A dynamic educational resource allocation method based on a multi-modal discipline knowledge graph

Abstract

The present application relates to the technical field of multi-modal data fusion, discipline education knowledge graph and intelligent scheduling of education resources, and specifically relates to a dynamic education resource allocation method based on multi-modal discipline knowledge graph, which extracts entities, relationships and attributes based on discipline curriculum standards and textbook systems, fuses expert experience, teaching cases and student cognitive rules, constructs a multi-granularity knowledge graph covering all school stages, realizes automatic recognition of terms and relationship extraction using a BERT-BiLSTM-CRF joint model, guarantees an accuracy of greater than or equal to 98.5% by combining an artificial verification mechanism, builds a scheduling engine relying on a particle swarm optimization algorithm, integrates data such as resource reserves, teaching needs, teaching progress and regional distribution of each school to complete modeling, relies on algorithm iteration and deduction to continuously optimize the allocation scheme, quickly obtains an optimal scheduling scheme, completes the on-demand transfer of teaching resources, the complement of the balance and the supplement of the demand, and realizes the dynamic allocation of education resources.

Description

Technical Field

[0001] This invention relates to the fields of multimodal data fusion, subject-based educational knowledge graphs, and intelligent scheduling of educational resources, specifically to a dynamic educational resource allocation method based on a multimodal subject-based knowledge graph. Background Technology

[0002] With the advancement of policies promoting digitalization and equalization in education, teaching resources in schools at all levels are characterized by increased quantity, diverse types, and dispersed distribution. Traditional allocation of educational resources often relies on manual statistics, experience-based scheduling, and fixed allocation models. Currently, most publicly available methods for constructing educational knowledge graphs use a single model or are constructed manually, which presents the following problems.

[0003] 1. Single data modality: Most models rely solely on text data and do not integrate multimodal information such as video, images, and learning behavior. The scenarios are not fully covered, the semantic expression is one-sided, the knowledge system is disconnected from teaching resources, and it is impossible to achieve accurate matching based on knowledge points, teaching progress and cognitive patterns.

[0004] 2. Static and fixed graph: The ontology framework is fixed, lacking dynamic updates and autonomous iteration mechanisms. The extraction of entities, relationships, and attributes relies on manual methods, resulting in low efficiency, poor consistency, difficulty in adapting to the rapid iteration of the new curriculum standards, inability to adapt to curriculum standard updates and changes in teaching progress, poor timeliness, and inability to support the unified organization and dynamic scheduling of multimodal teaching resources such as text, images, audio, and video.

[0005] 3. The fusion method is superficial and detached from actual allocation: It mostly adopts simple feature splicing and fusion, without deeply exploring the deep semantic associations across modalities. The knowledge association logic is weak, and resource scheduling lacks quantitative models and intelligent algorithm support. The supply and demand mismatch, low utilization rate, and regional imbalance are prominent, and it does not have the ability to coordinate and allocate resources.

[0006] 4. Lack of closed-loop optimization: The lack of a resource usage feedback mechanism makes it impossible to optimize the graph structure and matching strategy based on the application effect, resulting in a gradual decline in long-term performance.

[0007] For example, Chinese invention patent CN121706907A, "A Method and System for Constructing an Educational Knowledge Graph Based on a Large Model," discloses an automatic extraction method for knowledge point entities, semantic relationships, and attribute information from educational text data using a large language model. This method, combined with a pre-defined subject ontology framework, constructs a knowledge graph, enabling rapid organization of educational text knowledge and intelligent retrieval of basic teaching content, learning path planning, and other basic applications, reducing the workload of traditional manual graph construction. The 36th issue of the *Journal of Xi'an University of Engineering* in 2022 published an educational resource allocation model based on differential evolution algorithms. This model uses a differential evolution algorithm to allocate educational resource data in a district, providing a basis for resource allocation for education management departments. However, the scaling factor in the algorithm usually uses a fixed value, lacking an adaptive adjustment mechanism, which reduces the model's convergence speed or causes it to get stuck in local optima, thus affecting the globally optimal resource allocation scheme. Therefore, a dynamic educational resource allocation method based on a multimodal subject knowledge graph is proposed. Summary of the Invention

[0008] To address the technical problems existing in the prior art, this invention provides a dynamic educational resource allocation method based on multimodal subject knowledge graphs.

[0009] To address the aforementioned technical problems, this invention provides the following technical solution: a dynamic educational resource allocation method based on a multimodal subject knowledge graph, comprising the following steps:

[0010] S1, based on subject curriculum standards and textbook system, collects multi-source data in the education field, integrates expert experience, teaching cases and students' cognitive patterns, constructs a multimodal education field dataset and performs preprocessing;

[0011] S2 uses the BERT-BiLSTM-CRF joint model to extract term entities and relations from the preprocessed multimodal education domain dataset and outputs a set of knowledge triples. Then, the output set of knowledge triples is corrected through a manual verification mechanism.

[0012] S3, based on the revised set of knowledge triples, constructs a multi-granularity, multi-modal subject knowledge graph covering the entire learning stage, and establishes a dynamic update and version management mechanism for the graph;

[0013] S4 uses the particle swarm optimization algorithm to build an educational resource scheduling engine, performs scheduling modeling, and constructs a comprehensive resource evaluation function;

[0014] S5 uses the resource comprehensive evaluation function as the fitness criterion and iteratively solves the optimal scheduling scheme through the particle swarm optimization algorithm.

