Large model-based personalized learning generation method and device, equipment and medium

Abstract

This invention relates to the technical field of personalized learning. Specifically, it relates to a method, apparatus, device, and medium for generating personalized learning based on a large-scale model. The method includes feeding a learning objective and a cognitive state vector as joint input information into a large language model to generate a set of knowledge points corresponding to the learning objective and learning gap information. The knowledge point set is then aligned with prerequisite and dependency relationships in a preset educational knowledge graph to obtain an alignment result. When the difference between the updated cognitive state vector and the current cognitive state vector exceeds a preset cognitive state change threshold, semantic parsing and reasoning of the large language model are re-executed based on the updated cognitive state vector and the learning objective to obtain a new set of knowledge points. A new learning resource sequence is determined according to the updated personalized learning path, and the learning resource sequence is pushed to the user terminal. This invention achieves closed-loop optimization of the learning scheme.

Description

Technical Field

[0001] This invention relates to the technical field of personalized learning, and in particular to methods, apparatus, devices and media for generating personalized learning based on large models. Background Technology

[0002] With the development of artificial intelligence and big data analytics, the level of intelligent education is constantly improving, evolving from traditional question-based learning to personalized learning systems that can recommend content based on learner characteristics. Existing online education platforms typically use rule engines, knowledge point tags, or collaborative filtering algorithms to push content. These methods rely on users' historical behavioral data and shallow semantic matching, lacking deep modeling of the semantic structure of learning content and learners' cognitive states, resulting in insufficient targeting and accuracy of recommended content. Summary of the Invention

[0003] To achieve closed-loop optimization of the learning scheme, this application provides a method, apparatus, device, and medium for personalized learning generation based on a large model.

[0004] The above-mentioned objective of this application is achieved through the following technical solution:

[0005] A personalized learning generation method based on a large model, comprising:

[0006] Obtain user learning behavior information;

[0007] Based on learning behavior information, a cognitive state vector is calculated using a pre-trained cognitive state assessment model.

[0008] The learning objectives input by the user are obtained, and the learning objectives and cognitive state vectors are fed into the large language model as joint inputs to generate a set of knowledge points corresponding to the learning objectives and learning gap information.

[0009] Align the set of knowledge points with the prerequisite and dependency relationships in the pre-defined educational knowledge graph to obtain the alignment results;

[0010] Based on learning gap information, prerequisite knowledge points are inserted into the alignment results, the order of knowledge points and connection constraints are determined, and a personalized learning path that conforms to the logical order of knowledge is generated.

[0011] The learning resource sequence is determined based on the personalized learning path, and the learning resource sequence is pushed to the user terminal for learning execution. After the user completes the learning, the learning effect is obtained, and the cognitive state vector is updated based on the learning effect to obtain the updated cognitive state vector.

[0012] When the difference between the updated cognitive state vector and the cognitive state vector exceeds the preset cognitive state change threshold, the semantic parsing and reasoning of the large language model are re-executed based on the updated cognitive state vector and the learning objective to obtain a new set of knowledge points.

[0013] The new set of knowledge points is aligned with the preset educational knowledge graph to obtain an updated alignment result. An updated personalized learning path is generated based on the updated alignment result. A new learning resource sequence is determined according to the updated personalized learning path, and the learning resource sequence is pushed to the user terminal.

[0014] By adopting the above technical solution, we can achieve deep semantic personalized learning recommendations based on learners' cognitive states, fundamentally solving the problem that existing personalized learning systems rely on shallow rule matching and cannot dynamically reflect learners' true cognitive states. This solution acquires users' learning behavior information and constructs cognitive state vectors using a pre-trained cognitive state evaluation model. This enables the system to quantify learners' knowledge mastery, learning speed, and interest preferences, thereby achieving accurate cognitive modeling. By feeding the learning objective and the cognitive state vector as joint inputs into a large language model, combined with semantic reasoning capabilities, key knowledge points under the learning objective are identified, and learning gap information is calculated. This ensures that the recommendation results are no longer limited to surface similarity matching but generate more targeted learning content based on the learner's actual cognitive level. Furthermore, through a knowledge graph alignment mechanism and a pre-knowledge point insertion strategy, the learning path conforms to the logical order and dependencies of knowledge, avoiding fragmented or hierarchical learning content. During the learning execution and feedback stages, the system can dynamically update the cognitive state vector based on learning effect information and trigger adaptive adjustments to the learning plan through a threshold judgment mechanism, achieving continuous optimization and closed-loop iteration of the learning process. As a result, the entire learning process has shifted from static recommendation to dynamic evolution, significantly improving the rationality of personalized learning paths and the accuracy of recommended content, thus promoting learners' efficient progress and enhancing cognitive matching.

[0015] In a preferred embodiment, this application can be further configured as follows: the calculation of the cognitive state vector based on the learned behavior information using a pre-trained cognitive state evaluation model includes:

[0016] The learning behavior information is preprocessed, the learning events are organized in chronological order and abnormal information is removed to obtain structured learning behavior information;

[0017] Extract learning behavior feature parameters from structured learning behavior information to characterize users' learning habits and behavioral intensity;

[0018] The learning behavior feature parameters are input into a pre-trained cognitive state assessment model to calculate assessment information corresponding to the degree of knowledge mastery, learning speed, and interest preferences.

[0019] The evaluation information is normalized and combined according to preset weights to generate a cognitive state vector.

[0020] By adopting the above technical solutions, a quantitative representation of learners' cognitive states can be achieved based on learning behavior information, improving the accuracy and reliability of cognitive state assessment. By chronologically organizing and removing anomalies from the learning behavior information, the authenticity and consistency of the input data can be guaranteed. By extracting learning behavior feature parameters, the model can comprehensively reflect learners' learning habits and behavioral intensity. Through a pre-trained cognitive state assessment model, assessment information such as knowledge mastery, learning speed, and interest preferences can be calculated, characterizing learners' cognitive features in multiple dimensions. Finally, by normalizing and weighting, a cognitive state vector is generated, allowing learners' cognitive states to be expressed in a unified and computable manner, thereby achieving accurate modeling and dynamic reflection of individual learning characteristics.

[0021] In a preferred example, this application can be further configured as follows: obtaining the learning objective input by the user, feeding the learning objective and the cognitive state vector as joint input into a large language model, and generating a set of knowledge points corresponding to the learning objective and learning gap information, including:

[0022] Semantic analysis is performed on the learning objectives to extract thematic and competency information from them, thus obtaining the semantic information of the objectives.

[0023] The target semantic information and the cognitive state vector are fused to construct joint input information;

[0024] The joint input information is fed into the large language model, which performs semantic reasoning based on the pre-trained semantic understanding mechanism to obtain a set of knowledge points related to the learning objective.

[0025] By comparing the set of knowledge points with the cognitive state vector, the knowledge points that the user has not mastered or has not mastered sufficiently in the current cognitive state are identified, thus obtaining learning gap information.

[0026] By adopting the above technical solution, learning objective parsing based on the fusion of semantic understanding and cognitive state can be achieved, thereby accurately identifying learners' knowledge mastery gaps for specific objectives. By semantically parsing the learning objectives and extracting topical and competency requirements, the system can understand the core semantic structure of the learning task. By fusing the semantic information of the objectives with the cognitive state vector, the model considers both the learning content and the learner's cognitive level in semantic reasoning. Through the semantic reasoning capabilities of the large language model, it identifies the set of knowledge points corresponding to the learning objectives at the knowledge level, automatically establishing knowledge relationships. Then, by comparing the set of knowledge points with the cognitive state vector, it can accurately locate knowledge points that are not mastered or are insufficiently mastered, thus achieving quantitative identification of learners' learning gaps.

[0027] In a preferred embodiment, this application can be further configured as follows: the joint input information is input into a large language model, and the large language model performs semantic reasoning based on a pre-trained semantic understanding mechanism to obtain a set of knowledge points related to the learning objective, including:

[0028] Based on the joint input information, the semantic element information in the learning objective and the mastery feature information in the cognitive state vector are analyzed to establish semantic association;

[0029] Based on semantic associations, a large language model performs semantic reasoning operations to identify knowledge topic information that matches the semantics of the learning target.

[0030] Based on the knowledge topic information, retrieve the corresponding knowledge point information from the preset semantic knowledge base to form a knowledge point set;

[0031] Perform a relevance filter on the set of knowledge points to obtain a set of knowledge points that correspond to the learning objectives.

[0032] By adopting the above technical solutions, a deep semantic association analysis between the semantic content of the learning objective and the learner's cognitive characteristics can be achieved, thereby improving the accuracy and logical consistency of knowledge point identification. By analyzing the semantic elements and mastery features in the joint input information, the system can establish semantic associations, making the semantic structure of the learning objective and the learner's cognitive level mappable. Through reasoning operations based on a large language model of semantic associations, knowledge topic information that matches the semantics of the learning objective can be identified in a large-scale semantic space, ensuring contextual consistency in the knowledge reasoning results. Furthermore, by retrieving knowledge point information corresponding to the knowledge topic information from a pre-set semantic knowledge base, the knowledge point extraction process becomes structured and traceable. Finally, by performing relevance filtering on the knowledge point set, content that deviates from the semantics of the learning objective is removed, making the output knowledge point set more aligned with the goal orientation and achieving high semantic consistency between the learning content and the learning objective.

[0033] In a preferred embodiment, this application can be further configured as follows: aligning the set of knowledge points with the prerequisite and dependency relationships in a preset educational knowledge graph to obtain the alignment result includes:

[0034] Based on the knowledge point set, retrieve the target node information corresponding to the knowledge point set in the preset educational knowledge graph, and establish a mapping relationship between the knowledge point set and the target node information;

[0035] Based on the mapping relationship, we obtain the prerequisite relationship information and dependency relationship information corresponding to the knowledge point set;

[0036] Based on the prior and dependency information, the knowledge point set is logically sorted to obtain a structured knowledge point set containing the order information of the knowledge points;

[0037] Perform a consistency check on the structured knowledge point set to obtain the alignment result.

[0038] By adopting the above technical solutions, a one-to-one mapping can be established between the knowledge point set derived from semantic reasoning and the educational knowledge graph. Based on the mapping, prerequisite and dependency information can be accurately extracted, and an ordered, structured knowledge point set can be formed by combining topological sorting. Conflicts and redundancies are eliminated through integrity, closed-loop, and sequential consistency checks, so that the alignment results are more semantically accurate, structurally clearer, and more reliable in terms of traceability. This reduces manual intervention based on empirical rules, lowers the mismatch rate and maintenance costs caused by redundant construction, improves the consistency and scalability of knowledge organization, and facilitates the unified management and large-scale application of cross-disciplinary and multi-level content.

[0039] In a preferred example, this application can be further configured as follows: the step of inserting prerequisite knowledge points into the alignment result based on learning gap information, determining the order and connection constraints of knowledge points, and generating a personalized learning path that conforms to the logical order of knowledge includes:

[0040] Based on learning gap information, the location of the set of knowledge points with learning gaps in the alignment results is identified, and the corresponding insertion region is determined.

