Ancient poetry multi-mode teaching system based on HSK classification

Abstract

The invention discloses an ancient poetry multi-mode teaching system based on HSK classification, and belongs to the field of intelligent teaching. Comprising a learning evaluation portrait module, a learning path generation module, a physiological emotion monitoring module, a cognitive load analysis module, a vocabulary grammar enhancement module, a modal semantic analysis module, an artistic conception analogy generation module, a holographic scene generation module, a feedback interaction experience module, a content dynamic adjustment module, a recitation pronunciation correction module and an expansion knowledge generation module. A performance analysis feedback module and a review memory enhancement module; according to the method, learning tasks can be updated in real time along with the change of the ability of students, a one-step teaching mode is avoided, the learning efficiency and adaptability are improved, the learning experience and continuity are greatly improved, cultural faults can be effectively reduced, non-native Chinese language persons can understand the deep significance of ancient poetry more easily, and the learning efficiency is improved. The memory depth of the learner to the poetry artistic conception and vocabularies is greatly improved, and the learning motivation is also improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of smart teaching, in particular to a multi-modal teaching system for ancient poetry based on HSK grading. BACKGROUND

[0002] As an important carrier of Chinese traditional culture, ancient poetry contains highly condensed language structure, profound cultural connotation and multi-dimensional aesthetic characteristics. However, for learners whose second language is Chinese, ancient poetry learning often faces multiple challenges: the language structure is different from modern Chinese, the images in the poem are highly abstract, the cultural metaphors are dense, and the cultural background of the learners' mother tongue is significantly different, which limits the understanding of the content and the experience of the artistic conception. At the same time, learners of different HSK levels differ significantly in vocabulary, grammar mastery, cognitive load bearing capacity, etc., and the one-size-fits-all teaching method cannot meet the real learning needs. With the rapid development of artificial intelligence, quantum computing, virtual reality, neural symbolic reasoning, multi-modal interaction and other technologies, ancient poetry teaching is undergoing a revolution from "text indoctrination" to "intelligence, immersion, emotional resonance and cultural transfer". Therefore, the present application proposes a multi-modal teaching system for ancient poetry based on HSK grading.

[0003] The existing multi-modal teaching system for ancient poetry cannot update the learning tasks in real time as the students' ability changes, reduces the learning efficiency and adaptability, and has low learning experience and continuity, and there are many cultural gaps, and the depth of learners' memory of the artistic conception and vocabulary of the poem is reduced. Therefore, we propose a multi-modal teaching system for ancient poetry based on HSK grading. SUMMARY

[0004] The purpose of the present application is to solve the defects in the prior art, and to propose a multi-modal teaching system for ancient poetry based on HSK grading.

[0005] In order to achieve the above-mentioned purpose, the present application adopts the following technical solutions: The multi-modal teaching system for ancient poetry based on HSK grading comprises a learning evaluation portrait module, a learning path generation module, a physiological emotion monitoring module, a cognitive load analysis module, a vocabulary and grammar strengthening module, a modal semantic analysis module, an artistic conception analogy generation module, a holographic scene generation module, a feedback interaction experience module, a content dynamic adjustment module, a recitation pronunciation correction module, an extended knowledge generation module, a performance analysis feedback module and a review and memory strengthening module. The learning evaluation portrait module is used to generate individualized learning portraits of each student according to their HSK level, vocabulary, understanding of ancient poetry and mother tongue cultural background. The learning path generation module processes the individualized learning portraits of each student, predicts potential learning bottlenecks, and automatically plans individualized ancient poetry learning paths. The physiological emotion monitoring module collects physiological data of the learner in real time through a wearable device, and judges the learning emotional state of the student; The cognitive load analysis module is used for simulating the cognitive load of the student, and comprehensively judging the learning pressure of each student; The vocabulary and grammar strengthening module automatically screens key words in the poem according to the HSK level, and assists the student in reading and following reading training through voice recognition; The modal semantic analysis module is used for semantic decomposition of the ancient poem, and generates content explanation convenient for understanding; The artistic conception analogy generation module is used for constructing a cultural knowledge graph, and reasoning the cultural artistic conception behind the poem based on the cultural knowledge graph, and migrating it to the student's mother tongue culture system; The holographic scene generation module constructs a three-dimensional scene matched with the content of the poem through holographic projection according to the HSK level of the student; The feedback interaction experience module combines the holographic scene, and dynamically controls the interaction density and task complexity of the student according to different HSK levels; The content dynamic adjustment module dynamically adjusts the teaching strategy according to the real-time monitored emotional fluctuation of the student; The recitation pronunciation correction module is used for providing the student with standard recitation of the poem, voice rhythm analysis, rhyme perception learning, and automatically correcting the tone and pause problems in the student's recitation; The expanded knowledge generation module provides cultural expansion content corresponding to the artistic conception of the poem after the student understands the core content of the poem; The performance analysis feedback module is used for automatically recording the learning data of the student, and generating a visual report for the student to understand the progress and weak knowledge points of the student; The review and memory strengthening module draws the forgetting curve of each student, and automatically arranges the review time and content.

[0006] As a further scheme of the present application, the specific steps of the learning evaluation portrait module generating a personalized learning portrait of each student are as follows: S1.1: Collect HSK level, vocabulary estimate, ancient poem diagnostic test score and mother tongue culture label distribution from the student side, mark the data exceeding µ±3σ in each type of raw data as an abnormal value based on the 3σ principle, then respectively detect the range of each type of raw data, and mark the maximum and minimum values of each type of raw data; S1.2: Normalize each type of raw data by Min-Max normalization to obtain the standardized features of each type of data, based on expert experience, specify initial weights for each ability dimension of the student's language foundation, vocabulary mastery, poetic semantic understanding and cultural sensitivity, then group the standardized features according to the specified initial weights according to the ability dimensions, and linearly weight and sum the standardized features in the same dimension to obtain the ability score of the dimension; S1.3: Select a set of diagnostic questions for understanding ancient poetry, which contains different difficulty and different image types, and record the student's answer results and answer time for each question, and establish a probability model for each question based on the student's ability score, evaluate the student's understanding probability for each question, and according to the question answer probability, the difficulty parameters and the discrimination of the question are backstepped, then the probability results of the questions are summarized and fed back to the student's ability dimension; S1.4: Encode the student's cultural background questionnaire and cultural contact record into cultural semantic vector, and map the cultural image of the target ancient poetry into the same space image vector through expert annotation or semantic embedding, then calculate the cosine similarity between each cultural semantic vector and each image vector, if the cosine similarity is higher than the preset threshold, it means that the poetic conception is easier to find the corresponding analogy in the student's mother tongue culture, otherwise, it indicates that cross-cultural interpretation is needed, and each cosine similarity result is used as the inverse vector of the student's cultural transfer difficulty; S1.5: According to the student's answer rate and accuracy on the diagnostic question set, calculate the student's average reaction time and unit question error rate behavior indicators, map the two groups of behavior indicators to an instant risk score through linear regression prediction, and according to the preset threshold interval, label the risk score as high, medium and low three levels and output the corresponding confidence interval; S1.6: Collect the student's ability score, question level understanding probability summary statistics, cultural similarity and risk score portrait signals, based on the historical calibration results, assign a fusion weight to each type of portrait signal, and linearly fuse according to the weight to obtain the comprehensive score of each student's multi-dimensional portrait dimension, and generate corresponding visual labels according to the student's portrait dimension, wherein each label is accompanied by a confidence interval and the weight proportion of the constituent elements.

