Queue processing method and system based on dynamic forgetting feature extraction

Abstract

This invention relates to the field of data processing technology and discloses a queue processing method and system based on dynamic forgetting feature extraction. The method includes: collecting objective question answer records, vocabulary memorization records, and subjective question speech sequences input from a terminal; extracting features from each modality; using objective question answer records as the main record; aligning the matchable subjective expression multidimensional scoring vectors and vocabulary memorization records according to user identifiers and timestamps; marking unmatched components with null values ​​to generate a multidimensional learning feature set; mapping the multidimensional learning feature set to a multi-level state skip list; calculating dynamic forgetting feature parameters and performing node transitions to generate a phased forgetting feature partial order set; calculating individual data density indices based on the partial order set; performing branch extraction to generate a personalized push content queue; collecting interactive feedback based on the push results; calculating state degradation compensation factors; and reverse-calibrating the boundary thresholds of the multi-level state skip list. This invention achieves cross-module state linkage and dynamic calibration of phased push content through the joint representation of time intervals, historical errors, and capability fluctuations.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a queue processing method and system based on dynamic forgetting feature extraction. Background Technology

[0002] In language training systems based on adaptive learning algorithms, dynamic forgetting feature extraction and state transition mechanisms are commonly used to manage learners' knowledge mastery status. This system records learners' answers in modules such as vocabulary tests and voice input. Based on the accuracy rate within a single module and a fixed review frequency threshold, it categorizes vocabulary knowledge points into green or red squares, representing mastery levels, and generates a personalized anti-forgetting push list accordingly. When learners lack sufficient recent practice data, the system automatically extracts low-accuracy questions from the platform's global big data for supplementary pushes to maintain the continuity of training content.

[0003] After completing intensive training sessions, trainees often find that questions marked as green in earlier learning stages frequently reappear in their push notification lists. On-site observation reveals that the vocabulary for these questions has been mastered through repeated practice and testing, but the system mistakenly identifies them as requiring re-reinforcement due to sparse recent test-taking records. Similarly, when trainees repeatedly mispronounce a word in the speech module, the corresponding objective questions for that word remain consistently marked in green, failing to automatically revert to red to trigger targeted reinforcement. This disconnect between the pushed content and the trainees' actual weaknesses results in a large amount of repetitive push notifications of already mastered content, while newly exposed errors go unaddressed. Trainees are forced to expend energy on ineffective practice within limited training time, significantly weakening the efficiency of anti-forgetting training and hindering knowledge consolidation. Therefore, reducing the oscillations in content push notification status during cross-module data integration is a pressing issue that needs to be addressed in this field. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide a queue processing method based on dynamic forgetting feature extraction. This method collects multimodal learning data and extracts features, combines dynamic forgetting features with a multi-level state skip list to generate a personalized push queue, and updates the threshold based on feedback to achieve accurate, phased content push.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] In a first aspect, the present invention provides a queue processing method based on dynamic forgetting feature extraction, comprising: collecting objective question answer records, vocabulary memorization records, and subjective question speech sequences input from a terminal; extracting feature indicators for each modality respectively; and concatenating them to generate a multidimensional learning feature set using user identifiers and timestamps as alignment benchmarks; mapping the multidimensional learning feature set to a preset multi-level state skip list; calculating dynamic forgetting feature parameters in the multi-level state skip list; performing node transition operations based on the dynamic forgetting feature parameters to generate a phased forgetting feature partial order set; calculating individual data density indicators based on the phased forgetting feature partial order set; executing branch extraction instructions based on the individual data density indicators and the phased forgetting feature partial order set to generate a personalized push content queue; performing phased content push based on the personalized push content queue; collecting lower-level interactive feedback sequences and calculating state degradation compensation factors; and using the state degradation compensation factors to back-update the boundary thresholds of the multi-level state skip list.

[0007] Preferably, the step of extracting feature indicators for each modality and concatenating them with user identifier and timestamp as alignment references to generate a multidimensional learning feature set includes: calculating the single-question error frequency, continuous error dwell time, and total number of answers based on objective question answer records; concatenating the single-question error frequency, continuous error dwell time, and total number of answers into vectors to obtain objective error distribution features; using a speech recognition algorithm to convert the subjective question speech sequence into text to obtain a text sequence; calculating the logical coherence value based on the text sequence; calculating the pronunciation pause interval and speech rate variance based on the subjective question speech sequence; jointly calculating the logical coherence value, the pronunciation pause interval, and the speech rate variance into numerical values ​​to obtain a subjective expression multidimensional scoring vector; and aligning and concatenating the objective error distribution features, the subjective expression multidimensional scoring vector, and vocabulary memorization records with user identifier and timestamp to generate the multidimensional learning feature set.

[0008] Further, the step of aligning and concatenating the objective error distribution features, the subjective expression multidimensional scoring vector, and the vocabulary memorization records according to user identifiers and timestamps to generate the multidimensional learning feature set includes: using the user identifier as the first alignment key, filtering out data items belonging to the same user identifier among the objective error distribution features, the subjective expression multidimensional scoring vector, and the vocabulary memorization records; using the collection timestamp of the objective question answer record as the second alignment key, within the data item range of the same user identifier, retrieving subjective expression multidimensional scoring vectors and vocabulary memorization records whose absolute difference from the collection timestamp of the objective question answer record does not exceed 3600 seconds, and when there are multiple candidate records satisfying the time difference condition... When matching data items, the data item with the smallest absolute time difference is selected. If the absolute time differences are the same, the data item with the earlier collection timestamp is selected. Data items that have already been matched will not be matched again with other objective question answer records during the current alignment process. For successfully matched data types, the corresponding data is written into the corresponding component. For subjective expression multidimensional scoring vectors or vocabulary memorization records that are not matched, the corresponding component is set to a null value, and the current objective question answer record is retained. Based on this, the multidimensional learning feature set data items are generated. The objective wrong question distribution feature component, the subjective expression multidimensional scoring vector component, and the vocabulary memorization record component are horizontally concatenated according to the field order to generate the multidimensional learning feature set.

[0009] Further, the step of mapping the multidimensional learning feature set to a preset multi-level state skip list, calculating dynamic forgetting feature parameters in the multi-level state skip list, and performing node transition operations based on the dynamic forgetting feature parameters to generate a phased forgetting feature partial order set includes: constructing a multi-level state skip list, which includes a bottom-level full linked list, a middle-level consolidation linked list, and a top-level mastery linked list; sequentially inserting data items from the multidimensional learning feature set as nodes into the bottom-level full linked list of the multi-level state skip list; for each node in the multi-level state skip list, extracting the timestamp corresponding to the multidimensional learning feature set, calculating the time interval value based on the current timestamp, extracting the historical error count value based on the objective error distribution characteristics, and extracting the ability fluctuation variance based on the subjective expression multi-dimensional scoring vector; and calculating the dynamic forgetting feature parameters based on the time interval value, the historical error count value, and the ability fluctuation variance.

[0010] Further, the step of calculating the dynamic forgetting characteristic parameters based on the time interval value, the historical error count value, and the ability fluctuation variance includes: calculating the initial forgetting decay value using the time interval value; calculating the product of the initial forgetting decay value and the historical error count value to obtain the basic forgetting index; and coupling the basic forgetting index with the ability fluctuation variance to output the dynamic forgetting characteristic parameters.

[0011] Further, the step of performing node transition operations based on the dynamic forgetting feature parameters to generate a partial order set of staged forgetting features includes: comparing the dynamic forgetting feature parameters with a first threshold and a second threshold; if the dynamic forgetting feature parameters are greater than the first threshold, then the corresponding node is kept in the bottom full linked list; if the dynamic forgetting feature parameters are greater than or equal to the second threshold and less than or equal to the first threshold, then the corresponding node is moved to the middle consolidation linked list; if the dynamic forgetting feature parameters are less than the second threshold, then the corresponding node is moved to the top master linked list; the bottom full linked list is marked as an unmastered grid state, the middle consolidation linked list is marked as a review grid state, and the top master linked list is marked as a mastered grid state; the multi-level state jump list after the transition is marked as the partial order set of staged forgetting features.

[0012] Further, the step of calculating the individual data density index based on the partial order set of stage-based forgetting features, and executing branch extraction instructions based on the individual data density index and the partial order set of stage-based forgetting features to generate a personalized push content queue includes: traversing the bottom-level full linked list and the middle-level consolidation linked list of the partial order set of stage-based forgetting features, counting the total number of nodes and dividing by a preset period value to calculate the individual data density index; executing the branch extraction instructions according to the relationship between the individual data density index and the cold start threshold and the sufficiency threshold to extract questions and fill them into the push pool; and performing a difficulty smoothing sorting operation on the questions in the push pool to obtain the personalized push content queue.

[0013] Further, the step of executing the branch extraction instruction based on the relationship between the individual data density index and the cold start threshold and the sufficiency threshold includes: if the individual data density index is less than the cold start threshold, connecting to the platform's global error database, extracting platform basic questions with a global error rate greater than a global preset value and placing them into the push pool; if the individual data density index is greater than or equal to the cold start threshold and less than the sufficiency threshold, extracting individual persistent errors and placing them into the push pool, calculating the knowledge point coverage vector of the individual persistent errors, and extracting orthogonal candidate questions from the platform's global error database and merging them into the push pool; if the individual data density index is greater than or equal to the sufficiency threshold, extracting historical corresponding questions only from the partial order set of stage forgetting features in descending order of dynamic forgetting feature parameters and filling them into the push pool.

[0014] Further, the step of collecting the lower-level interactive feedback sequence and calculating the state degradation compensation factor includes: collecting the lower-level interactive feedback sequence generated by the terminal after completing the phased content push; extracting the error recurrence rate value from the lower-level interactive feedback sequence; retrieving the associated nodes in the top-level master list of the phased forgetting feature partial order set that are erroneous in the lower-level interactive feedback sequence, calculating the proportion of the number of the associated nodes, and obtaining the cross-order error rate value; and performing a weighted summation of the error recurrence rate value and the cross-order error rate value to obtain the state degradation compensation factor.

[0015] Further, the step of using the state degradation compensation factor to reverse update the boundary threshold of the multi-level state skip list includes: reading a first threshold benchmark value and a second threshold benchmark value predetermined according to the distribution of dynamic forgetting feature parameters of historical learners on the platform; in each learning cycle, subtracting the current state degradation compensation factor from the first threshold benchmark value and the second threshold benchmark value respectively to obtain the first threshold update value and the second threshold update value corresponding to the current cycle, thereby completing the dynamic calibration of the first threshold and the second threshold; if the second threshold update value is less than or equal to zero, then the second threshold update value is corrected to a preset minimum threshold; if the first threshold update value is lower than the sum of the second threshold update value and the preset safety distance, then the first threshold update value is corrected to the second threshold update value plus the preset safety distance.

[0016] Furthermore, the calculation of the knowledge point coverage vector for individual persistent errors includes: reading the knowledge point tag list corresponding to the question identifier of the node from the platform's question knowledge point tag library; arranging all knowledge point identifiers of the platform in a fixed order after removing duplicates to form a complete set of knowledge point identifiers; and performing a judgment on each knowledge point identifier in the complete set of knowledge point identifiers. If the knowledge point identifier appears in the knowledge point tag list of any node, the corresponding position in the binary vector is set to 1; otherwise, it is set to 0, thus obtaining the knowledge point coverage vector for the individual persistent errors.

