A large model-based education cognitive engine adaptive optimization method

Abstract

The application discloses an education cognition engine adaptive optimization method based on a large model, which comprises the following steps: S1, collecting teaching behavior data from an education database according to collection specifications, forming a unified format through an ETL tool, and mapping the behavior data into a cognitive state vector by combining a preset cognitive dimension through a cognitive signal encoder; S2, extracting cognitive correlations from the cognitive state vector through BERT, and obtaining an evolutionary knowledge graph with evolutionary markers through an evolutionary marker mechanism optimized by cognitive correlation weights; S3, based on the evolutionary knowledge graph, introducing an improved graph neural network to model cognitive evolution links and obtain a cognitive link evolution model, and generating an optimal candidate strategy list through the cognitive link evolution model; and S4, based on the optimal candidate strategy list, accessing a basic education cognition engine to automatically match teaching resources, effectively meeting the learning needs of different students, and improving the pertinence and effectiveness of teaching.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of education cognitive optimization, and in particular to an education cognitive engine adaptive optimization method based on a large model. BACKGROUND

[0002] With the penetration of artificial intelligence, knowledge graph and other technologies in the field of education, the digital transformation of education is accelerating. Learning data collection and analysis are possible, and student cognitive state, knowledge association and other data can support precise teaching. Teaching strategies need to be dynamically adjusted with the evolution of student cognition to build a complete closed loop of "teaching-learning-evaluation" and realize the iterative upgrade of the education ecosystem, which promotes the development of education towards intelligence and individualization.

[0003] Currently, in the aspect of cognitive modeling, traditional methods such as simple neural networks have obvious shortcomings. They often ignore the complex association of "knowledge point-knowledge point" and "student-knowledge point" in the knowledge graph, and only focus on the individual cognitive state of students, which cannot reflect the influence of knowledge association on cognitive development, such as the difficulty in measuring how the mastery of knowledge point B affects the complex application learning of knowledge point C. Moreover, the model has poor interpretability. After outputting the cognitive path probability, it cannot explain the underlying causes, and teachers cannot understand the reason for high probability, which makes it difficult to effectively apply to teaching decisions. The model cannot be updated in combination with the knowledge graph and real-time optimized with new learning data from students, and cannot keep up with the dynamic rhythm of student cognitive evolution. Therefore, an education cognitive engine adaptive optimization method based on a large model is proposed. SUMMARY

[0004] In order to overcome the above-mentioned defects of the prior art, in order to achieve the above-mentioned purpose, the present application proposes the following technical scheme:

[0005] An education cognitive engine adaptive optimization method based on a large model, comprising:

[0006] S1: Formulate collection specifications to collect teaching behavior data from an education database, form a unified format through an ETL tool, and map the behavior data to a cognitive state vector through a cognitive signal encoder combined with a preset cognitive dimension;

[0007] S2: Extract cognitive associations from the cognitive state vector through BERT, and obtain an evolution knowledge graph with evolution markers through an evolution marker mechanism optimized by cognitive association weights;

[0008] S3: Based on the evolution knowledge graph, introduce an improved graph neural network to model the cognitive evolution link and obtain a cognitive link evolution model, and generate an optimal candidate strategy list through the cognitive link evolution model;

[0009] The improved graph neural network realizes feature optimization and fusion coding by introducing node-level attention and relationship-level attention in the basic graph neural network;

[0010] S4: Accessing the basic education cognitive engine based on the optimal candidate strategy list to automatically match teaching resources.

[0011] The cognitive state vector acquisition process is:

[0012] The ETL tool is used to clean and convert the collected initial teaching behavior data to form a standardized data set;

[0013] The standardized teaching behavior data is disassembled to the preset cognitive dimensions by the cognitive signal encoder, the preset cognitive dimensions include the memory dimension, the understanding dimension and the application dimension, rules are designed for each dimension, the rule engine of the cognitive signal encoder is used to calculate the dimension values, and the cognitive state vector is spliced to realize the quantitative conversion of the behavior data to the cognitive state, and the formula is expressed as:

[0014] wherein, is the cognitive state vector of the i-th cognitive dimension, is the behavior data corresponding to the i-th cognitive dimension, is the weight of the i-th dimension, is a rule function;

[0015] The output cognitive state vector is wherein n is the number of cognitive dimensions.

