Learning path intelligent planning and adjustment method based on personalized assisted teaching

Abstract

The present application belongs to the technical field of intelligent education, and relates to a learning path intelligent planning and adjusting method based on personalized auxiliary teaching. The present application calculates a mastery probability vector and a cognitive load scalar by extracting a stem text tree and an interaction log; filters weak nodes in a knowledge graph mapping the vector, combines pre-node confidence and problem solving step edit distance to mark problem types, and constructs a causal adjacency matrix; accordingly maps a Markov decision process, strategy iteration outputs an optimal action sequence, and differentially dynamically adjusts difficulty or inserts pre-nodes according to accuracy; and uses Bayesian updating to correct the probability vector and adjust the sampling weight and resource allocation. The present application solves the problems of inaccurate state quantization, insufficient matching accuracy, path adjustment lag and limited resource scheduling caused by relying on single data, achieves the effects of accurate state perception, problem cause positioning, real-time dynamic path adjustment and key node resource tilt configuration, and improves teaching adaptability.

Description

Technical Field

[0001] This invention belongs to the field of intelligent education technology and relates to a method for intelligent planning and adjustment of learning paths based on personalized assisted teaching. Background Technology

[0002] With the development of artificial intelligence and information technology in teaching, online education platforms and smart learning terminals are increasingly being used to support teaching scenarios. Traditional teaching methods typically follow uniform content and a fixed learning sequence, making it difficult to adapt to individual differences among learners. To improve teaching adaptability, existing technologies are increasingly using methods such as collecting learning behavior data and analyzing learning process performance to assess learners' status and adjust learning content based on the assessment results.

[0003] Current personalized learning systems typically generate recommended learning plans based on information such as learning records, responses, and learning duration. In practical applications, learning outcomes are influenced by multiple factors, including knowledge base, learning process, and problem-solving methods; the same learning performance may correspond to different reasons. For example, a learner's incorrect answer may be related to their mastery of relevant foundational content, understanding of knowledge connections, or differences in the problem-solving process. When learning path adjustments are based solely on a single outcome, insufficient accuracy in matching learning content can easily occur, affecting the adaptability of the personalized learning plan.

[0004] Furthermore, the learning state during the learning process changes dynamically with the changes in learning content, and a coordinated relationship needs to be maintained between the difficulty of the learning content, the learning pace, and the task arrangement. Existing learning path planning methods are prone to path adjustment lags when facing continuous learning processes, resulting in a discrepancy between recommended content and the learner's current learning state, which affects the continuous optimization of the learning process.

[0005] Furthermore, within a complex knowledge system, there are certain connections between different learning contents. Problems encountered by learners at a certain learning stage may be influenced by related prior learning content. If the granularity of the correlation analysis is insufficient, it can easily affect the identification of problems and the effectiveness of subsequent learning content arrangement, leading to a deviation between the allocation of learning resources and learning needs. Summary of the Invention

[0006] In view of this, in order to solve the problems mentioned in the background technology, a learning path intelligent planning and adjustment method based on personalized auxiliary teaching is proposed.

[0007] The objective of this invention can be achieved through the following technical solution: a learning path intelligent planning and adjustment method based on personalized assisted teaching, including: S1, extracting the question stem text tree and option topology from the image uploaded from the terminal, parsing the terminal interaction log to obtain the answer result sequence and page dwell frequency, and calculating the mastery probability vector and cognitive load scalar.

[0008] S2. Map the mastery probability vector in the directed acyclic graph of knowledge, calculate the node confidence to screen weak nodes by graph walk, backtrack the distance between the product of the confidence of the previous nodes and the problem-solving logic steps, mark knowledge omissions or logical deviations, and construct a causal adjacency matrix.

[0009] S3. Extract the node sequence by topological sorting of the causal adjacency matrix, map the state space of the Markov decision process according to the cognitive load scalar, iterate the strategy to solve and output the optimal action sequence, and after executing the optimal action, analyze the answer result to calculate the accuracy difference. If the accuracy difference is greater than the difference threshold, the difficulty coefficient is increased. If the accuracy does not improve, the preceding node is inserted.

[0010] S4. Use Bayesian update to correct the mastery probability vector. If the corrected mastery probability vector is lower than the lower probability threshold, increase the sampling weight of the variant question. If it is higher than the upper probability threshold, set the state to zero, transfer the probability, and iterate the strategy again. Transfer the computing resources to the node with the highest degree centrality.

[0011] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The present invention extracts the question stem text tree and option topology and combines it with the terminal interaction log to analyze the answer result sequence and page dwell frequency to calculate the mastery probability vector and cognitive load scalar. This solves the problem that the existing methods rely on single data feedback and are difficult to accurately quantify the user's learning status. It realizes the accurate perception of the user's knowledge mastery and current cognitive pressure, and provides a reliable data foundation for subsequent personalized path planning.

