Intelligent Information Resource Management Methods and Systems Based on Large Models and Data-Driven Approaches
By constructing a knowledge graph and combining it with path reasoning algorithms, the problems of lagging semantic understanding and scheduling strategies in information resource management are solved, achieving efficient resource recommendation and management, and improving the system's adaptability and accuracy.
Patent Information
- Application Number
- CN202611177675.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-05
- Publication Date
- 2026-11-10
AI Technical Summary
Existing information resource management technologies cannot effectively adapt to multi-source data fusion, semantic association mining between resources, and intelligent scheduling. They lack a deep understanding of the semantics of resource content, resulting in high resource redundancy, low retrieval accuracy, lagging management strategies, and insufficient system generalization and adaptive adjustment capabilities.
Construct a knowledge graph that integrates behavioral features and semantic elements, and combine it with path reasoning algorithms to evaluate the confidence of resource nodes and sort them by selection, so as to realize the semantic understanding and intelligent scheduling of information resources in heterogeneous environments.
It enhances the ability to model potential relationships between resources, improves the system's adaptability and accuracy when dealing with complex semantic relationships and changes in behavioral patterns, realizes dynamic optimization of resource recommendation and management strategies, and avoids the rigidity of rules and scheduling lag in traditional systems.
Smart Images

Figure CN123068359A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of information resource management technology, and more specifically to intelligent information resource management methods and systems based on large models and data-driven approaches. Background Technology
[0002] Against the backdrop of rapid development in information technology, the generation and accumulation of massive heterogeneous data resources have brought unprecedented complexity and challenges to information resource management. Traditional information resource management methods rely heavily on static rule configuration and manual intervention mechanisms, which cannot effectively adapt to the actual needs of diverse data types, complex semantic structures, and dynamically evolving access behaviors. In particular, they have significant limitations in multi-source data fusion, semantic correlation mining between resources, and intelligent scheduling.
[0003] Existing information resource management technologies primarily focus on basic data collection, tag classification, and rule-based retrieval, lacking a deep understanding of the semantics of resource content. This prevents effective modeling of implicit semantic relationships between resources, leading to high resource redundancy, low retrieval accuracy, and lagging management strategies. Some studies have attempted to introduce machine learning methods for resource classification or recommendation, but due to the failure to construct a unified semantic representation structure and reliance on specific scenario rules, the systems lack generalization and adaptive adjustment capabilities, making it difficult to support collaborative scheduling of information resources in complex and dynamic environments.
[0004] In recent years, large-scale models, such as BERT and deep learning technologies, have achieved remarkable results in the fields of natural language processing and semantic modeling. Knowledge graph structures have also become an important form of expression for characterizing semantic relationships between resources. However, existing solutions generally suffer from three main problems: First, they lack integrated modeling of resource content and behavioral features, failing to effectively combine access behavior with semantic structure; second, resource scheduling and organization strategies lack data-driven optimization mechanisms, failing to achieve self-evolution and dynamic updates of strategies; and third, the semantic association reasoning process suffers from insufficient accuracy and a single path evaluation method, making it difficult to support high-precision resource recommendation and management decisions. Summary of the Invention
[0005] To address the aforementioned technical challenges, this invention proposes an intelligent information resource management method and system based on large models and data-driven approaches. It constructs a knowledge graph that integrates behavioral features and semantic elements, and combines path reasoning algorithms to evaluate the confidence level of resource nodes and perform screening and ranking. This enables precise mapping and strategy self-optimization of information resources across the entire chain, from semantic understanding to intelligent scheduling, in heterogeneous environments.
[0006] To achieve the above objectives, the present invention adopts the following technical solution:
[0007] In a first aspect, embodiments of the present invention provide an intelligent information resource management method based on a large model and data-driven approach, comprising the following steps: S1. Merge the information resource set and behavior dataset for preprocessing to generate a resource content sequence and a behavior record sequence; S2. Perform contextual semantic analysis on the resource content sequence based on the pre-trained BERT model to extract a set of semantic elements; S3. Using the semantic elements as nodes, calculate the support and path score the relationship edges between nodes, filter the relationship edges according to the support and path score, and perform consistency verification on each filtered node to construct a knowledge graph. S4. Extract resource co-occurrence features from the behavior record sequence, and perform time alignment between the behavior record sequence and each node in the knowledge graph to generate a behavior feature set; S5. Encode the behavioral feature set and map it to the knowledge graph to identify the mapped nodes and similar nodes; S6. Perform path frequency statistics and confidence scores on the reachable paths between the mapped node and the similar node, and filter the set of associated paths based on the frequency statistics and confidence scores to generate a target resource sequence; S7. Perform scheduling control on the information resource set according to the target resource sequence, write the scheduling result into the resource behavior set, and update the behavior record sequence.
[0008] Furthermore, in step S1, the information resource set represents a multi-source structured or unstructured data set containing timestamps to be managed, including documents, images, tables, logs, or audio; the behavior data set represents a set containing access behavior, operation records, and interaction log data; the preprocessing includes performing format normalization, noise reduction, and time calibration operations on the information resource set, and performing format standardization, missing value imputation, and time alignment operations on the behavior data set.