[0016] The implementation process of the cognitive association extraction is:

[0017] The pre-trained BERT model is used to identify the entity types in the data, the entity types include students, knowledge points and error types, and the BERT model outputs the entity probability distribution P;

[0018] The pre-trained BERT model is used to judge the relationship between entities, and the relationship probability is calculated to achieve: ;

[0019] The relationship type is output, and the attribute is marked to generate a triple set of cognitive associations wherein is the subject entity, is the object entity, r is the relationship, and a is the attribute.

[0020] The implementation process of the evolution marking mechanism based on the cognitive association weight optimization is:

[0021] Evolution metadata is added to each triple, and the association generation time is recorded The number of updates c is associated with the teaching effect , obtain the ternary form of the evolution label ;

[0022] The relationship weight w represents the influence of the association, and the initial weight is set by the statistical frequency, and each time the teaching effect feedback data is obtained, the weight is updated according to the following formula:

[0023] ;

[0024] wherein, , is the score weight, is the learning rate, is the test score improvement value, is the homework accuracy change, indicates the teaching comprehensive index;

[0025] The weight of the cognitive association is adjusted by the update formula If is positive, the association weight increases, and if is negative, the association weight decreases. After each weight update, an evolution label is added to the cognitive association.

[0026] Finally, the relationship weight is calculated and updated according to the weight update formula, an evolution label is added, and the extracted and processed entities, relationships, and label information are stored in the graph database. Through the development of query statements and the establishment of indexes, the dynamic evolution of the knowledge graph is realized, and the evolution knowledge graph is obtained.

[0027] The implementation process of the improved graph neural network is as follows:

[0028] Introduce node-level attention and relationship-level attention in the basic graph neural network, and perform feature optimization and fusion coding

[0029] Fuse the node features and relationship features processed by the attention mechanism, and encode the current path nodes and the graph G through multi-layer GNN convolution operation to obtain the evolution potential feature GNN p,G) of each path node p.

[0030] Based on the GNN encoding output of each path node GNN p,G), use the softmax function to calculate the future path probability distribution:

[0031] ;

[0032] wherein, indicates the conditional probability, indicates the GNN encoding of the path node and the encoding output of the knowledge graph G, represents an exponential function, P represents a set of all possible cognitive path nodes, represents a specific cognitive path node in the set P;

[0033] Based on the output probability distribution result, the improved graph neural network is obtained in combination with the weight information of the attention mechanism.

[0034] The introduced node-level attention process is:

[0035] The attention weight of the current path node and other nodes in the knowledge graph is represented as:

[0036] ;

[0037] wherein , is the feature vector of node i, j, is the learnable weight matrix of , is the learnable bias term of , is the attention weight of node i to node j;

[0038] The introduced relationship-level attention process is:

[0039] For different relationship types in the knowledge graph, the relationship attention weight is learned, and the formula is:

[0040] ;

[0041] wherein, is the relationship feature vector, is the learnable weight matrix of , is the learnable bias term of , is the attention weight of relationship r.

[0042] The cognitive link evolution model acquisition process is:

[0043] The process of outputting the probability distribution result based on the improved graph neural network is taken as the model output process, historical teaching data including cognitive state vectors of students at different time periods, corresponding knowledge graph associated relationships and real cognitive evolution paths are collected, and a training set and a verification set are constructed;

[0044] The cross-entropy loss function is used to measure the difference between the predicted probability distribution and the real path And use the Adam optimizer to iteratively optimize the model parameters, the learning rate is initially set to 0.001, dynamically adjusted according to the training condition, and the model performance is evaluated using the validation set, and the model structure is adjusted according to the evaluation result to complete the cognitive evolution link modeling, and the cognitive link evolution model is obtained.

