A job cognitive impairment diagnosis and adaptive correction system

Abstract

The application discloses a kind of job cognitive defects diagnosis and adaptive correction system, by constructing dynamic cognitive knowledge graph and combining structure causal model, the precise attribution to the deep knowledge defect behind student's wrong question is realized, and the limitation that traditional method can only carry out error classification is broken through;By introducing counterfactual reasoning mechanism, it can simulate "if certain knowledge point is mastered, how will the result change", so as to scientifically evaluate the expected gain of each intervention point, provide data-driven decision basis for personalized correction.The application significantly improves the interpretability, accuracy and practicability of the educational diagnosis system as a whole, can provide understandable, executable and efficient learning improvement path for each student in complex subject scenarios, effectively promote personalized learning and teaching quality improvement.

Description

Technical Field

[0001] This invention relates to the field of intelligent education technology, specifically to a system for diagnosing and adaptively correcting cognitive deficiencies in homework. Background Technology

[0002] Existing educational diagnostic systems largely rely on analyzing the surface characteristics of students' incorrect answers, with typical methods including semantic similarity comparison and error pattern clustering. These methods identify "where the mistake is" and "which type of error it belongs to" by matching student answers with standard answers textually or structurally, and then summarize common error types at the group level. However, such technologies only describe and categorize error phenomena, failing to delve into the underlying cognitive causes of errors. Especially when faced with complex problem-solving processes, they cannot accurately determine which missing or incorrect pre-existing knowledge point led to the final answer deviation. Furthermore, existing systems often use general strategies or push review materials based on empirical rules when recommending remedial learning content, lacking precise intervention path planning tailored to individual cognitive structure deficiencies, resulting in low remediation efficiency and insufficient personalization. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention aims to provide a system for diagnosing and adaptively correcting operational cognitive deficiencies.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: A system for diagnosing and adaptively correcting cognitive deficits in work includes a dynamic cognitive knowledge graph construction module, a causal diagnosis module, a counterfactual intervention reasoning module, a personalized correction path generation module, and an adaptive practice recommendation module. The dynamic cognitive knowledge graph construction module is used to automatically construct a dynamic knowledge graph containing knowledge point nodes and their interrelationships based on student answer data from subject textbooks, curriculum standards, and history; the interrelationships between the knowledge point nodes include at least prerequisite relationships and logical deduction relationships; The causal diagnosis module is used to identify the key knowledge points that cause the deviation from the correct solution path when a student makes a mistake. This is done by comparing the student's actual solution process for the current incorrect question with the standard solution path. The module also identifies all the prerequisite knowledge points for the key knowledge points based on the dynamic cognitive knowledge graph and constructs a structural causal model accordingly. The counterfactual intervention reasoning module is used to perform counterfactual simulations by combining structural causal models. It calculates the increase in the probability of students answering questions correctly after mastering each knowledge point, and then evaluates the corrective gain of each knowledge point, and determines the optimal intervention priority based on this ranking. The personalized correction path generation module is used to take the knowledge point with the highest correction gain as the target knowledge point, and combine the dependency path of the target knowledge point in the dynamic cognitive knowledge graph with the student's current mastery of each knowledge point to generate an optimal learning sequence to help the student reach the target knowledge point. The adaptive practice recommendation module is used to continuously monitor the changes in a student's mastery of each knowledge point in the optimal learning sequence as the student executes the optimal learning sequence based on the reinforcement learning algorithm, and dynamically recommend the most suitable practice questions accordingly.

[0005] Furthermore, the dynamic cognitive knowledge graph construction module includes a node extraction unit and a dependency mining unit. The node extraction unit is used to extract subject knowledge points from subject textbooks and curriculum standards as graph nodes. The dependency mining unit is used to utilize the step sequences of high-quality student problem-solving processes, and through sequence pattern mining and rule learning methods, automatically identify the prerequisites and logical order between knowledge points, and establish prerequisite relationship edges and logical deduction relationship edges between knowledge points.

