An adaptive learning method, storage medium, and device

Abstract

This invention discloses an adaptive learning method, storage medium, and device. By constructing a vector space mapping between the student's logical chain and the standard reasoning chain, and implementing a logical fallacy penalty mechanism, it can accurately identify weak links in the student's thinking process, effectively avoiding the phenomenon of "false mastery." The introduction of dynamic thresholds and confidence assessments based on the standard deviation of historical performance significantly improves the accuracy of intervention triggering and the robustness of state assessment. Simultaneously, by combining a localized knowledge graph vector library and a verification closed-loop mechanism, it not only ensures data privacy and offline availability but also guarantees the effectiveness of learning strategy adjustments and long-term tracking capabilities, thereby greatly improving the efficiency and quality of adaptive learning.

Description

Technical Field

[0001] This invention relates to the field of intelligent education technology, specifically to an adaptive learning method, storage medium, and device. Background Technology

[0002] Existing adaptive learning technologies primarily rely on judging the correctness of students' answers and matching knowledge point tags, often neglecting deep semantic analysis of the internal logical chains within students' problem-solving approaches. Traditional learning path planning schemes typically only push subsequent content based on whether the final answer is correct, failing to identify whether students have "coincidentally correct" answers or "logical gaps," and lacking dynamic consideration of fluctuations in students' cognitive states. This results in imprecise learning path planning, making it difficult to truly achieve personalized tutoring based on the thought process. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention aims to provide an adaptive learning method, storage medium, and device.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: An adaptive learning method includes the following steps: S1. Receive the student's response to the current learning task, the response including the final answer and the solution text; S2. Using a pre-built knowledge graph parsing model, extract concept nodes and logical relationships between concept nodes from the problem-solving text to construct a student logic chain. S3. Map the student logic chain to the standard reasoning chain in the standard knowledge graph using a vector space, and calculate the semantic logic integrity score. S4. When the semantic logic integrity score is lower than the first threshold, regardless of whether the final answer is correct, the Socratic active inquiry mechanism is triggered and the process proceeds to step S5. S5. Based on the node with the lowest confidence in the student logic chain, generate a targeted follow-up question and push it to the interactive interface; S6. Students respond to the follow-up question instructions on the interactive interface.

[0005] Furthermore, the above method also includes step S7: updating the student's learning status data based on the correctness of the student's feedback content and the semantic density of the response, and updating the student's personal ability vector using a deep knowledge tracking model based on the updated learning status data, and synchronizing it to the localized knowledge graph vector library; the localized knowledge graph vector library stores the student's personal ability vector, which is a high-dimensional vector representation calculated by the deep knowledge tracking model in combination with the student's learning status data.

[0006] Further, in step S2, the specific process of constructing the student's logical chain using the knowledge graph parsing model includes: performing dependency parsing on the problem-solving text, identifying entity words as concept nodes, identifying connectors and predicates as logical relation edges, and assigning an initial confidence value to each concept node, with the initial confidence value set to 1.0; if an ambiguous word is detected in a logical relation edge, the initial confidence value of the concept node connected to that logical relation edge is multiplied by a decay coefficient of 0.8; if the input device used by the student to input the answer supports stylus or voice input, the writing pressure sensitivity change rate or voice pause duration when the student inputs the answer is collected; if a sudden change in writing pressure sensitivity or a voice pause exceeding a preset duration is detected at a concept node, the initial confidence value of that concept node is additionally multiplied by a hesitation coefficient of 0.9.

[0007] Furthermore, in step S3, the specific process of scoring the semantic logical integrity is as follows: S3.1. Convert the student logic chain into a high-dimensional vector representation Vstudent, and the standard inference chain into a high-dimensional vector representation Vstandard; S3.2 Calculate the cosine similarity between the two. ; S3.3 Calculate the base value of the semantic logical integrity score: Score base = Sim × K; Where K is the logic chain coverage coefficient, which is equal to the ratio of the number of concept nodes in the student logic chain to the number of concept nodes in the standard reasoning chain. If the ratio is greater than 1, then it is 1.

