Intelligent association review method, storage medium and equipment
By constructing a knowledge graph network and using reinforcement learning algorithms, personalized review paths are dynamically generated, which solves the problem of fragmented knowledge forgetting and improves learning efficiency and effectiveness.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-17
- Publication Date
- 2026-04-10
AI Technical Summary
Existing technologies cannot effectively utilize the synergistic effect of knowledge networks and ignore the connections between knowledge points, leading to students easily forgetting fragmented knowledge, requiring repeated learning, and resulting in low learning efficiency.
We construct a knowledge graph network, calculate the overall forgetting rate of knowledge clusters and dynamically optimize review paths, combine reinforcement learning algorithms to generate personalized optimal review paths, dynamically adjust review intervals, and insert collaborative review links for cross-disciplinary related knowledge points.
It automatically recommends relevant knowledge points, intelligently optimizes review time, dynamically adjusts review intervals, generates personalized optimal review paths, and improves learning efficiency.
Smart Images
Figure SMS_2 
Figure SMS_3 
Figure SMS_4
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of intelligent education, in particular to an intelligent associated review method, a storage medium and equipment. BACKGROUND
[0002] For fragmented knowledge, students are prone to completely forget due to lack of timely and scientific review and consolidation, resulting in repeated learning, time waste and low learning efficiency. The traditional review system usually processes individual knowledge points independently, ignoring the relevance between knowledge points, and cannot effectively utilize the synergistic effect of knowledge network. In addition, the existing interval repetition algorithm mainly adjusts the review interval based on the time dimension, and fails to comprehensively consider multi-dimensional factors such as knowledge structure complexity, cross-disciplinary association and user learning state, making it difficult to provide personalized and efficient review strategies. SUMMARY
[0003] In view of the deficiencies of the prior art, the present application aims to provide an intelligent associated review method, a storage medium and equipment, which can realize more scientific memory consolidation by constructing a knowledge point network, calculating the forgetting degree of a knowledge cluster and dynamically optimizing a review path.
[0004] In order to achieve the above-mentioned purpose, the present application adopts the following technical solutions: An intelligent associated review method, comprising the following steps: S1, constructing a knowledge graph network, the knowledge graph network comprising a plurality of knowledge point nodes and edges representing the relationship between knowledge points; S2, clustering the plurality of knowledge points to form a knowledge cluster according to the relationship between the knowledge points in the knowledge graph network; S3, calculating the comprehensive forgetting degree of each knowledge cluster, the comprehensive forgetting degree being determined based on the memory retention value of each knowledge point in the knowledge cluster and the association strength weight between the knowledge points; the memory retention value of each knowledge point in the knowledge cluster being obtained according to the following formula: ; Wherein, Fi(t) represents the memory retention value of knowledge point i at time t, the value range being 0-1, 1 representing complete memory and 0 representing complete forgetting; a_i represents the importance of knowledge point i in the knowledge cluster, the value range being 0-1; β_i represents the difficulty coefficient of knowledge point i; S_i represents the stability index of knowledge point i; t represents the time from the last review of knowledge point i; Then, the memory retention values of the knowledge points in the knowledge cluster are weighted and summed according to the association strength weight to obtain the comprehensive forgetting degree score of the knowledge cluster; S4, generating an optimal review path according to the comprehensive forgetting degree of each knowledge cluster, the review path comprising the review order of the knowledge cluster and the combination mode of the associated knowledge points; Establish a state-action value function using reinforcement learning algorithms:
[0005] Among them, s t This is a vector representing the user's current knowledge state, containing the memory retention value and review history for each knowledge point; a t To select the knowledge clusters for review; r t The predicted knowledge is retained as an enhancement value after review; γ is the discount factor; Learn Q-value online using the following update rules:
[0006] It employs an ε-greedy strategy for action selection, striking a balance between exploring new strategies and utilizing existing experience, thereby dynamically generating a personalized optimal review path. S5. Dynamically adjust the review interval of knowledge clusters based on the user's actual review results.
[0007] Furthermore, the specific process of step S1 is as follows: S1.1 Extract knowledge point entities and the relationships between knowledge point entities from learning materials using natural language processing technology; S1.2 Assign association strength weights to the relationships between entities in each knowledge point, with weight values ranging from 0 to 1; S1.3. Using knowledge point entities as nodes and the relationships between knowledge point entities as weighted edges, construct a directed graph structure to obtain a knowledge graph network.