[0015] S6 analyzes teaching resource needs based on multi-granularity, multi-modal subject knowledge graphs, and then performs educational resource allocation through the optimal scheduling scheme to realize on-demand flow of educational resources, surplus complementarity and demand supplementation, thus completing dynamic educational resource allocation.

[0016] Preferably, the preprocessing in step S1 includes: segmenting text data, removing stop words, and standardizing curriculum standard keywords; extracting frames, reducing features, and labeling image and audio / video multimodal data; and filling missing values ​​in the structured learning data.

[0017] Preferably, the execution flow of the BERT-BiLSTM-CRF joint model in step S2 includes the following steps:

[0018] S21, the BERT layer performs semantic encoding on multimodal text and outputs context-dependent word vector representations;

[0019] S22, the BiLSTM layer performs bidirectional temporal feature extraction on word vector representations, capturing the context-dependent features of entities and relations to obtain feature sequences;

[0020] S23, the CRF layer decodes the label sequence of the feature sequence, classifies and extracts entities and relations, and outputs a set of knowledge triples. The expression is:

[0021] ;

[0022] Where h is the head entity, t is the tail entity, r is the relationship between entities, H is the set of head entities, R is the set of relationships, and T is the set of tail entities;

[0023] S24. Manually verify and correct the set of knowledge triples to ensure that the accuracy of the knowledge triples satisfies the following formula:

[0024] ;

[0025] Where TP represents correctly identified entities or relationships, TN represents correctly excluded non-entities, FP represents misidentification, and FN represents omission.

[0026] Preferably, in step S24, the F1 score is calculated simultaneously during verification to comprehensively evaluate the entity and relation extraction performance. The F1 score is used to distinguish the priority of manual verification, identify defective samples extracted by the model, and guide the iterative optimization of the model. A higher F1 score indicates better balance performance in entity and relation extraction. The F1 expression is:

[0027] ;

[0028] Where P is precision, which characterizes the model's ability to correctly identify entities or relationships; and Rec is recall, which characterizes the model's ability to completely detect all valid entities or relationships.

[0029] Preferably, step S3 includes the following steps:

[0030] S31, the set of knowledge triples corrected in step S23 Based on this, we supplement the extracted attribute information of entities, improve the three elements of subject knowledge—entities, relationships, and attributes—and form a set of structured knowledge triples with attached attributes. Where A is the set of attributes for the corresponding entity, and the specific attribute information includes the grade level, textbook version, difficulty of knowledge points, type of supporting resources, resource parameters, and adapted teaching progress, represented as:

[0031] ;

[0032] in, to The following are the single-class attribute fields within the entity attribute set, For each stage of education; This is a textbook version. Based on the difficulty of the knowledge points, To match the types of resources, For resource parameters, To adapt to the teaching progress, k is the total number of entity attributes; S32, based on the knowledge triple set Construct a multi-granularity, multi-modal subject knowledge graph, store the graph in a graph database, the graph contains concepts, knowledge points, teaching resources, cognitive nodes, and supports the association mapping and binding retrieval of multi-modal teaching resources such as text, images, audio, and video;

[0033] S33, establish a dynamic update and version management mechanism for the map, perform incremental updates and version management of the map, record update logs, and support historical version backtracking;

[0034] S34 calculates the entity similarity of the multi-granularity multimodal subject knowledge graph, which is used to match the surplus or shortage of teaching resources with similar knowledge points in the resource scheduling stage. It serves as the matching weight basis for cross-school resource sharing. The calculated entity similarity is used as the cross-school resource sharing matching weight and input into the scheduling feature model. In the resource gap matching stage, it prioritizes the scheduling of supporting resources with high similarity knowledge points to achieve complementary substitution of similar teaching resources.

[0035] Preferably, step S32, constructing a multi-granularity, multi-modal subject knowledge graph, specifically includes the following steps:

[0036] S321, three-level granularity layering: entities are layered according to subject category - unit chapter - subdivided knowledge points, generating knowledge nodes at three levels: macro, meso, and micro. Each of the three levels of knowledge nodes generated by the layering is assigned a globally unique knowledge point code and stored in the node attribute set A.

[0037] S322, Multimodal resource mounting: Establish a connection between teaching resources and knowledge nodes, and write the resource format, duration, and storage path into the attribute set A. Among them, teaching resources include lesson plans, pictures, audio, and video.

[0038] S323, Graph Structure Mapping Storage: The Neo4j graph database is used to map and store the graph. The mapping rules satisfy the following conditions:

[0039] ;

[0040] ;

[0041] ;

[0042] ;

[0043] in, This is an entity mapping function that enables a one-to-one mapping of subject entities to Neo4j knowledge nodes; This is a relation mapping function that maps entity relationships to connecting edges between knowledge nodes; This is an attribute mapping function that transforms the set of entity attributes into attribute fields of knowledge nodes; The database node obtained by mapping the head entity h; The database node obtained by mapping the tail entity t; and All of these are subject knowledge nodes stored in the Neo4j database, covering three types of nodes: subject concepts, sub-knowledge points, and teaching resources; N is the set of all knowledge node edges in Neo4j; E is the set of all relation edges. P is the attribute vector of the knowledge node generated by mapping A; P is the set of attributes of all knowledge nodes in Neo4j.

