Personalized education method and system based on large model

Abstract

The application discloses a kind of personalized education method and system based on large model.The method comprises the following steps: obtaining interaction dataset, pre-processing to obtain the cognitive index and interest index of the educated person;According to double index, generate a plurality of groups of knowledge unit sequences and map to interest index, calculate the push weight order to generate initial sequence set;Calculate knowledge unit transition probability, classified as inter-sequence probability and out-of-sequence probability;Under the constraint of preset cognitive index, iteration is carried out, and stage interaction feedback is input into large language model to update double index in real time and generate stage sequence set cyclically;When the inter-sequence probability and out-of-sequence probability corresponding to the target education sequence set cumulatively reach 1 under the preset condition, the target education sequence set is output.The application effectively solves the problems of traditional path generation rigidity, single probability control and lack of stable judgment in iteration, significantly improves the dynamic adaptation accuracy of personalized education scheme and the stability of system output.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a personalized education method and system based on large models. Background Technology

[0002] With the deepening application of artificial intelligence technology in the field of educational informatization, personalized learning path planning and knowledge sequence generation have become core research directions for intelligent education systems. In existing technologies, most intelligent education systems rely on static knowledge graphs or fixed rule engines for content delivery, and usually use a unified algorithm framework to process the interaction data of all learners, failing to fully establish a dynamic modeling mechanism for learners' real-time cognitive state and interest evolution.

[0003] Existing technologies reveal the following prominent problems in practical applications: First, traditional systems lack a continuous iterative exponential calibration mechanism when generating knowledge sequences. Systems often perform a one-time ability assessment in the initial stage, then push content along a fixed path. When learners experience a shift in cognitive load or a decline in interest-driven motivation, the system cannot capture the state shift in real time, leading learners to easily fall into a predicament of cognitive overload or learning motivation exhaustion.

[0004] Therefore, there is an urgent need for a personalized education method and system based on artificial intelligence, which can understand the characteristics of different learners and generate targeted education programs based on probability training. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a personalized education method and system based on a large model, which can deeply understand the characteristics of different educational groups and generate targeted education programs based on probability training.

[0006] A first aspect of the present invention provides a personalized education method based on a large model, comprising the following steps: S1. Obtain the interactive dataset of the current education stage, and preprocess the interactive dataset to obtain the cognitive index and interest index of the learners; S2. Based on the cognitive index and the interest index, generate multiple sets of knowledge unit sequences from a preset knowledge base, and each knowledge unit sequence is mapped to a different interest index; Calculate the push weight of each knowledge unit sequence, and sort the knowledge unit sequences based on the push weight to obtain an initial sequence set; S3. Calculate the transition probability of the knowledge units in the initial sequence set. The transition probability includes the in-sequence probability and the out-of-sequence probability. The in-sequence probability is the content transfer rate within the knowledge unit sequence determined by the current interest index. The out-of-sequence probability is the content transfer rate between different knowledge unit sequences determined by different interest indices. S4. Execute the iteration under the preset cognitive index constraint, and at the end of each stage of the iteration, input the interaction feedback data of the current stage into the large model, update the cognitive index and the interest index, and return to execute S2-S3 to obtain the stage sequence set. S5. Calculate the transition probability of the phase sequence set and determine whether the sum of the in-sequence probability and the out-of-sequence probability is 1 under the preset conditions; If it is not 1, return to S4 and execute the iteration; If the value is 1, output the current stage sequence set as the target education sequence set for the next education stage.

[0007] As a preferred method, when executing S3, the following steps are also performed: Based on the initial cognitive index and the interest index, a benchmark model for the construction phase is established, and the iterative process is divided into the cognitive construction phase, the structural reorganization phase, and the cross-domain migration phase. Assign a baseline weighting range to the in-sequence probability and the out-of-sequence probability for each period, and set a cognitive expectation growth curve and an interest dynamic decay baseline; When switching between iteration phases, extract the actual interaction feedback data of the current phase sequence set, and calculate the deviation of the real-time cognitive index and the interest index from the expected cognitive growth curve and the dynamic interest decay baseline. If the deviation falls within the preset deviation range, the current benchmark weight range is maintained to execute the next stage; If the deviation exceeds the preset deviation tolerance range, and the cognitive index lags, the weight of the in-sequence probability is increased and the frequency reduction factor of the out-of-sequence probability is injected to force convergence of the in-sequence path; When the interest index is lower than the interest dynamic decay baseline, the cross-domain transition weight of the out-of-sequence probability is activated, and the sequence is redirected to a high-matching knowledge unit sequence based on the volatility of the interest index. The redistribution parameters and deviation compensation coefficients are input into the large model to perform nonlinear mapping, generating the cognitive index and interest index for the next iteration, until the preset conditions are met.