[0041] In the position corresponding to the insertion area, the pre-defined prerequisite knowledge points in the educational knowledge graph are called and inserted into the alignment result to obtain the alignment result after insertion;

[0042] Based on the alignment results after insertion, extract the logical dependencies between each knowledge point and establish knowledge point order constraints.

[0043] Based on the knowledge point sequence constraints, determine the connection constraint information between adjacent knowledge points;

[0044] Based on the constraints of knowledge point sequence and connection, a personalized learning path that conforms to the logical order of knowledge is generated.

[0045] By adopting the above technical solutions, learning gaps can be accurately located based on the alignment results and learning gap information. Prerequisite knowledge points can be added with the help of educational knowledge graphs while maintaining the correctness of logical dependencies. An orderly and coherent personalized learning path can be constructed through dual constraints of knowledge point sequence constraints and connection constraints. This reduces hierarchical jumps and content fragmentation from the source, reduces the maintenance costs caused by manual rule configuration and repeated reordering, and makes the path structure more stable, consistent and reusable, adaptable to the large-scale generation and management of different knowledge granularities and subject scenarios.

[0046] In a preferred embodiment, this application can be further configured as follows: the step of determining a learning resource sequence based on a personalized learning path, pushing the learning resource sequence to the user terminal for learning execution, obtaining the learning effect after the user completes learning, and updating the cognitive state vector based on the learning effect to obtain the updated cognitive state vector includes:

[0047] Based on personalized learning paths, retrieve learning resource information corresponding to each knowledge point and establish a correspondence between knowledge points and learning resource information;

[0048] Based on the correspondence, determine the sequence of learning resources according to the order of knowledge points in the personalized learning path;

[0049] The learning resource sequence is pushed to the user terminal, and the corresponding learning execution method is configured according to the type of learning resource;

[0050] After the user completes the learning of the learning resource sequence, the learning effect information is obtained. The cognitive state vector is updated based on the learning effect information to obtain the updated cognitive state vector.

[0051] By adopting the above technical solutions, a complete closed-loop process can be achieved, from personalized learning paths to learning resource delivery and cognitive state updates, ensuring that the allocation of learning content dynamically aligns with the learner's actual cognitive level. By retrieving learning resource information based on personalized learning paths and establishing a correspondence between knowledge points and learning resource information, the system ensures that each knowledge point is matched with the most suitable learning resource. By generating a learning resource sequence according to the order of knowledge points, the logical continuity and hierarchical rationality of the learning content presentation are guaranteed. By configuring learning execution methods according to the type of learning resource, adapting to various learning formats such as videos, texts, exercises, and experiments, the learning process is made more aligned with the characteristics of the content. After the user completes learning, the cognitive state vector is updated through the collection and calculation of learning effect information. The system can reflect changes in the learner's cognition and maintain the timeliness and accuracy of the learning model, thus forming an adaptive cognitive feedback mechanism that enhances the accuracy and intelligence of learning content recommendation and execution.

[0052] The second objective of this invention is achieved through the following technical solution:

[0053] A personalized learning generation device based on a large model, comprising:

[0054] The learning behavior information acquisition module is used to acquire users' learning behavior information;

[0055] The cognitive state assessment module is used to calculate a cognitive state vector based on the learning behavior information using a pre-trained cognitive state assessment model.

[0056] The learning objective semantic parsing module is used to obtain the learning objective input by the user, and feed the learning objective and the cognitive state vector as joint input into the large language model to generate a set of knowledge points and learning gap information corresponding to the learning objective.

[0057] The educational knowledge graph alignment module is used to align the set of knowledge points with the prerequisite and dependency relationships in the preset educational knowledge graph to obtain the alignment result;

[0058] The learning path construction module is used to insert prerequisite knowledge points into the alignment results based on the learning gap information, determine the order of knowledge points and connection constraints, and generate a personalized learning path that conforms to the logical order of knowledge.

[0059] The learning resource push and feedback module is used to determine the learning resource sequence according to the personalized learning path, push the learning resource sequence to the user terminal for learning execution, obtain the learning effect after the user completes the learning, update the cognitive state vector based on the learning effect, and obtain the updated cognitive state vector.

[0060] The cognitive change detection module is used to re-execute the semantic parsing and reasoning of the large language model based on the updated cognitive state vector and the learning target when the difference between the updated cognitive state vector and the cognitive state vector exceeds the preset cognitive state change threshold, so as to obtain a new set of knowledge points.

[0061] The learning plan adaptive update module is used to align the new set of knowledge points with the preset educational knowledge graph to obtain an updated alignment result, generate an updated personalized learning path based on the updated alignment result, determine a new learning resource sequence according to the updated personalized learning path, and push the learning resource sequence to the user terminal.

[0062] By adopting the above technical solution, we can achieve deep semantic personalized learning recommendations based on learners' cognitive states, fundamentally solving the problem that existing personalized learning systems rely on shallow rule matching and cannot dynamically reflect learners' true cognitive states. This solution acquires users' learning behavior information and constructs cognitive state vectors using a pre-trained cognitive state evaluation model. This enables the system to quantify learners' knowledge mastery, learning speed, and interest preferences, thereby achieving accurate cognitive modeling. By feeding the learning objective and the cognitive state vector as joint inputs into a large language model, combined with semantic reasoning capabilities, key knowledge points under the learning objective are identified, and learning gap information is calculated. This ensures that the recommendation results are no longer limited to surface similarity matching but generate more targeted learning content based on the learner's actual cognitive level. Furthermore, through a knowledge graph alignment mechanism and a pre-knowledge point insertion strategy, the learning path conforms to the logical order and dependencies of knowledge, avoiding fragmented or hierarchical learning content. During the learning execution and feedback stages, the system can dynamically update the cognitive state vector based on learning effect information and trigger adaptive adjustments to the learning plan through a threshold judgment mechanism, achieving continuous optimization and closed-loop iteration of the learning process. As a result, the entire learning process has shifted from static recommendation to dynamic evolution, significantly improving the rationality of personalized learning paths and the accuracy of recommended content, thus promoting learners' efficient progress and enhancing cognitive matching.

[0063] The above-mentioned objective three of this application is achieved through the following technical solution:

[0064] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described personalized learning generation method based on a large model.

[0065] The fourth objective of this application is achieved through the following technical solution:

[0066] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described personalized learning generation method based on a large model.

[0067] In summary, this application includes at least one of the following beneficial technical effects:

[0068] 1. This solution enables deep semantic personalized learning recommendations based on learners' cognitive states, fundamentally solving the problem that existing personalized learning systems rely on shallow rule matching and cannot dynamically reflect learners' true cognitive states. By acquiring users' learning behavior information and constructing cognitive state vectors using a pre-trained cognitive state evaluation model, the system can quantify learners' knowledge mastery, learning speed, and interest preferences, thus achieving accurate cognitive modeling. By feeding the learning objective and cognitive state vector as joint inputs into a large language model, combined with semantic reasoning capabilities, key knowledge points under the learning objective are identified, and learning gap information is calculated. This ensures that the recommendation results are no longer limited to surface similarity matching but generate more targeted learning content based on the learner's actual cognitive level. Furthermore, through a knowledge graph alignment mechanism and a pre-knowledge point insertion strategy, the learning path conforms to the logical order and dependencies of knowledge, avoiding fragmented or disjointed learning content. During the learning execution and feedback stages, the system can dynamically update the cognitive state vector based on learning effect information and trigger adaptive adjustments to the learning plan through a threshold judgment mechanism, achieving continuous optimization and closed-loop iteration of the learning process. As a result, the entire learning process has shifted from static recommendation to dynamic evolution, significantly improving the rationality of personalized learning paths and the accuracy of recommended content, thus promoting learners' efficient progress and enhancing cognitive matching. Attached Figure Description

[0069] Figure 1 This is a flowchart of a personalized learning generation method based on a large model in one embodiment of this application;

[0070] Figure 2 This is a block diagram of a personalized learning generation device based on a large model in one embodiment of this application. Detailed Implementation

[0071] The present application will be further described in detail below with reference to the accompanying drawings.

[0072] In one embodiment, such as Figure 1 As shown, this application discloses a personalized learning generation method based on a large model, which specifically includes the following steps:

[0073] S10: Obtain information about the user's learning behavior.

[0074] In this embodiment, learning behavior information refers to the recordable behavioral data generated by users during their use of the learning platform. The behavioral data includes, but is not limited to, page access operations, question answering processes, material reading behaviors, audio and video learning operations, note-taking behaviors, and interactive commenting behaviors related to the learning process. This information can reflect the frequency of user operations, depth of participation, and behavioral habits in learning activities.

[0075] Specifically, the process of acquiring learning behavior information is achieved by setting behavior recording interfaces in the front-end interactive interface and back-end log system of the learning platform. The behavior recording interface refers to a program interface module used to receive and store learning operation event information, which can automatically generate behavior event entries when a user triggers a specific operation. The front-end interactive interface is used to collect click events, input operations, browsing trajectories, page dwell time, and playback control behaviors in the learning scenario. The back-end log system generates behavior recording information based on the event entries uploaded from the front end. Behavior recording information is a data structure composed of event type, event time, operation object, and operation parameters. Each record is linked by a unique user identifier field and a timestamp field, forming a behavior sequence. A behavior sequence refers to... A collection of user learning operation records arranged in chronological order is used to describe the temporal distribution characteristics of learning behaviors. During the data storage phase, the collected behavior records are organized into raw learning behavior record information according to time order. By parsing event codes, the behavior types of different learning stages are determined. Event codes are operation identifiers set to distinguish different learning scenarios such as course learning, question bank practice, material reading, and video viewing. By parsing event codes, the source of the behavior can be identified. Then, the behavior parameters are converted into structured records according to a preset format. Structured records refer to standardized data forms that describe learning behaviors using a field-based format, including behavior type field, operation time field, operation object field, and operation parameter field. Finally, all structured records are summarized to generate learning behavior information.

[0076] S20: Based on learning behavior information, the cognitive state vector is calculated using a pre-trained cognitive state assessment model.

[0077] In this embodiment, the cognitive state assessment model refers to a learner state estimation model built based on a deep learning framework. The model's input is learning behavior information, and its output is a multi-dimensional vector representation reflecting the learner's knowledge mastery, learning focus, and interest preferences. The cognitive state vector is a numerical vector generated by the assessment model to characterize the learner's current cognitive characteristics. Each dimension component corresponds to different cognitive dimension parameters, including knowledge comprehension parameters, operational proficiency parameters, learning focus parameters, and interest tendency parameters. During the training phase, the cognitive state assessment model establishes a mapping relationship between a large amount of sample learning behavior information and manual assessment labels. The pre-trained model parameters are fixed and used for subsequent cognitive state calculations.

