Machine learning-based basic education-oriented shih-chuan literary creation aesthetic education infiltration effectiveness evaluation system

Abstract

The application discloses a kind of based on machine learning's facing basic education poetic couplet text creative aesthetic education infiltration efficiency evaluation system, it is related to artificial intelligence education evaluation technical field, the present application includes data acquisition and window division module, text feature extraction module, CD-CVAE training module, mood hidden variable extraction and constraint clustering module, conditional entropy estimation and efficiency index calculation module and evaluation grade output module;The present application decouples the creation behavior into rhythm factor and mood factor by conditional variational autoencoder, the uncertainty of mood hidden variable is measured under the same constraint using conditional entropy, objectively quantifies the promotion degree of the compound ability of mood innovation under the constraint of rhythm from the information theory point of view, the evaluation process does not depend on subjective judgment, the result is reproducible, comparable, provides reliable technical means for the accurate measurement of aesthetic education effect.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence education assessment technology, and in particular relates to a machine learning-based evaluation system for the effectiveness of poetry, couplets, and cultural creative aesthetic education in basic education. Background Technology

[0002] With the increasing emphasis on the inheritance of excellent traditional Chinese culture in basic education, poetry and couplet-based cultural and creative aesthetic education activities have been widely integrated into primary and secondary school Chinese and art courses. Among the various forms of poetry and couplet creation, the embedded-character couplet creation requires students to embed specified characters in designated positions in the second line, while strictly adhering to metrical rules such as tonal correspondence and part-of-speech parallelism. The aim is to cultivate students' ability to innovate aesthetic concepts under certain formal constraints. In order to objectively evaluate the effect of this aesthetic education immersion, the student creative behavior analysis and ability modeling technology based on deep learning has gradually attracted attention. Existing mainstream technical solutions usually adopt sequence-to-sequence text generation evaluation models or semantic similarity measurement methods based on pre-trained language models. Among them, the pre-trained language model refers to a deep neural network trained on a large-scale corpus through self-supervised learning, which can map natural language text of arbitrary length into dense semantic vectors of fixed dimensions. The distance between these vectors in the semantic space is used to measure the content similarity or innovative deviation between texts.

[0003] However, the aforementioned general semantic vector evaluation methods face technical challenges arising from the unique nature of the creative form when applied to the specific aesthetic education scenario of embedded-character couplet creation. Specifically, the difficulty lies in the high coupling between metrical constraints and the ability to create artistic conception in text representation. A couplet containing embedded characters simultaneously contains two types of information: first, adherence to metrical rules such as tone patterns, parts of speech, and embedded character placement; and second, the ability to create artistic conception through imagery selection and rhetorical device construction under these constraints. Existing pre-trained language models, when semantically encoding the second line of the text, cannot actively distinguish these two types of information in the feature space. The degree of metrical compliance and the level of artistic conception expression are intertwined within the same semantic vector; this coupling is fundamentally different from the relationship between form and content in ordinary text evaluation. First, metrical constraints are mandatory discrete structural conditions, rather than flexible stylistic features. Violating metrical rules will directly lead to the work being unqualified, but existing semantic vectors are not sensitive enough to this discrete constraint. Second, the innovation of artistic conception under strict metrical constraints has the unique aesthetic attribute of freedom within constraints. The range of artistic conception changes in student works is far smaller than that in free creation. Subtle differences in themes are often submerged by the metrical compliance features that dominate the semantic model.