[0008] As a preferred method, executing S1 specifically includes the following steps: A large model is used to perform multimodal feature analysis on the interactive dataset of the current education stage, extracting the temporal dependency features of knowledge acquisition trajectory, the topological path features of logical reasoning, and the attention mask features of interactive behavior. The temporal dependency features and the topological path features are input into a multidimensional feature weighting network, the confidence weights are fused and normalization mapping is performed to calculate the cognitive index. The attention mask features are input into the time window decay model, and the interest index is calculated by combining the distribution parameters of dwell time and the polarity score of emotional feedback.

[0009] The preset conditions include iteration termination conditions and probability convergence verification conditions; The iteration termination condition is that the current iteration round number reaches the preset maximum iteration threshold, or the cumulative rate of change of the probability between sequences and the probability outside sequences in two adjacent iterations is lower than the preset convergence deviation threshold. The probability convergence verification condition is that the real-time updated cognitive index reaches the preset knowledge mastery threshold, and the fluctuation variance of the interest index falls into the preset stable range. When both the iteration termination condition and the probability convergence verification condition are met simultaneously, the stage sequence set is determined to converge into the target education sequence set. Normalization distribution verification is performed on the probabilities between sequences, cross-domain correlation compensation is performed on the probabilities outside the sequences, the weighted fusion value of the two types of probabilities is forcibly calibrated to 1 through the probability accumulation convergence function, and the target education sequence set is locked and output.

[0010] As a preferred approach, self-testing is also performed during the iteration process, which includes the following steps: The probability convergence trajectory of the sequence sets at each stage is compared in real time, and if the trajectories are consistent, the iteration continues. If a branch deviation occurs, the first self-test is performed, extracting a set of benchmark sequences with consistent sequence length and probability evolution characteristics from the historical database, constructing a test set and assigning different cognitive load weights and interest exploration weights, and outputting a joint index of calculation path completion and cognitive achievement rate. If the joint index does not fall within the joint confidence interval, a second self-test is performed to mark the abnormal probability nodes and calculate their perturbation coefficients on the overall probability accumulation. When the perturbation coefficient exceeds the preset blocking threshold, the sequence is reset. The knowledge unit sequence is adjusted according to the cognitive index deviation, the topological weight of the out-of-sequence probability jump path is reset according to the interest index decay coefficient, and S2-S4 are re-executed.

[0011] A second aspect of the present invention provides a personalized education system based on a large model, comprising: The data preprocessing module is configured to acquire the interactive dataset, perform multimodal feature parsing and decay modeling, and calculate the cognitive index and interest index. The sequence generation and sorting module is configured to generate multiple sets of knowledge unit sequences from a preset knowledge base based on dual indices and map them to interest indices, calculate push weights and sort them to generate an initial sequence set. The probability calculation and iteration module is configured to calculate the probabilities between sequences and the probabilities outside sequences, perform iterations under the constraint of the cognitive index, receive stage interactive feedback input to the large language model to update the double index in real time, and perform sequence generation and probability calculation in a loop to obtain the stage sequence set; The convergence determination and output module is configured to monitor preset conditions during the iteration process. When the cumulative probability between sequences and the probability outside sequences reaches 1, the path is locked and the target educational sequence set is output. The self-detection and correction module is configured to perform probability trajectory comparison, joint indicator verification and abnormal node disturbance compensation, and trigger sequence reset and probability redistribution when iterative deviation occurs.

[0012] Compared with the prior art, the present invention has the following advantages: This invention analyzes the interaction data of students or any learner at the current learning stage using a large-scale model, extracting the learning time sequence, logical path, and attention concentration. This allows for real-time calculation of a cognitive index representing knowledge mastery and an interest index representing learning motivation. This method accurately captures students' true state of mind regarding the difficulty and interest of learning, providing accurate data support for the subsequent generation and delivery of learning content. It effectively solves the problems of traditional educational software being slow to respond and unable to keep up with changes in students' states.