[0044] Preferably, the formula for calculating the inter-entity similarity of the multi-granularity, multi-modal subject knowledge graph in step S34 is as follows:

[0045] ;

[0046] in, , All are entity vectors output by BERT-BiLSTM-CRF.

[0047] Preferably, the step S4 of building the resource scheduling engine includes the following steps:

[0048] S41 collects data on resource reserves, teaching needs, teaching progress, and geographical distribution of each school, cleans and aligns the data, that is, classifies the standardized data according to the grade level and subject labels. After cleaning and alignment, it reads the knowledge point codes built into the graph nodes and completes the binding mapping between each school's teaching resources and subdivided knowledge points.

[0049] S42 uses five dimensions—school number, resource category, knowledge point, teaching level, and geographical location—as feature fields to quantify the resource inventory, demand gap, and transportation and allocation costs of each school as feature values, constructing a multi-dimensional educational resource scheduling feature model; among these, the knowledge point dimension is introduced in step S34 to calculate the entity similarity. As a quantitative feature, it characterizes the degree of substitutability and adaptability of teaching resources for different knowledge points;

[0050] S43 uses the multi-dimensional educational resource scheduling feature model as the input data source for the particle swarm optimization algorithm to construct a comprehensive resource evaluation function, the expression of which is:

[0051] );

[0052] The constraints are:

[0053] =1;

[0054] in, , , These are the weighting coefficients; For the degree of supply and demand matching, E(X) represents the regional equilibrium level, which is the efficiency of resource use. This is a resource scheduling decision vector, which includes the allocation quantity, deployment direction, and scheduling priority.

[0055] Preferably, in step S5, the optimal scheduling scheme is iteratively solved using a particle swarm optimization algorithm, with the resource comprehensive evaluation function as the fitness criterion. Specifically:

[0056] During the first In the next iteration, the flight velocity of the i-th particle on the d-dimensional decision variable... The updated formula is:

[0057] ;

[0058] Among them, flight speed The speed represents the adjustment range of the scheduling plan; the greater the speed, the greater the adjustment range of the number of resources allocated and the schools to be deployed. Particle i represents a complete global resource allocation plan. Let d be the value of the decision variable corresponding to the i-th complete global resource allocation scheme in the d-th dimension during the k-th iteration of the particle swarm optimization; d is the dimension index in the resource scheduling decision vector. For candidate allocation tasks determined by school number, resource category, knowledge point, teaching level, and geographical location, three decision variables are set: allocation quantity, deployment direction code, and scheduling priority, and they are flattened in a preset order to form the resource scheduling decision vector; k is the current iteration number. , For learning factors; , It is a random number; For the individual's optimal; It is the global optimum; The adaptive inertia weight is expressed as:

[0059] ;

[0060] in, , These are the maximum and minimum inertia weights, respectively. This represents the maximum total number of iterations.

[0061] exist At time t, the current position of the i-th particle on the d-dimensional decision variable, and its updated value. Represented as:

[0062] ;

[0063] in, This refers to the specific parameters of the entire global allocation scheme in the corresponding dimension; the entire vector... This constitutes a complete resource allocation scheme; the particle swarm optimization algorithm uses adaptive inertia weights. The process involves iteration, expanding the search range in the early stages and improving local convergence accuracy in the later stages.

[0064] Preferably, step S6 includes the following steps:

[0065] S61, based on the knowledge point and resource adaptation requirements corresponding to the teaching needs of multi-granularity and multi-modal subject knowledge graph analysis;

[0066] S62 generates cross-school or cross-regional resource allocation instructions according to the optimal scheduling scheme, so as to realize targeted replenishment of scarce resources and sharing of surplus resources;

[0067] S63 collects scheduling execution data and teaching effect data, feeds them back to the resource scheduling engine to form a closed-loop optimization, and constructs a closed-loop control loop from resource push to state perception to graph update. It collects and evaluates learners' behavior and state in real time, and feeds it back to the graph and particle swarm optimization algorithm to adjust the weight of resource nodes and particle search direction.

[0068] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0069] 1. This invention is based on subject curriculum standards and textbook system, integrates expert experience, teaching cases and students' cognitive patterns, extracts three elements: entities, relationships and attributes, and constructs a multi-granularity knowledge graph covering the entire learning stage; it adopts the BERT-BiLSTM-CRF joint model to realize automatic term recognition and relationship extraction, and combines a manual verification mechanism to ensure an accuracy rate of ≥98.5%; it supports dynamic updates and version management of the graph, adapts to the needs of new curriculum standard iteration, and deeply integrates multimodal subject knowledge graph with particle swarm intelligent optimization algorithm to achieve global optimal, dynamic balance and closed-loop iterative resource scheduling;

[0070] 2. This invention relies on the particle swarm optimization algorithm to build a scheduling engine, integrates data such as resource reserves, teaching needs, teaching progress and geographical distribution of various schools to complete the model, sets the resource allocation quantity, allocation direction and scheduling priority as variables, establishes a comprehensive resource evaluation function around the supply and demand matching degree, usage efficiency and resource balance level, and continuously optimizes the allocation scheme through algorithm iteration and deduction, quickly obtains the optimal scheduling scheme, completes the on-demand flow of teaching resources, the complementarity of surplus and the supplementation of demand, and realizes the dynamic allocation of educational resources. Attached Figure Description

[0071] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0072] The present invention will be further described below with reference to the accompanying drawings and embodiments, which illustrate the above and other technical features and advantages of the present invention. However, the following embodiments are merely preferred embodiments of the present invention and are not exhaustive.