[0007] As a further scheme of the present application, the various types of raw data in S1.1 are collected through the student's registration information, vocabulary test, understanding questions for a number of poetic sentences, and cultural background questionnaire.

[0008] The specific calculation formula of the Min-Max normalization processing in S1.2 is as follows: In the formula, represents the Standardized characteristics of the original data corresponding to each student type; Representing the Each student corresponds to the original data; Representing the The minimum value of the original data of each student type; Representing the The maximum value of the original data of each student type; The visual tags mentioned in S1.6 specifically include language foundation, semantic understanding, cultural transfer readiness, and emotional / behavioral stability; As a further aspect of the present invention, the specific calculation formula for the linear weighted summation described in S1.2 is as follows: In the formula, Representing the Weighted ability scores for each ability dimension; Used to constitute the Total data volume for each capability dimension; Representing the The first of the 1st ability dimension The weights of each standardized feature; Representing the One standardized feature; The specific formula for calculating the probability of correct understanding as described in S1.3 is as follows: In the formula, On behalf of the students, regarding the first Predict the probability of "understanding correctly" in the question; The corresponding ability score is projected onto the [number]th [level]. Student ability scores based on a competency scale that matches the questions; Representing the The difficulty parameter of the question; Representing the The criteria for the problem.

[0009] As a further aspect of the present invention, the specific steps of the learning path generation module in automatically planning a personalized ancient poetry learning path are as follows: S2.1: Receive and extract learning features from each student's personalized learning profile, and map each extracted learning feature to a set of limited candidate learning units. Each candidate unit in the candidate learning unit set represents a learning task. At the same time, set an initial cost and constraint penalty for each candidate unit. Then, calculate the scalar cost of each candidate unit based on a combination of various quantitative indicators, and calculate the overall cost of each candidate unit based on each scalar cost. S2.2: Each candidate unit is represented in the form of a binary variable "whether to select the unit", based on the selected candidate unit, a corresponding diagonal Hamiltonian is constructed according to the overall cost, and a corresponding penalty term is added to the Hamiltonian according to the hard constraint of the upper limit of the total learning time of the candidate unit, and the numerical matrix of the energy operator is established according to the established Hamiltonian, then a set of parameterized quantum circuits is selected, and the circuit state is initialized based on the initial parameter vector; S2.3: Based on the current energy operator numerical matrix, the probability distribution corresponding to different binary variable solutions is calculated by selecting the parameterized quantum circuit, and the quantum state of the parameterized quantum circuit is measured multiple times to calculate the expected value of the current corresponding Hamiltonian, and the expected energy and measurement samples based on the current parameter configuration are output; S2.4: According to the output expected energy and measurement samples, the parameterized quantum circuit parameters are updated by the gradient-free black box method, and parallel multiple starts are used to explore the parameter space and solution space at the same time, and in each round of parameter iteration, the corresponding candidate solution with an expected energy lower than a preset threshold is marked as a low-energy candidate solution, and a low-energy candidate solution pool is recorded for each iteration sampling, until the expected energy value converges to a preset range after multiple iterations, the iteration is stopped; S2.5: A set of low-energy bit string samples are obtained from the quantum state of the optimized parameterized quantum circuit, wherein each bit string corresponds to a set of selected learning units, and forms a combined representation of a candidate learning path, and the low-energy bit string samples are scored for explainability, then the comprehensive score of each candidate path is calculated, and the candidate path list sorted from high to low and the corresponding bottleneck prompt are output; S2.6: Select multiple groups of high-score candidate paths ranked at the top from the candidate path list, convert each candidate path to a sequential learning unit sequence through discretization processing, determine the prerequisite order and parallel learning possibility between units, establish corresponding integer timetable constraints according to the time window, the upper limit of daily learning time and student preferences, schedule the learning unit sequence based on the club to generate a specific daily schedule, and output the final executable path and the corresponding timetable, and mark the time points with high cognitive load in the path.

[0010] As a further scheme of the application, the combination of the quantitative indicators in S2.1 specifically includes difficulty mismatch inadaptability, cognitive load risk, cultural migration cost, and expected learning gain. The explainability score in S2.5 specifically includes coverage, coherence, prerequisite consistency and prediction risk.

[0011] As a further scheme of the application, the specific calculation formula of the overall cost in S2.1 is as follows: wherein, represents the overall cost of the candidate unit ; represents the difficulty mismatch index; represents the cognitive load risk index; represents the cultural transfer cost; represents the expected learning gain; , , and respectively represent the weight of each type of index; The specific form of the energy operator numerical matrix in S2.2 is as follows: wherein, represents a binary variable , i.e. the Hamiltonian; represents a binary decision variable, if represents selecting the candidate unit , otherwise not selecting; represents a single variable linear coefficient, i.e. the scalar after mapping of the overall cost of the candidate unit ; represents a two-variable interaction coefficient; represents the penalty weight of the th hard constraint; represents the th constraint function.