[0006] Furthermore, the execution process of the causal diagnosis module is as follows: The standard solution path for the current incorrect question is represented as a sequence of atomic operations of standard knowledge points applied in sequence, and the student's solution process for the current incorrect question is decomposed into a sequence of atomic operations of knowledge points of incorrect question; by comparing the atomic operation sequence of standard knowledge points and the atomic operation sequence of knowledge points of incorrect question, the key knowledge points that cause the atomic operation sequence of knowledge points of incorrect question to deviate from the atomic operation sequence of standard knowledge points are identified. Based on the dynamic cognitive knowledge graph, the prerequisite knowledge points on which the above key knowledge points depend are identified. A structural causal model is then constructed, and the state transfer equation is defined as: Mastery State (Ki) = f(Mastery State (Prereq1), Mastery State (Prereq2), ..., θi); Prereq1, Prereq2, ... represent the prerequisite knowledge points for knowledge point Ki, indicating that the mastery state of knowledge point Ki is jointly determined by the student's mastery state of the prerequisite knowledge points and their individual mastery threshold θi. The result prediction equation is: Probability of Correct Answer P(Correct) = g(K1,K2,...,), indicating that the probability of correct answer is calculated jointly by the mastery states of all knowledge points involved in the problem-solving process. Students' historical answer data are input into the structural causal model. The parameters in the state transfer equation f and the outcome prediction equation g are continuously adjusted and learned through Bayesian parameter learning or maximum likelihood estimation, and the individual mastery threshold θi of students for knowledge point Ki is fitted.

[0007] Furthermore, the counterfactual intervention reasoning module includes a counterfactual question generation unit, an intervention simulation calculation unit, and a gain calculation and ranking unit. The counterfactual question generation unit is used to propose hypothetical questions for each knowledge point Kd in the structural causal model: "If a student masters Kd, by how much will their probability of answering correctly increase?". The intervention simulation calculation unit is used to execute the intervention operation do (mastery state (Kd) = 1) in the trained structural causal model, keeping other variables unchanged, and re-predict the probability of answering correctly P' (correct). The gain calculation and ranking unit is used to calculate the corrective gain Gain(Kd) = P' (correct). P (Correct), and sort all knowledge points in descending order of correction gain value, and select the knowledge point with the highest correction gain as the primary target knowledge point for intervention.

[0008] Furthermore, the execution process of the personalized correction path generation module is as follows: after determining the target knowledge point for primary intervention, based on the relation edges in the dynamic cognitive knowledge graph, all prerequisite knowledge points required to reach the target knowledge point are searched in reverse. Then, based on the dependence strength between knowledge points and the student's existing knowledge point mastery, the path order is optimized to generate a minimum cost learning chain suitable for the student. The minimum cost learning chain is output in the form of a progressive learning task.

[0009] Furthermore, the specific process of generating the minimum-cost learning chain is as follows: Starting with the primary target knowledge point as the endpoint, the algorithm backtracks along the relational chain of the dynamic cognitive knowledge graph to identify the set of all prerequisite knowledge points required to reach the target knowledge point. Then, it removes the already mastered prerequisite knowledge points, retaining only those that are not yet mastered or are weakly mastered, thus constructing a subgraph of knowledge to be learned. Based on the dependence strength between knowledge points, the cognitive difficulty of the knowledge points themselves, and the estimated learning time, it assigns cost weights to the knowledge points and their connections in the subgraph of knowledge to be learned. Subsequently, within the subgraph of knowledge to be learned, a graph search algorithm is used to find an optimal path that minimizes the total learning cost of all knowledge points on this optimal path, while strictly satisfying the logical order in the dynamic cognitive knowledge graph. Finally, the knowledge points on this optimal path are arranged in sequential order, transforming it into a progressive learning task, i.e., generating a minimum-cost learning chain.