[0073] Example 1:

[0074] like Figure 1 As shown, this invention provides a dynamic educational resource allocation method based on a multimodal subject knowledge graph, comprising the following steps:

[0075] S1, based on subject curriculum standards and textbook system, collects multi-source data in the education field, integrates expert experience, teaching cases and students' cognitive patterns, constructs a multimodal education field dataset and performs preprocessing;

[0076] S2 uses the BERT-BiLSTM-CRF joint model to extract term entities and relations from the preprocessed multimodal education domain dataset and outputs a set of knowledge triples. Then, the output set of knowledge triples is corrected through a manual verification mechanism.

[0077] S3, based on the revised set of knowledge triples, constructs a multi-granularity, multi-modal subject knowledge graph covering the entire learning stage, and establishes a dynamic update and version management mechanism for the graph;

[0078] S4 uses the particle swarm optimization algorithm to build an educational resource scheduling engine, performs scheduling modeling, and constructs a comprehensive resource evaluation function;

[0079] S5 uses the resource comprehensive evaluation function as the fitness criterion and iteratively solves the optimal scheduling scheme through the particle swarm optimization algorithm.

[0080] S6 analyzes teaching resource needs based on multi-granularity, multi-modal subject knowledge graphs, and then performs educational resource allocation through the optimal scheduling scheme to realize on-demand flow of educational resources, surplus complementarity and demand supplementation, thus completing dynamic educational resource allocation.

[0081] In this embodiment, the preprocessing in step S1 includes: segmenting text data, removing stop words, and standardizing curriculum standard keywords; extracting frames, reducing features, and labeling image and audio / video multimodal data; and filling missing values ​​in the structured learning data.

[0082] In this embodiment, the execution flow of the BERT-BiLSTM-CRF joint model in step S2 includes the following steps:

[0083] S21, the BERT layer performs semantic encoding on multimodal text and outputs context-dependent word vector representations;

[0084] S22, the BiLSTM layer performs bidirectional temporal feature extraction on word vector representations, capturing the context-dependent features of entities and relations to obtain feature sequences;

[0085] S23, the CRF layer decodes the label sequence of the feature sequence, classifies and extracts entities and relations, and outputs a set of knowledge triples. The expression is:

[0086] ;

[0087] Where h is the head entity, t is the tail entity, r is the relationship between entities, H is the set of head entities, R is the set of relationships, and T is the set of tail entities;

[0088] S24, Manually verify and correct the set of knowledge triples to ensure that the accuracy of the knowledge triples satisfies the following formula:

[0089] ;

[0090] Where TP represents correctly identified entities or relationships, TN represents correctly excluded non-entities, FP represents misidentification, and FN represents omission.

[0091] During validation, the F1 score is calculated simultaneously to comprehensively evaluate the entity and relation extraction performance. The F1 score is used to distinguish the priority of manual validation, identify defective samples extracted by the model, and guide the iterative optimization of the model. A higher F1 score indicates better balanced performance in entity and relation extraction. The F1 expression is:

[0092] ;

[0093] Where P is precision, which characterizes the model's ability to correctly identify entities or relationships; and Rec is recall, which characterizes the model's ability to completely detect all valid entities or relationships.

[0094] Accuracy Acc and The dual-validation standard cannot distinguish between false positives and false negatives based solely on accuracy; after batch extraction of triples, Acc and... , All batches falling below the threshold were manually reviewed one by one. Only sampling and verification are performed on batches that meet the standards, thereby reducing manual workload while ensuring the quality of knowledge extraction.

[0095] In this embodiment, step S3 includes the following steps:

[0096] S31, the set of knowledge triples corrected in step S23 Based on this, we supplement the extracted attribute information of entities, improve the three elements of subject knowledge—entities, relationships, and attributes—and form a set of structured knowledge triples with attached attributes. Where A is the set of attributes for the corresponding entity, and the specific attribute information includes the grade level, textbook version, difficulty of knowledge points, type of supporting resources, resource parameters, and adapted teaching progress, represented as:

[0097] ;

[0098] in, to The following are the single-class attribute fields within the entity attribute set, For each stage of education; This is a textbook version. Based on the difficulty of the knowledge points, To match the types of resources, For resource parameters, To adapt to the teaching progress, k is the total number of entity attributes;

[0099] S32, based on knowledge triplet set Construct a multi-granularity, multi-modal subject knowledge graph and store the graph (i.e., multi-granularity, multi-modal subject knowledge graph) through a graph database. The graph includes concepts, knowledge points, teaching resources, and cognitive nodes, and supports the association mapping and binding retrieval of multi-modal teaching resources such as text, images, audio, and video.

