Teaching programming system and teaching recommendation method based on large language model assistance

Abstract

This invention relates to the field of intelligent education technology, and discloses a teaching programming system and teaching recommendation method based on a large language model. The system involves constructing a multi-dimensional ability graph and knowledge point topology network, utilizing a large language model for multi-round intent parsing and semantic reasoning to generate personalized learning path sequences. The system continuously tracks the learning trajectory, continuously optimizes the teaching strategy through dynamic calibration of strategy parameters, and proactively allocates specialized reinforcement resources based on trajectory prediction. This system achieves a breakthrough from static recommendation to dynamic adaptation. Through deep semantic understanding and a closed-loop optimization mechanism, it improves the accuracy of learning path planning and the timeliness of teaching intervention, enabling the programming teaching system to truly understand learning needs and adapt to changes in the learning state in real time, thereby improving the solidity of knowledge acquisition and learning efficiency.

Description

Technical Field

[0001] This invention relates to the field of intelligent education technology, specifically to a teaching programming system and teaching recommendation method based on a large language model. Background Technology

[0002] Through multi-turn intent parsing and semantic reasoning using a large language model, the generation process of teaching strategies can transcend keyword matching and rule constraints. By understanding natural language dialogue, the model can identify the learning difficulties and real needs implied in students' descriptions. Based on its built-in vast subject knowledge network, the model can perform semantic-level reasoning, constructing a dynamic learning path sequence that conforms to the inherent logical relationships of knowledge and individual cognitive patterns. This ensures that the recommended learning content not only matches ability tags but also aligns with students' thinking patterns and current optimal receptivity.

[0003] Based on the comparative analysis of real-time learning trajectories and preset standards, the weight parameters in the teaching strategy generation process are recalibrated, establishing a closed-loop optimization mechanism. The system can automatically adjust the weights of key parameters in ability dimension assessment, knowledge point priority, and resource recommendation strategies based on students' latest practice results, code complexity, and learning time. This data-driven continuous fine-tuning enables the system to respond sensitively to changes in students' knowledge status, overcoming the recommendation bias caused by student progress in static models, maintaining the timeliness and accuracy of teaching strategies, and achieving true adaptability in the teaching process.

[0004] By combining deep intent understanding with dynamic parameter optimization, the system can predict potential knowledge gaps from macro-level learning trajectories. Through analysis of long-term behavioral data, it identifies patterns deviating from expected paths and proactively allocates targeted reinforcement resources before students show obvious difficulties. This predictive, proactive intervention mechanism transforms the passive remedial model of traditional systems, pre-emptively supporting learning to more effectively consolidate knowledge and prevent knowledge gaps. Summary of the Invention

[0005] The purpose of this invention is to provide a teaching programming system and teaching recommendation method based on a large language model to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a teaching programming system based on a large language model, the system comprising:

[0007] Through the student feature analysis module, a learning element profile reflecting students' programming ability is constructed, and a student ability map containing multiple dimensions of ability indicators is output.

[0008] The learning resource management module receives student ability maps, collects real-time learning behavior data according to the preset knowledge system architecture, constructs a knowledge topology network that represents the relationship between knowledge points, and identifies the core learning resources and auxiliary learning resources of each node in the network.

[0009] The teaching recommendation engine module matches students' ability graphs with knowledge topology networks and generates personalized teaching strategies containing learning path sequences through multi-round intent parsing and semantic reasoning of a large language model.

[0010] The learning progress monitoring module continuously tracks students' implementation of personalized teaching strategies, collects actual learning trajectory data, and compares and analyzes it with preset learning standards.

[0011] The dynamic adjustment module recalibrates the weight parameters in the teaching strategy generation process based on the comparative analysis results, and outputs the optimized personalized teaching strategy.

[0012] The early warning and intervention module predicts weaknesses in knowledge acquisition based on overall learning trajectory data and allocates specific reinforcement resources for the detected weaknesses.

[0013] Preferably, the process of constructing a learning element profile reflecting students' programming abilities includes: collecting code submission records, error debugging logs, and concept test scores generated by students during the programming learning process; extracting features from these multi-source learning data to obtain code complexity features, error pattern features, and knowledge mastery features; organizing the above features according to time series; extracting learning trend features and fluctuation features through time series pattern analysis; and inputting the learning trend features and fluctuation features into a student ability assessment model trained with historical learning data to output a student ability map composed of multiple dimension ability indicators.