[0010] Furthermore, the adaptive practice recommendation module is implemented using a deep reinforcement learning framework, specifically including a state representation unit, an action space definition unit, a reward mechanism design unit, and a policy learning unit. The state representation unit encodes the student's current knowledge mastery state into a vector, forming the state s of a Markov decision process. The action space definition unit uses the set of recommendable practice questions as the action set a, with each question labeled with the knowledge point it tests, difficulty level, and question type. The reward mechanism design unit calculates immediate rewards based on the student's accuracy, response time, and breadth of knowledge coverage. The policy learning unit trains the recommendation agent using DQN, PPO, or other policy gradient algorithms, enabling it to learn the optimal policy function π(a|s), that is, to select questions that maximize long-term ability improvement for recommendation given state s.

[0011] The beneficial effects of this invention are as follows: 1. This invention, by constructing a dynamic cognitive knowledge graph and combining it with a structural causal model, achieves accurate attribution of the deep-seated knowledge deficiencies behind students' wrong answers, breaking through the limitation of traditional methods that can only classify errors; 2. By introducing a counterfactual reasoning mechanism, this invention can simulate "how the outcome would change if a certain knowledge point were mastered", thereby scientifically assessing the expected gains of each intervention point and providing a data-driven decision-making basis for personalized correction. This invention significantly improves the interpretability, accuracy, and practicality of educational diagnostic systems, enabling each student to access understandable, actionable, and efficient learning improvement paths in complex subject scenarios, thereby effectively promoting personalized learning and enhancing teaching quality. Detailed Implementation

[0012] The present invention will be further described below. It should be noted that this embodiment is based on the present technical solution and provides detailed implementation methods and specific operation processes, but the protection scope of the present invention is not limited to this embodiment.

[0013] This embodiment provides a system for diagnosing and adaptively correcting cognitive deficits in work, including a dynamic cognitive knowledge graph construction module, a causal diagnosis module, a counterfactual intervention reasoning module, a personalized correction path generation module, and an adaptive practice recommendation module. The dynamic cognitive knowledge graph construction module is used to automatically construct a dynamic knowledge graph containing knowledge point nodes and their interrelationships based on student answer data from subject textbooks, curriculum standards, and history; the interrelationships between the knowledge point nodes include at least prerequisite relationships and logical deduction relationships.

[0014] The causal diagnosis module is used to identify the key knowledge points that cause the deviation from the correct solution path when a student makes a mistake. This is done by comparing the student's actual solution process for the current incorrect question with the standard solution path. The module also identifies all the prerequisite knowledge points for the key knowledge points based on the dynamic cognitive knowledge graph and constructs a structural causal model accordingly. The counterfactual intervention reasoning module is used to perform counterfactual simulations by combining structural causal models. It calculates the increase in the probability of students answering questions correctly after mastering each knowledge point, and then evaluates the corrective gain of each knowledge point, and determines the optimal intervention priority based on this ranking. The personalized correction path generation module is used to take the knowledge point with the highest correction gain as the target knowledge point, and combine the dependency path of the target knowledge point in the dynamic cognitive knowledge graph with the student's current mastery status of each knowledge point (not yet mastered, weak mastery, and mastered) to generate an optimal learning sequence to help the student reach the target knowledge point. The adaptive practice recommendation module is used to continuously monitor the changes in a student's mastery of each knowledge point in the optimal learning sequence as the student executes the optimal learning sequence based on the reinforcement learning algorithm, and dynamically recommend the most suitable practice questions accordingly.

[0015] When a student makes a mistake in answering a question, the causal diagnosis module, the counterfactual intervention reasoning module, and the personalized correction path generation module will be triggered again to re-diagnose the knowledge points where the student's mastery is weak, determine the optimal intervention priority, and generate the optimal learning sequence, forming a closed loop of diagnosis-correction-assessment-adaptive practice recommendation.

[0016] In this embodiment, the dynamic cognitive knowledge graph construction module includes a node extraction unit and a dependency mining unit. The node extraction unit is used to extract subject knowledge points from subject textbooks and curriculum standards as graph nodes. The dependency mining unit is used to utilize the step sequence of high-quality student problem-solving processes, and through sequence pattern mining and rule learning methods, automatically identify the prerequisites and logical order between knowledge points, and establish prerequisite relationship edges and logical deduction relationship edges between knowledge points.