[0009] Furthermore, step S2 specifically includes: S21. The resource content sequence is sequentially segmented according to the preset text boundary rules to obtain multiple text fragments, which are then generated into a text fragment sequence after being marked with position and bound to the sequence. S22. Input the text segment sequence into the pre-trained BERT model, perform context encoding on each text segment, extract the word vector set, and arrange it into a word vector sequence in the original order; S23. Construct a word co-occurrence graph using word vectors as nodes and adjacency and co-occurrence relationships of words within a text segment as edges; the co-occurrence relationship is defined as follows: if two words are separated by no more than k words, then they have a co-occurrence relationship, where k is a preset positive integer; S24. Initialize the weight values of each node in the word co-occurrence graph, perform multiple rounds of TextRank iteration updates until convergence, and output the word ranking results in descending order of weight values; S25. Filter the word sorting results according to the preset mapping template to obtain a keyword set. Perform aggregation operation on the vector of each keyword and summarize them to form a semantic element set according to the original order of the resource content sequence.
[0010] Furthermore, step S3 specifically includes: S31. Using the semantic elements as nodes, traverse any two semantic elements to construct node pairs, and construct a set of directed relation edges based on the cosine similarity of each node pair and the position of each node in the set of semantic elements. S32. The PRA algorithm is used to calculate the support and path score of each relation edge, and relation edges whose support is lower than a preset reference value and whose path score is lower than a set score threshold are removed. S33. Use the TransE model to perform vector consistency verification on the node pairs corresponding to the remaining relation edges, retain the node pairs that meet the preset consistency conditions, and construct a set of connection edges. S34. Generate the knowledge graph based on the set of nodes and the set of connecting edges.
[0011] Furthermore, step S32 specifically includes: S321. Starting from the source node of each relation edge, perform directed path traversal in the knowledge graph, extract all semantic paths not exceeding the set maximum number of hops, and filter out the candidate path set containing the relation edge. S322. Count the number of times each relation edge appears in the candidate path set, and calculate the ratio of the number of occurrences to the total number of semantic paths, as the path frequency support of the relation edge. S323. Calculate the path semantic score for each relation edge in the following manner: The number of nodes in each candidate path is used as the path structure length value; Extract the resource representation pairs of each consecutive node pair in the candidate path from the resource representation set and calculate the vector similarity. Use the mean of all vector similarities as the path semantic coherence value. The path structure length value and the path semantic coherence value are weighted and superimposed according to a set weight to obtain the path semantic score value of the relation edge. S324. Remove relation edges where the path frequency support is lower than a preset reference value and the path semantic score is lower than a set score threshold.
[0012] Furthermore, step S4 specifically includes: S41. The behavior record sequence is expanded by time step, divided into multiple time segments according to the set sliding window length and step size, the accessed information resource pairs in each time segment are traversed, a resource co-occurrence pair set is constructed, and the frequency of their occurrence in the sliding window is counted to construct a co-occurrence frequency matrix. S42. Construct a resource co-occurrence graph based on the co-occurrence frequency matrix. The resource co-occurrence graph uses information resources as nodes and co-occurrence relationships as edges. Extract the nodes and edges of the resource co-occurrence graph, call the TF-IDF algorithm to calculate the weight value of each information resource and normalize it to generate a resource co-occurrence sequence. S43. Map the access time in the behavior record sequence to the resource co-occurrence sequence, use the dynamic time warping algorithm to pair the mapping result with the nodes in the knowledge graph, calculate the minimum alignment cost path, and obtain the alignment result set. S44. Merge the resource co-occurrence sequence with the alignment result set, bind each information resource to the corresponding alignment result at the tag level, construct a fused tag set, and arrange them in chronological order as the behavioral feature set.
[0013] Furthermore, step S42 specifically includes: Based on the co-occurrence frequency matrix, a resource co-occurrence graph is constructed with information resources as nodes and non-zero co-occurrence frequencies as edges. The resource co-occurrence graph is an undirected graph, and the edge weight is the corresponding co-occurrence frequency. Extract all nodes and edges from the resource co-occurrence graph, assign a unique resource identifier to each information resource and establish an index mapping relationship, and call the TF-IDF algorithm to calculate the weighted frequency of co-occurrence between each information resource and all co-occurring resources to generate a set of weight values.
[0014] Furthermore, step S5 specifically includes: S51. Use a multilayer perceptron to encode the behavioral feature set to generate an encoded feature set; S52. Calculate the cosine similarity between each encoded feature vector and all nodes in the knowledge graph, filter out nodes that are higher than a set threshold, and construct a set of mapped nodes. S53. Based on the mapped node set, call the SimRank algorithm to identify similar nodes in the knowledge graph and construct a similar node set; specifically including: Using each mapping node as the base node, traverse the knowledge graph to find nodes that have a direct connection edge or a two-hop path with it as candidate similar nodes. Count the number of adjacent nodes shared by each candidate similar node and the base node, and calculate the structural co-occurrence index; The co-occurrence indices of each structure are standardized using Z-Score and weighted by the edge weights in the knowledge graph to generate structural similarity values. Candidate nodes with structural similarity values higher than the preset similarity reference value are marked as similar nodes, and a set of similar nodes for the mapped node is constructed.
[0015] Furthermore, step S6 specifically includes: S61. Using the PCRA algorithm, construct all candidate node pairs between the mapping node set and the similar node set to generate a candidate node pair set; S62. In the knowledge graph, with each candidate node pair as the search range, extract all paths from the mapping node to the similar node that do not exceed the set maximum number of hops, and remove invalid paths after structural validity check to construct a set of reachable paths; S63. Count the frequency of each reachable path in all candidate node pairs, calculate the ratio of its occurrence to the total number of candidate node pairs as the frequency index, and generate a path frequency record table. S64. Extract the cosine similarity of the encoded features of consecutive nodes in each reachable path, calculate its mean as the path similarity value, and perform a weighted sum with the path length of the path to generate a path score value and form a path score record table. S65. Filter reachable paths with frequency indicators higher than preset reference values and path score values higher than set score thresholds to form an associated path set. Extract information resources corresponding to all endpoint nodes in the associated path set, sort them in descending order of path score values, and generate a target resource sequence.
