Intelligent education feedback method and system based on virtual character interaction

Abstract

The invention discloses an intelligent education feedback method and system based on virtual character interaction, and particularly relates to the technical field of human-computer interaction. The method comprises the following steps: obtaining multi-modal feedback data generated by a plurality of virtual characters for the same learning task to perform semantic standstill dual analysis, and quantizing the difference between a feedback language and an expression style to obtain a conflict feature set; mapping the priorities of the feedback semantic vectors according to the conflict feature set in combination with the learning task target weight and the virtual role teaching function weight to form a feedback priority sequence; calling a cognitive load gradient prediction model based on the feedback priority sequence to calculate an information integration load value of the learner to generate a cognitive load evaluation result; and performing expression style reconstruction and content simplification on the feedback semantic vector according to a minimum cognitive load principle based on a cognitive load evaluation result to generate a coordinated feedback semantic vector, synchronizing to multi-modal feedback data of the corresponding virtual character according to a feedback timestamp, and outputting a learner-oriented combined feedback.

Description

Technical Field

[0001] This invention relates to the field of human-computer interaction technology, and more specifically, to an intelligent educational feedback method and system based on virtual character interaction. Background Technology

[0002] In the current development of educational informatization, the participation of virtual characters in teaching interactions has become a common feature in intelligent education platforms. In scenarios where multiple virtual characters collaborate on the same learning task, different virtual characters may provide feedback from different teaching perspectives, and these various types of feedback are often generated independently in terms of semantic expression, emotional style, and behavioral performance, lacking a unified coordination mechanism.

[0003] In existing multi-virtual-role combined feedback scenarios, there is a lack of a mechanism to coordinate and regulate the inherent logical relationships and expressive differences between different feedbacks. This may lead to inconsistencies in the semantic stance and stylistic expression of the feedback content, thereby increasing the cognitive burden on learners to integrate the feedback information and making it difficult to achieve the optimal presentation of the feedback effect.

[0004] To address the aforementioned problems, a technical solution is provided. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide an intelligent educational feedback method and system based on virtual character interaction to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A smart educational feedback method based on virtual character interaction includes the following steps: S1: Obtain multimodal feedback data generated by multiple virtual characters for the same learning task, and construct a multimodal feedback dataset; S2: Perform semantic stance dual analysis on the multimodal feedback dataset, identify the stance repulsion relationship between feedback semantic vectors, quantify the differences between feedback language and facial expression style, and output a set of conflict features; S3: Based on the set of conflict features, and combining the weights of the learning task objectives and the teaching function weights of the virtual role, priority is mapped to each feedback semantic vector to generate a feedback priority sequence; S4: Based on the feedback priority sequence, the cognitive load gradient prediction model is called to calculate the learner's information integration load value and obtain the cognitive load assessment result; S5: Based on the cognitive load assessment results, the feedback semantic vector is reconstructed in terms of expression style and content simplification according to the principle of minimum cognitive load, and a coordinated feedback semantic vector is generated. S6: Synchronize the coordinated feedback semantic vector to the multimodal feedback data of the corresponding virtual character according to the feedback timestamp, and output combined feedback for learners.

[0007] In a preferred embodiment, S1 specifically refers to: Acquire multimodal feedback data generated by multiple virtual characters for the same learning task, including language feedback data, facial expression feedback data, and gesture feedback data; Based on the virtual character identity, language feedback data, facial expression feedback data, and gesture feedback data are collected and feedback timestamps are recorded; Sentence segmentation and semantic encoding are performed on the language feedback data to generate feedback semantic vectors; By associating the character identity, feedback timestamp, and feedback semantic vector with the corresponding facial expression feedback data and gesture feedback data, a multimodal feedback dataset is obtained.

[0008] In a preferred embodiment, S2 specifically refers to: Based on the multimodal feedback dataset, feedback semantic vectors are paired according to role identity and feedback timestamp to form a set of feedback semantic vector pairs; Based on the feedback semantic vector, semantic similarity and negation polarity differences are calculated for sets to identify the positional repulsion relationship between feedback semantic vectors; Language style vectors are formed by extracting sentence length distribution, emotional word ratio and instruction intensity features from language feedback data. Based on facial expression feedback data, facial expression category sequences and intensity features are extracted to form facial expression style vectors; The style difference degree is calculated based on the language style vector and the expression style vector, and then summarized and associated with the positional repulsion relationship to output a set of conflict features.

[0009] In a preferred embodiment, S3 specifically refers to: Based on the positional incompatibilities and style differences corresponding to each feedback semantic vector in the conflict feature set, a conflict assessment matrix is ​​constructed. The positional repulsion relationships in the conflict assessment matrix are weighted and corrected based on the learning task objective weights. A hierarchical mapping of style differences in the conflict assessment matrix is ​​performed based on the weights of virtual role teaching functions. The feedback semantic vectors are comprehensively sorted based on the weighted and corrected positional repulsion relationship and the style difference degree after hierarchical mapping to form a feedback priority sequence.

[0010] In a preferred embodiment, S4 specifically refers to: Based on the feedback priority sequence, feedback semantic vectors with adjacent priorities are paired sequentially to form a combination of feedback semantic vectors; For feedback semantic vector combinations that exhibit positional incompatibilities, the cognitive load gradient prediction model is invoked to calculate the learner's information integration load value based on the degree of semantic and style differences in the feedback semantic vectors, thereby generating cognitive load assessment results.

[0011] In a preferred embodiment, S5 specifically refers to: Based on the cognitive load assessment results, the feedback semantic vectors with positional incompatibilities and information integration load values ​​exceeding a preset threshold are identified in the feedback semantic vector combinations. Based on the principle of minimum cognitive load, the expressive style of the feedback semantic vector is reconstructed, the sentence length and the proportion of emotional words in the language feedback data are adjusted, and the expression category and intensity of the facial expression feedback data are corrected. Redundant semantic content is removed from the feedback semantic vector after the expression style is reconstructed, and a coordinated feedback semantic vector is generated.