[0012] As a further scheme of the present application, the specific steps of the physiological emotion monitoring module judging the learning emotional state of the student are as follows: P1.1: Real-time parallel acquisition of heart rate interval, skin conductance and eye movement data of students from wearable devices, time alignment of the three data streams using a unified clock, then division of each type of time series data into multiple groups of sliding time windows according to the preset fixed window length, and checking of data integrity and sampling quality in each window, if the quality of the sliding window is lower than the preset threshold, marking it as a low-quality window; P1.2: Generating corresponding aggregate statistics for each sliding window according to the collected various types of time series data, calculating the multi-modal features of each sliding window and each type of signal, extracting event-type features in each type of feature, and calculating the occurrence rate of each event-type feature, then normalizing the multi-modal features in each sliding window by the Z-score method, and recording the normalized features and the corresponding quality scores of each window; P1.3: According to the student HSK level, set the basic threshold for each type of activated physiological indicators by fitting the expert prior or historical data, then scale the basic threshold according to the student HSK level, and individualize the threshold according to the emotional sensitivity coefficient in the individualized portrait of different students, and finally generate a set of dynamic threshold values available for each sliding window; P1.4: Input the multi-modal features in each sliding window and the dynamic threshold set into the emotion inference model, and the emotion inference model outputs the emotion state probability through the forward propagation algorithm, and then temperature scaling and confidence calibration are performed on the output probability, and the final emotion probability distribution and confidence score are generated based on the quality weight of the corresponding sliding window; P1.5: Time series smoothing is performed on the emotion probability distribution of the continuous sliding window to reduce the emotional flicker caused by instantaneous noise, and if any negative emotion exceeds the given trigger probability threshold in the smoothed probability for consecutive multiple sliding windows and has high confidence, it is marked as "emotional abnormality, intervention needed", the trigger evidence is recorded, and the emotion probability curve in the sliding time period, the trigger event time point, the trigger reason, and the event confidence and explainable points are output.

[0013] As a further scheme of the present application, P1.2 specifies that the specific features of each sliding window and each type of signal include short-term heart rate variability, pulse number per unit time, pulse amplitude, average gaze duration, gaze switching rate, and mean pupil diameter. The event type features described in P1.2 include the number of short-term heart rate surges, short-term skin electricity surge peaks, and sudden gaze interruption events.

[0014] As a further scheme of the present application, the specific steps of the cognitive load analysis module for comprehensively judging the learning pressure of each student are as follows: S3.1: Collect the response time, gaze duration, and physiological signals of each student for each question or each interaction, and record the collection time of each type of data, then synchronize each type of data to a common time step through a unified clock interpolation, and then filter and denoise each type of data, and then aggregate and summarize the processed data of each type to generate synchronized windowed features; S3.2: Calculate the statistical features and dynamic features of each type of data respectively, standardize each type of feature to the [0, 1] interval, and record the quality weight of each feature in the corresponding time window, then map each type of feature in each time window to the corresponding explainable cognitive load sub-indicator, and after mapping, normalize each cognitive load sub-indicator; S3.3: Based on the preset time interval, the cognitive load sub-indices of different time periods are weighted and fused to obtain the student's instant overall cognitive load score, and the exponential weighted moving average retains the instantaneous impact and historical cumulative fatigue, then in the fusion, the load change rate and the load acceleration are calculated in parallel, and according to the calculation results, the instantaneous fluctuation and the rising load trend are identified; S3.4: According to the student's instant overall cognitive load score and the historical load change trend, when the instant overall cognitive load score or the historical load change rises, exceeds the preset threshold, it is determined that the current student's cognitive load is too high, and the corresponding risk category is output.

[0015] As a further scheme of the present application, the specific method of filtering and denoising in S3.1 is as follows: a band-pass filter is used for physiological signals to remove direct current drift and high-frequency noise; a median filter is used for gaze duration to remove instantaneous abnormalities such as blinking; and after detecting the extreme outliers in the response time of each question or each interaction, the response time is truncated or marked as abnormal. The statistical characteristics and dynamic characteristics in S3.2 specifically include: based on the response time of each question or each interaction, the mean, variance, shortest and longest response, and the slope of the reaction time are obtained; based on the gaze duration, the average gaze duration, the number of gaze segments and the longest continuous gaze are obtained; based on the physiological signals, the heart rate variability approximation index, the skin electricity peak rate and the respiratory cycle stability index are obtained. The cognitive load sub-indices in S3.2 specifically include processing load, perception load, emotional / physiological stress load, attention dispersion, etc.

[0016] As a further scheme of the present application, the specific steps of the artistic conception analogy generation module constructing a cultural knowledge graph and reasoning the cultural artistic conception behind the poem are as follows: S4.1: Identify the noun image, action, rhetoric or allusion tag and emotion tag in the ancient poetry, and generate a preliminary semantic representation for each identified element, and then map each semantic representation to a symbol unit, wherein the symbol unit includes an identifier, a type and a plurality of attributes, and each symbol unit generated is used as an initial candidate of a graph node, while a traceable mapping from the symbol to the original text segment is generated; S4.2: Calculate the confidence value of each symbol unit, and add the symbol units with confidence values exceeding a preset threshold to the graph node set, and then according to the explicit dependency relationship within the ancient poetry text, the semantic similarity and the external cultural source retrieval, establish candidate relationship edges for each symbol unit, and use text co-occurrence frequency, semantic embedding similarity and allusion matching degree as evidence to calculate edge weights, and establish the corresponding cultural knowledge graph based on the form of "node-edge-node"; S4.3: Induce candidate symbolic rules from the frequent sub-graphs, path patterns and edge type sequences in the existing cultural knowledge graph, then map the embedding vectors of the antecedent and consequent of each candidate symbolic rule, and calculate the compatibility score between the vectors, if the symbolic frequency and embedding compatibility score are higher than the preset threshold, then the rule is promoted to a high credible rule as a logical template that can be used for reasoning, otherwise, it is screened out, then a set of symbolic rules with confidence is generated, and the rules are added to the current cultural knowledge graph in a retrievable form; S4.4: Perform interpretable chain reasoning on the current cultural knowledge graph through the symbolic rule set to generate candidate mood conclusions, then use a neural inference based on a graph neural network to perform embedding-level inference on the whole cultural knowledge graph to generate a soft mood distribution, then align the neural output with the candidate mood conclusions, and calculate the consistency between them, if the consistency is greater than 80%, update the confidence of the corresponding mood conclusion; S4.5: Generate corresponding symbolic evidence chains, embedding support and combined confidence for each cultural mood in ancient poems, retrieve mother tongue cultural concepts with a similarity higher than a preset threshold to the confirmed mood candidate semantics in the cultural migration library, and generate multiple groups of candidate analogy pairs, calculate the mapping strength of each candidate analogy pair, then sort the mappings from high to low according to the strength, and select different depth migration strategies according to the current learning goal of the student, generate corresponding poetic mood explanations, analogy examples, two-sided evidence chains, and migration difficulty annotations according to the HSK level of the learner.