[0012] (2) This invention uses a graph walk to filter weak nodes by mapping the mastery probability vector to a directed acyclic graph of knowledge, and marks knowledge omissions or logical deviations by backtracking the confidence product of the preceding nodes and the edit distance of the problem-solving logic steps, thus constructing a causal adjacency matrix. This solves the problem that adjusting the learning path based solely on a single result information leads to insufficient matching accuracy and difficulty in distinguishing the causes of problems. It enables the location and type differentiation of learning problems, improving the matching accuracy between learning content and users' actual needs.

[0013] (3) This invention extracts the node sequence by sorting the causal adjacency matrix topology and solves the optimal action sequence through strategy iteration. After execution, the difficulty coefficient is dynamically increased or a preceding node is inserted based on the accuracy difference. This solves the problems of existing path planning methods lagging in adjustment and deviation between recommended content and the current learning state when facing continuous learning processes. It realizes real-time dynamic adjustment of the learning path and ensures the coordination between learning difficulty, pace and task arrangement.

[0014] (4) This invention employs Bayesian updating to correct the mastery probability vector, increases the sampling weight of variant questions based on probability thresholds or resets the state transition probability to zero and iterates the strategy again, and transfers computational resources to the node with the highest degree centrality. This solves the problems of insufficient granularity of association analysis, limited problem location, and limited resource allocation effectiveness in complex knowledge systems. It achieves resource-biased allocation to key nodes of the knowledge network, improving the accuracy of problem location and the overall scheduling efficiency of learning resources. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart of the intelligent planning and adjustment method for learning paths based on personalized assisted teaching in this invention;

[0017] Figure 2 This is a graph showing the difference in accuracy and the change in difficulty coefficient during the adjustment of the learning path in this invention.

[0018] Figure 3 This is a graph showing the update and change of the probability of mastering knowledge nodes during the learning process of this invention. Detailed Implementation

[0019] 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.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0021] The specific scheme of the intelligent planning and adjustment method for learning paths based on personalized assisted teaching provided by the present invention will be described in detail below with reference to the accompanying drawings.

[0022] Please see Figure 1 As shown, the implementation of the present invention includes S1 to S4: The present invention addresses the information correlation needs between learners' knowledge mastery status changes, learning resource matching, and path adjustment in the process of personalized assisted teaching, and proposes a learning path intelligent planning and adjustment method based on personalized assisted teaching.

[0023] This method extracts the question stem text tree and option topology, combines terminal interaction logs to obtain the sequence of answer results and page dwell frequency, calculates the mastery probability vector and cognitive load scalar, and realizes learning state perception. It further maps the mastery probability vector to a directed acyclic graph of knowledge, filters weak nodes, and identifies question types by combining the association relationship of preceding nodes and the edit distance of the problem-solving logic steps, constructing a causal adjacency matrix. Based on the causal adjacency matrix, it extracts the knowledge node sequence, constructs a Markov decision process based on the cognitive load state, and generates a learning path adjustment scheme through strategy iteration. Simultaneously, it uses answer feedback and Bayesian update mechanisms to correct the mastery probability vector, dynamically adjusting question sampling, state transition, and resource allocation strategies to form a continuous optimization closed loop for the learning path.

[0024] S1. Extract the question stem text tree and option topology from the image uploaded from the terminal, parse the terminal interaction log to obtain the answer result sequence and page dwell frequency, and calculate the mastery probability vector and cognitive load scalar.

[0025] Considering that personalized learning path planning requires real-time acquisition of user learning behavior characteristics and knowledge mastery status, existing learning analysis methods typically rely on manual input or single-type data feedback, making it difficult to address the data heterogeneity issues between question image information, user interaction behavior, and learning status data in a terminal environment. Furthermore, the question stem content, option structure, and knowledge point relationships in images uploaded to the terminal are usually in unstructured form, and user clicks, page dwell times, and other behavioral information are scattered across different interaction logs, making it difficult to accurately quantify the learning process status and affecting the reliability of subsequent knowledge deficit identification and path adjustment.

[0026] Therefore, this step involves parsing the text structure of the uploaded images and extracting the answering process information from the terminal interaction logs. This process constructs a question stem text tree, option topology, and answer data model, and further calculates the mastery probability vector and cognitive load scalar, providing a data foundation for subsequent learning state analysis and optimization decisions.

[0027] In one specific embodiment, firstly, the terminal receives the question image uploaded by the user, and directly uses an existing pre-trained text detection model such as the DBNet model to perform character region detection on the image, outputting the bounding box coordinates of each character region. Then, each bounding box region is cropped and sent to a convolutional recurrent neural network recognizer to perform text line recognition, and the initial text sequence is spliced ​​out in the reading order from top to bottom and from left to right.

