An intelligent review intervention system and method based on dual-engine judgment and cognitive drift detection

Abstract

This invention discloses an intelligent review intervention system and method based on dual-engine judgment and cognitive drift detection. The system includes an interval repetition scheduler, a cognitive drift detector, and a learner profile storage. The interval repetition scheduler uses a first engine and a second engine in conjunction to determine the timing of review: the first engine updates the ease factor and interval days of the review cards based on the recall quality score; the second engine overrides the first engine's judgment and forces review when the current mastery of a knowledge point is below a threshold. The cognitive drift detector monitors changes in students' knowledge status, detects cognitive drift patterns and generates warnings, adding the corresponding knowledge points to the priority review queue; when a high-severity drift is detected, an active clarification dialogue is triggered, and the profile is revised based on student feedback. The learner profile storage includes historical mastery records for each knowledge point, the last practice time, and answer records. This invention is applicable to intelligent education.

Description

Technical Field

[0001] This invention relates to the field of intelligent review intervention in intelligent education, specifically to an intelligent review intervention system and method based on dual-engine judgment and cognitive drift detection. Background Technology

[0002] With the increasing application of artificial intelligence technology in education, intelligent review systems have become an important tool to help students combat forgetting and consolidate knowledge. However, existing intelligent review systems have the following technical shortcomings:

[0003] Deficiency 1: Review scheduling relies solely on user self-assessment, lacking objective cognitive state evaluation. Existing spaced repetition systems (such as Anki and SuperMemo) use the SM-2 algorithm, where the determination of review timing depends entirely on students' subjective self-assessment (0-5 points) of the quality of their recall of review cards. When students overestimate or underestimate their own mastery, the review plan will deviate from their actual cognitive state—overestimation leads to delayed review and accelerated forgetting, while underestimation leads to over-review and inefficiency. The system cannot proactively detect students' actual cognitive decline and can only passively wait for users' self-assessment input.

[0004] Defect 2: The knowledge tracing model and the review scheduling system are completely disconnected, failing to form a closed loop. While knowledge tracing models such as Bayesian Knowledge Trace (BKT) can objectively assess the probability of students mastering each knowledge point, BKT is only used for visualizing knowledge status or recommending questions, and has never been used to drive the review scheduling system. There is no data path between the output of the knowledge tracing model (mastery probability) and the input of the interval repetition scheduler (recall quality score). The independent operation of the two prevents the review scheduling from utilizing objective mastery assessment results, resulting in a waste of cognitive state data.

[0005] Deficiency 3: The BKT model ignores forgetting, leading to inflated mastery assessments. The standard BKT model treats a student's knowledge status as a binary variable (learned / unlearned). Once the knowledge status shifts to "learned," it remains so indefinitely, without considering the impact of time on memory retention. When students haven't practiced a particular knowledge point for a long time, even if their historical mastery is high, their actual mastery level may significantly decline due to forgetting. However, the BKT model fails to reflect this decline process, resulting in inflated mastery assessments and a lack of targeted follow-up learning plans.

[0006] Deficiency 4: Lack of effective detection and proactive intervention mechanisms for students' cognitive abnormalities. Current systems monitor cognitive states only at the "accuracy rate" level, failing to identify deeper cognitive abnormalities—such as "contradictory drift" (suddenly forgetting something), "interest drift" (forgetting previously learned knowledge) due to lack of practice, and "conceptual confusion" (understanding the concept but not being able to use it)—where the concept is understood but applied incorrectly. More importantly, existing systems typically silently overwrite historical data or merely record abnormalities, never proactively initiating clarification dialogues with students to uncover the causes. This leads to accumulated misjudgments and a continued deviation from the learning path.

[0007] Defect 5: The learner profile is too narrowly dimensional to support multi-dimensional cognitive anomaly detection. Existing systems typically only maintain data on two dimensions: students' learning history (answer records) and proficiency in each knowledge point. They lack comprehensive collection and integration of multi-dimensional data such as cognitive ability, learning behavior, emotional attitude, and physiological rhythms. This results in a lack of multi-source data support for cognitive anomaly detection, leading to insufficient accuracy and interpretability of the detection results. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the prior art and provide an intelligent review intervention system and method based on dual-engine judgment and cognitive drift detection, so as to overcome the technical defects of the prior art, such as lack of objective cognitive state assessment in review scheduling, separation of knowledge tracking and review scheduling, BKT ignoring forgetting, lack of proactive intervention for cognitive abnormalities, and single profile dimension.

[0009] The present invention adopts the following technical solution to achieve the above objectives. In the first aspect, the present invention provides an intelligent review intervention system based on dual-engine judgment and cognitive drift detection, including an interval repetition scheduler, a cognitive drift detector, and learner profile storage.

[0010] The interval repetition scheduler is used to manage review cards for each knowledge point and to perform review judgments in a dual-engine linkage.

[0011] The data structure of the review cards includes the following fields: topic_id (knowledge point identifier), topic_name (knowledge point name), ease_factor (simplicity factor, initial default value is 2.5), interval_days (current interval days), repetitions (number of consecutive correct answers), next_review_date (next review date, ISO format), last_review_date (last review date), last_quality (last recall quality score, 0-5 points), mastery_at_review (BKT mastery value at the time of review), and a complete historical record list.

[0012] The interval repeat scheduler adopts a dual-engine architecture:

[0013] The first engine (SM-2 interval repetition engine): Based on the student's recall quality score (0-5 points) during review, it updates the ease factor, interval days, and next review date of the review card. Specifically: when the recall quality score is less than a third preset threshold of 3, the consecutive correct answers are reset to 0, and the interval days are set to 1 day; when the recall quality score is greater than or equal to 3, the consecutive correct answers are incremented by 1, and the interval days are updated to the value of the consecutive correct answers multiplied by the ease factor. The ease factor is updated as follows: the new ease factor is equal to the sum of 0.1 and the old ease factor minus a first product, the first product is 5 minus the recall quality score multiplied by a first coefficient, the first coefficient is 0.08 plus a second product, and the second product is 5 minus the recall quality score multiplied by 0.02; the lower limit of the new ease factor is 1.3.