[0045] The process of outputting the optimal candidate strategy list is:

[0046] According to the cognitive link evolution model, the path probability distribution and entity association data of the evolution knowledge graph output by the cognitive evolution link modeling are input into the reasoning engine, and the strategy is matched according to the following steps:

[0047] Extract the cognitive path with a probability greater than 50%, match the path rule, traverse the path-strategy mapping rule, check whether the current knowledge graph data meets the preset rule condition, if the condition is met, trigger the teaching strategy corresponding to the rule, and form a candidate strategy list

[0048] The present application has the following beneficial effects:

[0049] In the present application, first, the improved GNN is used to fuse the knowledge graph association and the student cognitive state, and the accurate prediction of the cognitive path is realized. On the one hand, the evolution path of the student from the basic concept to the complex application can be dynamically tracked, and compared with the traditional method, the prediction accuracy is improved. On the other hand, combined with the specific node-level attention and the relationship-level attention, the complex application path probability and other results can be explained, such as the reason that the mastery relationship weight of knowledge point B is high, so that the understanding rate of the teacher is improved, and the teacher can accurately grasp the student cognitive development;

[0050] Secondly, the knowledge graph is changed from a static knowledge base to a dynamic cognitive engine. Not only the knowledge association is stored, but also the dynamic influence of the student cognitive-knowledge association is mined through GNN, and the knowledge association value is deepened;

[0051] Finally, according to the path probability distribution output by the cognitive link evolution model, combined with the entity association data of the evolution knowledge graph, the reasoning engine can filter out the high-probability path, and match the corresponding teaching strategy, for example, for the student who has a higher probability to enter the complex application path of knowledge point B, but the mastery weight of knowledge point B is low, the complex application reinforcement package of knowledge point B is accurately pushed, and the learning needs of different students are effectively met, and the pertinence and effectiveness of teaching are improved. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 A method step diagram of an education cognitive engine self-adaptive optimization method based on a large model is provided. DETAILED DESCRIPTION

[0053] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative efforts belong to the scope of the present application.

[0054] Embodiment one

[0055] As shown in the figure, the self-adaptive optimization method of the education cognitive engine based on a large model comprises: Figure 1

[0056] S1: Formulate collection specifications to collect teaching behavior data from an education database, form a unified format through an ETL tool, and map the behavior data to a cognitive state vector through a cognitive signal encoder combined with preset cognitive dimensions;

[0057] Deploy multiple types of collection terminals (such as intelligent terminals and teaching system interfaces) to cover all teaching scenarios, collect teacher and student operation behaviors, interactive content, and time sequence raw data as teaching behavior data through formulated collection specifications, and the specific process is as follows:

[0058] The collection specifications include defining data collection fields (student ID, behavior type, and timestamp), formats (such as JSON format standardized storage), and setting data collection frequency, and the data is collected according to the submission event, the system log is collected incrementally every hour, the fields required to be collected by each data collection terminal and interface are determined according to the teaching scene requirements and data analysis targets, and the data format conversion tool (JSONSchema verification tool) is used to ensure that the collected data format conforms to the specifications;

[0059] After the collection specifications are formulated, each collection terminal executes collection from the education database according to the collection specifications, and outputs initial teaching behavior data;

[0060] An ETL tool (such as ApacheNiFi) is used to clean (remove noise and duplicate data), convert (unify format and coding) the collected initial teaching behavior data, and form a standardized data set;

[0061] The cognitive signal encoder is used to disassemble the standardized teaching behavior data to preset cognitive dimensions, the preset cognitive dimensions include memory dimension, understanding dimension, and application dimension, rules are designed for each dimension, the rule engine of the cognitive signal encoder is used to calculate the dimension values, and the cognitive state vector is spliced, so as to realize the quantitative conversion of behavior data to cognitive state, and the formula is expressed as:

[0062] , wherein, ​is the cognitive state vector of the i-th cognitive dimension, is the behavior data corresponding to the i-th cognitive dimension, is the weight of the i-th dimension (set according to the teaching objectives, expert experience, memory dimension weight , understanding dimension weight , application dimension weight ), is a rule function based on threshold, interval, etc. Logic output quantization value between 0-1;

[0063] Specifically, the definition of cognitive dimensions is determined by joint subject teachers and education experts to cover cognitive dimensions throughout the teaching process, such as:

[0064] Memory dimension is related to recitation question score, knowledge point repetition accuracy, etc. Behavior;

[0065] The understanding dimension is related to the concept explanation question correct rate, classroom question and answer logic score, etc. Behavior;

[0066] The application dimension is related to homework application question completion rate, practical task standard degree, etc. Behavior;

[0067] The final output cognitive state vector , where n is the number of cognitive dimensions.

[0068] S2: Extract cognitive associations from the cognitive state vector through BERT, and obtain an evolutionary knowledge graph with evolutionary markers through an evolutionary marker mechanism optimized by cognitive association weights;

[0069] Use the pre-trained BERT model to identify entity types in the data, including students, knowledge points, and error types. The BERT model outputs an entity probability distribution P, which is calculated by the softmax function:

[0070] ;

[0071] where, is the probability of entity e, is the weight matrix, and b is the bias. When , (the confidence threshold is preset to 0.8), the entity is extracted;

[0072] For example, from the student A wrong knowledge point B, the entities e1=student A (type: student), e2=knowledge point B (type: knowledge point) are identified;

[0073] Pre-train the BERT model to determine the relationship between entities by calculating the relationship probability Implementation:

[0074] ​ ;

[0075] Output the relationship type (such as error mastery association) and mark the attribute (such as error type concept misunderstanding) to generate a set of triples of cognitive associations , where is the subject entity, is the object entity, r is the relationship, and a is the attribute;

[0076] Example: For entities Student A and Knowledge Point B, extract the relationship r = Error-Knowledge Point, attribute a = Concept Misunderstanding, and triple = (Student A, Error-Knowledge Point, Knowledge Point B, Concept Misunderstanding);

[0077] The implementation process of the evolutionary marking mechanism based on cognitive association weight optimization is as follows:

[0078] Add evolutionary metadata to each triple and record the association generation time (such as homework submission time) Update the number of times c (initially 1, +1 each time the weight is updated) Teaching Effect Association Value (such as student performance changes, knowledge point mastery rate changes), and obtain the marked triple in the form of

[0079] ;

[0080] Define the relationship weight w to represent the influence of the association, and the initial weight is set by the statistical frequency (such as the initial weight of Error-Knowledge Point ), and each time the teaching effect feedback data (such as the test score improvement value , homework accuracy change ) is obtained, the weight is updated according to the following formula:

[0081] ;

[0082] where, , is the performance weight, such as 0.6, which integrates multi-dimensional feedback, is the learning rate (experience value takes 0.1-0.3, controls the weight change amplitude), represents the teaching comprehensive index;

[0083] Adjust the weight of the cognitive association through the update formula If the test score of Student A for Knowledge Point B improves significantly (positive ), the association weight of Student A-Mastery-Knowledge Point B increases, reflecting the positive contribution of the association to learning progress;

[0084] If the homework accuracy decreases (negative If the weight of the association between student A and knowledge point B is reduced (or marked in the opposite direction), the weight of the association between student A and error knowledge point B is reduced (or marked in the opposite direction), and the influence of the association is corrected;

[0085] After each weight update, add an evolution mark to the cognitive association, including:

[0086] Time mark: Record the time point of weight update (e.g. 2025-07-15 10:00), trace the evolution node of the association;

[0087] Feedback source: Mark the type of feedback data that triggers this update (e.g. exam score improvement, homework accuracy decline), and clarify the evolution driving force;