[0013] This invention cleverly divides the recommendation probability of learning content into in-sequence probability and out-of-sequence probability. In-sequence probability refers to the probability of continuing in-depth learning in the current area of ​​interest, while out-of-sequence probability refers to the probability of jumping to other areas of interest. During the learning process, the system compares the student's actual state with the expected goals in real time. If it finds that the student is learning slower than expected, resulting in cognitive lag, the system increases the in-sequence probability, encouraging the student to consolidate their learning at the current level and reducing jumps. If it finds that the student's interest is declining, the system increases the out-of-sequence probability, recommending new content that is more interesting to the student. This mechanism intelligently balances in-depth consolidation and breadth expansion, making the recommended learning path more closely match the student's real-time state.

[0014] This invention ensures the integrity and effectiveness of the learning process through a judgment logic that uses multiple conditions and probabilities to sum to 1. When recommending content, the system checks if the sum of the in-sequence probability and the out-of-sequence probability is 1. If not, it continues to adjust; only when they equal 1 is the current learning plan, i.e., the target educational sequence set, output to the next learning stage. Simultaneously, when various learning indicators are met, the system performs a final calibration and compensation on these two probabilities, forcing their sum to reach 1. This approach algorithmically avoids the system falling into invalid loops or false completion states, ensuring that the final output learning plan truly enables students to master knowledge and maintain stable learning interest, achieving a complete closed loop in the learning process.

[0015] This invention also performs real-time self-checks during the recommendation process. The system compares the current recommended path with historical standard paths. If a deviation from the normal trajectory is detected, a first round of self-checking is conducted to assess the student's learning progress and knowledge attainment rate. If the indicators are abnormal, a second round of in-depth self-checking is performed to identify the abnormal nodes causing the recommendation errors and calculate their impact on the overall recommendation. If the impact exceeds a set limit, the system automatically replans, readjusting the learning content and navigation paths based on the student's actual understanding and interests. This mechanism can detect and correct recommendation deviations as early as possible, preventing errors from accumulating and greatly improving the stability and reliability of personalized education programs in practical use. Attached Figure Description

[0016] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0017] Figure 1 This is a structural block diagram of the system provided in the embodiments of the present invention. Detailed Implementation

[0018] 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.

[0019] Example 1 This embodiment of the personalized education system based on artificial intelligence, such as Figure 1 As shown, the system includes a data preprocessing module, a sequence generation and sorting module, a probability calculation and iteration module, a convergence determination and output module, and a self-detection and correction module. These modules are interconnected via an internal high-speed data bus and distributed computing nodes, collaboratively performing the generation, iterative optimization, and quality verification of personalized education programs.

[0020] In this embodiment, the data preprocessing module is configured to acquire the interactive dataset of the current educational stage, perform multimodal feature parsing and decay modeling, and calculate the cognitive index and interest index. Specifically, the system collects multimodal interactive data of the current educational stage through the learner's terminal, including answer record text, gaze trajectory during video interaction, operation log timestamps, and voice interaction audio.

[0021] The module calls a pre-trained large model to perform feature vectorization analysis on the data, extracting temporal dependency features of the knowledge mastery trajectory, such as the time decay trend of continuous correct interactions; extracting topological path features of logical reasoning, such as the node connectivity density of problem-solving steps; and extracting attention mask features of interactive behavior, such as the normalized weight distribution of dwell time in hot areas of the interface. The temporal dependency features and topological path features are input into a multi-dimensional feature weighting network. The network assigns weight coefficients of 0.45, 0.35, and 0.20 to the temporal features, topological features, and interactive feedback data features according to the preset cognitive dimension confidence weights, and performs weighted summation. Then, Sigmoid normalization mapping is performed, and the cognitive index with a value range of [0,1] is output. This index represents the degree of structure of the learner's current knowledge network and cognitive load capacity.

[0022] Simultaneously, the attention mask features are input into the time window decay model, combined with the distribution parameters of dwell time, which are determined by the ratio of actual dwell time to standard cognitive time, and combined with the polarity score of emotional feedback. Through the sentiment analysis model, positive feedback is mapped to the interval [0.8, 1.0] and negative feedback is mapped to the interval [0.1, 0.3]. The interest index, which also has a value range of [0, 1], is calculated. This index represents the learner's intrinsic cognitive drive strength and willingness to explore.

[0023] The sequence generation and sorting module is configured to generate multiple sets of knowledge unit sequences from a preset knowledge base based on the cognitive index and interest index, map them to different interest indices, calculate the push weight of each knowledge unit sequence, and sort each knowledge unit sequence based on the push weight to obtain an initial sequence set.