[0078] Specifically, the process of calculating the cognitive state vector based on learning behavior information includes: first, arranging the input learning behavior information in chronological order and extracting learning behavior feature parameters. These parameters are a set of numerical indicators that reflect the changing patterns of learning behavior, including information on answer accuracy, learning duration, video viewing ratio, page switching frequency, and interaction response latency. The learning behavior feature parameters are then input into a pre-trained cognitive state assessment model. The activation values ​​of each layer of neural units are calculated through forward propagation of the neural network. The activation values ​​are the output signals after weighted summation based on the input and transformation by a nonlinear function in the model's neural units. After multi-layer propagation, a comprehensive representation vector is formed at the output layer. This comprehensive representation vector is then mapped to a fixed-dimensional space through normalization to obtain the cognitive state vector. Normalization refers to the process of eliminating numerical scale differences by linearly scaling or standard deviation standardizing the vector component values. The final output cognitive state vector is the calculation result of the learning behavior information in the cognitive state assessment model.

[0079] S30: Obtain the learning objective input by the user, and feed the learning objective and cognitive state vector as joint input information into the large language model to generate a set of knowledge points corresponding to the learning objective and learning gap information.

[0080] In this embodiment, the learning objective refers to the learning intention information actively input by the user in the learning platform. The learning objective can be obtained through text input boxes, voice recognition input, or preset option selection. The content includes specific course topics, knowledge domains, exam goals, or skill requirements. The joint input information refers to the input data structure formed by combining the learning objective information and the cognitive state vector in a fixed format. The learning objective information provides a semantic description of the learning intention, and the cognitive state vector provides a representation of the learner's current cognitive characteristics. The large language model refers to a semantic understanding and generation model based on large-scale corpus pre-training, which has the ability of semantic parsing, contextual reasoning, and knowledge association. It can generate structured knowledge reasoning results in the semantic space of the input information. The learning gap information refers to the difference information obtained by comparing the knowledge required for the learning objective with the knowledge already mastered by the learner in the current cognitive state. It is used to describe the gap range of the learner's knowledge mastery.

[0081] Specifically, the process of acquiring learning objectives is completed through the learning platform's input interface. The input interface receives learning objective information submitted by users in text or voice form. The voice input content is converted into text format by a speech recognition algorithm and then uniformly encoded and stored. Subsequently, the learning objective information and the cognitive state vector are concatenated according to a preset input format to form joint input information. The concatenation format follows the principle of dimensional alignment. The semantic part of the text is processed by word segmentation and vectorization, while the cognitive state vector is directly concatenated as a numerical vector to form a complete joint input matrix. The joint input matrix is ​​input into the large language model. The large language model parses the semantic element information of the learning objective through the semantic understanding layer and performs semantic reasoning in the internal knowledge embedding space. The reasoning process calculates semantic relevance weights through an attention mechanism. The weights reflect the degree of semantic matching between the learning objective and the cognitive state vector. Based on the semantic reasoning results, knowledge point information related to the learning objective is extracted from the model's knowledge base to form a knowledge point set. The knowledge point set is then compared with the corresponding knowledge mastery dimension information in the cognitive state vector. By calculating the difference, the knowledge content that the learner has not mastered or has insufficient mastery is identified, forming learning gap information.

[0082] S40: Align the set of knowledge points with the prerequisite and dependency relationships in the preset educational knowledge graph to obtain the alignment result.

[0083] In this embodiment, the educational knowledge graph refers to a directed graph structure constructed with knowledge points as nodes and logical connections between knowledge points as edges. The nodes in the graph contain fields such as knowledge name, subject, knowledge level, and content index. The direction of the edges is used to indicate the order of knowledge learning. Prerequisite relationship information refers to the set of information in the educational knowledge graph used to describe the corresponding relationship of the prerequisite knowledge points that need to be mastered before learning a knowledge point. Dependency relationship information refers to the logical dependence and content connection relationship between knowledge points, used to explain the dependency order between knowledge points. Alignment result refers to the structured matching result formed by matching the set of knowledge points with the nodes and edges of the educational knowledge graph, used to represent the logical position and association structure of the set of knowledge points in the educational knowledge graph.

[0084] Specifically, the process of aligning the knowledge point set with the educational knowledge graph includes: first, retrieving graph node information in the educational knowledge graph that matches the text content of each knowledge point in the knowledge point set; determining the optimal matching node through a semantic similarity calculation method, which calculates the cosine similarity between the knowledge point name vector and the graph node name vector based on a word vector embedding model; and identifying the node with the highest similarity as the matching node. After matching, establishing a mapping relationship between the knowledge point set and the graph nodes, which is a one-to-one correspondence table recording knowledge point identifiers and corresponding graph node identifiers. After establishing the mapping relationship, extracting the corresponding prerequisite and dependency information based on the directed edge information between graph nodes. The prerequisite information is used to identify the order of prior learning, and the dependency information is used to describe the logical connection path. Using the extracted prerequisite and dependency information as constraints, performing a logical sorting on the knowledge point set, which is arranged in order from basic knowledge to advanced knowledge, and eliminating duplicate or conflicting node mappings, finally obtaining a structurally complete and sequentially consistent alignment result.

[0085] S50: Based on learning gap information, insert prerequisite knowledge points into the alignment results, determine the order of knowledge points and connection constraints, and generate a personalized learning path that conforms to the logical order of knowledge.

[0086] In this embodiment, prerequisite knowledge points refer to the knowledge content that needs to be mastered before the learning target knowledge points, as determined by the educational knowledge graph. Knowledge point order refers to the arrangement order of knowledge points in the learning path. Connection constraints refer to the logical connection rules between adjacent knowledge points set to ensure the logical coherence of learning. Personalized learning path refers to the sequence of learning content that conforms to the logical order of knowledge, determined based on learning gap information and alignment results, and is used to describe the learning sequence and connection order between learning content.

[0087] Specifically, the process of generating personalized learning paths based on learning gap information includes first identifying the locations of learning gaps in the alignment results. Learning gap locations refer to knowledge nodes that are not included in the knowledge point set but need to be supplemented based on prerequisite or dependency relationships. The gaps are located by traversing the node sequence in the alignment results, and the corresponding insertion index is recorded. After determining the insertion index, the prerequisite knowledge point information related to the gap is retrieved from the educational knowledge graph, and the prerequisite knowledge points are inserted into the alignment results according to their index positions. The insertion operation is implemented by maintaining a knowledge node list, which is an ordered set containing knowledge point identifiers, content indexes, and hierarchical information. After the insertion operation is completed, the knowledge point sequence information is extracted based on the updated node list. The topological order of knowledge points in a directed graph is determined using a depth-first traversal algorithm. This topological order represents the arrangement of learning paths from basic to advanced knowledge. After obtaining the topological order, knowledge point order constraint information is established. This information is an ordered set of constraints that records the sequential relationships between knowledge points. Based on this information, the connection constraint information between adjacent knowledge points is calculated. This connection constraint information is obtained by calculating content relevance and knowledge level differences. Content relevance is determined using a text embedding similarity model, and knowledge level differences are determined by calculating node depth differences. Finally, a personalized learning path is generated based on the knowledge point order constraint information and the connection constraint information. This personalized learning path is represented as an ordered sequence of knowledge points, achieving a logically continuous arrangement of learning content.

[0088] S60: Determine the learning resource sequence based on the personalized learning path, and push the learning resource sequence to the user terminal for learning execution. After the user completes the learning, obtain the learning effect, update the cognitive state vector based on the learning effect, and obtain the updated cognitive state vector.

[0089] In this embodiment, the learning resource sequence refers to the set of learning resources arranged in order corresponding to the knowledge points in the personalized learning path. The learning resource information includes content formats such as e-textbooks, explanatory videos, question banks, interactive experiments, and mock tests. Each learning resource has a knowledge point identifier field and a resource type field to correspond one-to-one with the knowledge point order. The learning effect refers to the learning behavior feedback information generated after the user completes the learning resource sequence. The learning effect information includes learning completion rate information, question accuracy information, video viewing ratio information, and interaction participation frequency information, which are used to reflect the learner's behavioral characteristics after executing the learning resources.

[0090] Specifically, the process of determining the learning resource sequence based on the personalized learning path includes: first, retrieving the learning resource database according to the order of knowledge points in the personalized learning path. The learning resource database is a collection of learning resources stored according to knowledge point identifier indexes. The corresponding learning resource information is obtained by matching the knowledge point identifiers with the resource index table. The obtained learning resource information is then arranged in the order of knowledge points to form an initial learning resource sequence. After the initial learning resource sequence is formed, the learning resources are filtered by type and adjusted in duration. Type filtering is based on user learning preference parameters; for example, video resources are prioritized for users who prefer video learning, and document resources are prioritized for users who prefer reading learning. Duration adjustment is achieved through a resource segment extraction algorithm. This algorithm selects key learning segments based on the resource timestamp range without changing the integrity of the knowledge, thus controlling the learning duration. The learning process involves several steps. After the learning resource sequence is determined, it is sent to the user's terminal via the learning platform's push interface. The push interface determines the resource presentation method based on the terminal type, which includes mobile, web, and desktop clients. After the user completes the learning resource sequence, learning effectiveness information is collected. This information is obtained through the behavior recording module, which monitors learning completion markers, answer submission records, and interactive click events, generating learning effectiveness data records. In the data processing stage, the learning effectiveness data records are cleaned and normalized. The cleaned data is then input into a pre-trained cognitive state assessment model, and the cognitive state vector is updated through forward propagation. The updated cognitive state vector is a numerical vector that reflects the learner's current knowledge mastery status, calculated using new learning behavior feedback information. Finally, the updated cognitive state vector is output.

[0091] S70: When the difference between the updated cognitive state vector and the cognitive state vector exceeds the preset cognitive state change threshold, based on the updated cognitive state vector and the learning objective, the semantic parsing and reasoning of the large language model are re-executed to obtain a new set of knowledge points.

[0092] In this embodiment, the cognitive state change threshold refers to a numerical boundary parameter used to determine whether a significant change has occurred in the cognitive state. It measures the degree of difference between the updated cognitive state vector and the original cognitive state vector. The difference value refers to the numerical result calculated by the vector distance metric method. The vector distance metric method can adopt standard mathematical calculation methods such as Euclidean distance, cosine similarity inverse calculation, or Manhattan distance. The new knowledge point set refers to the set of knowledge points regenerated by the large language model after the cognitive state changes, which is used to reflect the learner's new knowledge needs.

[0093] Specifically, the process of semantic parsing and inference using the new large language model includes first calculating the difference between the updated cognitive state vector and the previous cognitive state vector. This difference is calculated by performing a difference operation on the corresponding components of the two vectors, squaring the differences of each component, summing them, and then taking the square root to obtain the Euclidean distance. Alternatively, the semantic difference can be obtained by inversely calculating the cosine of the angle between the vectors. The calculated difference is then compared with a preset cognitive state change threshold. When the difference exceeds the threshold, the semantic parsing and inference process is initiated, combining the updated cognitive state vector with the learning objective to form new joint input information. The joint input information is constructed according to the aforementioned input format. A new input matrix is ​​formed by concatenating text embeddings and numerical vectors. The new input matrix is ​​then input into the large language model. The large language model re-parses the semantic element information of the learning target based on the semantic understanding mechanism and performs semantic reasoning operations in the model's internal knowledge space. The semantic reasoning operations include attention weight allocation calculation and context semantic matching search. A new semantic reasoning result is obtained through weighted aggregation. Based on the semantic reasoning result, knowledge node information related to the learning target is retrieved from the model's knowledge base. The knowledge node information is filtered and aggregated according to semantic relevance, ultimately forming a new set of knowledge points.