[0008] Furthermore, in the weighted summation of step S3, the effective weight of a certain knowledge point node is aggregated from the association strength weights of all its corresponding weighted edges.
[0009] Furthermore, the specific process of step S5 is as follows: Based on the difference between the actual accuracy rate A and the expected accuracy rate E in the knowledge cluster test, adjust the review interval according to the following formula:
[0010] Where I_new represents the new review interval, I_old represents the original review interval, and η represents the learning rate parameter.
[0011] Furthermore, the above method also includes the following steps: Establish an interdisciplinary knowledge association model to map knowledge points from different disciplines to a unified vector space; identify interdisciplinary knowledge association clusters through dimensionality reduction algorithms; and insert collaborative review sessions of interdisciplinary related knowledge points into the review path.
[0012] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.
[0013] The present invention also provides a computer device, including a processor and a memory, wherein the memory is used to store a computer program; and the processor is used to execute the computer program to implement the above-described method.
[0014] The beneficial effects of this invention are as follows: 1. This invention can automatically recommend other knowledge points related to the currently reviewed knowledge points, forming a systematic review path and realizing knowledge association recommendation; 2. This invention can intelligently optimize review time and dynamically adjust review intervals based on the user's actual level of mastery and forgetting curve; 3. This invention dynamically generates a personalized optimal review path by integrating multiple dimensions. Detailed Implementation
[0015] 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.
[0016] This embodiment provides an intelligent associative review method, including the following steps: S1. Construct a knowledge graph network, wherein the knowledge graph network includes multiple knowledge point nodes and edges representing the relationships between knowledge points; S2. Based on the relationships between knowledge points in the knowledge graph network, cluster multiple knowledge points into knowledge clusters; S3. Calculate the overall forgetting rate of each knowledge cluster, wherein the overall forgetting rate is determined based on the memory retention value and association weight of each knowledge point within the knowledge cluster; S4. Generate the optimal review path based on the overall forgetting rate of each knowledge cluster. The review path includes the review order of knowledge clusters and the combination of related knowledge points. S5. Dynamically adjust the review interval of knowledge clusters based on the user's actual review results.
[0017] Specifically, the process of step S1 is as follows: S1.1 Extract knowledge point entities and the relationships between knowledge point entities from learning materials using natural language processing technology; S1.2 Assign association strength weights to the relationships between entities in each knowledge point, with weight values ranging from 0 to 1; S1.3. Using knowledge point entities as nodes and the relationships between knowledge point entities as weighted edges, construct a directed graph structure to obtain a knowledge graph network.
[0018] Specifically, in step S3, the memory retention value of each knowledge point within the knowledge cluster is obtained using the following formula: ; Where Fi(t) represents the memory retention value of knowledge point i at time t, and the value range is 0-1 (1 means completely remembered, 0 means completely forgotten). For example, the memory retention value of the trigonometric function formula after 3 days may be 0.65.
[0019] α_i represents the importance of knowledge point i within the knowledge cluster. Its value ranges from 0 to 1, reflecting the core nature of the knowledge point within the cluster. The value is either manually preset or automatically calculated based on the exam frequency of that knowledge point. For example, the importance of the Pythagorean theorem can be 0.9 (a core theorem), and the importance of trigonometric function values of special angles can be 0.7.
[0020] β_i represents the difficulty coefficient of knowledge point i, which characterizes the difficulty of understanding the knowledge point. The value can be dynamically adjusted based on the students' correct answer rate for knowledge point i, with a default value of 0.5. For example, the difficulty coefficient of finding the roots of a quadratic function is 0.6, and the difficulty coefficient of a linear equation in one variable is 0.3.
[0021] S_i represents the stability index of knowledge point i, which reflects the strength of memory. It increases as the number of times the student reviews knowledge point i increases, and the initial value is set to 1.0 (for newly learned knowledge points). For example, the stability index of the factorization method after 3 reviews can be dynamically adjusted to 2.5.
[0022] t represents the time elapsed since the last review of knowledge point i. For example, if the last review of the "perfect square formula" was 5 days ago, then t = 5.
[0023] In the above formula, -(β_i*t / S_i) determines the forgetting rate. The larger the difficulty coefficient β_i, the faster the forgetting. The higher the stability coefficient S_i, the slower the forgetting.
[0024] The memory retention values of each knowledge point within a knowledge cluster are weighted and summed based on the association strength weights to obtain a comprehensive forgetting score for the knowledge cluster. Thus, forgetting important, related knowledge points will increase the overall forgetting score of the knowledge cluster.