[0014] Preferably, the process of constructing a knowledge topology network representing the relationships between knowledge points includes: decomposing knowledge points into a basic concept layer, a syntax rule layer, and an algorithm application layer according to the hierarchical structure of the programming knowledge system; setting data collection points at each layer to continuously collect data on students' learning time, practice completion rate, and error frequency for each knowledge point; constructing an initial knowledge graph with knowledge points as nodes and the prerequisite / successor relationships between knowledge points as edges; and mapping the real-time collected learning behavior data to the corresponding nodes of the initial knowledge graph to form a knowledge topology network with learning behavior feature weights.

[0015] Preferably, the process of identifying the core learning resources and auxiliary learning resources of each node in the network includes: evaluating the learning resources associated with each node in the knowledge topology network from multiple dimensions, including resource difficulty coefficient, teaching suitability, and historical usage effect indicators; setting a resource quality threshold and classifying the learning resources of each node according to their quality; marking resources that meet the excellent quality standard as core learning resources and resources that do not meet the excellent quality standard but meet the basic teaching requirements as auxiliary learning resources; and storing the identification information of core learning resources and auxiliary learning resources for each node in the knowledge topology network.

[0016] Preferably, the process of generating a personalized teaching strategy containing a learning path sequence through multi-round intent parsing and semantic reasoning using a large language model includes: matching and analyzing the student's ability map with the knowledge topology network to identify the gap between the student's current ability and the requirements of the knowledge system; analyzing the gap through the large language model to generate a targeted learning objective sequence; searching for the optimal learning path in the knowledge topology network based on the learning objective sequence; ensuring that the generated learning path conforms to the student's learning habits through multi-round intent parsing, and verifying the rationality of the learning path through semantic reasoning; and finally outputting a personalized teaching strategy containing a specific learning activity sequence, resource allocation plan, and schedule.

[0017] Preferably, the process of continuously tracking students' execution of personalized teaching strategies includes: collecting learning trajectory data generated by students' execution of personalized teaching strategies in real time, including code writing behavior sequences, video viewing progress, and test completion status; comparing the actual learning trajectory data with the preset learning standards in the personalized teaching strategies, and calculating learning progress deviation and learning behavior deviation; triggering a strategy adjustment mechanism when the detected deviation exceeds a predetermined threshold; and classifying the learning stability level according to the magnitude of the deviation when the deviation is within the allowable range.

[0018] Preferably, the recalibration of weight parameters in the teaching strategy generation process based on the comparative analysis results includes: calculating the learning urgency coefficient based on learning progress deviation and learning behavior deviation; calculating the resource suitability coefficient based on students' actual performance at each knowledge point; establishing a dynamic weight adjustment model, taking the learning urgency coefficient and resource suitability coefficient as inputs, and outputting the adjustment amount of the weights of each objective function in the teaching strategy generation process; updating the parameters of the dynamic weight adjustment model according to a set period to ensure the timeliness of weight adjustment; and regenerating personalized teaching strategies using the adjusted weight parameters and updated learning data.

[0019] Preferably, the process of predicting weak points in knowledge mastery based on overall learning trajectory data includes: mapping the learning stability level of each knowledge point to the corresponding node of the knowledge topology network; constructing a dependency matrix between knowledge nodes and calculating the mutual influence coefficient between nodes; using a graph neural network to perform representation learning on the knowledge topology network with stability labels to explore potential risk propagation paths; outputting an overall risk prediction for the entire knowledge system and identifying a set of weak knowledge points that may generate chain risks.

[0020] Preferably, the process of configuring specific reinforcement resources for the detected weaknesses includes: setting an overall risk threshold; when the overall risk estimate exceeds the threshold, calculating the risk transmission coefficient of each knowledge point; calculating the comprehensive vulnerability index of each knowledge point based on the risk transmission coefficient and the local stability level; for knowledge points whose comprehensive vulnerability index exceeds a predetermined threshold, selecting suitable specific training materials from their associated auxiliary learning resources, combining them into reinforcement learning modules, and inserting them into the corresponding positions of the personalized teaching strategy.

[0021] Preferably, the present invention also includes a teaching recommendation method based on a large language model, the method comprising the following steps:

[0022] Construct a learning element profile that reflects students' programming abilities, and output a student ability map containing multiple dimensions of ability indicators;

[0023] Based on the pre-defined knowledge system architecture, real-time learning behavior data is collected to construct a knowledge topology network that represents the relationship between knowledge points, and the core learning resources and auxiliary learning resources of each node in the network are identified.

[0024] By matching student competency maps with knowledge topology networks, personalized teaching strategies containing learning path sequences are generated through multi-round intent parsing and semantic reasoning of a large language model.