[0017] Secondly, this invention provides a queue processing system based on dynamic forgetting feature extraction, comprising: a multidimensional learning feature set generation module, used to collect objective question answer records, vocabulary memorization records, and subjective question speech sequences input from the terminal, extract feature indicators for each modality respectively, and concatenate them with user identifier and timestamp as alignment reference to generate a multidimensional learning feature set; a state parameter and node transition management module, used to map the multidimensional learning feature set to a preset multi-level state jump table, calculate dynamic forgetting feature parameters in the multi-level state jump table, perform node transition operations according to the dynamic forgetting feature parameters, and generate a phased forgetting feature partial order set; an indicator calculation and branch extraction push module, used to calculate individual data density indicators according to the phased forgetting feature partial order set, execute branch extraction instructions based on the individual data density indicators and the phased forgetting feature partial order set, and generate a personalized push content queue; and a feedback compensation reverse update calibration module, used to perform phased content push based on the personalized push content queue, collect lower-level interactive feedback sequences and calculate state degradation compensation factors, and use the state degradation compensation factors to reverse update the boundary thresholds of the multi-level state jump table.

[0018] Compared to existing technologies, the beneficial effects of this invention are as follows: By collecting objective question answer records and subjective question speech sequences and splicing them to generate a multi-dimensional learning feature set, the alignment and state linkage of cross-module features of objective and subjective questions are achieved. This allows the weaknesses exposed in subjective speech answers to be quantified and reflected as the basis for backtracking mastered vocabulary in objective questions, overcoming the lag in state updates and push jump oscillations caused by fixed-number judgments within a single module. The dynamic forgetting feature parameters integrate time interval, historical error count, and ability fluctuation variance, while simultaneously considering perceptual memory decay, objective difficulty, and subjective cognitive shift, driving nodes to make precise jumps in a multi-level state jump table, eliminating the defect that mastered vocabulary cannot be automatically backtracked under long-term intermittent learning. Based on a three-branch extraction mechanism of individual data density index, global incorrect questions are used as a supplement when individual data is scarce, orthogonal blind spots are filled by knowledge point coverage vectors when data accumulates, and only individual stubborn incorrect questions are relied upon when data is abundant. This strictly binds the scope of global data reference to the richness of individual data, avoiding the continuous shift of global statistical characteristics to individual states after individual data becomes abundant, which leads to repeated push of mastered content. By calculating the state degradation compensation factor through lower-level interactive feedback and dynamically calibrating the skip list boundary threshold in reverse, a cross-step parameter closed loop is formed. This structurally cuts off the push state boundary offset and cumulative error, improving the serialization accuracy of adaptive anti-forgetting push and the intelligent adaptive learning efficiency in educational scenarios. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of the queue processing method based on dynamic forgetting feature extraction in this invention;

[0021] Figure 2 This is a schematic diagram of the splicing of multi-dimensional learning feature sets in an embodiment of the present invention;

[0022] Figure 3 This is a schematic diagram of the multi-level state jump table node transition operation in an embodiment of the present invention;

[0023] Figure 4 This is a schematic diagram of orthogonal extraction of knowledge point coverage vectors in an embodiment of the present invention;

[0024] Figure 5 This is a schematic diagram illustrating the smooth sorting of personalized push content queues based on difficulty, as described in an embodiment of the present invention.

[0025] Figure 6 This is a schematic diagram of cross-order error rate association node retrieval in an embodiment of the present invention;

[0026] Figure 7 This is a schematic diagram illustrating the dynamic calibration of the first threshold and the second threshold in an embodiment of the present invention;

[0027] Figure 8 This is a functional block diagram of the queue processing system based on dynamic forgetting feature extraction in this invention. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] Example 1:

[0030] Please see Figure 1 As shown, this embodiment provides a queue processing method based on dynamic forgetting feature extraction, including:

[0031] S1: Collect objective question answer records, vocabulary memorization records, and subjective question speech sequences from the terminal input. Extract feature indicators based on the objective question answer records, vocabulary memorization records, and subjective question speech sequences, and concatenate them to generate a multi-dimensional learning feature set.

[0032] In intelligent language training and comprehensive education intensive training scenarios, learners' learning behaviors span multiple interaction formats, including objective vocabulary testing and subjective voice interview answers. Existing dynamic forgetting feature extraction mechanisms mostly rely on the accuracy rate of answers within a single module and a fixed number of reviews for judgment. The objective and subjective question modules maintain their answer records independently, and there is no feature alignment or state linkage mechanism between the two types of modules. When the weaknesses exposed by learners in subjective voice answers cannot be mapped to the corresponding basic vocabulary marked as "mastered" in the objective question module, the vocabulary in the "mastered" state will not trigger regression due to a decline in subjective expression ability, causing subsequent anti-forgetting push content to deviate from the learner's true mastery status. This step aims to simultaneously collect objective question answer records, vocabulary memorization records, and subjective question voice sequences from the terminal, extract feature indicators for each modality, and use user identifiers and timestamps as alignment benchmarks to concatenate and generate a multi-dimensional learning feature set, providing a unified and complete data input benchmark for subsequent cross-modal state assessment.

[0033] Further, step S1 includes:

[0034] S11: Collects objective question answer records, vocabulary memorization records, and subjective question speech sequences input from the terminal.

[0035] This step aims to synchronously collect three types of raw interactive data from the student's terminal, forming the raw data source for subsequent feature extraction. Objective question answer records include the student's answers to each question in objective question modules such as vocabulary testing and chapter exercises, along with corresponding timestamps; vocabulary memorization records include the student's memorization results for each word in the anti-forgetting vocabulary learning module, along with corresponding timestamps; and subjective question speech sequences are the raw speech data streams entered by the student in subjective question answering or interview answering modules. Since these three types of data are collected at different times on the timeline and belong to different learning function modules, subsequent step S14 requires aligning and splicing the three types of data based on user identifiers and timestamps. Therefore, this step must synchronously record the user identifier and collection timestamp corresponding to each data point to ensure that cross-module data has an alignable index structure.

[0036] Specifically, the objective question answer records are stored in the form of a structured list. Each data item in the list contains four fields: user identifier, question identifier, answer result (correct or incorrect), and collection timestamp.

[0037] Vocabulary memorization records are stored in the form of a structured list. Each data item in the list contains four fields: user identifier, vocabulary identifier, memorization result (mastered or not mastered), and collection timestamp.

[0038] The audio sequences of subjective questions are stored in the form of audio data streams. Each data item in the audio data stream contains four fields: user identifier, question identifier, audio frame data, and collection timestamp.

[0039] The objective question answer records are used to generate historical error statistics indexed by question identifiers, while the vocabulary memorization records are used to independently store vocabulary mastery or lack of mastery results. Since their data objects and statistical scopes differ, they are collected as different data types. Subsequent multi-level status jump tables use the question identifiers corresponding to the objective question answer records as node keys, and the vocabulary identifiers in the vocabulary memorization records as association identifiers in the node value fields; these are not used to extract historical error counts from the distribution characteristics of objective incorrect questions.

[0040] S12: Calculate the single-question error frequency, consecutive error dwell time, and total number of answers based on the objective question answer records. Concatenate the single-question error frequency, consecutive error dwell time, and total number of answers into vectors to obtain the objective question error distribution characteristics.

[0041] Existing error correction mechanisms only determine the status of errors based on whether a single answer is correct or not, without quantifying the frequency of errors for the same question over a historical answering period or the duration of consecutive errors. After students have experienced a long period of intermittent learning, some questions, although marked as errors in recent answering records, are actually occasional mistakes rather than persistent lack of mastery; other questions, however, are consistently missed over multiple learning periods and are considered genuine persistent errors. If the temporal distribution characteristics of these two types of errors are not distinguished, the subsequent calculation of forgetting feature parameters will be unable to differentiate between occasional and persistent errors, leading to persistent errors and occasional mistakes being treated equally in the status jump table, affecting the accuracy of the pushed content. This step extracts three types of statistics from objective question answering records: single-question error frequency, consecutive error dwell time, and total number of answers, and concatenates these three into objective error distribution features, providing a structured vector representation basis for the extraction of historical error counts in step S22.

[0042] Specifically, for each question identifier in the objective question answer record, the total number of times the question identifier appears in the objective question answer record is counted, as well as the number of times the answer corresponding to the question identifier is incorrect. The ratio of the number of incorrect answers to the total number of answers is calculated to obtain the single question error frequency corresponding to the question identifier.

[0043] For each question identifier in the objective question answer record, all answer records for the question identifier are sorted in ascending order by the collection timestamp. The longest consecutive error segment in which consecutive answers are incorrect is identified. The difference between the collection timestamp of the last error record in the longest consecutive error segment and the collection timestamp of the first error record is calculated to obtain the consecutive error dwell time corresponding to the question identifier, in seconds.

[0044] For each question identifier in the objective question answer record, the total number of data items corresponding to the question identifier in the objective question answer record is counted to obtain the total number of answers corresponding to the question identifier, and the value is a positive integer.

[0045] The error frequency per question, the duration of consecutive errors, and the total number of answers corresponding to each question identifier are arranged in order of question identifier and concatenated in the form of a three-dimensional vector to obtain the objective error distribution characteristics. The objective error distribution characteristics are a three-dimensional matrix, the number of rows of which is equal to the total number of duplicate question identifiers in the objective question answer records after deduplication. The first column of the matrix is ​​the error frequency per question corresponding to each question identifier, with a value range of [0, 1]; the second column of the matrix is ​​the duration of consecutive errors corresponding to each question identifier, with a value range of [0, +∞), in seconds; and the third column of the matrix is ​​the total number of answers corresponding to each question identifier, with a value of a positive integer.

[0046] S13: Use a speech recognition algorithm to convert the speech sequence of subjective questions into a text sequence. Calculate the logical coherence value based on the text sequence, and calculate the pronunciation pause interval and speech rate variance based on the speech sequence of subjective questions. Perform numerical joint calculation of the logical coherence value, pronunciation pause interval, and speech rate variance to obtain a multi-dimensional scoring vector for subjective expression.

[0047] The quality of subjective question responses via speech encompasses two distinct types of information: logical organization at the textual level and fluency at the speech level. Textual logical coherence reflects a student's ability to express themselves at the level of understanding knowledge points, while speech pauses and speech rate variance reflect fluency at the level of oral output. If subjective question response quality is judged solely by textual comparison, weaknesses in student expression due to poor fluency will go undetected; similarly, if only speech physical features are relied upon, textual logical organization deficiencies will remain unquantified. This step quantifies both the continuous physical features of speech and the discrete logical features of text into a multi-dimensional scoring vector for subjective expression. This allows subsequent subjective question evaluation to move beyond simple textual comparison and provides a cross-modal feature quantification benchmark for calculating the dynamic forgetting feature parameters of the multi-level state skip table in step S2.