[0094] S80: Align the new set of knowledge points with the preset educational knowledge graph to obtain an updated alignment result, generate an updated personalized learning path based on the updated alignment result, determine a new learning resource sequence according to the updated personalized learning path, and push the learning resource sequence to the user terminal.

[0095] In this embodiment, the updated alignment result refers to the knowledge point mapping relationship and logical order structure re-established during the matching process between the new knowledge point set and the educational knowledge graph; the updated personalized learning path refers to the learning path sequence regenerated based on the updated alignment result and the new learning gap information; and the new learning resource sequence refers to the learning resource sequence set corresponding to each knowledge point in the updated personalized learning path.

[0096] Specifically, the alignment process includes first extracting knowledge point identifiers from the new set of knowledge points; retrieving corresponding node content from the knowledge graph database based on the node index of the educational knowledge graph; storing prerequisite and dependency relationships between knowledge units in the form of node identifiers and relation tables; establishing new knowledge point mapping relationships through node identifier matching; after the mapping relationships are determined, extracting the dependency order information between each knowledge point from the relation table in the educational knowledge graph; performing logical sorting and consistency verification on the newly generated knowledge point mapping relationships; logical sorting is completed using a topological sorting algorithm, and consistency verification is achieved by checking whether there is a closed-loop structure in the prerequisite dependencies; obtaining updated alignment results; and after the updated alignment results are generated, identifying knowledge gap positions and adjusting the knowledge points based on the new learning gap information. The learning path is updated by arranging knowledge points in a specific order and inserting prerequisite knowledge points to complete the learning chain. During the generation of the personalized learning path, a path building algorithm is invoked. This algorithm generates an ordered index based on dependency constraints and outputs a path structure table. After the updated personalized learning path is determined, corresponding learning resource information is retrieved from the learning resource database according to the order of knowledge points in the path structure table. This information is then arranged sequentially to form a new learning resource sequence. Finally, the new learning resource sequence is pushed to the user's terminal through the learning platform push module. The push module determines the push format based on the terminal type (mobile, web, and desktop), and the push format includes resource links, task lists, or course schedules, thus completing the push of the updated learning resource sequence.

[0097] In one embodiment, step S20, namely, calculating the cognitive state vector based on the learned behavior information using a pre-trained cognitive state evaluation model, includes:

[0098] S201: Preprocess the learning behavior information, organize the learning events in chronological order and remove abnormal information to obtain structured learning behavior information.

[0099] In this embodiment, structured learning behavior information refers to a data set that can be identified and calculated by the model after being formatted, time-synchronized, and outlier-handled. A learning event refers to a single learning operation record in the learning behavior information that has a time identifier and behavior type attribute.

[0100] Specifically, the preprocessing of learning behavior information includes first sorting the information according to the timestamp field, unifying the timestamp format to a standard time format to ensure consistency in the time dimension, then grouping the learning events according to user identifiers to form a learning behavior sequence arranged in chronological order. After the learning behavior sequence is established, the learning events undergo integrity checks and anomaly identification. Integrity checks are performed by verifying that each learning event record contains the behavior type field, timestamp field, and event parameter field simultaneously. If any fields are missing or the format is inconsistent, corrections are made by filling in default values ​​or deleting abnormal records. Anomaly identification is achieved through statistical analysis and rule matching. Statistical analysis methods include those based on behavior duration. Z-Score outlier detection is used to discover abnormal records that exceed the normal learning duration range. The rule matching method performs comparisons based on a predefined behavior pattern library to identify logically conflicting learning events, such as learning start time being later than end time or mutually exclusive behavior events occurring within the same time period. After removing outlier information, the learning event parameters are formatted and standardized. The formatting operation includes converting text-type parameters into a unified encoding format, and the standardization operation is achieved through numerical scaling or interval normalization to ensure consistency of dimensions for different parameters. Finally, the preprocessed learning behavior sequence is converted into a tabular structure with fields including user identifier, behavior type, timestamp, behavior parameters, and behavior duration, forming structured learning behavior information.

[0101] S202: Extract learning behavior feature parameters from structured learning behavior information to characterize users' learning habits and behavioral intensity.

[0102] In this embodiment, learning behavior characteristic parameter information refers to a set of parameters extracted from structured learning behavior information and obtained through statistics or calculation, which can reflect the user's learning pattern, level of learning engagement and behavioral stability.

[0103] Specifically, the process of extracting learning behavior feature parameters from structured learning behavior information includes first classifying and aggregating the learning behavior information according to the behavior type field, grouping similar learning events into the same behavior category, such as question-answering events, video-watching events, document reading events, and interactive commenting events. After the behavior classification is completed, key behavior parameter fields are extracted for each type of learning event. For question-answering events, the number of questions answered, the accuracy rate of answers, and the time taken to answer questions are extracted; for video-watching events, the number of plays, the proportion of viewing time, and the number of pauses are extracted; for document reading events, the number of pages read, the reading dwell time, and the scrolling depth are extracted; and for interactive commenting events, the number of comments and the reply frequency are extracted. After the behavior parameters are extracted, the behavior frequency index is calculated based on the timestamp field. The rate metric is obtained by counting the number of behavioral events occurring within a unit of time, and is used to characterize the user's learning pace. Simultaneously, a behavior persistence metric is calculated, which is achieved by calculating the variance of the distribution intervals of learning events in the time series, and is used to describe the stability of user learning. Subsequently, all behavioral parameters are normalized to eliminate dimensional differences. The normalization process uses a min-max scaling algorithm to linearly map each parameter value to the range of 0 to 1. After parameter normalization, the feature parameters of different behavior types are summarized and combined to generate learning behavior feature parameter information containing multi-dimensional parameter fields, including answer feature fields, viewing feature fields, reading feature fields, and interaction feature fields, forming learning behavior feature parameter information that can characterize user learning habits and behavioral intensity.

[0104] S203: Input the learning behavior feature parameter information into the pre-trained cognitive state assessment model to calculate the assessment information corresponding to the degree of knowledge mastery, learning speed and interest preference.

[0105] Specifically, the process of inputting learning behavior feature parameter information into a pre-trained cognitive state assessment model includes: first, organizing the learning behavior feature parameter information according to the model input requirements; arranging the parameter fields in dimensional order to form an input vector; and performing feature dimension verification to ensure that the input dimension is consistent with the model structure. After the input vector is constructed, it is fed into the input layer of the cognitive state assessment model. The input layer receives all parameter fields and performs weighted normalization processing. Feature weights are calculated using a weight matrix to reduce the correlation between features. Subsequently, the model enters the hidden layer, which consists of multiple fully connected neural networks. Each layer performs linear transformations and non-linear activation operations. The linear transformations are achieved through matrix multiplication. The method is implemented using ReLU or Sigmoid functions for nonlinear activation, which is used to capture the nonlinear relationship between learning behavior feature parameters and cognitive performance. After the hidden layer calculation is completed, the model output layer generates corresponding output results through a Softmax classifier and a linear regression node. The Softmax classifier is used to calculate the distribution of interest preference categories, and the linear regression node is used to calculate the continuous scores of knowledge mastery and learning speed. The output results are numerically smoothed and range-limited by the post-processing module. The numerical smoothing is achieved through a moving average algorithm, and the range limitation is achieved through standardized interval mapping. Finally, the evaluation information corresponding to knowledge mastery, learning speed, and interest preference is obtained.

[0106] S204: Based on the evaluation information, normalize the data and combine it according to preset weights to generate a cognitive state vector.

[0107] Specifically, the process of normalizing and combining assessment information according to preset weights includes first extracting the raw numerical results of knowledge mastery, learning speed, and interest preference from the assessment information; organizing the different assessment results into an initial assessment matrix according to their corresponding feature dimensions, where each column of the initial assessment matrix corresponds to a different cognitive feature dimension and each row corresponds to learning records from different time periods; after the initial assessment matrix is ​​established, normalization is performed on the values ​​in the matrix using a standard deviation normalization algorithm. This involves calculating the mean and standard deviation of each feature dimension, subtracting the mean from the raw values, and then dividing by the standard deviation to obtain a standardized result, thus eliminating differences between different feature units. After the normalization process is completed... Then, each feature dimension is weighted and combined according to preset weight parameters. The preset weight parameters are a fixed set of parameters determined by feature importance analysis during the model training phase. The weighting and combination is achieved through matrix weighting operations, that is, multiplying the normalized feature matrix with the weight vector to obtain the weighted result matrix. Then, a dimension aggregation operation is performed on the weighted result matrix. The dimension aggregation operation obtains the comprehensive cognitive value representing the current learning stage by weighted averaging of feature values ​​at different time periods. Finally, the aggregated comprehensive cognitive values ​​are arranged in the order of feature dimensions to form a cognitive state vector. The cognitive state vector contains numerical elements such as knowledge mastery dimension value, learning speed dimension value, and interest preference dimension value, thus completing the generation of the cognitive state vector.

[0108] In one embodiment, in step S30, the learning objective input by the user is obtained, and the learning objective and the cognitive state vector are fed as joint input into the large language model to generate a set of knowledge points corresponding to the learning objective and learning gap information, including:

[0109] S301: Perform semantic analysis on the learning objectives to extract thematic and competency information from the learning objectives, thereby obtaining the semantic information of the objectives.

[0110] Specifically, the semantic parsing process for the learning objectives includes first performing encoding conversion and sentence segmentation on the learning objective text. Encoding conversion is implemented through a standardized UTF-8 interface, and sentence segmentation uses sentence-end identifier rules to divide the text into independent sentence units for subsequent semantic segmentation. After sentence segmentation, the learning objective text undergoes lexical annotation processing, which is achieved by loading a pre-defined educational domain dictionary. This dictionary contains key subject terms and their corresponding semantic tags to ensure consistency between the parsing results and the terminology of the educational knowledge graph. Next, topic identification is performed based on the TF-IDF keyword extraction algorithm. By statistically analyzing the weight distribution of terms in the learning objective text within the global dictionary, the terms with the highest weight values ​​are selected as candidate topic terms. Finally, word vector similarity is calculated using a topic clustering algorithm, and semantics are merged. Similar candidate keywords generate the final topic information. After the topic information is determined, ability requirements are identified using a semantic template matching method. The semantic template contains five levels of ability: "mastery, understanding, application, analysis, and evaluation" and their corresponding verb sets. By matching verb phrases in the learning target text, ability requirement words are identified and labeled with their corresponding levels. Then, a semantic correspondence between topic information and ability requirement information is established. This semantic correspondence is achieved using a context window sliding algorithm. The length of the sliding window is dynamically adjusted based on the sentence length to determine the dependency direction between topic words and ability verbs. Finally, target semantic information is generated according to the mapping result between topic information and ability requirement information. The target semantic information is output in the structure of "topic field - ability level field - dependency direction field", thus completing the semantic parsing of the learning target.