[0025] Continuously track students' implementation of personalized teaching strategies, collect actual learning trajectory data, and compare and analyze them with preset learning standards;

[0026] Based on the comparative analysis results, the weight parameters in the teaching strategy generation process are recalibrated, and the optimized personalized teaching strategy is output.

[0027] Based on the overall learning trajectory data, we predict the weak points in knowledge acquisition and allocate special reinforcement resources for the detected weak points.

[0028] Compared with the prior art, the beneficial effects of the present invention are:

[0029] Through multi-turn intent parsing and semantic reasoning using a large language model, the generation process of teaching strategies can transcend keyword matching and rule constraints. By understanding natural language dialogue, the model can identify the learning difficulties and real needs implied in students' descriptions. Based on its built-in vast subject knowledge network, the model can perform semantic-level reasoning, constructing a dynamic learning path sequence that conforms to the inherent logical relationships of knowledge and individual cognitive patterns. This ensures that the recommended learning content not only matches ability tags but also aligns with students' thinking patterns and current optimal receptivity.

[0030] Based on the comparative analysis of real-time learning trajectories and preset standards, the weight parameters in the teaching strategy generation process are recalibrated, establishing a closed-loop optimization mechanism. The system can automatically adjust the weights of key parameters in ability dimension assessment, knowledge point priority, and resource recommendation strategies based on students' latest practice results, code complexity, and learning time. This data-driven continuous fine-tuning enables the system to respond sensitively to changes in students' knowledge status, overcoming the recommendation bias caused by student progress in static models, maintaining the timeliness and accuracy of teaching strategies, and achieving true adaptability in the teaching process.

[0031] By combining deep intent understanding with dynamic parameter optimization, the system can predict potential knowledge gaps from macro-level learning trajectories. Through analysis of long-term behavioral data, it identifies patterns deviating from expected paths and proactively allocates targeted reinforcement resources before students show obvious difficulties. This prediction-based proactive intervention mechanism transforms the passive remedial model of traditional systems, pre-emptively supporting learning to solidify knowledge acquisition and prevent knowledge gaps. Attached Figure Description

[0032] Figure 1 This is a schematic diagram illustrating the working principle of the teaching programming system based on a large language model as described in this invention.

[0033] Figure 2 A flowchart for constructing a profile of learning elements that reflects students' programming abilities;

[0034] Figure 3 A flowchart illustrating the process of constructing a knowledge topology network that represents the relationships between knowledge points;

[0035] Figure 4 A graph showing the dynamic change trend of weight factors during the weight parameter adjustment phase;

[0036] Figure 5 This is a trend chart showing the overall risk threshold dynamic adjustment effect during the threshold optimization phase. Detailed Implementation

[0037] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] Please see Figure 1 This invention provides a teaching programming system based on a large language model. The system includes: a student feature analysis module, a learning resource management module, a teaching recommendation engine module, a learning progress monitoring module, a dynamic adjustment module, and an early warning and intervention module. The student feature analysis module is responsible for constructing a learning element profile reflecting students' programming abilities and outputting a student ability map containing multiple dimensions of ability indicators. The learning resource management module receives the student ability map and collects real-time learning behavior data according to a preset knowledge system architecture to construct a knowledge topology network representing the relationship between knowledge points and identify the core learning resources and auxiliary learning resources of each node in the network. The teaching recommendation engine module matches the student ability map with the knowledge topology network and uses multi-round intent parsing and semantic reasoning of the large language model to generate personalized teaching strategies containing learning path sequences. The learning progress monitoring module continuously tracks the students' execution of the personalized teaching strategies and collects actual learning trajectory data for comparison and analysis with preset learning standards. The dynamic adjustment module recalibrates the weight parameters in the teaching strategy generation process based on the comparison and analysis results and outputs the optimized personalized teaching strategies. The early warning and intervention module predicts weak points in knowledge mastery based on the overall learning trajectory data and configures special reinforcement resources for the detected weak points.

[0039] Example 1: See Figure 2In specific implementation, code submission records include code version information and submission timestamps; error debugging logs record exception types and debugging operation sequences during program runtime; and concept test scores store the scores for multiple-choice and fill-in-the-blank questions. Feature extraction is performed on these multi-source learning data to obtain code complexity features, error pattern features, and knowledge mastery features. Code complexity features are calculated by analyzing the code abstract syntax tree to determine the control flow complexity index; error pattern features statistically analyze the distribution of error types and the number of repair steps; and knowledge mastery features calculate the accuracy rate based on the mapping relationship between test questions and knowledge points. These features are organized according to a time series, and learning trend and fluctuation features are extracted through time-series pattern analysis. The time-series pattern analysis uses a sliding window method to calculate the feature statistics within each window. The learning trend and fluctuation features are input into a student ability assessment model trained on historical learning data. The student ability assessment model uses a gradient boosting decision tree algorithm to output a student ability graph composed of multiple-dimensional ability indicators, including code efficiency score, logical rigor score, and concept comprehension level. In some embodiments, multi-source learning data is obtained from the database interface of an online programming platform, and the feature extraction process uses static code analysis tools.