[0004] Under the interference of the coupling between prosody and artistic conception, existing semantic vector evaluation methods suffer from serious performance measurement bias. On the one hand, from a static evaluation perspective, if the cosine similarity between the semantic vector of the second line and the reference text is directly used as an indicator of innovation, the change in similarity is likely to reflect fluctuations in students' prosodic norms rather than a genuine improvement in their artistic conception innovation ability. The progress brought about by aesthetic education immersion may manifest as a shift from singular to richer artistic conception expression under the same prosodic accuracy, but existing methods cannot isolate the influence of prosodic factors to independently quantify changes in artistic conception factors. On the other hand, from a dynamic tracking perspective, aesthetic education immersion is a long-term, gradual process that requires comparing changes in students' artistic conception expression under the same prosodic constraints before and after the teaching period. However, since different creative tasks have different word embedding requirements and prosodic difficulty, directly comparing the distance or dispersion of semantic vectors before and after cannot eliminate the interference caused by differences in constraint difficulty, resulting in evaluation results that cannot truly reflect the effect of aesthetic education immersion on improving the core ability of innovation within constraints. This background technology mentions innovation within constraints, specifically referring to the ability of students to conceive couplets with novel imagery and aesthetic appeal even when faced with inviolable formal restrictions such as tonal patterns, parallelism, and word embedding. If the assessment system cannot decouple the ability to adhere to prosody from the ability to create artistic conception at the representational level, it cannot accurately measure whether aesthetic education truly promotes the growth of students' aesthetic creativity; therefore, the following solutions are proposed to address the above issues. Summary of the Invention

[0005] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: This invention is a machine learning-based evaluation system for the effectiveness of poetry and couplet cultural and creative aesthetic education in basic education, including a data acquisition and window segmentation module, a text feature extraction module, a CD-CVAE training module, an imagery latent variable extraction and constrained clustering module, a conditional entropy estimation and effectiveness index calculation module, and an evaluation level output module. The workflow of the evaluation system is as follows: Step S1: Collect the text data of the second line of students in multiple rounds of the embedded character couplet creation activity, record the timestamp and embedded character constraint label, and divide the creation sequence of each student into the initial time window and the later time window according to the teaching cycle. Step S2: Automatically extract the tonal sequence, part-of-speech sequence, and embedded character position mask for each second line of the poem, encode them into a metrical constraint vector, and at the same time use a pre-trained language model to extract the sentence-level semantic vector of the second line; Step S3: Construct a constraint-decoupling conditional variational autoencoder. The variational autoencoder includes an artistic conception posterior network, a metrical posterior network, a metrical prior network, and a generative network. It is trained using a loss function composed of reconstruction loss, artistic conception decoupling regularization term, and metrical encoding regularization term. The expression patterns related to metrical rules in the second couplet are encoded into the metrical latent variable, and the artistic conception innovation that is free from metrical constraints is separated into the artistic conception latent variable. Step S4: Fix the encoder parameters after training, input the semantic vectors of the lower couplet created by each student in the initial window and the later window into the posterior network of artistic conception, take the posterior mean as the latent variable of artistic conception, and divide the metrical constraint categories according to the similarity of the metrical constraint vectors to form the latent variable sample set of artistic conception under each constraint category in each window. Step S5: Use the K-nearest neighbor entropy estimation algorithm to calculate the local conditional entropy of the initial window and the later window under each metrical constraint category, and then use the sample proportion of each category as the weight to sum the global conditional entropy. Calculate the aesthetic education immersion effectiveness index based on the difference ratio between the initial global conditional entropy and the later global conditional entropy. Step S6: Based on the comparison results between the aesthetic education immersion effectiveness index and the preset threshold, automatically generate and output the evaluation level.

[0006] Furthermore, the data acquisition and window partitioning module, text feature extraction module, contextual latent variable extraction and constrained clustering module, conditional entropy estimation and efficiency index calculation module, and evaluation level output module are connected in sequence. The output end of the text feature extraction module is unidirectionally connected to the input end of the CD-CVAE training module, and the output end of the CD-CVAE training module is unidirectionally connected to the input end of the contextual latent variable extraction and constrained clustering module. The data acquisition and window division module is used to automatically collect the second line text submitted by students in multiple rounds of the embedded character couplet creation activity, record the timestamp and embedded character constraint label, and divide the creation sequence of each student into an initial time window and a later time window according to the teaching cycle, so as to provide structured data for subsequent time series comparison analysis. The text feature extraction module is used to perform metrical analysis and semantic encoding on each second line of the couplet, outputting metrical constraint vectors and sentence-level semantic vectors, and transforming the unstructured couplet text into a numerical feature representation that can be processed by machine learning models. The CD-CVAE training module is used to construct and train a constraint-decoupled conditional variational autoencoder. Through joint optimization of reconstruction loss, mood decoupling regularization term and metrical encoding regularization term, it automatically learns to decompose the second line of the work into mood latent variables and metrical latent variables, thereby achieving decoupling of mood innovation features and metrical constraint features. The latent variable extraction and constraint clustering module is used to fix the encoder parameters after training, extract the latent variables of the artistic conception created by each student in the initial and later windows, and cluster them according to the similarity of the metrical constraint vectors to form a sample set of latent variables of artistic conception under each window and each constraint category for use in conditional entropy calculation. The conditional entropy estimation and efficacy index calculation module estimates the local conditional entropy under each constraint category in the early and later stages by using the K-nearest neighbor entropy estimation algorithm. The global conditional entropy is obtained by weighted summation. Then, the aesthetic education immersion efficacy index is calculated based on the entropy reduction rate to quantify the changes in students' artistic conception innovation ability under the constraints of metrical rules. The evaluation level output module is used to compare the effectiveness index with a preset threshold, automatically determine the immersion level, and output the evaluation results to the teaching management terminal to provide data support for personalized teaching.

