Intelligent partner training adjusting system and method

Abstract

The invention belongs to the technical field of artificial intelligence, and discloses an intelligent partner training adjustment system and method. The problems that in the prior art, only a preset script and simulation training data are relied on, so that the data collection dimension is single, and the user cognitive load state is difficult to comprehensively evaluate are solved; the method comprises the steps of collecting multi-dimensional data of users, behaviors and physiology, comprehensively extracting structured features and cognitive load sequences, processing to obtain cross-dimensional features, difficulty constraints and capability vectors, correcting the capability vectors in combination with an improved model, analyzing, generating and correcting an enhanced task queue by using a TOPS IS algorithm, dynamically adjusting the weight of a learning task, and obtaining a learning task. Finally, a compound learning task is obtained, and comprehensive evaluation and task optimization of the cognitive load state of the user are achieved. According to the method, the enhanced task queue is generated through the TOPS IS algorithm, the task weight is dynamically adjusted in combination with cross-dimension features and difficulty constraints, a composite learning task is formed, and the problem that single capability assessment is insufficient is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence, and more specifically, to an intelligent coaching adjustment system and method. Background Art

[0002] With the penetration of artificial intelligence technology in the field of education, the limitations of the traditional standardized coaching mode have become increasingly prominent.

[0003] Chinese Patent Application No. CN113627801A discloses an intelligent coaching method, device, electronic device and storage medium. The method includes: a tutor builds simulation practice scripts for different service scenarios at the tutor end; a training center trains the intents and quality inspection rules corresponding to the simulation practice scripts to obtain an intent model and a quality inspection rule model, and sends the intent model and the quality inspection rule model to the student end; a student simulates and trains the simulation practice scripts for different service scenarios at the student end to obtain simulation training information; the student end scores the student's simulation training information in real time according to the intent model and the quality inspection rule model to obtain a first score; the quality inspection rules at least include: emotion recognition rules, interrupting speech monitoring rules, speech rate monitoring rules and service taboo words. This invention saves the labor cost of training.

[0004] Although the above method can meet most scenarios, through research and practical application of the above method and the prior art, it is found that the above method and the prior art have at least the following partial defects:

[0005] Relying only on preset scripts and simulation training data results in a single data collection dimension, making it difficult to comprehensively evaluate the user's cognitive load state; only scoring the simulation training results through the intent model and quality inspection rules lacks a dynamic task adjustment logic based on real-time data; relying only on predefined simulation scripts and fixed scoring rules leads to insufficient refinement of personalized measurement.

[0006] In view of this, the present invention proposes an intelligent coaching adjustment system and method to solve the above problems. Summary of the Invention

[0007] To overcome the above defects of the prior art and to achieve the above object, the present invention provides the following technical solutions:

[0008] An intelligent coaching adjustment method, comprising the following steps:

[0009] Collect user data, behavior data and physiological data;

[0010] Extract comprehensive features from the user data, behavior data and physiological data to obtain structured features; extract comprehensive features from the behavior data and physiological data to obtain a cognitive load sequence;

[0011] Process the structured features to obtain cross-dimensional features; analyze the cognitive load sequence to obtain difficulty constraints;

[0012] According to the predefined knowledge graph, process the structured features to obtain an ability vector; based on the improved Ebbinghaus model, correct the ability vector to obtain a corrected ability vector;

[0013] Based on the TOPSIS multi-criteria decision-making algorithm, analyze the structured features according to the corrected ability vector to obtain a strengthened task queue; process the strengthened task queue according to the cross-dimensional features to obtain a corrected task queue;

[0014] Dynamically adjust the learning task weights according to the cognitive load index sequence to obtain a task weight matrix;

[0015] Analyze according to the difficulty constraints, the corrected task queue, and the task weight matrix to obtain a composite learning task.

[0016] Furthermore, the user data includes oral recordings and answer records; the answer records include learning task type labels, question IDs, user answers, correctness labels, and response times. Among them, the learning task type labels in the answer records include grammar, vocabulary, and listening. The correctness label of 1 indicates a correct answer, and the correctness label of 0 indicates a wrong answer; the behavior data includes learning session timestamps; the physiological data is RR interval data.

[0017] Furthermore, the method for obtaining the structured features includes:

[0018] Convert the oral recording into oral text, compare it with the preset answer text, count the total number of words and the number of wrong words in the oral text, calculate the ratio of the number of wrong words to the total number of words to obtain the oral error density;

[0019] Statistically calculate the average value and standard deviation of the number of words per second in the oral text, calculate the ratio of the standard deviation of the number of words per second to the average value of the number of words per second to obtain the speech rate volatility; count the number of occurrences of preset filler words and adjacent repeated phrases within a unit time, calculate the ratio of the number of occurrences to the total number of words in the oral text to obtain the coherence index; use the oral text as the input of the pre-trained speech model to obtain the phoneme error rate;

[0020] Obtain the number of questions with the learning task type label of vocabulary and the correctness label of 1 in the last R questions, calculate the ratio of the number of questions with the correctness label of 1 to R to obtain the vocabulary accuracy rate; count the total number of reviews of vocabulary by the user in the last H weeks, calculate the ratio of the total number of reviews to H to obtain the average number of reviews per week;

[0021] Obtain the number of questions with the most recent E-channel learning task type label being grammar and the correctness label being 1, calculate the ratio of the number of questions with the correctness label being 1 to E, and obtain the grammar accuracy rate;

[0022] Obtain all the question IDs with the correctness label being 0, obtain the knowledge point labels and the corresponding knowledge point basic coefficients corresponding to all the question IDs with the correctness label being 0 according to the predefined knowledge graph, count the number of errors corresponding to the same knowledge point label, and calculate the product of the number of errors and the knowledge point basic coefficient to obtain the knowledge point weight; count all the knowledge point weights to obtain the defect matrix;

[0023] Obtain the response time of the preset first number, calculate the standard deviation and the mean of the response time of the preset first number, calculate the ratio of the standard deviation to the mean of the response time of the preset first number, and obtain the answer time dispersion; obtain the physiological data of the preset second number, calculate the standard deviation and the mean of the physiological data of the preset second number, calculate the ratio of the standard deviation to the mean of the physiological data of the preset second number, and obtain the physiological characteristics; weight the answer time dispersion and the physiological characteristics to obtain the cognitive load index;

[0024] Concatenate the oral error density, speech rate volatility, coherence index, phoneme error rate, vocabulary accuracy rate, weekly average review times, grammar accuracy rate, defect matrix, and cognitive load index, as well as the corresponding timestamps, to obtain the structured features.

