is constructed.

[0113] In the specific implementation process, the digital asset multi-source fusion dataset is grouped according to the asset identifier. Standardized asset records corresponding to the same asset identifier are divided into the same asset record group, and each asset record group is mapped to an asset node. Subject records and behavior records are extracted from each asset record group. Subject records with the same subject identifier are mapped to the same associated node, and each behavior record is mapped to the corresponding associated node according to the behavior identifier.

[0114] Read the field correspondence in the asset record group, subject record and behavior record, establish node relationship edges between nodes with the same asset identifier, subject identifier or behavior association identifier, and write the relationship type, starting node identifier, ending node identifier and time field into the node relationship edge;

[0115] The time field is converted into a unified time sequence number at the millisecond level. The node relationship edges connected to the same asset node are sorted in ascending order according to the unified time sequence number. The node relationship edges with adjacent time and consecutive node identifiers are connected in sequence to form an asset relationship time sequence chain. The asset node, associated node, node relationship edge and asset relationship time sequence chain are written into the graph database to construct a digital asset time sequence knowledge graph.

[0116] In this embodiment, the improved BLIP-2 model includes a multimodal input encoding module, a query bridging module, an ownership evidence guidance layer, a temporal semantic interaction structure, and a matching calculation module. The ownership evidence guidance layer and the temporal semantic interaction structure constitute the improved structure of the improved BLIP-2 model.

[0117] The retrieval request and the digital asset time-series knowledge graph are input into the multimodal input encoding module to extract the request content and retrieval conditions from the retrieval request. The request content is converted into a sequence of request terms, and the retrieval conditions are converted into condition feature vectors.

[0118] Extract asset nodes, associated nodes, node relationship edges, and asset relationship time-series chains from the digital asset time-series knowledge graph. Convert the content fields corresponding to asset nodes and associated nodes into node feature sequences, and convert node relationship edges and asset relationship time-series chains into relationship feature sequences to obtain multimodal input features.

[0119] The request term sequence, conditional feature vector, and multimodal input features are input into the query bridging module. Self-attention calculation is performed on the request term sequence, and cross-attention calculation is performed using the request term sequence as the query term and the multimodal input features as the key and value terms.

[0120] The cross-attention calculation results are concatenated with the conditional feature vectors dimension by dimension. After linear mapping, query bridging feature vectors corresponding one-to-one with each asset node are obtained. These vectors are then arranged in the order of the asset nodes to obtain the query bridging feature sequence.

[0121] Input the query bridging feature sequence and digital asset time-series knowledge graph into the ownership evidence guidance layer. Calculate the evidence association value of each asset node based on the relationship type and connection path corresponding to the node relationship edge. Divide each evidence association value by the sum of evidence association values ​​to obtain the graph relationship weight. Multiply the graph relationship weight by the query bridging feature vector of the corresponding asset node dimension by dimension to obtain the ownership guidance feature sequence.

[0122] Input the ownership guidance feature sequence and asset relationship time-series chain into the time-series semantic interaction structure, calculate the time interval between adjacent node relationship edges according to the unified time sequence number, perform time decay weighting on the ownership guidance feature sequence, and perform cross-attention calculation with the weighting result and the corresponding relationship feature vector to obtain the time-series interaction feature sequence;

[0123] Input the temporal interaction feature sequence into the matching calculation module, calculate the vector similarity between the retrieval request and each asset node, accumulate and normalize the vector similarity corresponding to the same asset node, arrange each asset node according to the normalization result, and obtain the asset semantic matching result.

[0124] In the specific implementation process, the improved BLIP-2 model inherits the multimodal encoding and query bridging framework of the traditional BLIP-2 model, including a multimodal input encoding module, a query bridging module, an ownership evidence guidance layer, a temporal semantic interaction structure, and a matching calculation module. Among these, the ownership evidence guidance layer and the temporal semantic interaction structure are new additions. First, the retrieval request is segmented, and the request content is truncated or padded to 128 words. After word embedding mapping, a 768-dimensional request feature is obtained. The retrieval conditions are encoded in a fixed order to obtain a 128-dimensional condition feature vector. For asset nodes and associated nodes in the digital asset temporal knowledge graph, the image content is scaled to 224×224 pixels, keyframes are extracted from the video content at 5-second intervals, and the first 256 words of the text content are truncated. After multimodal input encoding, a 768-dimensional node feature vector is obtained. The relationship type of the node relationship edge is mapped to a 128-dimensional relationship feature vector and arranged according to a unified time sequence number. The combination yields the multimodal input features.

[0125] The query bridging module uses 32 query vectors and 12 attention heads. It performs self-attention calculation on the request term sequence, then performs cross-attention calculation using the request term sequence as the query term and multimodal input features as the key and value terms. The cross-attention result is concatenated dimension-by-dimensionally with the conditional feature vector, and a linear mapping is performed to obtain the query bridging feature vector, which is then arranged according to the asset node order to form the query bridging feature sequence. This query bridging feature sequence is input into the ownership evidence guidance layer. Evidence association values ​​are calculated based on relationship type, connection path, and temporal continuity. Each evidence association value is divided by the sum of evidence association values ​​to obtain the graph relationship weight, which is then multiplied dimension-by-dimensionally with the corresponding query bridging feature vector to obtain the ownership guidance feature sequence.