[0007] Further, step S1 includes the following steps: Step S11: Set the initial time window as the data set consisting of the first 4 creations of each student within a complete teaching cycle, and set the later time window as the data set consisting of the last 4 creations of each student within the same teaching cycle. Step S12: The embedding constraint tag is used to identify the list of specified characters required to be embedded in the second line in each creation task, as well as the specific position information of each specified character in the second line sentence.

[0008] Furthermore, step S2 includes the following steps: Step S21: Using an embedding layer, map the tone category to a tone embedding vector, map the part-of-speech category to a part-of-speech embedding vector, and map the position mask of the embedded character in the sentence to a position embedding vector. Concatenate the above embedding vectors to form a vector with dimension [missing information]. Regulation constraint vector ; Step S22: Extract sentence-level semantic vectors of the following text using the RoBERTa pre-trained language model. The semantic vector dimension is denoted as .

[0009] Furthermore, in step S3, the network structure of the constraint decoupling conditional variational autoencoder includes the following sub-networks: Positive A posteriori network Only semantic vectors As input, output latent variables of mood / context. mean Sum of logarithmic variance ; Metrical A posteriori network semantic vectors and metrical constraint vector As input, output implicit metrical variables mean Sum of logarithmic variance ; Metrical A priori network Only using metrical constraint vectors As input, output implicit metrical variables prior mean Sum of logarithmic variance ; Generative Networks Implicit variables in artistic conception With prosody hidden variables Semantic vectors are reconstructed after concatenation and decoding. .

[0010] Furthermore, in step S3, to simultaneously achieve semantic reconstruction, mood decoupling, and separation of prosody information, a decoupling variational loss function is used to train the constrained decoupling conditional variational autoencoder end-to-end. The loss function... The formula is as follows: ; The loss function is composed of the following terms: First item To reconstruct the loss term, mean squared error is used in the implementation. The alternative is the negative log-likelihood; this term enables generative networks to extract latent variables from the context. With prosody hidden variables Accurately restore the original semantic vector ; Second item For decoupling the artistic conception, the regularization term is used, where The prior distribution is a standard normal distribution. Denotes KL divergence, a hyperparameter. This term controls the approximation strength between the posterior distribution of the latent mood variables and the uninformative prior distribution; it is used to remove the latent mood variables. The structured information contained therein related to metrical constraints allows the latent variables of artistic conception to encode only pure content innovation features; Third item Encode regular terms for metrical patterns, where It is composed of a metrical prior network and metrical constraint vectors The prior distribution of the conditional output, hyperparameters This term controls the approximation strength between the posterior distribution of the latent metrical variables and the conditional prior distribution; it is used to encode the metrical constraint information into the latent metrical variables. In this process, the elements of artistic conception and metrical structure are decoupled.

[0011] Furthermore, in step S4, the extraction of latent variables of artistic conception and the classification of prosody constraints include the following steps: Step S41: Create the semantic vector of the next line for each creation in the initial and later windows. Input the values ​​into the trained posterior network of mood and take the posterior mean. As a latent variable in the artistic conception of corresponding creations The dimension of the latent variable of artistic conception is denoted as ; Step S42: Cluster all created metrical constraint vectors, grouping similar metrical constraint vectors into the same metrical constraint category, and labeling the category as... All categories constitute a set ; Step S43: Summarize the results according to the initial window and the later window to form a sample set of latent variables of artistic conception under each metrical constraint category in each window.