[0100] S33, establish a dynamic update and version management mechanism for the map, perform incremental updates and version management of the map, record update logs, and support historical version backtracking;

[0101] S34 calculates the inter-entity similarity of the multi-granularity, multi-modal subject knowledge graph. This similarity is used in the resource scheduling phase to match surplus or shortage teaching resources with similar knowledge points. It serves as the matching weight for cross-school resource sharing. The calculated entity similarity is used as the cross-school resource sharing matching weight and input into the scheduling feature model. In the resource gap matching phase, resources with high similarity knowledge points are prioritized for scheduling, achieving complementary substitution of similar teaching resources. The formula for calculating the inter-entity similarity of the multi-granularity, multi-modal subject knowledge graph is as follows:

[0102] ;

[0103] in, , All are entity vectors output by BERT-BiLSTM-CRF.

[0104] In this embodiment, step S32, constructing a multi-granularity, multi-modal subject knowledge graph, specifically includes the following steps:

[0105] S321, three-level granularity layering: entities are layered according to subject category - unit chapter - subdivided knowledge points, generating knowledge nodes at three levels: macro, meso, and micro. Each of the three levels of knowledge nodes generated by the layering is assigned a globally unique knowledge point code and stored in the node attribute set A.

[0106] S322, Multimodal resource mounting: Establish a connection between teaching resources and knowledge nodes, and write the resource format, duration, and storage path into the attribute set A. Among them, teaching resources include lesson plans, pictures, audio, and video.

[0107] S323, Graph Structure Mapping Storage: The graph is stored using the Neo4j graph database, and the mapping rules meet the following conditions:

[0108] ;

[0109] ;

[0110] ;

[0111] ;

[0112] in, This is an entity mapping function that enables a one-to-one mapping of subject entities to Neo4j knowledge nodes; This is a relation mapping function that maps entity relationships to connecting edges between knowledge nodes; This is an attribute mapping function that transforms the set of entity attributes into attribute fields of knowledge nodes; The database node obtained by mapping the head entity h; The database node obtained by mapping the tail entity t; and All of these are subject knowledge nodes stored in the Neo4j database, covering three types of nodes: subject concepts, sub-knowledge points, and teaching resources; N is the set of all knowledge nodes in Neo4j; E is the set of all relation edges. P is the attribute vector of the knowledge node generated by mapping A; P is the set of attributes of all knowledge nodes in Neo4j.

[0113] In this embodiment, the step S4 of building the resource scheduling engine includes the following steps:

[0114] S41 collects data on resource reserves, teaching needs, teaching progress, and geographical distribution of each school, cleans and aligns the data, that is, classifies the standardized data according to the grade level and subject labels. After cleaning and alignment, it reads the knowledge point codes built into the graph nodes and completes the binding mapping between each school's teaching resources and subdivided knowledge points.

[0115] S42 uses five dimensions—school number, resource category, knowledge point, teaching level, and geographical location—as feature fields to quantify the resource inventory, demand gap, and transportation and allocation costs of each school as feature values, constructing a multi-dimensional educational resource scheduling feature model; among these, the knowledge point dimension is introduced in step S34 to calculate the entity similarity. As a quantitative feature, it characterizes the degree of substitutability and adaptability of teaching resources for different knowledge points;

[0116] S43 uses the multi-dimensional educational resource scheduling feature model as the input data source for the particle swarm optimization algorithm to construct a comprehensive resource evaluation function, the expression of which is:

[0117] );

[0118] The constraints are:

[0119] =1;

[0120] in, , , These are the weighting coefficients; For the degree of supply and demand matching, E(X) represents the regional equilibrium level, which is the efficiency of resource use. This is a resource scheduling decision vector, which includes the allocation quantity, deployment direction, and scheduling priority.

[0121] In this embodiment, step S5 uses the resource comprehensive evaluation function as the fitness criterion and iteratively solves the optimal scheduling scheme through particle swarm optimization algorithm, specifically as follows:

[0122] During the first In the next iteration, the flight velocity of the i-th particle on the d-dimensional decision variable... The updated formula is:

[0123] ;

[0124] Among them, flight speed The adjustment range of the scheduling scheme represents the amount of adjustment of this decision variable between two adjacent iterations. The greater the speed, the greater the adjustment range of the number of resources allocated and the schools to be deployed. Particle i is a complete global resource allocation scheme. For the k-th iteration of the particle swarm optimization, the value of the decision variable corresponding to the d-th dimension of the complete global resource allocation scheme of the i-th set is taken; d is the dimension index in the resource scheduling decision vector; for the candidate allocation task determined by school number, resource category, knowledge point, teaching level and geographical location, three decision variables are set respectively: allocation quantity, deployment direction code and scheduling priority, and flattened in a preset order to form the resource scheduling decision vector; k is the current iteration number; , For learning factors; , It is a random number; For the individual's optimal; It is the global optimum; The adaptive inertia weight is expressed as:

[0125] ;

[0126] in, , These are the maximum and minimum inertia weights, respectively. This represents the maximum total number of iterations.

[0127] exist At time t, the current position of the i-th particle on the d-dimensional decision variable (i.e., the position of the i-th particle in this global resource allocation scheme). (the current value of each decision variable), and its updated value. Represented as:

[0128] ;

[0129] in, This refers to the specific parameters of the global allocation scheme in the corresponding dimension; the entire vector... This constitutes a complete resource allocation scheme; the particle swarm optimization algorithm uses adaptive inertia weights. The process involves iteration, expanding the search range in the early stages and improving local convergence accuracy in the later stages.