[0017] Compared with the prior art, the present application has the following advantages: The application firstly collects the basic data of the learner's HSK level, vocabulary, ancient poetry diagnosis test results and mother tongue cultural background, and generates standardized features, according to the weight formulated by experts, each feature is divided into the corresponding ability dimension, and the preliminary ability score is obtained through linear weighting, then the probability model is established by using the diagnostic questions of multi-difficulty and multi-image type, the difficulty and discrimination of the questions are inferred, so as to further calibrate each ability dimension, the cultural background of the learner is coded as a cultural semantic vector, and the similarity is calculated with the image vector of the poem, so as to judge the mood migration difficulty; at the same time, the regression model is constructed by combining the behavior signals such as answering speed and error rate, and the real-time risk score is generated to identify the potential learning bottleneck, the signals such as ability score, question statistics, cultural similarity and risk score are integrated according to the fusion weight, and the multi-dimensional learning portrait with confidence interval is formed, the portrait is screened from the learning unit pool, and the cost and penalty term are defined for each task, then the Hamiltonian form is input into the quantum optimization model, the parameterized quantum circuit is used to search for the lowest energy solution, so as to obtain the personalized and efficient learning path, after the explainability analysis, the path is discretized into a learning task sequence, and the final schedule is generated by combining the learning time and time arrangement, in the learning process, the physiological data such as heart rate, skin electricity and eye movement are collected in real time, and time window processing, feature extraction and dynamic threshold correction are carried out, the learning emotional state is inferred by the emotion model and the continuous negative fluctuation is detected, the cognitive load is estimated by the reaction time and gaze duration, the load surge is identified and the risk category is output, finally, the ancient poetry culture knowledge graph is constructed, the image, rhetoric and allusion symbol structure are extracted, the cultural mood interpretation is generated through neural symbol reasoning, and the analogy concept of the student's mother tongue culture is matched by combining the cross-cultural migration library, the cross-cultural poem learning content conforming to the HSK grading and having explainability is generated, so that the learning task can be updated in real time with the change of the student's ability, the "one-size-fits-all" teaching method is avoided, the learning efficiency and adaptability are improved, the learning experience and continuity are greatly improved, the cultural gap can be effectively reduced, the non-native Chinese speakers can more easily understand the deep meaning of the ancient poetry, the memory depth of the learners on the mood and vocabulary of the poem is greatly improved, and the learning motivation is improved. BRIEF DESCRIPTION OF DRAWINGS

[0018] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, together with the embodiments of the application, to explain the application, and do not constitute a limitation on the application.

[0019] Figure 1 The system block diagram of the ancient poetry multi-modal teaching system based on HSK grading proposed in the application. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments of the present application.

[0021] Embodiment 1 With reference to Figure 1 , the HSK level-based ancient Chinese poetry multi-modal teaching system comprises a learning evaluation portrait module, a learning path generation module, a physiological emotion monitoring module, a cognitive load analysis module, a vocabulary and grammar strengthening module, a modal semantic analysis module, a artistic conception analogy generation module, a holographic scene generation module, a feedback interaction experience module, a content dynamic adjustment module, a recitation pronunciation correction module, an extended knowledge generation module, a performance analysis feedback module and a review and memory strengthening module.

[0022] The learning evaluation portrait module is used to generate individualized learning portraits of students according to the HSK level, vocabulary, ancient Chinese poetry understanding ability and mother tongue cultural background of the students.

[0023] Specifically, the HSK level, vocabulary estimate, ancient poetry diagnostic test score, and mother tongue culture label distribution of each student are collected as raw data. Based on the 3σ principle, data exceeding µ±3σ is marked as an outlier. Range detection is performed on each type of raw data, and the maximum and minimum values of each type of raw data are marked. Min-Max normalization is performed on each type of raw data to obtain standardized features. Based on expert experience, initial weights are assigned to the language foundation, vocabulary mastery, poetry semantic understanding, and cultural sensitivity of each student. The standardized features are then grouped according to the specified initial weights, and the standardized features within the same dimension are linearly weighted and summed to obtain the ability score of that dimension. A set of diagnostic questions for ancient poetry understanding with different difficulties and different image types is selected, and the student's answer results and answer time for each question are recorded. A probability model based on the student's ability score is established for each question to evaluate the student's understanding probability. The question difficulty parameters and discrimination are inferred from the question answer probability. The question probability results are then summarized and fed back to the student's ability dimension. The cultural background questionnaire and cultural exposure of each student are encoded as cultural semantic vectors. The cultural images of the target ancient poetry are mapped to image vectors in the same space through expert annotation or semantic embedding. The cosine similarity between each cultural semantic vector and each image vector is calculated. If the cosine similarity is higher than the preset threshold, it means that the poetic image can be easily found in the student's mother tongue culture, otherwise, it indicates that cross-cultural interpretation is needed. The cosine similarity results are used as the inverse vector of the student's cultural transfer difficulty. The average reaction time and unit question error rate of each student are calculated based on the answer speed and accuracy on the diagnostic question set. Linear regression prediction is used to map the two groups of behavior indicators to an instant risk score. According to the preset threshold interval, the risk score is labeled as high, medium, and low, and the corresponding confidence interval is output. The ability score, question level understanding probability summary statistics, cultural similarity, and risk score of each student are collected as portrait signals. Based on historical calibration results, each type of portrait signal is assigned a fusion weight, and linear fusion is performed according to the weight to obtain the comprehensive score of each student's multi-dimensional portrait dimension. The corresponding visual labels are generated according to the student's portrait dimension, and each label is accompanied by a confidence interval and the weight proportion of the constituent elements.

[0024] It should be further explained that each type of raw data is collected through the student's registration information, vocabulary test, understanding questions for several poems, and cultural background questionnaire. The visual labels include language foundation, semantic understanding, cultural transfer readiness, and emotional / behavioral stability. The specific calculation formula of Min-Max normalization is as follows: wherein, represent the standardized features of the raw data corresponding to the th student; represent the raw data corresponding to the th student; represent the raw data corresponding to the th student; represent the minimum value of the raw data corresponding to the th student; The specific calculation formula of the linear weighted sum is as follows: wherein, represent the weighted ability score of the th ability dimension; represent the total amount of data used to constitute the th ability dimension; represent the weight of the th standardized feature in the th ability dimension; represent the th standardized feature; The specific calculation formula of the probability of correct understanding is as follows: wherein, represent the predicted probability of the student's correct understanding of the th question; represent the student's ability score after the corresponding ability score is projected to the ability scale matched with the th question; represent the difficulty parameter of the th question; represent the discrimination parameter scale of the th question.