[0012] Furthermore, in step S5, the calculation process of the aesthetic education immersion effectiveness index includes the following steps: Step S51: To quantify the uncertainty of students' expression of artistic conception under specific metrical constraints, the K-nearest neighbor entropy estimation algorithm is used to calculate the local conditional entropy under each constraint category; for each metrical constraint category... Let the initial window contain the latent variable sample set of mood under this category. The initial local conditional entropy is estimated using the following formula: ; In the formula, For the dimension of hidden variables in artistic conception, Category within the initial window The number of samples, For the first The sample to the first The Euclidean distance between the nearest neighbors, with constant terms canceled out in the difference calculation between subsequent stages; The same method is used to calculate the same metrical constraint category within the later window. Local conditional entropy ; Step S52: To obtain a comprehensive measure reflecting the overall uncertainty of each constraint category, the local conditional entropy is weighted and summed using the proportion of each category's sample size to the total number of samples in its respective stage. This yields the initial global conditional entropy and the later global conditional entropy. The weighting formula for the global conditional entropy is as follows: ; In the formula, For the current stage of metrical constraints The number of samples, This represents the total number of samples in the current stage. This is the estimated value of the local conditional entropy for the corresponding stage in step S51; Step S53: Based on the relationship between the changes in global conditional entropy in the initial and later stages, define the aesthetic education immersion efficacy index. The aesthetic education immersion efficacy index reflects the relative decrease in the uncertainty of students' expression of artistic conception under the same metrical constraints, thereby quantifying the effect of aesthetic education immersion on improving innovation ability under constraints. The formula is as follows: ; In the formula, This represents the absolute decrease in global conditional entropy. As the initial global conditional entropy baseline, the ratio form makes the aesthetic education immersion effectiveness index comparable among different students and with different constraint complexities.

[0013] Furthermore, in step S6, the rule for generating the evaluation level is as follows: When the effectiveness index of aesthetic education immersion When the value is greater than or equal to the first preset threshold, it is determined to be highly immersive; When the effectiveness index of aesthetic education immersion If the threshold value is greater than or equal to the second preset threshold and less than the first preset threshold, it is determined to be moderate infiltration. When the effectiveness index of aesthetic education immersion If the value is less than the second preset threshold, it is determined to be insufficient immersion.

[0014] The present invention has the following beneficial effects: 1. This invention decouples creative behavior into prosody factors and mood factors through a conditional variational autoencoder. It uses the conditional entropy of the latent mood variables under the same constraints to measure their uncertainty. From the perspective of information theory, it objectively quantifies the degree of improvement of the composite ability to create mood under prosody constraints. The evaluation process does not rely on subjective judgment, and the results are reproducible and comparable, providing a reliable technical means for the accurate measurement of the effectiveness of aesthetic education.

[0015] 2. This invention forces the model to encode metrical information into specific latent variables, making the latent variables of artistic conception independent of metrical conditions, extracting pure innovative features that are not affected by constrained forms, thereby accurately reflecting the real impact of aesthetic education on aesthetic creativity and avoiding evaluation bias caused by different metrical difficulties.

[0016] 3. This invention defines an efficacy index based on the ratio of conditional entropy between preceding and subsequent stages, which can dynamically track the evolution trend of aesthetic education immersion. Without the need for pre-setting complex scoring rubrics, it can automatically calculate the relative decrease in the uncertainty of artistic conception during the immersion process using only natural data from students' multiple creations. It is highly sensitive and possesses normalization characteristics, enabling horizontal comparisons across students and tasks, providing continuous and stable quantitative feedback for personalized aesthetic education intervention.

[0017] 4. The evaluation process of this invention is fully automated and requires no manual annotation. From the extraction of prosody features, semantic encoding, model training to entropy calculation and grade determination, all are completed autonomously by the algorithm, which significantly reduces the human cost and time consumption of the evaluation. It is suitable for large-scale deployment on online poetry and couplet cultural and creative education platforms and can achieve normalized monitoring of aesthetic education effectiveness without increasing the burden on teachers.