[0130] In this embodiment, step S6 includes the following steps:

[0131] S61, based on the knowledge point and resource adaptation requirements corresponding to the teaching needs of multi-granularity and multi-modal subject knowledge graph analysis;

[0132] S62 generates cross-school or cross-regional resource allocation instructions according to the optimal scheduling scheme, so as to realize targeted replenishment of scarce resources and sharing of surplus resources;

[0133] S63: Collect scheduling execution data and teaching effect data, and use Pearson correlation coefficient + weighted linear regression algorithm for back-correction. Feedback is sent to the resource scheduling engine to form a closed-loop optimization, constructing a closed-loop control loop from resource push to state perception to graph update. Real-time collection and evaluation of learners' behavior and state are collected and fed back to the graph and particle swarm optimization algorithm to adjust the weight of resource nodes (corresponding to teaching resource nodes such as lesson plans, pictures, audio and video that are associated with knowledge nodes in step S322) and particle search direction.

[0134] In this embodiment, a dynamic educational resource allocation system based on a multimodal subject knowledge graph is also disclosed, including a data acquisition module, a data preprocessing module, a BERT-BiLSTM-CRF extraction module, a manual verification module, a graph storage module, a version management module, a data integration module, a particle swarm optimization engine module, an evaluation function calculation module, an optimal strategy output module, a resource allocation module, and a feedback optimization module.

[0135] The data acquisition module collects multi-source data in the education field, integrates expert experience, teaching cases, and student cognitive patterns, and transmits it to the data preprocessing module. The data preprocessing module preprocesses the multimodal education data. The BERT-BiLSTM-CRF extraction module extracts terms, entities, and relationships from the preprocessed multimodal education dataset. The manual verification module corrects the knowledge triple set through a manual verification mechanism. The knowledge graph storage module stores the knowledge graph. The version management module performs incremental updates and version management of the knowledge graph. The data integration module collects data on resource reserves, teaching needs, teaching progress, and geographical distribution of various schools. The particle swarm optimization engine module builds a resource scheduling engine. The evaluation function calculation module constructs a comprehensive resource evaluation function. The optimal strategy output module iteratively solves for the optimal scheduling scheme. The resource allocation module executes the allocation of educational resources through the optimal scheduling scheme. The feedback optimization module collects and evaluates learners' behavior and status in real time, feeding back to the knowledge graph and particle swarm optimization algorithm to adjust the weights of resource nodes and the particle search direction.

[0136] In this embodiment, a computer-readable storage medium is also disclosed, on which a computer program is stored, and when the computer program is executed by a processor, it implements the above-described method.

[0137] In this embodiment, an electronic device is also disclosed, including a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described method.

[0138] Example 2:

[0139] In this embodiment, the host connects to the resource servers of various primary and secondary schools in the area via a gigabit Ethernet interface, an HDMI interface connects to a large visual management screen, a USB interface connects to a student learning data collection terminal, and a PoE port powers the campus resource monitoring equipment. Simultaneously, it connects to teachers' and students' mobile terminals via a wireless module to achieve multi-terminal data exchange and management. When a primary school raises an urgent need for micro-lesson resources in mathematics, the host responds through the following steps:

[0140] S1: Collect data on teachers, laboratories, multimedia equipment, curriculum resources, teaching progress, and geographical location of 20 schools within the region;

[0141] S2: The decision vector X is the resource allocation quantity, deployment direction, and scheduling priority;

[0142] S3: Construct a multi-objective evaluation function F(X), and take... =0.4, =0.3, =0.3;

[0143] S4: Set the particle population size to 30 and iterate 200 times. =0.9, =0.4, ;

[0144] S5: Particle swarm optimization, output the optimal scheduling scheme, specifically: select three schools with surplus micro-lesson resources in the region, transfer a total of 127 math micro-lesson resources, prioritize the allocation to nearby schools with shortages, and store the remaining resources in the regional resource public database.

[0145] S6: According to the scheduling plan, send resource retrieval instructions to schools with idle resources, complete the batch transmission of micro-lesson resources through the network interface, complete cross-school resource scheduling, and update the status of multimodal subject knowledge graph resources in sync. Display the resource flow and transmission progress on the visualization management screen, and push the allocation completion notification to the management personnel through the mobile terminal. The whole process is expected to take 4 minutes, realize the rapid filling of resource gaps, and match the data in Table 1.

[0146] Through actual deployment and verification, this invention can stably achieve a knowledge extraction accuracy rate of ≥98.5%, a resource scheduling response time of ≤3 seconds, and a significant improvement in resource balance. It is suitable for regional education resource coordination platforms. Specific data are shown in Table 1.

[0147] Table 1. Performance Comparison Data of Dynamic Educational Resource Allocation Methods

[0148]

[0149] The above description is merely a preferred embodiment of the present invention and is illustrative rather than restrictive. Those skilled in the art will understand that many changes, modifications, and even equivalents can be made within the spirit and scope defined by the claims of the present invention, all of which will fall within the protection scope of the present invention.