[0111] S302: Information fusion is performed between target semantic information and cognitive state vector to construct joint input information.

[0112] Specifically, the process of information fusion between target semantic information and cognitive state vectors includes first aligning the data formats of the target semantic information and cognitive state vectors. The target semantic information is represented in the form of a text semantic feature matrix, and the cognitive state vector is represented in the form of a numerical feature vector. Data format alignment is achieved by unifying the feature dimension length and numerical range. The feature dimension length is expanded by a vector padding algorithm to increase insufficient dimensions, and the numerical range is adjusted to the 0-1 range by a min-max normalization algorithm. After format alignment, text vectorization processing is performed on the target semantic information. Text vectorization processing is implemented using a pre-trained word embedding model, converting the topic field, capability level field, and dependency direction field in the target semantic information into corresponding vector representations to generate semantic feature vectors. Subsequently, Semantic feature vectors and cognitive state vectors are concatenated and fused using a vector concatenation operator. During the concatenation process, the boundaries of each feature partition are kept non-overlapping and arranged in a fixed order, ensuring the independence of semantic and cognitive features in the input space. After concatenation, a feature weighting operation is performed on the concatenated vectors. This operation is implemented through an attention weight allocation mechanism, which adjusts the weight ratio of semantic and cognitive features based on the importance parameter of the feature dimensions. The weighted fusion result is calculated using matrix multiplication. Finally, the weighted fusion result is output as joint input information, which exists as a high-dimensional dense vector containing two subspaces: semantic feature components and cognitive feature components. This vector is used for semantic parsing and inference input in subsequent large language models.

[0113] S303: Input the joint input information into the large language model, and the large language model performs semantic reasoning based on the pre-trained semantic understanding mechanism to obtain a set of knowledge points related to the learning objective.

[0114] Specifically, the process of inputting joint input information into the large language model includes first loading the joint input information into the model input interface. The joint input information is represented in the form of a high-dimensional dense vector. The input interface performs dimensionality checking and masking operations on the input vector. Dimensionality checking is used to ensure that the input vector is consistent with the structure of the model input layer. Masking is achieved by padding with zeros at missing feature positions to ensure the integrity of the input data. After the input stage is completed, the large language model performs multi-layer encoding operations on the joint input information through a pre-trained semantic understanding mechanism. The semantic understanding mechanism is implemented using a multi-head attention network structure based on the Transformer architecture. The multi-head attention network structure includes a self-attention layer, a feedforward neural layer, and a layer normalization unit. The self-attention layer determines the dependency strength between semantic features by calculating the dot product correlation between the query vector, key vector, and value vector. The feedforward neural layer achieves non-linear semantic mapping through linear mapping and activation function processing. The layer normalization unit is used for... To maintain the stability of feature distribution, after multi-layer encoding, the model generates a semantic representation matrix. Each vector unit in the semantic representation matrix corresponds to a semantic concept relationship. Semantic decoding is performed by the matrix decoding unit. The semantic decoding process calculates the attention weight distribution based on the context window and extracts the concept node vectors most relevant to the learning target semantics from the semantic representation matrix. Then, the knowledge retrieval submodule is called to match the concept node vectors with the internal knowledge index. The knowledge index is based on educational knowledge units and their semantic indexes. The matching operation is achieved by calculating the cosine similarity between the concept node vectors and the knowledge index vectors. Knowledge index vectors with similarity values ​​higher than a preset relevance threshold are filtered and output as candidate results. Finally, deduplication and clustering operations are performed on the candidate results. Deduplication is achieved by vector hash comparison, and clustering is achieved by using the K-Means algorithm to divide knowledge units into categories based on semantic distance. The set of center vectors after clustering is the set of knowledge points related to the learning target.

[0115] S304: Compare the set of knowledge points with the cognitive state vector to identify the knowledge points that the user has not mastered or has not mastered sufficiently in the current cognitive state, and obtain learning gap information.

[0116] Specifically, the process of comparing the knowledge point set with the cognitive state vector includes first establishing a knowledge mastery index mapping relationship for each knowledge point in the knowledge point set. This mapping relationship describes the correspondence between knowledge points and various dimensions of the cognitive state vector. Specifically, the knowledge mastery level dimension reflects the depth of understanding of the knowledge point in the cognitive state, the learning speed dimension reflects the learning efficiency of the knowledge point, and the interest preference dimension reflects the matching degree between the knowledge point and learning interests. After establishing the mapping relationship, the knowledge point set is vectorized. This vectorization is implemented through a knowledge encoding module, which uses a word embedding model to convert the semantic content of each knowledge point into a fixed-length knowledge point vector representation. Subsequently, the knowledge point vector and the cognitive state vector are input into a similarity calculation unit. The similarity calculation unit calculates the cosine similarity between the two vectors. The similarity score or Euclidean distance determines the matching degree between knowledge points and cognitive states. The matching degree is used to measure the learner's mastery of the knowledge point. After the similarity calculation is completed, the matching degree scores of all knowledge points are compared with the preset mastery threshold. The mastery threshold is an empirical parameter obtained from the model training phase and is used to distinguish between mastery and non-mastery states. When the matching degree score is lower than the mastery threshold, the corresponding knowledge point is marked as a non-mastery knowledge point. When the matching degree score is between the mastery threshold and the upper limit of the mastery threshold, the corresponding knowledge point is marked as a insufficiently mastered knowledge point. Finally, all non-mastered knowledge points and insufficiently mastered knowledge points are summarized according to knowledge point identification and mastery level to form learning gap information. The learning gap information is represented in tabular form, including knowledge point identification field, mastery state field, and matching degree score field, thus completing the generation of learning gap information.

[0117] In one embodiment, in step S302, the joint input information is input into a large language model, which performs semantic reasoning based on a pre-trained semantic understanding mechanism to obtain a set of knowledge points related to the learning objective, including:

[0118] S3021: Based on the joint input information, analyze the semantic element information in the learning objective and the mastery feature information in the cognitive state vector to establish semantic association.

[0119] Specifically, the process of parsing semantic element information and mastering feature information based on joint input information includes firstly, partitioning and reading the semantic feature components and cognitive feature components in the joint input information. The semantic feature components correspond to the topic field, ability level field, and dependency direction field in the target semantic information, while the cognitive feature components correspond to the knowledge mastery dimension, learning speed dimension, and interest preference dimension in the cognitive state vector. After partitioning and reading, element extraction is performed on the semantic feature components. The element extraction operation is implemented through a semantic component analysis algorithm. The semantic component analysis algorithm identifies semantic core words and modification relationships based on the dependency syntax tree structure, thereby determining the boundary range and hierarchical structure of the semantic element information. Subsequently, feature parsing is performed on the cognitive feature components. The feature parsing operation is implemented through a feature decomposition function, mapping the numerical components in the cognitive feature components to knowledge. The mastery feature space is used to extract feature variables that reflect the level of cognition. After element extraction and feature parsing, semantic alignment processing is performed on semantic element information and mastery feature information. Semantic alignment processing is achieved through a bidirectional attention matching mechanism. The bidirectional attention matching mechanism determines the semantic correspondence between the two types of features in the multidimensional space by calculating the weighted correlation between the semantic element vector and the mastery feature vector. After semantic alignment, a semantic association matrix is ​​generated through the relevance matrix calculation module. The row vectors of the semantic association matrix correspond to semantic element information, the column vectors correspond to mastery feature information, and the element values ​​in the matrix represent the semantic relevance values ​​between the two. Finally, the semantic association matrix is ​​normalized and output as semantic associations. The semantic associations are stored in matrix form and used for feature matching and knowledge point generation in the subsequent semantic reasoning stage.

[0120] S3022: Based on semantic association, the large language model performs semantic reasoning operations to identify knowledge topic information that matches the semantics of the learning target.

[0121] Specifically, the process of performing semantic reasoning operations based on semantic associations includes first inputting the semantic associations into the semantic reasoning unit of the large language model. The semantic reasoning unit uses a multi-head attention network structure as its core computational module to establish a context-dependent mapping between semantic element information and mastered feature information. During the input phase, the large language model performs embedding expansion on the semantic associations. This embedding expansion is achieved through a vector mapping layer, converting the semantic association matrix into a high-dimensional feature representation for semantic computation. Subsequently, feature interaction operations are performed in the multi-head attention network structure. These operations are achieved by calculating the weighted correlation between the query vector, key vector, and value vector. Each attention head independently learns the correspondence between semantic element information and mastered feature information. Weighted convergence generates a semantic fusion representation. After the semantic fusion representation is generated... The semantic fusion representation is input to the feedforward neural layer. The feedforward neural layer extracts latent semantic relationships through linear transformation and activation function operations. The linear transformation is implemented through matrix multiplication, and the ReLU function is used to enhance the nonlinear feature capture capability. After multi-layer semantic inference is completed, the model outputs a semantic relevance vector. Each component of the semantic relevance vector represents a latent knowledge topic semantic weight. Subsequently, the saliency value of each component in the semantic relevance vector is calculated by the topic determination module. The saliency value is calculated through the Softmax function and is used to determine the importance ranking of each semantic candidate topic. Finally, semantic candidate topics with saliency values ​​higher than the preset topic determination threshold are selected, and their corresponding text labels are extracted as knowledge topic information. The knowledge topic information is output in the form of text labels for semantic mapping in the subsequent knowledge point retrieval stage.

[0122] S3023: Based on the knowledge topic information, retrieve the corresponding knowledge point information from the preset semantic knowledge base to form a knowledge point set.

[0123] Specifically, the process of retrieving knowledge point information from a pre-defined semantic knowledge base based on knowledge topic information includes: firstly, performing semantic encoding on the knowledge topic information. This semantic encoding is implemented using a word embedding model, converting the text identifiers in the knowledge topic information into fixed-length semantic vector representations. The word embedding model is pre-trained based on educational domain corpora, ensuring that the semantic vectors remain consistent with the conceptual semantic space of the educational knowledge domain. After semantic encoding, the encoded knowledge topic vectors are input into the knowledge retrieval module. The knowledge retrieval module then connects to the pre-defined semantic knowledge base for matching queries. The pre-defined semantic knowledge base is a database constructed based on knowledge point entries. Each knowledge point entry in the database includes a knowledge point name, a semantic description of the knowledge point, a subject identifier, and related conceptual fields. All entries in the semantic knowledge base are pre-vectorized using the same word embedding model, and their corresponding semantic vectors are stored. In the semantic indexing process, during matching queries, the knowledge retrieval module calculates the cosine similarity between the knowledge topic vector and each knowledge point vector in the knowledge base. Similarity is calculated by dividing the vector dot product by the modulus product. The similarity value measures semantic similarity. Then, all knowledge point entries are sorted in descending order based on their similarity values. Entries with similarity values ​​higher than a preset similarity threshold are selected as candidate results. After candidate result filtering, duplicate detection and semantic clustering are performed. Duplicate detection is achieved through hash index comparison to remove semantically equivalent entries. Semantic clustering uses the K-Means algorithm to group candidate results based on semantic distance, obtaining representative knowledge points for each semantic category. Finally, the representative knowledge points from each group are summarized to form a knowledge point set, stored in a list structure. Each element corresponds to a knowledge point name and its semantic description field, thus completing the construction of the knowledge point set.