[0028] Secondly, dependency parsing techniques such as Stanford Parser are used to analyze the initial text sequence, extracting the core components with the predicate as the core and the dependency subtrees with modifiers such as attributives and adverbs as modifiers. The question stem text tree is constructed with the predicate node as the root node and the modifiers as child nodes. At the same time, the text content of each option is extracted by matching the option identifiers "A", "B", "C", and "D" using regular expressions. The option topology is constructed with the option identifiers as nodes and the logical mutual exclusion relationships between options as edges.

[0029] At the same time, the system analyzes the terminal interaction logs, extracts the spatial coordinates and timestamps of user click events, performs point inclusion determination on the click coordinates and the bounding boxes of each option area in the aforementioned option topology, merges click events falling into the same option area into selection operations for that option, and generates a sequence of answer results in ascending order of timestamps, where each element records the option identifier and the correctness mark of the answer.

[0030] In addition, the entry timestamp and exit timestamp of each page are extracted from the interaction logs, and the difference between the two is calculated to obtain the single page dwell time. The number of loading events of the same page is counted as the loading frequency. The dwell time and loading frequency are jointly defined as the page dwell frequency.

[0031] Furthermore, when calculating the mastery probability vector and cognitive load scalar, the knowledge point tags in the question stem text tree are aligned with the sequence of answer results. The aligned correct answer sequence is extracted and input together with the question difficulty parameter into the three-parameter logistic item response theory model to calculate the mastery probability vector.

[0032] The specific process is as follows: The aligned sequence of correct answers is used as the observation variable, and the question difficulty parameter is... Discrimination parameters With guessed parameters As a known quantity, the capability parameter to be estimated is introduced. The parameters are the inherent attribute parameters of each question in the question bank that are pre-labeled, and are directly read and called from the question bank during calculation.

[0033] Construct the probability function for the correct answer: .

[0034] in, To guess the parameters, This represents the upper limit of the probability that a subject answers based on their actual ability after excluding guesswork factors; This is the logistic exponent, reflecting the difference between the ability parameter and the difficulty parameter; This represents the probability of answering a question correctly based on actual ability. This represents the total probability of answering a question correctly after comprehensively considering both guessing factors and actual ability.

[0035] For an observation sequence formed by multiple responses, the likelihood function is obtained by multiplying the probabilities of each observation: .

[0036] in, Indicates the index number of the responding sample in the observation sequence; Indicates the first The observation results of each response (1 for a correct answer and 0 for an incorrect answer); This represents the total number of responding samples in the observation sequence.

[0037] Taking the logarithm of the likelihood function yields the log-likelihood function: .

[0038] Based on capability parameters Find the partial derivatives for the variables and solve for the stationary points by setting the partial derivatives to zero. If no explicit solution exists for the stationary points, update the capability parameters incrementally along the gradient direction of the log-likelihood function with a step size of 0.01. The process continues until the change in capability parameters between two consecutive iterations is less than a convergence threshold. This convergence threshold is set according to the algorithm's convergence requirements. The capability parameters obtained from the final iteration are then... Substituting into the aforementioned probability function, the probability of a user mastering the knowledge point is calculated; arranging the calculated mastery probabilities of each knowledge point according to the knowledge point dimension constitutes the mastery probability vector.

[0039] Finally, the difference between the current page dwell frequency and the average historical page dwell frequency is calculated. The difference is then input into the normalized exponential function softmax, and the output value ranges from [0, 1] to reflect the cognitive load scalar, which reflects the current cognitive stress.

[0040] S2. Map the mastery probability vector in the directed acyclic graph of knowledge, calculate the node confidence to screen weak nodes by graph walk, backtrack the distance between the product of the confidence of the previous nodes and the problem-solving logic steps, mark knowledge omissions or logical deviations, and construct a causal adjacency matrix.

[0041] Given that learning knowledge systems typically exhibit hierarchical connections and dependencies, the level of mastery of a single knowledge point cannot accurately reflect a user's weak areas and their causes. Furthermore, existing methods struggle to distinguish between problems stemming from a lack of foundational knowledge and those caused by deviations in the problem-solving process, resulting in a lack of targeted learning resource recommendations. Simultaneously, the complex relationships between knowledge nodes necessitate analyzing the degree of influence between nodes within the context of knowledge propagation paths to avoid erroneous judgments based solely on partial information.

[0042] Therefore, this step maps the mastery probability vector to a directed acyclic graph of knowledge, uses graph walks to calculate node confidence and filter weak nodes, combines the product of the confidence of the preceding nodes and the edit distance of the problem-solving logic steps to determine the type, and further constructs a causal adjacency matrix.

[0043] In one specific embodiment, firstly, the probability values ​​of each knowledge point dimension in the calculated mastery probability vector are assigned one by one to the nodes corresponding to the knowledge points in the pre-constructed directed acyclic graph of knowledge, as the initial confidence of each node; the directed acyclic graph of knowledge is pre-constructed according to the course knowledge outline, the nodes in the graph represent knowledge points, the directed edges represent the prerequisite dependencies between knowledge points, and the direction of the edges points from the prerequisite knowledge points to the subsequent knowledge points.