[0088] Version iteration: Record the number of weight updates (e.g. 3rd update), reflect the maturity of the association;

[0089] Effect note: Briefly describe the impact of the update on teaching analysis (e.g. the weak warning priority of the association to knowledge point B is from high to medium);

[0090] Synchronize the new weight and evolution marked cognitive association to the knowledge graph in real time, including:

[0091] New association: If the teaching feedback discovers a new association (e.g. student A - application - knowledge point B), extract, assign weight and mark the first creation according to the initial process;

[0092] Association iteration: If the association weight and mark change (e.g. student A - error - knowledge point B weight reduction), update the attributes (weight value, evolution mark field) of the corresponding edge in the graph;

[0093] Specifically, choose a graph database (e.g. Neo4j) to store the knowledge graph, use its graph structure query advantage, develop Cypher query statements, support fast retrieval of complex relationships such as the best teaching strategy of the weak knowledge point association of a student and the mastery evolution path of a knowledge point in the class, and establish indexes (e.g. student ID index, knowledge point coding index) to optimize query performance and ensure single query response time ≤ 500ms;

[0094] Finally, calculate and update the relationship weight according to the weight update formula, add the evolution mark, store the extracted and processed entities, relationships and mark information in the graph database, and realize the dynamic evolution of the knowledge graph through the development of query statements and the establishment of indexes, and obtain the evolution knowledge graph.

[0095] S3: Based on the evolution knowledge graph, introduce an improved graph neural network to model the cognitive evolution link and obtain a cognitive link evolution model, and generate an optimal candidate strategy list through the cognitive link evolution model;

[0096] Extracting student cognition-related entity and relationship data from the evolutionary knowledge graph to form a cognitive association set , wherein is a subject entity, is an object entity, and r is a relationship, including a student cognition node (such as student A-cognition state: basic concept learning), a knowledge point mastery relationship (such as student A-mastery-knowledge point B), and a teaching behavior association (such as teacher C-explanation-knowledge point B);

[0097] Modeling the cognitive evolution link by an improved graph neural network learning path evolution rule using the current cognition state vector V of the student and the cognitive association set G;

[0098] Fusing node attribute features (such as knowledge point difficulty coefficients and student learning time) of the knowledge graph into the input, for a student node, concatenating the cognition state vector with the learning time and historical mistake number features, and for a knowledge point node, concatenating the knowledge point ID with the difficulty, associated knowledge point number, and other features to construct a spatio-temporal feature sequence;

[0099] Specifically, if there are multiple consecutive cognition state vectors of a student, concatenate them in chronological order to form a sequence input to assist the model in learning the temporal trend of cognitive evolution;

[0100] The improved graph neural network acquisition process is as follows:

[0101] Based on a graph neural network, node-level attention and relationship-level attention are integrated and feature optimization and fusion coding are performed, the implementation process is as follows:

[0102] Integrate node-level attention:

[0103] Calculate the attention weight of the current path node and other nodes in the knowledge graph, highlighting nodes that have a large impact on cognitive evolution, represented as:

[0104] ;

[0105] wherein , are the feature vectors of nodes i and j, is a learnable weight matrix of , is a learnable bias term of , is the attention weight of node i to node j, through which the neighbor node features are aggregated to enhance the influence of key associations;

[0106] Specifically, the attention weight of the current path node , the nodes with high weights have greater influence during feature aggregation, thus focusing on key nodes for cognitive evolution, such as knowledge points that students have just learned and have a high error rate, strengthening the influence of these nodes on subsequent path prediction, making the prediction more in line with real learning conditions, thereby avoiding irrelevant node interference and improving the ability to capture key cognitive nodes;

[0107] Then integrate the relationship-level attention:

[0108] For different relationship types in the knowledge graph (such as mastery, association, misunderstanding), learn the relationship attention weight, the formula is:

[0109] ;

[0110] where, is the relationship feature vector, is the learnable weight matrix of , is the learnable bias term of ,