[0126] The ownership-guided feature sequence and asset relationship time-series chain are input into the temporal semantic interaction structure. The time interval between adjacent node relationship edges is divided by 30 days and the negative exponent is taken to obtain the time decay weight, which is then multiplied dimension-by-dimensionally with the corresponding ownership-guided feature vector. The weighted result is then subjected to cross-attention calculation with the relationship feature sequence to obtain the temporal interaction feature sequence. The matching calculation module calculates the cosine similarity between the retrieval request and each asset node, averages the similarities corresponding to the same asset node to obtain the semantic matching value, and sorts them in descending order of semantic matching value to obtain the asset semantic matching result. Compared with the traditional BLIP-2 model, this implementation incorporates ownership relationship and relationship time into the feature update process, reducing the impact of expired authorization relationships and early transfer records, and improving the reliability of the retrieval results.

[0127] In this implementation, the query bridging feature sequence and digital asset time-series knowledge graph are input into the ownership evidence guiding layer, specifically as follows:

[0128] Receive query bridging feature sequences and digital asset time-series knowledge graphs, extract node relationship edges between each asset node and associated nodes, and determine the connection order of each node relationship edge according to the asset relationship time-series chain;

[0129] Starting with each asset node, extract the connection path along the node relationship edge, and gather the connection paths corresponding to the same asset node into a set of evidence paths.

[0130] Count the number of node relationship edges and the number of relationship types contained in each connection path, and divide the number of relationship types by the number of node relationship edges to obtain the relationship coverage value;

[0131] Calculate the time interval between adjacent node relationship edges according to the unified time sequence number, divide each time interval by the preset time interval upper limit, adjust the ratio greater than one to one, and obtain the interval ratio; calculate the average of all interval ratios corresponding to the same connection path, and subtract the average value from one to obtain the time continuity value.

[0132] Add the relationship coverage value and the time continuity value corresponding to the same connection path and divide by two to obtain the path evidence value. Add up all the path evidence values ​​corresponding to the same asset node to obtain the evidence association value.

[0133] Divide the evidence association value corresponding to each asset node by the sum of all evidence association values ​​to obtain the graph relationship weight;

[0134] Extract the query bridging feature vector corresponding to each asset node, multiply the graph relationship weight with the corresponding query bridging feature vector dimension by dimension, and arrange them according to the order of asset nodes to obtain the ownership guidance feature sequence;

[0135] In the specific implementation process, the ownership evidence guiding layer receives the query bridging feature sequence and the digital asset time-series knowledge graph, and calculates the graph relationship weights according to the order of asset nodes. First, taking each asset node as the starting node, the connection path is extended to the associated nodes along the node relationship edges. A single connection path contains one to four node relationship edges. When the connection path has duplicate nodes or the number of node relationship edges exceeds four, the extension stops, and all connection paths corresponding to the same asset node are gathered into an evidence path set.

[0136] For each connection path, count the number of node-related edges and the number of relationship types. Divide the number of relationship types by the number of node-related edges to obtain the relationship coverage value. For example, if a connection path contains four node-related edges and involves three relationship types, the relationship coverage value is 0.75. Calculate the time interval between adjacent node-related edges according to a unified time sequence number. Divide each time interval by 30 days, adjusting the ratio for intervals exceeding 30 days to 1 to obtain the interval percentage. Sum all interval percentages and divide by the number of interval percentages to obtain the average interval percentage. Subtract the average interval percentage from 1 to obtain the time continuity value.

[0137] The path evidence value is obtained by adding the relationship coverage value and the time continuity value corresponding to the same connection path and dividing by 2. The evidence association value is obtained by summing all path evidence values ​​corresponding to the same asset node. If no valid connection path is extracted, the evidence association value is set to 0. The graph relationship weight is obtained by dividing the evidence association value corresponding to each asset node by the sum of all evidence association values. When the sum of evidence association values ​​is 0, the graph relationship weight is set to the reciprocal of the number of asset nodes.

[0138] The query bridging feature vectors corresponding to each asset node are extracted. The graph relationship weights are then multiplied dimension-by-dimensionally with the corresponding query bridging feature vectors to obtain ownership-guided feature vectors. These vectors are then arranged according to the asset node order to obtain an ownership-guided feature sequence. Compared to the traditional BLIP-2 model, which only relies on content semantics to form query bridging features, the ownership evidence-guided layer incorporates both relationship coverage and temporal continuity into the graph relationship weights. This ensures that asset nodes with complete ownership paths and diverse relationship types achieve a higher feature retention rate, while suppressing asset nodes with interrupted ownership records or singular evidence, thereby improving the credibility of asset semantic matching results.

[0139] In this embodiment, S4 specifically refers to:

[0140] Read the asset semantic matching results, extract the semantic matching value corresponding to each asset node, and divide each semantic matching value by the sum of all semantic matching values ​​to obtain the normalized semantic matching value;

[0141] Extract ownership association data corresponding to each asset node from the digital asset time-series knowledge graph, connect the corresponding node relationship edges according to the subject identifier and time sequence in the ownership association data, and form ownership association paths by connecting the node relationship edges that can be continuously connected.

[0142] The number of node relationship edges included in the ownership association path is counted, and the number of node relationship edges is divided by the total number of node relationship edges included in the ownership association data of the corresponding asset node to obtain the complete value of the ownership path.

[0143] The valid time value is obtained by dividing the number of records in the ownership association data that are within the valid time range by the total number of records in the ownership association data.

[0144] The number of records with the same ownership entity in the ownership association data is counted. The number of records corresponding to the ownership entity with the largest number is divided by the total number of records in the ownership association data to obtain the record consistency value.