[0008] Furthermore, after calculating the baseline value of the semantic logical integrity score, a logical fallacy penalty mechanism is used to correct the baseline value. The specific process is as follows: The system uses a pre-defined rule base for detecting logical fallacies to identify whether logical fallacies exist in students' logical chains. These fallacies include affirmative consequents, equivocation, and circular reasoning. If a logical fallacy is identified, the base score is then calculated. base Deduct penalty points P from the base score fallacy The final semantic logical integrity score is obtained; where P fallacy The weight is proportional to the severity of the logical fallacy type.

[0009] Furthermore, in step S4, the first threshold is a dynamic threshold T. dynamic The current formula for calculating the first threshold is:

[0010] Where Tbase is the base threshold and α is the adjustment coefficient. The standard deviation of the student's semantic logical integrity score over the past N learning tasks.

[0011] Furthermore, the calculation process for the response semantic density in step S7 is as follows: (1) Calculate the ratio of the effective information entropy of the students' feedback content to the text length; (2) If the students’ feedback contains repetitive and meaningless words, then reduce the ratio obtained in step (1); (3) If the response semantic density adjusted in step (2) is lower than the preset second threshold, the student is determined to be in a state of cognitive ambiguity, the difficulty level of the subsequent learning task is forcibly reduced by one level, and the corresponding knowledge point is marked as "requires manual review".

[0012] Furthermore, the calculation process for the effective information entropy is as follows: A1. Preprocessing and word segmentation: Perform word segmentation and stop word filtering on the student feedback content; A2. Key Concept Identification: Using a knowledge graph parsing model to identify entity words and logical relation words in the feedback content; A3. Probability Distribution Construction: Statistically analyze the frequency of entity words and logical relation words identified in step A2 in the feedback content, and construct a probability distribution P(x); A4. Entropy Calculation: The entropy of the probability distribution P(x) is calculated based on Shannon's information entropy formula, which is the effective information entropy.

[0013] Furthermore, after step S7, a verification closed-loop step is also included: in the subsequent learning tasks, a micro-test question that is isomorphic to the concept node but with a different context is embedded; if the student's logical chain integrity score in the micro-test question recovers, the "requires manual review" mark for the corresponding knowledge point is removed; if the logical chain integrity score in two consecutive micro-test questions does not recover, a learning report is generated and pushed to the teacher's end.

[0014] 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.

[0015] 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.

[0016] The beneficial effects of this invention are as follows: By constructing a vector space mapping between student logical chains and standard reasoning chains and a logical fallacy penalty mechanism, this invention can accurately identify weak links in students' thinking processes and effectively avoid the phenomenon of "false mastery"; by introducing dynamic thresholds and confidence assessments based on the standard deviation of historical performance, it significantly improves the accuracy of intervention triggering and the robustness of state assessment; at the same time, by combining a localized knowledge graph vector library and a verification closed-loop mechanism, it not only ensures data privacy and offline availability, but also ensures the effectiveness of learning strategy adjustments and long-term tracking capabilities, thereby greatly improving the efficiency and quality of adaptive learning. Detailed Implementation

[0017] 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.

[0018] This embodiment provides an adaptive learning method, including the following steps: S1. Receive the student's response to the current learning task, the response including the final answer and the solution text; S2. Using a pre-built knowledge graph parsing model, extract concept nodes and logical relationships between concept nodes from the problem-solving text to construct a student logic chain. S3. Map the student logic chain to the standard reasoning chain in the standard knowledge graph using a vector space, and calculate the semantic logic integrity score. S4. When the semantic logic integrity score is lower than the first threshold, regardless of whether the final answer is correct, the Socratic active inquiry mechanism is triggered and the process proceeds to step S5. S5. Based on the node with the lowest confidence level (i.e. the weak link) in the student's logical chain, generate targeted follow-up questions and push them to the interactive interface. S6. Students respond to the follow-up question instructions on the interactive interface.