[0044] Secondly, starting from the root node with an in-degree of zero in the directed acyclic graph of knowledge, each node is traversed sequentially according to topological sorting. When traversing to each node, the arithmetic mean of the updated confidence of all its parent nodes is taken, and then this mean is multiplied by the node's own confidence. The product is used as the updated confidence of the node, thus simulating the cumulative dependency relationship of knowledge points. Since topological sorting is used for traversal, each node is visited only after all its parent nodes have been updated. After the traversal is completed, all nodes in the directed acyclic graph of knowledge have obtained a definite updated confidence.

[0045] Next, the arithmetic mean of the updated confidence scores of all nodes in the directed acyclic graph is calculated as the confidence mean. Nodes whose updated confidence scores are lower than this confidence mean are marked as weak nodes, forming a set of weak nodes.

[0046] Therefore, after forming the set of weak nodes, it is necessary to further distinguish the causes of defects in each weak node and clarify the propagation and influence relationships between weak nodes, so as to provide a structured basis for subsequent learning path adjustments. To this end, for each weak node in the set of weak nodes, all predecessor nodes are extracted by traversing backwards along its incoming edge direction, and the product of the updated confidence scores of each predecessor node is calculated to obtain the product of the confidence scores of the predecessor nodes.

[0047] Simultaneously, the standard problem-solving logic step sequence, pre-annotated manually, is extracted from the question bank for the knowledge point corresponding to the weak node. The user's click trajectory on the question for that knowledge point is replayed from the terminal interaction log. The actual problem-solving logic step sequence is extracted in the order of click timestamps. A dynamic programming algorithm is used to calculate the minimum number of editing operations between the standard problem-solving logic step sequence and the actual problem-solving logic step sequence. The editing operations include inserting a step, deleting a step, and replacing a step. The cost of each operation is set to 1. The calculated minimum number of editing operations is used as the problem-solving logic step edit distance.

[0048] Furthermore, the confidence product of the preceding nodes is compared with the first decision boundary, which is the difference between the mean and standard deviation of the confidence product of the preceding nodes in the historical learning data. If the confidence product of the preceding nodes is less than the first decision boundary, it indicates that the preceding knowledge support is insufficient, and the weak node is marked as a knowledge omission type.

[0049] If the product of the confidence scores of the preceding nodes is greater than or equal to the first decision threshold, the edit distance of the problem-solving logic steps is further compared with the second decision threshold, which is the sum of the mean and standard deviation of the edit distance of the problem-solving logic steps of the historical answer samples of the same question type. If the edit distance of the problem-solving logic steps exceeds the second decision threshold, it indicates that there is a logical deviation in the problem-solving process, and it is marked as a logical deviation type.

[0050] Finally, extract the complete paths from all marked weak nodes in the directed acyclic graph of knowledge, traversing backwards along the incoming edges to the root node. Merge the nodes contained in each path and remove duplicate nodes. Sort the deduplicated node set according to the topological sorting rules of the directed edges to obtain a node sequence. Use the nodes in this node sequence as the row and column indices of the matrix in sequence. If there is a directed edge between a row node and a column node, set the corresponding matrix element to 1; otherwise, set it to 0. Construct a causal adjacency matrix, which serves as the input for the subsequent construction of the state space of the Markov decision process.

[0051] S3. Extract the node sequence by topological sorting of the causal adjacency matrix, map the state space of the Markov decision process according to the cognitive load scalar, iterate the strategy to solve and output the optimal action sequence, and after executing the optimal action, analyze the answer result to calculate the accuracy difference. If the accuracy difference is greater than the difference threshold, the difficulty coefficient is increased. If the accuracy does not improve, the preceding node is inserted.

[0052] Because different users have dynamic changes in their knowledge level, learning progress, and cognitive load, a fixed-order approach to pushing learning resources is difficult to meet the needs of individual differences. Furthermore, the lack of a mechanism to adjust the learning difficulty and knowledge supplementation path based on real-time feedback during the learning process can easily lead to a mismatch between the learning content and the user's actual ability.

[0053] Therefore, this step extracts node sequences that conform to knowledge dependencies based on the causal adjacency matrix, constructs the state space of the Markov decision process by combining the cognitive load scalar, obtains the optimal action sequence through policy iteration, calculates the change in accuracy based on the answer results after execution, and dynamically adjusts the difficulty coefficient or the preceding nodes to provide feedback for continuous optimization of the learning path.

[0054] In one specific embodiment, firstly, a node sequence is extracted according to the topological sorting of the causal adjacency matrix. For each node in the node sequence, the number of words in the learning resource corresponding to that node is read from the question bank resource table. Multimedia complexity level and the number of related knowledge points The theoretical load value of this node is calculated using the following formula: .