[0048] Specifically, the speech recognition algorithm uses an open-source end-to-end automatic speech recognition model (such as OpenAI's open-source Whisper model) to process the subjective question speech sequence. This end-to-end automatic speech recognition model has been initially trained on a massive multilingual and multi-task audio dataset, possessing robust cross-lingual general automatic speech recognition and acoustic feature extraction capabilities. In this embodiment, the model takes audio frame data from the subjective question speech sequence as input and outputs the corresponding text transcription result, i.e., the text sequence.

[0049] The logical coherence score is calculated as follows: the text sequence is segmented into a list of sentences according to sentence boundaries, and the semantic similarity between adjacent sentence pairs is calculated using cosine similarity. Specifically, sentence embedding vectors are extracted from each adjacent sentence in the sentence list. These sentence embedding vectors are output by an open-source pre-trained text encoding model (such as the BERT model). This pre-trained text encoding model has completed initial masked language model training on large-scale general corpora such as Wikipedia and BookCorpus, and has the general feature representation ability to map discrete sentence texts into dense vectors containing rich contextual semantic information. In this embodiment, the pre-trained text encoding model takes two adjacent sentence texts as input objects and outputs sentence embedding vectors with a dimension of 768. Then, the cosine similarity between the two sentence embedding vectors is calculated. After traversing all adjacent sentence pairs in the sentence list, the arithmetic mean of all obtained cosine similarities is taken to obtain the logical coherence score. If the sentence list contains only one sentence, the logical coherence score of that sentence is recorded as 0.

[0050] The method for calculating the pronunciation pause interval is as follows: Silence detection is performed on the audio frame data in the subjective question speech sequence. The condition for determining a silence frame is that the root mean square energy value of consecutive audio frames is lower than a preset silence energy threshold. The preset silence energy threshold is determined based on the root mean square energy level of the background noise in the recording environment, specifically 1.5 times the root mean square energy level of the background noise. Intervals with consecutive silence frames lasting more than 200 milliseconds are identified as pause intervals. The duration of all pause intervals in the speech sequence is calculated, and the arithmetic mean of the durations of all pause intervals is taken to obtain the pronunciation pause interval in milliseconds. If there are no pause intervals in the speech sequence that meet the condition, the pronunciation pause interval is set to 0 milliseconds.

[0051] The method for calculating speech rate variance is as follows: The text sequence is divided into a list of time windows, each lasting 5 seconds, according to the corresponding speech time position. The last time window, which is less than 5 seconds, is recorded according to its actual duration. The ratio of the number of syllables in each time window to the actual duration of that time window is calculated to obtain the instantaneous speech rate corresponding to each time window, in units of syllables per second. When the time window list contains more than two time windows, the variance of the set of instantaneous speech rates corresponding to all time windows is calculated to obtain the speech rate variance. When the time window list contains only one time window, the speech rate variance is set to 0.

[0052] The logical coherence score, pronunciation pause interval, and speech rate variance are normalized and then concatenated into a 3D vector to obtain a multi-dimensional subjective expression score vector. The normalization method uses min-maximum linear normalization, specifically: using the minimum and maximum values ​​of logical coherence, pronunciation pause interval, and speech rate variance in historical subjective question answer records as normalization boundaries, the current value is linearly mapped to the [0,1] interval. The normalized values ​​of logical coherence, pronunciation pause interval, and speech rate variance are all dimensionless. If the maximum value in the current normalization boundary of any dimension equals the minimum value, the normalized value of that dimension in the current record is set to 0 to avoid a denominator of 0 in the normalization calculation. The minimum and maximum values ​​of each dimension in the historical subjective question answer records are initialized with pre-stored platform historical data statistics during system initialization and updated using a sliding window method after each new subjective question answer record. The sliding window size is set to 1000 recent historical records. When updating the minimum and maximum values ​​of each dimension using a sliding window approach, outlier filtering is performed on newly added records: if the standard deviation of the corresponding dimension within the current sliding window is greater than 0, and the value of that dimension in a newly added record exceeds the range of plus or minus three standard deviations of the mean of that dimension within the current sliding window, it is considered an outlier, excluded from the sliding window statistics, and the normalization boundary of that dimension is not updated; when the standard deviation of the corresponding dimension within the current sliding window is equal to 0, the three-standard-deviation constant filtering is not performed, and the newly added record is normally included in the sliding window; other non-outlier records are normally included in the sliding window, and the normalization boundary is updated according to the original rules. This filtering mechanism prevents extreme outlier recording data from polluting the normalization boundary, ensuring that the normalization benchmark of data from different periods in the multidimensional learning feature set remains stable and comparable.

[0053] The subjective expression multidimensional scoring vector concatenates the continuous physical features of speech and the discrete logical features of text into a unified 3D vector space, providing a unified comparison benchmark for the two types of information with different properties in the subsequent calculation of dynamic forgetting feature parameters in step S2. In the single-modal assessment method that relies solely on text comparison, the weak language fluency exposed by students due to frequent pauses and drastic fluctuations in speaking speed will not affect the status assessment of corresponding vocabulary in objective questions. Even if there are obvious signs of weakness in subjective expression, vocabulary that has been classified as green mastered grids will not trigger status regression. After introducing pronunciation pause intervals and speaking speed variance, the physical feature components of the subjective expression multidimensional scoring vector can encode the fluctuations in students' speaking ability into values ​​that can participate in the calculation of ability fluctuation variance in step S23, thus providing a quantitative basis for the downgrade regression of green mastered grid nodes in the node transition determination in step S24.

[0054] S14: Align and concatenate the objective error distribution features, subjective expression multidimensional scoring vectors, and vocabulary memorization records according to user identifiers and timestamps to generate a multidimensional learning feature set.

[0055] Objective error distribution characteristics, subjective expression multidimensional scoring vectors, and vocabulary memorization records belong to three different learning function modules, and their collection timestamps are not synchronized. If alignment is not based on user identifiers and timestamps, the three types of feature data cannot be correlated at the same learning record granularity, resulting in a lack of a unified temporal anchor point for cross-modal state evaluation in step S2. This step aims to use user identifiers and timestamps as dual alignment keys, with objective question answer records as the master record for the current learning period. Subjective expression multidimensional scoring vectors and vocabulary memorization records that can be obtained within the matching range are written into the corresponding fields. For data types that are not generated or matched in the current learning period, their field positions are preserved by marking them as null values. This generates a multidimensional learning feature set without discarding valid objective question records, providing a complete data structure for inserting multi-level state skip lists into nodes in step S21.

[0056] Specifically, using the user identifier as the first alignment key, data items belonging to the same user identifier are filtered out from objective error distribution characteristics, subjective expression multidimensional scoring vectors, and vocabulary memorization records.

[0057] Using the timestamp as the second alignment key, within the data item range of the same user identifier, and taking the collection timestamp of the objective question answer record as the matching benchmark, subjective expression multidimensional scoring vectors and vocabulary memorization records corresponding to subjective question speech sequences with an absolute time difference not exceeding 3600 seconds are retrieved. If only one data item of the same type satisfies the time difference condition, that data item is matched with the current objective question answer record; if multiple data items of the same type satisfy the time difference condition, the data item with the smallest absolute time difference between its collection timestamp and the current objective question answer record is selected for matching according to the nearest neighbor timestamp matching principle; if two or more data items have the same absolute time difference, the data item with the earlier collection timestamp is selected for matching. Once a subjective expression multidimensional scoring vector or vocabulary memorization record has been matched once in the current alignment process, it will not be matched again with other objective question answer records. If no data from either the subjective expression multidimensional scoring vector or the vocabulary memorization record is matched within the 3600-second matching range corresponding to the current objective question answer record, then the current objective question answer record is retained, and the component corresponding to the unmatched data type is marked as null. For the matched data type, the corresponding data is written into the corresponding component, and a multidimensional learning feature set data item corresponding to the current objective question answer record is generated accordingly. The null value marker is only used to indicate that no corresponding modality data was obtained in the current learning period and does not participate in the numerical calculation of the corresponding modality feature.

[0058] Using the current objective question answer record as the main record, its corresponding objective incorrect question distribution feature component is written into the multidimensional learning feature set data item. Successfully matched subjective expression multidimensional score vectors and vocabulary memorization records are written into their respective components. Unmatched data types are marked with null values ​​in their corresponding components. The multidimensional learning feature set is stored in a structured list format. Each data item in the list contains five fields: user identifier, alignment timestamp, objective incorrect question distribution feature component, subjective expression multidimensional score vector component, and vocabulary memorization record component. The alignment timestamp is taken from the collection timestamp of the objective question answer record in the current data item, which serves as the time matching benchmark.

[0059] See Figure 2 This is a schematic diagram of the splicing of multi-dimensional learning feature sets provided in an embodiment of this application. For example... Figure 2As shown, the objective error distribution features, subjective expression multidimensional scoring vectors, and vocabulary memorization records are aligned and concatenated. The user identifier is used as the first alignment key, and the timestamp as the second. Within the data item range of the same user identifier, the collection timestamp of the objective question answer record is used as the matching benchmark. Within a 3600-second time difference, the subjective expression multidimensional scoring vector and vocabulary memorization record are selected according to the nearest neighbor timestamp matching principle. When multiple candidate data items exist, the data item with the smallest absolute time difference is selected; when the absolute time differences are the same, the data item with the earlier collection timestamp is selected. Data items that have already been matched are not reused in the current alignment process. The objective question answer record is used as the master record to generate the corresponding multidimensional learning feature set data item; successfully matched subjective expression multidimensional scoring vectors and vocabulary memorization records are written into the corresponding components, and unmatched data types are marked with null values. The final generated multidimensional learning feature set contains five categories of fields for each data item: user identifier, alignment timestamp, objective error distribution feature component, subjective expression multidimensional scoring vector component, and vocabulary memorization record component. The alignment timestamp is the collection timestamp of the objective question answer record, which serves as the matching benchmark. This approach maintains the temporal correlation between data across modules and prevents the discarding of valid objective question records due to the absence of certain modalities in the current learning period, providing continuous data input for subsequent nodes to insert multi-level state skip lists.

[0060] S2: Map the multidimensional learning feature set to a preset multi-level state skip list, calculate the dynamic forgetting feature parameters in the multi-level state skip list, perform node transition operations according to the dynamic forgetting feature parameters, and generate a phased forgetting feature partial order set.

[0061] The existing red-yellow-green grid classification mechanism relies on a fixed review threshold (e.g., nine reviews) to determine vocabulary mastery. This single criterion fails to account for natural forgetting and decay over time, error accumulation due to differences in objective question difficulty, and cross-module state shifts caused by fluctuations in subjective expression ability. When students experience prolonged periods of intermittent learning, the fixed review threshold cannot revert the status of vocabulary that has been reviewed nine times but forgotten over time. Furthermore, when newly revealed weaknesses in subjective speech questions are related to vocabulary already classified as mastered (green) in objective questions, the accuracy assessment within a single module cannot trigger cross-module state downgrading. A multi-level state jump list is a hierarchical linked list structure with a partial order, consisting of a bottom-level full list, a middle-level consolidation list, and a top-level mastery list. These three levels correspond to the red (not mastered) grid state, the yellow (reviewed) grid state, and the green (mastered) grid state, respectively. The direction of node transitions between the three levels is driven by a comparison between dynamic forgetting feature parameters and two thresholds. This step aims to map data items in the multidimensional learning feature set to nodes in a multi-level state skip list. By combining the time interval value, historical error count value, and ability fluctuation variance, dynamic forgetting feature parameters are calculated, driving nodes to dynamically transition in the three-level linked list and generating a staged forgetting feature partial order set. This provides state-accurate structured input for the calculation of individual data density index and branch extraction in step S3.