[0014] The second engine (BKT mastery monitoring engine) tracks the current BKT mastery probability of each knowledge point in real time. When the current mastery probability falls below a first preset threshold (preferably 0.7), a forgetting risk signal is generated, overriding the interval determination result of the first engine. The corresponding knowledge point is then forcibly marked as needing review, triggering a review reminder. This "overriding" linkage mechanism enables proactive intervention of BKT mastery probability on the SM-2 time interval—even if the SM-2 interval has not yet expired, once the mastery probability decays below the threshold, the system immediately triggers review, effectively combating the Ebbinghaus forgetting curve.

[0015] The interval repetition scheduler also provides an expired review retrieval method, which iterates through all review cards and checks two conditions: condition one is that the current date is later than or equal to the next review date (SM-2 interval has expired), and condition two is that the current BKT mastery probability is lower than a first preset threshold (mastery decay triggered). Cards that meet either condition are added to the review list and sorted by urgency—those with the lowest mastery probability are given priority.

[0016] The cognitive drift detector is used to monitor changes in students' knowledge status, detect at least one cognitive drift pattern, generate a drift warning when cognitive drift is detected, and send a trigger signal to the interval repetition scheduler to add the corresponding knowledge point to the priority review queue.

[0017] The at least one cognitive drift pattern detected by the cognitive drift detector includes:

[0018] Mastery Conflict Detection Mode: This mode records historical mastery snapshots for each knowledge point. The cognitive drift detector continuously records BKT mastery probability snapshots (including timestamps and mastery values) for each knowledge point after each update. When the current mastery level decreases more than a second preset threshold (preferably 30%) relative to the most recent historical snapshot, it is determined to be a mastery conflict drift, generating a first-type drift warning. When the decrease exceeds a sixth preset threshold (preferably 50%), the corresponding drift warning is marked as a high-severity warning, triggering an active clarification dialogue process. This detection mechanism identifies the abnormal state of "suddenly not understanding"—students who previously had good mastery but whose recent performance has dropped sharply, and the system actively inquires about the reasons instead of silently overwriting historical records.

[0019] Interest Shift Detection Mode: This mode periodically scans the last practice time for each knowledge point. When a knowledge point has not been practiced for more than a preset number of days (preferably 14 days) and its historical mastery level is higher than a third preset threshold (preferably 0.6), it is determined to be an interest shift drift, generating a second type of drift warning. The cognitive drift detector adds the corresponding knowledge point to the priority review queue of the interval repetition scheduler and generates a gentle reminder to inform students that the knowledge point has not been touched for a long time. This detection mechanism identifies the risk of forgetting due to "long-term neglect," preventing highly mastered knowledge points from being naturally forgotten due to prolonged inactivity.

[0020] Conceptual Confusion Detection Mode: This mode analyzes the question type distribution in students' recent answer records. The cognitive drift detector categorizes answer records by question type (conceptual questions, calculation questions, application questions, etc.) and calculates the accuracy rate for each type. When the accuracy rate for the first type of question (conceptual questions) exceeds a fourth preset threshold (preferably 70%) and the number of consecutive incorrect answers for the second type of question (calculation or application questions) exceeds a fifth preset threshold (preferably 3 times), it is determined to be conceptual confusion drift—the student understands the meaning of the concept but is confused about the specific application steps. A third type of drift warning is generated, and the corresponding knowledge point is routed to step-by-step practice, pushing a problem-solving demonstration question instead of more complete questions of the same type. This detection mechanism identifies deep cognitive problems of "understanding but not knowing how to use."

[0021] The cognitive drift detector is also used to trigger an active clarification dialogue process when high-severity drift is detected. Specifically, the active clarification dialogue process involves the cognitive drift detector generating a clarification dialogue request, which includes a first question and multiple candidate feedback options. These options include a first option indicating forgetting, a second option indicating a different question type, and a third option indicating the need for review. The detector receives the student's selected feedback option and determines the type of drift cause based on it: when the drift cause is the first type (forgetting), the corresponding knowledge point is immediately added to the priority review queue of the interval repetition scheduler; when the drift cause is the second type (different question type), specific practice for the corresponding question type is pushed; and when the drift cause is the third type (needs review), review materials for that knowledge point are pushed. This mechanism upgrades "passive monitoring" to "active interaction," revising the student profile based on the clarification feedback received, rather than silently overwriting historical records.

[0022] The learner profile storage is used to store multi-dimensional profile data of students. This multi-dimensional profile data includes at least the historical mastery records for each knowledge point, the last practice time for each knowledge point, and the answer records. The learner profile storage adopts a three-layer data separation architecture:

[0023] Real-time data layer: Used to store direct measurement data collected in real time through tracking points, including answer records (including question type, correctness, and answer time), learning duration, number of times each knowledge point was practiced, emotion tags and cognitive load levels output by the WCCE engine, and other directly quantifiable data, which are written to the profile storage in real time.

[0024] The proxy variable layer stores indirect measurement data inferred through proxy variables. Proxy variable inference methods include: using the accuracy rate of reasoning questions to represent logical reasoning ability; using average answer time to represent information processing speed; using the redo rate of incorrect questions to represent error handling methods; using the distribution of learning time for each subject to represent time allocation patterns; and using the rate of improvement in accuracy with the number of practice sessions to represent learning efficacy. The proxy variable layer data is updated periodically, supplementing profile dimensions that cannot be directly collected by the real-time data layer.