[0012] In a preferred embodiment, S6 specifically refers to: Based on the feedback timestamp and virtual character identity associated with the coordinated feedback semantic vector, the corresponding language feedback data, facial expression feedback data and gesture feedback data are determined; Based on the feedback timestamp, the coordinated feedback semantic vector is synchronously fused with the determined language feedback data, facial expression feedback data, and gesture feedback data to generate combined feedback for learners.

[0013] On the other hand, the present invention provides an intelligent educational feedback system based on virtual character interaction, comprising: Data acquisition module: Acquire multimodal feedback data generated by multiple virtual characters for the same learning task, and construct a multimodal feedback dataset; Conflict Analysis Module: Performs semantic stance dual analysis on multimodal feedback datasets, identifies the stance repulsion relationship between feedback semantic vectors, quantifies the differences in feedback language and facial expression style, and outputs a set of conflict features; Priority mapping module: Based on the conflict feature set, combined with the learning task objective weight and the virtual role teaching function weight, priority is mapped to each feedback semantic vector to generate a feedback priority sequence; Load assessment module: Based on the feedback priority sequence, it calls the cognitive load gradient prediction model to calculate the learner's information integration load value and obtain the cognitive load assessment result; Style Reconstruction Module: Based on the cognitive load assessment results, the feedback semantic vector is reconstructed in terms of expression style and content simplification according to the principle of minimum cognitive load, and a coordinated feedback semantic vector is generated. Feedback Fusion Module: Synchronizes the coordinated feedback semantic vector to the multimodal feedback data of the corresponding virtual character according to the feedback timestamp, and outputs combined feedback for learners.

[0014] The technical effects and advantages of the intelligent educational feedback method and system based on virtual character interaction of this invention are as follows: By acquiring multimodal feedback data generated by multiple virtual characters on the same learning task and constructing a multimodal feedback dataset, the identity of virtual characters and feedback time information are uniformly organized. Through semantic position dual analysis of the multimodal feedback dataset, the positional incompatibilities between feedback semantic vectors are identified, and the differences in feedback language and facial expression styles are quantified, enabling the characterization of potential conflicts in combined feedback from a semantic and expressive perspective. Based on the conflict feature set, learning task objective weights and virtual character teaching function weights are introduced to map priorities to feedback semantic vectors and form a feedback priority sequence, enabling controllable coordination of multi-virtual character feedback in terms of presentation order and importance. Based on the anti- The feedback priority sequence calls the cognitive load gradient prediction model to calculate the learner's information integration load value and generate cognitive load assessment results, thus quantifying the impact of feedback conflict on the learner's comprehension cost. Based on the cognitive load assessment results, the feedback semantic vector is reconstructed in terms of expression style and content simplification according to the principle of minimum cognitive load, and a coordinated feedback semantic vector is generated, so that the combined feedback is unified in terms of semantic consistency and expression coordination. Finally, the coordinated feedback semantic vector is synchronized to the multimodal feedback data of the corresponding virtual role according to the feedback timestamp and the combined feedback is output to the learner, thereby improving the comprehensibility and guidance stability of multi-virtual role interactive feedback. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of an intelligent educational feedback method based on virtual character interaction according to the present invention; Figure 2 This is a schematic diagram of the structure of an intelligent educational feedback system based on virtual character interaction according to the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0017] Example 1: Figure 1 This invention presents an intelligent educational feedback method based on virtual character interaction, which includes the following steps: S1: Obtain multimodal feedback data generated by multiple virtual characters for the same learning task, and construct a multimodal feedback dataset; S2: Perform semantic stance dual analysis on the multimodal feedback dataset, identify the stance repulsion relationship between feedback semantic vectors, quantify the differences between feedback language and facial expression style, and output a set of conflict features; S3: Based on the set of conflict features, and combining the weights of the learning task objectives and the teaching function weights of the virtual role, priority is mapped to each feedback semantic vector to generate a feedback priority sequence; S4: Based on the feedback priority sequence, the cognitive load gradient prediction model is called to calculate the learner's information integration load value and obtain the cognitive load assessment result; S5: Based on the cognitive load assessment results, the feedback semantic vector is reconstructed in terms of expression style and content simplification according to the principle of minimum cognitive load, and a coordinated feedback semantic vector is generated. S6: Synchronize the coordinated feedback semantic vector to the multimodal feedback data of the corresponding virtual character according to the feedback timestamp, and output combined feedback for learners.

[0018] S1: Obtain multimodal feedback data generated by multiple virtual characters for the same learning task, and construct a multimodal feedback dataset, including: Acquire multimodal feedback data generated by multiple virtual characters for the same learning task; Specifically, a unified learning task is set, such as solving mathematical equations. The identities of multiple virtual roles participating in the feedback are determined, including but not limited to virtual teachers, virtual peers, and virtual tutors. Multimodal feedback data includes verbal feedback data, facial expression feedback data, and gestural feedback data. Verbal feedback data consists of the textual data output through the virtual character's voice dialogue module during interaction with the learner, such as prompts, questions, or evaluations from the virtual character regarding the learning task. Facial expression feedback data includes facial expression data during virtual character interaction, such as smiles, frowns, surprise, and corresponding intensity parameters. Gesture feedback data includes the types and trajectories of gestures formed by the virtual character's body movements during interaction, such as gestures indicating, inviting, and refusing, and corresponding motion trajectory parameters.

[0019] Based on the virtual character identity, language feedback data, facial expression feedback data, and gesture feedback data are collected and feedback timestamps are recorded; Specifically, a virtual character identity refers to a unique identifier pre-assigned to each virtual character in the educational interaction system. This identifier can be generated using a combination of letters and numbers, for example, "Role_A01". The feedback timestamp represents the precise moment when the virtual character outputs various feedback data, recorded to the millisecond level using a standard time format. Based on the virtual character identity, separate data storage tables are established for verbal feedback data, facial expression feedback data, and gesture feedback data. The feedback timestamp is used as a unified index parameter to achieve a one-to-one correspondence between these three types of data. This method ensures that each piece of verbal feedback data accurately corresponds to specific facial expression feedback data and gesture feedback data.