[0124] S3024: Perform relevance filtering on the knowledge point set to obtain the knowledge point set corresponding to the learning objective.

[0125] Specifically, the process of performing relevance screening on the knowledge point set includes: first, constructing a semantic matching template based on the target semantic information of the learning objective. The semantic matching template consists of a topic field, a capability level field, and a dependency direction field, used to limit the semantic matching range between knowledge points and the learning objective. After the semantic matching template is constructed, semantic feature extraction is performed on each knowledge point in the knowledge point set. Semantic feature extraction is achieved through a text feature encoding model, converting the semantic description of the knowledge point into a feature vector representation. Each dimension of the feature vector corresponds to a feature dimension in the semantic knowledge space. Subsequently, the relevance between the semantic matching template vector and the knowledge point feature vector is calculated using a cosine similarity function. The calculation formula is the vector dot product divided by the product of the magnitudes of the two vectors. The resulting relevance value is used to measure the semantic fit between the knowledge point and the learning objective. After the relevance calculation is completed, all knowledge points are... The relevance value of each knowledge point is compared with a preset relevance threshold. When the relevance value is higher than the threshold, the corresponding knowledge point is marked as a valid match. When the relevance value is lower than the threshold, it is marked as an irrelevant knowledge point. After marking, all valid matching knowledge points are sorted in descending order based on their relevance value, and then a secondary sort is performed based on the ability level field. This ensures that knowledge points with high ability requirements are placed at the top of the same topic. After the sorting results are generated, an aggregation and filtering operation is performed. The aggregation and filtering merges semantically similar knowledge points into a group by setting a topic aggregation radius to eliminate knowledge duplication and semantic redundancy. Finally, the filtered and aggregated knowledge point items are reorganized to generate a knowledge point set corresponding to the learning objective. The knowledge point set is output in a structured table format, with fields including knowledge point name, semantic similarity value, and ability level identifier, thus completing the determination of the knowledge point set corresponding to the learning objective.

[0126] In one embodiment, step S40 involves aligning the set of knowledge points with the prerequisite and dependency relationships in a preset educational knowledge graph to obtain an alignment result, including:

[0127] S401: Based on the knowledge point set, retrieve the target node information corresponding to the knowledge point set in the preset educational knowledge graph, and establish a mapping relationship between the knowledge point set and the target node information.

[0128] Specifically, the process of retrieving educational knowledge graphs based on knowledge point sets includes first reading the name field of each knowledge point in the knowledge point set, inputting the knowledge point name as a key identifier into the educational knowledge graph index module, and then using an inverted index structure to achieve efficient querying. The inverted index structure uses the knowledge point name as the key and the knowledge graph node number as the index item. After input, a node matching operation is performed. This operation calculates the edit distance between the knowledge point name and the knowledge graph node label. Nodes with an edit distance less than a preset matching threshold are considered matched nodes. The matching threshold is set based on knowledge naming conventions. After node matching is complete, the corresponding node attribute information is extracted, including the node name field. The system includes fields for node ID, prerequisite node list, and dependent node list, used to establish mapping relationships later. A one-to-one correspondence is then established between knowledge point names and their corresponding node IDs to generate an initial mapping table. This initial mapping table is represented as key-value pairs, where the key is the knowledge point name and the value is the corresponding node ID. After the initial mapping table is established, consistency verification is performed by comparing the semantic similarity between node names and knowledge point names. Mapping entries with similarity below the verification threshold are marked as pending confirmation. Pending confirmation entries are manually or rule-based verified to complete the final mapping table. Each record in the final mapping table contains a knowledge point name, node ID, and a node attribute summary, ultimately establishing a mapping relationship between the knowledge point set and the target node information.

[0129] S402: Based on the mapping relationship, obtain the prerequisite relationship information and dependency relationship information corresponding to the knowledge point set.

[0130] Specifically, the process of obtaining prerequisite and dependency information based on the mapping relationship includes first reading the node number field in the mapping relationship, locating the corresponding node record in the educational knowledge graph database using the node number, and the node record containing three core fields: node number, prerequisite node list, and dependency node list. After node location, the prerequisite relationship extraction operation is performed, which is achieved by parsing the prerequisite node list field in the node record. The prerequisite node list field records the set of prerequisite knowledge node numbers required by the current node. The system sequentially reads each prerequisite node number and establishes directed connections between nodes to represent prerequisite dependencies in the learning order. After the prerequisite relationship extraction is completed, the dependency relationship extraction operation is performed. The extraction operation is achieved by parsing the dependency node list field in the node record. The dependency node list field records the set of subsequent knowledge node numbers that depend on the current node. By traversing the dependency node numbers, the corresponding dependency connection relationship is established to describe the backward association between knowledge. After extracting the two types of relationship information, the extraction results are merged and deduplicated. The merging process is achieved by matching node numbers, keeping only one record of duplicate node pairs. After deduplication, a complete node relationship table is formed. The node relationship table includes two types of data structures: the first type is the prerequisite relationship record, and the second type is the dependency relationship record. Each record contains the source node number, the target node number, and the relationship type identifier field. Finally, the node relationship table is output as prerequisite relationship information and dependency relationship information.

[0131] S403: Based on the prior relationship information and dependency relationship information, the knowledge point set is logically sorted to obtain a structured knowledge point set containing the knowledge point order information.

[0132] Specifically, the process of logically sorting the knowledge point set based on prerequisite and dependency information includes: first, constructing a directed graph structure of knowledge points based on the prerequisite and dependency information. The directed graph uses knowledge point numbers as nodes and prerequisite and dependency relationships as directed edges, with the edges pointing from the prerequisite node to the subsequent node. After the directed graph is constructed, an in-degree counting operation is performed. This operation iterates through all nodes and records the number of times each node is pointed to by other nodes. Nodes with an in-degree value of zero represent initial knowledge points without prerequisite knowledge requirements. After the in-degree counting is completed, a topological sorting algorithm is used to sort the directed graph. The topological sorting algorithm selects nodes with an in-degree of zero for output and removes the corresponding outgoing edges, then updates the in-degree values ​​of the remaining nodes, repeating this process until all nodes have been output. The sorting result is the order of knowledge points that conforms to the knowledge dependency logic. During the sorting process, to avoid sorting anomalies caused by circular dependencies, a loop detection mechanism is set up. The loop detection mechanism is implemented through a depth-first traversal algorithm. When a path that can return to itself from a certain node is detected, the path is recorded and marked as a logical conflict node group. Logical conflict node groups are handled through manual or rule-based correction. After correction, the topological sorting is re-executed to ensure the orderliness of the sorting result. After sorting, the output knowledge point order result is associated with the original knowledge point set to form a structured knowledge point set. The structured knowledge point set is represented in tabular form, with fields including knowledge point number, knowledge point name, predecessor node number, successor node number, and sorting sequence number. Finally, a structured knowledge point set containing knowledge point order information is obtained.

[0133] S404: Perform a consistency check on the structured knowledge point set to obtain the alignment result.

[0134] Specifically, the process of performing consistency verification on the structured knowledge point set includes first reading and matching the sorting sequence number, predecessor node number, and successor node number fields in the structured knowledge point set to generate a node dependency table. This table describes the dependency direction and hierarchical position between knowledge points. After the node dependency table is generated, a topology integrity check is performed. This check verifies whether each knowledge point has a valid predecessor or successor node record by traversing all node relationships. For isolated nodes or broken-chain nodes, an anomaly flag is recorded in the verification log. After the topology integrity check is completed, a dependency loop closure check is performed. This check is implemented using a depth-first traversal algorithm. The algorithm maintains an access stack while traversing the path. When duplicate nodes are detected in the access stack, a dependency loop is identified, and the loop path information is output to the exception record table for subsequent correction. Then, sequential consistency verification is performed, which is achieved by comparing the sorting sequence number with the dependency direction. When the sequence number of the predecessor node of a certain knowledge point is greater than the sequence number of the subsequent node, it is identified as an order error and the error pair is recorded. After all verification items are completed, the results of topology integrity check, dependency loop detection, and sequential consistency verification are summarized and analyzed. If no exception records are found in all checks, a verification record with the consistency flag "passed" is generated; otherwise, a verification report containing the exception node number and error type fields is generated. Finally, the verification results are bound to the structured knowledge point set and output to form an alignment result.

[0135] In one embodiment, step S50, which involves inserting prerequisite knowledge points into the alignment result based on learning gap information, determining the order and connection constraints of knowledge points, and generating a personalized learning path that conforms to the logical order of knowledge, includes:

[0136] S501: Based on learning gap information, identify the location of the set of knowledge points with learning gaps in the alignment results and determine the corresponding insertion area.

[0137] Specifically, the process of identifying learning gap locations based on learning gap information includes first reading the knowledge point identifier field and mastery status field from the learning gap information. The mastery status field records the mastery status identifier for each knowledge point, which includes two types: "not mastered" and "insufficient mastery." After the data is read, the learning gap information is compared with the structured knowledge point set in the alignment results. The comparison process achieves one-to-one matching through the knowledge point number field. Knowledge points that do not appear in the alignment results or whose mastery status is "not mastered" are identified as learning gaps. After the learning gaps are identified, the position of each learning gap in the alignment results is located. Position location is achieved by looking up the sorting sequence number of the knowledge point number in the structured knowledge point set. When a learning void defect has no sorting index, its neighboring node number is recorded as a reference node number. Then, the position index of the reference node number in the structured knowledge point set is calculated. Based on this position index, the insertion region range is determined. The insertion region range includes two parts: the predecessor node number and the successor node number, which are used to indicate the connection position between the preceding knowledge and the subsequent knowledge. After the insertion region is determined, the learning void defect number and the insertion region range information are combined to form an insertion mapping table. The insertion mapping table records each learning void defect and its corresponding insertion region position field. The fields include the learning void defect number, the predecessor node number, the successor node number, and the insertion priority identifier. Finally, the insertion mapping table is output as the input basis for subsequent preceding knowledge point insertion operations, thus completing the determination of the insertion region.

[0138] S502: In the position corresponding to the insertion area, call the prerequisite knowledge points in the preset educational knowledge graph, and insert the prerequisite knowledge points into the alignment result to obtain the alignment result after insertion.