[0145] The ownership credibility score is obtained by adding the normalized semantic matching value, ownership path integrity value, time validity value and record consistency value corresponding to the same asset node and then dividing by four.

[0146] Asset nodes whose ownership trust scores reach a preset trust threshold are retained and sorted in descending order of ownership trust scores to obtain a set of trustworthy candidate assets;

[0147] In the specific implementation process, the semantic matching value in the asset semantic matching result is read, and each semantic matching value is divided by the sum of all semantic matching values ​​to obtain the normalized semantic matching value; the ownership association data corresponding to each asset node is extracted from the digital asset time-series knowledge graph, and the node relationship edges are connected according to the subject identifier and time order. The number of node relationship edges included in the ownership association path is divided by the total number of corresponding node relationship edges to obtain the complete value of the ownership path.

[0148] Valid records are those with valid status fields and current times falling between the start and end times. The number of valid records is divided by the total number of records in the ownership association data to obtain the valid time value. The number of records corresponding to the same ownership entity is counted, and the maximum number of records is divided by the total number of records to obtain the record consistency value.

[0149] The ownership trust score is obtained by adding the normalized semantic matching value, ownership path integrity value, time validity value, and record consistency value and then dividing by four. When the total number of records is zero, the corresponding time validity value and record consistency value are set to zero. The preset trust threshold is set to 0.65, and asset nodes with ownership trust scores of not less than 0.65 are retained and sorted in descending order of ownership trust scores to obtain a set of trustworthy candidate assets.

[0150] In this embodiment, S5 specifically refers to:

[0151] Parse the retrieval request, extract the retrieval criteria that limit the retrieval scope from the retrieval request, and record the original sorting position of each asset node in the set of trusted candidate assets;

[0152] Each search condition item is either deleted or adjusted, with only one search condition item changed at a time, resulting in multiple perturbation search requests.

[0153] Each perturbation retrieval request and the digital asset temporal knowledge graph are input into the improved BLIP-2 model to obtain the asset semantic matching results corresponding to each perturbation retrieval request;

[0154] Based on the asset semantic matching results and ownership association data corresponding to each perturbation retrieval request, the ownership credibility score is recalculated, and the asset nodes are arranged according to the recalculated ownership credibility score to obtain the perturbation sorting sequence.

[0155] Extract the perturbation ranking position of each asset node in the set of trustworthy candidate assets in the perturbation ranking sequence; when an asset node does not appear in the perturbation ranking sequence, add one to the total number of asset nodes in the perturbation ranking sequence as the corresponding perturbation ranking position;

[0156] Calculate the absolute difference between the original sort position and each disturbed sort position of the same asset node;

[0157] The total number of asset nodes in the set of credible candidate assets is compared with the total number of asset nodes in the corresponding perturbation sorting sequence plus one, and the larger value is taken as the sorting base; each absolute difference is divided by the corresponding sorting base to obtain the normalized sorting difference.

[0158] Calculate the average of all normalized sort differences corresponding to the same asset node, and subtract the average from one to obtain the retrieval stability evaluation value of the corresponding asset node;

[0159] In the specific implementation process, the search criteria in the search request are parsed, and the original sorting position of each asset node in the set of credible candidate assets is recorded. For text search criteria, one qualifier is deleted each time. For time boundary search criteria, the start time and end time are adjusted forward or backward by 10% of the original time range. For numerical boundary search criteria, the boundary value is increased or decreased by 10% each time. Only one search criteria is changed each time, resulting in no more than six perturbation search requests.

[0160] Each perturbation retrieval request and the digital asset time-series knowledge graph are input into the improved BLIP-2 model to obtain the asset semantic matching results again. The ownership credibility score is then recalculated according to the calculation process of S4 to obtain the perturbation ranking sequence.

[0161] Calculate the absolute difference between the original ranking position and each perturbation ranking position of the same asset node; when an asset node does not appear in the corresponding perturbation ranking sequence, add one to the total number of asset nodes in the perturbation ranking sequence as the perturbation ranking position; compare the total number of asset nodes in the set of credible candidate assets with the total number of asset nodes in the corresponding perturbation ranking sequence plus one, take the larger value as the ranking base, divide each absolute difference by the ranking base to obtain the normalized ranking difference; calculate the average of all normalized ranking differences corresponding to the same asset node, and subtract the average value from one to obtain the retrieval stability evaluation value.

[0162] In this embodiment, S6 specifically refers to:

[0163] Input the asset semantic matching results, ownership credibility score and retrieval stability evaluation value into the DPP algorithm to extract the semantic matching value corresponding to each asset node in the set of credible candidate assets.

[0164] Divide the semantic matching value, ownership credibility score and retrieval stability evaluation value by the sum of their corresponding values, and add the three ratios corresponding to the same asset node and divide by three to obtain the candidate asset quality value.

[0165] Extract the node feature sequence corresponding to each asset node from the digital asset time-series knowledge graph, and then sum the feature vectors in the node feature sequence dimension by dimension and divide by the number of feature vectors to obtain the candidate asset feature vector.

[0166] Multiply the values ​​of the corresponding dimensions of the feature vectors of the two candidate assets one dimension at a time and sum them up. Divide the summed result by the product of the magnitudes of the feature vectors of the two candidate assets to obtain the similarity of the candidate assets.