[0019] In this embodiment, the method further includes step S7: updating the student's learning status data based on the correctness of the student's feedback content and the semantic density of the response, and updating the student's personal ability vector using a deep knowledge tracking model based on the updated learning status data, and synchronizing it to the localized knowledge graph vector library.

[0020] The localized knowledge graph vector library stores students' individual ability vectors. These vectors are high-dimensional representations calculated using a deep knowledge tracing model, combined with students' learning status data.

[0021] In this embodiment, step S2, the specific process of constructing the student logical chain by the knowledge graph parsing model includes: performing dependency parsing on the problem-solving text, identifying entity words as concept nodes, identifying connectors and predicates as logical relation edges, and assigning an initial confidence value to each concept node, with the initial confidence value set to 1.0.

[0022] Furthermore, if a vague word is detected in a logical relation edge, the initial confidence value of the concept node connected to that logical relation edge is multiplied by a decay coefficient of 0.8.

[0023] Furthermore, if the input device used by the student to input the answer supports stylus or voice input, the writing pressure sensitivity change rate or voice pause duration when the student inputs the answer is collected; if a sudden change in writing pressure sensitivity or a voice pause exceeding a preset duration is detected at a certain concept node, the initial confidence value of that concept node is multiplied by a hesitation coefficient of 0.9.

[0024] In this embodiment, the specific process of scoring the semantic logical integrity in step S3 is as follows: S3.1. Convert the student logic chain into a high-dimensional vector representation Vstudent, and the standard inference chain into a high-dimensional vector representation Vstandard; S3.2 Calculate the cosine similarity between the two. ; S3.3 Calculate the base value of the semantic logical integrity score: Score base = Sim × K; Where K is the logic chain coverage coefficient, which is equal to the ratio of the number of concept nodes in the student logic chain to the number of concept nodes in the standard reasoning chain. If the ratio is greater than 1, then it is 1.

[0025] Furthermore, in this embodiment, after calculating the base value of the semantic logical integrity score, a logical fallacy penalty mechanism is used to correct the base value. The specific process is as follows: The system uses a pre-defined rule base for detecting logical fallacies to identify whether logical fallacies exist in students' logical chains. These fallacies include affirmative consequents, equivocation, and circular reasoning. If a logical fallacy is identified, the base score is then calculated. base Deduct penalty points P from the base score fallacy The final semantic logical integrity score is obtained; where P fallacy The weight is proportional to the severity of the logical fallacy type.

[0026] In this embodiment, in step S4, the first threshold is a dynamic threshold T. dynamic The current formula for calculating the first threshold is:

[0027] Where Tbase is the base threshold and α is the adjustment coefficient. The standard deviation of the student's semantic logical integrity score over the past N learning tasks.

[0028] Standard deviation This reflects the stability of student performance. A larger standard deviation indicates more unstable performance (fluctuating wildly); a smaller standard deviation indicates more stable performance. If student performance fluctuates greatly (…), it indicates… If the value is large, the calculated Tdynamic will increase, which means that the threshold for triggering the Socratic active inquiry mechanism will need to be raised.

[0029] In this embodiment, when generating targeted follow-up questions in step S5, verification questions that are strongly related to the concept node and have a difficulty level lower than the current learning task can be retrieved from the question bank; or, the concept node can be converted into a natural language question, such as "Please explain why A can deduce B", where A and B are the preceding and following concept nodes connected to the concept node.

[0030] In this embodiment, the calculation process of the response semantic density in step S6 is as follows: (1) Calculate the ratio of effective information entropy of student feedback content to text length. This ratio is a quantitative expression of response semantic density. By measuring the relationship between effective knowledge content and expression length in student feedback, it helps to distinguish between true understanding, false understanding, or vague expression. This serves as a key parameter to correct students' learning status data and personal ability vectors, ensuring the accuracy of adaptive learning.