[0111] Specifically, by learning the relationship attention weight , when aggregating relationship information, key relationships (such as students' misunderstanding of certain knowledge points) are given greater weight, and when analyzing students' cognitive paths, if there is a misunderstanding relationship, the strategy package generation process will focus on this relationship to determine whether the student will "correct misunderstanding and deepen cognition" or "get stuck in a learning bottleneck due to misunderstanding", improving the accuracy of cognitive path branch prediction;

[0112] Then perform feature optimization and fusion encoding, the implementation process is:

[0113] Fuse the node features and relationship features processed by the attention mechanism, and encode the current path node and the graph G through multi-layer GNN convolution (GraphSAGE convolution) to obtain the evolution potential feature GNN p,G) of each path node p, which is represented as:

[0114] , is the feature of node i at the kth layer, is the new feature vector of node i after the (k+1)th layer convolution, which is the result of concatenation, N i) is the neighbor set of node i, and the deep features of the node and its associated nodes are learned through iterative convolution;

[0115] Specifically, by integrating the node-level and relationship-level attention processed features, the different dimensional features are spliced to form richer input features, which contain the key information of nodes and relationships in the knowledge graph, providing comprehensive materials for subsequent encoding;

[0116] At the same time, through multi-layer GNN convolution operation, the node features are iteratively updated Each layer of convolution combines the neighbor node features, allowing the node features to continuously integrate surrounding associated information and learn more abstract and more reflective cognitive evolution trend features GNN layer by layer p,G), improving the effectiveness of path prediction;

[0117] Based on the GNN encoding output of each path node GNN p,G), the softmax function is used to calculate the future path probability distribution:

[0118] ;

[0119] Wherein, P (G | G, G) represents the conditional probability, i.e. the probability of evolving to path node and knowledge graph G under the condition of current path node , GNN (graph neural network) encoding output of path node and knowledge graph G, GNN will analyze the node association in the knowledge graph (such as the mastery relationship of students to knowledge points, the association between knowledge points) and score each path, the higher the score, the greater the evolution potential, P (G | G, G) represents the exponential function, P in the formula represents the set of all possible cognitive path nodes, represents a specific cognitive path node in set P.

[0120] Based on the output probability distribution result, combined with the weight information of the attention mechanism, the paths with higher probability are explained, including explaining which nodes and relationships contribute more to the path prediction (such as "the'mastery' relationship of knowledge point B contributes significantly to the 'complex application' path probability due to the high attention weight"), to help teachers understand the prediction logic and obtain an improved graph neural network;

[0121] The above process based on the improved graph neural network output probability distribution result is used as the model output process, historical teaching data is collected, including students' cognitive state vectors at different times, corresponding knowledge graph association relationships and real cognitive evolution paths (artificially annotated or derived through teaching records), to build training set and validation set;

[0122] The cross-entropy loss function is used to measure the difference between the predicted probability distribution and the real path And use the Adam optimizer to iteratively optimize the model parameters, the learning rate is initially set to 0.001, and is dynamically adjusted according to the training situation, and the model performance is evaluated using the validation set, and the model structure is adjusted (increase the number of convolution layers, adjust the attention mechanism parameters) to complete the cognitive evolution link modeling, and obtain the cognitive link evolution model;

[0123] The process of outputting the optimal candidate strategy list is:

[0124] According to the cognitive link evolution model, the path probability distribution output by the cognitive evolution link modeling and the entity association data of the evolution knowledge graph are input into the reasoning engine, and the strategy is matched according to the following steps:

[0125] Screening high-probability paths: Extract cognitive paths with a probability greater than a threshold (50%) as the core target of strategy reasoning (such as the "complex application" path probability 65%, selected);

[0126] Matching path rules: Traverse the "path-strategy mapping rule" and check whether the current knowledge graph data meets the preset rule conditions (such as whether the mastery weight of student B to knowledge point B is <0.6);

[0127] Generating candidate strategies: If the conditions are met, trigger the teaching strategy corresponding to the rule (such as pushing the "knowledge point B complex application reinforcement package"), and form a candidate strategy list.