[0025] Specifically, in this embodiment, the effective weight of a knowledge point node is aggregated from the association strength weights of all its corresponding weighted edges (such as summation, averaging, or normalized values adjusted based on importance).
[0026] For example, the association strength weights of all weighted edges corresponding to knowledge point node i can be summed using the following formula:
[0027] Wherein, knowledge point N(i) is the set of neighboring nodes connected to node i, weight(i,j) is the association strength weight of edge (i,j), and the final calculated Wi is the effective weight of knowledge point node i.
[0028] The above calculation process will be illustrated with examples below.
[0029] Suppose we are reviewing the knowledge cluster of "plane geometry proofs", which includes three knowledge points: Pythagorean theorem, similar triangle criteria, and power of a circle theorem.
[0030] Pythagorean theorem (α_1=0.9, β_1=0.5, S_1=2.0, t=4 days) F1 = 0.9×e (-0.5×4 / 2) ≈ 0.55 Similar triangle determination (α_2=0.7, β_2=0.6, S_2=1.8, t=3 days) F2 = 0.7×e (-0.6×3 / 1.8) ≈ 0.47 Power theorem (α_3=0.6, β_3=0.7, S_3=1.5, t=5 days) F3 = 0.6×e (-0.7×5 / 1.5) ≈ 0.18 Therefore, the overall forgetting rate of this knowledge cluster = (0.55×0.8+0.47×0.6+0.18×0.4) / (0.8+0.6+0.4) ≈ 0.43 The algorithm in step S3 quantifies students' retention of knowledge points, calculates a comprehensive forgetting score for knowledge clusters, and intelligently determines which knowledge clusters most need priority review. Its core is based on the forgetting curve theory, combining the importance, difficulty, and stability of knowledge points to dynamically assess students' forgetting levels.
[0031] Specifically, in this embodiment, in step S4, a state-action value function is established using a reinforcement learning algorithm (Q-learning):
[0032] Among them, s t This is a vector representing the user's current knowledge state, containing the memory retention value and review history for each knowledge point; a t To select the knowledge clusters for review; r t The predicted knowledge is retained as an enhancement value after review; γ is the discount factor.
[0033] Learn Q-value online using the following update rules:
[0034] It employs an ε-greedy strategy for action selection, striking a balance between exploring new strategies and utilizing existing experience, thereby dynamically generating a personalized optimal review path.
[0035] The aforementioned reinforcement learning algorithm is used to intelligently generate the optimal review path, dynamically adjusting the review strategy by simulating a closed-loop process of "learning-feedback-optimization". Its core value lies in: Personalized adaptation: based on the student's real-time knowledge status. t Recommended knowledge cluster a, which is most suitable for review t ; Maximizing long-term returns: through reward mechanisms r t Ensure that each review session maximizes knowledge retention and improvement; Exploring and Utilizing the Balance: The ε-greedy strategy ensures both the consolidation of regular knowledge and the discovery of new and effective review combinations; Status s t (Current knowledge status) uses digital representation to indicate the student's mastery of each knowledge point, such as: Trigonometric function proficiency: 72 / 100 Geometric proof proficiency: 85 / 100 Time difference from last review: 3 days Dynamic changes: The state matrix is updated after each review session. Action a t (Review Selection) is used to decide which specific knowledge clusters to review, and its operation process is as follows: Beginner's Step: Select the "Quadratic Equations" knowledge cluster. Advanced actions: Combine the review of related knowledge clusters of "function graphs" and "inequalities". For example, when the forgetting of trigonometric functions is detected to be accelerated, the "unit circle application" sub-cluster is automatically added.
[0036] Rewards t (Knowledge retention prediction), its calculation logic is as follows: r t = ΔMemory retention value × Association strength weight - Cognitive load cost For example: When you successfully review interdisciplinary knowledge points related to "Algebra-Geometry", you will receive a reward of +1.5 points. When reviewing knowledge points that have already been mastered, the reward value is reduced by 0.3 (to avoid ineffective review). Furthermore, the principle of the ε-greedy strategy is as follows: There is a 90% probability (1-ε) of choosing the current optimal review path. 10% probability (ε) to randomly explore new review combinations For example: Conventional selection: Review "Factorization" according to the forgetting curve (optimal path). Exploratory selection: Try new associations between "Summation of sequences and function graphs" (may discover a more efficient review mode).