[0024] Specifically, the pre-defined knowledge base stores atomic knowledge units categorized by subject area, difficulty gradient, and cognitive dimension. The module filters units exceeding the current cognitive load using the current cognitive index as the retrieval threshold, and uses the interest index as the cluster center to call a graph database to generate multiple sets of candidate knowledge unit sequences. For example, some sequences focus on consolidating basic concepts, while others focus on interdisciplinary expansion. Each set of sequences is mapped to a corresponding interest index interval. The module uses a weighted utility function to calculate the push weight of each sequence. The formula for this function is W = α·(1 - |Interest Index - Sequence Center Interest Value|) + β·(Cognitive Index / Average Cognitive Difficulty of Sequence), where the adjustment coefficients α and β are 0.6 and β is 0.4. Based on the calculated push weight W, the multiple sets of sequences are sorted in descending order, and the top four sequences by weight are selected to form the initial sequence set. A sequence index table containing sequence identifiers, node topology, and initial weights is generated.

[0025] The probability calculation and iteration module is configured to calculate the in-sequence probability and out-of-sequence probability of knowledge units in the initial sequence set, perform iteration under the preset cognitive index constraint, and input the interaction feedback data of the current stage into the large model at the end of each iteration stage to update the cognitive index and interest index. The sequence generation and probability calculation are performed in a loop to obtain the stage sequence set.

[0026] The module constructs a transition matrix to calculate the transition probability of knowledge units and strictly classifies them into intra-sequence probability and out-of-sequence probability. The intra-sequence probability is defined as the content transfer rate within the knowledge unit sequence determined by the current interest index; the out-of-sequence probability is defined as the content transfer rate between different knowledge unit sequences determined by different interest indices.

[0027] During the iteration process, a baseline model is first constructed based on the initial cognitive index and interest index, dividing the complete learning cycle into a cognitive construction period, a structural reorganization period, and a cross-domain transfer period. Baseline weight ranges for in-sequence and out-of-sequence probabilities are assigned to each period. For example, in the cognitive construction period, the weight range for in-sequence probabilities is set to [0.7, 0.9], and the weight range for out-of-sequence probabilities is set to [0.1, 0.3]. Simultaneously, a cognitive expected growth curve is established, which initially exhibits a linear increasing trend (growth slope of 0.05 / round), and a dynamic interest decay baseline is set, which exhibits an exponential decay trend over time (decay coefficient of 0.02 / round).

[0028] During the iteration phase, at the switching node, the actual interaction feedback data of the current stage sequence set is extracted, and the deviation Δ of the real-time cognitive index and interest index relative to the expected growth curve of cognition and the dynamic decay baseline of interest is calculated. If the deviation satisfies |Δ|≤0.08, the baseline weight range is maintained to enter the next stage. If the deviation exceeds this tolerance range, a probability redistribution mechanism is triggered: when the cognitive index lags (Δ>0.08), the weight of the probability within the sequence is significantly increased, and a frequency reduction decay factor is injected into the probability outside the sequence to force convergence of the path within the sequence to solidify the knowledge base; when the interest index is lower than the decay baseline (Δ<-0.1), the cross-domain transition weight of the probability outside the sequence is activated (increased to 0.4), and redirected to the high-matching knowledge unit sequence according to the volatility of the interest index (sliding window variance greater than 0.05). The module inputs the redistribution parameters and deviation compensation coefficients into the large model to perform nonlinear mapping processing, and outputs the updated cognitive index and interest index until the preset conditions are met.

[0029] The convergence determination and output module is configured to calculate the transition probability of the stage sequence set, determine whether the sum of the in-sequence probability and the out-of-sequence probability is 1 under the preset conditions. If it is not 1, the iteration is returned. If it is 1, the path is locked and the current stage sequence set is output as the target education sequence set to the next education stage.

[0030] Specifically, the preset conditions include an iteration termination condition and a probability convergence verification condition. The iteration termination condition is set as follows: the current iteration rounds reach 20 rounds, or the cumulative change rate of the in-sequence probability and out-of-sequence probability in two adjacent iterations is less than 0.01. The probability convergence verification condition is set as follows: the updated cognitive index reaches 0.85, and the sliding window fluctuation variance of the interest index falls within the range of [0.01, 0.04]. When the system simultaneously meets the above two conditions, it is determined that the current stage sequence set has converged and evolved into the target education sequence set. The module then performs a normalized distribution verification on the in-sequence probability to ensure that the sum of the in-sequence probability meets the unitization requirement, and performs cross-domain correlation compensation on the out-of-sequence probability, introducing an attention compensation coefficient to balance the information loss during the cross-sequence jump process. The weighted fusion value of the two types of probabilities is forcibly calibrated through the probability cumulative convergence function so that its sum is exactly 1. After calibration, the node path and jump index of the target education sequence set are locked, and it is output as the target education sequence set to the next education stage.