Claims

1. A dynamic educational resource allocation method based on multimodal subject knowledge graphs, characterized in that, Includes the following steps: S1, based on subject curriculum standards and textbook system, collects multi-source data in the education field, integrates expert experience, teaching cases and students' cognitive patterns, constructs a multimodal education field dataset and performs preprocessing; S2 uses the BERT-BiLSTM-CRF joint model to extract term entities and relations from the preprocessed multimodal education domain dataset and outputs a set of knowledge triples. Then, the output set of knowledge triples is corrected through a manual verification mechanism. S3, based on the revised set of knowledge triples, constructs a multi-granularity, multi-modal subject knowledge graph covering the entire learning stage, and establishes a dynamic update and version management mechanism for the graph; S4 uses the particle swarm optimization algorithm to build an educational resource scheduling engine, performs scheduling modeling, and constructs a comprehensive resource evaluation function; S5 uses the resource comprehensive evaluation function as the fitness criterion and iteratively solves the optimal scheduling scheme through the particle swarm optimization algorithm. S6 analyzes teaching resource needs based on multi-granularity, multi-modal subject knowledge graphs, and then executes educational resource allocation through the optimal scheduling scheme to realize on-demand flow of educational resources, surplus complementarity and demand supplementation, thus completing dynamic educational resource allocation.

2. The dynamic educational resource allocation method based on multimodal subject knowledge graph as described in claim 1, characterized in that, The preprocessing in step S1 includes: segmenting text data, removing stop words, and standardizing curriculum standard keywords; extracting frames, reducing features, and labeling multimodal data of images and audio / video; and filling missing values ​​in structured learning data.

3. The dynamic educational resource allocation method based on multimodal subject knowledge graph as described in claim 1, characterized in that, The execution process of the BERT-BiLSTM-CRF joint model in step S2 includes the following steps: S21, the BERT layer performs semantic encoding on multimodal text and outputs context-dependent word vector representations; S22, the BiLSTM layer performs bidirectional temporal feature extraction on word vector representations, capturing the context-dependent features of entities and relations to obtain feature sequences; S23, the CRF layer decodes the label sequence of the feature sequence, classifies and extracts entities and relations, and outputs a set of knowledge triples. The expression is: ； Where h is the head entity, t is the tail entity, r is the relationship between entities, H is the set of head entities, R is the set of relationships, and T is the set of tail entities; S24, Manually verify and correct the set of knowledge triples to ensure that the accuracy of the knowledge triples satisfies the following formula: ； Where TP represents correctly identified entities or relationships, TN represents correctly excluded non-entities, FP represents misidentification, and FN represents omission.

4. The dynamic educational resource allocation method based on multimodal subject knowledge graph as described in claim 2, characterized in that, In step S24, during verification, the F1 score is calculated simultaneously to comprehensively evaluate the entity and relation extraction performance. The F1 score is used to distinguish the priority of manual verification, identify defective samples extracted by the model, and guide the iterative optimization of the model. A higher F1 score indicates better balanced performance in entity and relation extraction. The F1 expression is: ； Where P is precision, which characterizes the model's ability to correctly identify entities or relationships; and Rec is recall, which characterizes the model's ability to completely detect all valid entities or relationships.

5. The dynamic educational resource allocation method based on multimodal subject knowledge graph as described in claim 1, characterized in that, Step S3 includes the following steps: S31, the set of knowledge triples corrected in step S23 Based on this, we supplement the extracted attribute information of entities, improve the three elements of subject knowledge—entities, relationships, and attributes—and form a set of structured knowledge triples with attached attributes. Where A is the set of attributes for the corresponding entity, and the specific attribute information includes the grade level, textbook version, difficulty of knowledge points, type of supporting resources, resource parameters, and adapted teaching progress, represented as: ； in, to The following are the single-class attribute fields within the entity attribute set, For each stage of education; This is a textbook version. Based on the difficulty of the knowledge points, To match the types of resources, For resource parameters, To adapt to the teaching progress, k is the total number of entity attributes; S32, based on knowledge triplet set Construct a multi-granularity, multi-modal subject knowledge graph, store the graph in a graph database, the graph contains concepts, knowledge points, teaching resources, cognitive nodes, and supports the association mapping and binding retrieval of multi-modal teaching resources such as text, images, audio, and video; S33, establish a dynamic update and version management mechanism for the map, perform incremental updates and version management of the map, record update logs, and support historical version backtracking; S34 calculates the entity similarity of the multi-granularity multimodal subject knowledge graph, which is used to match the surplus or shortage of teaching resources with similar knowledge points in the resource scheduling stage. It serves as the matching weight basis for cross-school resource sharing. The calculated entity similarity is used as the cross-school resource sharing matching weight and input into the scheduling feature model. In the resource gap matching stage, it prioritizes the scheduling of supporting resources with high similarity knowledge points to achieve complementary substitution of similar teaching resources.