[0062] Further, step S2 includes:

[0063] S21: Construct a multi-level state jump list, which includes a bottom-level full linked list, a middle-level consolidation linked list, and a top-level mastery linked list; insert the data items in the multi-dimensional learning feature set as nodes into the bottom-level full linked list of the multi-level state jump list in sequence.

[0064] The multi-level state skip list is a custom hierarchical linked list data structure, consisting of three ordered linked lists: a bottom-level full linked list, a middle-level consolidation linked list, and a top-level mastery linked list. These three lists share node entities but each maintains its own independent linked list pointers. During the initial node insertion phase, the bottom-level full linked list contains all nodes, corresponding to the red "not yet mastered" cells, representing learning content that the student has not yet mastered or has forgotten. The middle-level consolidation linked list contains nodes whose dynamic forgetting characteristic parameters are in the middle range, corresponding to the yellow "review" cells, representing learning content that the student has partially mastered but not yet solidified. The top-level mastery linked list contains nodes whose dynamic forgetting characteristic parameters are below the second threshold, corresponding to the green "mastered" cells, representing learning content that the student has solidified. Each data item in the multidimensional learning feature set uses the question identifier corresponding to the objective question answer record (which serves as the time matching benchmark) as the node key, and the corresponding user identifier, alignment timestamp, objective incorrect question distribution feature component, subjective expression multidimensional scoring vector component, and vocabulary memorization record component as the node value domain. The question identifier corresponding to the subjective question speech sequence and the vocabulary identifier corresponding to the vocabulary memorization record are stored as association identifiers in the node value domain and are not used as node keys in the multi-level state jump list. Initially, nodes are uniformly placed in the bottom-level full linked list, and the transition operation is performed in step S24 after the dynamic forgetting feature parameters are calculated in step S23.

[0065] Specifically, iterate through all data items in the multidimensional learning feature set, use the question identifier corresponding to the objective question answer record as the time matching benchmark as the node key, construct node entities, and append the node entities to the end of the underlying full linked list in sequence to complete the initial node insertion.

[0066] S22: For each node in the multi-level state jump list, extract the timestamp corresponding to the multi-dimensional learning feature set, calculate the time interval value by combining the current timestamp, extract the historical error count value based on the objective error distribution characteristics, and extract the ability fluctuation variance based on the subjective expression multi-dimensional scoring vector.

[0067] The calculation of dynamic forgetting feature parameters in step S23 requires three types of raw inputs, all extracted from the multidimensional learning feature set generated in step S1. The time interval value reflects the amount of time elapsed since the last learning record was created, serving as the physical basis for natural forgetting decay; the historical error count value reflects the cumulative number of errors made in historical answers to the objective question corresponding to the current node key, providing a quantitative representation of the objective difficulty of the knowledge point; and the ability fluctuation variance reflects the degree of fluctuation in the learner's subjective expression ability, providing a quantitative representation of cross-module cognitive state shifts. These three elements must fully cover the time, objective, and subjective dimensions to ensure sufficient input for the calculation of dynamic forgetting feature parameters.

[0068] Specifically, each node in the underlying full linked list of the multi-level state skip list is traversed, and the aligned timestamp field in the node's value field is read to obtain the historical timestamp corresponding to the node; the current system timestamp is obtained; the difference between the current timestamp and the historical timestamp is calculated to obtain the time interval value corresponding to the node, in hours.

[0069] Read the single-question error frequency and total number of answers corresponding to the question identifier of the node key from the objective error distribution feature component in the node value domain. Multiply the single-question error frequency and the total number of answers, and take the integer value of the product to obtain the historical error count value.

[0070] When the subjective expression multidimensional scoring vector component in the node value domain is not marked as null, the normalized value of the pronunciation pause interval and the normalized value of the speech rate variance are read from the subjective expression multidimensional scoring vector component; when the component is marked as null, the corresponding value is not extracted from the current data item. The normalized value of the pronunciation pause interval and the normalized value of the speech rate variance are both mapped to the [0,1] interval after the minimum-maximum linear normalization processing described in step S13, and are both dimensionless values. The values ​​of each time in the historical subjective question answer record are obtained by traversing all data items of the same user identifier in the multidimensional learning feature set, ignoring data items marked as null in the subjective expression multidimensional scoring vector component, and after aligning the timestamps in ascending order, the normalized value of the pronunciation pause interval and the normalized value of the speech rate variance are extracted respectively, forming two independent dimensionless scalar historical sequences. The variances of the historical sequences of normalized pronunciation pause intervals and normalized speech rate variances are calculated separately. The two dimensionless variance values ​​are then added together to obtain the scalar form of the ability fluctuation variance. When the number of valid data items in any historical sequence is less than 2, the variance corresponding to that historical sequence is set to 0. The ability fluctuation variance is a dimensionless numerical value, ranging from [0, +∞). A larger value indicates a greater fluctuation in the student's subjective expressive ability throughout the historical records.

[0071] S23: Calculate the dynamic forgetting characteristic parameters based on the time interval value, historical error count value, and ability fluctuation variance.

[0072] The three types of original input values ​​extracted in step S22 are merged into a single numerical variable, the dynamic forgetting feature parameter, providing a unified judgment benchmark for the node transition threshold comparison in step S24. Existing mechanisms use the correct answer rate within a single module or a fixed number of reviews as the basis for state determination, failing to simultaneously perceive forgetting decay caused by the passage of time, differences in persistence caused by the accumulation of historical errors, and cross-module state shifts caused by fluctuations in subjective ability. This step couples these three types of information into a dynamic forgetting feature parameter through a three-stage serial computation, enabling the parameter value to shift towards higher risk directions as time intervals increase, historical error accumulation increases, or ability fluctuations intensify, driving nodes to perform corresponding transitions in the multi-level state jump table.

[0073] Further, step S23 includes:

[0074] S231: Calculate the initial forgetting decay value using the time interval value.

[0075] This step aims to convert the time interval value into an initial forgetting decay value, quantifying the degree of natural forgetting from the last learning to the current moment. The longer the time interval, the higher the degree of natural forgetting, and the initial forgetting decay value should increase with the increase of the time interval. A mapping relationship between the time interval value and the initial forgetting decay value is established using the reciprocal form of exponential decay, so that the initial forgetting decay value is close to 0 when the time interval is short, and rises rapidly and tends to 1 as the time interval increases. This is consistent with the physical law described by the Ebbinghaus forgetting curve, which shows that the forgetting rate is faster in the early stages and slows down in the later stages.

[0076] Specifically, the initial forgetting decay value The calculation formula is:

[0077]

[0078] in, The time interval value calculated in step S22 is in hours; The forgetting rate coefficient is determined by using historical student answer records from the platform as the data source. It calculates the average error rate when students answer correctly once and then re-answer at different intervals. The time interval corresponding to an average error rate of 0.5 is used as the value. As a calibration point, let For example, if platform statistics show that the average error rate reaches 0.5 after a 48-hour interval, then .

[0079] The initial forgetting decay value The value range is [0, 1). The closer the value is to 1, the longer the time interval and the higher the risk of natural forgetting.

[0080] S232: Calculate the product of the initial forgetting decay value and the historical error count value to obtain the basic forgetting index.

[0081] By coupling the initial forgetting decay value with the historical error count, a basic forgetting index is generated that simultaneously reflects both the temporal forgetting decay and the objective difficulty of the knowledge point. If only the initial forgetting decay value is used as the basis for forgetting assessment, words with a single occasional error within the same time interval and persistently misspelled words with multiple historical errors will be assigned the same forgetting risk score, failing to distinguish the essential differences in mastery between the two types of words. After introducing the historical error count as a multiplicative factor, words with more accumulated historical errors receive a higher basic forgetting index within the same time interval. This allows persistently misspelled words to maintain a higher forgetting risk assessment result in the multi-level state skip table under the dual effects of temporal decay and historical error accumulation.

[0082] Specifically, the basic forgetting index The calculation formula is:

[0083]

[0084] in, The initial forgetting decay value calculated in step S231; The historical error count value extracted in step S22 is a non-negative integer.

[0085] The basic forgetting index The value range is [0, +∞), when the historical error count value When it is 0, the basic forgetting index A value of 0 always indicates that the node has never recorded an error, and the risk of basic forgetting is zero.

[0086] For example, assuming the time interval value of the objective question node corresponding to a certain word is small, the forgetting rate coefficient is: if the historical error count value is: then the basic forgetting index is: if the historical error count value of the objective question node corresponding to another word is: the basic forgetting index is: under the same time interval, the difference between the two basic forgetting indices reflects the distinction between persistent misspelled words and occasional misspelled words in the quantitative results of forgetting risk.

[0087] S233: Couples the basic forgetting index with the variance of ability fluctuation to calculate and output dynamic forgetting characteristic parameters.

[0088] The basic forgetting index output in step S232 is coupled with the ability fluctuation variance extracted in step S22 to output the final dynamic forgetting feature parameters. The basic forgetting index has encoded information from two dimensions: the passage of time and the accumulation of objective errors, but it has not yet incorporated the influence of the learner's subjective cognitive ability fluctuations. When learners exhibit increased ability fluctuations in subjective voice-based answers, the ability fluctuation variance increases, and the dynamic forgetting feature parameters should be increased accordingly. This should prompt nodes already included in the top-level mastery list to trigger a downgrade transition in step S24, thereby achieving cross-module state linkage updates.

[0089] Specifically, dynamic forgetting feature parameters The calculation formula is:

[0090]

[0091] in, The basic forgetting index is calculated in step S232; The capability fluctuation variance extracted in step S22; The ability fluctuation coupling coefficient is used to adjust the contribution weight of ability fluctuation variance to dynamic forgetting characteristic parameters. Its value is determined by using historical student data from the platform as a sample, and using the ratio of the 75th percentile of ability fluctuation variance in subjective question answers to the mean of the baseline forgetting index as the coefficient. The initial calibration value, for example, It can be initially set to 2.0.

[0092] The dynamic forgetting feature parameters An additive structure is adopted, with the two items corresponding to objective temporal forgetting risk and subjective cognitive state shift risk, respectively. An increase in either type of risk can independently drive the dynamic forgetting characteristic parameter to move towards the higher risk direction, so that the node transition judgment in step S24 remains sensitive to both objective forgetting accumulation and subjective ability fluctuations.