[0031] (2) If the students’ feedback contains repetitive and meaningless words, then reduce the ratio obtained in step (1); (3) If the response semantic density adjusted in step (2) is lower than the preset second threshold, the student is determined to be in a state of cognitive ambiguity, the difficulty level of the subsequent learning task is forcibly reduced by one level, and the corresponding knowledge point is marked as "requires manual review".

[0032] In this embodiment, the calculation process of the effective information entropy is as follows: A1. Preprocessing and word segmentation: Perform word segmentation and stop word filtering on the student feedback content.

[0033] A2. Key Concept Identification (Validity Screening): This step utilizes a knowledge graph parsing model to identify entity words (concept nodes) and logical relation words in the feedback content. Through this step, only those words that match standard concept nodes in the standard knowledge graph or are related to the current problem-solving logic chain are considered valid information; irrelevant filler words, modifiers, or unrelated concepts are eliminated or have their weight reduced.

[0034] A3. Probability Distribution Construction: Statistically analyze the frequency of entity words and logical relation words identified in step A2 in the feedback content, and construct a probability distribution P(x).

[0035] A4. Entropy Calculation: The entropy of the probability distribution P(x) is calculated based on Shannon's information entropy formula, which is the effective information entropy. If the feedback content repeatedly uses a small number of simple words, the entropy value will be low (low information content). If the feedback content is rich, covers multiple different key concepts, and is evenly distributed, the entropy value will be high (high information content).

[0036] In this embodiment, after step S6, a closed-loop verification step is also included: embedding micro-test questions that are isomorphic to the concept node but have different contexts in the subsequent learning tasks; if the student's logical chain integrity score in the micro-test questions recovers, the "requires manual review" mark for the corresponding knowledge point is removed; if the logical chain integrity score in two consecutive micro-test questions does not recover, a learning report is generated and pushed to the teacher's end.

[0037] Example 2 This embodiment provides an application example of the method described in Embodiment 1.

[0038] Xiaoming, a second-year junior high school student, uses a smart learning tablet equipped with the method described in Example 1. The tablet supports handwriting input and voice interaction, and has a built-in localized mathematical knowledge graph vector library. When proving that "two triangles are congruent if two sides and their included angle are equal" (an application of the SAS criterion theorem), the student is in offline mode.

[0039] The specific process is as follows: Step 1: Receiving the Answer. Xiaoming views the question on his tablet: In △ABC and △DEF, AB=DE, ∠B=∠E, BC=EF. Prove that △ABC... △DEF. Xiaoming used a stylus to write down his solution in the answer area: 1. Because AB = DE (given) 2. Because ∠B = ∠E (given) 3. Because BC = EF (given) 4. Therefore, △ABC △DEF (congruent) Step 2: Constructing the student logical chain and assigning confidence scores Text Analysis: OCR and dependency parsing were performed on Xiaoming's handwritten problem-solving approach text.

[0040] Concept node extraction: Extract "AB=DE", "∠B=∠E", "BC=EF", and "△ABC". Concept nodes such as △DEF.

[0041] Logical relationship edge extraction: Identify "known" and "therefore" as connecting edges.

[0042] Initial confidence level: All concept nodes are initially set to 1.0.

[0043] Multimodal correction: Reading the sensor data from the stylus, it was found that Xiaoming was writing step 4, "Therefore △ABC". When the value is △DEF, the pen tip hovers for more than 3 seconds (preset duration), and the pressure sensitivity fluctuates significantly. It is determined that the student hesitates at this point, so the initial confidence level of the node in step 4 is multiplied by the hesitation coefficient of 0.9, and the confidence level of this node becomes 0.9.

[0044] Step 3: Calculate the semantic logical integrity score Vector mapping: Transform Xiaoming's learning logic chain into a vector Vstudent, and retrieve the standard SAS proof logic chain vector Vstandard from the standard knowledge graph.

[0045] Similarity calculation: Calculate the cosine similarity between vectors Vstudent and Vstandard, Sim=0.95.