[0025] Further, the method for obtaining the cognitive load sequence includes:

[0026] Obtain the cognitive load index, data source label, and the corresponding timestamp in the structured features within the preset time period, and then concatenate them to obtain the cognitive load sequence;

[0027] The method for obtaining the cross-dimensional features includes:

[0028] Normalize the structured features to obtain the standardized features, use the standardized features as the input of the feature extraction model, and obtain the cross-dimensional features;

[0029] The method for obtaining the difficulty constraint includes:

[0030] Obtain the real-time cognitive load index. When the real-time cognitive load index is greater than the second preset load upper limit, adjust the maximum new knowledge ratio to the preset first threshold, and the first difficulty constraint interval is [d - X, d), where d is the user's current ability and X is the first adjustment value; when the real-time cognitive load index is less than the second preset load lower limit, adjust the maximum new knowledge ratio to the preset second threshold, and the second difficulty constraint interval is [d, d + Y), where Y is the second adjustment value.

[0031] Further, the method for obtaining the ability vector includes:

[0032] Weight the oral error density, speech rate volatility, coherence index, and phoneme error rate to obtain the oral proficiency;

[0033] Weight the vocabulary accuracy rate and the average number of reviews per week to obtain the vocabulary mastery rate;

[0034] According to the knowledge graph, obtain the knowledge point weights corresponding to the questions with the learning task type label of grammar in the structured features, and the knowledge point basic coefficients of the corresponding knowledge points in the knowledge graph. Weight the knowledge point weights and knowledge point basic coefficients of each knowledge point to obtain the grammar mastery degree;

[0035] Obtain the user's historical listening practice records, count the correct rates at different speech rates, obtain the highest speech rate with a correct rate not lower than the preset highest threshold as the recognition upper limit, and obtain the lowest speech rate with a correct rate not lower than the preset lowest threshold as the recognition lower limit; calculate the difference between the recognition upper limit and the recognition lower limit as the recognition speech rate range;

[0036] Count the listening text and the list of words the user has learned, count the number of unlearned new words, and calculate the ratio of the number of new words to the total number of words in the listening text as the new word frequency;

[0037] Weight the recognition speech rate range and the new word frequency to obtain the listening mastery rate;

[0038] Concatenate the oral proficiency, vocabulary mastery rate, grammar mastery degree, and listening mastery rate to obtain the ability vector.

[0039] Furthermore, the method for obtaining the corrected ability vector includes:

[0040] Calculate the memory retention rate based on the cumulative knowledge base coefficient and the correct rate of the last C questions, and obtain the corrected ability vector based on the product of the memory retention rate and the ability vector, where the cumulative knowledge base coefficient is obtained by weighting the knowledge base coefficients of the knowledge points already mastered by the user in the knowledge graph.

[0041] Furthermore, the method for obtaining the enhanced task queue includes:

[0042] Extract the corrected oral proficiency, corrected vocabulary mastery rate, corrected grammar mastery degree, and corrected listening comprehension threshold in the corrected ability vector, and calculate the grammar ability deviation degree and listening ability deviation degree corresponding to the corrected oral proficiency, corrected vocabulary mastery rate, corrected grammar mastery degree, and corrected listening comprehension threshold;

[0043] Calculate the cognitive load adaptability based on the cognitive load index;

[0044] After standardizing the deviation degrees of grammar ability and listening ability and taking the reciprocal, the standardized deviation degrees are obtained; after standardizing the knowledge point basic coefficients, the standard knowledge coefficients are obtained; after standardizing the cognitive load adaptability, the standard cognitive load is obtained; the standardized deviation degrees, standard knowledge coefficients, and standard cognitive loads are weighted to obtain the standard matrix; the maximum values of the standardized deviation degrees, standard knowledge coefficients, and standard cognitive loads are obtained, weighted to obtain the ideal solution, the minimum values of the standardized deviation degrees, standard knowledge coefficients, and standard cognitive loads are obtained, weighted to obtain the negative ideal solution; calculate the distances from the standard matrix to the ideal solution and the negative ideal solution respectively, calculate the closeness based on the distances from the standard matrix of each knowledge point to the ideal solution and the negative ideal solution, and sort the knowledge points in descending order based on the closeness to obtain the reinforcement task queue.

[0045] Further, the method for obtaining the corrected task queue includes:

[0046] Encode the knowledge points in the reinforcement task queue in ascending order of sorting, calculate the contribution weights of each knowledge point to the cross-dimensional features based on the attention mechanism, multiply the contribution weights by the encodings of the corresponding knowledge points in the reinforcement task queue to obtain the corrected encodings, re-sort the knowledge points in the reinforcement task queue in ascending order according to the corrected encodings, and use the re-sorted result as the corrected task queue.

[0047] Further, the method for obtaining the task weight matrix includes:

[0048] Count the ratio of the number of questions with the correctness label of 1 to the number of questions in the corresponding dimension in the learning tasks of different dimensions to obtain the accuracy rates of the learning tasks of different dimensions, and calculate the actual task load coefficients in combination with the preset task load coefficients; where different dimensions include speaking, vocabulary, grammar, and listening;

[0049] Calculate the updated load coefficient according to the preset task load coefficient, the cognitive load index at the current moment, and the smoothing factor;

[0050] Calculate the task weights in combination with the sensitivity coefficient, the cognitive load index at the current moment, the preset task load coefficient, and the updated load coefficient, and splice the task weights of different dimensions to obtain the task weight matrix.

[0051] Further, the method for obtaining the compound learning task includes:

[0052] Obtain the first N knowledge points in the corrected task queue, generate a task sequence, and design the task combination ratio in the task sequence in combination with the task weights to obtain the compound task;

[0053] Query the associated knowledge points in the knowledge graph, and design vertical progressive tasks and horizontal linkage tasks respectively. The vertical progressive tasks mean learning N knowledge points in sequence, and the horizontal linkage tasks mean learning N knowledge points simultaneously; obtain the user's current cognitive load index, design the time allocation ratio of learning tasks in different dimensions according to the knapsack algorithm, divide the user's cognitive load level according to the preset cognitive load level, and set the number of questions in different cognitive load levels in each dimension according to the preset ratio.

[0054] Further, the method for obtaining the time allocation ratio of the learning tasks in different dimensions includes:

[0055] Obtain the weight value of the learning tasks in each dimension, the difficulty constraint interval of each dimension, and the duration corresponding to the user's learning period. Take the learning tasks in each dimension as items, the value of the learning task is the task weight, and the weight of the learning task is the estimated time required to complete the learning task; when selecting tasks, ensure that the difficulty parameter of the selected learning task meets the preset difficulty constraint interval; on the premise that the total time does not exceed the estimated total time, select a combination of learning tasks to maximize the total weight; calculate the time proportion allocated to each task type according to the total weight of the task combination, count the time proportion of the learning tasks in different dimensions, and take the time proportion as the time allocation ratio of the learning tasks in different dimensions;

[0056] The method for obtaining the difficulty parameter of the learning task includes:

[0057] Obtain the user's most recent Z answering data, count the number of correct answers, calculate the ratio of the number of correct answers to Z, and obtain the user's answering accuracy rate;

[0058] Based on maximum likelihood estimation, build a model of user answering accuracy rate, user ability, and question difficulty. By fixing the question difficulty, calculate and obtain the user ability based on the user's answering accuracy rate; according to the user ability and the preset question difficulty, calculate the prediction accuracy rate, count the actual accuracy rate of the user's most recent M questions, and calculate the difficulty adjustment parameter based on the prediction accuracy rate and the actual accuracy rate. If the prediction accuracy rate is greater than the actual accuracy rate, update the difficulty parameter by adding the difficulty adjustment parameter; if the prediction accuracy rate is less than the actual accuracy rate, update the difficulty parameter by subtracting the difficulty adjustment parameter; wherein, the user ability is also updated according to the updated answering record.