[0020] Sentence segmentation and semantic encoding are performed on the language feedback data to generate feedback semantic vectors; Specifically, the language feedback data is segmented into sentences, for example, using sentence segmentation algorithms based on punctuation and grammatical structure to divide the language feedback data into independent semantic expression units. Semantic encoding is then performed on each segmented sentence using a hybrid encoding method combining word embedding and sentence embedding to convert the sentence's text content into vector form. For example, each sentence is segmented into words, and a word sequence is obtained using a general dictionary. Word embedding methods, such as Word2Vec or GloVe models, are used to convert each word into a high-dimensional vector representation. Then, a weighted average of all word vectors in the sentence is taken, or a recurrent neural network is used to fuse them, ultimately generating a feedback semantic vector that reflects the sentence's semantic meaning. The feedback semantic vector is typically a high-dimensional real-number vector; the dimension is determined by optimizing the parameters of the semantic encoding model on the language feedback data in the training set, based on the principle of maximizing semantic recognition accuracy. For example, the feedback semantic vector is set to a 128-dimensional or 256-dimensional real-number vector representation.

[0021] By associating the character identity, feedback timestamp, and feedback semantic vector with the corresponding facial expression feedback data and gesture feedback data, a multimodal feedback dataset is obtained. Specifically, using the feedback timestamp and virtual character identity as common index fields, the feedback semantic vector is associated with the corresponding facial expression feedback data and gesture feedback data to finally generate a multimodal feedback dataset, including virtual character identity, feedback timestamp, feedback semantic vector, corresponding facial expression feedback data (including facial expression category sequence and intensity features) and corresponding gesture feedback data (including gesture type and motion trajectory parameters).

[0022] S2: Perform semantic stance dual analysis on the multimodal feedback dataset to identify the stance repulsion relationship between feedback semantic vectors, quantify the differences in feedback language and facial expression style, and output a set of conflict features, including: Based on the multimodal feedback dataset, feedback semantic vectors are paired according to role identity and feedback timestamp to form a set of feedback semantic vector pairs; Specifically, feedback semantic vectors are extracted from the multimodal feedback dataset. Each feedback semantic vector is associated with a specific virtual character identity and a feedback timestamp. Therefore, the feedback semantic vectors in the multimodal feedback dataset are grouped according to the virtual character identity. Within the same virtual character identity group, adjacent feedback semantic vectors are combined according to the chronological order of the feedback timestamps to form a set of feedback semantic vector pairs. The set of feedback semantic vector pairs consists of multiple sets of feedback semantic vector pairs, and each set contains two feedback semantic vectors corresponding to different feedback timestamps.

[0023] Based on the feedback semantic vector, semantic similarity and negation polarity differences are calculated for sets to identify the positional repulsion relationship between feedback semantic vectors; Specifically, for each pair of feedback semantic vectors in the set, semantic similarity is calculated using a cosine similarity algorithm. This involves performing a dot product operation on the high-dimensional real vector representation of the feedback semantic vectors, then dividing by the product of the magnitudes of the respective feedback semantic vectors. The cosine similarity value ranges from 0 to 1; a value closer to 0 indicates a greater semantic difference, while a value closer to 1 indicates greater semantic similarity. For each pair of feedback semantic vectors in the set, the difference in negation polarity is quantified by labeling and statistically analyzing negation words in the language feedback data. A negation lexicon is established in the language feedback data, including but not limited to negation words such as "not," "no," "impossible," and "none." The frequency of occurrence of negation words in the language feedback data corresponding to the feedback semantic vectors is identified, and the degree of difference in the frequency of negation word usage between the two feedback semantic vectors in the pair is calculated. For example, the quantification of negation polarity difference is represented by the ratio of the difference in the frequency of negation words in the language feedback data corresponding to the two feedback semantic vectors to the total number of words, ranging from 0 to 1; a larger value indicates a more significant difference in negation polarity. Based on semantic similarity and negation polarity difference, a threshold combination for determining positional repulsion is set. The threshold is determined by conducting multiple experiments on the previously labeled training dataset to analyze the feature combinations between feedback semantic vectors that are manually judged as positional repulsion, and determining the optimal threshold combination of semantic similarity and negation polarity difference. For example, when the semantic similarity is less than 0.3 and the negation polarity difference is greater than 0.5, it is determined that there is a positional repulsion relationship between the feedback semantic vector pairs.

[0024] Language style vectors are formed by extracting sentence length distribution, emotional word ratio and instruction intensity features from language feedback data. Specifically, the sentence length distribution represents the statistical distribution of the number of characters in each sentence of the language feedback data, such as extracting the mean, variance, and the number of characters in the longest and shortest sentences; the emotion word ratio represents the proportion of words containing emotion words in the language feedback data to the total number of words, including but not limited to positive emotion words such as "happy," "satisfied," and "praise," as well as negative emotion words such as "disappointed," "confused," and "criticism"; the instruction intensity feature represents the proportion of sentences expressed as imperative or command sentences in the language feedback data to the total number of sentences. The language feedback data is segmented sentence by sentence, and the number of characters in each sentence is counted to form the sentence length distribution; a pre-defined emotion lexicon is called to match and count each word in the language feedback data, calculating the emotion word ratio; then, the number of command or imperative sentences in the language feedback data is counted, and the instruction intensity feature is calculated; finally, the sentence length distribution, emotion word ratio, and instruction intensity feature are combined to form a high-dimensional real vector, i.e., a language style vector, for example, the language style vector is set as a three-dimensional real vector.