[0040] In practical implementation, time series pattern analysis uses the root mean square error formula to quantify fluctuation characteristics, which is expressed as:

[0041]

[0042] in: Indicates the fluctuation characteristic value, This represents the total number of data points within the time window. This represents the feature value at the i-th time point. This represents the average value of the feature values ​​within the time window. Code complexity features are calculated using the Holstead volume metric to determine the total number of operators and operands. Error pattern features are analyzed by constructing an error transition matrix to understand error correction paths. Knowledge mastery features are calibrated using a project-response theory model to align question difficulty with student ability. The learning trend features from time-series pattern analysis are fitted with the slope of feature values ​​over time using linear regression. Before inputting the learning trend features and fluctuation features into the student ability assessment model, feature standardization is performed. The training of the student ability assessment model uses 10-fold cross-validation to evaluate generalization ability. In some embodiments, code commit record parsing uses an abstract syntax tree analyzer, and error debugging log parsing uses regular expressions to match error patterns. It is understood that the output dimension of the student ability assessment model is aligned with the requirements of the knowledge topology network. Optionally, fluctuation feature calculation can be supplemented with range statistics. It is understood that feature standardization uses the Z-score normalization method.

[0043] Example 2: See Figure 3The process of constructing a knowledge topology network representing the relationships between knowledge points decomposes knowledge points into a basic concept layer, a syntax rule layer, and an algorithm application layer based on the hierarchical structure of the programming knowledge system. In specific implementation, the basic concept layer includes the definitions of variables, data types, and operators; the syntax rule layer covers conditional statements, loop structures, and function definitions; and the algorithm application layer involves sorting, searching, and recursion. Data collection points are set up at each level to continuously collect data on students' learning time, practice completion rate, and error frequency for each knowledge point. Learning time is obtained by recording the interaction time between students and teaching resources related to the knowledge point. Practice completion rate is determined by whether the programming exercises submitted by students pass the test cases. Error frequency is calculated by counting the number of times the same type of error occurs during code debugging. An initial knowledge graph is constructed using knowledge points as nodes and the prerequisite / successor relationships between knowledge points as edges. The prerequisite / successor relationships are determined according to the teaching syllabus; for example, mastering variables is a prerequisite for learning conditional statements. The real-time collected learning behavior data is mapped to the corresponding nodes of the initial knowledge graph to form a knowledge topology network with learning behavior feature weights. The learning behavior feature weights are calculated using a weighted average formula, expressed as:

[0044]

[0045] in: This represents the normalized weights of learning behavior features. This represents the standardized learning time. This indicates the percentage of practice completion. This represents the normalized error frequency. , , These are pre-defined weighting coefficients, all of which are dimensionless values. In some embodiments, data collection points are integrated into an online programming environment, and the initial knowledge graph is stored and retrieved using a graph database. It can be understood that the calculation cycle for the learning behavior feature weights is set to daily batch processing.