[0016] Secondly, embodiments of the present invention provide an intelligent information resource management system based on large models and data-driven approaches, as described in the first aspect, the system comprising: The data acquisition module is used to acquire information resource sets and behavior datasets, merge them for preprocessing, and generate resource content sequences and behavior record sequences. The semantic analysis module is used to perform contextual semantic analysis on the resource content sequence based on a pre-trained BERT model and extract a set of semantic elements. The graph construction module is used to calculate the support and path score of the relationship edges between the semantic elements as nodes, filter the relationship edges according to the support and path score, and perform consistency verification on each filtered node to construct a knowledge graph. The feature extraction module is used to extract resource co-occurrence features from the behavior record sequence, align the behavior record sequence with each node in the knowledge graph in time, and generate a set of behavior features. The similarity recognition module is used to encode the behavioral feature set and map it to the knowledge graph, and identify the mapped nodes and similar nodes; The path filtering module is used to perform path frequency statistics and confidence scores on reachable paths between the mapped node and the similar node, filter the set of associated paths based on the frequency statistics and confidence scores, and generate a target resource sequence. The scheduling control module is used to schedule and control the information resource set according to the target resource sequence, write the scheduling result into the resource behavior set, and update the behavior record sequence.
[0017] The beneficial effects of the above-described technical solutions provided in the embodiments of the present invention include at least the following: This invention enables contextual understanding and keyword extraction of information resource content, and can extract a high-quality set of semantic elements from raw unstructured text, providing an accurate and rich semantic foundation for subsequent knowledge graph construction, and effectively improving the ability to model potential relationships between resources.
[0018] This invention achieves deep integration of resource behavior features and semantic graphs, improving the system's adaptability and accuracy when facing complex semantic associations and changes in behavior patterns.
[0019] This invention constructs a complete scheduling and control chain from resource representation to target resource sequence, realizing a resource recommendation and management strategy update mechanism driven by semantic understanding and data feedback. The scheduling results can be dynamically fed back to the behavioral data set, enabling continuous optimization and evolution of management strategies. This effectively avoids the problems of rigid rules and scheduling lag in traditional information systems, and improves the overall intelligence level, response efficiency and scheduling accuracy of information resource management. Attached Figure Description
[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0021] Figure 1 This is a flowchart of an intelligent information resource management method based on a large model and data-driven approach, provided in an embodiment of the present invention.
[0022] Figure 2 This is a flowchart of the BERT model pre-training process for an intelligent information resource management system based on a large model and data-driven approach, provided in an embodiment of the present invention.
[0023] Figure 3 This is a flowchart of path reasoning and semantic scheduling for an intelligent information resource management system based on a large model and data-driven approach, provided in an embodiment of the present invention. Detailed Implementation
[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0025] Example 1: like Figure 1 As shown, this invention discloses an intelligent information resource management method based on large models and data-driven approaches: Step S1: Obtain the information resource set and the behavior data set, and preprocess them to generate the resource content sequence and the behavior record sequence; Step S2: Call the pre-trained BERT model to perform contextual semantic analysis on the resource content sequence, extract word vector sequences and construct a word co-occurrence graph, use the TextRank algorithm to iteratively calculate the word co-occurrence graph, and convert the iteration results into a semantic element set according to the preset mapping template; Step S3: Using semantic elements as nodes, initialize the relationship edges between each node pair, use the PRA algorithm to perform support calculation and path scoring operations on the relationship edges, remove relationship edges with support lower than the preset reference value and path score lower than the set scoring threshold, and perform consistency verification on each node to construct a knowledge graph. Step S4: Perform resource co-occurrence feature extraction on the behavior record sequence through the sliding window mechanism, call the TF-IDF algorithm to perform importance evaluation and weight normalization on the extraction results, generate resource co-occurrence sequence, and at the same time use the dynamic time warping algorithm to perform time alignment operation on the behavior record sequence and each node in the knowledge graph, fuse and bind the resource co-occurrence sequence and the alignment result to construct a behavior feature set; Step S5: Perform feature encoding operation on the behavioral feature set to generate an encoded feature set, and perform cosine similarity calculation with the nodes in the knowledge graph. Select nodes with cosine similarity higher than a set threshold to form a mapping node set. Use the SimRank algorithm to perform similar node recognition operation in the knowledge graph based on the mapping node set. Step S6: Combine the PCRA algorithm to perform path frequency statistics and confidence scoring operations on reachable paths between mapped nodes and similar nodes, filter the set of associated paths with confidence scores higher than the set confidence threshold, and generate the target resource sequence; Step S7: Perform scheduling control operations on the information resource set according to the target resource sequence, write the scheduling results into the resource behavior set, and update the behavior record sequence.
[0026] In this invention, the information resource set in step S1 represents a multi-source structured or unstructured data set containing timestamps to be managed, including documents, images, tables, logs or audio. The behavior data set represents a set containing access behavior, operation records and interaction log data. Preprocessing includes performing format normalization, noise reduction and time calibration operations on the information resource set, and performing format standardization, missing value imputation and time alignment operations on the behavior data set.