[0059] The intelligent coaching adjustment system implements the intelligent coaching adjustment method, including:

[0060] Data collection module: Collect user data, behavior data, and physiological data;

[0061] The first analysis module: extract comprehensive features from user data, behavior data, and physiological data to obtain structured features; extract comprehensive features from behavior data and physiological data to obtain a cognitive load sequence;

[0062] The second analysis module: process the structured features to obtain cross-dimensional features; analyze the cognitive load sequence to obtain difficulty constraints;

[0063] The third analysis module: process the structured features according to a predefined knowledge graph to obtain an ability vector; correct the ability vector based on an improved Ebbinghaus model to obtain a corrected ability vector;

[0064] The fourth analysis module: analyze the structured features based on the TOPSIS multi-criteria decision-making algorithm according to the corrected ability vector to obtain a strengthened task queue; process the strengthened task queue according to the cross-dimensional features to obtain a corrected task queue;

[0065] The fifth analysis module: dynamically adjust the learning task weights according to the cognitive load index sequence to obtain a task weight matrix;

[0066] The task adjustment module: analyze according to the difficulty constraints, corrected task queue, and task weight matrix to obtain a composite learning task.

[0067] The technical effects and advantages of the intelligent coaching adjustment system and method of the present invention:

[0068] The present invention breaks through the limitations of fixed-script training by fusing physiological data to monitor cognitive load in real time, combining the Ebbinghaus model to correct the ability vector, and accurately matching the learning progress with the memory law; relying on the knowledge graph to locate the ability short board, combining load constraints to dynamically adjust the question difficulty and task rhythm, avoiding the problem of poor adaptability caused by standardized training; generating a strengthened task queue through the TOPSIS algorithm, combining cross-dimensional features and difficulty constraints to dynamically adjust the task weights to form a composite learning task, solving the problem of insufficient single-ability assessment; at the same time, iteratively updating the model parameters through real-time load sequences and learning effect data to ensure the effectiveness of the long-term learning process and make up for the defect of the lack of a model update mechanism. Description of the Drawings

[0069] Figure 1 It is a schematic diagram of the intelligent coaching adjustment method of Embodiment 1 of the present invention;

[0070] Figure 2 It is a schematic diagram of the intelligent coaching adjustment method of Embodiment 2 of the present invention;

[0071] Figure 3 It is a block diagram of the intelligent coaching adjustment system of Embodiment 3 of the present invention. Detailed Embodiments

[0072] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0073] Embodiment 1

[0074] Please refer to Figure 1 As shown, this embodiment provides an intelligent coaching adjustment method, including the following steps:

[0075] Collect user data, behavior data, and physiological data; user data includes oral recordings and answer records. The oral recordings are the oral practice audio collected by the microphone and the preset answer text. By analyzing the oral practice audio and the preset answer text, pronunciation and content integrity can be accurately evaluated, the feedback mechanism and practice content of oral coaching can be optimized, and the expression ability can be targeted improved; the answer records include learning task type tags (including grammar, vocabulary, and listening), question IDs, user answers, correctness tags (correct is 1, wrong is 0), and response times. Collecting answer records can locate knowledge weak points such as vocabulary, grammar, and listening based on the task type, answer right or wrong, and response time, and intelligently push matching intensive practice tasks to achieve precise learning; behavior data includes learning session timestamps, and the learning session timestamps include start timestamps and end timestamps, which can analyze the learning rhythm and concentration according to the learning session, dynamically adjust the practice duration and task interval, and optimize the coherence of the learning process; physiological data is the RR interval data collected by a wearable device such as a smart watch, which can monitor the cognitive load in real time, automatically adjust the difficulty of coaching questions, avoid too high or too low load, and maintain an efficient learning state.

[0076] Perform comprehensive feature extraction on user data, behavior data, and physiological data to obtain structured features, and perform comprehensive feature extraction on behavior data and physiological data to obtain a cognitive load sequence; by performing comprehensive feature extraction on user data, behavior data, and physiological data to form structured features, the user's knowledge mastery level and learning behavior pattern can be comprehensively characterized; at the same time, the cognitive load sequence extracted from behavior data and physiological data accurately reflects the user's real-time load state. Based on these data supports, intelligent coaching dynamically adjusts the difficulty of practice content and the learning rhythm, realizes personalized adaptation of the practice plan, and improves learning efficiency and experience.

[0077] The method for obtaining structured features includes:

[0078] Convert the spoken language recording into a spoken language text, compare it with the preset answer text, count the total number of words and the number of incorrect words in the spoken language text, calculate the ratio of the number of incorrect words to the total number of words, and obtain the spoken language error density;

[0079] Count the average value and standard deviation of the number of words per second in the spoken language text, calculate the ratio of the standard deviation of the number of words per second to the average value of the number of words per second, and obtain the speech rate volatility;

[0080] Count the number of occurrences of preset filler words and adjacent repeated phrases within a unit of time, calculate the ratio of the number of occurrences to the total number of words in the spoken language text, and obtain the coherence index;

[0081] Use the spoken language text as the input of a pre-trained speech model (such as Wav2Vec 2.0 or a localized model) to obtain the phoneme error rate;

[0082] Obtain the number of questions with the learning task type label of vocabulary and the correctness label of 1 in the most recent R questions, calculate the ratio of the number of questions with the correctness label of 1 to R, and obtain the vocabulary accuracy rate;

[0083] Count the total number of times the user has reviewed vocabulary in the most recent H weeks, calculate the ratio of the total number of reviews to H, and obtain the average number of reviews per week;

[0084] Obtain the number of questions with the learning task type label of grammar and the correctness label of 1 in the most recent E questions, calculate the ratio of the number of questions with the correctness label of 1 to E, and obtain the grammar accuracy rate;

[0085] Obtain all the question IDs with the correctness label of 0, obtain the knowledge point labels and the corresponding knowledge point basic coefficients corresponding to all the question IDs with the correctness label of 0 according to the predefined knowledge graph, count the number of errors corresponding to the same knowledge point label, calculate the product of the number of errors and the knowledge point basic coefficient to obtain the knowledge point weight; count all the knowledge point weights to obtain the defect matrix;

[0086] Obtain the response time of the preset first number of times, calculate the standard deviation and mean of the response time of the preset first number of times, calculate the ratio of the standard deviation of the response time of the preset first number of times to the mean, and obtain the answer time dispersion;

[0087] Obtain the physiological data of the preset second number of times, calculate the standard deviation and mean of the physiological data of the preset second number of times, calculate the ratio of the standard deviation of the physiological data of the preset second number of times to the mean, and obtain the physiological characteristics;

[0088] Weight the answer time dispersion and the physiological characteristics to obtain the cognitive load index; among them, the weighting coefficient can be optimized through a nature-inspired optimization algorithm;

[0089] Concatenate the spoken error density, speech rate volatility, coherence index, phoneme error rate, vocabulary accuracy, average weekly review times, grammar accuracy, defect matrix, cognitive load index, and the corresponding timestamps to obtain structured features.