[0093] For example, following the numerical example of step S232, if the basic forgetting index of an objective question node for a certain word is the ability fluctuation variance and the coupling coefficient, then the dynamic forgetting characteristic parameter is; if the basic forgetting index of an objective question node for another word is the ability fluctuation variance, then, although the latter has less historical error accumulation, due to the greater subjective ability fluctuation, the dynamic forgetting characteristic parameter is still higher than the result determined purely by the historical error count value, reflecting the supplementary role of subjective cognitive state shift in forgetting risk assessment.

[0094] S24: Compare the dynamic forgetting feature parameter with the first threshold and the second threshold. If the dynamic forgetting feature parameter is greater than the first threshold, the corresponding node is kept in the bottom full linked list. If the dynamic forgetting feature parameter is greater than or equal to the second threshold and less than or equal to the first threshold, the corresponding node is moved to the middle consolidation linked list. If the dynamic forgetting feature parameter is less than the second threshold, the corresponding node is moved to the top master linked list. The bottom full linked list corresponds to the red unmastered grid state, the middle consolidation linked list corresponds to the yellow review grid state, and the top master linked list corresponds to the green mastered grid state. The multi-level state jump list after the transition is marked as a partial order set of stage forgetting features.

[0095] Based on the dynamic forgetting feature parameters calculated in step S23, nodes are driven to perform transitions between the three-layer linked lists of the multi-level state skip list through dual threshold comparison. The discrete red-yellow-green grids are abstracted into a hierarchical skip list structure with partial order relations, and the multi-level state skip list after the transition is marked as a stage-based forgetting feature partial order set, providing a well-ordered input with clear state boundaries for step S3. The first threshold and the second threshold jointly define the state boundaries of the three-layer linked list. The initial values ​​of both are calibrated by the platform's historical data and are dynamically calibrated in reverse by the state degradation compensation factor in step S45.

[0096] Specifically, the first threshold The initial determination method is as follows: using the distribution of dynamic forgetting characteristic parameters of all historical learners on the platform as samples, and taking the first threshold. This is the 67th percentile of the sample distribution, ensuring that approximately 67% of the nodes' dynamic forgetting feature parameters are at the first threshold. The following are examples, The initial value can be set to 2.5.

[0097] Second threshold The initial determination method is as follows: take the 33rd percentile of the sample distribution, so that the dynamic forgetting feature parameters of approximately 33% of the nodes are at the second threshold. The following are examples. The initial value can be set to 0.8.

[0098] Traverse each node in the multi-level state skip list and read the dynamic forgetting feature parameters corresponding to that node. :like The node is then retained in the underlying full linked list and marked as a red, unmastered cell; if If so, move the node from its current linked list to the middle-level consolidation linked list and mark it as a yellow review cell; if If a node is found to be in a state that has been successfully transitioned, it is moved from its current linked list to the top-level mastered linked list and marked as a green mastered cell. After all node transitions are completed, the multi-level state jump list is marked as a partially ordered set of stage-forgotten features.

[0099] See Figure 3 This is a schematic diagram of the multi-level state jump table node transition operation provided in the embodiments of this application. For example... Figure 3As shown, the multi-level state skip list is a hierarchical linked list structure with a partial order relationship, namely, a partially ordered set of stage-based forgetting features, consisting of three layers: a bottom-level full linked list, a middle-level consolidation linked list, and a top-level mastery linked list. A first threshold and a second threshold jointly define the state boundaries of the three linked lists. Dynamic forgetting feature parameters serve as the input comparison criteria, driving nodes to perform transitions between the three linked lists of the multi-level state skip list. If the dynamic forgetting feature parameter corresponding to a node is greater than the first threshold, the corresponding node is kept in the bottom-level full linked list; if the dynamic forgetting feature parameter is greater than or equal to the second threshold and less than or equal to the first threshold, it is moved to the middle-level consolidation linked list; if the dynamic forgetting feature parameter is less than the second threshold, it is moved to the top-level mastery linked list. By jointly calculating and outputting the dynamic forgetting feature parameters to drive the dynamic transitions of nodes in the three linked lists, the lag in state updates caused by sudden leaps in individual ability, as seen in traditional fixed review frequency criteria, is eliminated. This overcomes the hidden defects of cross-level misjudgments caused by a single dimension of state evaluation in long-term learning scenarios.

[0100] The dynamic forgetting feature parameters are jointly calculated using three inputs: time interval value, historical error count value, and ability fluctuation variance. The coupling of these three factors allows the dynamic forgetting feature parameters to simultaneously perceive the passage of time, objective difficulty, and subjective state shift. The joint calculation of the parameters drives the dynamic transition of nodes in a three-level linked list. The transition mechanism abstracts the discrete red-yellow-green grid classification into a hierarchical skip list structure with a partial order relationship, eliminating the state update lag phenomenon caused by the traditional fixed review number judgment standard when individual ability suddenly changes, and overcoming the hidden defect of cross-order misjudgment caused by the single dimension of state evaluation in long-term learning scenarios.

[0101] Specifically, in the fixed nine-review threshold criterion, the only condition for a node to advance from the red "unmastered" state to the green "mastered" state is that it has accumulated nine correct reviews. This criterion is fixed after the student completes nine reviews, and the node will not automatically regress regardless of the subsequent time interval or the magnitude of subjective ability fluctuations. In scenarios where students experience intermittent learning over long periods, the objective question nodes corresponding to vocabulary that have completed nine reviews tend to approach 1 due to the increased time interval. If the historical error count is greater than 0, it will continue to increase over time. After being combined with the contribution of ability fluctuation variance, it may exceed the first threshold, triggering the node to be downgraded from the top-level mastered linked list to the bottom-level full linked list. This achieves long-term forgetting regression detection that cannot be accomplished by the fixed-review criterion. The triggering condition corresponding to this degradation path is entirely determined by the mathematical structure of the dynamic forgetting feature parameters, without the need to introduce an additional regression detection module.

[0102] S3: Calculate the individual data density index based on the partial order set of stage-based forgetting features, execute branch extraction instructions based on the individual data density index and the partial order set of stage-based forgetting features, and generate a personalized push content queue.

[0103] The amount of individual historical practice data that the platform can acquire varies significantly at different stages of a student's learning process. In the early stages, individual historical data is extremely scarce, making it impossible to support a push strategy primarily based on individual historical errors. In the middle stages, individual historical data gradually accumulates, but it is still insufficient to cover all knowledge points, posing a risk of uneven knowledge point coverage. In the later stages, individual historical data is abundant, allowing push notifications to be executed entirely based on individual historical status. The existing mechanism, when introducing platform-wide data during the stage of insufficient individual data, does not weight and isolate the global error rate distribution from the individual long-term forgetting curve parameters. Coarse-grained global statistical characteristics directly overshadow fine-grained individual learning time-series status, causing vocabulary already classified as mastered (green) to be re-included in the push pool by the global low-accuracy data. This step uses the individual data density index as a conditional variable, dividing the question extraction strategy into three branches, corresponding to the three stages of scarce, accumulated, and abundant individual data. This ensures a strict conditional binding between the scope of global data introduction and the richness of individual data, preventing the statistical characteristics of global data from continuing to deviate from individual status even as individual data becomes more abundant.

[0104] Further, step S3 includes:

[0105] S31: Traverse the bottom-level full linked list and the middle-level consolidation linked list of the partial order set of stage forgetting features, count the total number of nodes and divide by the preset period value to calculate the individual data density index.

[0106] The richness of effective historical practice data for current learners within the partial-order set of stage-based forgetting characteristics is quantified using individual data density indicators, providing quantifiable input for branch selection in step S32. Nodes in the bottom-level full linked list and the middle-level consolidation linked list correspond to questions learners haven't fully mastered, while nodes in the top-level mastery linked list correspond to content already firmly grasped. The latter contributes limitedly to the formulation of the push strategy; therefore, only the total number of nodes in the first two layers is counted as an estimate of the effective historical practice data volume. Dividing by a preset period value normalizes the total number of nodes to a data density dimension corresponding to a unit learning period, making individual data density indicators comparable across different learning period lengths.

[0107] Specifically, traverse all nodes of the underlying full linked list in the partial order set of stage-based forgetting features, and count the number of nodes, denoted as . Traverse all nodes of the middle-level consolidation linked list and count the number of nodes. ; Calculate the total number of nodes .

[0108] Preset period value The method for determining the value is as follows: take the median number of questions answered by all historical students on the platform within a single standard training cycle as the benchmark value, and round the benchmark value to the nearest integer as the preset cycle value. For example, the preset cycle value can be set to 50.

[0109] Individual data density index The calculation formula is:

[0110]

[0111] The individual data density index The value ranges from [0, +∞), and the larger the value, the richer the effective historical practice data accumulated by the current student.

[0112] S32: Execute the branch extraction instruction based on the relationship between the individual data density index and the cold start threshold and the sufficiency threshold.

[0113] This step aims to select the corresponding question extraction branch based on the individual data density index calculated in step S31, perform question extraction, and populate the push pool, thus creating a strict conditional binding between the question extraction strategy and the richness of the statistical distribution of individual data. The cold start threshold and the sufficiency threshold together define three conditional intervals: below the cold start threshold corresponds to the individual data scarcity stage, above the sufficiency threshold corresponds to the individual data abundance stage, and in between corresponds to the individual data accumulation stage. The cold start threshold is determined by taking the 20th percentile of the distribution of the individual data density index of historical learners on the platform; for example, the cold start threshold can be set to 0.3. The sufficiency threshold is determined by taking the 80th percentile of the distribution of the individual data density index of historical learners on the platform; for example, the sufficiency threshold can be set to 1.5.

[0114] Further, step S32 includes:

[0115] S321: If the individual data density index is less than the cold start threshold, trigger the platform extraction branch: connect to the platform's global wrong question database, extract the platform's basic questions with a global error rate greater than the global preset value and put them into the push pool.

[0116] When the individual data density index is below the cold start threshold, the number of nodes in the bottom-level full linked list and the middle-level consolidation linked list in the partial-order set of stage-based forgetting features is insufficient to support personalized push notifications. This necessitates addressing the cold start problem when data is extremely scarce. In the early stages of a student's learning, if questions are forcibly extracted from an individual's historical incorrect answers, the push notification pool will be unable to fill completely due to insufficient available questions, resulting in a large amount of duplicate content. In this case, a platform-wide incorrect question database is used as a supplementary source. Using a global error rate greater than a global preset value as a filtering condition, basic questions with generally high error rates at the platform level are extracted. This ensures that the push notification content can still cover high-difficulty knowledge points even when individual data is extremely scarce.

[0117] Specifically, the system connects to the platform's global error database. This database uses the question identifier as the primary key, and the global error rate field is the ratio of the number of incorrect answers to that question identifier across all historical student responses on the platform to the total number of responses. The global preset value is determined by taking the 75th percentile of the global error rate distribution across all questions in the global error database. For example, the global preset value can be set to 0.6. All questions in the global error database with a global error rate greater than the global preset value are filtered out, and the list of question identifiers for these questions is written to the push pool.

[0118] S322: If the individual data density index is greater than or equal to the cold start threshold and less than the sufficiency threshold, trigger the hybrid extraction branch: extract the individual persistent wrong questions corresponding to the bottom full linked list and put them into the push pool. At the same time, calculate the knowledge point coverage vector of the individual persistent wrong questions, extract the platform's low accuracy questions that are orthogonal to the knowledge point coverage vector from the platform's global wrong question database, and merge them into the push pool.