[0027] In this embodiment of the invention, step S2 specifically includes: Step S21: Perform sequential segmentation on the resource content sequence according to preset text boundary rules to obtain multiple text fragments and perform position marking and sequence binding to generate a text fragment sequence; Step S22: Input the text segment sequence into the pre-trained BERT model, perform context encoding operation on each text segment, extract the word vector set of the text segment, and arrange them in the original order to form a word vector sequence; Step S23: Using word vectors as nodes and adjacency and co-occurrence relationships of words in the word vector sequence within the text segment as edges, construct a word co-occurrence graph. Here, co-occurrence relationship means that if the interval between two words does not exceed k words, then there is a co-occurrence relationship between the two words, where k represents a preset positive integer value. Step S24: Initialize the weight values of each node in the word co-occurrence graph and perform TextRank iterative calculation. Update the weight values of each node during multiple iterations to generate a word ranking result sorted in descending order of weight values. Specifically, this includes: Initialize weight values for all nodes in the word co-occurrence graph; In each iteration, for each node, a weighted update operation is performed based on the weight values of all its neighboring nodes and the number of connecting edges, combined with a preset transition probability factor, to update the weight value of the node. After each iteration, the change in weight values of all nodes is statistically analyzed, and the change in weight values of each node is compared between the previous and current iterations. If the change in weight values of all nodes is lower than the preset convergence threshold or reaches the preset maximum number of iterations, the iteration process is terminated, and the next iteration is directly entered after the first iteration without performing the convergence judgment operation. After the iteration is completed, the weight values updated in the last iteration are constructed, all nodes are sorted in descending order, and the word ranking results are generated. Step S25: Perform matching and filtering operations on the words in each word sorting result according to the preset mapping template to obtain the keyword set and perform vector aggregation operation. Summarize all aggregation results and build a semantic element set according to the order of resource content sequence.
[0028] like Figure 2As shown in this embodiment, the pre-training process of the BERT model specifically includes: A general corpus was constructed and subjected to unified encoding, special character removal, and multilingual filtering to obtain standard text sequences. The general corpus consists of online encyclopedia texts, book chapters, and news content. A masking operation is performed on the standard text sequence. In each standard text, a pseudo-random number generator is used to randomly mask multiple words according to a set ratio. The masked words are used as prediction targets, and the unmasked words are used as context information to construct a word prediction sample set. Extract consecutive sentence pairs from the standard text sequence according to sentence boundaries to construct positive samples with contextual relationships. At the same time, randomly select non-adjacent sentence pairs to construct negative samples. Assign binary labels to the positive and negative samples respectively to construct a sentence pair sample set. The binary label is used to indicate whether the sentence pair is a consecutive sentence pair with contextual relationships. The word prediction samples and sentence pair samples are input into the BERT model, which is composed of a multi-layer bidirectional Transformer structure. The context modeling operation is performed on each word in the two types of samples to generate the corresponding semantic representation. Based on semantic representation, word prediction operation is performed on word prediction samples, and the error value between the prediction result and the corresponding prediction target is calculated. At the same time, based on semantic representation, sentence relationship judgment operation is performed on sentence pair samples, and the error value between the judgment result and the binary label is calculated. The loss function is constructed with the goal of minimizing the two types of error values, and the parameters of the BERT model are iteratively updated through the backpropagation mechanism until the loss function converges, thus completing the pre-training of the BERT model.
[0029] In this embodiment, step S25 specifically includes: Step S251: Read words sequentially from the word sorting results, determine whether each word meets the keyword conditions according to the preset mapping template, and use the words that meet the conditions as keywords to construct a keyword set; Step S252: For each keyword, extract the corresponding word vector, and based on the position of the keyword in the resource content sequence as the center point, extract the word vectors corresponding to all words within a window with a fixed length as the radius, and construct a vector group; Step S253: Perform a weighted aggregation operation on each vector group to generate semantic elements. The aggregation weight coefficient of the weighted aggregation operation is set according to the distance between each word position and the center point and the corresponding weight value. Step S254: Rearrange all semantic elements according to the order of the resource content sequence to construct a semantic element set.
[0030] In this embodiment, step S3 specifically includes: Step S31: Using semantic elements as nodes, traverse any two semantic elements in the semantic element set to construct a set of node pairs, and construct a set of directed relation edges based on the cosine similarity between each node pair and the position of each node in the semantic element set. Step S32: Use the PRA algorithm to perform support calculation and path scoring operations on each relation edge. Perform analysis and comparison operations on the relation edges based on the support and path scores, and remove relation edges with support lower than the preset reference value and path scores lower than the set scoring threshold from the relation edge set. Step S33: Using the TransE model, perform vector consistency verification on the node pairs corresponding to the unremoved relation edges, retain the node pairs that meet the preset consistency conditions, and construct a set of connection edges. Specifically, this includes: For each relation edge that has not been removed, extract the node pair corresponding to the relation edge, and obtain the resource representation of the source node and the target node in the resource representation set respectively. Here, the source node represents the starting point of the directed relation edge, the target node represents the ending point of the directed relation edge, and the resource representation set represents the vectorized feature set generated by the behavior feature set after feature encoding operation. The resource representation is a d-dimensional real vector, where d represents the preset encoding dimension. Based on the direction information of each relation edge, an offset vector consistent with the resource representation dimension is constructed, and the resource representation of the source node and the offset vector are added together in the feature dimension to generate the predicted representation of the relation edge. The predicted representation and the resource representation of the target node are calculated by performing a difference operation on each feature dimension. The square operation is performed on each feature dimension of the difference result and then summed to generate the vector distance value of the relation edge. Node pairs whose vector distance values are lower than the preset distance reference value are marked as meeting the preset consistency condition, and a set of connecting edges is constructed. Step S34: Generate a knowledge graph based on the set of nodes and the set of connecting edges.