[0017] In this embodiment, the execution process of the causal diagnosis module is as follows: The standard solution path for the current incorrect question is represented as a sequence of atomic operations of standard knowledge points applied in sequence, and the student's solution process for the current incorrect question is decomposed into a sequence of atomic operations of knowledge points of incorrect question; by comparing the atomic operation sequence of standard knowledge points with the atomic operation sequence of knowledge points of incorrect question, the key knowledge points (such as missing or disordered knowledge points) that cause the atomic operation sequence of knowledge points of incorrect question to deviate from the atomic operation sequence of standard knowledge points are identified. Based on the dynamic cognitive knowledge graph, the prerequisite knowledge points upon which the above key knowledge points depend are identified. A structural causal model is then constructed, and the state transfer equation (structural equation) is defined as: Mastery State (Ki) = f(Mastery State (Prereq1), Mastery State (Prereq2), ..., θi); Prereq1, Prereq2, ... represent the prerequisite knowledge points for knowledge point Ki, indicating that the mastery state of knowledge point Ki is jointly determined by the student's mastery state of the prerequisite knowledge points and their individual mastery threshold θi. The prediction equation is: the probability of answering correctly P(correct) = g(K1,K2,...), indicating that the probability of answering correctly is calculated jointly by the mastery states of all knowledge points involved in the problem-solving process.

[0018] Students' historical answer data (including students' actual problem-solving steps and final correct or incorrect results) are input into the structural causal model. The parameters in the state transfer equation f and the outcome prediction equation g are continuously adjusted and learned through Bayesian parameter learning or maximum likelihood estimation, and the individual mastery threshold θi of students for knowledge point Ki is fitted.

[0019] In this embodiment, the counterfactual intervention reasoning module includes a counterfactual question generation unit, an intervention simulation calculation unit, and a gain calculation and ranking unit. The counterfactual question generation unit is used to propose a hypothetical question for each knowledge point Kd in the structural causal model: "If a student masters Kd, by how much will their probability of answering correctly increase?". The intervention simulation calculation unit is used to execute the intervention operation do (mastery state (Kd) = 1) in the trained structural causal model, keeping other variables unchanged, and re-predict the probability of answering correctly P' (correct). The gain calculation and ranking unit is used to calculate the corrective gain Gain(Kd) = P' (correct). P (Correct), and sort all knowledge points in descending order of correction gain value, and select the knowledge point with the highest correction gain as the primary target knowledge point for intervention.

[0020] In this embodiment, the personalized correction path generation module performs the following steps: After determining the target knowledge point for primary intervention, it searches backwards for all prerequisite knowledge points required to reach the target knowledge point based on the relational edges in the dynamic cognitive knowledge graph. Then, based on the dependence strength between knowledge points and the student's existing knowledge point mastery, it optimizes the path order and generates a minimum-cost learning chain suitable for the student. The minimum-cost learning chain is output in the form of progressive learning tasks, such as: "Review the trigonometric function values ​​of special angles → Practice the side-angle relationship of right triangles → Apply to force decomposition calculations".

[0021] Furthermore, in this embodiment, the specific process of generating the minimum cost learning chain is as follows: Starting with the primary target knowledge point as the endpoint, the algorithm backtracks along the relational chain of the dynamic cognitive knowledge graph to identify the set of all prerequisite knowledge points required to reach the target knowledge point. Then, knowledge points already mastered are removed from the prerequisite knowledge point set, retaining only those not yet mastered or weakly mastered, constructing a subgraph of knowledge points to be learned specifically for the student. Based on factors such as the strength of dependencies between knowledge points, the cognitive difficulty of the knowledge points themselves, and the estimated learning time, cost weights are assigned to the knowledge points and their connections in the subgraph (e.g., the greater the difficulty of understanding and the longer the practice time required, the higher the learning cost weight). Subsequently, within the subgraph, a graph search algorithm (such as the shortest path algorithm combined with topological sorting) is used to find an optimal path, ensuring that the total learning cost (total weight) of all knowledge points on this optimal path is minimized, while strictly satisfying the logical order in the dynamic cognitive knowledge graph (i.e., prerequisite knowledge points must be learned before subsequent knowledge points). Finally, the knowledge points on this optimal path are arranged in sequential order, transforming it into a progressive learning task, i.e., generating a minimum-cost learning chain.