[0025] The learning path generation module processes the individualized learning portraits of each student, predicts potential learning bottlenecks, and automatically plans individualized learning paths for classical poetry.

[0026] Specifically, the learning features in the personalized learning portrait of each student are received and extracted, and the extracted learning features are mapped to a set of candidate learning units, wherein each candidate unit in the candidate learning unit set represents a learning task, and an initial cost and a constraint penalty term are set for each candidate unit. Then, the scalar cost of each candidate unit is calculated based on a combination of various quantitative indicators, the overall cost of each candidate unit is calculated based on the scalar cost, and each candidate unit is represented in the form of a binary variable to indicate whether the unit is selected. Based on the selected candidate units, a corresponding diagonal Hamiltonian is constructed according to the overall cost, and a corresponding penalty term is added to the Hamiltonian according to the hard constraint of the upper limit of the total learning time of the candidate units. Meanwhile, an energy operator numerical matrix is established based on the established Hamiltonian. Then, a set of parameterized quantum circuits is selected, and the circuit state is initialized based on an initial parameter vector. Based on the current energy operator numerical matrix, the probability distribution of the solution corresponding to different binary variables is calculated by selecting the parameterized quantum circuit. The quantum state of the parameterized quantum circuit is measured multiple times to calculate the expected value of the current corresponding Hamiltonian, and the expected energy and measurement samples under the current parameter configuration are output. According to the output expected energy and measurement samples, the parameterized quantum circuit parameters are updated by the gradient-free black box method, and parallel multiple starts are used to explore the parameter space and solution space simultaneously. In each round of parameter iteration, the candidate solution with an expected energy value lower than a preset threshold is marked as a low-energy candidate solution, and a pool of low-energy candidate solutions obtained by sampling in each iteration is recorded. When the change value of the expected energy value converges to a preset range in multiple iterations, the iteration is stopped. A set of low-energy bit string samples is obtained from the quantum state of the optimized parameterized quantum circuit, wherein each bit string corresponds to a set of selected learning units and forms a combined representation of a candidate learning path. The low-energy bit string samples are scored for explainability. Then, the comprehensive score of each candidate path is calculated, and a candidate path list sorted from high to low and corresponding bottleneck prompts are output. From the candidate path list, multiple groups of high-score candidate paths ranked at the top are selected, and each candidate path is converted to a sequential learning unit sequence through discretization processing. The prerequisite order and parallel learning possibility between units are determined, and the corresponding integer timetable constraints are established according to the time window, the upper limit of the daily learning time, and the student's preference. Based on the club, the learning unit sequence is scheduled to generate a specific daily schedule, and the final executable path and the corresponding schedule are output, with the time points with high cognitive load marked in the path.

[0027] In addition, in the present embodiment, the combination of quantitative indicators specifically includes difficulty mismatch, cognitive load risk, cultural transfer cost, and expected learning gain; and the explainability score specifically includes coverage, coherence, prerequisite consistency, and prediction risk. The specific calculation formula of the overall cost is as follows: wherein, represents the overall cost of the candidate unit ; represents the difficulty mismatch index; represents the cognitive load risk index; represents the cultural transfer cost; represents the expected learning gain; , , and respectively represent the weights of various indexes; The specific form of the energy operator numerical matrix is as follows: wherein, represents a binary variable , i.e. the Hamiltonian; represents a binary decision variable, if , the candidate unit is selected, otherwise it is not selected; represents a single variable linear coefficient, i.e. the scalar after the overall cost mapping of the candidate unit ; represents a two-variable interaction coefficient; represents the penalty weight of the th hard constraint; represents the th constraint function.

[0028] The physiological emotion monitoring module collects physiological data of the learner in real time through a wearable device, and judges the learning emotional state of the student.

[0029] In addition, as a further description of the invention, the student's heart rate interval, skin conductance and eye movement data are collected in real time from the wearable device, the three data streams are time-aligned using a unified clock, and then each type of time series data is divided into multiple groups of sliding time windows according to the preset fixed window length, and the data integrity and sampling quality are checked in each window. If the quality of the sliding window is lower than the preset threshold, it is marked as a low-quality window. According to the collected various types of time series data, an aggregated statistical value is generated for each sliding window, the multi-modal features of each sliding window and each type of signal are calculated, and the event-type features in each type of feature are extracted, and the occurrence rate of each event-type feature is counted. Then, the multi-modal features in each sliding window are normalized by the Z-score method, and the normalized features and the corresponding quality scores of each window are recorded. According to the student's HSK level, the expert prior or historical data fitting is used to set the basic threshold for each type of activated physiological indicator, and then the basic threshold is scaled according to the student's HSK level, and the threshold is individually corrected according to the emotional sensitivity coefficient in the individual portrait of different students. Finally, a set of dynamic threshold values available for each sliding window is generated, and the multi-modal features in each sliding window and the dynamic threshold set are used as input data and transmitted to the emotion inference model. The emotion inference model outputs the emotion state probability through the forward propagation algorithm, and then the output probability is temperature scaled and confidence calibrated. Based on the quality weight of the corresponding sliding window, the final emotion probability distribution and confidence score are generated. The emotion probability distribution of the continuous sliding window is time series smoothed to reduce the emotional flicker caused by instantaneous noise. If any negative emotion exceeds the given trigger probability threshold and has high confidence in the smoothed probability for consecutive multiple sliding windows, it is marked as "emotion abnormality, intervention needed", and the trigger evidence is recorded. The emotion probability curve, trigger event time point, trigger reason, event confidence and explainable points in the sliding time period are output.

[0030] It should be noted that the specific multi-features of each sliding window and each type of signal include short-term heart rate variability, pulse number per unit time, pulse amplitude, average fixation duration, fixation switching rate, and mean pupil diameter; event-type features include short-term heart rate surge frequency, short-term skin conductance peak, sudden fixation interruption event, etc.