[0139] Specifically, the process of calling prerequisite knowledge points and performing insertion operations in the insertion area includes first reading the learning gap / deficiency number, predecessor node number, and successor node number fields from the insertion mapping table; then retrieving the associated prerequisite knowledge point information from the educational knowledge graph database based on the learning gap / deficiency number. This prerequisite knowledge point information is obtained through a node association table, which uses the node number as an index and records a list of predecessor node numbers and a node name field. After the retrieval is complete, a set of prerequisite knowledge point numbers is obtained, and a duplicate detection operation is performed on each node in the set. Duplicate detection is achieved by comparing the node number with the existing node numbers in the alignment results. If the node already exists, the insertion operation is skipped; if the node does not exist, it is marked as a node to be inserted. Subsequently, the insertion position index is determined based on the predecessor node number and successor node number in the insertion mapping table. The insertion position index represents the prerequisite knowledge point. The insertion order of nodes in the structured knowledge point set is determined by incrementing the predecessor node's sorting index. After the insertion position is determined, the nodes to be inserted are inserted sequentially according to the dependency order in the educational knowledge graph, and the alignment results are stored in a table structure. The insertion operation is implemented through a table row insertion function, and the sorting index field and node connection relationship field are updated simultaneously to ensure the continuity of the logical structure. After all nodes to be inserted are inserted, a consistency adjustment is performed on the order of the inserted nodes. The consistency adjustment is implemented by recalculating the dependency direction and updating the node reference index to ensure that the dependency relationship between the newly inserted node and the original node remains correct. Finally, the alignment result after insertion is generated and output in the form of an updated structured knowledge point set, including the newly added predecessor knowledge point record and the corrected sorting index field, which is used for subsequent knowledge point order constraint extraction operations.

[0140] S503: Based on the alignment results after insertion, extract the logical dependencies between each knowledge point and establish knowledge point order constraints.

[0141] Specifically, the process of extracting logical dependencies and establishing knowledge point order constraints based on the alignment results after insertion includes first reading the node number field, predecessor node number field, and successor node number field from the alignment results after insertion, and importing these fields into the dependency analysis module. The dependency analysis module generates a knowledge point dependency graph based on the directed connections between nodes. The knowledge point dependency graph uses node numbers as vertices and directed edges from predecessor nodes to successor nodes as dependency paths. After the dependency graph is generated, a relation resolution operation is performed. The relation resolution operation is implemented through adjacency matrix calculation. Each element in the adjacency matrix records the dependency direction and level difference between two knowledge points. After the relation resolution is completed, redundancy detection is performed on the dependency paths. Redundancy detection is achieved by comparing the order of nodes that appear repeatedly in multiple paths. The algorithm implements a column-based approach, marking redundant edges in duplicate paths as invalid and deleting them to ensure the uniqueness and clarity of dependencies. Then, it performs a hierarchy extraction operation, traversing the dependency graph using a depth-first search algorithm. During traversal, the hierarchy depth value of each node is recorded, representing the relative position of knowledge points in the logical order. After hierarchy extraction, knowledge point order constraints are generated based on the node's hierarchy depth value and dependency direction. These constraints are represented as rules, each consisting of a preceding node number, a subsequent node number, and a sequence index field. Finally, the generated knowledge point order constraints are bound to the knowledge point set to form a sequence constraint table. This table is output in a structured format as input for subsequent steps in determining knowledge point connection constraints.

[0142] S504: Based on the knowledge point sequence constraints, determine the connection constraint information between adjacent knowledge points.

[0143] Specifically, the process of determining the connection constraint information based on the knowledge point sequence constraint includes first reading the preceding node number field, the subsequent node number field, and the sequence index field from the knowledge point sequence constraint table, and importing these fields into the connection relationship calculation module. The connection relationship calculation module analyzes the arrangement relationship of adjacent nodes in the knowledge point sequence constraint, identifies the logical adjacency between each pair of adjacent nodes, and determines that two knowledge point pairs are adjacent knowledge point pairs when the difference in their sequence indices is equal to one, recording the node combination information. After the adjacent knowledge point pairs are determined, the knowledge content connection strength is calculated for each pair of nodes. The knowledge content connection strength is achieved by comparing the semantic description vectors of the corresponding nodes in the educational knowledge graph. The semantic description vectors are pre-generated by a word embedding model, and the connection strength calculation uses a cosine similarity function. The calculation results are used for... Reflecting the semantic coherence of adjacent knowledge points, after the connection strength calculation is completed, the consistency of execution direction is checked for all adjacent knowledge point pairs. The direction consistency check is achieved by comparing the dependency direction field of the preceding node number and the subsequent node number. When the dependency direction is consistent with the order constraint direction, it is marked as a valid connection relationship; otherwise, it is marked as a reverse relationship. Subsequently, the valid connection relationships are filtered by a threshold. The filtering condition is that the connection strength value is higher than the preset connection strength threshold. The threshold is set by the model training phase based on the semantic coherence statistics. After the filtering is completed, all valid connection relationships are recorded as connection constraint information. The connection constraint information is output in a table structure, with fields including the preceding node number, the subsequent node number, the connection strength value, and the connection direction identifier, forming the connection constraint information between adjacent knowledge points.

[0144] S505: Based on the knowledge point sequence constraints and connection constraints, generate a personalized learning path that conforms to the logical order of knowledge.

[0145] Specifically, the process of generating a personalized learning path based on knowledge point sequence constraints and connection constraints includes first reading the preceding node number field, subsequent node number field, sequence index field, and connection strength value field from the knowledge point sequence constraint table and connection constraint information table, and importing these two types of data into the path generation module. The path generation module constructs a learning path graph using knowledge point sequence constraints as the main thread and connection constraints as the correction basis. The learning path graph uses knowledge point numbers as nodes, sequence constraints as basic edges, and connection strength as weighted edge weights. After the learning path graph is constructed, path weight calculation is performed. The path weight calculation is achieved by combining the sequence index difference and the connection strength value. The calculation formula is: Path weight = α × Connection strength value + β×sequence matching coefficient, where α and β are preset weight coefficients used to balance the importance of semantic coherence and sequential rationality. After the path weight is calculated, an optimal path search is performed on the learning path graph. The optimal path search is implemented through an improved Dijkstra algorithm. The algorithm selects the node access order according to the principle of minimum weight, ensuring that the arrangement of knowledge points in the path conforms to both the prior logic and semantic connection. Then, a redundancy elimination operation is performed on the searched path sequence. Redundancy elimination is achieved by detecting consecutive duplicate nodes or semantically equivalent nodes. The semantic equivalence determination is based on the similarity calculation results between the semantic description vectors of knowledge points. When the similarity is higher than a set threshold, nodes are merged. After redundancy elimination, the path sequence is hierarchically optimized. Hierarchical optimization is achieved by adjusting the knowledge span between different learning modules to ensure the hierarchical progression of the ability of adjacent knowledge points. Finally, the optimized path sequence is output as a personalized learning path. The personalized learning path is represented in a list structure, with fields including knowledge point number, learning order number, hierarchical identifier, and connection weight value, thus completing the generation of a personalized learning path that conforms to the logical order of knowledge.

[0146] In one embodiment, in step S60, a learning resource sequence is determined based on the personalized learning path, and the learning resource sequence is pushed to the user terminal for learning execution. After the user completes the learning, the learning effect is obtained, and the cognitive state vector is updated based on the learning effect to obtain the updated cognitive state vector, including:

[0147] S601: Based on personalized learning paths, retrieve learning resource information corresponding to each knowledge point and establish a correspondence between knowledge points and learning resource information.

[0148] Specifically, the process of retrieving learning resource information based on a personalized learning path includes first reading the knowledge point number field and learning order field from the personalized learning path, and then inputting the knowledge point number as a search keyword into the learning resource database. The learning resource database is a repository storing categorized and organized teaching content information, including various types of learning resource information such as video course information, text textbook information, exercise information, and interactive experiment information. Each piece of learning resource information includes a resource number, resource type, knowledge point tag, and difficulty level field. During the retrieval process, the resource tag field is matched according to the knowledge point number, and the difficulty level field is used to filter out resource information that matches the learning stage of the knowledge points in the personalized learning path. A hash index mechanism is used to achieve fast retrieval, where the hash function is based on the knowledge point number... The combination of the number and resource type is used as an input variable to generate an index key value. Among the selected learning resource information, the resource arrangement order is further determined based on the learning order field. Multiple learning resources under the same knowledge point are sorted according to the teaching type priority, which is as follows: video course information, text textbook information, exercise information, and interactive experiment information. After the resources are sorted, a correspondence between knowledge points and learning resource information is established. The correspondence is stored in the form of a mapping table, which includes a knowledge point number field, a resource number field, a resource type field, and a priority weight field. In this way, each knowledge point corresponds to several learning resource information with different presentation formats and difficulty levels, thus forming a complete knowledge point-learning resource information correspondence structure.

[0149] S602: Based on the correspondence, determine the sequence of learning resources according to the order of knowledge points in the personalized learning path.

[0150] Specifically, the process of determining the learning resource sequence according to the knowledge point order in the personalized learning path includes: first, reading the knowledge point number field and the learning order field from the personalized learning path; then, extracting the learning resource information matching each knowledge point number from the mapping table of correspondence between knowledge points and learning resource information; after extraction, sorting the learning resource information according to the learning order field, arranging the resource sets corresponding to each knowledge point in ascending order; and simultaneously, performing secondary sorting within each knowledge point according to the resource type field and priority weight field, ensuring that different types of learning resources are arranged in a progressive order of teaching logic within the same knowledge point; after completing the two-level sorting, concatenating the learning resource information of each knowledge point according to the knowledge point number to form a preliminary learning resource sequence. After the learning resource sequence is generated, a consistency check is performed on the sequence. The consistency check is achieved by verifying whether the knowledge point numbers in the learning resource sequence correspond one-to-one with the sequence fields in the personalized learning path. When there are omissions or duplicates, the index matching is re-executed according to the mapping table to ensure the correct order. After the consistency check passes, a sequence number field and a learning stage identifier field are further assigned to each resource in the learning resource sequence. The learning stage identifier field is determined according to the hierarchical position of the knowledge point in the path, such as the basic stage, the extension stage, or the consolidation stage. The final output learning resource sequence is stored in a list structure, and the fields include resource sequence number, knowledge point number, resource type, priority weight, and learning stage identifier, realizing the generation of learning resource sequences arranged in the order of the personalized learning path.

[0151] S603: Push the learning resource sequence to the user terminal and configure the corresponding learning execution method according to the learning resource type.

[0152] Specifically, the process of pushing learning resource sequences to user terminals and configuring learning execution methods includes first importing the resource sequence number, knowledge point number, resource type, and learning stage identifier fields from the learning resource sequence into the push control module. The push control module is a functional unit for managing the transmission and display configuration of learning resources, and it has task scheduling and resource distribution interfaces. Before pushing, the learning resource sequences are grouped according to the learning stage identifier field, grouping learning resources of the same stage into the same push batch, and generating a push task identifier for each batch. Then, the learning resource type is determined according to the resource type field. Learning resource types include four types: video course information, text textbook information, exercise information, and interactive experiment information. Each type corresponds to a different learning execution method. When the push task is executed, the video course... Information is pushed to the user terminal player module via a streaming media transmission protocol, and playback control parameters are configured to enable resume playback and speed adjustment. Text textbook information is transmitted to the user terminal document module via a file download protocol, and reading mode parameters are configured to support pagination and annotation functions. Exercise information is pushed to the user terminal testing module via an interactive interface, and answer logic parameters are configured to support instant feedback and automatic scoring. Interactive experiment information is pushed to the user terminal experiment module via a simulation interface, and simulation running parameters are configured to support multi-scenario experiment loading. After all learning resources are pushed, push log information is recorded, including push timestamp, task identifier, user terminal identifier, and resource status fields, for subsequent learning progress recording and feedback management, realizing the terminal push and learning execution mode configuration process of the learning resource sequence.