[0025] The manual calibration layer stores periodic calibration data obtained through student self-assessment or teacher input, including student metacognitive self-assessment reports (including self-assessment of learning attitude, self-report of difficulties, and self-assessment of learning goals) and teacher-entered learning attitude evaluations and classroom performance observation records. The data in the manual calibration layer is updated at a low frequency (e.g., weekly or monthly) to periodically calibrate and correct the data in the real-time data layer and the proxy variable layer.

[0026] The multi-dimensional profile data includes data from seven dimensions: cognitive ability (including knowledge structure graph, information processing speed, and logical reasoning ability), learning behavior (including deep learning ratio, time allocation pattern, and error handling method), emotional attitude (including emotional history and learning efficacy), professional ability (including core skills matrix and cross-domain transferability), development environment (including resource accessibility score), circadian rhythm (including learning time preference and daytime fluctuation pattern), and development potential (including growth trajectory and progress rate). All data is persisted to the student profile directory in JSON format, supporting incremental updates and version control.

[0027] Secondly, this invention provides an intelligent review intervention method based on dual-engine judgment and cognitive drift detection, comprising the following steps:

[0028] Step S1: The interval repetition scheduler obtains the review cards and current mastery level of each knowledge point.

[0029] Step S2: The interval repetition scheduler performs the first engine determination: based on the next review date in the review card, it determines whether the current date is the right time to review.

[0030] Step S3: The interval repetition scheduler executes the second engine judgment: determining whether the current mastery level of each knowledge point is lower than the first preset threshold.

[0031] Step S4: When the current date reaches the next review date or the current mastery level is lower than the first preset threshold, the corresponding knowledge point is marked as needing review and a review task is generated.

[0032] Step S5: The cognitive drift detector monitors changes in mastery of each knowledge point, the last practice time, and answer records to detect cognitive drift patterns. Specifically, this includes: performing mastery discrepancy detection, recording historical mastery snapshots of each knowledge point, generating a first-type drift warning when the current mastery decreases by more than 30% compared to the most recent historical snapshot, and marking it as a high-severity warning when it exceeds 50%; performing interest shift detection, scanning the last practice time of each knowledge point, generating a second-type drift warning when there has been no practice for more than 14 days and the historical mastery is higher than 0.6, and adding the corresponding knowledge point to the priority review queue; performing concept confusion detection, analyzing the question type distribution of students' recent answer records, generating a third-type drift warning when the accuracy rate for concept questions is higher than 70% and calculation or application questions are answered incorrectly more than 3 times consecutively, and routing the corresponding knowledge point to step-by-step breakdown practice.

[0033] Step S6: When cognitive drift is detected, generate a drift warning.

[0034] Step S7: When the severity of the drift warning exceeds a preset severity threshold, a proactive clarification dialogue process is triggered. A clarification dialogue request is sent to the student, the student's clarification feedback is received, and the student profile is revised based on the clarification feedback. The proactive clarification dialogue process specifically involves: generating a clarification dialogue request, sending the student an inquiry statement and three candidate feedback options ("May have forgotten," "Different question types," "Needs to review"); receiving the student's selected feedback option; when the student selects the "forgotten" option, the corresponding knowledge point is immediately added to the priority review queue; when the student selects the "different question types" option, specific practice exercises corresponding to that question type are pushed; when the student selects the "needs to review" option, review materials for that knowledge point are pushed.

[0035] Step S8: Add the knowledge points with drift warnings to the priority review queue and execute the review task through the interval repetition scheduler.

[0036] Step S9: The learner profile storage adopts a three-layer data separation architecture to store multi-dimensional profile data.

[0037] The technical solution described in this invention has the following advantages compared with the prior art:

[0038] This invention employs a dual-engine, "overall" linkage to achieve review scheduling based on actual cognitive states. It deeply integrates the SM-2 interval repetition algorithm with the BKT mastery probability decay mechanism, using a BKT mastery probability threshold (0.7) to cover the SM-2 time interval determination—even if the interval hasn't expired, review is forcibly triggered once mastery decays below the threshold. Compared to the SM-2 algorithm, which relies solely on user subjective self-assessment, this invention upgrades the review scheduling mechanism from "subjectively driven" to "objectively driven," effectively combating the Ebbinghaus forgetting curve.

[0039] For the first time, knowledge tracking and review scheduling achieve a closed-loop linkage. This invention breaks down the long-standing technical barrier between the Knowledge Tracking Model (BKT) and the Spaced Repetition System (SM-2), establishing a direct data pathway from the BKT mastery probability to the review scheduler. The output of BKT is no longer merely used for visualization, but serves as the core signal driving review behavior, realizing a closed loop between knowledge status assessment and review behavior intervention, and solving the technical problem of wasted cognitive state data in existing technologies.

[0040] Three types of cognitive drift detection enable multi-dimensional anomaly monitoring. This invention is the first to introduce the concept of "concept drift" detection from the field of machine learning into the field of educational cognitive monitoring, and defines three specific and quantifiable drift modes for educational scenarios: mastery contradiction detection (a decrease of >30% triggers an alert, >50% triggers a high-severity alert), interest shift detection (more than 14 days without practice and mastery >0.6 triggers an alert), and concept confusion detection (a correct answer rate of >70% for concept questions and >3 consecutive incorrect answers to calculation questions triggers an alert), realizing an upgrade from single accuracy monitoring to multi-dimensional cognitive anomaly monitoring.