[0037] The above reinforcement learning algorithm is illustrated by the following specific examples.
[0038] (1) Breakthrough of weak links: When the knowledge state s t shows that the correct rate of "Binary linear equations" < 60%: Automatically increase the appearance frequency of this knowledge cluster; Strengthen the basic knowledge point of "Equality property".
[0039] (2) Optimization during pre-exam sprint: When it is detected that there are 7 days until the math final exam: Adjust the reward function r t Focus on the high-frequency exam points; Generate a composite review unit of "Geometric proof + Algebraic calculation".
[0040] (3) Interdisciplinary linkage: When reviewing the "Force analysis on inclined plane" in physics: Trigger collaborative review related to "Trigonometric functions" in mathematics; The calculation of the reward value includes the cross-disciplinary gain coefficient.
[0041] In this embodiment, the specific process of step S5 is as follows: According to the difference between the actual correct rate A and the expected correct rate E of the knowledge cluster test, adjust the review interval according to the following formula:
[0042] Among them, I_new represents the new review interval, I_old represents the original review interval, and η represents the learning rate parameter. When A > E, extend the review interval; when A < E, shorten the review interval.
[0043] It should be noted that the algorithm of step S5 is used to intelligently adjust the review interval time of the knowledge cluster, and dynamically optimize the review plan by comparing the difference between the actual mastery degree of students and the expected goal. Its core value is reflected in: Personalized adaptation: Adjust the review frequency according to the real-time test performance of each student; Efficiency maximization: Avoid over-reviewing the mastered content and strengthen the weak links; Scientific memory: Conform to the memory law of the Ebbinghaus forgetting curve.
[0044] More specifically, the actual correct rate A can be the true scoring rate of the student in the most recent test for this knowledge cluster. For example, if Xiaoming answers 8 out of 10 questions in the "Quadratic equation of one variable" test, then A = 80%.
[0045] The expected correct rate E can be the achievable level automatically predicted based on historical data. For newly learned knowledge points, the initial value of E is 60%, and it increases by 5 - 10% each time it is reviewed (e.g., after the first review, E = 70%). For example, after 3 reviews of the "perfect square formula", E = 85%.
[0046] The original review interval I_old is the number of days of the review interval currently in use. If it is the first time learning, the default review interval is 1 day.
[0047] The learning rate η is a sensitivity coefficient that controls the adjustment amplitude, and its typical value is 0.2 - 0.5. The adjustment principle of the learning rate is: for science knowledge points, η takes a larger value (e.g., for mathematics, η = 0.4), and for liberal arts knowledge points, η takes a smaller value (e.g., for ancient poems, η = 0.2).
[0048] The difference ratio (A - E) / E is used to quantify the degree to which a student's performance exceeds / falls below the expectation. For example, if A = 75% and E = 80%, then the difference ratio = (75 - 80) / 80 = -0.0625.
[0049] 1 + η(A - E) / E represents the adjustment coefficient. When A > E, the coefficient > 1 (lengthen the interval), and when A < E, the coefficient < 1 (shorten the interval).
[0050] The following illustrates the calculation of the above algorithm through examples.
[0051] Scenario: η = 0.3, A = 90%, E = 80%, I_old = 5 days Calculation: I_new = 5 * (1 + 0.3 * (0.9 - 0.8) / 0.8) = 5 * 1.0375 ≈ 5.2 days Example: 1. Foundation consolidation stage: Grade 7 students just start learning "rational number operations" Parameter settings: E_initial = 60%, η = 0.4 Adjustment process: First test A = 55% → The review interval is shortened from 3 days to 2.6 days Second test A = 68% → The interval is adjusted to 3.2 days 2. Pre - exam sprint stage: Review of "Archimedes' principle" in physics for Grade 9 students Special treatment: Increase E to 90% (high standard for exams) Set η = 0.5 (quick response) Effect: Make weak knowledge points receive more intensive reviews 3. Interdisciplinary linkage: "Function graphs" in mathematics and "kinematic graphs" in physics Adjustment strategy: When any subject A < E, synchronously shorten the interval between the two subjects Calculate the value of A using the average correct rate of associated knowledge points.
[0052] Furthermore, the method of this embodiment further includes the following steps: Establish an interdisciplinary knowledge association model, map knowledge points of different subjects to a unified vector space; identify interdisciplinary knowledge association clusters through a dimensionality reduction algorithm; insert a collaborative review session of interdisciplinary associated knowledge points in the review path.