[0119] When individual data density metrics fall between the cold start threshold and the sufficiency threshold, a balance needs to be struck between the diversity of pushed content and the coverage of knowledge points. While a certain number of persistent errors have accumulated in the partial-order set of stage-based forgetting characteristics, the scope of knowledge points covered may still be limited to those already encountered by the learner, posing a risk of unaddressed knowledge gaps. Therefore, while extracting persistent errors, a knowledge point coverage vector is used to identify questions in the platform's global error database that correspond to knowledge points orthogonal to the individual's practiced knowledge points. These questions are then merged into the push pool, allowing the pushed content to reinforce individual weaknesses while smoothly transitioning to previously unexplored knowledge gaps.

[0120] Specifically, the method for extracting individual persistent errors is as follows: traverse all nodes in the bottom-level full linked list of the partial order set of stage-based forgetting features, and sort them according to the dynamic forgetting feature parameters. Sort from largest to smallest; if the number of nodes in the underlying full linked list is greater than or equal to... Then take the former. The list of question identifiers corresponding to each node is written to the push pool; if the number of nodes in the underlying full linked list is less than... Then, the list of question identifiers corresponding to all nodes in the underlying full linked list will be written into the push pool. The method for determining this is as follows: take the preset period value. The 60% integer value, for example, if the preset period value ,but .

[0121] The method for constructing the knowledge point coverage vector is as follows: Read the knowledge point tag list corresponding to all nodes of the question identifier in the bottom-level full linked list and the middle-level consolidation linked list from the platform's question knowledge point tag library. The platform's question knowledge point tag library uses the question identifier as the primary key and the knowledge point identifier list as the knowledge point tag field. After deduplicating all knowledge point identifiers on the platform, arrange them in a fixed order to form a complete set of knowledge point identifiers, with the dimension denoted as [dimension not specified]. Construct a A binary vector of dimension 1 is used. For each knowledge point identifier in the complete set of knowledge point identifiers, if the identifier appears in the knowledge point label list of any node in the bottom-level full linked list or the middle-level consolidation linked list, the corresponding position is set to 1; otherwise, it is set to 0. This results in a knowledge point coverage vector. The data structure of the knowledge point coverage vector is as follows: A two-dimensional integer vector with values ​​ranging from 1 to 2. .

[0122] The method for extracting low-accuracy questions from the platform that are orthogonal to the knowledge point coverage vector is as follows: In the platform's global error question database, questions with a global error rate greater than a global preset value are selected. The knowledge point tag list for each question is then read to construct a list of candidate questions. A binary knowledge point vector is generated. The inner product of the knowledge point vector and the knowledge point coverage vector for each candidate question is calculated. A candidate question with an inner product of 0 indicates that none of the knowledge points involved in the candidate question appear within the scope of the individual's practiced knowledge points; that is, the candidate question and the knowledge point coverage vector are orthogonal. The list of question identifiers for all candidate questions with an inner product of 0 is merged into the push pool. If the number of candidate questions with an inner product of 0 is zero, the orthogonality constraint is relaxed. Instead, among the candidate questions that satisfy a global error rate greater than a global preset value, they are sorted in ascending order by the inner product of the candidate question's knowledge point vector and the knowledge point coverage vector, and the top-ranked questions are selected. Candidate questions with the smallest inner product are merged into the push pool; if the total number of candidate questions with a global error rate greater than the global preset value is still insufficient... If the platform's question bank is small, all candidate questions that meet the global error rate constraint are merged into the push pool. This fallback strategy ensures that the push pool can still be filled with valid questions even when the platform's question bank is small or the knowledge point system is limited, without affecting the normal execution of step S33, which involves smoothing the difficulty ranking. See also Figure 4 This is a schematic diagram of orthogonal extraction of knowledge point coverage vectors provided in an embodiment of this application. For example... Figure 4As shown, the knowledge point tags corresponding to all nodes in the bottom-level full linked list and the middle-level consolidation linked list are read to construct a knowledge point coverage vector. Simultaneously, the knowledge point tags of candidate questions are read from the platform's global error database, and candidate question knowledge point vectors are constructed accordingly. Next, the inner product of the candidate question knowledge point vector and the knowledge point coverage vector is calculated. An inner product of 0 within the box in the diagram indicates that none of the knowledge points involved in the candidate question appear within the individual's practiced knowledge point range; that is, the candidate question is orthogonal to the knowledge point coverage vector. Finally, these candidate questions with an inner product of 0 are merged into the push pool below. This hybrid extraction branch, while extracting persistent individual errors, also extracts low-accuracy platform questions corresponding to orthogonal knowledge points with an inner product of 0 from the global error database. This expands the push pool to a range of knowledge points the individual has not yet encountered, allowing the pushed content to smoothly transition to previously unexplored knowledge blind spots. This ensures that the introduced platform questions are high-difficulty questions at the platform level rather than low-value fillers, avoiding the defect that the diversity of knowledge points in the pushed content continuously shrinks as the individual's practice cycle extends.

[0123] The three-branch conditional extraction mechanism divides the individual data density index into three stages: data scarcity, data accumulation, and data abundance, corresponding to the three stages of the learning process. The boundary between these three stages is formed by the cold start threshold and the sufficiency threshold, creating a strict conditional binding between the question extraction strategy and the richness of the statistical distribution of individual data. The hybrid extraction branch, while extracting persistent incorrect questions, calculates the knowledge point coverage vector and extracts low-accuracy questions from the platform corresponding to orthogonal knowledge points with an inner product of 0 from the global database. This ensures that the pushed content covers individual weaknesses while smoothly transitioning to previously unaddressed knowledge gaps.

[0124] Specifically, when the individual data density index is between the cold start threshold and the sufficiency threshold, if only questions are extracted from the individual's persistent mistakes, the knowledge point tags of the questions in the push pool will always be limited to the range of questions that the student has already encountered. All questions corresponding to knowledge point identifiers with a value of 0 in the knowledge point coverage vector will never enter the push pool. Under the constraint of the limited individual historical question bank, a knowledge point information cocoon effect is formed, and the knowledge point diversity of the pushed content continues to shrink as the individual's practice cycle extends. The orthogonal knowledge point extraction mechanism expands the push pool to the range of knowledge points that the individual has never encountered by using the screening condition that the inner product of the knowledge point coverage vector and the candidate question knowledge point vector is 0. At the same time, it ensures that the introduced platform questions are high-difficulty questions at the platform level rather than low-value fillers by using the premise that the global error rate is greater than the global preset value. This avoids the defect of a sharp drop in knowledge point diversity at the structural level in the medium data density stage.

[0125] S323: If the individual data density index is greater than or equal to the sufficiency threshold, trigger the independent extraction branch: extract the corresponding historical questions from the bottom full linked list and the middle consolidation linked list of the partial order set of stage forgetting features in descending order of dynamic forgetting feature parameters, and fill them into the push pool.

[0126] Once the individual data density index reaches a sufficient threshold, the number of nodes in the bottom-level full linked list and the middle-level consolidation linked list of the partial order set of staged forgetting features is sufficient to support the complete filling of the push queue, without the need to introduce external data supplementation. At this stage, pushes are executed entirely based on the individual's historical state, avoiding the introduction of global platform data through hybrid extraction or platform extraction branches when individual data is already sufficiently abundant. This prevents the global error rate distribution from continuously interfering with the refined forgetting state assessment results of individuals after sufficient individual data has been obtained.

[0127] Specifically, iterate through all nodes in the bottom-level full linked list of the partial order set of stage-based forgetting features, according to the dynamic forgetting feature parameters. Sort the nodes from largest to smallest, and write the corresponding question identifiers to the push pool in sequence. If the number of questions in the push pool is less than the preset number of pushes, continue to traverse all nodes in the middle-level consolidation list, and write them according to the dynamic forgetting feature parameters. Sort the questions from largest to smallest, and append the corresponding question identifiers to the push pool in sequence until the number of questions in the push pool reaches the preset push limit. The preset push limit is determined by taking a preset period value. As a limit on the number of questions pushed out each time, for example, the preset number of questions can be set to 50.

[0128] S33: Perform a smooth sorting operation on the difficulty of the questions in the push pool to obtain a personalized push content queue.

[0129] The system performs a smooth difficulty sorting on questions from the underlying full linked list, the intermediate consolidation linked list, or the platform's global error database in the push pool. This ensures that the difficulty of questions in the personalized push content queue progresses smoothly from low to high, preventing the disorderly alternation of high-difficulty, persistent errors and low-difficulty beginner questions in the push queue, which could affect the continuity of students' practice during the intensive training phase. The difficulty smoothing operation is based on the question difficulty coefficient, which is provided by the platform's question metadata database. It uses the question identifier as the primary key and stores the difficulty coefficient field. The difficulty coefficient ranges from [0,1], with a larger value indicating a higher difficulty level.

[0130] Specifically, the difficulty coefficient field corresponding to the question identifier of each question in the push pool is read from the platform's question metadata database. All questions in the push pool are sorted in ascending order according to the difficulty coefficient, generating a personalized push content queue. The personalized push content queue is stored in the form of an ordered list. Each data item in the list contains two fields: question identifier and difficulty coefficient, and is sorted in ascending order of difficulty coefficient.

[0131] See Figure 5 This is a schematic diagram illustrating the smooth sorting of personalized push content queue difficulty provided in an embodiment of this application. For example... Figure 5 As shown, the push pool contains questions scattered from the bottom-level full linked list, the middle-level consolidation linked list, or the platform's global error database. Each question includes a corresponding question identifier and difficulty coefficient. A smooth difficulty sorting operation is performed on the questions in the push pool, primarily sorting all questions in ascending order of difficulty coefficient. The sorted questions are then sequentially added to the bottom layer in a step-by-step manner, forming a personalized push content queue. Based on the generated personalized push content queue, the difficulty of the questions in the queue is smoothly progressive from low to high, preventing highly difficult, persistent errors from appearing randomly alongside easy beginner questions, thus maintaining the continuity of student practice during the intensive training phase.

[0132] S4: Based on the personalized push content queue, perform phased content push, collect the lower-level interaction feedback sequence and calculate the state degradation compensation factor, and use the state degradation compensation factor to update the boundary threshold of the multi-level state skip table in reverse.

[0133] After a round of content delivery, students' actual performance on the delivered questions may deviate from the state evaluation results of the multi-level state jump list in step S2. Specifically, some questions corresponding to nodes already included in the top-level mastery list may show errors again in this round of delivery, indicating that the transition threshold boundary is too lenient, causing some nodes that have not yet been firmly mastered to be promoted to the top-level mastery list prematurely. If the actual answer feedback from this round of delivery is not applied back to the threshold boundary of the multi-level state jump list, the forgetting state determination in subsequent cycles will continue to use the already biased threshold boundary, and the error will continue to accumulate as the learning cycle lengthens. This step extracts the error recurrence rate and cross-order error rate values ​​from the lower-level interactive feedback sequence, calculates the state degradation compensation factor, and applies the state degradation compensation factor back to the first and second thresholds, forming a cross-step parameter dynamic dependency closed loop, so that the forgetting state determination in subsequent cycles is strictly controlled by the delivery test results of the previous cycle.