[0025] Based on facial expression feedback data, facial expression category sequences and intensity features are extracted to form facial expression style vectors; Specifically, the expression category sequences recorded in the facial expression feedback data are processed by category encoding. Different categories of expressions are vectorized using one-hot encoding. For example, a smiling expression is encoded as [1,0,0], a frowning expression as [0,1,0], and a surprised expression as [0,0,1]. Expression intensity features are extracted. Expression intensity features are quantitative indicators of the amplitude or strength of the virtual character's movements when displaying various expressions. The value ranges from 0 to 1, with a larger value indicating a higher expression intensity. The expression category sequence encoding and intensity features are then combined to form a high-dimensional real number vector, i.e., an expression style vector. For example, it can be set as a real number vector with a dimension equal to the sum of the number of expression categories and the dimension of the intensity features.

[0026] The style difference degree is calculated based on the language style vector and the facial expression style vector, and then summarized and associated with the positional repulsion relationship to output a set of conflict features. Specifically, the Euclidean distance algorithm is used to calculate the dissimilarity between language style vectors and facial expression style vectors. The sum of squared differences in the corresponding dimensions of the language style vectors and facial expression style vectors is calculated, and the square root of the sum of squares is taken to obtain the style dissimilarity. The calculated style dissimilarity is then correlated with positional repulsion relationships; that is, for sets of feedback semantic vector pairs with positional repulsion relationships, the corresponding style dissimilarity is extracted. Finally, the positional repulsion relationships and the corresponding style dissimilarity are combined to form a conflict feature set. The conflict feature set includes a unique identifier for the feedback semantic vector pair, the determination result of the positional repulsion relationship, and the style dissimilarity.

[0027] S3: Based on the conflict feature set, and combining the learning task objective weights and the virtual role's teaching function weights, map priorities to each feedback semantic vector to generate a feedback priority sequence, including: Based on the positional incompatibilities and style differences corresponding to each feedback semantic vector in the conflict feature set, a conflict assessment matrix is ​​constructed. Specifically, the conflict assessment matrix is ​​a two-dimensional structured data table based on a set of conflict features, with feedback semantic vectors as row indices and positional incompatibility and style difference as column indices. Each element in the conflict assessment matrix stores the positional incompatibility judgment value and the quantified value of style difference corresponding to a single feedback semantic vector. The positional incompatibility judgment value is a binary variable; it takes a value of 1 when the semantic similarity is less than 0.3 and the negation polarity difference is greater than 0.5, indicating the existence of a positional incompatibility relationship; otherwise, it takes a value of 0, indicating the absence of a positional incompatibility relationship. The style difference is a continuous variable, with its value calculated using the Euclidean distance algorithm. It represents the square root of the sum of the squares of the differences between the corresponding dimensions of the language style vector and the expression style vector, for example, a calculated style difference of 0.65 or 1.32.

[0028] The positional repulsion relationships in the conflict assessment matrix are weighted and corrected based on the learning task objective weights. Specifically, the learning task target weight refers to a pre-set coefficient used to adjust the importance of positional conflict for a specific learning task type. The method for determining the learning task target weight is as follows: Establish a learning task category library, including but not limited to knowledge memorization tasks, problem-solving tasks, and creative thinking tasks; based on the differences in sensitivity to positional conflict among different learning task categories, determine the initial weight values ​​using expert scoring, then refine and optimize using historical interaction data, finally determining the learning task target weight corresponding to each learning task category. For example, knowledge memorization tasks have a higher requirement for consistent feedback, therefore a larger weight, for example, set between 0.8 and 1.0; creative thinking tasks require diverse feedback perspectives, so the weight is relatively smaller, for example, set between 0.2 and 0.5. Extract positional conflict judgment values ​​from the conflict assessment matrix, and use the learning task target weight to weight and correct these judgment values. The weighting calculation method is: the weighted positional conflict value equals the original positional conflict judgment value multiplied by the learning task target weight. For example, if the original judgment value is 1 and the learning task target weight is 0.9, then the weighted corrected positional conflict value is 0.9.

[0029] A hierarchical mapping of style differences in the conflict assessment matrix is ​​performed based on the weights of virtual role teaching functions. Specifically, the teaching function weight of virtual characters is a weighting coefficient assigned based on the role responsibilities undertaken by the virtual character in teaching activities. The method for determining the teaching function weight of virtual characters is as follows: Define teaching function categories, such as, but not limited to, guiding roles, auxiliary roles, and assessment roles; based on the teaching function positioning and style consistency requirements of each role, determine the teaching function weight of virtual characters through expert evaluation or interactive feedback historical data analysis. For example, guiding roles need to maintain a consistent style and have a higher weight, such as set between 0.7 and 1.0; auxiliary roles allow for a certain degree of style difference and have a lower weight, such as set between 0.3 and 0.6. Based on the determined teaching function weight of virtual characters, the original style difference degree in the conflict assessment matrix is ​​subjected to graded mapping processing. Specifically, multiple style difference degree thresholds are set; for example, style difference degree levels within the range of 0 to 0.5 are defined as low level, style difference degree levels within the range of 0.5 to 1.0 are defined as medium level, and style difference degree levels above 1.0 are defined as high level; then, the levels are adjusted a second time in conjunction with the teaching function weight of virtual characters, by increasing or decreasing the style difference degree level. For example, for high-weight virtual characters, the style difference value at the medium level is increased to the high level, while for low-weight virtual characters, the high-level style difference value is reduced to the medium level or remains unchanged.