[0041] The proactive clarification dialogue mechanism represents a leap from passive monitoring to proactive intervention. When high-severity cognitive drift is detected, this invention does not silently overwrite historical records, but proactively initiates a clarification dialogue, offering students three options: "forgotten / different question type / needs to review." Based on student feedback, differentiated intervention strategies are implemented (prioritized review / specific practice / review). This mechanism upgrades the system from a "passive monitor" to a "proactive intervener," effectively preventing learning path deviations caused by algorithmic misjudgments.

[0042] The seven-dimensional, three-layer separation profile provides multi-source data support for cognitive drift detection. This invention's learner profile covers seven dimensions: cognitive ability, learning behavior, emotional attitude, professional skills, developmental environment, physiological rhythms, and developmental potential. It employs a three-layer data separation architecture—real-time data acquisition, proxy variable calculation, and manual calibration—to provide comprehensive, accurate, and interpretable data support for cognitive drift detection, thereby improving the accuracy and robustness of the detection. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the overall architecture of an intelligent review intervention system based on dual-engine judgment and cognitive drift detection provided by the present invention. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0045] Example 1: System Overall Architecture

[0046] Reference Figure 1 The intelligent review intervention system based on dual-engine judgment and cognitive drift detection described in this invention includes an interval repetition scheduler, a cognitive drift detector, and learner profile storage.

[0047] The interval repetition scheduler includes a review card management module, an SM-2 algorithm engine, a BKT monitoring engine, and an overdue review retrieval module. The review card management module maintains the review card data structure for each knowledge point; the SM-2 algorithm engine executes interval update logic based on recall quality scores; the BKT monitoring engine tracks the BKT mastery probability of each knowledge point in real time, generating a forgetting risk signal when the mastery rate falls below a threshold of 0.7 and overriding the SM-2 time determination; the overdue review retrieval module iterates through all cards and returns a list of knowledge points that meet the condition of "SM-2 expired or BKT < 0.7" and require review.

[0048] The cognitive drift detector comprises a mastery contradiction detection module, an interest shift detection module, a concept confusion detection module, a drift warning management module, and a proactive clarification dialogue module. The three detection modules respectively identify three types of drift patterns; the drift warning management module maintains an active warning list and supports time window auditing; the proactive clarification dialogue module generates dialogue requests and processes student feedback when a high-severity warning is triggered.

[0049] The learner profile storage includes a HOT data layer (real-time data collection), a PROXY data layer (data inferred from proxy variables), and a MANUAL data layer (data manually calibrated), which together constitute a complete data view of the seven-dimensional profile.

[0050] Data flow: The BKT tracking module provides real-time mastery data to the BKT monitoring engine of the interval repetition scheduler; the interval repetition scheduler provides mastery snapshots and last practice time to the cognitive drift detector; the cognitive drift detector sends drift warning signals to the interval repetition scheduler and adds knowledge points to the priority review queue; the learner profile storage provides multi-dimensional profile data to the cognitive drift detector to support various detection methods; after proactive clarification, the cognitive drift detector writes the corrected results back to the learner profile storage.

[0051] Example 2: Review Card Data Structure

[0052] In one specific embodiment of the present invention, the data structure of the review card is shown in Table 1 below.

[0053] Table 1. Data Structure of Review Cards

[0054]

[0055] In the initial state, the ease_factor=2.5, interval_days=1, repetitions=0, and next_review_date is the current date plus 1 day for the new knowledge point.

[0056] Example 3: SM-2 Algorithm Update Process

[0057] The update process of the SM-2 algorithm engine of the interval repeat scheduler 100 described in this invention includes the following steps:

[0058] Step A01: Obtain the recall quality score for this review, ranging from 0 to 5 points. The scoring criteria are as follows: 5 points indicates complete correctness and confidence; 4 points indicates correctness but slight hesitation; 3 points indicates partial correctness but significant forgetting; 2 points indicates incorrect but with some recollection; 1 point indicates complete forgetting; and 0 points indicates no answer.

[0059] Step A02: Read the ease_factor, interval_days, and repetitions of the current review card.

[0060] Step A03: Determine if quality is less than 3. If quality < 3, proceed to step A04; if quality ≥ 3, proceed to step A05.

[0061] Step A04 (Reset Path): Reset repetitions to 0, set interval_days to 1 day, and proceed to step A06.

[0062] Step A05 (Incremental Path): Increment repetitions by 1, and update interval_days to repetitions × ease_factor.

[0063] Step A06: Update ease_factor. The calculation formula is:

[0064] ease_factor_new = ease_factor_old + (0.1 - (5 - quality) × (0.08 + (5 - quality) × 0.02))

[0065] If ease_factor_new < 1.3, then ease_factor_new = 1.3.

[0066] Step A07: Calculate the next review date: next_review_date = current date + interval_days.

[0067] Step A08: Update the card fields: last_review_date = current date, last_quality = quality, mastery_at_review = current BKT mastery level. Append this review record to the history list.

[0068] In a specific application example of this invention, a student reviews the knowledge point of "logarithmic function" and receives a recall quality score of 4 (correct but slightly hesitant). Currently, ease_factor = 2.5, repetitions = 2, interval_days = 5. Since quality = 4 ≥ 3, repetitions become 3, and interval_days = 3 × 2.5 = 7.5 days. ease_factor_new = 2.5 + (0.1 - (5-4) × (0.08 + (5-4) × 0.02)) = 2.5 + (0.1 - 1 × 0.1) = 2.5. The next review date is the current date plus 7.5 days (approximately 8 days).

[0069] In another application example, a student reviews the knowledge point of "trigonometric functions" and receives a recall quality score of 2 (incorrect but remembered). Since quality = 2 < 3, repetitions is reset to 0, and interval_days is reset to 1 day. ease_factor_new = 2.5 + (0.1 - (5-2) × (0.08 + (5-2) × 0.02)) = 2.5 + (0.1 - 3 × 0.14) = 2.5 - 0.32 = 2.18. The next review date is the current date plus one day.