[0134] Further, step S4 includes:

[0135] S41: Distribute the personalized push content queue to the terminal in chronological order, driving the terminal to perform phased content push.

[0136] The personalized push content queue generated in step S33 is sent to the student's terminal in chronological order, triggering the terminal's question-answering interaction process and providing the prerequisite for the collection of the lower-level interactive feedback sequence in step S42. The questions in the personalized push content queue are arranged in ascending order of difficulty, and are pushed one by one in the list order during the intensive training phase. This ensures that students complete the exercises in order of difficulty during the intensive training phase, guaranteeing that the lower-level interactive feedback sequence collected in subsequent step S42 has a temporal structure that corresponds one-to-one with the personalized push content queue.

[0137] Specifically, the system reads the list of question identifiers from the personalized push content queue, sends the question content to the terminal in the order of the list, and records the sending timestamp of each question. After receiving the question content, the terminal displays it to the student. After the student completes the answer, he / she uploads the answer to the system. The system records the sending record of each question using the question identifier and sending timestamp as the primary key, and waits for step S42 to collect the lower-level interaction feedback sequence.

[0138] S42: The sequence of lower-level interactive feedback generated by the acquisition terminal after completing the phased content push.

[0139] After the student's terminal completes answering all questions in this round of content delivery, the generated answer feedback data is collected to construct the next-level interactive feedback sequence. This provides the raw data input for extracting the error recurrence rate and cross-order error rate values ​​in step S43. The next-level interactive feedback sequence must be collected all at once after all questions in this round have been answered to ensure that the statistical indicators cover the entire delivery cycle.

[0140] Specifically, the system waits for the student's terminal to complete the answers to all questions in the personalized content push queue. It then reads the answer results for each question from the system's distribution records and sorts all questions by their question identifiers, answer results (correct or incorrect), and answer timestamps in ascending order of timestamp, forming a lower-level interactive feedback sequence. This lower-level interactive feedback sequence is stored as an ordered list, where each data item contains three fields: question identifier, answer result, and answer timestamp.

[0141] S43: Extract the error recurrence rate from the lower-level interactive feedback sequence; retrieve the associated nodes in the top-level master list in the partial order set of stage forgetting features but which are erroneous in the lower-level interactive feedback sequence, calculate the proportion of associated nodes, and obtain the cross-order error rate.

[0142] Two types of quantitative indicators are extracted from the lower-level interactive feedback sequence: the error recurrence rate reflects the proportion of errors that reappear in the questions pushed in this round, and is a global statistic of the overall mastery of the pushed content; the cross-order error rate specifically counts the proportion of nodes that have been judged as green mastered grids in step S24 that answer incorrectly in this round of push, and is a precise quantitative indicator of the degree to which the threshold boundary is too lenient. The two are combined to form the original input for calculating the state degradation compensation factor in step S44.

[0143] Specifically, the error recurrence rate value The calculation method is as follows: count the number of data items with incorrect answers in the lower-level interactive feedback sequence, divide the number by the total number of data items in the lower-level interactive feedback sequence, and obtain the error recurrence rate. The value range is [0, 1].

[0144] Cross-order error rate The calculation method is as follows: traverse all nodes in the top-level master list of the partial order set of stage-based forgetting features, read the question identifier of each node, retrieve the answer corresponding to the question identifier in the lower-level interactive feedback sequence, filter out nodes with incorrect answers, record these nodes as associated nodes, and count the number of associated nodes. When the total number of nodes in the top-level controlled linked list is greater than 0, the ratio of the number of associated nodes to the total number of nodes in the top-level controlled linked list is calculated to obtain the cross-order error rate value; when the total number of nodes in the top-level controlled linked list is equal to 0, the cross-order error rate value is set to 0. The value range of the cross-order error rate value is [0,1]. See also Figure 6 This is a schematic diagram of cross-order error rate association node retrieval provided in an embodiment of this application. For example... Figure 6 As shown, the left side represents the lower-level interactive feedback sequence, which contains data items identifying questions with incorrect answers. The right side represents the top-level mastery list of the partial order set of stage-based forgetting features, containing multiple interconnected question identifier nodes. By traversing all nodes in the top-level mastery list of the partial order set of stage-based forgetting features, the corresponding answer result for each question identifier is retrieved from the lower-level interactive feedback sequence. Nodes in the right-hand list are then matched with data items indicating incorrect answers extracted from the lower-level interactive feedback sequence, thereby filtering out nodes with incorrect answers. These nodes are further recorded as associated nodes, and the cross-order error rate is calculated based on the proportion of associated nodes. Cross-order error rate statistics are performed on associated nodes that have been judged as having a green mastered grid state, specifically reflecting the degree of misjudgment of green mastered grid state nodes, providing a precise scalar input that can constrain error accumulation for dynamic calibration of threshold boundaries.

[0145] S44: The error recurrence rate and the cross-order error rate are weighted and summed to obtain the state degradation compensation factor.

[0146] This step aims to extract the error recurrence rate value from step S43. and cross-order error rate values The state degradation compensation factor is merged into a single numerical variable, providing a scalar input for the dynamic calibration of the threshold boundary in step S45. The error recurrence rate reflects the overall error level of the pushed content, while the cross-order error rate specifically reflects the degree of misjudgment of the green grid state nodes. The two have different weights in contributing to the threshold calibration, and the cross-order error rate has a more direct impact on the threshold boundary, so it is given a higher weight coefficient.

[0147] Specifically, state degradation compensation factor The calculation formula is:

[0148]

[0149] The error recurrence rate is the value calculated in step S43; the cross-order error rate is the value calculated in step S43; both weighting coefficients are non-negative and their sum equals 1. The weighting coefficients are determined as follows: using the error recurrence rate and cross-order error rate values ​​from multiple historical learning cycles as samples, under the constraint that both weighting coefficients are not less than 0 and their sum equals 1, constrained least squares method is used to fit the threshold calibration effect target. The reduction in cross-order error rate in the next cycle after calibration is taken as the target variable, and the error recurrence rate and cross-order error rate are used as independent variables to construct a regression matrix. The solution with the smallest sum of squared residuals within the solution space satisfying the constraints is taken as the estimated value of the two weighting coefficients. For example, the two weighting coefficients can be initially set to 0.3 and 0.7, respectively.

[0150] The state degradation compensation factor The value range is [0, 1]. The larger the value, the higher the degree of error detected in this round of push, and the threshold boundary needs to be tightened and calibrated more significantly.

[0151] S45: Feedback the state degradation compensation factor back to step S24. Using the first threshold reference value and the second threshold reference value as calibration references, calculate the first threshold update value and the second threshold update value corresponding to the current learning cycle based on the current state degradation compensation factor, and complete the dynamic calibration of the first threshold and the second threshold.

[0152] This step aims to apply the state degradation compensation factor calculated in step S44 as a back feedback to the boundary threshold in step S24, enabling the state boundary of the multi-level state skip list to adaptively adjust according to the actual error feedback of the current learning cycle. The first and second thresholds initially determined in step S24 are saved as the first and second threshold baseline values, respectively. Each learning cycle recalibrates the thresholds of the current cycle using these baseline values, without further accumulating the thresholds already updated in the previous learning cycle. When the error recurrence rate and cross-order error rate increase in the current cycle, the state degradation compensation factor increases, and the first and second thresholds of the current cycle further decrease relative to their respective baseline values, tightening the conditions for nodes to be promoted to the middle-level consolidation list and the top-level mastery list. When the error level decreases in subsequent learning cycles, the state degradation compensation factor decreases, and the first and second thresholds of the current cycle correspondingly increase relative to the previous learning cycle. When the state degradation compensation factor is 0, the first and second thresholds of the current cycle return to their corresponding baseline values, thus allowing the boundary thresholds to change bidirectionally with the feedback state.

[0153] Specifically, read the current learning cycle state degradation compensation factor calculated in step S44; read the first threshold benchmark value and the second threshold benchmark value determined and saved in step S24 based on the distribution of dynamic forgetting characteristic parameters of historical learners on the platform; and perform the following update calculations:

[0154]

[0155]

[0156] The first and second threshold reference values ​​remain unchanged during the threshold dynamic calibration process in each learning cycle. In the current learning cycle, the corresponding threshold update value is calculated only based on the state degradation compensation factor of this cycle.

[0157] The updated first and second threshold values ​​after boundary correction are written back to the current valid threshold storage in step S24 and used as the current valid first and second thresholds when step S24 is executed in the next learning cycle. When step S45 is executed next time, the first and second threshold reference values ​​are read again for calculation, and the threshold update values ​​of the previous learning cycle are not used as the reference for the next subtraction calculation.

[0158] If the second threshold update value is less than or equal to 0, the second threshold update value will be corrected to the preset minimum threshold. The preset minimum threshold is determined by taking the 5th percentile of the distribution of dynamic forgetting feature parameters of historical students on the platform. For example, the preset minimum threshold can be set to 0.1 to avoid the second threshold exceeding the effective value range when the current period state degradation compensation factor is large, which would lead to the boundary failure situation where there are no nodes to accommodate the top-level master chain.

[0159] If the updated value of the first threshold is lower than the sum of the updated value of the second threshold and the preset safety distance, the updated value of the first threshold is corrected to the updated value of the second threshold plus the preset safety distance. The preset safety distance is set to 0.2 to ensure that the first threshold is always greater than the second threshold, preventing the two thresholds from being reversed and causing the interval comparison logic in step S24 to fail. If the updated value of the first threshold after the safety distance correction exceeds the preset maximum threshold, the updated value of the first threshold is corrected to the preset maximum threshold. The preset maximum threshold is determined by taking the 95th percentile of the distribution of the dynamic forgetting feature parameters of historical students on the platform. For example, the preset maximum threshold can be set to 5.0.

[0160] The reverse feedback mechanism transforms the lower-level interactive feedback sequence into a state degradation compensation factor. Using a first and second threshold benchmark as fixed calibration benchmarks, it generates the effective threshold for the current period based on the actual feedback state of each learning cycle. When the error recurrence rate and cross-order error rate increase in the current push, the state degradation compensation factor increases, and the two current effective thresholds are adjusted downwards relative to the benchmark values, tightening the conditions for nodes to transition to higher mastery states. When the error severity decreases in subsequent learning cycles, the state degradation compensation factor decreases, and the two current effective thresholds rise back towards the benchmark values ​​relative to the previous learning cycle. Therefore, the dynamic threshold calibration in this embodiment does not result in a unidirectional decrease that accumulates continuously with the learning cycle. Instead, it adjusts the threshold based on historical data and the current learning state, allowing the threshold boundary to continuously track changes in the individual learning state. See also... Figure 7 This is a schematic diagram of dynamic calibration of the first and second thresholds provided in an embodiment of this application. Figure 7 As shown, the first and second threshold benchmarks constitute the starting boundary for threshold calibration in the current learning cycle. The larger the state degradation compensation factor, the greater the downward adjustment of the current effective first and second thresholds relative to the corresponding benchmarks. When the state degradation compensation factor decreases, the current effective thresholds rise back towards the corresponding benchmarks. Simultaneously, by setting a minimum threshold and a safe distance, boundary constraints are imposed on the update results, thereby ensuring that the three-layer state interval remains effective at all times.