[0128] S4: Based on the optimal candidate strategy list, access the basic education cognitive engine to automatically match teaching resources;

[0129] Access the optimal candidate strategy list to the basic learning engine, and execute according to the strategy type:

[0130] Resource recommendation strategy: Call the resource management interface to push the specified learning package to the student;

[0131] Supervision strategy: Trigger the teaching supervision system to generate pending tasks for teachers (such as "check the complex application of knowledge point B by student A").

[0132] And collect the teaching effect data after the strategy execution (such as student resource access rate, knowledge point mastery rate change), and return to:

[0133] Evolution knowledge graph: Update the associated weight (such as the weight of "student A - mastery - knowledge point B" from 0.5 to 0.7);

[0134] Re-train the model with new data to optimize the accuracy of cognitive path prediction.

[0135] In the application, several formulas involved are calculated by taking the values of the dimensionless formulas, and the establishment of the formulas is obtained by collecting a large amount of data to simulate the most recent real situation, and part of the coefficients or weights in the formulas is set by the person skilled in the art according to the actual situation, so here is not more elaborated.

[0136] The above embodiments can be implemented wholly or partially by software, hardware, firmware or any other combination. When implemented by software, the above embodiments can be implemented wholly or partially in the form of a computer program product. Those of ordinary skill in the art can realize that the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are executed in hardware or software depends on the specific application and design constraints of the technical solutions.

[0137] Although the embodiments of the present application have been shown and described, it can be understood by those of ordinary skill in the art that various changes, modifications, replacements and variations can be made to the embodiments without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. An adaptive optimization method for an educational cognitive engine based on a large model, characterized in that, include: S1: Establish collection standards to collect teaching behavior data from educational databases, form a unified format through ETL tools, and map the behavior data into cognitive state vectors by combining the cognitive signal encoder with the preset cognitive dimensions. S2: Extract cognitive associations from cognitive state vectors using BERT, and obtain an evolutionary knowledge graph with evolutionary labels through an evolutionary labeling mechanism that optimizes cognitive association weights; The evolutionary labeling mechanism optimized through cognitive association weights is implemented as follows: Add evolutionary metadata to each triple and record the time when the association was generated. Update count c, teaching effectiveness correlation value After obtaining the marked triplet form ; Define the relation weight w as the influence of the association, and the initial weight. Based on the statistical frequency setting, the weights are updated each time teaching effectiveness feedback data is obtained, according to the following formula: ； in, , As a weighting factor for grades, For learning rate, To improve exam scores For changes in homework accuracy, Indicates comprehensive teaching indicators; By updating the formula Adjust the weights of cognitive associations, if If it is positive, the association weight increases; if... If the value is negative, the association weight decreases, and an evolutionary marker is added to the cognitive association after each weight update. Finally, the relation weights are calculated and updated according to the weight update formula, evolutionary tags are added, and the extracted and processed entity, relation and tag information is stored in the graph database. By developing query statements and building indexes, the dynamic evolution of the knowledge graph is realized, and an evolutionary knowledge graph is obtained. S3: Based on the evolutionary knowledge graph, an improved graph neural network is introduced to model the cognitive evolutionary link to obtain the cognitive link evolutionary model, and the optimal candidate strategy list is generated through the cognitive link evolutionary model. The improved graph neural network is achieved by introducing node-level attention and relation-level attention into the basic graph neural network for feature optimization and fusion encoding. S4: Based on the optimal candidate strategy list, access the basic education cognitive engine to automatically match teaching resources.