[0046] The process of identifying core and auxiliary learning resources for each node in the network involves multi-dimensional evaluation of the learning resources associated with each node in the knowledge topology network. This multi-dimensional evaluation includes resource difficulty coefficients, instructional suitability, and historical usage effectiveness indicators. Resource difficulty coefficients are categorized into beginner, intermediate, and advanced levels based on the comprehension difficulty of the resource content and quantified numerically. Instructional suitability is obtained by calculating the semantic similarity between the resource description and the teaching objectives of the knowledge points. Historical usage effectiveness indicators are calculated based on the improvement in students' test scores for related knowledge points after using the resource. A resource quality threshold is set to classify the quality of learning resources for each node. This threshold is dynamically determined based on the distribution of evaluation scores for all learning resources. Resources that meet the excellent quality standard are marked as core learning resources, while resources that do not meet the excellent standard but meet basic teaching requirements are marked as auxiliary learning resources. Basic teaching requirements refer to the accuracy of the resource content and its coverage of the core concepts of the knowledge points. Identification information for core and auxiliary learning resources is stored for each node in the knowledge topology network. This identification information includes a resource ID and a resource type code. In practical implementation, multi-dimensional evaluation uses a weighted summation model to generate a comprehensive score. The resource difficulty coefficient is determined by subject matter experts using Bloom's Taxonomy. Instructional fit is calculated using a word embedding model to determine cosine similarity. The specific implementation of the word embedding model involves using natural language processing technology to digitally represent the descriptive text of learning resources and the instructional target text of knowledge points. A trained word embedding model maps words in the text to vectors in a high-dimensional feature space. This model is trained on a large-scale corpus and can capture semantic relationships and contextual information between words. The calculation of instructional fit first involves word segmentation and preprocessing of the resource description and instructional target texts, then inputting them into the word embedding model to generate corresponding feature vectors. Finally, the cosine value of the angle between the two vectors is calculated, serving as a quantitative indicator of instructional fit, reflecting the semantic matching degree between the resource content and the instructional target. Optionally, resource quality thresholds can be set separately according to knowledge point categories. In some embodiments, core learning resources are used to generate the main learning path, while auxiliary learning resources are used for supplementation and reinforcement. It can be understood that the storage of identification information is achieved through the node attributes of an extended graph database. Optionally, resource identification information is associated with real-time learning data to achieve dynamic updates. It is understandable that all multi-dimensional evaluation indicators are normalized to eliminate the influence of dimensions.

[0047] Example 3: In specific implementation, matching analysis is achieved by calculating the Euclidean distance between the ability indicators of each dimension in the student's ability graph and the ability thresholds required by the nodes in the knowledge topology network. The gap identification results are stored in the form of gap vectors. A large language model is used to analyze the gaps and generate a targeted learning objective sequence. The large language model adopts a Transformer-based architecture, taking the gap vector and the structural information of the knowledge topology network as input, and outputting a series of ordered learning objectives, each corresponding to one or more nodes in the knowledge topology network. Based on the learning objective sequence, the optimal learning path is searched in the knowledge topology network. The search algorithm uses the A* algorithm combined with the prerequisite and successor relationships of knowledge points and the weights of learning behavior features. The evaluation function considers path length and resource quality. Multi-round intent parsing ensures that the generated learning path conforms to the student's learning habits. Multi-round intent parsing collects student feedback through conversational interaction, adjusts the path order and resource type, and verifies the rationality of the learning path through semantic reasoning, checking whether the logical dependencies between knowledge points in the path are satisfied. The final output includes a personalized teaching strategy comprising a specific sequence of learning activities, a resource allocation plan, and a schedule. The learning activity sequence specifies the order of learning tasks for each knowledge point, the resource allocation plan clarifies the ratio of core learning resources to auxiliary learning resources, and the schedule sets the completion time for each stage. In some embodiments, the training data for the large language model includes historical learning paths and success stories, and the intent parsing module integrates a natural language understanding component. It is understood that the semantic reasoning module uses knowledge graph embedding technology to verify relational consistency.

[0048] The process of continuously tracking students' implementation of personalized teaching strategies involves real-time collection of learning trajectory data generated by students' execution of these strategies. This includes coding behavior sequences, video viewing progress, and quiz completion status. In practice, coding behavior sequences record the type and timestamp of code editing events; video viewing progress tracks the percentage of video playback completed; and quiz completion status collects quiz scores and submission times. The actual learning trajectory data is compared with the preset learning standards in the personalized teaching strategies to calculate learning progress deviations and learning behavior deviations. The learning progress deviation is calculated using the following formula:

[0049]

[0050] in: This indicates the deviation value of learning progress. Indicates the actual completion time. The expected completion time is indicated. Learning behavior deviation is calculated by comparing the difference between the actual resource usage pattern and the recommended pattern, with cosine similarity used as the difference metric. When a deviation exceeds a predetermined threshold, a strategy adjustment mechanism is triggered, which sends a signal to the dynamic adjustment module requesting a regeneration of the teaching strategy. When the deviation is within the allowable range, learning stability levels are assigned based on the magnitude of the deviation, with three levels: high, medium, and low, based on the percentile of the deviation value. In some embodiments, data collection is achieved through event tracking technology, and the comparison and analysis cycle is set to once a day. It is understood that the predetermined threshold is dynamically adjusted based on historical data. Optionally, the learning stability level can be used for risk assessment by the early warning module.