[0090] The methods for obtaining the cognitive load sequence include:

[0091] Obtain the cognitive load index, data source label, and the corresponding timestamps in the structured features within a preset time period, and then concatenate them to obtain the cognitive load sequence.

[0092] Process the structured features to obtain cross-dimensional features; analyze the cognitive load sequence to obtain difficulty constraints. The cross-dimensional features obtained by processing the structured features can integrate multi-dimensional information such as the user's knowledge mastery and behavior patterns, and deeply depict the user's learning characteristics and needs. The difficulty constraints obtained by analyzing the cognitive load sequence define the range of question difficulties that the user's current cognition can accept. The intelligent tutoring relies on the cross-dimensional features to accurately locate the user's ability weaknesses, and dynamically matches the question difficulty in combination with the difficulty constraints. It can not only push intensive training for knowledge weak points, but also avoid low learning efficiency caused by excessive difficulty beyond the load, realizing scientific adjustment of the practice content, difficulty, and rhythm, ensuring that the learning process is both targeted and in line with the cognitive tolerance, and optimizing the learning experience and effect.

[0093] The methods for obtaining cross-dimensional features include:

[0094] Normalize the structured features to obtain standardized features, use the standardized features as the input of the feature extraction model, and obtain cross-dimensional features;

[0095] The training method of the feature extraction model includes:

[0096] Pre-collect P groups of training data, where the training data includes standardized features and the corresponding cross-dimensional features;

[0097] Use the standardized features as the input of the feature extraction model, and the cross-dimensional features as the output of the feature extraction model. With the goal of minimizing the error between the output cross-dimensional features and the actual cross-dimensional features, optimize the network parameters of the feature extraction model through a natural inspiration optimization algorithm, obtain the network parameters corresponding to minimizing the error between the cross-dimensional features output by the feature extraction model and the actual cross-dimensional features, and construct the feature extraction model with the corresponding network parameters as the trained feature extraction model.

[0098] The methods for obtaining difficulty constraints include:

[0099] Obtain the real-time cognitive load index. When the real-time cognitive load index is greater than the second preset load upper limit, adjust the maximum proportion of new knowledge to the preset first threshold. The first difficulty constraint interval is [d - X, d), where d is the user's current ability and X is the first adjustment value, such as 0.1. When the real-time cognitive load index is greater than the second preset load upper limit, it indicates that the cognitive pressure currently borne by the user is at a relatively high level. To avoid the user from experiencing learning fatigue or resistance due to excessive load, it is necessary to reduce the difficulty of the learning content at this time. Adjusting the maximum proportion of new knowledge to the preset first threshold can control the proportion of new knowledge in the learning content, thereby alleviating the cognitive load. The first difficulty constraint interval is set to [d - X, d), which means that the difficulty of the learning content will be limited to within a range of at most X (such as 0.1) based on the user's current ability d. This is because at high cognitive loads, appropriately reducing the difficulty helps the user better understand and absorb knowledge, making the learning process smoother. Through such settings, a negative correlation is established between the cognitive load index and the difficulty constraint interval, that is, as the cognitive load index exceeds the upper limit, the difficulty constraint interval is adjusted accordingly in the direction of lower difficulty to adapt to the user's current cognitive state and ensure the learning effect. When the real-time cognitive load index is less than the second preset load lower limit, adjust the maximum proportion of new knowledge to the preset second threshold, and the second difficulty constraint interval is [d, d + Y), where Y is the second adjustment value, such as 0.2.

[0100] Process the structured features according to the predefined knowledge graph to obtain the ability vector; correct the ability vector based on the improved Ebbinghaus model to obtain the corrected ability vector. Generating the ability vector by processing the structured features based on the knowledge graph can quantify the user's knowledge mastery level in four dimensions: speaking, vocabulary, grammar, and listening. Then, correct the ability vector through the memory retention rate, incorporate the memory forgetting law, and accurately evaluate knowledge retention and weak points. It can targetedly plan the practice direction according to the user's real ability state during the intelligent training process, such as preferentially strengthening the knowledge points with fast memory decline, ensuring that the training content not only fits the current ability but also meets the memory optimization requirements.

[0101] The methods for obtaining the ability vector include:

[0102] Weight the speaking error density, speech rate volatility, coherence index, and phoneme error rate to obtain the speaking mastery degree;

[0103] Weight the vocabulary accuracy rate and the average weekly review times to obtain the vocabulary mastery rate;

[0104] According to the knowledge graph, obtain the knowledge point weights corresponding to the questions with the learning task type label of grammar in the structured features, and the knowledge point basic coefficients of the corresponding knowledge in the knowledge graph. Weight the knowledge point weights and knowledge point basic coefficients of each knowledge point to obtain the grammar mastery degree;

[0105] Obtain the user's historical listening practice records, count the correct rates at different speech rates, obtain the highest speech rate with a correct rate not lower than the preset highest threshold as the recognition upper limit, and obtain the lowest speech rate with a correct rate not lower than the preset lowest threshold as the recognition lower limit; calculate the difference between the recognition upper limit and the recognition lower limit as the recognition speech rate range;

[0106] Count the listening text and the list of words the user has learned, count the number of unfamiliar words, and calculate the ratio of the number of unfamiliar words to the total number of words in the listening text as the unfamiliar word frequency;

[0107] Weight the recognition speech rate range and the unfamiliar word frequency to obtain the listening mastery rate;

[0108] Concatenate the speaking mastery degree, vocabulary mastery rate, grammar mastery degree, and listening mastery rate to obtain the ability vector.