[0031] The self-detection and correction module is configured to perform probability trajectory comparison, joint indicator verification, and abnormal node disturbance compensation, and to trigger sequence reset and probability redistribution when iterative deviation occurs.

[0032] During the iteration process, the module compares the probability convergence trajectory of the sequence sets at each stage in real time and records the evolution curves of the in-sequence probability and out-of-sequence probability in each round. If the trajectories of multiple computational sub-threads remain consistent, the regular iteration continues. If a branch deviation occurs, the first self-test is performed: extracting a set of benchmark sequences with the same sequence length and consistent probability evolution characteristics from the historical education scheme library, constructing a dual test set and assigning high cognitive load weight and high interest exploration weight respectively, performing path deduction respectively, and outputting a joint index J = 0.5 × path completion degree + 0.5 × cognitive achievement rate. If the joint index J does not fall within the joint confidence interval [0.75, 0.95], the second self-test is performed: marking abnormal probability nodes and calculating their perturbation coefficient η = |node probability change| / total overall probability. When the perturbation coefficient η>0.15, the sequence reset mechanism is triggered: the granularity of the knowledge unit sequence is dynamically adjusted according to the cognitive index deviation, and the longer sequence is divided into shorter micro-unit sequences; the topological weight of the probability jump path outside the sequence is reset according to the interest index decay coefficient, the weight allocation of cross-domain connection edges is reduced, and then the module clears the abnormal cache state, re-initializes the initial sequence set, and instructs the system to re-execute the sequence generation, probability calculation and iterative convergence process to ensure that the output scheme is not affected by local model drift.

[0033] In a specific application scenario of this embodiment, the system receives an interactive dataset of a high school student's current educational stage. The data preprocessing module calculates an initial cognitive index of 0.55 and an interest index of 0.72. Based on this, the sequence generation and sorting module generates five sets of sequences from the knowledge base, focusing on the deepening of function concepts and the application of physical modeling. After weight calculation and sorting, an initial sequence set is obtained. The probability calculation and iteration module determines that the system is currently in a structural reorganization period, setting the probability benchmark within the sequence to a higher weight and the probability benchmark outside the sequence to a lower weight. When the iteration reaches the middle stage, feedback data shows that the student's interest has plummeted, with the interest index falling below the decay baseline and the deviation exceeding the tolerance range. The system immediately triggers a reallocation, increasing the cross-domain transition weight of the probability outside the sequence and injecting a visual interactive experiment sequence. Subsequently, the module continues to iterate. When the iteration enters the later stage, the cognitive index stabilizes at 0.88, the interest variance falls into the stable range of [0.02, 0.03], and the probability change rate is less than 0.008, satisfying all preset conditions. The convergence determination and output module calculates the transition probability, determines that the sum of the in-sequence probability and the out-of-sequence probability is 1, performs normalization and compensation, and locks the output target educational sequence set containing multiple core nodes to the next educational stage. During this process, the self-detection and correction module detects an abnormal surge in the out-of-sequence probability of a certain sub-path, calculates the perturbation coefficient to be 0.18, exceeding the blocking threshold, and immediately performs sequence reset and topology weight reallocation, effectively preventing the accumulation of erroneous paths, and finally outputs the optimal learning plan that highly matches the student's real-time cognitive and interest state.

[0034] This embodiment thoroughly overcomes the technical defects of traditional education systems, such as rigid paths, single probability models, and uncontrollable convergence, by deeply integrating dual-exponential dynamic driving, dual-dimensional probabilistic classification control, multi-dimensional convergence verification, and trajectory self-detection mechanisms. It achieves high-precision generation and high-robust output of personalized education programs.