[0051] Example 4: Recalibrating the weighting parameters in the teaching strategy generation process based on the comparative analysis results includes calculating the learning urgency coefficient based on learning progress deviation and learning behavior deviation. In specific implementation, the learning urgency coefficient is calculated using the following formula:

[0052]

[0053] in: Indicates the learning urgency coefficient. This indicates the deviation value of learning progress. Indicates the learning behavior deviation value. and It is a preset weighting factor. and The values ​​are determined through historical data regression analysis. The resource fit coefficient is calculated based on students' actual performance at each knowledge point, using their test scores and resource usage frequency. Actual performance data is extracted from learning trajectory data. A dynamic weight adjustment model is established, which is a multiple linear regression model. The learning urgency coefficient and resource fit coefficient are used as input variables, and the output is the adjustment amount of the objective function weights during the teaching strategy generation process. The objective function weights include path length weights and resource quality weights. The parameters of the dynamic weight adjustment model are updated weekly, using a grid search method to optimize the model parameters and ensure the timeliness of weight adjustments. Personalized teaching strategies are regenerated using the adjusted weight parameters and updated learning data, with the regeneration process calling the algorithm of the teaching recommendation engine module. In some embodiments, learning progress deviation values ​​and learning behavior deviation values ​​are obtained from the learning progress monitoring module, and a time decay factor is introduced in the calculation of the resource fit coefficient to emphasize recent performance. It is understood that leave-one-out cross-validation is used in the training of the dynamic weight adjustment model to prevent overfitting. Optionally, weight factors λ1 and λ2 can be dynamically adjusted according to the characteristics of the subject.

[0054] Referring to Table 1, the process of predicting knowledge mastery weaknesses based on overall learning trajectory data includes mapping the learning stability level of each knowledge point to the corresponding node in the knowledge topology network. The learning stability level is divided into three levels: high, medium, and low. The mapping process is achieved by adding a stability attribute to each knowledge node. A dependency matrix between knowledge nodes is constructed. The dependency matrix is ​​based on the edge relationships of the knowledge topology network. The matrix elements represent the pre- and post-relationship strength between nodes. The mutual influence coefficient between nodes is calculated. The mutual influence coefficient is calculated using the matrix eigenvector centrality algorithm, reflecting the influence of the node in the network. A graph neural network is used to perform representation learning on the knowledge topology network with stability labels. The graph neural network adopts a graph convolutional network structure. The input node features include stability level and mutual influence coefficient. The neighbor node information is aggregated through multi-layer convolution to mine potential risk propagation paths. The risk propagation path is identified by calculating the path probability between nodes. The output is an overall risk prediction for the entire knowledge system. The overall risk prediction is represented in the form of a risk score. The set of weak knowledge points that may cause chain risks is identified. The set of weak knowledge points is the group of nodes whose risk scores exceed a threshold. In practice, the dependency matrix is ​​constructed using an adjacency matrix for storage, and the mutual influence coefficients are calculated using a variant of the PageRank algorithm.

[0055] Table 1: Partial Parameter Table of the Weight Dynamic Adjustment Model

[0056]

[0057] In some embodiments, the graph neural network is trained using a stochastic gradient descent optimizer, and the risk score threshold is set using the percentile method. It is understood that the stability attribute mapping process is linked to the learning progress monitoring module. Optionally, the risk propagation path mining can be combined with a community detection algorithm. Optionally, the output format of the overall risk prediction is a structured list.

[0058] See Figure 4 In the dynamic change analysis of weight factors during the weight parameter adjustment phase, the urgency weight is learned. ) and behavioral bias weights ( The dynamic trend reflects the execution effect of the dynamic weight adjustment model. In specific operation, the time period is in weeks, and the weight factor value range corresponds to the set range of 0.1-0.9 in Table 1. The change process from week 1 to week 8 shows a two-way fluctuation characteristic: the learning urgency weight (λ1) drops to a low point (0.3) in week 3, then gradually rises and reaches 0.6 in week 8; the behavioral deviation weight (λ2) rises to a high point (0.7) in week 3, then fluctuates and decreases to 0.4 in week 8. The two cross in week 7, reflecting the real-time calibration of the weight of the teaching strategy objective function weight by the dynamic weight adjustment model based on the input of the learning urgency coefficient and the resource suitability coefficient. During the parameter change process, the update cycle is consistent with the 7-day setting in Table 1. The fluctuation range of the weight factor corresponds to the dynamic calculation results of learning progress deviation and learning behavior deviation. Its trend change can be directly related to the deviation data collection of the learning progress monitoring module and the parameter output of the dynamic weight adjustment model.