[0031] In this embodiment, step S32 specifically includes: Step S321: For each relation edge, take the source node of each relation edge as the starting point of the path, perform a directed path traversal operation in the knowledge graph, extract all semantic paths that do not exceed the set maximum number of hops, and filter out the candidate path set that contains the relation edge in the semantic path. Step S322: Count the number of times each relation edge appears in the candidate path set, and calculate the ratio of the number of occurrences to the total number of semantic paths to generate the support of the relation edge. The support represents the frequency of the relation edge as a semantic propagation path in the knowledge graph. Step S323: Perform a path score calculation operation for each relation edge, specifically including: The path length is determined by the number of nodes in each candidate path. Extract the resource representation pairs in the resource representation set for each consecutive node pair in the candidate path, calculate the cosine similarity, and take the mean of all cosine similarities in the candidate path as the path similarity value. The path length value and the path similarity value are weighted and superimposed according to the set weights to calculate the path score value of the relation edge; Step S324: Remove relationship edges whose path support value is lower than the preset reference value and whose path score value is lower than the set score threshold from the relationship edge set.
[0032] In this embodiment, step S4 specifically includes: Step S41: Perform time step expansion operation on the behavior record sequence, divide the behavior record sequence into multiple time segments according to the set sliding window length and step size, traverse all information resource pairs accessed in each time segment, construct a resource co-occurrence pair set, and construct a co-occurrence frequency matrix according to the frequency of occurrence of resource co-occurrence pairs in the sliding window. Step S42: Construct a resource co-occurrence graph based on the co-occurrence frequency matrix, extract the nodes and edges in the resource co-occurrence graph, call the TF-IDF algorithm to calculate the weight value of each information resource, and perform normalization processing to generate a resource co-occurrence sequence. The resource co-occurrence graph uses information resources as nodes and the sharing relationships corresponding to resource co-occurrence pairs as edges, specifically including: Based on the co-occurrence frequency matrix, traverse the entire set of resource co-occurrence pairs, using information resources as nodes, assign edges to any two resource co-occurrence pairs with non-zero co-occurrence frequencies in the co-occurrence frequency matrix, and construct a resource co-occurrence graph, where the resource co-occurrence graph is an undirected graph and the weight of the edge is the corresponding co-occurrence frequency; Extract all nodes and edges from the resource co-occurrence graph, assign a unique resource identifier to each information resource in the information resource set, establish an index mapping relationship, and use the TF-IDF algorithm to calculate the weighted co-occurrence frequency between each information resource and all corresponding co-occurring resources, generating a set of weight values, specifically including: Assign a unique resource identifier to each node in the resource co-occurrence graph and establish an index mapping relationship between the resource identifier and the node location; For each resource co-occurrence pair, extract the co-occurrence frequency of the two corresponding information resources, and calculate the confidence weight of the co-occurrence frequency in the whole graph. The confidence weight represents the significance of the information resource in all co-occurrence relationships. Summarize the confidence weight values between each information resource and all corresponding co-occurring resources, sort them according to the index mapping relationship, and generate the corresponding resource weight vector; All resource weight vectors are aggregated into a set of weight values; Perform Min-Max normalization on the set of weight values and construct a weight mapping table based on the index mapping relationship; Weight injection is performed on all nodes in the resource co-occurrence graph according to the weight mapping table, and the nodes are arranged according to the original time order of the information resources to generate a resource co-occurrence sequence. Step S43: Map the access time information recorded in the behavior record sequence to the resource co-occurrence sequence, use the dynamic time warping algorithm to pair the mapping result with the nodes in the knowledge graph, perform the minimum alignment cost path calculation operation on the pairing result, and obtain the alignment result set; Step S44: Perform a fusion operation on the resource co-occurrence sequence and the alignment result set, bind each information resource in the resource co-occurrence sequence with the corresponding alignment result at the tag level, construct a fusion tag set, and arrange it into a behavioral feature set in chronological order.
[0033] In this embodiment, step S43 specifically includes: Step S431: Extract the access time tag corresponding to each information resource from the behavior record sequence, and fill the blank time placeholder value for behavior records with missing time tags. Step S432: Map each information resource in the resource co-occurrence sequence to its corresponding access time tag according to the resource identifier order, and generate a resource co-occurrence sequence with timestamps as the mapping result; Step S433: Extract the timestamps of each node in the knowledge graph, construct the node time series, and perform time format standardization and granularity unification operations; Step S434: Call the dynamic time warping algorithm to perform pairing calculations on the mapping results and node time series, and generate resource node paths with the minimum path cost as the alignment target; Step S435: Perform path cost statistics operation on each paired node in the resource node path, filter paired nodes whose path cost is lower than the set cost threshold, and form an alignment result set, where the paired node represents the information resource and the corresponding knowledge graph node.