[0030] Based on the weighted and corrected positional repulsion relationship and the hierarchical mapping of style differences, the feedback semantic vectors are comprehensively sorted to form a feedback priority sequence. Specifically, the weighted and corrected positional incompatibility values ​​and the style difference levels after hierarchical mapping are extracted from the conflict assessment matrix. A comprehensive ranking evaluation function is determined, which is a weighted combination function of the positional incompatibility values ​​and style difference levels, specifically: Feedback Priority Score = (Positional Incompatibility Value × Positional Influence Coefficient) + (Style Difference Level × Style Difference Influence Coefficient). The positional influence coefficient and style difference influence coefficient are preset weighting coefficients used to balance the importance of positional relationship and style difference. They are determined by analyzing the impact of feedback on learners' cognitive effects through historical interaction data in teaching feedback effectiveness evaluation experiments, thereby optimizing the values ​​of the positional influence coefficient and style difference influence coefficient. For example, after optimization, the positional influence coefficient is determined to be 0.6, and the style difference influence coefficient is 0.4. The feedback priority score of all feedback semantic vectors is calculated using the evaluation function, and the feedback priority scores are sorted from high to low to form a feedback priority sequence. The feedback priority sequence is a sequence record of feedback semantic vectors and feedback priority scores.

[0031] S4: Based on the feedback priority sequence, the cognitive load gradient prediction model is invoked to calculate the learner's information integration load value, obtaining the cognitive load assessment results, including: Based on the feedback priority sequence, feedback semantic vectors with adjacent priorities are paired sequentially to form a combination of feedback semantic vectors; Specifically, all feedback semantic vectors are extracted from the feedback priority sequence. Adjacent feedback semantic vectors are grouped into sets according to their feedback priority scores in descending order. Each set contains two feedback semantic vectors, each with a specific feedback priority score in the feedback priority sequence. For example, three feedback semantic vectors with priority scores of 0.95, 0.89, and 0.80 in the feedback priority sequence are combined into two sets. The first set contains vectors with priority scores of 0.95 and 0.89, and the second set contains vectors with priority scores of 0.89 and 0.80.

[0032] For feedback semantic vector combinations that have a positional conflict relationship, the cognitive load gradient prediction model is invoked to calculate the learner's information integration load value based on the degree of semantic difference and style difference of the feedback semantic vectors, and the cognitive load assessment result is generated. Specifically, for each set of feedback semantic vector combinations, the two feedback semantic vectors are used to determine whether there is a positional conflict between them, based on the semantic similarity threshold and the negation polarity difference threshold. For example, if the semantic similarity between the two feedback semantic vectors in a set of feedback semantic vector combinations is less than 0.3 and the negation polarity difference is greater than 0.5, then the set of feedback semantic vectors is determined to be a set of feedback semantic vectors with a positional conflict.

[0033] For feedback semantic vector combinations determined to have a positional conflict, the semantic difference and style difference between the two feedback semantic vectors within the combination are calculated. The semantic difference is calculated using cosine distance, which is calculated as 1 minus the cosine similarity between the two feedback semantic vectors in the combination. For example, if the cosine similarity between two feedback semantic vectors is 0.2, the cosine distance is 0.8, indicating a significant semantic difference. The style difference is calculated using the Euclidean distance algorithm, which involves calculating the square root of the sum of the squared differences between the corresponding dimensional values ​​of the language style vector and the facial expression style vector of the two feedback semantic vectors in the combination.

[0034] The cognitive load gradient prediction model calculates the learner's information integration load by using the semantic and style differences within the feedback semantic vector combination as input. This model is a computational model obtained through supervised learning on a large amount of prior training data. Its inputs are the semantic and style differences, and its output is the information integration load, representing the cognitive load required for the learner to integrate feedback information. The implementation method of the cognitive load gradient prediction model is as follows: based on a large amount of historical data collected from interactions between virtual characters and learners, training samples labeled with cognitive load levels are obtained by analyzing learners' task performance under different feedback semantic and style differences. The training samples are then trained using support vector regression or gradient boosting tree algorithms to optimize and obtain the best cognitive load gradient prediction model. Key parameters during model training include kernel function type, regularization coefficient, and the number and depth of trees. For example, the kernel function for support vector regression is set to radial basis function, and the regularization coefficient is set between 0.1 and 1.0, such as 0.5. If gradient boosting tree algorithm is used, the number of trees is set to 100 to 200, and the tree depth is set to 3 to 5 layers, such as 150 trees and a depth of 4 layers. After model training is completed, the generalization performance of the model is determined by cross-validation. When the prediction error of cross-validation is less than a preset accuracy threshold, such as a mean squared error of less than 0.05, the cognitive load gradient prediction model is determined to be usable.

[0035] Using a cognitive load gradient prediction model, the degree of semantic difference and style difference are input to obtain the information integration load value. The information integration load value ranges from 0 to 1; for example, the closer the information integration load value is to 1, the higher the learner's cognitive load when processing the corresponding feedback semantic vector combination, and the greater the integration difficulty; the closer the information integration load value is to 0, the lower the learner's cognitive load when integrating the feedback, and the easier the integration. Finally, the calculated information integration load value is correlated with the feedback semantic vector combination to generate the cognitive load assessment result. The cognitive load assessment result includes a unique identifier for the feedback semantic vector combination, the degree of semantic difference, the style difference, and the corresponding information integration load value.

[0036] S5: Based on the cognitive load assessment results, the feedback semantic vector is reconstructed in terms of expression style and simplified in accordance with the principle of minimum cognitive load, generating a reconciled feedback semantic vector, including: Based on the cognitive load assessment results, the feedback semantic vectors with positional incompatibilities and information integration load values ​​exceeding a preset threshold are identified in the feedback semantic vector combinations. Specifically, the positional incompatibility determination results and information integration load values ​​corresponding to each set of feedback semantic vector combinations are extracted from the cognitive load assessment results. Feedback semantic vector combinations are then screened based on a preset information integration load threshold to determine the combinations to be reconstructed. The method for setting the information integration load threshold is as follows: based on historical learning interaction data, the optimal threshold for information integration load is determined by recording learners' task completion accuracy and response time performance under different information integration load values ​​and using performance-load curve analysis. For example, statistical analysis revealed that when the information integration load value exceeds 0.75, learners' task completion accuracy significantly decreases, indicating that an information integration load value exceeding 0.75 represents a high cognitive load state; therefore, the information integration load threshold is set to 0.75. Once determined, when the information integration load value corresponding to a feedback semantic vector combination is greater than 0.75 and a positional incompatibility relationship exists simultaneously, the feedback semantic vector combination is identified as the feedback semantic vector combination requiring expression style reconstruction.