[0059] Example 5: The process of allocating specific reinforcement resources to detected weaknesses includes setting an overall risk threshold. In practice, the overall risk threshold is set based on the statistical distribution of historical learning data, for example, using the 90th percentile of the overall risk estimate as the threshold. When the overall risk estimate exceeds the threshold, the risk transmission coefficient of each knowledge point is calculated. In practice, the risk transmission coefficient is calculated by analyzing the dependencies between nodes in the knowledge topology network, using network flow theory to simulate the risk propagation path, and calculating the comprehensive vulnerability index of each knowledge point based on the risk transmission coefficient and the local stability level. The comprehensive vulnerability index is calculated using the formula:

[0060]

[0061] in: Indicating the comprehensive vulnerability index Indicates the risk transmission coefficient. Indicates the local stability level. and These are preset weighting coefficients, whose values ​​are determined through historical data regression analysis. For knowledge points whose comprehensive vulnerability index exceeds a predetermined threshold, suitable specialized training materials are selected from their associated auxiliary learning resources, combined into reinforcement learning modules, and inserted into the corresponding positions of personalized teaching strategies. In some embodiments, the overall risk threshold can be dynamically adjusted based on real-time risk data, such as updating the threshold according to the fluctuation range of recent risk estimates. It can be understood that the calculation of the risk transmission coefficient uses the attention mechanism in graph neural networks to capture the influence between nodes.

[0062] In practical implementation, the selection process for specialized training materials is based on semantic matching between resource tags and knowledge points. For example, for the knowledge point of "variable definition," resources containing explanations of basic concepts and example code are selected from the auxiliary learning resource library. The combination of reinforcement learning modules involves resource sorting and integration. The module content includes targeted exercises and instant feedback mechanisms. When inserting personalized teaching strategies into appropriate positions, the system determines the insertion point based on the prerequisite and successor relationships of knowledge points. For example, a reinforcement module may be inserted before the learning of the direct successor node of a weak knowledge point. In some embodiments, the resource selection algorithm considers the suitability of the difficulty gradient of the resources to the student's current ability. Optionally, the predetermined threshold of the comprehensive vulnerability index can be adjusted according to individual student differences, for example, by setting a personalized threshold based on the student's historical performance data.

[0063] See Figure 5 In the analysis of the overall risk threshold dynamic adjustment effect during the threshold optimization phase, the relationship between the actual overall risk value, static risk threshold, dynamic risk threshold, and risk-exceeding areas is presented using the teaching week as the time dimension. Specifically, the static risk threshold (0.65) serves as a fixed baseline, while the dynamic risk threshold is adaptively adjusted based on real-time risk data. The actual overall risk value reaches its peak (0.75) in week 3, significantly exceeding both the static and dynamic thresholds, and the corresponding area is identified as a risk-exceeding area. After week 4, the dynamic risk threshold adjusts synchronously with the decrease in the actual risk value, and the fit between the two gradually improves, reflecting the responsiveness of the dynamic threshold to actual risk fluctuations. The core value of this figure lies in its intuitive quantification of the effectiveness of the dynamic risk threshold adjustment mechanism. Through the flexible adaptation of the dynamic threshold, the duration and magnitude of the actual risk value exceeding the standard are effectively controlled, providing data support for subsequent risk warning and reinforcement resource allocation for weak knowledge points.

[0064] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0065] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A teaching programming system based on a large language model, characterized in that, The system includes: The student characteristic analysis module constructs a learning element profile reflecting students' programming abilities and outputs a student ability map containing multiple dimensions of ability indicators. The learning resource management module receives student ability maps, collects real-time learning behavior data according to the preset knowledge system architecture, constructs a knowledge topology network that represents the relationship between knowledge points, and identifies the core learning resources and auxiliary learning resources of each node in the network. The teaching recommendation engine module matches students' ability graphs with knowledge topology networks and generates personalized teaching strategies containing learning path sequences through multi-round intent parsing and semantic reasoning of a large language model. The learning progress monitoring module collects learning trajectory data generated by students implementing personalized teaching strategies in real time; compares the actual learning trajectory data with the preset learning standards in the personalized teaching strategies, and calculates learning progress deviation and learning behavior deviation; when the deviation exceeds the predetermined threshold, the strategy adjustment mechanism is triggered; when the deviation is within the allowable range, the learning stability level is divided according to the magnitude of the deviation. The dynamic adjustment module calculates the learning urgency coefficient based on learning progress deviations and learning behavior deviations; it calculates the resource suitability coefficient based on students' actual performance at each knowledge point; it establishes a weight dynamic adjustment model, taking the learning urgency coefficient and resource suitability coefficient as inputs, and outputting the adjustment amount of the weights of each objective function during the generation of teaching strategies; it updates the parameters of the weight dynamic adjustment model according to a set period to ensure the timeliness of weight adjustments; and it regenerates personalized teaching strategies using the adjusted weight parameters and updated learning data. The early warning and intervention module maps the learning stability level of each knowledge point to the corresponding nodes in the knowledge topology network; constructs a dependency matrix between knowledge nodes and calculates the mutual influence coefficient between nodes; uses a graph neural network to perform representation learning on the knowledge topology network with stability labels to explore potential risk propagation paths; outputs an overall risk prediction for the entire knowledge system and identifies a set of weak knowledge points that may cause cascading risks; sets an overall risk threshold, and calculates the risk transmission coefficient of each knowledge point when the overall risk prediction exceeds the threshold; calculates the comprehensive vulnerability index of each knowledge point based on the risk transmission coefficient and the local stability level; for knowledge points whose comprehensive vulnerability index exceeds the predetermined threshold, selects suitable specialized training materials from their associated auxiliary learning resources, combines them into a reinforcement learning module, and inserts it into the corresponding position of the personalized teaching strategy.