[0031] Embodiment 2: Referring to Figure 1 , the HSK classification-based multi-modal teaching system of classical Chinese poetry includes a learning evaluation portrait module, a learning path generation module, a physiological emotion monitoring module, a cognitive load analysis module, a vocabulary and grammar strengthening module, a modal semantic analysis module, a mood analogy generation module, a holographic scene generation module, a feedback interaction experience module, a content dynamic adjustment module, a recitation pronunciation correction module, an extended knowledge generation module, a performance analysis feedback module, and a review and memory strengthening module.

[0032] The cognitive load analysis module is used to simulate the cognitive load of students and comprehensively judge the learning pressure of each student.

[0033] Specifically, the response time, gaze duration, and physiological signals of each student for each question or each interaction are collected, and the collection time of various types of data is recorded. Then, the various types of data are synchronized to a common time step through a unified clock interpolation, and then the various types of data are filtered and denoised. After that, the processed various types of data are aggregated and summarized to generate synchronized windowed features. The statistical features and dynamic features of each type of data are calculated, each type of feature is standardized to the [0, 1] interval, and the quality weight of each feature in the corresponding time window is recorded. Then, each type of feature in each time window is mapped to the corresponding interpretable cognitive load sub-index. After the mapping is completed, each cognitive load sub-index is normalized. Based on the preset time interval, the cognitive load sub-indices of different time periods are weighted and fused to obtain the student's instantaneous overall cognitive load score. The exponential weighted moving average retains the instantaneous impact and historical cumulative fatigue. Then, during the fusion, the load change rate and load acceleration are calculated in parallel. According to the calculation results, the instantaneous fluctuation and rising load trend are identified. According to the student's instantaneous overall cognitive load score and the historical load change trend, if the instantaneous overall cognitive load score or the historical load change rises above the preset threshold, it is determined that the current student's cognitive load is too high, and the corresponding risk category is output.

[0034] In this embodiment, the specific filtering and denoising method is as follows: for physiological signals, a band-pass filter is used to remove direct current drift and high-frequency noise; for gaze duration, median filtering is used to remove instantaneous abnormalities such as blinking; for the response time of each question or each interaction, extreme outliers are detected and truncated or marked as abnormal; statistical features and dynamic features include: based on the response time of each question or each interaction, the mean, variance, shortest and longest response, and the slope of the reaction time are obtained; based on the gaze duration, the average gaze duration, the number of gaze segments, and the longest continuous gaze are obtained; based on the physiological signals, the heart rate variability approximation index, the skin electricity peak rate, and the respiratory cycle stability index are obtained; the cognitive load sub-indexes include processing load, perception load, emotional / physiological stress load, attention dispersion, etc.

[0035] The vocabulary grammar reinforcement module automatically filters key words in the poem according to the HSK level, and assists students in reading and following reading training through voice recognition; the modal semantic analysis module is used to decompose the semantic of the ancient poem, and generate content explanation for easy understanding; the artistic conception analogy generation module is used to build a cultural knowledge graph, and infer the cultural artistic conception behind the poem based on the cultural knowledge graph, and at the same time, migrate it to the student's mother tongue cultural system.

[0036] Specifically, the noun image, action, rhetoric or allusion tag and emotion tag in the ancient poetry are identified, and a preliminary semantic representation is generated for each identified element. Each semantic representation is mapped to a symbol unit, which includes an identifier, a type, and multiple sets of attributes. The generated symbol units are used as initial candidates for graph nodes, and a traceable mapping from symbols to original text fragments is generated. The confidence values of each symbol unit are calculated, and the symbol units with confidence values exceeding a preset threshold are added to the graph node set. Then, based on the explicit dependency relationship within the ancient poetry text, semantic similarity, and external cultural source retrieval, candidate relationship edges are established for each symbol unit. The edge weights are calculated using various evidence such as text co-occurrence frequency, semantic embedding similarity, and allusion matching degree. The corresponding cultural knowledge graph is established based on the "node-edge-node" form. From the existing cultural knowledge graph, frequent subgraphs, path patterns, and edge type sequences are induced to generate candidate symbol rules. The embedding vectors of the antecedent and consequent of each candidate symbol rule are mapped, and the compatibility score between the vectors is calculated. If the symbol frequency and embedding compatibility score are higher than the preset threshold, the rule is promoted to a high-confidence rule, which is used as a logical template for reasoning. Otherwise, the rule is excluded. A set of symbol rules with confidence is generated, and the rules are added to the current cultural knowledge graph in a retrievable form. The current cultural knowledge graph is subjected to interpretable chain reasoning using the symbol rule set, producing candidate artistic conception conclusions. The entire cultural knowledge graph is subjected to embedding-level inference using a graph neural network-based neural reasoner, producing a soft artistic conception distribution. The neural output is aligned with the candidate artistic conception conclusions, and the consistency between them is calculated. If the consistency is greater than 80%, the confidence of the corresponding artistic conception conclusion is updated. Symbol evidence chains, embedding support, and combined confidence are generated for each cultural artistic conception in the ancient poetry. The mother tongue cultural concepts with a semantic similarity higher than a preset threshold to the confirmed artistic conception candidates are retrieved from the cultural migration library, and multiple candidate analogies are generated. The mapping strength of each candidate analogy is calculated, and the mappings are sorted from high to low based on the strength. Different depth migration strategies are selected based on the student's current learning goals, and corresponding poetic artistic conception explanations, analogy examples, two-sided evidence chains, and migration difficulty annotations based on the learner's HSK level are generated.

[0037] The holographic scene generation module constructs a three-dimensional scene matching the content of the poem through holographic projection according to the student's HSK level. The feedback interaction experience module dynamically controls the student's interaction density and task complexity based on different HSK levels in combination with the holographic scene.

[0038] The content dynamic adjustment module dynamically adjusts the teaching strategy according to the real-time monitoring of the emotional fluctuations of the students; the recitation pronunciation correction module is used for providing standard recitation of poems, voice rhythm analysis, and rhyme perception learning for the students, and automatically correcting the tone and pause problems in the recitation of the students; the knowledge expansion generation module provides cultural expansion content corresponding to the artistic conception of the poems after the students understand the core content of the poems.

[0039] The performance analysis feedback module is used for automatically recording the learning data of the students, and generating a visual report for the students to understand their progress and weak knowledge points; the review and memory reinforcement module draws the forgetting curve of each student, and automatically arranges the review time and content.