[0046] Coverage coefficient: Xiaoming's learning logic chain has 4 concept nodes, the standard reasoning chain has 4 concept nodes, and K=1. Score is the base value for semantic logical integrity. base =0.95×1=0.95.

[0047] Logical fallacy detection: It was detected that Xiaoming directly wrote "congruent" in step 4 without clearly indicating the basis for the judgment "(SAS)". Although the conclusion is correct, the logical chain is missing the key logical support of "reference to the judgment theorem", which is a minor logical fallacy of "insufficient argumentation".

[0048] Penalty: Deduct penalty points Pfallacy = 0.15.

[0049] The final semantic logic integrity score was 0.95. 0.15 = 0.80.

[0050] Step 4: Dynamic Threshold Determination and Trigger Inquiry Historical data retrieval: Retrieve Xiaoming's semantic logic integrity scores from his past 5 geometric proof tasks: [0.9, 0.6, 0.95, 0.5, 0.85].

[0051] First threshold calculation: Calculate the standard deviation of historical scores ≈0.18. Let the base threshold Tbase = 0.75 and the adjustment coefficient α = 0.5. Therefore, Tdynamic = 0.75 × (1 + 0.5 × 0.18) = 0.8175.

[0052] The current semantic logic integrity score is 0.80. <Tdynamic(0.8175) 。

[0053] Therefore, even though the answer is correct, it still triggers a Socratic active inquiry mechanism.

[0054] Step 5: Generate follow-up question instructions Weaknesses were identified: the confidence level of the conclusion node in step 4 was the lowest (0.9, and it contained a logical fallacy).

[0055] Generate instruction: Call the natural language generation template to generate follow-up questions: "Please explain why step 3 (equal sides) combined with steps 1 and 2 can lead to step 4 (triangle congruence)? What is the specific theorem name on which this is based?" The follow-up question instruction is displayed on the interactive interface and accompanied by voice playback.

[0056] Step Six: Feedback Processing and Status Update Student feedback: Xiaoming answered in voice: "Because the two sides enclose an angle, it is the SAS theorem, so they are congruent." Response semantic density calculation: Analyze the speech-to-text content. If the effective information entropy is high and there are no repetitive or meaningless words, the response semantic density is higher than the second threshold.

[0057] Status Update: Once it's determined that the student has grasped the weak areas, the learning status data is updated, marking "SAS Theorem Application" as "Mastered," but retaining "Logical Expression Standardization" as "Needs Improvement." Based on the updated learning status data, the student's individual ability vector is updated using a deep knowledge tracking model and synchronized to the localized knowledge graph vector library.

[0058] Step 7: Verify closed loop and long-term evolution Embedded Micro-Test: In the next push of practice questions, embed a fill-in-the-blank question that requires the words "SAS" to be explicitly written (isomorphic micro-test).

[0059] Results verification: If Xiaoming answers the next question correctly, the "Needs to be strengthened" mark will be removed; if he misses it again, a report will be generated and sent to the teacher.

[0060] 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. An adaptive learning method, characterized in that, Includes the following steps: S1. Receive the student's response to the current learning task, the response including the final answer and the solution text; S2. Using a pre-built knowledge graph parsing model, extract concept nodes and logical relationships between concept nodes from the problem-solving text to construct a student logic chain. S3. Map the student logic chain to the standard reasoning chain in the standard knowledge graph using a vector space, and calculate the semantic logic integrity score. S4. When the semantic logic integrity score is lower than the first threshold, regardless of whether the final answer is correct, the Socratic active inquiry mechanism is triggered and the process proceeds to step S5. S5. Based on the node with the lowest confidence in the student logic chain, generate a targeted follow-up question and push it to the interactive interface; S6. Students respond to the follow-up question instructions on the interactive interface.

2. The method according to claim 1, characterized in that, Step S7 includes: updating the student's learning status data based on the correctness of the student's feedback content and the semantic density of the response, and updating the student's personal ability vector using a deep knowledge tracking model based on the updated learning status data, and synchronizing it to the localized knowledge graph vector library; the localized knowledge graph vector library stores the student's personal ability vector, which is a high-dimensional vector representation calculated by the deep knowledge tracking model in combination with the student's learning status data.