[0109] The method for obtaining the corrected ability vector includes:

[0110] Calculate the memory retention rate based on the cumulative knowledge base coefficient and the correct rates of the last C questions, and obtain the corrected ability vector based on the product of the memory retention rate and the ability vector; if the memory retention rate Among them, τ is the cumulative knowledge base coefficient, which is obtained by weighting the knowledge point base coefficients of the knowledge points the user has mastered in the knowledge graph. The knowledge point base coefficient is determined according to the inherent attributes of the knowledge point, including difficulty (scored by educational experts based on the curriculum syllabus, 1-5 points) and importance (determined by the score ratio of this knowledge point in the exam syllabus, 1-5 points). The calculation formula for the cumulative knowledge base coefficient τ is τ = (0.6×difficulty score + 0.4×importance score - 1) / (5 - 1) (normalized to the 0-1 range); the cumulative knowledge base coefficient can also be corrected according to the user's historical learning performance. For example, if the accuracy rate of the user's last 3 times of learning this knowledge point ≥ 80%, then correct the cumulative knowledge base coefficient τ ' = τ×0.9, if the accuracy rate < 50%, then correct the cumulative knowledge base coefficient τ '= τ × 1.1; γ is the correct rate of the most recent C questions; Δt is the acquisition time interval of the most recent C questions, and exp(·) represents the exponential function; the parameter 1.5 is an empirical correction factor, which is obtained by fitting the experimental data of the memory retention rate of 1000 learners. By comparing the deviation between the actual memory retention rate and the theoretical value (without correction) (which can be calculated by an improved Ebbinghaus model), it is found that when the correction factor is 1.5, the mean square error between the model prediction value and the actual value is the smallest (error ≤ 3%). Therefore, this parameter is selected; the corrected ability vector GS' = GS × MR; where GS is the ability vector; the memory retention rate formula combines the cumulative knowledge base coefficient τ, the recent correct rate γ, and the time interval Δt, and simulates the Ebbinghaus forgetting curve through the exponential function to quantify the degree of knowledge forgetting over time. For example, if the user has a high cumulative knowledge base coefficient and a high recent answering correct rate, the formula will calculate a higher memory retention rate, otherwise it reflects the risk of knowledge forgetting; by correcting the ability vector with the memory retention rate, the ability assessment can not only reflect the current knowledge level but also incorporate the memory factor. For example, if the original ability vector of the user shows that a certain dimension is well mastered but the memory retention rate is low (such as not reviewed for a long time), the weight of this dimension will be reduced in the corrected ability vector, which more realistically reflects the "retrievable state" of knowledge and provides a more accurate basis for subsequent learning task planning.

[0111] Based on the TOPSIS multi-criteria decision-making algorithm, analyze the structured features according to the corrected ability vector to obtain the enhanced task queue; process the enhanced task queue according to the cross-dimensional features to obtain the corrected task queue. Use the TOPSIS multi-criteria decision-making algorithm to analyze the corrected ability vector, screen out the enhanced task queue with high priority, and clarify the key points of knowledge consolidation; then further process it in combination with the cross-dimensional features to eliminate unmatched tasks and supplement appropriate tasks to form the corrected task queue. This makes the task push of intelligent coaching more scientific, focusing on both the core weak points and the balance of learning dimensions.

[0112] The method for obtaining the enhanced task queue includes:

[0113] Extract the corrected oral proficiency, corrected vocabulary mastery rate, corrected grammar proficiency, and corrected listening comprehension threshold in the corrected ability vector, and calculate the grammar ability deviation and listening ability deviation corresponding to the corrected oral proficiency, corrected vocabulary mastery rate, corrected grammar proficiency, and corrected listening comprehension threshold;

[0114] Calculate the cognitive load adaptability based on the cognitive load index, such as Among them, CLIM is cognitive load adaptability; parameter 5 is the normalization coefficient, which is used to adjust the sensitivity of the cognitive load index CLI to the cognitive load adaptability CLIM; when parameter = 5, the actual distribution of cognitive load adaptability is 0.076-0.924 points (covering 90% of user scenarios). This normalization parameter makes the impact of CLI changes on CLIM neither too gentle (such as the normalization parameter is too small, CLIM discrimination is insufficient), nor too extreme (such as the normalization parameter is too large, CLIM is prone to concentrate at both ends), retaining the discrimination of the score, meeting the actual needs of "adaptability" evaluation; parameter 0.5 is used to calibrate the center position of the function, so that when CLI = 0.5, CLIM is in the middle state, balancing the impact of high and low CLI values on CLIM, and ensuring that the evaluation logic conforms to the setting of "negative correlation between cognitive load and adaptability" (that is, the higher the CLI, the lower the CLI); CLI is the cognitive load index, CLI = ω CLD ×CLD+ω CLP ×CLP, where ω CLD is the CLD weight, which can be obtained by optimizing the natural inspiration optimization algorithm; CLD is the answer time discreteness (normalized to the range of 0-1); ω CLP is the CLP weight, which can be obtained by searching for the optimal value through the nature-inspired optimization algorithm; CLP is the comprehensive value of physiological characteristics (standardized to the range of 0-1); the reason for selecting RR interval (heart rate variability index) as the physiological characteristic is that, according to medical research, the correlation between RR interval and cognitive load (such as Pearson correlation coefficient 0.82) is significantly higher than other physiological indicators (such as skin galvanic response correlation coefficient 0.75). For a very small number of users whose RR interval has no significant change, the system automatically switches to skin galvanic response data and adjusts the calculation method of CLP to CLP = 0.5 × skin galvanic response + 0.5 × pupil diameter change rate (the correlation of the two indicators is ≥ 0.7, and other physiological indicators with correlation higher than 0.7 can also be selected according to the actual calculation results); by converting the cognitive load index into cognitive load adaptability, the user's current load state can be quantitatively evaluated; by mapping the abstract cognitive load index to an adaptability score in the [0,1] interval, the matching degree between task difficulty and user load can be intuitively reflected; and the higher the cognitive load index, the lower the cognitive load adaptability score, and the system can dynamically adjust the task difficulty accordingly. For example, when the CLI is higher than the preset CLI threshold, the CLIM score is low. At this time, the push of high-difficulty tasks is reduced to prevent users from experiencing learning fatigue due to overload. Conversely, if the load is low, challenging tasks are added to make full use of cognitive resources and balance learning efficiency and experience.

[0115] After standardizing the deviation degrees of grammar ability and listening ability and taking the reciprocal, the standardized deviation degree is obtained; after standardizing the knowledge point basic coefficient, the standard knowledge coefficient is obtained; after standardizing the cognitive load adaptability, the standard cognitive load is obtained; the standardized deviation degree, the standard knowledge coefficient and the standard cognitive load are weighted to obtain the standard matrix; the maximum values of the standardized deviation degree, the standard knowledge coefficient and the standard cognitive load are obtained, weighted to obtain the ideal solution, the minimum values of the standardized deviation degree, the standard knowledge coefficient and the standard cognitive load are obtained, weighted to obtain the negative ideal solution; the distances from the standard matrix to the ideal solution and the negative ideal solution are calculated respectively, the closeness degree is calculated according to the distances from the standard matrix of each knowledge point to the ideal solution and the negative ideal solution, and the knowledge points are sorted in descending order based on the closeness degree to obtain the reinforcement task queue.

[0116] The method for obtaining the corrected task queue includes:

[0117] The knowledge points in the reinforcement task queue are encoded in ascending order of sorting. Based on the attention mechanism, the contribution weights of each knowledge point to the cross-dimensional features are calculated. The contribution weights are multiplied by the encodings of the corresponding knowledge points in the reinforcement task queue to obtain the corrected encodings. The knowledge points in the reinforcement task queue are re-sorted in ascending order according to the corrected encodings, and the re-sorted result is used as the corrected task queue.