Claims

1. A multimodal teaching system for classical Chinese poetry based on HSK level assessment, characterized in that: It includes modules for learning assessment profiling, learning path generation, physiological and emotional monitoring, cognitive load analysis, vocabulary and grammar reinforcement, modal semantic parsing, contextual analogy generation, holographic scene generation, feedback and interactive experience, dynamic content adjustment, recitation pronunciation correction, extended knowledge generation, performance analysis and feedback, and review and memory reinforcement. The learning assessment profile module is used to generate personalized learning profiles for each student based on their HSK level, vocabulary, understanding of classical Chinese poetry, and native language and cultural background. The learning path generation module processes each student's personalized learning profile, predicts potential learning bottlenecks, and automatically plans a personalized learning path for classical Chinese poetry. The physiological emotion monitoring module collects learners' physiological data in real time through wearable devices and judges students' learning emotional state. The cognitive load analysis module is used to simulate the cognitive load of students and comprehensively judge the learning pressure of each student. The vocabulary and grammar enhancement module automatically filters keywords in poems according to HSK levels and uses speech recognition to assist students in reading aloud and repeating training. The modal semantic parsing module is used to perform semantic decomposition on ancient poems and generate content explanations that are easy to understand. The imagery analogy generation module is used to construct a cultural knowledge graph, deduce the cultural imagery behind poems, and transfer it to the students' native language cultural system. The holographic scene generation module constructs a three-dimensional scene that matches the content of the poem through holographic projection based on the student's HSK level. The feedback and interaction experience module combines a holographic scene and dynamically controls the student interaction density and task complexity according to different HSK levels. The content dynamic adjustment module dynamically adjusts teaching strategies based on real-time monitoring of students' emotional fluctuations; The recitation pronunciation correction module is used to provide students with standard recitation of poems, speech rhythm analysis, rhyme perception learning, and automatically correct the tone and pause problems in the students' recitation; The extended knowledge generation module provides cultural extension content corresponding to the artistic conception of the poems after students understand the core content of the poems. The performance analysis and feedback module is used to automatically record students' learning data and generate visual reports so that students can understand their progress and weak knowledge points. The review and memory enhancement module plots the forgetting curve for each student and automatically schedules review time and content.

2. The multimodal teaching system for classical Chinese poetry based on HSK leveling as described in claim 1, characterized in that, The specific steps for the learning assessment profile module to generate personalized learning profiles for each student are as follows: S1.1: Collect various raw data from students, including HSK level, vocabulary estimation, ancient poetry diagnostic test scores, and native language cultural label distribution. Based on the 3σ principle, mark data exceeding µ±3σ in each type of raw data as outliers. Then, perform range detection on each type of raw data and mark the maximum and minimum values ​​of each type of raw data. S1.2: Perform Min-Max normalization on each type of raw data to obtain standardized features of each type of data. Based on expert experience, assign initial weights to each ability dimension of students’ language foundation, vocabulary mastery, poetry semantic understanding and cultural sensitivity. Then, group each standardized feature according to the assigned initial weights into ability dimensions, and perform linear weighted summation on the standardized features within the same dimension to obtain the ability score for that dimension. S1.3: Select a set of diagnostic questions for understanding ancient poems and lyrics, which contain different difficulties and different image types, and record the students' answers and answering time for each question. At the same time, establish a probability model based on the students' ability scores for each question to evaluate the probability of students understanding each question correctly. Based on the probability of answering the question correctly, infer the question difficulty parameters and discrimination. Then, summarize the question probability results and feed them back to the students' ability dimension. S1.4: Encode students' cultural background questionnaires and cultural contact records into cultural semantic vectors. At the same time, through expert annotation or semantic embedding, map the cultural imagery of the target ancient poems into image vectors in the same space. Then calculate the cosine similarity between each cultural semantic vector and each image vector. If the cosine similarity is higher than the preset threshold, it means that the imagery of the poem can be easily found in the student's native culture. Otherwise, it indicates that cross-cultural interpretation is needed. At the same time, the cosine similarity results are used as the inverse vector of the difficulty of students' cultural transfer. S1.5: Based on the student's answering speed and accuracy on the diagnostic question set, calculate the student's average reaction time and unit question error rate as behavioral indicators. Map the two sets of behavioral indicators to an instant risk score through linear regression prediction. According to the preset threshold range, label the risk score as high, medium and low levels and output the corresponding confidence interval. S1.6: Collect various profile signals such as students' ability scores, question-level comprehension probability summary statistics, cultural similarity, and risk scores. Based on historical calibration results, assign fusion weights to each type of profile signal and perform linear fusion according to the weights to obtain a comprehensive score for each student's multi-dimensional profile dimensions. At the same time, generate corresponding visual labels according to the student profile dimensions. Each label is accompanied by a confidence interval and the weight ratio of the constituent elements.

3. The HSK-based multimodal teaching system for classical Chinese poetry according to claim 2, characterized in that, The specific calculation formula for the linear weighted summation described in S1.2 is as follows: In the formula, Representing the Weighted ability scores for each ability dimension; Used to constitute the Total data volume for each capability dimension; Representing the The first of the 1st ability dimension The weights of each standardized feature; Representing the One standardized feature; The specific formula for calculating the probability of correct understanding as described in S1.3 is as follows: In the formula, On behalf of the students, regarding the first Predict the probability of "understanding correctly" in the question; The corresponding ability score is projected onto the [number]th [level]. Student ability scores based on a competency scale that matches the questions; Representing the The difficulty parameter of the question; Representing the The criteria for the problem.