[0167] The evidence overlap value between two asset nodes is calculated based on the evidence coverage constraint. The evidence overlap value is added to the corresponding candidate asset similarity and then divided by two to obtain the initial corrected candidate asset similarity.

[0168] Write the square of the candidate asset quality value into the main diagonal position of the candidate asset kernel matrix, and write the product of the two candidate asset quality values ​​and the corresponding initial modified candidate asset similarity into the non-main diagonal position to construct the candidate asset kernel matrix;

[0169] The candidate asset kernel matrix is ​​decomposed by eigenvalue decomposition, and the eigenvalues ​​less than zero are adjusted to zero. The candidate asset kernel matrix is ​​then reconstructed based on the adjusted eigenvalues.

[0170] Add the asset node with the highest candidate asset quality value to the selected asset set, and add the remaining asset nodes to the principal sub-matrix of the candidate asset core matrix corresponding to the selected asset set. Calculate the difference in determinant between the principal sub-matrix of the candidate asset core matrix before and after the addition.

[0171] Add the asset node with the largest determinant difference to the selected asset set, and update the candidate asset kernel matrix according to the evidence coverage constraint. Repeat the process of adding asset nodes and updating the candidate asset kernel matrix until the selected asset set reaches the preset number of assets or all determinant differences are not greater than zero. Then, determine the selected asset set as the target digital asset set.

[0172] In the specific implementation process, the semantic matching results of assets, the ownership credibility score, and the retrieval stability evaluation value are input into the DPP algorithm. First, the semantic matching value, ownership credibility score, and retrieval stability evaluation value corresponding to each asset node are extracted from the set of credible candidate assets. The sum of the three types of values ​​is calculated for each asset node. The value corresponding to each individual asset node is divided by the corresponding sum to obtain three normalized values. The three normalized values ​​are added together and divided by 3 to obtain the candidate asset quality value.

[0173] Extract the node feature sequence corresponding to each asset node from the digital asset time-series knowledge graph, accumulate the feature vectors in the node feature sequence dimension by dimension, and divide by the number of feature vectors to obtain the candidate asset feature vector; calculate the product of the corresponding dimensions of the feature vectors of two candidate assets and accumulate them to obtain the vector dot product; calculate the magnitude of the feature vectors of two candidate assets respectively, divide the vector dot product by the product of the two magnitudes to obtain the candidate asset similarity.

[0174] The evidence overlap value between two asset nodes is calculated based on the evidence coverage constraint. This overlap value is then added to the candidate asset similarity and divided by 2 to obtain the initial revised candidate asset similarity. The square of the candidate asset quality value is written to the main diagonal position of the candidate asset kernel matrix, and the product of the two candidate asset quality values ​​and the initial revised candidate asset similarity is written to the corresponding non-main diagonal position to construct the candidate asset kernel matrix. Eigenvalue decomposition is performed on the candidate asset kernel matrix, adjusting any eigenvalues ​​less than zero to zero, and the candidate asset kernel matrix is ​​reconstructed based on the adjusted eigenvalues.

[0175] Select the asset node with the highest candidate asset quality value and add it to the selected asset set. Add the remaining asset nodes to the principal submatrix of the corresponding candidate asset core matrix and calculate the difference in determinant before and after addition. Add the asset node with the largest difference in determinant to the selected asset set and update the candidate asset core matrix according to the evidence coverage constraint. Repeat the above process until the selected asset set has 20 asset nodes or all determinant differences are not greater than zero. In the floating-point calculation process, the determinant difference with an absolute value not exceeding 0.000001 is treated as zero, and the selected asset set is determined as the target digital asset set.

[0176] Compared to traditional retrieval methods that sort solely by similarity, this implementation uses the DPP algorithm to consider both the quality of candidate assets and the degree of asset differences, thereby reducing the probability of highly similar digital assets repeatedly appearing in the retrieval results.

[0177] In this implementation, the candidate asset kernel matrix is ​​updated according to the evidence coverage constraint. The process of adding asset nodes and updating the candidate asset kernel matrix is ​​repeated. Specifically:

[0178] Using each asset node in the set of credible candidate assets as the starting point of the path, extract the set of evidence paths along the node relationship edges in the digital asset temporal knowledge graph;

[0179] The starting node identifier, ending node identifier, and relationship type of each node relationship edge are combined into an evidence edge identifier. The evidence edge identifiers corresponding to the same asset node are collected to obtain the evidence edge identifier set.

[0180] Perform intersection and union operations on the evidence edge identifier sets corresponding to two asset nodes, and divide the number of evidence edge identifiers in the intersection by the number of evidence edge identifiers in the union to obtain the evidence overlap value.

[0181] Add the evidence overlap value to the corresponding candidate asset similarity and divide by two to obtain the initial corrected candidate asset similarity.

[0182] By aggregating the set of evidence edge identifiers corresponding to the selected asset set, we obtain the set of covered evidence edge identifiers;

[0183] Perform a difference operation between the set of evidence edge identifiers corresponding to the asset node to be added and the set of evidence edge identifiers already covered. Divide the number of evidence edge identifiers contained in the difference set by the total number of evidence edge identifiers corresponding to the asset node to be added to obtain the evidence increment value.

[0184] The evidence duplication coefficient is obtained by subtracting the evidence increment from one. The similarity of the initial corrected candidate assets is multiplied by the evidence duplication coefficient to obtain the similarity of the corrected candidate assets.

[0185] The two symmetric non-main diagonal positions in the candidate asset kernel matrix corresponding to the node to be added are updated to the product of the quality values ​​of the two candidate assets and the similarity of the modified candidate assets.