[0118] According to the cognitive load index sequence, the learning task weights are dynamically adjusted to obtain the task weight matrix; according to the cognitive load index sequence, the learning task weights are dynamically adjusted to match the user's current cognitive load state in real time. When the load is high, the weights of complex tasks are reduced, and when the load is low, the weights of challenging tasks are increased, so that the task allocation of the intelligent companion is more in line with the user's physiological and psychological tolerance, avoiding learning fatigue or low efficiency caused by unreasonable task weights, and maintaining the comfort and concentration of the learning process.

[0119] The method for obtaining the task weight matrix includes:

[0120] The ratio of the number of questions with the correctness label of 1 to the number of questions in the corresponding dimension in the learning tasks of different dimensions is statistically obtained to obtain the accuracy rates of the learning tasks in different dimensions. The actual task load coefficient is calculated by combining the preset task load coefficient; among them, different dimensions include speaking, vocabulary, grammar and listening.

[0121] The updated load coefficient is calculated based on a preset task load coefficient, the cognitive load index at the current moment, and a smoothing factor; the smoothing factor is used to adjust the cognitive load index at the current moment and the preset task load coefficient, making the calculation result smoother and more reasonable, avoiding drastic changes in the updated load coefficient caused by instantaneous fluctuations in the cognitive load index, making the entire calculation process smoother, and reducing the adverse effects brought by data mutations.

[0122] The task weight is calculated by combining the sensitivity coefficient, the cognitive load index at the current moment, the preset task load coefficient, and the updated load coefficient, and the task weight matrix is obtained by splicing the task weights of different dimensions; the sensitivity coefficient is used to measure the sensitivity of the task weight to changes in the cognitive load index at the current moment and the preset task load coefficient; it determines to what extent the changes in the cognitive load index and the preset task load coefficient can affect the final task weight result when calculating the task weight. The higher the sensitivity coefficient, the more sensitive the task weight is to changes in these factors, and vice versa.

[0123] Analyze according to the corrected task queue, the task weight matrix, and the difficulty constraint to obtain a composite learning task. Integrate the difficulty constraint (defining the acceptable question difficulty range), the corrected task queue (clarifying the core practice content), and the task weight matrix (matching the task priorities adapted to the load), and finally generate a composite learning task. This enables the intelligent training partner to output a personalized task combination with "reasonable difficulty, accurate content, and scientific weight", which not only guarantees the knowledge reinforcement effect but also adapts to the cognitive load, realizes the deep fit of the learning content, rhythm, and the user's state, and maximizes the learning efficiency.

[0124] The methods for obtaining a composite learning task include:

[0125] Obtain the first N knowledge points in the corrected task queue, generate a task sequence, combine the task weights, design the task combination ratio in the task sequence, and obtain a composite task.

[0126] Query the associated knowledge points in the knowledge graph, and design vertical progressive tasks and horizontal linkage tasks respectively. The vertical progressive task means learning N knowledge points in sequence, and the horizontal linkage task means learning N knowledge points simultaneously; obtain the user's current cognitive load index, design the time allocation ratio of learning tasks in different dimensions according to the knapsack algorithm, divide the user's cognitive load level according to the preset cognitive load level, and set the number of questions at different cognitive load levels in each dimension according to the preset ratio. The design ratio of each dimension can be optimized based on the natural inspiration optimization algorithm.

[0127] The methods for obtaining the time allocation ratio of learning tasks in different dimensions include:

[0128] Obtain the weight values of the learning tasks for each dimension, the difficulty constraint intervals for each dimension, and the duration corresponding to the user's learning period. Consider the learning tasks for each dimension as items, where the value of a learning task is the task weight and the weight of a learning task is the time required to complete the corresponding learning task. When selecting tasks, ensure that the difficulty parameters of the selected learning tasks meet the preset difficulty constraint intervals. On the premise that the total time does not exceed the available time, select a combination of learning tasks to maximize the total weight. Calculate the proportion of the time allocated to each task type based on the total weight of the task combination, and count the proportion of the time of the learning tasks in different dimensions. Use the proportion of time as the time allocation ratio of the learning tasks in different dimensions.

[0129] The methods for obtaining the difficulty parameters of learning tasks include:

[0130] Obtain the user's most recent Z question-answering data, count the number of correct answers, calculate the ratio of the number of correct answers to Z, and obtain the user's question-answering accuracy rate.

[0131] Based on maximum likelihood estimation, build a model of the user's question-answering accuracy rate, user ability, and question difficulty. By fixing the question difficulty, calculate and obtain the user's ability based on the user's question-answering accuracy rate.

[0132] According to the user's ability and the preset question difficulty, calculate the prediction accuracy rate, count the actual accuracy rate of the user's most recent M questions, calculate the difficulty adjustment parameter based on the prediction accuracy rate and the actual accuracy rate. If the prediction accuracy rate is greater than the actual accuracy rate, update the difficulty parameter by adding the difficulty adjustment parameter to obtain the difficulty parameter of the learning task. If the prediction accuracy rate is less than the actual accuracy rate, update the difficulty parameter by subtracting the difficulty adjustment parameter to obtain the difficulty parameter of the learning task. Among them, the user's ability is also updated according to the updated question-answering record.

[0133] Embodiment 2

[0134] Refer to Figure 2 , this embodiment provides a composite learning task optimization method, including the following steps:

[0135] Collect interruption events, where the interruption events consist of a timestamp sequence of pause events and continue events;

[0136] Optimize the structured features according to the interruption events to obtain optimized structured features;

[0137] The methods for obtaining optimized structured features include:

[0138] Obtain the learning period timestamp and interruption events, aggregate them through a preset time window, count the learning duration within the preset time window, calculate the ratio of the learning duration to the preset time of the preset time window, and obtain the proportion of the effective learning duration;

[0139] Count the number of interruption events occurring within a preset time window, and calculate the ratio of the number of occurrences to the preset time window to obtain the interruption frequency;

[0140] Concatenate the proportion of effective learning duration and the interruption frequency with the structured features to obtain optimized structured features.

[0141] Process the proportion of effective learning duration and the historical learning period in the optimized structured features respectively to obtain the recommended learning period; process the interruption frequency and the historical record of task types in the optimized structured features to obtain the interruption risk score of the learning task;

[0142] The methods for obtaining the recommended learning period include:

[0143] Divide the daily time periods, obtain the daily time period to which the learning period timestamp belongs, count the learning period with the highest proportion of effective learning duration among all daily time periods, use the learning period with the highest proportion of effective learning duration as the recommended learning period, and concatenate the start time, end time, and proportion of effective duration of the recommended learning period to obtain the learning endurance feature.