[0153] S604: After the user completes the learning of the learning resource sequence, the learning effect information is obtained, and the cognitive state vector is updated based on the learning effect information to obtain the updated cognitive state vector.

[0154] Specifically, the process of updating the cognitive state vector after a user completes learning a learning resource sequence includes first matching the resource sequence number field in the learning resource sequence with the interaction logs generated by the user's terminal to extract the user's learning records for each learning resource. These records include fields such as video course viewing time, text textbook reading progress, correct answer rate for practice questions, and completion rate of interactive experiments. These fields are then organized into learning behavior feedback information. After collecting the learning behavior feedback information, time normalization and validity screening are performed on each type of feedback information. Time normalization is achieved by mapping the learning behavior feedback information to a unified time axis according to the time interval of the learning stage identifier field. Validity screening is achieved by removing outlier records. Outlier determination is based on a dual condition of a learning duration threshold and an operation frequency threshold. After data preprocessing, the learning behavior feedback information is then analyzed. A learning effectiveness information vector is constructed, which is a multi-dimensional vector structure formed by combining four sub-vectors—viewing completion rate, reading completion rate, answer accuracy rate, and experiment completion rate—with preset weights. This vector is used to quantify the user's mastery of different learning resource types. The learning effectiveness information vector is then input into the cognitive state assessment model to perform vector update operations. The update operations include three steps: feature extraction, weight adjustment, and state recalculation. Feature extraction uses convolutional layers to extract core features from the learning effectiveness information. Weight adjustment adjusts the weights of each sub-vector in the overall cognitive state according to the distribution ratio of learning resource types. State recalculation generates a new cognitive state vector through the model output layer. After the cognitive state vector is updated, normalization is performed on the updated cognitive state vector to ensure that the numerical range of the vector components remains consistent with the model input, ultimately forming the updated cognitive state vector.

[0155] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0156] In one embodiment, a personalized learning generation apparatus based on a large model is provided, which corresponds one-to-one with the personalized learning generation method based on a large model in the above embodiments. For example... Figure 2 As shown, the personalized learning generation device based on a large model includes a learning behavior information acquisition module, a cognitive state assessment module, a learning goal semantic parsing module, an educational knowledge graph alignment module, a learning path construction module, a learning resource push and feedback module, a cognitive change detection module, and a learning plan adaptive update module.

[0157] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A large model-based personalized learning generation method, characterized in that, The personalized learning generation method based on large models includes: Obtain user learning behavior information; Based on learning behavior information, a cognitive state vector is calculated using a pre-trained cognitive state assessment model. The learning objectives input by the user are obtained, and the learning objectives and cognitive state vectors are fed into the large language model as joint input information to generate a set of knowledge points corresponding to the learning objectives and learning gap information. The process of obtaining the learning objective input by the user involves feeding the learning objective and the cognitive state vector as joint input information into a large language model to generate a set of knowledge points corresponding to the learning objective and learning gap information, including: Semantic analysis is performed on the learning objectives to extract thematic and competency information from them, thus obtaining the semantic information of the objectives. The target semantic information and the cognitive state vector are fused to construct joint input information; Based on the joint input information, the semantic element information in the learning objective and the mastery feature information in the cognitive state vector are analyzed to establish semantic association; Based on semantic associations, a large language model performs semantic reasoning operations to identify knowledge topic information that matches the semantics of the learning target. Based on the knowledge topic information, retrieve the corresponding knowledge point information from the preset semantic knowledge base to form a knowledge point set; Perform a relevance filter on the knowledge point set to obtain the knowledge point set that corresponds to the learning objective; By comparing the set of knowledge points with the cognitive state vector, the knowledge points that the user has not mastered or has not mastered sufficiently in the current cognitive state are identified, and learning gap information is obtained. Align the set of knowledge points with the prerequisite and dependency relationships in the pre-defined educational knowledge graph to obtain the alignment results; Based on learning gap information, prerequisite knowledge points are inserted into the alignment results, the order of knowledge points and connection constraints are determined, and a personalized learning path that conforms to the logical order of knowledge is generated. The learning resource sequence is determined based on the personalized learning path, and the learning resource sequence is pushed to the user terminal for learning execution. After the user completes the learning, the learning effect is obtained, and the cognitive state vector is updated based on the learning effect to obtain the updated cognitive state vector. When the difference between the updated cognitive state vector and the cognitive state vector exceeds the preset cognitive state change threshold, the semantic parsing and reasoning of the large language model are re-executed based on the updated cognitive state vector and the learning objective to obtain a new set of knowledge points. The new set of knowledge points is aligned with the preset educational knowledge graph to obtain an updated alignment result. An updated personalized learning path is generated based on the updated alignment result. A new learning resource sequence is determined according to the updated personalized learning path, and the learning resource sequence is pushed to the user terminal.

2. The large model-based personalized learning generation method according to claim 1, characterized in that, The cognitive state vector, calculated using a pre-trained cognitive state assessment model based on learning behavior information, includes: The learning behavior information is preprocessed, the learning events are organized in chronological order and abnormal information is removed to obtain structured learning behavior information; Extract learning behavior feature parameters from structured learning behavior information to characterize users' learning habits and behavioral intensity; The learning behavior feature parameters are input into a pre-trained cognitive state assessment model to calculate assessment information corresponding to the degree of knowledge mastery, learning speed, and interest preferences. The evaluation information is normalized and combined according to preset weights to generate a cognitive state vector. 3.The large model-based personalized learning generation method according to claim 1, characterized in that, The process of aligning the set of knowledge points with the prerequisite and dependency relationships in a pre-defined educational knowledge graph to obtain the alignment result includes: Based on the knowledge point set, retrieve the target node information corresponding to the knowledge point set in the preset educational knowledge graph, and establish a mapping relationship between the knowledge point set and the target node information; Based on the mapping relationship, we obtain the prerequisite relationship information and dependency relationship information corresponding to the knowledge point set; Based on the prior and dependency information, the knowledge point set is logically sorted to obtain a structured knowledge point set containing the order information of the knowledge points; Perform a consistency check on the structured knowledge point set to obtain the alignment result. 4.The large model-based personalized learning generation method of claim 1, wherein, The process of inserting prerequisite knowledge points into the alignment results based on learning gap information, determining the order and connection constraints of knowledge points, and generating a personalized learning path that conforms to the logical order of knowledge includes: Based on learning gap information, the location of the set of knowledge points with learning gaps in the alignment results is identified, and the corresponding insertion region is determined. In the position corresponding to the insertion area, the pre-defined prerequisite knowledge points in the educational knowledge graph are called and inserted into the alignment result to obtain the alignment result after insertion; Based on the alignment results after insertion, extract the logical dependencies between each knowledge point and establish knowledge point order constraints. Based on the knowledge point sequence constraints, determine the connection constraint information between adjacent knowledge points; Based on the constraints of knowledge point sequence and connection, a personalized learning path that conforms to the logical order of knowledge is generated.

5. The personalized learning generation method based on a large model according to claim 1, characterized in that, The process involves determining a learning resource sequence based on a personalized learning path, pushing the learning resource sequence to the user terminal for learning execution, obtaining the learning effect after the user completes learning, and updating the cognitive state vector based on the learning effect to obtain the updated cognitive state vector, including: Based on personalized learning paths, retrieve learning resource information corresponding to each knowledge point and establish a correspondence between knowledge points and learning resource information; Based on the correspondence, determine the sequence of learning resources according to the order of knowledge points in the personalized learning path; The learning resource sequence is pushed to the user terminal, and the corresponding learning execution method is configured according to the type of learning resource; After the user completes the learning of the learning resource sequence, the learning effect information is obtained. The cognitive state vector is updated based on the learning effect information to obtain the updated cognitive state vector.

6. A personalized learning generation device based on a large model, characterized in that, The personalized learning generation device based on a large model includes: The learning behavior information acquisition module is used to acquire users' learning behavior information; The cognitive state assessment module is used to calculate the cognitive state vector based on learning behavior information using a pre-trained cognitive state assessment model. The learning objective semantic parsing module is used to obtain the learning objective input by the user, and feed the learning objective and cognitive state vector as joint input information into the large language model to generate a set of knowledge points corresponding to the learning objective and learning gap information. The process of obtaining the learning objective input by the user involves feeding the learning objective and the cognitive state vector as joint input information into a large language model to generate a set of knowledge points corresponding to the learning objective and learning gap information, including: Semantic analysis is performed on the learning objectives to extract thematic and competency information from them, thus obtaining the semantic information of the objectives. The target semantic information and the cognitive state vector are fused to construct joint input information; Based on the joint input information, the semantic element information in the learning objective and the mastery feature information in the cognitive state vector are analyzed to establish semantic association; Based on semantic associations, a large language model performs semantic reasoning operations to identify knowledge topic information that matches the semantics of the learning target. Based on the knowledge topic information, retrieve the corresponding knowledge point information from the preset semantic knowledge base to form a knowledge point set; Perform a relevance filter on the knowledge point set to obtain the knowledge point set that corresponds to the learning objective; By comparing the set of knowledge points with the cognitive state vector, the system identifies the knowledge points that the user has not mastered or has insufficient mastery of in their current cognitive state, thus obtaining learning gap information. The educational knowledge graph alignment module is used to align a set of knowledge points with the prerequisite and dependency relationships in a preset educational knowledge graph to obtain the alignment result. The learning path construction module is used to insert prerequisite knowledge points into the alignment results based on learning gap information, determine the order of knowledge points and connection constraints, and generate a personalized learning path that conforms to the logical order of knowledge. The learning resource push and feedback module is used to determine the learning resource sequence based on the personalized learning path, push the learning resource sequence to the user terminal for learning execution, obtain the learning effect after the user completes the learning, and update the cognitive state vector based on the learning effect to obtain the updated cognitive state vector. The cognitive change detection module is used to re-execute the semantic parsing and reasoning of the large language model based on the updated cognitive state vector and the learning target when the difference between the updated cognitive state vector and the cognitive state vector exceeds the preset cognitive state change threshold, so as to obtain a new set of knowledge points. The adaptive update module for the learning plan is used to align the new set of knowledge points with the preset educational knowledge graph to obtain the updated alignment result, generate an updated personalized learning path based on the updated alignment result, determine a new learning resource sequence according to the updated personalized learning path, and push the learning resource sequence to the user terminal.

7. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the personalized learning generation method based on a large model as described in any one of claims 1 to 5.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the personalized learning generation method based on a large model as described in any one of claims 1 to 5.

Images