[0037] Based on the principle of minimum cognitive load, the expressive style of the feedback semantic vector is reconstructed, the sentence length and the proportion of emotional words in the language feedback data are adjusted, and the expression category and intensity of the facial expression feedback data are corrected. Specifically, the principle of minimum cognitive load refers to minimizing the learner's cognitive load when integrating feedback information by adjusting the way feedback information is expressed. Based on the principle of minimum cognitive load, style reconstruction is performed on the language feedback data and facial expression feedback data corresponding to the feedback semantic vectors. Style reconstruction of language feedback data includes adjusting sentence length and the proportion of emotional words. The method for adjusting sentence length is as follows: by performing syntactic analysis on the language feedback data corresponding to the feedback semantic vectors, the length of each sentence is calculated. Based on the information comprehension efficiency curve of learners under different sentence length conditions, the optimal range of sentence length is determined. For example, experimental analysis found that when the length of each sentence is controlled within the range of 8 to 15 characters, the learner's information comprehension efficiency is the highest. Therefore, by reconstructing sentences on the language feedback data corresponding to the feedback semantic vectors, long sentences with more than 15 characters are broken into multiple short sentences, and short sentences with less than 8 characters are expanded to more than 8 characters through information supplementation. The method for adjusting the proportion of emotion words is as follows: Emotion words are identified in the language feedback data corresponding to the feedback semantic vector. Based on a preset emotion word library, emotion words in the language feedback data are identified, including positive emotion words (such as "satisfied", "excellent", "encouraging") and negative emotion words (such as "disappointed", "needs improvement", "insufficient"). The proportion of emotion words in the language feedback data to the total vocabulary is counted. The optimal range of emotion word proportion is determined based on historical data analysis. For example, if the optimal range of emotion word proportion is determined to be 10% to 20% through historical interaction analysis, the proportion of emotion words in the original language feedback data is adjusted to the optimal range of emotion word proportion, specifically by deleting excessive emotion words or adding an appropriate number of emotion words.

[0038] Style reconstruction of facial expression feedback data includes revising expression categories and intensity. The expression category revision method involves statistically analyzing the correlation between learners' attention span and expression categories in response to different feedback content based on historical learning interaction data. This determines appropriate expression category selection principles for reducing cognitive load. For example, when the feedback is guiding, learners respond more positively to expressions like "smile" and "encouraging nod"; when the feedback is corrective, learners are more focused on expressions like "serious" and "attentive." By adjusting the original expression categories corresponding to the feedback semantic vectors to suitable categories based on the feedback type, learners' cognitive load is reduced. The expression intensity revision method involves statistically analyzing the expression intensity in the facial feedback data corresponding to the feedback semantic vectors. The optimal range for expression intensity is determined using learners' historical response data. For example, analysis results show that learners experience the lowest cognitive load when expression intensity is between 0.4 and 0.7; therefore, the expression intensity in the facial feedback data corresponding to the feedback semantic vectors is adjusted accordingly.

[0039] Redundant semantic content is removed from the feedback semantic vector after the expression style is reconstructed, and a reconciled feedback semantic vector is generated. Specifically, redundant semantic content refers to content fragments in the feedback semantic vector that express similarities, repetitions, or fail to contribute to the current learning task objective. The method for deleting redundant semantic content is as follows: Perform semantic similarity analysis again on the feedback semantic vector after style reconstruction; perform word vector semantic similarity analysis on the language feedback data corresponding to the feedback semantic vector; if multiple content fragments have a semantic similarity higher than 0.85, delete the corresponding low-similarity fragments. Based on a preset learning task objective relevance threshold (determined through historical learning task objective relevance annotation data, for example, setting the threshold to 0.6), calculate the semantic relevance of each semantic fragment in the feedback semantic vector to the current learning task objective; content fragments with a relevance lower than the learning task objective relevance threshold are identified as redundant semantic content and deleted. Through the above processing, a coordinated feedback semantic vector is obtained.

[0040] S6: Synchronize the coordinated feedback semantic vector to the corresponding virtual character's multimodal feedback data according to the feedback timestamp, and output combined feedback for learners, including: Based on the feedback timestamp and virtual character identity associated with the coordinated feedback semantic vector, the corresponding language feedback data, facial expression feedback data and gesture feedback data are determined; Specifically, the reconciled feedback semantic vector retains the feedback timestamp and virtual character identity corresponding to the initial generation. The feedback timestamp records the moment the virtual character generates feedback data; the virtual character identity is a unique identifier pre-defined for the virtual character within the educational interaction system. Based on a relational database query method, the reconciled feedback semantic vector uses the feedback timestamp and virtual character identity as a joint index to perform a database search, obtaining the corresponding language feedback data, facial expression feedback data, and gesture feedback data.

[0041] When the coordinated feedback semantic vector is indexed using the feedback timestamp and virtual character identity, precise retrieval of language feedback data is achieved. Each record in the language feedback data storage table contains the feedback text content, feedback timestamp, and virtual character identity. A dual-condition query using the feedback timestamp and virtual character identity fields uniquely retrieves the corresponding language feedback text content, such as "Please carefully check the third step in the formula derivation process." The retrieval methods for facial expression feedback data and gesture feedback data are consistent with those for language feedback data. A joint index query is performed in their respective data storage tables to obtain the facial expression feedback data and gesture feedback data under the corresponding feedback timestamp and virtual character identity. For example, the results of facial expression feedback data retrieval include the expression category and expression intensity parameter, such as an expression category of "attention" and an expression intensity of 0.65; the results of gesture feedback data retrieval include the gesture type and motion trajectory parameter, such as a gesture type of "indication" and a motion trajectory parameter that is a two-dimensional coordinate sequence representing the virtual character's gesture movement path within the virtual interaction space.