6. A dynamic educational resource allocation method based on a multimodal subject knowledge graph as described in claim 5, characterized in that, Step S32, constructing a multi-granularity, multi-modal subject knowledge graph, specifically includes the following steps: S321, three-level granularity layering: entities are layered according to subject category - unit chapter - subdivided knowledge points, generating knowledge nodes at three levels: macro, meso, and micro. Each of the three levels of knowledge nodes generated by the layering is assigned a globally unique knowledge point code and stored in the node attribute set A. S322, Multimodal resource mounting: Establish a connection between teaching resources and knowledge nodes, and write the resource format, duration, and storage path into the attribute set A. Among them, teaching resources include lesson plans, pictures, audio, and video. S323, Graph Structure Mapping Storage: The graph is stored using the Neo4j graph database, and the mapping rules meet the following conditions: ； ； ； ； in, This is an entity mapping function that enables a one-to-one mapping of subject entities to Neo4j knowledge nodes. This is a relational mapping function that maps entity relationships to connecting edges between knowledge nodes. This is an attribute mapping function that transforms the set of entity attributes into attribute fields of knowledge nodes; The database node obtained by mapping the head entity h; The database node obtained by mapping the tail entity t; and All of these are subject knowledge nodes stored in the Neo4j database, covering three types of nodes: subject concepts, sub-knowledge points, and teaching resources; N is the set of all knowledge nodes in Neo4j; E is the set of all relation edges. P is the attribute vector of the knowledge node generated by mapping A; P is the set of attributes of all knowledge nodes in Neo4j.

7. The dynamic educational resource allocation method based on multimodal subject knowledge graph as described in claim 5, characterized in that, The formula for calculating the inter-entity similarity of the multi-granularity, multi-modal subject knowledge graph in step S34 is as follows: ； in, , All are entity vectors output by BERT-BiLSTM-CRF.

8. The dynamic educational resource allocation method based on multimodal subject knowledge graph as described in claim 1, characterized in that, The step S4 of building the resource scheduling engine includes the following steps: S41 collects data on resource reserves, teaching needs, teaching progress, and geographical distribution of each school, cleans and aligns the data, that is, classifies the standardized data according to the grade level and subject labels. After cleaning and alignment, it reads the knowledge point codes built into the graph nodes and completes the binding mapping between each school's teaching resources and subdivided knowledge points. S42 uses five dimensions—school number, resource category, knowledge point, teaching level, and geographical location—as feature fields to quantify the resource inventory, demand gap, and transportation and allocation costs of each school as feature values, constructing a multi-dimensional educational resource scheduling feature model; among these, the knowledge point dimension is introduced in step S34 to calculate the entity similarity. As a quantitative feature, it characterizes the degree of substitutability and adaptability of teaching resources for different knowledge points; S43 uses the multi-dimensional educational resource scheduling feature model as the input data source for the particle swarm optimization algorithm to construct a comprehensive resource evaluation function, the expression of which is: )； The constraints are: =1； in, , , These are the weighting coefficients; For the degree of supply and demand matching, E(X) represents the regional equilibrium level, which is the efficiency of resource use. This is a resource scheduling decision vector, which includes the allocation quantity, deployment direction, and scheduling priority.

9. The dynamic educational resource allocation method based on multimodal subject knowledge graph as described in claim 1, characterized in that, In step S5, the optimal scheduling scheme is iteratively solved using the particle swarm optimization algorithm, with the resource comprehensive evaluation function as the fitness criterion. Specifically: During the first In the next iteration, the flight velocity of the i-th particle on the d-dimensional decision variable... The updated formula is: ； Among them, flight speed The speed represents the adjustment range of the scheduling plan; the greater the speed, the greater the adjustment range of the number of resources allocated and the schools to be deployed. Particle i represents a complete global resource allocation plan. Let d be the value of the decision variable corresponding to the i-th complete global resource allocation scheme in the d-th dimension during the k-th iteration of the particle swarm optimization; d is the dimension index in the resource scheduling decision vector. For candidate allocation tasks determined by school number, resource category, knowledge point, teaching level, and geographical location, three decision variables are set: allocation quantity, deployment direction code, and scheduling priority, and they are flattened in a preset order to form the resource scheduling decision vector; k is the current iteration number. , For learning factors; , It is a random number; For the individual's optimal; It is the global optimum; The adaptive inertia weight is expressed as: ； in, , These are the maximum and minimum inertia weights, respectively. This represents the maximum total number of iterations. exist At time t, the current position of the i-th particle on the d-dimensional decision variable, and its updated value. Represented as: ； in, This refers to the specific parameters of the entire global allocation scheme in the corresponding dimension; the entire vector... This constitutes a complete resource allocation scheme; the particle swarm optimization algorithm uses adaptive inertia weights. The process involves iteration, expanding the search range in the early stages and improving local convergence accuracy in the later stages.

10. The dynamic educational resource allocation method based on multimodal subject knowledge graph as described in claim 1, characterized in that, Step S6 includes the following steps: S61, Based on the knowledge points and resource adaptation requirements corresponding to the teaching needs of graph analysis; S62 generates cross-school or cross-regional resource allocation instructions according to the optimal scheduling scheme, so as to realize targeted replenishment of scarce resources and sharing of surplus resources; S63 collects scheduling execution data and teaching effect data, feeds them back to the resource scheduling engine to form a closed-loop optimization, and constructs a closed-loop control loop from resource push to state perception to graph update. It collects and evaluates learners' behavior and state in real time, and feeds it back to the graph and particle swarm optimization algorithm to adjust the weight of resource nodes and particle search direction.

Images