[0070] Example 4: Dual-engine linkage determination process

[0071] The dual-engine linkage determination process of the interval repeat scheduler described in this invention includes the following steps:

[0072] Step B01: Obtain the review cards for the target knowledge points, and extract next_review_date and the current BKT mastery level m.

[0073] Step B02: Execute the first engine (SM-2 interval judgment): Determine if the current date is greater than or equal to next_review_date. If true, mark the result of the first engine as True (requires review); otherwise, mark it as False.

[0074] Step B03: Execute the second engine (BKT mastery assessment): Determine if the current BKT mastery level m is less than the first preset threshold of 0.7. If m < 0.7, mark the second engine result as True (requires review); otherwise, mark it as False.

[0075] Step B04: Perform a coverage-based logical judgment: If the result of the first engine is True or the result of the second engine is True, mark the corresponding knowledge point as needing review; otherwise, mark it as not needing review.

[0076] Key mechanism: When the result of the second engine is True, even if the result of the first engine is False (i.e., the SM-2 interval has not yet expired), the judgment result of the BKT monitoring engine overrides the time judgment of SM-2 - the system forces the knowledge point to be added to the review list.

[0077] Step B05: Add the knowledge points marked as needing review to the due review list, sorting them from lowest to highest level of mastery (those with the lowest level of mastery are the most urgent).

[0078] Step B06: If the second engine is triggered (i.e., m < 0.7), publish the FORGETTING_RISK (Forgot Risk Warning) event to the capability event bus.

[0079] In a specific application example of this invention, a student's review card for the "quadratic function" knowledge point shows next_review_date as 2026-08-05 (the current date is 2026-07-23, not yet due), and the current BKT mastery level is 0.62 (<0.7). The first engine result is False (not due), and the second engine result is True (mastery level is below the threshold). The overlay logic takes effect—this knowledge point is forcibly marked as needing review, added to the due review list, and a forgetting risk warning event is issued. Although there are still 13 days until the review time scheduled by SM-2, the system detects that the mastery level has dropped below the threshold and proactively triggers the review in advance.

[0080] In another application example, the student's next_review_date for the "Derivatives" knowledge point is 2026-07-20 (3 days past the expiration date), and their current BKT mastery level is 0.82 (≥0.7). The first engine result is True (expired), and the second engine result is False (mastery level is above the threshold). The first engine takes effect—this knowledge point is marked as needing review due to the expiration of the interval.

[0081] Example 5: Mastering the Conflict Detection Process

[0082] The detection process of the mastery contradiction detection module of the cognitive drift detector described in this invention includes the following steps:

[0083] Step C01: Retrieve a list of historical snapshots of mastery levels for the target knowledge points from the learner profile storage. Each time the BKT is updated, the cognitive drift detector automatically records a snapshot in the format {timestamp, mastery_value}.

[0084] Step C02: Obtain the current mastery value m_current (latest BKT output) and the mastery value m_old from the most recent historical snapshot.

[0085] Step C03: Calculate the reduction delta = (m_old - m_current) / m_old×100%.

[0086] Step C04: Determine if delta is greater than the second preset threshold of 30%. If delta ≤ 30%, it is determined to be a non-contradictory drift, and the detection ends.

[0087] Step C05: If delta > 30%, it is determined to be a mastery degree conflict drift, and a first-type drift warning is generated.

[0088] Step C06: Further determine whether delta is greater than the sixth preset threshold of 50%. If delta ≤ 50%, mark it as a normal severity warning; if delta > 50%, mark it as a high severity warning.

[0089] Step C07: Publish the COGNITIVE_DRIFT event to the capability event bus. The event data includes: alert_type="contradiction", severity (low / medium / high), old_mastery, new_mastery, delta, and suggested_clarification (suggested clarification language). High severity alerts trigger an active clarification dialogue process.

[0090] In a specific application example of this invention, a student's BKT mastery of the "sequences" knowledge point plummeted from 0.85 to 0.52 within 10 days, a decrease of (0.85-0.52) / 0.85 ≈ 38.8% > 30% but ≤ 50%. The cognitive drift detector determined this to be a mastery discrepancy drift (moderate severity), generated an alert, and issued an event.

[0091] In another application example, a student's BKT mastery of the "vector" knowledge point plummeted from 0.78 to 0.28 within 3 days, a decrease of (0.78-0.28) / 0.78 ≈ 64.1% > 50%. The cognitive drift detector identified this as a contradictory drift in mastery (high severity), generated an alert, and issued an event, triggering a proactive clarification dialogue process—the system asked the student: "You previously had a very good grasp of the 'vector' knowledge point, but your performance on recent tests has declined. What happened?"

[0092] Example 6: Interest Transfer Detection Process

[0093] The detection process of the interest shift detection module of the cognitive drift detector described in this invention includes the following steps:

[0094] Step D01: Retrieve the last practice time (last_practice_date) and historical mastery level (m_historical) for each knowledge point from the learner profile storage.

[0095] Step D02: Calculate the difference in days between the current date and last_practice_date, gap_days.

[0096] Step D03: Determine if gap_days is greater than the preset number of days, 14 days. If gap_days ≤ 14, it is determined that there is no risk of interest transfer, and the detection ends.

[0097] Step D04: If gap_days > 14, determine if the historical mastery level m_historical is greater than the third preset threshold of 0.6. If m_historical ≤ 0.6, the knowledge points with low mastery are considered to be naturally idle and no warning is generated (it is normal for knowledge points with low mastery not to be practiced).

[0098] Step D05: If m_historical > 0.6, it is determined to be interest shift drift, and a second type of drift warning is generated. This warning indicates that "students may have forgotten knowledge points that they once mastered well due to long-term lack of exposure."