2. The adaptive optimization method for an educational cognitive engine based on a large model according to claim 1, characterized in that, The process of obtaining the cognitive state vector is as follows: ETL tools were used to clean and transform the collected initial teaching behavior data to form a standardized dataset; Through a cognitive signal encoder, standardized teaching behavior data is decomposed into pre-defined cognitive dimensions, including memory, comprehension, and application. Rules are designed for each dimension, and the rule engine of the cognitive signal encoder calculates the dimension values, which are then concatenated into a cognitive state vector. This achieves a quantitative transformation from behavioral data to cognitive states, expressed by the following formula: ,in, Let i be the cognitive state vector of the i-th cognitive dimension. For the behavioral data corresponding to the i-th cognitive dimension, Let i be the weight of the i-th dimension. For regular functions; Output cognitive state vector , where n is the number of cognitive dimensions.

3. The adaptive optimization method for an educational cognitive engine based on a large model according to claim 1, characterized in that, The process of extracting cognitive associations is as follows: The pre-trained BERT model is used to identify entity types in the data, including students, knowledge points, and error types. The BERT model outputs the entity probability distribution P. Based on a pre-trained BERT model, relationships between entities are determined by calculating relationship probabilities. accomplish: ; Output the relation type and tag the attributes to generate a set of cognitive association triples. ,in As the main entity, Let r be the object entity, r be the relation, and a be the attribute.

4. The adaptive optimization method for an educational cognitive engine based on a large model according to claim 1, characterized in that, The improved graph neural network implementation process is as follows: Node-level attention and relation-level attention are introduced into the basic graph neural network, and feature optimization and fusion encoding are performed; The node features and relationship features processed by the attention mechanism are fused, and then processed through multi-layer GNN convolution operations on the nodes of the current path. By encoding the graph G, the evolutionary potential features of each path node are obtained. ; Evolutionary potential features of each path node based on GNN encoding output. (G), calculate the future path probability distribution using the softmax function: ； in, Represents conditional probability. This indicates that GNN pairs of path nodes. And the evolutionary potential characteristics of knowledge graph G, Represents an exponential function. In Represents the set of all possible cognitive path nodes; An improved graph neural network is obtained by combining the weight information of the attention mechanism with the output probability distribution results.

5. The adaptive optimization method for an educational cognitive engine based on a large model according to claim 4, characterized in that, The process of introducing node-level attention is as follows: Calculate the current path node The attention weights with other nodes in the knowledge graph are represented as follows: ； in , Let i be the feature vectors of nodes i and j. for The learnable weight matrix, for Learnable bias terms, Let be the attention weight of node i to node j; The process of introducing relation-level attention is as follows: For different relation types in a knowledge graph, the attention weights of the learned relations are calculated using the following formula: ； in, For relational feature vectors, for The learnable weight matrix, for Learnable bias terms, Let r be the attention weight of relation r.

6. The adaptive optimization method for an educational cognitive engine based on a large model according to claim 5, characterized in that, The process of obtaining the cognitive link evolution model is as follows: The process of outputting probability distribution results based on improved graph neural networks is used as the model output process. Historical teaching data is collected, including students' cognitive state vectors at different time periods, corresponding knowledge graph relationships and real cognitive evolution paths, to construct training and validation sets. The cross-entropy loss function is used to measure the difference between the predicted probability distribution and the actual path. The Adam optimizer was used to iteratively optimize the model parameters. The learning rate was initially set to 0.001 and dynamically adjusted according to the training situation. The model performance was evaluated using a validation set. Based on the evaluation results, the model structure was adjusted to complete the cognitive evolutionary link modeling and obtain the cognitive link evolution model.

7. The adaptive optimization method for an educational cognitive engine based on a large model according to claim 6, characterized in that, The process of generating the optimal candidate strategy list is as follows: Based on the cognitive link evolution model, the path probability distribution output by the cognitive evolution link modeling and the entity association data of the evolutionary knowledge graph are input into the inference engine, and the matching strategy is performed according to the following steps: Extract cognitive paths with a probability greater than 50%, match them with path rules, traverse the path-strategy mapping rules, check whether the current knowledge graph data meets the preset rule conditions, and if the conditions are met, trigger the teaching strategy corresponding to the rule to form a candidate strategy list.

Images