[0186] In the specific implementation process, the asset nodes in the set of trusted candidate assets are used as the starting point of the path, and the evidence path set is extracted along the node relationship edges in the digital asset time-series knowledge graph; the number of node relationship edges contained in a single evidence path is limited to 1 to 4, and the expansion stops when the path has duplicate nodes or reaches 4 node relationship edges.

[0187] The starting node identifier, ending node identifier, and relationship type of each node relationship edge are combined to obtain the evidence edge identifier. The evidence edge identifiers corresponding to the same asset node are collected, and duplicate identifiers are removed to obtain the evidence edge identifier set. The intersection and union operations are performed on the evidence edge identifier sets corresponding to two asset nodes. The number of evidence edge identifiers in the intersection is divided by the number of evidence edge identifiers in the union to obtain the evidence overlap value.

[0188] Add the evidence overlap value to the candidate asset similarity value and divide by 2 to obtain the initial revised candidate asset similarity value. Obtain the evidence edge identifier set corresponding to the selected asset set, perform a difference operation with the evidence edge identifier set corresponding to the asset node to be added, and divide the difference set count by the total number of evidence edge identifiers corresponding to the asset node to be added to obtain the evidence increment value.

[0189] Subtract the evidence increment from 1 to obtain the evidence duplication coefficient. Multiply the initial corrected candidate asset similarity with the evidence duplication coefficient to obtain the corrected candidate asset similarity. Update the two symmetric non-main diagonal positions in the candidate asset kernel matrix corresponding to the asset node to be added to the product of the two candidate asset quality values ​​and the corrected candidate asset similarity.

[0190] By using evidence coverage constraints, digital assets with different ownership paths and transfer records are given a higher probability of selection, reducing the number of digital assets with highly duplicated ownership evidence entering the target digital asset set at the same time.

[0191] In this embodiment, S7 specifically refers to:

[0192] Arrange the corresponding digital assets according to the order in which each asset node in the target digital asset collection was added;

[0193] Extract the digital asset content, associated data, and asset identifiers corresponding to each digital asset from the multi-source fusion dataset of digital assets;

[0194] The digital asset retrieval results are generated by associating digital asset content, associated data, and asset identifiers with corresponding asset semantic matching results, ownership credibility scores, and retrieval stability evaluation values.

[0195] Output digital asset retrieval results;

[0196] In the specific implementation process, the corresponding digital assets are arranged according to the order in which each asset node in the target digital asset set is added to the selected asset set; the digital asset content, associated data and asset identifiers corresponding to each digital asset are extracted from the multi-source fusion dataset of digital assets, and the corresponding asset semantic matching results, ownership credibility scores and retrieval stability evaluation values ​​are read.

[0197] The digital asset content, associated data, asset identifier, asset semantic matching results, ownership credibility score, and retrieval stability evaluation value are written into the digital asset retrieval results. The digital asset retrieval results are stored in JSON format, with each record corresponding to a target digital asset. The digital asset retrieval results are generated and output according to the order of the target digital asset set.

[0198] Example 1: To verify the feasibility of this invention in practice, it was applied to a digital asset retrieval scenario on a digital content operation platform. This platform centrally manages 86,400 items, including images, videos, promotional texts, design source files, and combined materials. These include different resolution versions, partially modified versions, and repeatedly uploaded versions of the same material. The platform also stores 52,700 ownership records and 318,000 authorization and transfer records. Existing retrieval methods primarily rely on file names, manual tags, and content keywords to return results. When a search request simultaneously includes the content topic, scope of use, and authorization period, problems easily arise such as related content but expired authorization, inconsistent ownership entities, and duplicate returns of different versions of the same asset. Search personnel typically need to open multiple files and verify ownership records item by item; the manual verification time for a single complex search is approximately 4 to 8 minutes.

[0199] In this scenario, digital assets and related data within the platform are incrementally acquired at 30-minute intervals, and the field names, time formats, and asset identifiers from different data sources are standardized. For digital assets lacking asset identifiers, the SHA-256 algorithm is used to calculate the hash value of the digital asset content, and this hash value is used as the asset identifier. After merging duplicate records according to the asset identifier, a multi-source fused dataset of digital assets is obtained. Based on the data relationships between digital assets, ownership entities, and behavior records, asset nodes, related nodes, and node relationship edges are established, and a temporal chain of asset relationships is formed according to the time sequence of authorization, transfer, and access records, thus constructing a temporal knowledge graph of digital assets.

[0200] During testing, a search request was inputting "city night scene promotional materials, permitted for commercial use, with an authorization period of no less than six months." The improved BLIP-2 model jointly processes the search request and the digital asset temporal knowledge graph. The graph relationship weights of each asset node are calculated through an ownership evidence-guided layer, and the feature weights of early authorized records and expired transfer records are reduced through a temporal semantic interaction structure to obtain asset semantic matching results. An ownership credibility score is calculated based on the ownership path integrity value, time validity value, and record consistency value. A preset credibility threshold of 0.65 is set, and asset nodes with an ownership credibility score of 0.65 are retained, resulting in a set of credible candidate assets.

[0201] By removing the commercial use condition, adjusting the authorization period boundary, and replacing the content topic qualifiers from the search requests, six perturbation search requests were generated. The search stability evaluation value was calculated based on the change in the ranking position of each asset node before and after the perturbation. The asset semantic matching results, ownership credibility score, and search stability evaluation value were input into the DPP algorithm. Evidence coverage constraints were used to correct the similarity of candidate assets between asset nodes with similar content and overlapping ownership evidence. Finally, 20 digital assets were selected from the set of credible candidate assets to obtain the digital asset search results.