[0022] In this embodiment, the adaptive practice recommendation module is implemented using a deep reinforcement learning framework, specifically including a state representation unit, an action space definition unit, a reward mechanism design unit, and a policy learning unit. The state representation unit encodes the student's current knowledge mastery state into a vector, forming the state s of a Markov decision process. The action space definition unit uses the set of recommendable practice questions as the action set a, with each question labeled with the knowledge point it tests, difficulty level, and question type. The reward mechanism design unit calculates the immediate reward based on the student's accuracy, response time, and breadth of knowledge coverage. The policy learning unit trains the recommendation agent using DQN, PPO, or other policy gradient algorithms to learn the optimal policy function π(a|s), that is, to select questions that maximize long-term ability improvement for recommendation given state s.

[0023] For those skilled in the art, various corresponding changes and modifications can be made based on the above technical solutions and concepts, and all such changes and modifications should be included within the protection scope of the claims of this invention.

Claims

1. A system for diagnosing and adaptively correcting cognitive deficits in work, characterized in that, It includes a dynamic cognitive knowledge graph construction module, a causal diagnosis module, a counterfactual intervention reasoning module, a personalized correction path generation module, and an adaptive practice recommendation module; The dynamic cognitive knowledge graph construction module is used to automatically construct a dynamic knowledge graph containing knowledge point nodes and their interrelationships based on student answer data from subject textbooks, curriculum standards, and history; the interrelationships between the knowledge point nodes include at least prerequisite relationships and logical deduction relationships; The causal diagnosis module is used to identify the key knowledge points that cause the deviation from the correct solution path when a student makes a mistake. This is done by comparing the student's actual solution process for the current incorrect question with the standard solution path. The module also identifies all the prerequisite knowledge points for the key knowledge points based on the dynamic cognitive knowledge graph and constructs a structural causal model accordingly. The counterfactual intervention reasoning module is used to perform counterfactual simulations by combining structural causal models. It calculates the increase in the probability of students answering questions correctly after mastering each knowledge point, and then evaluates the corrective gain of each knowledge point, and determines the optimal intervention priority based on this ranking. The personalized correction path generation module is used to take the knowledge point with the highest correction gain as the target knowledge point, and combine the dependency path of the target knowledge point in the dynamic cognitive knowledge graph with the student's current mastery of each knowledge point to generate an optimal learning sequence to help the student reach the target knowledge point. The adaptive practice recommendation module is used to continuously monitor the changes in a student's mastery of each knowledge point in the optimal learning sequence as the student executes the optimal learning sequence based on the reinforcement learning algorithm, and dynamically recommend the most suitable practice questions accordingly.

2. The system according to claim 1, characterized in that, The dynamic cognitive knowledge graph construction module includes a node extraction unit and a dependency mining unit; the node extraction unit is used to extract subject knowledge points from subject textbooks and curriculum standards as graph nodes; The dependency mining unit utilizes the step sequences of high-quality student problem-solving processes. Through sequence pattern mining and rule learning methods, it automatically identifies the prerequisites and logical order between knowledge points, and establishes prerequisite relationship edges and logical deduction relationship edges between knowledge points.