[0053] The process of establishing the interdisciplinary knowledge association model includes: (1) Use a pre-trained language model (such as Sentence-BERT) to encode the text descriptions of knowledge points in each subject and map them to a unified semantic vector space; (2) Calculate the semantic similarity between each knowledge point, and construct an interdisciplinary semantic association graph with the similarity relationship between knowledge points whose semantic similarity exceeds a preset threshold; (3) Use UMAP or t-SNE to reduce the dimensionality of the semantic vectors, and use DBSCAN or spectral clustering algorithm to identify interdisciplinary knowledge association clusters; (4) When generating a review path, detect whether the current knowledge point to be reviewed belongs to a certain interdisciplinary association cluster. If so, insert a corresponding collaborative review session in the review sequence. The collaborative review session includes parallel review, comparative explanation, or interdisciplinary application exercises; (5) Dynamically adjust the interdisciplinary association weight according to the user's mastery feedback of collaborative knowledge points to achieve personalized knowledge network evolution.
[0054] For those skilled in the art, various corresponding changes and deformations can be given based on the above technical solutions and concepts, and all these changes and deformations should be included within the protection scope of the claims of this invention.
Claims
1. An intelligent associative review method, characterized in that, Includes the following steps: S1. Construct a knowledge graph network, wherein the knowledge graph network includes multiple knowledge point nodes and edges representing the relationships between knowledge points; S2. Based on the relationships between knowledge points in the knowledge graph network, cluster multiple knowledge points into knowledge clusters; S3. Calculate the overall forgetting rate of each knowledge cluster, which is determined based on the memory retention value of each knowledge point within the knowledge cluster and the association strength weight between each knowledge point; obtain the memory retention value of each knowledge point within the knowledge cluster using the following formula: ； Where Fi(t) represents the memory retention value of knowledge point i at time t, with a value range of 0-1, where 1 represents complete memorization and 0 represents complete forgetting; α_i represents the importance of knowledge point i in the knowledge cluster, with a value range of 0-1; β_i represents the difficulty coefficient of knowledge point i; S_i represents the stability index of knowledge point i; and t represents the time since the last review of knowledge point i. Then, based on the association strength weight, the memory retention values of each knowledge point within the knowledge cluster are weighted and summed to obtain the comprehensive forgetting score of the knowledge cluster; S4. Generate the optimal review path based on the overall forgetting rate of each knowledge cluster. The review path includes the review order of knowledge clusters and the combination of related knowledge points. Establish a state-action value function using reinforcement learning algorithms: Among them, s t This is a vector representing the user's current knowledge state, containing the memory retention value and review history for each knowledge point; a t To select the knowledge clusters for review; r t The predicted knowledge is retained as an enhancement value after review; γ is the discount factor; Learn Q-value online using the following update rules: It employs an ε-greedy strategy for action selection, striking a balance between exploring new strategies and utilizing existing experience, thereby dynamically generating a personalized optimal review path. S5. Dynamically adjust the review interval of knowledge clusters based on the user's actual review results.
2. The method according to claim 1, characterized in that, The specific process of step S1 is as follows: S1.1 Extract knowledge point entities and the relationships between knowledge point entities from learning materials using natural language processing technology; S1.2 Assign association strength weights to the relationships between entities in each knowledge point, with weight values ranging from 0 to 1; S1.
3. Using knowledge point entities as nodes and the relationships between knowledge point entities as weighted edges, construct a directed graph structure to obtain a knowledge graph network.
3. The method according to claim 2, characterized in that, In the weighted summation of step S3, the effective weight of a knowledge point node is formed by aggregating the association strength weights of all its corresponding weighted edges.
4. The method according to claim 1, characterized in that, The specific process of step S5 is as follows: Based on the difference between the actual accuracy rate A and the expected accuracy rate E in the knowledge cluster test, adjust the review interval according to the following formula: Where I_new represents the new review interval, I_old represents the original review interval, and η represents the learning rate parameter.
5. The method according to claim 1, characterized in that, It also includes the following steps: Establish an interdisciplinary knowledge association model to map knowledge points from different disciplines to a unified vector space; identify interdisciplinary knowledge association clusters through dimensionality reduction algorithms; and insert collaborative review sessions of interdisciplinary related knowledge points into the review path.
6. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method described in any one of claims 1-5.
7. A computer device, characterized in that, It includes a processor and a memory, the memory being used to store a computer program; the processor being used to execute the computer program to implement the method of any one of claims 1-5.