3. The method according to claim 1, characterized in that, In step S2, the specific process of constructing the student's logical chain using the knowledge graph parsing model includes: performing dependency parsing on the problem-solving text, identifying entity words as concept nodes, identifying connectors and predicates as logical relation edges, and assigning an initial confidence value to each concept node, with the initial confidence value set to 1.0; if an ambiguous word is detected in a logical relation edge, the initial confidence value of the concept node connected to that logical relation edge is multiplied by a decay coefficient of 0.8; if the input device used by the student to input the answer supports stylus or voice input, the writing pressure sensitivity change rate or voice pause duration when the student inputs the answer is collected; if a sudden change in writing pressure sensitivity or a voice pause exceeding a preset duration is detected at a concept node, the initial confidence value of that concept node is additionally multiplied by a hesitation coefficient of 0.

9.

4. The method according to claim 1, characterized in that, In step S3, the specific process of scoring the semantic logical integrity is as follows: S3.

1. Convert the student logic chain into a high-dimensional vector representation Vstudent, and the standard inference chain into a high-dimensional vector representation Vstandard; S3.2 Calculate the cosine similarity between the two. ; S3.3 Calculate the base value of the semantic logical integrity score: Score base = Sim × K; Where K is the logic chain coverage coefficient, which is equal to the ratio of the number of concept nodes in the student logic chain to the number of concept nodes in the standard reasoning chain. If the ratio is greater than 1, then it is 1. After calculating the baseline value of the semantic logical integrity score, a logical fallacy penalty mechanism is further used to correct the baseline value: The system uses a pre-defined rule base for detecting logical fallacies to identify whether logical fallacies exist in students' logical chains. These fallacies include affirmative consequents, equivocation, and circular reasoning. If a logical fallacy is identified, the base score is then calculated. base Deduct penalty points P from the base score fallacy The final semantic logical integrity score is obtained; where P fallacy The weight is proportional to the severity of the logical fallacy type.

5. The method according to claim 1, characterized in that, In step S4, the first threshold is a dynamic threshold T. dynamic The current formula for calculating the first threshold is: Where Tbase is the base threshold and α is the adjustment coefficient. The standard deviation of the student's semantic logical integrity score over the past N learning tasks.

6. The method according to claim 2, characterized in that, The calculation process of the response semantic density in step S7 is as follows: (1) Calculate the ratio of the effective information entropy of the students' feedback content to the text length; (2) If the students’ feedback contains repetitive and meaningless words, then reduce the ratio obtained in step (1); (3) If the response semantic density adjusted in step (2) is lower than the preset second threshold, the student is determined to be in a state of cognitive ambiguity, the difficulty level of the subsequent learning task is forcibly reduced by one level, and the corresponding knowledge point is marked as "requires manual review".

7. The method according to claim 6, characterized in that, The calculation process for the effective information entropy is as follows: A1. Preprocessing and word segmentation: Perform word segmentation and stop word filtering on the student feedback content; A2. Key Concept Identification: Using a knowledge graph parsing model to identify entity words and logical relation words in the feedback content; A3. Probability Distribution Construction: Statistically analyze the frequency of entity words and logical relation words identified in step A2 in the feedback content, and construct a probability distribution P(x); A4. Entropy Calculation: The entropy of the probability distribution P(x) is calculated based on Shannon's information entropy formula, which is the effective information entropy.

8. The method according to claim 2, characterized in that, After step S7, there is also a verification closed-loop step: in the subsequent learning tasks, embed micro-test questions that are isomorphic to the concept node but have different contexts; if the student's logical chain integrity score in the micro-test question recovers, the "requires manual review" mark for the corresponding knowledge point is removed; if the logical chain integrity score in two consecutive micro-test questions does not recover, a learning report is generated and pushed to the teacher.

9. 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 1-8.

10. 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 1-8.

Images