[0055] in, , , The preset weighting coefficients correspond to the degree of influence of word count, multimedia complexity, and the number of related knowledge points on the theoretical workload, respectively. Their values ​​are determined empirically or through regression fitting of historical data, and satisfy the following conditions: The multimedia complexity level is assigned a value based on the media format in the resource: 1 for plain text, 2 for images, and 3 for videos. Implementers can set the value according to their actual needs.

[0056] Next, the cognitive load scalar is added to the theoretical load value of the current node to obtain the total load value. This total load value is then compared with the maximum recorded value of the cognitive load scalar during the user's historical learning period. If the total load value is greater than the maximum recorded value, all nodes following the current node in the topological sort are removed from the node sequence, the current node is retained, and the current node and the remaining nodes are arranged in their original order to form the initial sequence. If the total load value is less than or equal to the maximum recorded value, the current node is retained, and the same comparison is performed on the next node until all nodes in the node sequence have been traversed to generate the initial sequence.

[0057] Next, each node in the initial sequence is defined as a state node of a Markov decision process, forming a set of states. ,in, The total number of nodes in the initial sequence; for each state node The system retrieves all question resources under the corresponding knowledge point from the question bank, encodes the storage address of each question resource as a resource pointer, and uses the difficulty coefficient of the question as an action parameter. The combination of the resource pointer and the difficulty coefficient is defined as the action set for this state, as follows: ,in, For the first The resource pointer for the question. No. The difficulty level of the question is set to 1 initially. This represents the total number of questions under this knowledge point.

[0058] For the state transition probability, extract the state nodes from the causal adjacency matrix. For each column index in the row with a value of 1, determine the subsequent state corresponding to that column index. As The transitional state will The transition probabilities to each successor state are initialized as follows: ,in, for The total number of successor states is used to construct the state space of the Markov decision process.

[0059] Then, policy iteration is performed in this state space to solve for the optimal action sequence. The specific process is as follows: First, the value functions of all states are initialized to zero. Then, the current policy is used to evaluate the policy for each state, i.e., according to the Bellman equation. Iteratively update the value function of each state.

[0060] in, This is the current state node; Current state Execute action The subsequent state node that may be transferred to; The action selected from the action set in the current state; To start from the current state Execute action Transition to successor state The probability, This is an immediate reward function defined as taking 1 for a correct answer and 0 for an incorrect answer. The discount factor is set to 0.9; For the successor state The value function. This equation represents the current state. The value is equal to the value of all possible successor states. The sum of the immediate reward and the discounted subsequent value, weighted by the transfer probability.

[0061] After each policy evaluation, a policy improvement is performed, selecting the action that maximizes the action value function for each state as the new policy. After each policy improvement, the value functions of all states in the state space are scanned, and the maximum change in value for all states between the two scans is calculated.

[0062] The maximum value is compared with the stopping threshold, which is set based on the floating-point precision of the computing hardware. If the maximum value is less than the stopping threshold, it indicates that the policy value function has stabilized, and the current policy is determined to have reached convergence, outputting the optimal action sequence including resource pointers and difficulty coefficients. If the maximum value is not less than the stopping threshold, it indicates that the policy has not yet converged, and the improved policy is used to continue the next round of policy evaluation and policy improvement.

[0063] Furthermore, after executing the actions in the optimal action sequence, the system parses the most recent N responses from the terminal interaction log, calculates the accuracy rate of the most recent N responses as the post-execution accuracy rate, and extracts the accuracy rate of the most recent N responses from the node's history as the pre-execution accuracy rate. The difference between the post-execution accuracy rate and the pre-execution accuracy rate is used as the accuracy rate difference. In this embodiment, N is set to a positive integer not less than 5 to eliminate the randomness of a single response; the implementer can set this value as needed.

[0064] like Figure 2 As shown, this invention calculates the accuracy difference based on the feedback results after performing the learning action, adjusts the subsequent learning difficulty parameters according to the accuracy change trend, and optimizes the learning path by combining a preceding node insertion mechanism. In the figure, the horizontal axis represents the learning adjustment rounds, and the vertical axis represents the accuracy change percentage. Positive changes indicate that the current learning resource matching is effective, while negative changes indicate that preceding knowledge nodes need to be added or the learning difficulty adjusted. Dynamic optimization of the learning path is achieved through continuous feedback.

[0065] When the accuracy difference is greater than the difference threshold, which is the sum of the mean and standard deviation of the consecutive accuracy differences for the same knowledge point in the historical learning data, it indicates that the user has a good grasp of the current learning content and has room for improvement. Then, the difficulty coefficient in the subsequent optimal action sequence will be increased by a step size of 0.1 based on the difficulty gradient smoothness requirement.