[0042] Based on the feedback timestamp, the coordinated feedback semantic vector is synchronously fused with the determined language feedback data, facial expression feedback data and gesture feedback data to generate combined feedback for learners. Specifically, synchronous fusion is defined as aligning the language feedback data, facial expression feedback data, and gesture feedback data corresponding to the coordinated feedback semantic vector in the time dimension through feedback timestamps. This ensures that the language feedback data, facial expression feedback data, and gesture feedback data are synchronized on the time axis using a unified feedback timestamp as a benchmark. For example, for a coordinated feedback semantic vector with a feedback timestamp of "2025-02-11 14:30:15.325", the corresponding language feedback data contains "Please carefully check the third step in the formula derivation process", the facial expression feedback data contains the category "attention" and the intensity 0.65, and the gesture feedback data contains the type "instruction" and the motion trajectory parameters are the corresponding spatial coordinate sequence. Then, the language feedback data, facial expression feedback data, and gesture feedback data are synchronously output to the learner at "2025-02-11 14:30:15.325", generating a combined feedback for the learner.

[0043] The combined feedback generation process is as follows: Language feedback data is transmitted to the learner in real-time by converting text content into audio data through the virtual character's voice output module. A preset speech synthesis algorithm, such as an end-to-end speech synthesis method based on deep learning, is used. The input is the text content of the language feedback data, and the output is the corresponding audio waveform data. Key parameters of the speech synthesis algorithm include acoustic model type, speech synthesis speed, and speech style parameters. For example, the acoustic model uses a Transformer-based end-to-end model, the speech synthesis speed is set between 150 and 180 words per minute, and the speech style parameter is set to a "guided" style. The synthesized speech feedback data is played to the learner in the voice of the virtual character at the time marked by the feedback timestamp.

[0044] Facial expression feedback data is presented to learners in real time within a virtual interactive environment through the facial expression output method of a virtual character. Based on real-time rendering technology using a 3D virtual character model, the expression category and intensity parameters are mapped to the virtual character's facial expression animation parameters. A preset facial animation parameter mapping model is used, such as a linear or nonlinear interpolation model, to map the category "attention" and intensity 0.65 of the facial expression feedback data to an animation sequence in a preset facial animation database. Key parameters of the model include the type of facial animation interpolation function and the animation frame rate setting. For example, a cubic spline interpolation function is used, and the animation frame rate is set to 30 frames per second, rendering the real-time virtual character facial expression output that generates the facial expression feedback data.

[0045] Gesture feedback data is presented to learners in real time within the virtual interactive space through the gesture output method of the virtual character. Using a pre-defined virtual character gesture control algorithm, the gesture type and motion trajectory parameters in the gesture feedback data are mapped to a sequence of limb movements of the virtual character. For example, using an inverse kinematics algorithm, the input is the gesture type "indication" and the motion trajectory coordinate sequence parameters, and the output is a sequence of specific limb movement animations of the virtual character. Key parameters include the virtual character's skeletal structure model, the motion smoothness coefficient, and the trajectory interpolation method. For instance, the skeletal structure model uses a standard human skeletal structure, the motion smoothness coefficient is set to 0.85, and the trajectory interpolation method uses Bézier curve interpolation to ensure that the virtual character's gesture movements are smoothly presented to the learner.

[0046] By synchronously integrating the above-mentioned language feedback data, facial expression feedback data, and gesture feedback data, the speech content of the language feedback data, the facial animation content of the facial expression feedback data, and the body movement content of the gesture feedback data are simultaneously output to the learner, forming a unified and coordinated combined feedback that comprehensively conveys information to the learner. The combined feedback is presented as follows: at the designated feedback time "2025-02-11 14:30:15.325", the virtual character speaks the language feedback content in a guiding voice style, while simultaneously displaying a focused facial expression with moderate intensity, and performing gesture actions to execute instructions. This helps the learner focus on the key points of the learning task and the feedback content, achieving a coordinated and unified expression of educational feedback information across the three modalities of language, facial expression, and gesture, ensuring that the learner can accurately understand the feedback content with minimal cognitive load.

[0047] Example 2: The difference between Example 2 and Example 1 is that this example introduces an intelligent educational feedback system based on virtual character interaction.

[0048] Figure 2 A schematic diagram of the structure of an intelligent educational feedback system based on virtual role interaction according to the present invention is provided. The intelligent educational feedback system based on virtual role interaction includes: Data acquisition module: Acquire multimodal feedback data generated by multiple virtual characters for the same learning task, and construct a multimodal feedback dataset; Conflict Analysis Module: Performs semantic stance dual analysis on multimodal feedback datasets, identifies the stance repulsion relationship between feedback semantic vectors, quantifies the differences in feedback language and facial expression style, and outputs a set of conflict features; Priority mapping module: Based on the conflict feature set, combined with the learning task objective weight and the virtual role teaching function weight, priority is mapped to each feedback semantic vector to generate a feedback priority sequence; Load assessment module: Based on the feedback priority sequence, it calls the cognitive load gradient prediction model to calculate the learner's information integration load value and obtain the cognitive load assessment result; Style Reconstruction Module: Based on the cognitive load assessment results, the feedback semantic vector is reconstructed in terms of expression style and content simplification according to the principle of minimum cognitive load, and a coordinated feedback semantic vector is generated. Feedback Fusion Module: Synchronizes the coordinated feedback semantic vector to the multimodal feedback data of the corresponding virtual character according to the feedback timestamp, and outputs combined feedback for learners.