[0144] The methods for obtaining the interruption risk score include:

[0145] Obtain the interruption frequency in the optimized structured features and the corresponding learning task type labels, and use logistic regression by the user to predict the interruption probability of different task types, and use the interruption probability as the interruption risk score; as where, P(n i |i) is the interruption probability of the i-th task type; e is the mathematical constant; α is the adjustment coefficient, obtained by optimizing based on the nature-inspired optimization algorithm, used as the baseline offset of the model to adjust the basic level of the interruption probability; β is the frequency coefficient, obtained by optimizing based on the nature-inspired optimization algorithm, used to measure the influence weight of the interruption frequency on the interruption probability; if the absolute value of β is large, it means that the influence of the interruption frequency on the interruption probability is significant, that is, the higher the interruption frequency, the more obvious the increase in the interruption probability; conversely, if β approaches 0, the influence of the interruption frequency on the interruption probability can be ignored. The two work together to make the model more accurately capture the association between the task type and the interruption behavior and improve the reliability of the interruption risk prediction; n i is the interruption frequency of the i-th task type; n avg is the user's historical average number of interruptions (taking the average of the last 10 learning sessions); ∈ is an extremely small positive number, used to avoid n avg= 0. Converting the interruption frequency into a measurable interruption risk score can intuitively reflect the likelihood of different task types triggering user interruption of learning. For example, if the calculated interruption probability of a certain type of task (such as high-difficulty grammar questions) is high, it indicates that users are prone to interruption when performing this task, which can provide a decision-making basis for the intelligent tutoring system. For task types with high interruption risks, adjust the task difficulty, duration, or presentation method (such as splitting complex tasks) to reduce the probability of learning interruption and enhance learning continuity.

[0146] Optimize the composite learning tasks according to the recommended learning time period and interruption risk score;

[0147] The methods for optimizing the composite learning tasks include:

[0148] The available time of the composite learning task is obtained according to the recommended learning time period. Interruption monitoring is implemented according to the interruption risk score. If the current interruption risk score is greater than the preset interruption probability threshold, a rest reminder is automatically inserted or the learning task type is switched.

[0149] Embodiment 3

[0150] Please refer to Figure 3 As shown, this embodiment provides an intelligent tutoring adjustment system, including:

[0151] Data acquisition module: Acquire user data, behavior data, and physiological data;

[0152] First analysis module: Extract comprehensive features from user data, behavior data, and physiological data to obtain structured features; extract comprehensive features from behavior data and physiological data to obtain a cognitive load sequence;

[0153] Second analysis module: Process the structured features to obtain cross-dimensional features; analyze the cognitive load sequence to obtain difficulty constraints;

[0154] Third analysis module: Process the structured features according to the predefined knowledge graph to obtain an ability vector; correct the ability vector based on the improved Ebbinghaus model to obtain a corrected ability vector;

[0155] Fourth analysis module: Analyze the structured features based on the TOPSIS multi-criteria decision-making algorithm according to the corrected ability vector to obtain a strengthened task queue; process the strengthened task queue according to the cross-dimensional features to obtain a corrected task queue;

[0156] Fifth analysis module: Dynamically adjust the learning task weights according to the cognitive load index sequence to obtain a task weight matrix;

[0157] Task adjustment module: Analyze according to the difficulty constraints, corrected task queue, and task weight matrix to obtain a composite learning task.

[0158] As described above, it is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the said claims.

[0159] Finally: The above is only the preferred embodiment of the present invention and is not used to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. Intelligent coaching adjustment method, characterized in that It includes the following steps: Collect user data, behavior data, and physiological data; Extract comprehensive features from the user data, behavior data, and physiological data to obtain structured features; Extract comprehensive features from the behavior data and physiological data to obtain a cognitive load sequence; Process the structured features to obtain cross-dimensional features; Analyze the cognitive load sequence to obtain difficulty constraints; Process the structured features according to a predefined knowledge graph to obtain an ability vector; correct the ability vector based on an improved Ebbinghaus model to obtain a corrected ability vector; Based on the TOPSIS multi-criteria decision-making algorithm, analyze the structured features according to the corrected ability vector to obtain a strengthened task queue; Process the strengthened task queue according to the cross-dimensional features to obtain a corrected task queue; Dynamically adjust the learning task weights according to the cognitive load index sequence to obtain a task weight matrix; Analyze according to the difficulty constraints, corrected task queue, and task weight matrix to obtain a composite learning task.

2. The intelligent accompaniment training adjustment method according to claim 1, wherein The method for obtaining the structured features includes: Convert the oral recording into oral text, compare it with the preset answer text, count the total number of words and the number of incorrect words in the oral text, calculate the ratio of the number of incorrect words to the total number of words to obtain the oral error density; Statistically calculate the average value and standard deviation of the number of words per second in the oral text, calculate the ratio of the standard deviation of the number of words per second to the average value of the number of words per second to obtain the speech rate volatility; count the number of occurrences of preset filler words and adjacent repeated phrases within a unit time, calculate the ratio of the number of occurrences to the total number of words in the oral text to obtain a coherence index; use the oral text as the input of a pre-trained speech model to obtain the phoneme error rate; Obtain the number of questions with the learning task type label of vocabulary and the correctness label of 1 in the most recent R questions, calculate the ratio of the number of questions with the correctness label of 1 to R to obtain the vocabulary accuracy rate; count the total number of times the user has reviewed vocabulary in the most recent H weeks, calculate the ratio of the total number of review times to H to obtain the average weekly review times; Obtain the number of questions with the learning task type label of grammar and the correctness label of 1 in the most recent E questions, calculate the ratio of the number of questions with the correctness label of 1 to E to obtain the grammar accuracy rate; Obtain the IDs of all questions with the correctness label of 0, obtain the knowledge point labels and corresponding knowledge point basic coefficients corresponding to the IDs of all questions with the correctness label of 0 according to a predefined knowledge graph, count the number of errors corresponding to the same knowledge point label, calculate the product of the number of errors and the knowledge point basic coefficient to obtain the knowledge point weight; count all the knowledge point weights to obtain a defect matrix; Obtain the response times of a preset first number of times, calculate the standard deviation and mean of the response times of the preset first number of times, calculate the ratio of the standard deviation of the response times of the preset first number of times to the mean to obtain the answer time dispersion; obtain the physiological data of a preset second number of times, calculate the standard deviation and mean of the physiological data of the preset second number of times, calculate the ratio of the standard deviation of the physiological data of the preset second number of times to the mean to obtain the physiological characteristics; weight the answer time dispersion and the physiological characteristics to obtain the cognitive load index; Concatenate the spoken error density, speech rate volatility, coherence index, phoneme error rate, vocabulary accuracy, average weekly review times, grammar accuracy, defect matrix, and cognitive load index, as well as the corresponding timestamps, to obtain structured features.

3. The intelligent coaching adjustment method according to claim 2, wherein The method for obtaining the cognitive load sequence includes: Obtain the cognitive load index, data source label, and corresponding timestamp in the structured features within a preset time period, and then concatenate them to obtain the cognitive load sequence; The method for obtaining cross-dimensional features includes: Normalize the structured features to obtain standardized features, use the standardized features as the input of the feature extraction model, and obtain cross-dimensional features; The method for obtaining difficulty constraints includes: Obtain the real-time cognitive load index. When the real-time cognitive load index is greater than the second preset load upper limit, adjust the maximum proportion of new knowledge to the preset first threshold. The first difficulty constraint interval is [d - X, d), where d is the user's current ability and X is the first adjustment value. When the real-time cognitive load index is less than the second preset load lower limit, adjust the maximum proportion of new knowledge to the preset second threshold. The second difficulty constraint interval is [d, d + Y), where Y is the second adjustment value.