[0066] Finally, the historical page dwell frequency is extracted from the interaction logs and its average is calculated as the average page dwell frequency. When the page dwell frequency is greater than a multiple of the average page dwell frequency, such as 1.5 times, and the accuracy difference is less than or equal to the difference threshold, it indicates that the user has invested a lot of time in the current node but has not achieved an improvement in accuracy and there is a lack of prior knowledge. In this case, the direct predecessor node of the current node is extracted from the causal adjacency matrix as the predecessor node, and the predecessor node is inserted in front of the current node position in the initial node sequence. The updated initial node sequence and the corrected difficulty coefficient are used as the input for subsequent Bayesian updates.

[0067] S4. Use Bayesian update to correct the mastery probability vector. If the corrected mastery probability vector is lower than the lower probability threshold, increase the sampling weight of the variant question. If it is higher than the upper probability threshold, set the state to zero, transfer the probability, and iterate the strategy again. Transfer the computing resources to the node with the highest degree centrality.

[0068] Considering that a user's knowledge mastery status evolves continuously with ongoing learning, practice exercises, and feedback, the mastery probability obtained from the initial assessment may not accurately reflect the user's current cognitive level in the long term. Furthermore, during the learning path adjustment process, it is necessary to avoid unreasonable resource allocation due to lag in state estimation. In addition, different nodes in the knowledge network have varying degrees of importance; if computational resources cannot be dynamically adjusted based on the degree of node correlation, the effectiveness of the path optimization process may be reduced.

[0069] Therefore, this step uses a Bayesian update mechanism to correct the mastery probability vector based on the latest answer results, and adjusts the sampling weight of variant questions or re-executes the strategy iteration according to the probability threshold. At the same time, it calculates the degree centrality of nodes and allocates computing resources to support the dynamic updating of the learning path and the optimization of resource allocation.

[0070] In one specific embodiment, firstly, the correctness marker of the latest answer is extracted from the sequence of answer results and used as observation data. The probability value corresponding to the current knowledge point in the mastery probability vector is taken as the prior probability, and a Bayesian update is performed to obtain the corrected mastery probability vector. The specific process is as follows: The prior probability... Through linear mapping Converted into capability parameters ,in, and Let be the linear transformation parameter. This mapping is used to linearly transform the probability interval [0, 1] to a commonly used range of values ​​for the capability parameter, such as [-3, 3]. and The value of should satisfy: when hour Take the lower limit of the interval, when hour Take the upper limit of the interval.

[0071] Simultaneously set , representing the capability parameter in the unmastered state, which is related to about Symmetry is used to maintain the relative positions of unmastered and mastered states within the capability space.

[0072] Will Substituting into the aforementioned probability function, the likelihood probability is calculated as follows: The corrected probability calculated using Bayes' theorem is: ,Will Replace the probability value corresponding to the current knowledge point in the mastery probability vector to obtain the corrected mastery probability vector.

[0073] like Figure 3As shown, this invention uses a Bayesian update mechanism to dynamically adjust the learner's mastery probability of each knowledge node. The horizontal axis represents the number of learning interaction rounds, and the vertical axis represents the mastery probability of each knowledge node. As response feedback data is continuously input, the mastery probability corresponding to each knowledge node is adjusted according to the latest observation results. Nodes with low mastery probability gradually increase after continuous training, while nodes with high mastery probability maintain a stable trend, thus forming a dynamic state update oriented towards the learning process.

[0074] Secondly, the probability value of the current knowledge point in the corrected mastery probability vector is compared with the lower probability threshold of 0.4. This lower probability threshold is a basic mastery threshold set based on teaching practice experience, used to define whether the user has a basic understanding of the knowledge point. If the corrected mastery probability vector is lower than the lower probability threshold, it indicates that the user has not reached the basic mastery level. Then, all variant questions with the same knowledge point tag as the current node are retrieved from the question bank, and the probability weights of each variant question in the sampling distribution are adjusted from the original weights. Increment to This increases the sampling probability for targeted training. The variant questions include basic question types corresponding to the knowledge points of the current node.

[0075] Next, the probability value of the current knowledge point in the corrected mastery probability vector is compared with the upper probability threshold. This upper probability threshold is a critical value for mastery set based on teaching practice experience, used to define whether the user has achieved complete mastery of the knowledge point. Then, in the Markov decision process state space, the current state node is marked as a mastered state, the node is removed from the set of current transitional target states, and all transition probabilities starting or ending at this state are simultaneously set to zero, thus achieving logical masking of this state node.

[0076] After the masking is completed, based on the reduced state space, the policy evaluation and policy improvement are re-executed according to the policy iteration process described in step S3. Since the masked node has been removed from the state space, the policy iteration is carried out within a finite state space, ensuring the convergence of the iteration process. After the policy iteration is completed again, the updated optimal action sequence is output.