[0018] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

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

[0020] Figure 1 This is a schematic diagram of the framework of a machine learning-based evaluation system for the immersion effectiveness of poetry and couplet-based cultural and creative aesthetic education in basic education according to the present invention. Figure 2 This is a flowchart illustrating the effectiveness evaluation system for the immersion of poetry, couplets, and cultural creative arts education in basic education based on machine learning, according to the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figure 1-2 As shown, the present invention is a machine learning-based evaluation system for the immersion effect of poetry and couplet cultural and creative aesthetic education in basic education, including a data acquisition and window division module, a text feature extraction module, a CD-CVAE training module, an imagery latent variable extraction and constrained clustering module, a conditional entropy estimation and efficacy index calculation module, and an evaluation level output module. The workflow of the evaluation system is as follows: Step S1: Collect the text data of the second line of students in multiple rounds of the embedded character couplet creation activity, record the timestamp and embedded character constraint label, and divide the creation sequence of each student into the initial time window and the later time window according to the teaching cycle. Step S2: Automatically extract the tonal sequence, part-of-speech sequence, and embedded character position mask for each second line of the poem, encode them into a metrical constraint vector, and at the same time use a pre-trained language model to extract the sentence-level semantic vector of the second line; Step S3: Construct a constraint-decoupling conditional variational autoencoder. The variational autoencoder includes an artistic conception posterior network, a metrical posterior network, a metrical prior network, and a generative network. It is trained using a loss function composed of reconstruction loss, artistic conception decoupling regularization term, and metrical encoding regularization term. The expression patterns related to metrical rules in the second couplet are encoded into the metrical latent variable, and the artistic conception innovation that is free from metrical constraints is separated into the artistic conception latent variable. Step S4: Fix the encoder parameters after training, input the semantic vectors of the lower couplet created by each student in the initial window and the later window into the posterior network of artistic conception, take the posterior mean as the latent variable of artistic conception, and divide the metrical constraint categories according to the similarity of the metrical constraint vectors to form the latent variable sample set of artistic conception under each constraint category in each window. Step S5: Use the K-nearest neighbor entropy estimation algorithm to calculate the local conditional entropy of the initial window and the later window under each metrical constraint category, and then use the sample proportion of each category as the weight to sum the global conditional entropy. Calculate the aesthetic education immersion effectiveness index based on the difference ratio between the initial global conditional entropy and the later global conditional entropy. Step S6: Based on the comparison results between the aesthetic education immersion effectiveness index and the preset threshold, automatically generate and output the evaluation level.

[0023] The data acquisition and window partitioning module, text feature extraction module, mood latent variable extraction and constrained clustering module, conditional entropy estimation and performance index calculation module, and evaluation level output module are connected in sequence. The output end of the text feature extraction module is unidirectionally connected to the input end of the CD-CVAE training module, and the output end of the CD-CVAE training module is unidirectionally connected to the input end of the mood latent variable extraction and constrained clustering module. The data collection and window division module is used to automatically collect the second line text submitted by students in multiple rounds of the embedded character couplet creation activity, record the timestamp and embedded character constraint tags, and divide each student's creation sequence into an initial time window and a later time window according to the teaching cycle, providing structured data for subsequent time series comparison analysis. The text feature extraction module is used to perform metrical analysis and semantic encoding on each second line of the couplet, outputting metrical constraint vectors and sentence-level semantic vectors, transforming the unstructured couplet text into a numerical feature representation that can be processed by machine learning models; The CD-CVAE training module is used to build and train a constraint-decoupled conditional variational autoencoder. Through joint optimization of reconstruction loss, mood decoupling regularization term and metrical encoding regularization term, it automatically learns to decompose the second line of the work into mood latent variables and metrical latent variables, thereby achieving decoupling of mood innovation features and metrical constraint features. The latent variable extraction and constraint clustering module is used to fix the encoder parameters after training, extract the latent variables of the artistic conception created by each student in the initial and later windows, and cluster them according to the similarity of the metrical constraint vectors to form a sample set of latent variables of artistic conception under each window and constraint category for use in conditional entropy calculation. The module for estimating conditional entropy and calculating efficiency index estimates the local conditional entropy under each constraint category in the early and later stages by using the K-nearest neighbor entropy estimation algorithm. The global conditional entropy is obtained by weighted summation. Then, the efficiency index of aesthetic education immersion is calculated based on the entropy reduction rate to quantify the changes in students’ ability to create artistic conception under the constraints of metrical rules. The assessment level output module compares the effectiveness index with preset thresholds, automatically determines the immersion level, and outputs the assessment results to the teaching management terminal to provide data support for personalized teaching.