[0099] Step D06: Generate a gentle reminder message, such as: "You haven't practiced 'knowledge point names' for 14 days. Want to warm up with a simple question?"

[0100] Step D07: Add the corresponding knowledge point to the priority review queue of the interval repetition scheduler to ensure that the knowledge point is given priority in the recent review tasks.

[0101] In a specific application example of this invention, a student's last practice session on the "trigonometric functions" knowledge point was on July 1, 2026, and the current date is July 23, 2026. The gap days are 22 days, which is greater than 14 days, and the historical mastery level is 0.75, which is greater than 0.6. The cognitive drift detector determines this as a risk of interest shift and generates a gentle reminder: "You haven't practiced 'trigonometric functions' for 22 days. Would you like to warm up with a simple question?" Simultaneously, this knowledge point is added to the priority review queue.

[0102] Example 7: Concept Confusion Detection Process

[0103] The detection process of the concept confusion detection module of the cognitive drift detector described in this invention includes the following steps:

[0104] Step E01: Retrieve the student's answer records for the 10 most recent questions on the target knowledge point from the learner profile storage, including the question type (conceptual question / calculation question / application question) and the correctness indicator for each question.

[0105] Step E02: Calculate the accuracy rate (rate_concept) for concept questions and the answer records for calculation questions and / or application questions.

[0106] Step E03: Determine if rate_concept is greater than the fourth preset threshold of 70%. If rate_concept ≤ 70%, it is determined that there is no conceptual confusion (the accuracy rate of the conceptual questions is not high, indicating that the students have not mastered the concepts), and the test ends.

[0107] Step E04: If rate_concept > 70%, determine whether there are cases where the number of consecutive incorrect answers to calculation or application problems exceeds the fifth preset threshold of 3 times.

[0108] Step E05: If there are no more than 3 consecutive incorrect answers to calculation / application questions, it is determined that there is no conceptual confusion, and the test ends.

[0109] Step E06: If there are 4 or more consecutive (>3 times) incorrect answers to calculation or application problems, it is judged as conceptual confusion drift - the student understands the meaning of the concept but is confused about the specific application steps, and a third type of drift warning is generated.

[0110] Step E07: Route the corresponding knowledge point to the step-by-step practice, and push the problem-solving process demonstration questions instead of pushing more complete questions of the same type.

[0111] In a specific application example of this invention, a student's recent answer record on the "Trigonometric Identities" knowledge point is as follows: all 5 conceptual questions were answered correctly (concept = 100%), while all 4 calculation questions were answered incorrectly (number of consecutive errors = 4 > 3). The cognitive drift detector determined this to be conceptual confusion—the student understood the formula for sin(A+B), but confused the steps when substituting the angles for calculation. The system generated a conceptual confusion warning, routed the knowledge point to a step-by-step decomposition exercise, and pushed a guided exercise that breaks down the expansion process of sin(75°) into four steps: "memorizing the formula → substituting the angle → simplifying → verifying the result".

[0112] Example 8: Proactive Clarification Dialogue Process

[0113] The active clarification dialogue module of the cognitive drift detector described in this invention includes the following steps:

[0114] Step F01: Receive a high-severity drift warning (originating from a decrease of >50% in mastery contradiction detection, or a long-term lack of practice of highly mastered knowledge points in interest shift detection, or severe confusion in concept confusion detection), and obtain the corresponding student_id and topic_id.

[0115] Step F02: Generate a clarification dialogue request. The clarification dialogue request includes a first question and three candidate feedback options. The first question is: "You used to have a good grasp of the knowledge point '{topic_name}', but recently your performance has declined / you haven't practiced for a long time / you make mistakes when doing questions. What do you think is the reason?" The three candidate feedback options are: Option A - "You may have forgotten it", Option B - "The question types are different", and Option C - "You need to review it again".

[0116] Step F03: Send a clarification dialogue request to the student and wait for the student's response.

[0117] Step F04: Receive feedback options selected by the student.

[0118] Step F05: Determine the selection result. If option A (forgotten), proceed to step F06; if option B (different question type), proceed to step F07; if option C (needs to be reviewed), proceed to step F08.

[0119] Step F06 (Forgetting Path): Immediately add the corresponding knowledge point to the priority review queue of the interval repetition scheduler 100, reset the review card to the initial state (repetitions=0, interval_days=1), and record the drift reason as "forgetting".

[0120] Step F07 (Different Question Types): Call the intelligent question generation module to push the corresponding practice questions (5-8 variations of the same question type) for this question type, and record the drift reason as "question_type_mismatch".

[0121] Step F08 (Path needs to be reviewed): Call the knowledge base retrieval module, push the revised materials for this knowledge point (including concept explanations, formula derivations and typical examples), and record the reason for drift as "need_review".

[0122] Step F09: Revise the student profile based on clarification feedback—write the reason for the drift into the "Cognitive Anomaly Record" field of the profile for subsequent analysis. If the student selects "forgotten" and their mastery improves after review, the accuracy of the drift warning is verified; if the student selects "different question types" and their accuracy improves after targeted practice, the diagnosis of conceptual confusion is confirmed. In either case, the system does not silently overwrite historical records but retains the complete data chain of drift warnings, clarification dialogues, and student feedback.

[0123] In a specific application example of this invention, a student's BKT mastery of the "vector" knowledge point plummeted from 0.78 to 0.28 within 3 days (a drop of 64.1% > 50%), triggering a high-severity mastery discrepancy warning. The proactive clarification dialogue module sent a question to the student: "You previously had a very good grasp of the 'vector' knowledge point, but recently your performance on questions has declined significantly. What do you think is the reason?" along with three option buttons.