[0077] As a further explanation of the aforementioned logical masking mechanism, masked, mastered state nodes will no longer appear as target nodes in subsequent learning path planning. The system will automatically skip these knowledge points and concentrate learning and computing resources on weak nodes that have not yet been mastered. The state masking operation only affects the set of transferable targets in the current learning path and does not change the basic structure of the knowledge system represented by the causal adjacency matrix, thus ensuring that knowledge dependencies can still be invoked in subsequent learning rounds.

[0078] Meanwhile, the sum of the number of incoming edges and outgoing edges of each node in the causal adjacency matrix is ​​used as the degree centrality. All nodes are sorted in descending order of degree centrality, and the computing power quota of the system's recommendation algorithm is allocated to each node proportionally. The node with the highest degree centrality is allocated 40% of the total computing power quota, and the remaining nodes are allocated the remaining 60% of the computing power quota according to the proportion of their degree centrality to the sum of the degree centralities of all nodes. This optimizes the response speed of key knowledge points and the overall processing efficiency of the system.

[0079] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0080] Those skilled in the art will recognize that the algorithmic steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented 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 to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0081] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0082] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0083] Finally, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A learning path intelligent planning and adjustment method based on personalized assisted teaching, characterized in that, include: S1. Extract the question stem text tree and option topology from the image uploaded from the terminal, parse the terminal interaction log to obtain the answer result sequence and page dwell frequency, and calculate the mastery probability vector and cognitive load scalar. S2. Map the mastery probability vector in the directed acyclic graph of knowledge, calculate the node confidence to screen weak nodes by graph walk, backtrack the distance between the product of the confidence of the previous nodes and the problem-solving logic steps, mark knowledge omissions or logical deviations, and construct a causal adjacency matrix. S3. Extract the node sequence by topological sorting of the causal adjacency matrix, map the state space of the Markov decision process according to the cognitive load scalar, solve the optimal action sequence by strategy iteration, analyze the answer result after executing the optimal action, calculate the accuracy difference, if the accuracy difference is greater than the difference threshold, increase the difficulty coefficient, if the accuracy does not improve, insert the previous node. S4. Use Bayesian update to correct the mastery probability vector. If the corrected mastery probability vector is lower than the lower probability threshold, increase the sampling weight of the variant question. If it is higher than the upper probability threshold, set the state to zero, transfer the probability, and iterate the strategy again. Transfer the computing resources to the node with the highest degree centrality.

2. The intelligent planning and adjustment method for learning paths based on personalized assisted teaching as described in claim 1, characterized in that, The process of extracting the question stem text tree and option topology in S1, and parsing the terminal interaction log to obtain the answer result sequence and page dwell frequency, specifically includes: Perform character region detection and text line recognition on the image uploaded by the terminal to obtain the initial text sequence; Dependency parsing is performed on the initial text sequence to extract the main components and modifiers to construct the question stem text tree, and option identifiers and option content are extracted to construct the option topology. User click events are extracted from terminal interaction logs, the spatial coordinates of user click events are matched with the option areas in the option topology, the target option identifier is identified, and the answer result sequence is generated by sorting by timestamp; The entry and exit timestamps of each page are extracted from the terminal interaction logs. The difference is calculated to obtain the page dwell time. The number of loading events for each page is counted to obtain the loading frequency. The dwell time and loading frequency are combined to form the page dwell frequency.

3. The intelligent planning and adjustment method for learning paths based on personalized assisted teaching as described in claim 1, characterized in that, The calculation of the mastery probability vector and cognitive load scalar in S1 specifically includes: Align the knowledge point tags in the question text tree with the sequence of answer results, input the aligned correct answer sequence and question difficulty parameters into the project response theory model, and output the mastery probability vector; Calculate the difference between the current page dwell frequency and the average historical page dwell frequency, input the difference into the normalized exponential function, and output the cognitive load scalar.

4. The intelligent planning and adjustment method for learning paths based on personalized assisted teaching as described in claim 3, characterized in that, The calculation process of the project's reaction theory model includes: The aligned sequence of correct answers is used as the observed variable, the question difficulty parameter is used as the known quantity, and the ability parameter to be estimated is introduced to construct the probability function of correct answer; For an observation sequence formed by multiple responses, the product of the probabilities of each observation is calculated to obtain the likelihood function; Take the logarithm of the likelihood function, differentiate it with the capability parameter as the variable, and solve for the stationary point by setting the derivative to zero; If there is no explicit solution for the stationary point, the capability parameter is updated step by step along the gradient direction of the log-likelihood function until the change in capability parameter between two iterations is less than the convergence threshold. The capability parameters obtained from the final iteration are mapped to the dimensions of each knowledge point to form a mastery probability vector.