[0024] Step S1 includes the following steps: Step S11: Set the initial time window as the data set consisting of the first 4 creations of each student within a complete teaching cycle, and set the later time window as the data set consisting of the last 4 creations of each student within the same teaching cycle. Step S12: The embedding constraint tag is used to identify the list of specified characters required to be embedded in the second line in each creation task, as well as the specific position information of each specified character in the second line sentence.

[0025] Step S2 includes the following steps: Step S21: Using an embedding layer, map the tone category to a tone embedding vector, map the part-of-speech category to a part-of-speech embedding vector, and map the position mask of the embedded character in the sentence to a position embedding vector. Concatenate the above embedding vectors to form a vector with dimension [missing information]. Regulation constraint vector ; Step S22: Extract sentence-level semantic vectors of the following text using the RoBERTa pre-trained language model. The semantic vector dimension is denoted as .

[0026] In step S3, the network structure of the constraint-decoupling conditional variational autoencoder includes the following sub-networks: Positive A posteriori network Only semantic vectors As input, output latent variables of mood / context. mean and logarithmic variance ; Metrical A posteriori network semantic vectors and metrical constraint vector As input, output implicit metrical variables mean and logarithmic variance ; Metrical A priori network Only using metrical constraint vectors As input, output implicit metrical variables prior mean and logarithmic variance ; Generative Networks Implicit variables in artistic conception With prosody hidden variables Semantic vectors are reconstructed after concatenation and decoding. .

[0027] In step S3, to simultaneously achieve semantic reconstruction, mood decoupling, and separation of prosody information, a decoupling variational loss function is used to train the constrained decoupling conditional variational autoencoder end-to-end. The loss function... The formula is as follows: ; The loss function is composed of the following terms: First item To reconstruct the loss term, mean squared error is used in the implementation. The alternative is the negative log-likelihood; this term enables generative networks to extract latent variables from the context. With prosody hidden variables Accurately restore the original semantic vector ; Second item For decoupling the artistic conception, the regularization term is used, where The prior distribution is a standard normal distribution. Denotes KL divergence, a hyperparameter. This term controls the approximation strength between the posterior distribution of the latent mood variables and the uninformative prior distribution; it is used to remove the latent mood variables. The structured information contained therein related to metrical constraints allows the latent variables of artistic conception to encode only pure content innovation features; Third item Encode regular terms for metrical patterns, where It is composed of a metrical prior network and metrical constraint vectors The prior distribution of the conditional output, hyperparameters This term controls the approximation strength between the posterior distribution of the latent metrical variables and the conditional prior distribution; it is used to encode the metrical constraint information into the latent metrical variables. In this process, the elements of artistic conception and metrical structure are decoupled.

[0028] In step S4, the extraction of latent variables of artistic conception and the classification of prosody constraints include the following steps: Step S41: Create the semantic vector of the next line for each creation in the initial and later windows. Input the values ​​into the trained posterior network of mood and take the posterior mean. As a latent variable in the artistic conception of corresponding creations The dimension of the latent variable of artistic conception is denoted as ; Step S42: Cluster all created metrical constraint vectors, grouping similar metrical constraint vectors into the same metrical constraint category, and labeling the category as... All categories constitute a set ; Step S43: Summarize the results according to the initial window and the later window to form a sample set of latent variables of artistic conception under each metrical constraint category in each window.