[0035] Example 2 Providing a personalized education approach based on a large model includes the following steps: S1. Obtain the interactive dataset of the current education stage, and preprocess the interactive dataset to obtain the cognitive index and interest index of the learners; S2. Based on the cognitive index and the interest index, generate multiple sets of knowledge unit sequences from a preset knowledge base, and each knowledge unit sequence is mapped to a different interest index; Calculate the push weight of each knowledge unit sequence, and sort the knowledge unit sequences based on the push weight to obtain an initial sequence set; S3. Calculate the transition probability of the knowledge units in the initial sequence set. The transition probability includes the in-sequence probability and the out-of-sequence probability. The in-sequence probability is the content transfer rate within the knowledge unit sequence determined by the current interest index. The out-of-sequence probability is the content transfer rate between different knowledge unit sequences determined by different interest indices. S4. Execute the iteration under the preset cognitive index constraint, and at the end of each stage of the iteration, input the interaction feedback data of the current stage into the large model, update the cognitive index and the interest index, and return to execute S2-S3 to obtain the stage sequence set. S5. Calculate the transition probability of the phase sequence set and determine whether the sum of the in-sequence probability and the out-of-sequence probability is 1 under the preset conditions; If it is not 1, return to S4 and execute the iteration; If the value is 1, output the current stage sequence set as the target education sequence set for the next education stage.

[0036] As a preferred method, when executing S3, the following steps are also performed: Based on the initial cognitive index and the interest index, a benchmark model for the construction phase is established, and the iterative process is divided into the cognitive construction phase, the structural reorganization phase, and the cross-domain migration phase. Assign a baseline weighting range to the in-sequence probability and the out-of-sequence probability for each period, and set a cognitive expectation growth curve and an interest dynamic decay baseline; When switching between iteration phases, extract the actual interaction feedback data of the current phase sequence set, and calculate the deviation of the real-time cognitive index and the interest index from the expected cognitive growth curve and the dynamic interest decay baseline. If the deviation falls within the preset deviation range, the current benchmark weight range is maintained to execute the next stage; If the deviation exceeds the preset deviation tolerance range, and the cognitive index lags, the weight of the in-sequence probability is increased and the frequency reduction factor of the out-of-sequence probability is injected to force convergence of the in-sequence path; When the interest index is lower than the interest dynamic decay baseline, the cross-domain transition weight of the out-of-sequence probability is activated, and the sequence is redirected to a high-matching knowledge unit sequence based on the volatility of the interest index. The redistribution parameters and deviation compensation coefficients are input into the large model to perform nonlinear mapping, generating the cognitive index and interest index for the next iteration, until the preset conditions are met.

[0037] As a preferred method, executing S1 specifically includes the following steps: A large model is used to perform multimodal feature analysis on the interactive dataset of the current education stage, extracting the temporal dependency features of knowledge acquisition trajectory, the topological path features of logical reasoning, and the attention mask features of interactive behavior. The temporal dependency features and the topological path features are input into a multidimensional feature weighting network, the confidence weights are fused and normalization mapping is performed to calculate the cognitive index. The attention mask features are input into the time window decay model, and the interest index is calculated by combining the distribution parameters of dwell time and the polarity score of emotional feedback.

[0038] The preset conditions include iteration termination conditions and probability convergence verification conditions; The iteration termination condition is that the current iteration round number reaches the preset maximum iteration threshold, or the cumulative rate of change of the probability between sequences and the probability outside sequences in two adjacent iterations is lower than the preset convergence deviation threshold. The probability convergence verification condition is that the real-time updated cognitive index reaches the preset knowledge mastery threshold, and the fluctuation variance of the interest index falls into the preset stable range. When both the iteration termination condition and the probability convergence verification condition are met simultaneously, the stage sequence set is determined to converge into the target education sequence set. Normalization distribution verification is performed on the probabilities between sequences, cross-domain correlation compensation is performed on the probabilities outside the sequences, the weighted fusion value of the two types of probabilities is forcibly calibrated to 1 through the probability accumulation convergence function, and the target education sequence set is locked and output.

[0039] As a preferred approach, self-testing is also performed during the iteration process, which includes the following steps: The probability convergence trajectory of the sequence sets at each stage is compared in real time, and if the trajectories are consistent, the iteration continues. If a branch deviation occurs, the first self-test is performed, extracting a set of benchmark sequences with consistent sequence length and probability evolution characteristics from the historical database, constructing a test set and assigning different cognitive load weights and interest exploration weights, and outputting a joint index of calculation path completion and cognitive achievement rate. If the joint index does not fall within the joint confidence interval, a second self-test is performed to mark the abnormal probability nodes and calculate their perturbation coefficients on the overall probability accumulation. When the perturbation coefficient exceeds the preset blocking threshold, the sequence is reset. The knowledge unit sequence is adjusted according to the cognitive index deviation, the topological weight of the out-of-sequence probability jump path is reset according to the interest index decay coefficient, and S2-S4 are re-executed.