[0202] Sixty-one search requests containing content and ownership criteria were selected for testing, including 240 image requests, 150 video requests, and 210 text and design document requests. For each search request, three individuals with digital asset review experience independently evaluated the first ten search results. A result was considered valid only if at least two reviewers determined it to be relevant and valid. The keyword search method, vector similarity-based search method, the original BLIP-2 model search method, and the method of this invention were compared, resulting in Table 1.

[0203] Table 1. Comparison of Accuracy and Reliability of Digital Asset Retrieval

[0204] Keyword search methods 71.8 67.5 27.6 20.3 286 0.31 Vector similarity retrieval method 78.9 74.2 22.4 16.8 231 0.45 Original BLIP-2 search method 84.3 80.1 18.7 12.5 176 0.62 Method of the present invention 92.1 88.7 6.9 2.6 68 0.88

[0205] As shown in Table 1, the accuracy of the first 10 items of the method of this invention reaches 92.1%, which is 7.8 percentage points higher than the original BLIP-2 model retrieval method, and the effective asset recall rate increases from 80.1% to 88.7%. The ownership evidence guiding layer weights the query bridging features according to the relationship type and connection path, so that digital assets with incomplete ownership paths or inconsistent subject records receive lower graph relationship weights; the temporal semantic interaction structure performs attenuation processing according to the time interval of authorization records and circulation records, so that the impact of expired or long-interval ownership association data on the asset semantic matching results is reduced, thus reducing the proportion of invalid ownership assets from 12.5% ​​to 2.6%.

[0206] After updating the candidate asset kernel matrix through evidence coverage constraints, digital assets with similar content and the same ownership path no longer cluster in the target digital asset set, reducing the proportion of duplicate assets from 18.7% in the original BLIP-2 model retrieval method to 6.9%. Due to the significant reduction in invalid ownership assets and duplicate assets in the search results, the average manual review time for the top 10 results decreased from 176 seconds to 68 seconds. Although the average response time of the method of this invention increased to 0.88 seconds, it is still less than 1 second, meeting the interactive retrieval requirements of digital content operation platforms. The above test results demonstrate that this invention can improve content matching accuracy, ownership credibility, and search result diversity in a massive digital asset environment, while reducing the workload of manual verification.

[0207] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A digital asset retrieval method based on big data processing, characterized in that, Includes the following steps: S1. Acquire digital assets and related data, and obtain a multi-source fusion dataset of digital assets through standardized processing; S2. Construct a digital asset time-series knowledge graph based on the data association relationships in the multi-source fusion dataset of the digital assets; S3. Input the retrieval request and the digital asset temporal knowledge graph into the improved BLIP-2 model, map the graph relationship weights through the ownership evidence guidance layer, and fuse request features, asset features and temporal features through the temporal semantic interaction structure to obtain the asset semantic matching result; S4. Calculate the ownership credibility score based on the asset semantic matching result and the ownership association data in the digital asset time-series knowledge graph to obtain a set of credible candidate assets; S5. Perform conditional perturbation on the search request, and calculate the search stability evaluation value based on the change in the ranking of the set of credible candidate assets before and after the perturbation; S6. Input the asset semantic matching result, the ownership credibility score and the retrieval stability evaluation value into the DPP algorithm, construct the candidate asset kernel matrix through evidence coverage constraints, and jointly screen the credible candidate asset set to obtain the target digital asset set. S7. Based on the target digital asset set, generate and output digital asset retrieval results.

2. The digital asset retrieval method based on big data processing according to claim 1, characterized in that, Specifically, S1 is: Acquire digital assets and related data, and extract digital asset content, asset attribute records, and related records respectively; The digital asset content, the asset attribute records, and the associated records are format-converted, field-mapped, and time-unified to obtain standardized asset records. Extract the asset identifier from the standardized asset record. When the asset identifier is missing, perform a hash calculation on the corresponding digital asset content and use the hash value as the asset identifier. Standardized asset records belonging to the same digital asset are merged according to the asset identifier, and duplicate records are deleted to obtain a multi-source fusion dataset of digital assets.

3. The digital asset retrieval method based on big data processing according to claim 1, characterized in that, Specifically, S2 is: The digital asset multi-source fusion dataset is divided into asset record groups according to the asset identifier, and each asset record group is mapped to an asset node. Extract the subject records and behavior records associated with the digital assets from the asset record group, and map the subject records and behavior records as associated nodes; Establish node relationship edges based on the field correspondence between the asset record group, the subject record, and the behavior record, and convert the time field associated with the node relationship edge into a unified time sequence number; Arrange the node relationship edges connected to the same asset node according to the unified time sequence number to form an asset relationship time sequence chain; By aggregating the asset nodes, the associated nodes, the node relationship edges, and the asset relationship time-series chains, a digital asset time-series knowledge graph is constructed.