[0029] In step S5, the calculation process of the aesthetic education immersion effectiveness index includes the following steps: Step S51: To quantify the uncertainty of students' expression of artistic conception under specific metrical constraints, the K-nearest neighbor entropy estimation algorithm is used to calculate the local conditional entropy under each constraint category; for each metrical constraint category... Let the initial window contain the latent variable sample set of mood under this category. The initial local conditional entropy is estimated using the following formula: ; In the formula, For the dimension of hidden variables in artistic conception, Category within the initial window The number of samples, For the first The sample to the first The Euclidean distance between the nearest neighbors, with constant terms canceled out in the difference calculation between subsequent stages; The same method is used to calculate the same metrical constraint category within the later window. Local conditional entropy ; Step S52: To obtain a comprehensive measure reflecting the overall uncertainty of each constraint category, the local conditional entropy is weighted and summed using the proportion of each category's sample size to the total number of samples in its respective stage. This yields the initial global conditional entropy and the later global conditional entropy. The weighting formula for the global conditional entropy is as follows: ; In the formula, For the current stage of metrical constraints The number of samples, This represents the total number of samples in the current stage. This is the estimated value of the local conditional entropy for the corresponding stage in step S51; Step S53: Based on the relationship between the changes in global conditional entropy in the initial and later stages, define the aesthetic education immersion efficacy index. The aesthetic education immersion efficacy index reflects the relative decrease in the uncertainty of students' expression of artistic conception under the same metrical constraints, thereby quantifying the effect of aesthetic education immersion on improving innovation ability under constraints. The formula is as follows: ; In the formula, This represents the absolute decrease in global conditional entropy. As the initial global conditional entropy baseline, the ratio form makes the aesthetic education immersion effectiveness index comparable among different students and with different constraint complexities.

[0030] In step S6, the rules for generating the evaluation level are as follows: When the effectiveness index of aesthetic education immersion When the value is greater than or equal to the first preset threshold, it is determined to be highly immersive; When the effectiveness index of aesthetic education immersion If the threshold value is greater than or equal to the second preset threshold and less than the first preset threshold, it is determined to be moderate infiltration. When the effectiveness index of aesthetic education immersion If the value is less than the second preset threshold, it is determined to be insufficient immersion.

[0031] One specific application of this embodiment is: This embodiment uses a creative arts education course on embedded-character couplets, taught in two seventh-grade classes with a total of 86 students, as the application scenario. The teaching cycle is 8 weeks, with one embedded-character couplet creation activity conducted each week. Each activity requires students to create a 7-character second line based on a given first line and embedded character constraints. The specific implementation process is as follows: Step S1: Data Acquisition and Time Window Division The system automatically collected 688 valid second-line texts submitted by 86 students in 8 creative activities through the online teaching platform. Each data entry included a unique student identifier, a creation timestamp, and an embedding constraint tag. The embedding constraint tag clearly indicated the specified characters and corresponding positions required to be embedded in the second line for each task.

[0032] The creative sequence for each student is divided into two time windows according to the teaching cycle: Initial time window The data set consisting of each student's 1st-4th creations, totaling... Sample; Later Time Window The data set consisting of each student's 5th-8th creations, totaling... Sample; Step S2: Extracting features from the following text For each line of the couplet, extract metrical constraint features and semantic features, and transform the unstructured text into a numerical vector representation: Step S21: Encoding the Regulation Constraint Vector For the 7-character second line, three types of metrical features are defined and mapped to dense vectors through an embedding layer: Tonal characteristics: Map the three tonal categories (level and oblique tones) into a 16-dimensional tonal embedding vector; Part-of-speech features: Mapping 12 categories of Chinese parts of speech, including nouns, verbs, and adjectives, into 32-dimensional part-of-speech embedding vectors; Embedded character position features: A 7-dimensional one-hot mask is used to identify the embedded character position, which is mapped to a 16-dimensional position embedding vector; Concatenating the above three types of embedding vectors yields a vector with dimension [dimensional value missing]. Regulation constraint vector .

[0033] Step S22: Sentence-level semantic vector extraction The Chinese RoBERTa-base pre-trained language model is used to semantically encode the second part of the text, and the average output value of the tokens is extracted as the sentence-level semantic vector. Semantic vector dimension .