[0124] The student selected option B – “Different question types.” Based on this, the system determined that the student had not completely forgotten vector knowledge, but rather encountered unfamiliar vector application problems (such as combining spatial vectors with solid geometry). The system then provided five targeted practice problems on vector application. After completing the practice, the student's accuracy rate reached 80%, and their BKT mastery level rebounded to 0.69. The system stored the complete chain of this drift event (detection → warning → clarification → feedback → correction → recovery) in the student's profile's “cognitive anomaly record,” forming a traceable intervention loop.

[0125] Example 9: Due Review Retrieval and Priority Queue

[0126] The retrieval process of the expiration review retrieval module of the interval repeat scheduler described in this invention is as follows:

[0127] Step G01: Obtain a list of review cards for all knowledge points for the target student, with a total of N cards.

[0128] Step G02: Initialize the review list due_list to be empty.

[0129] Step G03: Traverse each review card and perform dual-engine judgment for each card (as described in Example 4).

[0130] Step G04: If the card is marked as needing review (triggered by the first or second engine), add the card to due_list.

[0131] Step G05: After the traversal is complete, sort the due_list according to the following rules: First, sort by BKT mastery level from low to high (the lowest mastery level is the most urgent). If the mastery levels are the same, sort by next_review_date from early to late (the earliest due date is reviewed first).

[0132] Step G06: Return the sorted due_list.

[0133] Step G07: If the length of due_list exceeds the preset maximum number of returns (default 5), only the top 5 most urgent knowledge points will be returned.

[0134] In a specific application example of the present invention, the situation of a student's review cards is shown in Table 2 below.

[0135] Table 2. A student's review cards

[0136]

[0137] `due_list = [Logarithmic Functions (0.62), Derivatives (0.45), Sequences (0.55)]`. Sorted by mastery level, it is: [Derivatives (0.45), Sequences (0.55), Logarithmic Functions (0.62)]. The top 3 most urgent knowledge points are returned for student review. Although there are still 13 days until the scheduled date of SM-2 (August 5, 2026), "Logarithmic Functions" is forcibly triggered for review because the BKT mastery level has dropped to 0.62 < 0.7.

[0138] Example 10: Three-layer data separation for a seven-dimensional learner profile

[0139] The three-layer data separation architecture for storing learner profiles described in this invention is as follows:

[0140] Real-time data layer:

[0141] Data source: Real-time data collection from system embedded points;

[0142] Data content: Answer records (including question type, correctness, answer time, answer duration), learning duration (cumulative learning time for each knowledge point), number of practice sessions, emotion tags (POSITIVE / NEGATIVE / NEUTRAL) output by the WCCE engine, cognitive load level (HIGH / MODERATE / LOW), and historical snapshots of BKT mastery of each knowledge point;

[0143] Update frequency: Real-time (written as soon as a piece of data is generated);

[0144] Storage format: JSON file, stored in chunks by date.

[0145] Proxy variable layer:

[0146] Data source: Calculated through proxy variables, or by timed or event-triggered calculations.

[0147] Proxy variable mapping rules:

[0148] Information processing speed = average answering time (seconds) for the last 10 questions, reverse indicator (the shorter the time, the faster the speed);

[0149] Logical reasoning ability = Correct answer rate (%) of the most recent 20 reasoning questions;

[0150] Error handling method = Correct rate of redoing wrong questions (number of correct answers after redoing / total number of redoings × 100%)

[0151] Deep learning percentage = (Number of questions with a dwell time exceeding 30 seconds / Total number of questions) × 100%

[0152] Time allocation pattern = distribution of learning time for each subject / knowledge point;

[0153] Learning efficacy = the slope of accuracy as the number of practice sessions increases;

[0154] Update frequency: Recalculation is triggered every 10 new answer records;

[0155] Storage format: JSON file, including calculation time and data source tags;

[0156] Artificial calibration layer:

[0157] Data sources: student self-evaluation and teacher input;

[0158] Data content: Students' metacognitive self-assessment reports (self-assessment of learning attitude 1-5 points, self-report of difficulties, self-assessment of learning goals), teacher-entered evaluations of learning attitudes, and classroom performance observation records;

[0159] Update frequency: once a week (student self-assessment) or once a month (teacher entry);

[0160] Storage format: JSON file, containing entry time, entry identifier, and version number.

[0161] The seven-dimensional profile data is generated by aggregating data from three layers, as shown in Table 3 below.

[0162] Table 3 Dimensions and Data Sources

[0163]

[0164] All data is persisted in JSON format to the student_profiles directory, with each student having their own independent folder. Incremental updates and version control are supported (each update generates a new version, and the 10 most recent versions are retained).

[0165] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. An intelligent review intervention system based on dual-engine judgment and cognitive drift detection, characterized in that, include: The interval repetition scheduler uses the first engine and the second engine to work together to determine the timing of reviewing each knowledge point. The first engine is an interval repetition engine, which updates the simplicity factor, interval days, and next review date of the review cards based on the student's recall quality score during review. The second engine is a mastery monitoring engine, which obtains the current mastery of each knowledge point. When the current mastery is lower than a first preset threshold, a forgetting risk signal is generated and overwrites the judgment result of the first engine, forcibly marking the corresponding knowledge point as needing review and triggering a review reminder. A cognitive drift detector is used to monitor changes in students' knowledge status, detect at least one cognitive drift pattern, generate a drift warning when cognitive drift is detected, and send a trigger signal to the interval repetition scheduler to add the corresponding knowledge point to the priority review queue. The cognitive drift detector is also used to trigger an active clarification dialogue process when a high degree of drift is detected, obtain the student's clarification feedback, and correct the student profile based on the clarification feedback, rather than silently overwriting the history. Learner profile storage is used to store multi-dimensional profile data of students. The multi-dimensional profile data includes at least the historical mastery of each knowledge point, the last practice time and answer record of each knowledge point.