4. The digital asset retrieval method based on big data processing according to claim 1, characterized in that, The improved BLIP-2 model includes a multimodal input encoding module, a query bridging module, an ownership evidence guidance layer, a temporal semantic interaction structure, and a matching calculation module. The ownership evidence guidance layer and the temporal semantic interaction structure constitute the improved structure of the improved BLIP-2 model. The retrieval request and the digital asset time-series knowledge graph are input into the multimodal input encoding module to extract the request content and retrieval conditions from the retrieval request. The request content is converted into a request word sequence, and the retrieval conditions are converted into a condition feature vector. Extract asset nodes, associated nodes, node relationship edges, and asset relationship time-series chains from the digital asset time-series knowledge graph. Convert the content fields corresponding to the asset nodes and associated nodes into node feature sequences, and convert the node relationship edges and asset relationship time-series chains into relationship feature sequences to obtain multimodal input features. The request term sequence, the conditional feature vector, and the multimodal input features are input into the query bridging module. Self-attention calculation is performed on the request term sequence, and cross-attention calculation is performed using the request term sequence as the query term and the multimodal input features as the key and value terms. The cross-attention calculation result is concatenated with the conditional feature vector dimension by dimension, and linearly mapped to obtain the query bridging feature vector corresponding to each asset node. The vectors are then arranged in the order of the asset nodes to obtain the query bridging feature sequence. The query bridging feature sequence and the digital asset time-series knowledge graph are input into the ownership evidence guidance layer. The evidence association value of each asset node is calculated according to the relationship type and connection path corresponding to the node relationship edge. The evidence association value is divided by the sum of the evidence association values ​​to obtain the graph relationship weight. The graph relationship weight is multiplied dimension by dimension by the query bridging feature vector of the corresponding asset node to obtain the ownership guidance feature sequence. The ownership guidance feature sequence and the asset relationship time-series chain are input into the time-series semantic interaction structure. The time interval between adjacent node relationship edges is calculated according to the unified time sequence number. The ownership guidance feature sequence is subjected to time decay weighting. The weighting result is cross-attention calculation with the corresponding relationship feature vector to obtain the time-series interaction feature sequence. The temporal interaction feature sequence is input into the matching calculation module, and the vector similarity between the retrieval request and each asset node is calculated respectively. The vector similarity corresponding to the same asset node is accumulated and normalized. The asset nodes are arranged according to the normalization result to obtain the asset semantic matching result.

5. The digital asset retrieval method based on big data processing according to claim 4, characterized in that, The step of inputting the query bridging feature sequence and the digital asset time-series knowledge graph into the ownership evidence guiding layer specifically involves: Receive the query bridging feature sequence and the digital asset time-series knowledge graph, extract the node relationship edges between each asset node and associated nodes, and determine the connection order of each node relationship edge according to the asset relationship time-series chain; Starting from each asset node, the connection path is extracted along the node relationship edge, and the connection paths corresponding to the same asset node are aggregated into an evidence path set. Count the number of node relationship edges and the number of relationship types contained in each connection path, and divide the number of relationship types by the number of node relationship edges to obtain the relationship coverage value; Calculate the time interval between adjacent node relationship edges according to the unified time sequence number, divide each time interval by the preset time interval upper limit, adjust the ratio greater than one to one, and obtain the interval ratio; calculate the average value of all interval ratios corresponding to the same connection path, and subtract the average value from one to obtain the time continuity value. Add the relationship coverage value and the time continuity value corresponding to the same connection path and divide by two to obtain the path evidence value. Add up all the path evidence values ​​corresponding to the same asset node to obtain the evidence association value. Divide the evidence association value corresponding to each asset node by the sum of all evidence association values ​​to obtain the graph relationship weight; Extract the query bridging feature vector corresponding to each asset node, multiply the graph relationship weight with the corresponding query bridging feature vector dimension by dimension, and arrange them according to the order of asset nodes to obtain the ownership guidance feature sequence.

6. The digital asset retrieval method based on big data processing according to claim 1, characterized in that, Specifically, S4 is: Read the asset semantic matching results, extract the semantic matching value corresponding to each asset node, and divide each semantic matching value by the sum of all semantic matching values ​​to obtain the normalized semantic matching value; Extract ownership association data corresponding to each asset node from the digital asset time-series knowledge graph, and connect the corresponding node relationship edges according to the subject identifier and time sequence in the ownership association data. Form ownership association paths by connecting the node relationship edges that can be continuously connected. The number of node relationship edges included in the ownership association path is counted, and the number of node relationship edges is divided by the total number of node relationship edges included in the ownership association data of the corresponding asset node to obtain the complete value of the ownership path. The effective time value is obtained by dividing the number of records in the ownership association data that are within the effective time range by the total number of records in the ownership association data. The number of records with the same ownership subject in the ownership association data is counted. The number of records corresponding to the ownership subject with the largest number is divided by the total number of records in the ownership association data to obtain the record consistency value. The ownership credibility score is obtained by adding the normalized semantic matching value, the ownership path integrity value, the time validity value, and the record consistency value corresponding to the same asset node and then dividing by four. Asset nodes whose ownership trust scores reach a preset trust threshold are retained and sorted in descending order according to the ownership trust scores to obtain a set of trustworthy candidate assets.