3. The system according to claim 1, characterized in that, The execution process of the causal diagnosis module is as follows: The standard solution path for the current incorrect question is represented as a sequence of atomic operations of standard knowledge points applied in sequence, and the student's solution process for the current incorrect question is decomposed into a sequence of atomic operations of knowledge points of incorrect question; by comparing the atomic operation sequence of standard knowledge points and the atomic operation sequence of knowledge points of incorrect question, the key knowledge points that cause the atomic operation sequence of knowledge points of incorrect question to deviate from the atomic operation sequence of standard knowledge points are identified. Based on the dynamic cognitive knowledge graph, the prerequisite knowledge points on which the above key knowledge points depend are identified. A structural causal model is then constructed, and the state transfer equation is defined as: Mastery State (Ki) = f(Mastery State (Prereq1), Mastery State (Prereq2), ..., θi); Prereq1, Prereq2, ... represent the prerequisite knowledge points for knowledge point Ki, indicating that the mastery state of knowledge point Ki is jointly determined by the student's mastery state of the prerequisite knowledge points and their individual mastery threshold θi. The result prediction equation is: Probability of Correct Answer P(Correct) = g(K1,K2,...,), indicating that the probability of correct answer is calculated jointly by the mastery states of all knowledge points involved in the problem-solving process. Students' historical answer data are input into the structural causal model. The parameters in the state transfer equation f and the outcome prediction equation g are continuously adjusted and learned through Bayesian parameter learning or maximum likelihood estimation, and the individual mastery threshold θi of students for knowledge point Ki is fitted.

4. The system according to claim 3, characterized in that, The counterfactual intervention reasoning module includes a counterfactual question generation unit, an intervention simulation calculation unit, and a gain calculation and ranking unit. The counterfactual question generation unit is used to propose a hypothetical question for each knowledge point Kd in the structural causal model: "If a student masters Kd, by how much will their probability of answering correctly increase?". The intervention simulation calculation unit is used to execute the intervention operation do (mastery state (Kd) = 1) in the trained structural causal model, keeping other variables unchanged, and re-predict the probability of answering correctly P' (correct). The gain calculation and ranking unit is used to calculate the corrective gain Gain(Kd) = P' (correct). P (Correct), and sort all knowledge points in descending order of correction gain value, and select the knowledge point with the highest correction gain as the primary target knowledge point for intervention.

5. The system according to claim 1, characterized in that, The execution process of the personalized correction path generation module is as follows: After determining the target knowledge point for primary intervention, based on the relation edges in the dynamic cognitive knowledge graph, all prerequisite knowledge points required to reach the target knowledge point are searched in reverse. Then, based on the dependence strength between knowledge points and the student's existing knowledge point mastery, the path order is optimized to generate a minimum cost learning chain suitable for the student. The minimum cost learning chain is output in the form of a progressive learning task.

6. The system according to claim 5, characterized in that, The specific process of generating the minimum cost learning chain is as follows: Starting with the primary target knowledge point as the endpoint, the algorithm backtracks along the relational chain of the dynamic cognitive knowledge graph to identify the set of all prerequisite knowledge points required to reach the target knowledge point. Then, it removes the already mastered prerequisite knowledge points, retaining only those that are not yet mastered or are weakly mastered, thus constructing a subgraph of knowledge to be learned. Based on the dependence strength between knowledge points, the cognitive difficulty of the knowledge points themselves, and the estimated learning time, it assigns cost weights to the knowledge points and their connections in the subgraph of knowledge to be learned. Subsequently, within the subgraph of knowledge to be learned, a graph search algorithm is used to find an optimal path that minimizes the total learning cost of all knowledge points on this optimal path, while strictly satisfying the logical order in the dynamic cognitive knowledge graph. Finally, the knowledge points on this optimal path are arranged in sequential order, transforming it into a progressive learning task, i.e., generating a minimum-cost learning chain.

7. The system according to claim 1, characterized in that, The adaptive practice recommendation module is implemented using a deep reinforcement learning framework, which includes a state representation unit, an action space definition unit, a reward mechanism design unit, and a policy learning unit. The state representation unit encodes the student's current knowledge mastery state into a vector, which constitutes the state s of the Markov decision process. The action space definition unit is used to define the set of recommended practice questions as the action set a, with each question labeled with the knowledge point it tests, difficulty level, and question type; the reward mechanism design unit is used to calculate the immediate reward based on the student's accuracy, response time, and breadth of knowledge coverage; the policy learning unit is used to train the recommending agent using DQN, PPO, or other policy gradient algorithms to learn the optimal policy function π(a|s), that is, to select the question that maximizes the long-term ability improvement effect for recommendation under a given state s.

Images