2. The intelligent review intervention system based on dual-engine judgment and cognitive drift detection according to claim 1, characterized in that, The scheduling logic of the first engine is as follows: When the recall quality score is less than the third preset threshold, the number of consecutive correct answers is reset to 0, and the interval in days is set to the first preset value. When the recall quality score is greater than or equal to the third preset threshold, the number of consecutive correct answers is incremented by 1, and the interval in days is updated to the value of the number of consecutive correct answers multiplied by the simplicity factor. The method for updating the simplicity factor is as follows: the new simplicity factor is equal to the sum of the second preset value and the simplicity factor. The second preset value is the first parameter minus the second difference. The second difference is the fourth preset threshold minus the memory quality score multiplied by the third parameter. The third parameter is the fourth preset threshold minus the memory quality score multiplied by the fourth parameter.

3. The intelligent review intervention system based on dual-engine judgment and cognitive drift detection according to claim 1, characterized in that, The specific method for the dual-engine linkage determination of the interval repeat scheduler is as follows: The interval repetition scheduler maintains the scheduling logic of the first engine, updates the ease factor and the interval number of days based on the recall quality score, and calculates the next review date; the recall quality score ranges from 0 to 5. The interval repetition scheduler maintains the monitoring logic of the second engine and tracks the current mastery of each knowledge point in real time. When the current mastery is lower than the first preset threshold, a forgetting risk signal is generated, which overrides the interval judgment result of the first engine and forcibly marks the corresponding knowledge point as needing review.

4. The intelligent review intervention system based on dual-engine judgment and cognitive drift detection according to claim 1, characterized in that, The at least one cognitive drift pattern detected by the cognitive drift detector includes: The mastery contradiction detection mode is used to record the historical mastery snapshot of each knowledge point. If the current mastery decreases more than the second preset threshold compared to the most recent historical snapshot, it is judged as mastery contradiction drift and a first type of drift warning is generated. The interest shift detection mode is used to periodically scan the last practice time of each knowledge point. If a knowledge point has not been practiced for more than a preset number of days and the historical mastery is higher than the third preset threshold, it is judged as interest shift drift and a second type of drift warning is generated. The concept confusion detection mode is used to analyze the question type distribution of students' recent answer records. If the correct rate of the first type of question is higher than the fourth preset threshold and the number of consecutive incorrect answers to the second type of question exceeds the fifth preset threshold, it is judged as concept confusion drift, and a third type of drift warning is generated.

5. The intelligent review intervention system based on dual-engine judgment and cognitive drift detection according to claim 4, characterized in that, In the mastery conflict detection mode, when the decrease exceeds the sixth preset threshold, the corresponding drift warning is marked as a high severity warning, triggering the proactive clarification dialogue process.

6. The intelligent review intervention system based on dual-engine judgment and cognitive drift detection according to claim 1, characterized in that, The proactive clarification dialogue process is as follows: The cognitive drift detector generates a clarification dialogue request, which includes a first question and multiple candidate feedback options. Receive feedback options selected by students and determine the type of drift cause based on the feedback options; When the cause of the drift is of the first type, the corresponding knowledge point is immediately added to the priority review queue of the interval repetition scheduler; When the cause of the drift is the second type, specific practice questions for the corresponding question type will be pushed. When the cause of the drift is the third type, a revised version of the material for that knowledge point will be pushed out.

7. The intelligent review intervention system based on dual-engine judgment and cognitive drift detection according to claim 6, characterized in that, The multiple candidate feedback options include a first option indicating forgetting, a second option indicating a different question type, and a third option indicating that the question needs to be reviewed again.

8. The intelligent review intervention system based on dual-engine judgment and cognitive drift detection according to claim 1, characterized in that, The learner profile storage adopts a three-layer data separation architecture, the three layers being: The real-time data layer is used to store direct measurement data collected in real time through embedded points; The proxy variable layer is used to store indirect measurement data inferred through proxy variables; The manual calibration layer is used to store periodic calibration data obtained through student self-assessment or teacher input.

9. The intelligent review intervention system based on dual-engine judgment and cognitive drift detection according to claim 8, characterized in that, The multi-dimensional profile data includes data from seven dimensions: cognitive ability, learning behavior, emotional attitude, professional ability, development environment, physiological rhythm, and development potential.

10. An intelligent review intervention method based on dual-engine decision-making and cognitive drift detection, applied to the intelligent review intervention system based on dual-engine decision-making and cognitive drift detection as described in any one of claims 1-9, characterized in that, Includes the following steps: Step S1: The interval repetition scheduler obtains the review cards and current mastery level for each knowledge point; Step S2: The interval repetition scheduler performs the first engine judgment: based on the next review date in the review card, it determines whether the current date is the right time to review; Step S3: The interval repetition scheduler executes the second engine judgment: determining whether the current mastery level of each knowledge point is lower than the first preset threshold; Step S4: When the current date reaches the next review date or the current mastery level is lower than the first preset threshold, mark the corresponding knowledge point as needing review and generate a review task; Step S5: The cognitive drift detector monitors the changes in mastery of each knowledge point, the last practice time, and the answer records to detect cognitive drift patterns; Step S6: When cognitive drift is detected, generate a drift warning; Step S7: When the severity of the drift warning exceeds the preset severity threshold, the active clarification dialogue process is triggered, a clarification dialogue request is sent to the student, the student's clarification feedback is received, and the student profile is corrected based on the clarification feedback. Step S8: Add the knowledge points with drift warnings to the priority review queue and execute the review task through the interval repetition scheduler.

Images