4. The HSK-based multimodal teaching system for classical Chinese poetry according to claim 2, characterized in that, The specific steps of the learning path generation module in automatically planning a personalized learning path for classical Chinese poetry are as follows: S2.1: Receive and extract learning features from each student's personalized learning profile, and map each extracted learning feature to a set of limited candidate learning units. Each candidate unit in the candidate learning unit set represents a learning task. At the same time, set an initial cost and constraint penalty for each candidate unit. Then, calculate the scalar cost of each candidate unit based on a combination of various quantitative indicators, and calculate the overall cost of each candidate unit based on each scalar cost. S2.2: Each candidate unit is represented by a binary variable to indicate whether the unit is selected. Based on the selected candidate unit, a corresponding diagonal Hamiltonian is constructed according to its total cost. According to the hard constraint of the upper limit of the total learning time of the candidate unit, the corresponding penalty term is added to the Hamiltonian. At the same time, an energy operator numerical matrix is ​​established based on the established Hamiltonian. Then, a set of parameterized quantum circuits is selected, and the circuit state is initialized based on the initial parameter vector. S2.3: Based on the current energy operator numerical matrix, the probability distribution of the measurement distribution under the selected parameterized quantum circuit is calculated to correspond to the solution of different binary variables. Multiple measurements are performed based on the quantum state of the parameterized quantum circuit to calculate the expected value of the current Hamiltonian and output the expected energy and measurement sample based on the current parameter configuration. S2.4: Based on the expected output energy and the measurement samples, update the parameterized quantum circuit parameters using a gradient-free black-box method. Simultaneously, adopt parallel multi-startup to explore both the parameter space and the solution space. In each round of parameter iteration, the corresponding candidate solutions with expected energy below a preset threshold are marked as low-energy candidate solutions. Record the pool of low-energy candidate solutions obtained from each iteration until the change in expected energy value over multiple iterations converges to a preset range, at which point the iteration stops. S2.5: Obtain a set of low-energy bit string samples from the quantum states of the optimized parameterized quantum circuit, where each bit string corresponds to a set of selected learning units and forms a combined representation of a candidate learning path. Score the interpretability of the low-energy bit string samples, then calculate the comprehensive score of each candidate path, and output the candidate path list sorted from high to low and the corresponding bottleneck prompts. S2.6: Select multiple high-scoring candidate paths from the candidate path list, and convert each candidate path into a sequential learning unit sequence through discretization. Determine the prerequisite order and parallel learning possibility between units. Based on the time window, the daily learning time limit and student preferences, establish corresponding integer timetable constraints. Based on this, schedule the learning unit sequence to generate a specific schedule. Then output the final executable path and the corresponding timetable, and mark the time points with high cognitive load in the path.

5. The HSK-based multimodal teaching system for classical Chinese poetry according to claim 4, characterized in that, The specific formula for calculating the overall cost mentioned in S2.1 is as follows: In the formula, Representative candidate unit The overall cost; This indicates a mismatch between the difficulty level and the indicator. Represents an indicator of cognitive load risk; Represents the cost of cultural migration; Represents the expected learning gain; , , as well as These represent the weights of various indicators; The specific representation of the energy operator numerical matrix described in S2.2 is as follows: In the formula, Represents binary variables That is, the Hamiltonian; Represents binary decision variables, if Indicates the selection of candidate units Otherwise, do not select; Represents univariate linear coefficients, i.e., candidate units The scalar after mapping the total cost; Represents the interaction coefficient between the two variables; Representing the The penalty weight for a hard constraint; Representing the Constraint functions.

6. The multimodal teaching system for classical Chinese poetry based on HSK leveling as described in claim 1, characterized in that, The specific steps taken by the cognitive load analysis module to comprehensively determine the learning pressure of each student are as follows: S3.1: Collect the response time, gaze duration, and physiological signals of each student for each question or interaction, and record the collection time of each type of data. Then, synchronize each type of data to a common time step through a unified clock interpolation, filter and denoise each type of data separately, and then aggregate and summarize the processed data to generate synchronized windowed features. S3.2: Calculate the statistical and dynamic characteristics of each type of data, standardize each type of feature to the [0, 1] interval, and record the quality weight of each feature in the corresponding time window. Then, map each type of feature in each time window to the corresponding interpretable cognitive load sub-indicator. After the mapping is completed, normalize each cognitive load sub-indicator. S3.3: Based on a preset time interval, the cognitive load sub-indicators of different time periods are weighted and fused to obtain the student's real-time overall cognitive load score. At the same time, the exponentially weighted moving average retains the instantaneous impact and historical cumulative fatigue. Then, during the fusion, the load change rate and load acceleration are calculated in parallel, and the instantaneous fluctuation and rising load trend are identified based on the calculation results. S3.4: Based on the student's real-time overall cognitive load score and historical load change trend, if the real-time overall cognitive load score or historical load change rises above the preset threshold, it is determined that the student's current cognitive load is too high, and the corresponding risk category is output.

7. The multimodal teaching system for classical Chinese poetry based on HSK leveling as described in claim 1, characterized in that, The specific steps of the imagery analogy generation module in constructing a cultural knowledge graph and inferring the cultural imagery behind poems are as follows: S4.1: Identify noun images, actions, rhetorical or allusion tags and emotional tags in classical Chinese poetry and prose, generate a preliminary semantic representation for each identified element, and then map each semantic representation to a symbol unit. The symbol unit contains an identifier, type and multiple sets of attributes. The generated symbol units serve as initial candidates for graph nodes, and at the same time, a traceable mapping from symbols to original text fragments is generated. S4.2: Calculate the confidence value of each symbol unit, and add the symbol units with confidence values ​​exceeding the preset threshold to the graph node set. Then, based on the explicit dependency relationship in the ancient poetry text, semantic similarity, and external cultural source retrieval, establish candidate relation edges for each symbol unit, and use various evidence such as text co-occurrence frequency, semantic embedding similarity, and allusion matching degree to calculate edge weights, and establish the corresponding cultural knowledge graph based on the form of "node-edge-node". S4.3: Summarize candidate symbol rules from frequent subgraphs, path patterns, and edge type sequences in the existing cultural knowledge graph. Then, map the embedding vectors of the antecedent and consequent of each candidate symbol rule and calculate the compatibility score between the vectors. If the symbol frequency and embedding compatibility score are higher than a preset threshold, the rule is promoted to a high-confidence rule and used as a logical template for reasoning. Otherwise, it is filtered out. Then, a set of symbol rules with confidence is generated and the rules are added to the current cultural knowledge graph in a searchable form. S4.4: Perform interpretable chain reasoning on the current cultural knowledge graph through the symbol rule set to generate candidate mood conclusions. Then, use a neural inference engine based on graph neural network to perform embedding-level inference on the entire cultural knowledge graph to generate soft mood distribution. After that, align the neural output with the candidate mood conclusions and calculate their consistency. If the consistency is >80%, update the confidence of the corresponding mood conclusion. S4.5: Generate corresponding symbolic evidence chains, embedding support, and combinatorial confidence for each cultural imagery in classical Chinese poetry. Retrieve native language cultural concepts from the cultural transfer database that have a semantic similarity to confirmed imagery candidates that exceeds a preset threshold, and generate multiple sets of candidate analogies. Simultaneously, calculate the mapping strength of each candidate analogy, sort the mappings from high to low strength, and select transfer strategies of different depths according to the student's current learning objectives. Based on the selected transfer strategies, generate corresponding poetic imagery descriptions, analogy examples, bilateral evidence chains, and transfer difficulty labels generated according to the learner's HSK level.

Images