7. The digital asset retrieval method based on big data processing according to claim 1, characterized in that, Specifically, S5 is: Parse the retrieval request, extract the retrieval condition items that limit the retrieval scope in the retrieval request, and record the original sorting position of each asset node in the set of trusted candidate assets; Each search condition item is either deleted or adjusted, with only one search condition item changed at a time, resulting in multiple perturbation search requests. Each perturbation retrieval request and the digital asset temporal knowledge graph are input into the improved BLIP-2 model to obtain the asset semantic matching results corresponding to each perturbation retrieval request; The ownership credibility score is recalculated based on the asset semantic matching results and ownership association data corresponding to each perturbation retrieval request, and the asset nodes are arranged according to the recalculated ownership credibility score to obtain the perturbation sorting sequence. Extract the perturbation ranking position of each asset node in the set of trusted candidate assets in the perturbation ranking sequence; when an asset node does not appear in the perturbation ranking sequence, add one to the total number of asset nodes in the perturbation ranking sequence as the corresponding perturbation ranking position; Calculate the absolute difference between the original sort position and each disturbed sort position of the same asset node; The total number of asset nodes in the set of trusted candidate assets is compared with the total number of asset nodes in the corresponding perturbation sorting sequence plus one, and the larger value is taken as the sorting base; each absolute difference is divided by the corresponding sorting base to obtain the normalized sorting difference. Calculate the average of all normalized sort differences corresponding to the same asset node, and subtract the average from one to obtain the retrieval stability evaluation value of the corresponding asset node.

8. The digital asset retrieval method based on big data processing according to claim 1, characterized in that, Specifically, S6 is: The asset semantic matching result, the ownership credibility score and the retrieval stability evaluation value are input into the DPP algorithm to extract the semantic matching value corresponding to each asset node in the set of credible candidate assets. Divide the semantic matching value, the ownership credibility score, and the retrieval stability evaluation value by the sum of their corresponding values, and add the three ratios corresponding to the same asset node and divide by three to obtain the candidate asset quality value. Extract the node feature sequence corresponding to each asset node from the digital asset time-series knowledge graph, and then sum the feature vectors in the node feature sequence dimension by dimension and divide by the number of feature vectors to obtain the candidate asset feature vector. Multiply the values ​​of the corresponding dimensions of the feature vectors of the two candidate assets one dimension at a time and sum them up. Divide the summed result by the product of the magnitudes of the feature vectors of the two candidate assets to obtain the similarity of the candidate assets. The evidence overlap value between two asset nodes is calculated based on the evidence coverage constraint. The evidence overlap value is added to the corresponding candidate asset similarity and then divided by two to obtain the initial corrected candidate asset similarity. The square of the candidate asset quality value is written into the main diagonal position of the candidate asset kernel matrix, and the product of the two candidate asset quality values ​​and the corresponding initial modified candidate asset similarity is written into the non-main diagonal position to construct the candidate asset kernel matrix. The candidate asset kernel matrix is ​​decomposed by eigenvalue decomposition, and the eigenvalues ​​less than zero are adjusted to zero. The candidate asset kernel matrix is ​​then reconstructed based on the adjusted eigenvalues. Add the asset node with the highest candidate asset quality value to the selected asset set, and add the remaining asset nodes to the principal sub-matrix of the candidate asset core matrix corresponding to the selected asset set. Calculate the difference in determinant between the principal sub-matrix of the candidate asset core matrix before and after adding them. The asset node with the largest determinant difference is added to the selected asset set, and the candidate asset kernel matrix is ​​updated according to the evidence coverage constraint. The process of adding asset nodes and updating the candidate asset kernel matrix is ​​repeated until the selected asset set reaches the preset number of assets or all determinant differences are not greater than zero. The selected asset set is then determined as the target digital asset set.

9. A digital asset retrieval method based on big data processing according to claim 8, characterized in that, The step of updating the candidate asset kernel matrix based on the evidence coverage constraint, and repeatedly performing asset node addition and candidate asset kernel matrix update, specifically involves: Using each asset node in the set of trusted candidate assets as the starting point of the path, extract the set of evidence paths along the node relationship edges in the digital asset time-series knowledge graph. The starting node identifier, ending node identifier, and relationship type of each node relationship edge are combined into an evidence edge identifier. The evidence edge identifiers corresponding to the same asset node are collected to obtain the evidence edge identifier set. Perform intersection and union operations on the evidence edge identifier sets corresponding to two asset nodes, and divide the number of evidence edge identifiers in the intersection by the number of evidence edge identifiers in the union to obtain the evidence overlap value. The initial corrected candidate asset similarity is obtained by adding the evidence overlap value to the corresponding candidate asset similarity and then dividing by two. By aggregating the set of evidence edge identifiers corresponding to the selected asset set, a set of covered evidence edge identifiers is obtained. Perform a difference operation between the set of evidence edge identifiers corresponding to the asset node to be added and the set of evidence edge identifiers that have been covered. Divide the number of evidence edge identifiers contained in the difference set by the total number of evidence edge identifiers corresponding to the asset node to be added to obtain the evidence increment value. The evidence duplication coefficient is obtained by subtracting the evidence increment from one. The similarity of the initial corrected candidate assets is multiplied by the evidence duplication coefficient to obtain the similarity of the corrected candidate assets. The two symmetrical non-main diagonal positions in the candidate asset kernel matrix corresponding to the asset node to be added are updated to the product of the quality values ​​of the two candidate assets and the similarity of the modified candidate assets.

10. A digital asset retrieval method based on big data processing according to claim 1, characterized in that, Specifically, S7 is: Arrange the corresponding digital assets according to the order in which each asset node in the target digital asset set is added; Extract the digital asset content, associated data, and asset identifiers corresponding to each digital asset from the multi-source fusion dataset of digital assets; The digital asset content, the associated data, and the asset identifier are associated with the corresponding asset semantic matching results, the ownership credibility score, and the retrieval stability evaluation value to form digital asset retrieval results; Output the digital asset retrieval results.

Images