[0040] The specific execution steps in this embodiment are similar to the steps for configuring and executing each module in Embodiment 1, and will not be repeated here.

[0041] The foregoing description and accompanying drawings fully illustrate embodiments of this disclosure to enable those skilled in the art to practice them. Other embodiments may include structural, logical, electrical, procedural, and other changes. The embodiments represent variations only. Individual components and functions are optional unless explicitly required, and the order of operations may vary. Parts and features of some embodiments may be included in or replace parts and features of other embodiments. Moreover, the terminology used in this application is for describing embodiments only and is not intended to limit the claims. As used in the description of embodiments and claims, the singular forms “a,” “an,” and “the” are intended to equally include the plural forms unless the context clearly indicates otherwise. Similarly, the term “and / or” as used in this application means including one or more of the associated listed combinations. Additionally, when used in this application, the term "comprise" and its variations "comprises" and / or "comprising" refer to the presence of stated features, integrals, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or groups thereof. Without further limitations, an element defined by the phrase "comprises a..." does not exclude the presence of other identical elements in the process, method, or apparatus that includes said element. In this document, each embodiment may focus on the differences from other embodiments, and similar or identical parts between embodiments can be referred to mutually. For methods, products, etc., disclosed in the embodiments, if they correspond to the method section disclosed in the embodiments, the relevant parts can be referred to the description of the method section.

[0042] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented using electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods for each specific application to achieve the described functions, but such implementation should not be considered beyond the scope of the embodiments of this disclosure. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the described devices, apparatuses, and units can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0043] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, function, and operation of implementations of apparatus, methods, and computer program products according to embodiments of the present disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions indicated in the blocks may occur in a different order than indicated in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. In the descriptions corresponding to the flowcharts and block diagrams in the accompanying drawings, the operations or steps corresponding to different blocks may also occur in a different order than disclosed in the description; sometimes there is no specific order between different operations or steps. For example, two consecutive operations or steps may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. Each block in a block diagram and / or flowchart, and combinations of blocks in a block diagram and / or flowchart, can be implemented using a dedicated hardware-based device that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

Claims

1. A personalized education method based on a large model, characterized in that, Includes the following steps: S1. Obtain the interactive dataset of the current education stage, and preprocess the interactive dataset to obtain the cognitive index and interest index of the learners; The execution of S1 specifically includes the following steps: A large model is used to perform multimodal feature analysis on the interactive dataset of the current education stage, extracting the temporal dependency features of knowledge acquisition trajectory, the topological path features of logical reasoning, and the attention mask features of interactive behavior. The temporal dependency features and the topological path features are input into a multidimensional feature weighting network, the confidence weights are fused and normalized mapping is performed to calculate the cognitive index. The attention mask features are input into a time window decay model, and the interest index is calculated by combining the distribution parameters of dwell time and the polarity score of emotional feedback. S2. Based on the cognitive index and the interest index, generate multiple sets of knowledge unit sequences from a preset knowledge base, and each knowledge unit sequence is mapped to a different interest index; Calculate the push weight of each knowledge unit sequence, and sort the knowledge unit sequences based on the push weight to obtain an initial sequence set; S3. Calculate the transition probability of the knowledge units in the initial sequence set. The transition probability includes the in-sequence probability and the out-of-sequence probability. The in-sequence probability is the content transfer rate within the knowledge unit sequence determined by the current interest index. The out-of-sequence probability is the content transfer rate between different knowledge unit sequences determined by different interest indices. S4. Execute the iteration under the preset cognitive index constraint, and at the end of each stage of the iteration, input the interaction feedback data of the current stage into the large model, update the cognitive index and the interest index, and return to execute S2-S3 to obtain the stage sequence set. S5. Calculate the transition probability of the phase sequence set and determine whether the sum of the in-sequence probability and the out-of-sequence probability is 1 under the preset conditions; If it is not 1, return to S4 and execute the iteration; If the value is 1, output the current stage sequence set as the target education sequence set for the next education stage.