[0161] Example 2:

[0162] This embodiment, based on Embodiment 1, provides a queue processing system based on dynamic forgetting feature extraction, such as... Figure 8 As shown, it includes:

[0163] The multidimensional learning feature set generation module is used to collect objective question answer records, vocabulary memorization records and subjective question speech sequences input from the terminal, extract feature indicators of each modality, and concatenate them with user identifier and timestamp as alignment benchmark to generate a multidimensional learning feature set;

[0164] The state parameter and node transition management module is used to map the multi-dimensional learning feature set to a preset multi-level state skip list, calculate dynamic forgetting feature parameters in the multi-level state skip list, perform node transition operations according to the dynamic forgetting feature parameters, and generate a phased forgetting feature partial order set.

[0165] The indicator calculation and branch extraction push module is used to calculate the individual data density indicator based on the partial order set of stage forgetting features, execute the branch extraction instruction based on the individual data density indicator and the partial order set of stage forgetting features, and generate a personalized push content queue.

[0166] The feedback compensation reverse update calibration module is used to perform phased content push based on the personalized push content queue, collect the lower-level interaction feedback sequence and calculate the state degradation compensation factor, and use the state degradation compensation factor to reverse update the boundary threshold of the multi-level state skip table.

Claims

1. A queue processing method based on dynamic forgetting feature extraction, characterized in that, include: The system collects objective question answer records, vocabulary memorization records, and subjective question speech sequences from the terminal input, extracts feature indicators for each modality, and concatenates them with user identifier and timestamp as alignment benchmarks to generate a multidimensional learning feature set. The multidimensional learning feature set is mapped to a preset multi-level state skip list. Dynamic forgetting feature parameters are calculated in the multi-level state skip list. Node transition operations are performed according to the dynamic forgetting feature parameters to generate a phased forgetting feature partial order set. Calculate the individual data density index based on the partial order set of stage-based forgetting features, and execute branch extraction instructions based on the individual data density index and the partial order set of stage-based forgetting features to generate a personalized push content queue. Based on the personalized push content queue, phased content push is performed, the lower-level interaction feedback sequence is collected and the state degradation compensation factor is calculated, and the boundary threshold of the multi-level state skip table is updated in reverse using the state degradation compensation factor.

2. The queue processing method based on dynamic forgetting feature extraction according to claim 1, characterized in that, The concatenation to generate a multidimensional learning feature set includes: The error frequency per question, the duration of consecutive errors, and the total number of answers are calculated based on the objective question answer records. The error frequency per question, the duration of consecutive errors, and the total number of answers are then concatenated into vectors to obtain the objective question error distribution characteristics. A speech recognition algorithm is used to convert the speech sequence of subjective questions into a text sequence. The logical coherence value is calculated based on the text sequence. The pronunciation pause interval and speech rate variance are calculated based on the speech sequence of subjective questions. The logical coherence value, the pronunciation pause interval, and the speech rate variance are numerically and jointly calculated to obtain a multi-dimensional score vector for subjective expression. The objective error distribution features, the subjective expression multidimensional scoring vector, and the vocabulary memorization records are aligned and concatenated according to user identifiers and timestamps to generate the multidimensional learning feature set.

3. The queue processing method based on dynamic forgetting feature extraction according to claim 2, characterized in that, The generation of the multidimensional learning feature set includes: Using the user identifier as the first alignment key, data items belonging to the same user identifier in the objective error distribution features, the subjective expression multidimensional scoring vector, and the vocabulary memorization records are filtered out. Using the collection timestamp of the objective question answer record as the second alignment key, within the data item range of the same user identifier, the collection timestamp of the subjective expression multidimensional scoring vector and the collection timestamp of the vocabulary memorization record are aligned with the current objective question answer record according to the nearest neighbor timestamp matching principle; The objective error distribution feature component corresponding to the current objective question answer record is horizontally concatenated with the successfully matched subjective expression multidimensional score vector and / or vocabulary memorization record in field order; for unmatched data types, the corresponding component is set as a null value marker to generate the multidimensional learning feature set.

4. The queue processing method based on dynamic forgetting feature extraction according to claim 1, characterized in that, The generated partial order set of stage-based forgetting features includes: Construct a multi-level state jump list, which includes a bottom-level full linked list, a middle-level consolidation linked list, and a top-level mastery linked list; insert the data items in the multi-dimensional learning feature set as nodes into the bottom-level full linked list of the multi-level state jump list in sequence; For each node in the multi-level state jump list, extract the timestamp corresponding to the multi-dimensional learning feature set, calculate the time interval value by combining the current timestamp, extract the historical error count value based on the objective error distribution characteristics, and extract the ability fluctuation variance based on the subjective expression multi-dimensional scoring vector. The dynamic forgetting characteristic parameters are calculated based on the time interval value, the historical error count value, and the ability fluctuation variance.

5. The queue processing method based on dynamic forgetting feature extraction according to claim 4, characterized in that, The calculation of dynamic forgetting characteristic parameters based on the time interval value, the historical error count value, and the ability fluctuation variance includes: The initial forgetting decay value is calculated using the aforementioned time interval value; The basic forgetting index is obtained by multiplying the initial forgetting decay value by the historical error count value. The basic forgetting index is coupled with the ability fluctuation variance to calculate and output the dynamic forgetting feature parameters.

6. The queue processing method based on dynamic forgetting feature extraction according to claim 1, characterized in that, The generated partial order set of stage-based forgetting features includes: Compare the dynamic forgetting feature parameters with the first threshold and the second threshold; If the dynamic forgetting feature parameter is greater than the first threshold, the corresponding node will be retained in the underlying full linked list. If the dynamic forgetting feature parameter is greater than or equal to the second threshold and less than or equal to the first threshold, then the corresponding node is moved to the middle-level consolidation list; If the dynamic forgetting feature parameter is less than the second threshold, then the corresponding node is moved to the top-level master list; The bottom-level full linked list is marked as the unmastered grid state, the middle-level consolidation linked list is marked as the review grid state, and the top-level mastered linked list is marked as the mastered grid state. The multi-level state jump list after the transition is marked as the partial order set of the stage forgetting features.

7. The queue processing method based on dynamic forgetting feature extraction according to claim 1, characterized in that, The step of calculating an individual data density index based on the partial order set of staged forgetting features, and executing a branch extraction instruction based on the individual data density index and the partial order set of staged forgetting features to generate a personalized push content queue includes: Traverse the bottom-level full linked list and the middle-level consolidation linked list of the partial order set of the stage forgetting features, count the total number of nodes and divide by the preset period value to calculate the individual data density index. Based on the relationship between the individual data density index and the cold start threshold and the sufficiency threshold, the branch extraction instruction is executed to extract questions and fill them into the push pool; Perform a difficulty smoothing sorting operation on the questions in the push pool to obtain the personalized push content queue.

8. The queue processing method based on dynamic forgetting feature extraction according to claim 7, characterized in that, Executing the branch extraction instruction includes: If the individual data density index is less than the cold start threshold, connect to the platform's global error database and extract basic platform questions with a global error rate greater than the global preset value and put them into the push pool. If the individual data density index is greater than or equal to the cold start threshold and less than the sufficiency threshold, the individual persistent wrong questions are extracted and placed into the push pool. The knowledge point coverage vector of the individual persistent wrong questions is calculated. Orthogonal candidate questions are extracted from the platform's global wrong question database and merged into the push pool. If the individual data density index is greater than or equal to the sufficiency threshold, historical corresponding questions are extracted from the partial set of stage-based forgetting features in descending order of dynamic forgetting feature parameters and added to the push pool.

9. The queue processing method based on dynamic forgetting feature extraction according to claim 1, characterized in that, The process of collecting the lower-level interactive feedback sequence and calculating the state degradation compensation factor includes: The sequence of lower-level interactive feedback generated by the acquisition terminal after completing the phased content push; Extract the error recurrence rate value from the lower-level interactive feedback sequence; Retrieve the associated nodes in the top-level master list of the partial order set of the stage forgetting features that are located in the lower-level interactive feedback sequence but exhibit errors, calculate the proportion of the number of such associated nodes, and obtain the cross-order error rate value. The state degradation compensation factor is obtained by weighted summing of the error recurrence rate and the cross-order error rate.

10. The queue processing method based on dynamic forgetting feature extraction according to claim 1, characterized in that, The step of using the state degradation compensation factor to reverse update the boundary threshold of the multi-level state skip list includes: Read the first and second threshold benchmark values ​​that are pre-determined and saved based on the distribution of dynamic forgetting characteristic parameters of historical learners on the platform; The state degradation compensation factor of the current learning cycle is subtracted from the first threshold reference value and the second threshold reference value respectively to obtain the first threshold update value and the second threshold update value corresponding to the current learning cycle, thus completing the dynamic calibration of the first threshold and the second threshold. If the second threshold update value is less than or equal to zero, then the second threshold update value is corrected to the preset minimum threshold. If the first threshold update value is lower than the sum of the second threshold update value and the preset safety distance, then the first threshold update value is corrected to the second threshold update value plus the preset safety distance.

11. The queue processing method based on dynamic forgetting feature extraction according to claim 8, characterized in that, The knowledge point coverage vector for calculating individual persistent errors includes: Retrieves a list of knowledge point tags corresponding to the question identifier of the node from the platform's question knowledge point tag library; After deduplicating all knowledge point identifiers on the platform, they are arranged in a fixed order to form a complete set of knowledge point identifiers. For each knowledge point identifier in the complete set of knowledge point identifiers, a judgment is made. If the knowledge point identifier appears in the knowledge point label list of any node, the corresponding position in the binary vector is set to 1; otherwise, it is set to 0, thus obtaining the knowledge point coverage vector of the individual persistent wrong question.

12. A queue processing system based on dynamic forgetting feature extraction, used to implement the method according to any one of claims 1-11, characterized in that, include: The multidimensional learning feature set generation module is used to collect objective question answer records, vocabulary memorization records and subjective question speech sequences input from the terminal, extract feature indicators of each modality, and concatenate them with user identifier and timestamp as alignment benchmark to generate a multidimensional learning feature set; The state parameter and node transition management module is used to map the multi-dimensional learning feature set to a preset multi-level state skip list, calculate dynamic forgetting feature parameters in the multi-level state skip list, perform node transition operations according to the dynamic forgetting feature parameters, and generate a phased forgetting feature partial order set. The indicator calculation and branch extraction push module is used to calculate the individual data density indicator based on the partial order set of stage forgetting features, execute the branch extraction instruction based on the individual data density indicator and the partial order set of stage forgetting features, and generate a personalized push content queue. The feedback compensation reverse update calibration module is used to perform phased content push based on the personalized push content queue, collect the lower-level interaction feedback sequence and calculate the state degradation compensation factor, and use the state degradation compensation factor to reverse update the boundary threshold of the multi-level state skip table.

Images