[0049] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0050] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0051] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0052] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0053] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0054] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0055] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0056] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0057] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A smart educational feedback method based on virtual character interaction, characterized in that, Includes the following steps: S1: Obtain multimodal feedback data generated by multiple virtual characters for the same learning task, and construct a multimodal feedback dataset; S2: Perform semantic stance dual analysis on the multimodal feedback dataset, identify the stance repulsion relationship between feedback semantic vectors, quantify the differences between feedback language and facial expression style, and output a set of conflict features; S3: Based on the set of conflict features, and combining the weights of the learning task objectives and the teaching function weights of the virtual role, priority is mapped to each feedback semantic vector to generate a feedback priority sequence; S4: Based on the feedback priority sequence, the cognitive load gradient prediction model is called to calculate the learner's information integration load value and obtain the cognitive load assessment result; S5: Based on the cognitive load assessment results, the feedback semantic vector is reconstructed in terms of expression style and content simplification according to the principle of minimum cognitive load, and a coordinated feedback semantic vector is generated. S6: Synchronize the coordinated feedback semantic vector to the multimodal feedback data of the corresponding virtual character according to the feedback timestamp, and output combined feedback for learners.

2. The intelligent educational feedback method based on virtual character interaction according to claim 1, characterized in that, S1, specifically: Acquire multimodal feedback data generated by multiple virtual characters for the same learning task, including language feedback data, facial expression feedback data, and gesture feedback data; Based on the virtual character identity, language feedback data, facial expression feedback data, and gesture feedback data are collected and feedback timestamps are recorded; Sentence segmentation and semantic encoding are performed on the language feedback data to generate feedback semantic vectors; By associating the character identity, feedback timestamp, and feedback semantic vector with the corresponding facial expression feedback data and gesture feedback data, a multimodal feedback dataset is obtained.

3. The intelligent educational feedback method based on virtual character interaction according to claim 2, characterized in that, S2, specifically: Based on the multimodal feedback dataset, feedback semantic vectors are paired according to role identity and feedback timestamp to form a set of feedback semantic vector pairs; Based on the semantic similarity and negation polarity difference of the feedback semantic vector set, the positional repulsion relationship between the feedback semantic vectors is identified. Language style vectors are formed by extracting sentence length distribution, emotional word ratio and instruction intensity features from language feedback data. Based on facial expression feedback data, facial expression category sequences and intensity features are extracted to form facial expression style vectors; The style difference degree is calculated based on the language style vector and the expression style vector, and then summarized and associated with the positional repulsion relationship to output a set of conflict features.

4. The intelligent educational feedback method based on virtual character interaction according to claim 3, characterized in that, S3, specifically: Based on the positional incompatibilities and style differences corresponding to each feedback semantic vector in the conflict feature set, a conflict assessment matrix is ​​constructed. The positional repulsion relationships in the conflict assessment matrix are weighted and corrected based on the learning task objective weights. A hierarchical mapping of style differences in the conflict assessment matrix is ​​performed based on the weights of virtual role teaching functions. The feedback semantic vectors are comprehensively sorted based on the weighted and corrected positional repulsion relationship and the style difference degree after hierarchical mapping to form a feedback priority sequence.

5. The intelligent educational feedback method based on virtual character interaction according to claim 4, characterized in that, S4, specifically: Based on the feedback priority sequence, feedback semantic vectors with adjacent priorities are paired sequentially to form a combination of feedback semantic vectors; For feedback semantic vector combinations that exhibit positional incompatibilities, the cognitive load gradient prediction model is invoked to calculate the learner's information integration load value based on the degree of semantic and style differences in the feedback semantic vectors, thereby generating cognitive load assessment results.

6. The intelligent educational feedback method based on virtual character interaction according to claim 5, characterized in that, S5, specifically: Based on the cognitive load assessment results, the feedback semantic vectors with positional incompatibilities and information integration load values ​​exceeding a preset threshold are identified in the feedback semantic vector combinations. Based on the principle of minimum cognitive load, the expressive style of the feedback semantic vector is reconstructed, the sentence length and the proportion of emotional words in the language feedback data are adjusted, and the expression category and intensity of the facial expression feedback data are corrected. Redundant semantic content is removed from the feedback semantic vector after the expression style is reconstructed, and a coordinated feedback semantic vector is generated.

7. The intelligent educational feedback method based on virtual character interaction according to claim 6, characterized in that, S6, specifically: Based on the feedback timestamp and virtual character identity associated with the coordinated feedback semantic vector, the corresponding language feedback data, facial expression feedback data and gesture feedback data are determined; Based on the feedback timestamp, the coordinated feedback semantic vector is synchronously fused with the determined language feedback data, facial expression feedback data, and gesture feedback data to generate combined feedback for learners.

8. An intelligent educational feedback system based on virtual character interaction, used to implement the intelligent educational feedback method based on virtual character interaction as described in any one of claims 1-7, characterized in that, include: Data acquisition module: Acquire multimodal feedback data generated by multiple virtual characters for the same learning task, and construct a multimodal feedback dataset; Conflict Analysis Module: Performs semantic position dual analysis on multimodal feedback datasets, identifies positional repulsion relationships between feedback semantic vectors, quantifies differences in feedback language and facial expression styles, and outputs a set of conflict features; Priority Mapping Module: Based on the conflict feature set, combined with the learning task objective weight and the virtual role teaching function weight, priority is mapped to each feedback semantic vector to generate a feedback priority sequence; Load assessment module: Based on the feedback priority sequence, it calls the cognitive load gradient prediction model to calculate the learner's information integration load value and obtain the cognitive load assessment result; Style Reconstruction Module: Based on the cognitive load assessment results, the feedback semantic vector is reconstructed in terms of expression style and content simplification according to the principle of minimum cognitive load, and a coordinated feedback semantic vector is generated. Feedback Fusion Module: Synchronizes the coordinated feedback semantic vector to the multimodal feedback data of the corresponding virtual character according to the feedback timestamp, and outputs combined feedback for learners.

Images