2. The personalized education method based on a large model according to claim 1, characterized in that, When performing S3, the following steps are also included: Based on the initial cognitive index and the interest index, a benchmark model for the construction phase is built, and the iterative process is divided into the cognitive construction phase, the structural reorganization phase, and the cross-domain migration phase. Assign a baseline weighting range to the in-sequence probability and the out-of-sequence probability for each period, and set a cognitive expectation growth curve and an interest dynamic decay baseline; When switching between iteration phases, extract the actual interaction feedback data of the current phase sequence set, and calculate the deviation of the real-time cognitive index and the interest index from the expected cognitive growth curve and the dynamic interest decay baseline. If the deviation falls within the preset deviation range, the current benchmark weight range is maintained to execute the next stage; If the deviation exceeds the preset deviation tolerance range, and the cognitive index lags, the weight of the in-sequence probability is increased and the frequency reduction factor of the out-of-sequence probability is injected to force convergence of the in-sequence path; When the interest index is lower than the interest dynamic decay baseline, the cross-domain transition weight of the out-of-sequence probability is activated, and the sequence is redirected to a high-matching knowledge unit sequence based on the volatility of the interest index. The redistribution parameters and deviation compensation coefficients are input into the large model to perform nonlinear mapping, generating the cognitive index and interest index for the next iteration, until the preset conditions are met.

3. The personalized education method based on a large model according to claim 2, characterized in that, The preset conditions include iteration termination conditions and probability convergence verification conditions; The iteration termination condition is that the current iteration round number reaches a preset maximum iteration threshold, or the cumulative rate of change of the in-sequence probability and the out-of-sequence probability in two adjacent iterations is lower than a preset convergence deviation threshold. The probability convergence verification condition is that the updated cognitive index reaches the preset knowledge mastery threshold, and the fluctuation variance of the interest index falls into the preset stable range. When both the iteration termination condition and the probability convergence verification condition are met, the stage sequence set is determined to converge into the target education sequence set. Normalization distribution verification is performed on the intra-sequence probability, cross-domain correlation compensation is performed on the extra-sequence probability, the weighted fusion value of the two types of probabilities is forcibly calibrated to 1 through the probability accumulation convergence function, and the target education sequence set is locked and output.

4. The personalized education method based on a large model according to claim 3, characterized in that, During the iteration process, a self-test is also performed, which includes the following steps: The probability convergence trajectory of the sequence sets at each stage is compared in real time. If the probability convergence trajectories are consistent, the iteration continues. If a branch deviation occurs, the first self-test is performed, extracting a set of benchmark sequences with consistent sequence length and probability evolution characteristics from the historical database, constructing a test set and assigning different cognitive load weights and interest exploration weights, and outputting a joint index of calculated path completion and cognitive achievement rate. If the joint index does not fall within the joint confidence interval, a second self-detection is performed to mark the abnormal probability nodes and calculate their perturbation coefficients on the overall probability accumulation. When the perturbation coefficient exceeds the preset blocking threshold, the sequence is reset. The knowledge unit sequence is adjusted according to the cognitive index deviation. The topological weight of the out-of-sequence probability jump path is reset according to the interest index decay coefficient. Then, the execution returns to S2-S4.

5. A personalized education system based on a large model, characterized in that, For implementing the personalized education method as described in any one of claims 1 to 4, the large-model-based personalized education system comprises: The data preprocessing module is configured to: acquire the interactive dataset of the current educational stage, perform multimodal feature parsing and decay modeling, and calculate the cognitive index and interest index; The sequence generation and sorting module is configured to: generate multiple sets of knowledge unit sequences from a preset knowledge base and map them to different interest indices based on the cognitive index and the interest index; calculate the push weight of each knowledge unit sequence; and sort each knowledge unit sequence based on the push weight to obtain an initial sequence set. The probability calculation and iteration module is configured to: calculate the in-sequence probability and out-of-sequence probability of knowledge units in the initial sequence set, perform iteration under the constraint of a preset cognitive index, and input the interaction feedback data of the current stage to the large model at the end of each stage of the iteration, update the cognitive index and the interest index, and repeatedly perform sequence generation and probability calculation to obtain the stage sequence set; The convergence determination and output module is configured to: calculate the transition probability of the stage sequence set, determine whether the sum of the in-sequence probability and the out-of-sequence probability is 1 under preset conditions; if it is not 1, return to iteration; if it is 1, lock the path and output the current stage sequence set as the target education sequence set to the next education stage. The self-detection and correction module is configured to perform probability trajectory comparison, joint indicator verification and abnormal node disturbance compensation, and trigger sequence reset and probability redistribution when iterative deviation occurs.

Images