5. The intelligent planning and adjustment method for learning paths based on personalized assisted teaching as described in claim 1, characterized in that, The steps for screening weak nodes include: The probability values ​​of the master probability vector are used as the initial confidence of the corresponding nodes in the directed acyclic graph of knowledge. Starting from the root node, the nodes are traversed in the topological sorting order. The arithmetic mean of the confidence of all parent nodes of each node is multiplied by the node's own confidence to update the confidence of the child nodes. Calculate the average confidence of all nodes in the directed acyclic graph of knowledge, and mark nodes whose updated confidence is lower than the average confidence as weak nodes.

6. The intelligent planning and adjustment method for learning paths based on personalized assisted teaching as described in claim 1, characterized in that, In S2, knowledge omissions or logical deviations are marked, and a causal adjacency matrix is ​​constructed, specifically including: Traverse backwards along the incoming edges of weak nodes to extract all predecessor nodes, calculate the product of the confidence scores of each predecessor node, and obtain the product of the confidence scores of the predecessor nodes. Extract the standard problem-solving logic step sequence and the actual problem-solving logic step sequence corresponding to the weak node, and calculate the minimum number of editing operations between the two as the problem-solving logic step editing distance; If the product of the confidence scores of the preceding nodes is less than the first decision threshold, the weak node is marked as a knowledge omission type. If the product of the confidence scores of the preceding nodes is greater than or equal to the first decision threshold, and the edit distance of the problem-solving logic steps exceeds the second decision threshold, the weak node is marked as a logic deviation type. Extract the reverse traversal path of the marked weak nodes in the directed acyclic graph of knowledge, perform topological sorting on the nodes along the path, and obtain the node sequence. Using the nodes in the node sequence as rows and columns, the directed edge connections are mapped to matrix elements to form a causal adjacency matrix.

7. The intelligent planning and adjustment method for learning paths based on personalized assisted teaching as described in claim 1, characterized in that, The steps for obtaining the optimal action sequence include: Extract the node sequence by topological sorting of the causal adjacency matrix; Calculate the theoretical load value of the resource pointer corresponding to each node in the node sequence. If the sum of the cognitive load scalar and the theoretical load value is greater than the historical maximum value of the cognitive load scalar, then remove all nodes after the current node from the node sequence, retain the current node, and form the initial sequence with the current node and the remaining nodes in the original order. The nodes in the initial sequence are used as the state nodes of the Markov decision process, and the resource pointer and difficulty coefficient are used as the action set to construct the state space of the Markov decision process. The strategy iteration is performed in the state space of the Markov decision process. When the strategy converges, the optimal action sequence containing resource pointers and difficulty coefficients is output.

8. The intelligent planning and adjustment method for learning paths based on personalized assisted teaching as described in claim 7, characterized in that, The method for determining policy convergence during policy iteration is as follows: After each policy improvement, the value function of all states in the state space is scanned, and the maximum value change of all states between the two scans is calculated. Compare the maximum value with the stop limit; If the maximum value is less than the stopping limit, the current policy is determined to have reached the convergence state, the iteration is stopped and the current policy is output as the optimal policy; If the maximum value is not less than the stopping limit, then the next round of strategy evaluation and strategy improvement will continue with the improved strategy.

9. The intelligent planning and adjustment method for learning paths based on personalized assisted teaching as described in claim 1, characterized in that, The process of executing the optimal action in S3, followed by parsing the answer result and adjusting the learning path, specifically includes: Execute the resource pointers and difficulty coefficients in the optimal action sequence to push the corresponding learning resources and questions, and obtain the results of the most recent answers; The accuracy rate is calculated by analyzing the ratio of the number of correct answers to the total number of answers in the most recent responses, and the difference between the accuracy rate of the last response and the accuracy rate of the previous response is calculated as the accuracy rate difference. When the accuracy difference is greater than the difference threshold, the difficulty coefficient in the subsequent optimal action sequence will be increased by one step. The mean of page dwell frequency is calculated. When the page dwell frequency is greater than a multiple of the mean page dwell frequency and the accuracy difference is less than or equal to the difference threshold, the direct predecessor node of the current node in the causal adjacency matrix is ​​taken as the predecessor node and inserted in front of the current position in the initial node sequence.

10. The intelligent planning and adjustment method for learning paths based on personalized assisted teaching as described in claim 1, characterized in that, The S4 step includes: The latest answer in the sequence of answer results is used as the observation data, and the mastery probability vector is used as the prior distribution. A Bayesian update is performed to obtain the corrected mastery probability vector. If the corrected probability vector is lower than the lower probability threshold, then extract variant questions that are the same as the knowledge points of the current node from the question bank, and increase the probability weight value of the variant questions in the sampling distribution. If the corrected probability vector exceeds the upper probability threshold, then all transition probabilities in the Markov decision process state space that start or end at the current state node are set to zero, and the policy iteration is re-executed. The sum of the number of incoming edges and outgoing edges of each node in the causal adjacency matrix is ​​used as the degree centrality, and computing resources are allocated to the node with the highest degree centrality.

Images