[0034] like Figure 3 As shown, in this embodiment, step S5 specifically includes: Step S51: Use a multilayer perceptron to perform feature encoding operations on all behavioral features to generate an encoded feature set; Step S52: Perform cosine similarity calculation on each encoded feature vector and all nodes in the knowledge graph, filter out nodes with cosine similarity higher than a set threshold, and construct a set of mapped nodes; Step S53: Based on the mapped node set, call the SimRank algorithm to perform similar node identification in the knowledge graph and construct a similar node set, specifically including: Using each mapping node as a base node, all nodes in the knowledge graph that are directly connected to the base node or connected by a two-hop path are traversed as candidate similar nodes. The number of adjacent nodes shared by each candidate similar node and the base node in the knowledge graph is counted, and the structural co-occurrence index between the base node and the candidate node is calculated based on the statistical results. Z-Score standardization is performed on the structural co-occurrence index between all candidate nodes and base nodes, and weighted calculation is performed in combination with the edge weights in the knowledge graph to generate the corresponding structural similarity value; Candidate nodes with structural similarity values higher than the preset similarity reference value are marked as similar nodes, and a set of similar nodes for the mapped nodes is constructed.
[0035] In this embodiment, step S6 specifically includes: Step S61: Using the PCRA algorithm, perform node pair construction operation on the mapping node set and the similar node set to construct a candidate node pair set between all mapping nodes and all corresponding similar nodes; Step S62: In the knowledge graph, take each candidate node pair as the search range, set the maximum number of hops, perform path traversal operation, extract all paths from the mapping node to the corresponding similar node with a path length not exceeding the maximum number of hops, and perform structural validity check operation to remove paths with missing edges or directional conflicts, and construct a set of reachable paths. Step S63: Perform path frequency statistics operation for each reachable path, count the number of times each reachable path appears in all candidate node pairs, and calculate the ratio with the total number of candidate node pairs to obtain the frequency index of the corresponding reachable path. Record all frequency indices to form a path frequency record table. Step S64: Extract the encoded feature pairs corresponding to consecutive nodes in each reachable path and calculate the cosine similarity. Take the mean of all cosine similarities in the reachable path as the path similarity value, and perform a weighted summation operation in combination with the corresponding path length to generate a path score value and form a path score record table. Step S65: Select reachable paths from the path frequency record table and the path score record table that have a frequency index higher than the preset reference value and a path score value higher than the set score threshold to form an associated path set. Extract the information resources corresponding to all endpoint nodes in the associated path set, sort them in descending order of path score value, and generate a target resource sequence.
[0036] In this embodiment, step S7 specifically includes: Step S71: Construct a set of scheduling instructions based on the target resource sequence and execute them sequentially; Step S72: Collect the running status information and feedback data of each information resource during the execution of scheduling control instructions, extract the scheduling response results and construct an execution record set, wherein the execution record set includes information resource identifier, scheduling time, execution status and response delay; Step S73: Write the execution record set into the resource behavior set, and align and match it with the behavior record sequence according to the resource identifier and timestamp. Insert execution records to update the structure of the behavior record sequence.
[0037] The following example illustrates the intelligent information resource management method based on large models and data-driven approaches of the present invention: To verify the feasibility of this invention in practice, it was applied to an enterprise's internal knowledge asset management system. This system performs unified semantic modeling and intelligent scheduling optimization on massive amounts of technical documents, R&D reports, structured experimental data, and employee interaction records. The main problem in this scenario is the complexity of information resource types, including a large amount of unstructured text, tables, images, and behavioral logs. Traditional retrieval mechanisms based on keyword matching and fixed classification rules cannot meet the needs of in-depth semantic correlation mining, resulting in low resource reuse rates, long search times, and poor recommendation accuracy, severely restricting the efficient circulation and value transformation of knowledge assets.
[0038] In this embodiment, various information resources and corresponding access behavior data are first collected. After format normalization, noise reduction, and timestamp alignment, resource content sequences and behavior record sequences are generated. A pre-trained BERT model is invoked to perform semantic encoding on the resource content, extracting context word vectors and constructing a word co-occurrence graph. The TextRank algorithm is then used to iteratively generate a set of semantic elements. Based on this, an information resource knowledge graph is constructed. The PRA algorithm is used to evaluate the path support and score of semantic relationship edges, and the TransE consistency mechanism is used to remove untrusted node pairs, resulting in a structured semantic graph.
[0039] To enhance the fusion of semantics and behavior, resource co-occurrence pairs are extracted from behavior record sequences to construct a resource co-occurrence graph. The TF-IDF algorithm is used to calculate resource weights, which are then aligned with knowledge graph nodes using a dynamic time warping algorithm to generate a fused set of behavioral features. This set is then encoded using a multilayer perceptron structure to obtain a resource representation set. The cosine similarity between the representation set and the knowledge graph nodes is further calculated, and structurally similar nodes are identified using the SimRank algorithm. During the inference phase, the PCRA algorithm is used to perform frequency statistics and confidence scoring on the paths between mapped nodes and similar nodes, ultimately selecting the target resource sequence and generating a scheduling control strategy.
[0040] In actual testing, by comparing the performance of traditional keyword matching with that of this invention in a resource management scenario, multiple dimensions such as query response time, recommendation accuracy, resource scheduling efficiency, and system dynamic adaptability were evaluated, verifying that this invention has significant performance advantages. Table 1 is a statistical table of key performance data collected during the experiment: Table 1 Comparative Test Data of Information Resource Intelligent Management System
[0041] As can be seen from the comparative data in Table 1, this invention outperforms traditional rules in several key indicators, particularly in recommendation accuracy, response timeliness, and resource fusion capabilities. The average daily access volume increased by 29.7%, indicating higher resource allocation efficiency; the recommendation accuracy improved from 62.3% to 81.5%, demonstrating that the large model and knowledge graph structure significantly enhanced personalized recommendation performance in semantic reasoning and intelligent scheduling; the average retrieval time decreased by 43.8%, significantly improving user experience. Through automatic semantic path identification and knowledge graph structure optimization, the knowledge graph update cycle was compressed to 17 minutes, achieving higher dynamic adaptability; the improved resource utilization and heterogeneous fusion rate also indicate that the system possesses stronger knowledge organization and reuse capabilities.