2. The teaching programming system based on a large language model as described in claim 1, characterized in that, The process of constructing a learning element profile reflecting students' programming abilities includes: collecting code submission records, error debugging logs, and concept test scores generated by students during the programming learning process; extracting features from these multi-source learning data to obtain code complexity features, error pattern features, and knowledge mastery features; organizing the above features according to time series; extracting learning trend features and fluctuation features through time series pattern analysis; and inputting the learning trend features and fluctuation features into a student ability assessment model trained with historical learning data to output a student ability map composed of multiple dimensions of ability indicators.

3. The teaching programming system based on a large language model as described in claim 2, characterized in that, The process of constructing a knowledge topology network representing the relationships between knowledge points includes: decomposing knowledge points into a basic concept layer, a syntax rule layer, and an algorithm application layer according to the hierarchical structure of the programming knowledge system; setting data collection points at each layer to continuously collect data on students' learning time, practice completion rate, and error frequency for each knowledge point; constructing an initial knowledge graph with knowledge points as nodes and the prerequisite and successor relationships between knowledge points as edges; and mapping the real-time collected learning behavior data to the corresponding nodes of the initial knowledge graph to form a knowledge topology network with learning behavior feature weights.

4. The teaching programming system based on a large language model as described in claim 3, characterized in that, The process of identifying the core learning resources and auxiliary learning resources of each node in the network includes: evaluating the learning resources associated with each node in the knowledge topology network from multiple dimensions, including resource difficulty coefficient, teaching suitability, and historical usage effect indicators; setting resource quality thresholds and classifying the learning resources of each node according to quality; marking resources that meet the excellent quality standard as core learning resources and resources that do not meet the excellent quality standard but meet the basic teaching requirements as auxiliary learning resources; and storing the identification information of core learning resources and auxiliary learning resources for each node in the knowledge topology network.

5. The teaching programming system based on a large language model as described in claim 4, characterized in that, The process of generating a personalized teaching strategy containing a learning path sequence through multi-round intent parsing and semantic reasoning using a large language model includes: matching and analyzing the student's ability map with the knowledge topology network to identify the gap between the student's current ability and the requirements of the knowledge system; analyzing the gap through the large language model to generate a targeted learning objective sequence; searching for the optimal learning path in the knowledge topology network based on the learning objective sequence; ensuring that the generated learning path conforms to the student's learning habits through multi-round intent parsing, and verifying the rationality of the learning path through semantic reasoning; and finally outputting a personalized teaching strategy containing a specific learning activity sequence, resource allocation plan, and schedule.

6. A teaching recommendation method based on a large language model, applied to a teaching programming system based on a large language model as described in any one of claims 1-5, characterized in that, Includes the following steps: Construct a learning element profile that reflects students' programming abilities, and output a student ability map containing multiple dimensions of ability indicators; Based on the pre-defined knowledge system architecture, real-time learning behavior data is collected to construct a knowledge topology network that represents the relationship between knowledge points, and the core learning resources and auxiliary learning resources of each node in the network are identified. By matching student competency maps with knowledge topology networks, personalized teaching strategies containing learning path sequences are generated through multi-round intent parsing and semantic reasoning of a large language model. Continuously track students' implementation of personalized teaching strategies, collect actual learning trajectory data, and compare and analyze them with preset learning standards; Based on the comparative analysis results, the weight parameters in the teaching strategy generation process are recalibrated, and the optimized personalized teaching strategy is output. Based on the overall learning trajectory data, we predict the weak points in knowledge acquisition and allocate special reinforcement resources for the detected weak points.

Images