4. The intelligent coaching adjustment method according to claim 3, characterized in that The method for obtaining the ability vector includes: Weight the spoken error density, speech rate volatility, coherence index, and phoneme error rate to obtain the spoken language mastery degree; Weight the vocabulary accuracy and average weekly review times to obtain the vocabulary mastery rate; According to the knowledge graph, obtain the knowledge point weights corresponding to the questions with the learning task type label of grammar in the structured features, and the knowledge point basic coefficients of the corresponding knowledge points in the knowledge graph. Weight the knowledge point weights and knowledge point basic coefficients of each knowledge point to obtain the grammar mastery degree; Obtain the user's historical listening practice records, count the correct rates at different speech rates, obtain the highest speech rate with a correct rate not lower than the preset highest threshold as the recognition upper limit, and obtain the lowest speech rate with a correct rate not lower than the preset lowest threshold as the recognition lower limit. Calculate the difference between the recognition upper limit and the recognition lower limit as the recognition speech rate range; Count the listening text and the user's learned word list, count the number of unlearned new words, and calculate the ratio of the number of new words to the total number of words in the listening text as the new word frequency; Weight the recognition speech rate range and the new word frequency to obtain the listening mastery rate; Concatenate the spoken language mastery degree, vocabulary mastery rate, grammar mastery degree, and listening mastery rate to obtain the ability vector.

5. The intelligent coaching adjustment method according to claim 4, characterized in that, The method for obtaining the corrected ability vector includes: Calculate the memory retention rate based on the cumulative knowledge base coefficient and the correct rates of the last C questions, and obtain the corrected ability vector based on the product of the memory retention rate and the ability vector, where the cumulative knowledge base coefficient is obtained by weighting the knowledge base coefficients of the knowledge points already mastered by the user in the knowledge graph.

6. The intelligent accompaniment training adjustment method according to claim 5, wherein The method for obtaining the reinforcement task queue includes: Extract the corrected spoken language mastery degree, corrected vocabulary mastery rate, corrected grammar mastery degree, and corrected listening comprehension threshold in the corrected ability vector, and calculate the grammar ability deviation degree and listening ability deviation degree corresponding to the corrected spoken language mastery degree, corrected vocabulary mastery rate, corrected grammar mastery degree, and corrected listening comprehension threshold; Calculate the cognitive load adaptability based on the cognitive load index; After standardizing the deviation degrees of grammar ability and listening ability and taking the reciprocal, the standardized deviation degrees are obtained; after standardizing the knowledge point basic coefficients, the standard knowledge coefficients are obtained; after standardizing the cognitive load adaptability, the standard cognitive load is obtained; the standardized deviation degrees, standard knowledge coefficients and standard cognitive load are weighted to obtain the standard matrix; the maximum values of the standardized deviation degrees, standard knowledge coefficients and standard cognitive load are obtained, weighted to obtain the ideal solution, the minimum values of the standardized deviation degrees, standard knowledge coefficients and standard cognitive load are obtained, weighted to obtain the negative ideal solution; calculate the distances from the standard matrix to the ideal solution and the negative ideal solution respectively, calculate the closeness degree according to the distances from the standard matrix of each knowledge point to the ideal solution and the distances from the standard matrix to the negative ideal solution, and sort the knowledge points in descending order based on the closeness degree to obtain the reinforcement task queue.

7. The intelligent coaching adjustment method according to claim 6, wherein The method for obtaining the corrected task queue includes: Encoding the knowledge points in the reinforcement task queue in ascending order, calculating the contribution weights of each knowledge point to the cross-dimensional features based on the attention mechanism, multiplying the contribution weights by the encodings of the corresponding knowledge points in the reinforcement task queue to obtain the corrected encodings, and re-sorting the knowledge points in the reinforcement task queue in ascending order according to the corrected encodings, and taking the re-sorted result as the corrected task queue.

8. The intelligent coaching adjustment method according to claim 7, characterized in that The method for obtaining the task weight matrix includes: Counting the ratio of the number of questions with the correctness label of 1 to the number of questions in the corresponding dimension in the learning tasks of different dimensions to obtain the accuracy rates of the learning tasks of different dimensions, and calculating the actual task load coefficients in combination with the preset task load coefficients; where the different dimensions include speaking, vocabulary, grammar and listening; Calculating the updated load coefficient according to the preset task load coefficient, the cognitive load index at the current moment and the smoothing factor; Calculating the task weights in combination with the sensitivity coefficient, the cognitive load index at the current moment, the preset task load coefficient and the updated load coefficient, and splicing the task weights of different dimensions to obtain the task weight matrix.

9. The intelligent coaching adjustment method according to claim 8, wherein, The method for obtaining the composite learning task includes: Obtaining the first N knowledge points in the corrected task queue, generating a task sequence, and designing the task combination ratio in the task sequence in combination with the task weights to obtain the composite task; Querying the associated knowledge points in the knowledge graph, designing the vertical progression tasks and horizontal linkage tasks respectively, where the vertical progression tasks represent learning N knowledge points in sequence, and the horizontal linkage tasks represent learning N knowledge points simultaneously; obtaining the user's current cognitive load index, designing the time allocation ratio of the learning tasks of different dimensions according to the knapsack algorithm, dividing the user's cognitive load level according to the preset cognitive load level, and setting the number of questions in different cognitive load levels in each dimension according to the preset ratio.

10. An intelligent coaching adjustment system that implements the intelligent coaching adjustment method according to any one of claims 1-9, characterized in that, Including: Data acquisition module: acquiring user data, behavior data and physiological data; First analysis module: comprehensively extracting the features of the user data, behavior data and physiological data to obtain the structured features; Comprehensively extracting the features of the behavior data and physiological data to obtain the cognitive load sequence; The second analysis module: processes the structured features to obtain cross-dimensional features; analyzes the cognitive load sequence to obtain difficulty constraints; The third analysis module: processes the structured features according to the predefined knowledge graph to obtain an ability vector; corrects the ability vector based on the improved Ebbinghaus model to obtain a corrected ability vector; The fourth analysis module: analyzes the structured features based on the TOPSIS multi-criteria decision-making algorithm according to the corrected ability vector to obtain a strengthened task queue; processes the strengthened task queue according to the cross-dimensional features to obtain a corrected task queue; The fifth analysis module: dynamically adjusts the learning task weights according to the cognitive load index sequence to obtain a task weight matrix; The task adjustment module: analyzes according to the difficulty constraints, the corrected task queue, and the task weight matrix to obtain a composite learning task.

Images