[0042] Overall, the data in Table 1 verifies that the method of the present invention can effectively solve the problems of weak semantic understanding, strong behavioral fragmentation, and rigid scheduling strategies in traditional information resource management in actual business scenarios, and has good application and promotion value.
[0043] Example 2: Based on the same inventive concept, embodiments of the present invention also provide an intelligent information resource management system based on a large model and data-driven approach, the system comprising: The data acquisition module is used to acquire information resource sets and behavior datasets, merge them for preprocessing, and generate resource content sequences and behavior record sequences. The semantic analysis module is used to perform contextual semantic analysis on the resource content sequence based on a pre-trained BERT model and extract a set of semantic elements. The graph construction module is used to calculate the support and path score of the relationship edges between the semantic elements as nodes, filter the relationship edges according to the support and path score, and perform consistency verification on each filtered node to construct a knowledge graph. The feature extraction module is used to extract resource co-occurrence features from the behavior record sequence, align the behavior record sequence with each node in the knowledge graph in time, and generate a set of behavior features. The similarity recognition module is used to encode the behavioral feature set and map it to the knowledge graph, and identify the mapped nodes and similar nodes; The path filtering module is used to perform path frequency statistics and confidence scores on reachable paths between the mapped node and the similar node, filter the set of associated paths based on the frequency statistics and confidence scores, and generate a target resource sequence. The scheduling control module is used to schedule and control the information resource set according to the target resource sequence, write the scheduling result into the resource behavior set, and update the behavior record sequence.
[0044] Since these systems and the principles they address are similar to the aforementioned intelligent information resource management methods based on large models and data-driven approaches, the implementation of these systems can refer to the implementation of the aforementioned methods, and the repetitions will not be repeated.
[0045] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.
[0046] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A smart information resource management method based on large models and data-driven approaches, characterized in that: Includes the following steps: S1. Merge the information resource set and behavior dataset for preprocessing to generate a resource content sequence and a behavior record sequence; S2. Perform contextual semantic analysis on the resource content sequence based on the pre-trained BERT model to extract a set of semantic elements; S3. Using the semantic elements as nodes, calculate the support and path score the relationship edges between nodes, filter the relationship edges according to the support and path score, and perform consistency verification on each filtered node to construct a knowledge graph. S4. Extract resource co-occurrence features from the behavior record sequence, and perform time alignment between the behavior record sequence and each node in the knowledge graph to generate a behavior feature set; S5. Encode the behavioral feature set and map it to the knowledge graph to identify the mapped nodes and similar nodes; S6. Perform path frequency statistics and confidence scores on the reachable paths between the mapped node and the similar node, and filter the set of associated paths based on the frequency statistics and confidence scores to generate a target resource sequence; S7. Perform scheduling control on the information resource set according to the target resource sequence, write the scheduling result into the resource behavior set, and update the behavior record sequence.
2. The method as described in claim 1, characterized in that, In step S1, the information resource set represents a multi-source structured or unstructured data set containing timestamps to be managed, including documents, images, tables, logs, or audio; the behavior data set represents a set containing access behavior, operation records, and interaction log data; the preprocessing includes performing format normalization, noise reduction, and time calibration operations on the information resource set, and performing format standardization, missing value imputation, and time alignment operations on the behavior data set.
3. The method as described in claim 1, characterized in that, Step S2 specifically includes: S21. The resource content sequence is sequentially segmented according to the preset text boundary rules to obtain multiple text fragments, which are then generated into a text fragment sequence after being marked with position and bound to the sequence. S22. Input the text segment sequence into the pre-trained BERT model, perform context encoding on each text segment, extract the word vector set, and arrange it into a word vector sequence in the original order; S23. Construct a word co-occurrence graph using word vectors as nodes and adjacency and co-occurrence relationships of words within a text segment as edges; the co-occurrence relationship is defined as follows: if two words are separated by no more than k words, then they have a co-occurrence relationship, where k is a preset positive integer; S24. Initialize the weight values of each node in the word co-occurrence graph, perform multiple rounds of TextRank iteration updates until convergence, and output the word ranking results in descending order of weight values; S25. Filter the word sorting results according to the preset mapping template to obtain a keyword set. Perform aggregation operation on the vector of each keyword and summarize them to form a semantic element set according to the original order of the resource content sequence.
4. The method as described in claim 1, characterized in that, Step S3 specifically includes: S31. Using the semantic elements as nodes, traverse any two semantic elements to construct node pairs, and construct a set of directed relation edges based on the cosine similarity of each node pair and the position of each node in the set of semantic elements. S32. The PRA algorithm is used to calculate the support and path score of each relation edge, and relation edges whose support is lower than a preset reference value and whose path score is lower than a set score threshold are removed. S33. Use the TransE model to perform vector consistency verification on the node pairs corresponding to the remaining relation edges, retain the node pairs that meet the preset consistency conditions, and construct a set of connection edges. S34. Generate the knowledge graph based on the set of nodes and the set of connecting edges.
5. The method as described in claim 4, characterized in that, Step S32 specifically includes: S321. Starting from the source node of each relation edge, perform directed path traversal in the knowledge graph, extract all semantic paths not exceeding the set maximum number of hops, and filter out the candidate path set containing the relation edge. S322. Count the number of times each relation edge appears in the candidate path set, and calculate the ratio of the number of occurrences to the total number of semantic paths, as the path frequency support of the relation edge. S323. Calculate the path semantic score for each relation edge in the following manner: The number of nodes in each candidate path is used as the path structure length value; Extract the resource representation pairs of each consecutive node pair in the candidate path from the resource representation set and calculate the vector similarity. Use the mean of all vector similarities as the path semantic coherence value. The path structure length value and the path semantic coherence value are weighted and superimposed according to a set weight to obtain the path semantic score value of the relation edge. S324. Remove relation edges where the path frequency support is lower than a preset reference value and the path semantic score is lower than a set score threshold.
6. The method as described in claim 1, characterized in that, Step S4 specifically includes: S41. The behavior record sequence is expanded by time step, divided into multiple time segments according to the set sliding window length and step size, the accessed information resource pairs in each time segment are traversed, a resource co-occurrence pair set is constructed, and the frequency of their occurrence in the sliding window is counted to construct a co-occurrence frequency matrix. S42. Construct a resource co-occurrence graph based on the co-occurrence frequency matrix. The resource co-occurrence graph uses information resources as nodes and co-occurrence relationships as edges. Extract the nodes and edges of the resource co-occurrence graph, call the TF-IDF algorithm to calculate the weight value of each information resource and normalize it to generate a resource co-occurrence sequence. S43. Map the access time in the behavior record sequence to the resource co-occurrence sequence, use the dynamic time warping algorithm to pair the mapping result with the nodes in the knowledge graph, calculate the minimum alignment cost path, and obtain the alignment result set. S44. Merge the resource co-occurrence sequence with the alignment result set, bind each information resource to the corresponding alignment result at the tag level, construct a fused tag set, and arrange them in chronological order as the behavioral feature set.
7. The method as described in claim 6, characterized in that, Step S42 specifically includes: Based on the co-occurrence frequency matrix, a resource co-occurrence graph is constructed with information resources as nodes and non-zero co-occurrence frequencies as edges. The resource co-occurrence graph is an undirected graph, and the edge weight is the corresponding co-occurrence frequency. Extract all nodes and edges from the resource co-occurrence graph, assign a unique resource identifier to each information resource and establish an index mapping relationship, and call the TF-IDF algorithm to calculate the weighted frequency of co-occurrence between each information resource and all co-occurring resources to generate a set of weight values.
8. The method as described in claim 1, characterized in that, Step S5 specifically includes: S51. Use a multilayer perceptron to encode the behavioral feature set to generate an encoded feature set; S52. Calculate the cosine similarity between each encoded feature vector and all nodes in the knowledge graph, filter out nodes that are higher than a set threshold, and construct a set of mapped nodes. S53. Based on the mapped node set, call the SimRank algorithm to identify similar nodes in the knowledge graph and construct a similar node set; specifically including: Using each mapping node as the base node, traverse the knowledge graph to find nodes that have a direct connection edge or a two-hop path with it as candidate similar nodes. Count the number of adjacent nodes shared by each candidate similar node and the base node, and calculate the structural co-occurrence index; The co-occurrence indices of each structure are standardized using Z-Score and weighted by the edge weights in the knowledge graph to generate structural similarity values. Candidate nodes with structural similarity values higher than the preset similarity reference value are marked as similar nodes, and a set of similar nodes for the mapped node is constructed.
9. The method as described in claim 1, characterized in that, Step S6 specifically includes: S61. Using the PCRA algorithm, construct all candidate node pairs between the mapping node set and the similar node set to generate a candidate node pair set; S62. In the knowledge graph, with each candidate node pair as the search range, extract all paths from the mapping node to the similar node that do not exceed the set maximum number of hops, and remove invalid paths after structural validity check to construct a set of reachable paths; S63. Count the frequency of each reachable path in all candidate node pairs, calculate the ratio of its occurrence to the total number of candidate node pairs as the frequency index, and generate a path frequency record table. S64. Extract the cosine similarity of the encoded features of consecutive nodes in each reachable path, calculate its mean as the path similarity value, and perform a weighted sum with the path length of the path to generate a path score value and form a path score record table. S65. Filter reachable paths with frequency indicators higher than preset reference values and path score values higher than set score thresholds to form an associated path set. Extract information resources corresponding to all endpoint nodes in the associated path set, sort them in descending order of path score values, and generate a target resource sequence.
10. An intelligent information resource management system based on large models and data-driven approaches, characterized in that: The system comprises: The data acquisition module is used to acquire information resource sets and behavior datasets, merge them for preprocessing, and generate resource content sequences and behavior record sequences. The semantic analysis module is used to perform contextual semantic analysis on the resource content sequence based on a pre-trained BERT model and extract a set of semantic elements. The graph construction module is used to calculate the support and path score of the relationship edges between the semantic elements as nodes, filter the relationship edges according to the support and path score, and perform consistency verification on each filtered node to construct a knowledge graph. The feature extraction module is used to extract resource co-occurrence features from the behavior record sequence, align the behavior record sequence with each node in the knowledge graph in time, and generate a set of behavior features. The similarity recognition module is used to encode the behavioral feature set and map it to the knowledge graph, and identify the mapped nodes and similar nodes; The path filtering module is used to perform path frequency statistics and confidence scores on reachable paths between the mapped node and the similar node, filter the set of associated paths based on the frequency statistics and confidence scores, and generate a target resource sequence. The scheduling control module is used to schedule and control the information resource set according to the target resource sequence, write the scheduling result into the resource behavior set, and update the behavior record sequence.