[0034] Step S3: Construction and training of the constraint-decoupled conditional variational autoencoder (CD-CVAE) A CD-CVAE model containing four sub-networks is constructed, and the decoupling of artistic conception and metrical features is achieved through a joint loss function: Network structure parameter settings Positive A posteriori network : A 2-layer fully connected network with an input dimension of 768, a hidden layer dimension of 256, and an output of a mood / conceptual hidden variable. mean Sum of logarithmic variance Artistic conception, hidden variable dimension ; Metrical A posteriori network : 2-layer fully connected network, input dimension Hidden layer dimension 256, output metrical hidden variables mean Sum of logarithmic variance The implicit variable dimension of prosody is 64; Metrical A priori network A 2-layer fully connected network with an input dimension of 64 and a hidden layer dimension of 128, outputting the mean of the prior distribution of the latent variable of metrical pattern. Sum of logarithmic variance ; Generative Networks : 2-layer fully connected network, input dimension The hidden layer has a dimension of 512, and the output is a 768-dimensional reconstructed semantic vector. ; Model training configuration Training dataset: Metrical constraint vectors and semantic vectors of all 688 couplet texts Optimizer: AdamW, Learning Rate Weight decay ; Batch size: Training rounds ; Loss function hyperparameters: , ; Decoupling variational loss function End-to-end training is performed using the following loss function: ; Wherein: the reconstruction loss term is implemented using mean squared error: ; In the decoupling regularization term of artistic conception, It follows a standard normal prior distribution; In the regular expression for metrical encoding, The conditional normal distribution is the output of the metrical prior network; After training, the model is able to encode the metrical constraints of the second line of the poem into a centralized encoding. Encoding pure artistic conception and innovative information into This achieves complete decoupling of the two types of features.

[0035] Step S4: Extraction of latent variables for mood and clustering based on prosody constraints With the encoder parameters fixed after training, perform the following operations: Step S41: Extraction of latent variables for mood Will and semantic vectors of all lower couplets within the window Input the values ​​into the posterior network of the mood and take the posterior mean. As a latent variable in the artistic conception of corresponding creations This yields a set of latent variables for the mood with a dimension of 128.

[0036] Step S42: Classification of Metrical Constraints Temperament constraint vector for all 688 samples Perform K-Means clustering, number of clusters (Corresponding to 5 common constraint types: single-character embedding in 1-3 positions, single-character embedding in 4-7 positions, adjacent double-character embedding, spaced double-character embedding, and triple-character embedding), resulting in a set of metrical constraint categories. .

[0037] Step S43: Sample set construction The samples are aggregated according to time window and metrical constraint category to form a latent variable sample set of artistic conception under each window and category. In this embodiment, the sample quantity distribution of each window and category is as follows: Metrical Constraint Category In the initial time window The sample size was 68, and the later time window The sample size is 72; categories exist The sample size is 76. The sample size is 74; categories exist The sample size is 82. The sample size is 80; categories exist The sample size is 58. The sample size is 56; categories exist The sample size is 60. The sample size was 62; the total sample size for both windows was 344.

[0038] Step S5: Conditional Entropy Estimation and Calculation of Aesthetic Education Immersion Effectiveness Index The K-nearest neighbor entropy estimation algorithm is adopted. Quantify the uncertainty of the expression of artistic conception, and then calculate the effectiveness index: Step S51: Calculation of local conditional entropy For each category of metrical constraints Calculate separately and Local conditional entropy under the window. Window Taking categories as an example, the formula for estimating local conditional entropy is: ; In the formula, For the first The Euclidean distance from each latent variable sample to its 5th nearest neighbor, with a constant term. They cancel each other out in subsequent difference calculations.

[0039] Calculate using the same method The local conditional entropy for various categories under the window is calculated as follows in this embodiment: Regulation Constraint Category The initial local conditional entropy is 4.52, and the later local conditional entropy is 2.98; category The initial local conditional entropy is 4.37, and the later local conditional entropy is 2.85; category The initial local conditional entropy is 4.19, and the later local conditional entropy is 2.71; category The initial local conditional entropy is 4.68, and the later local conditional entropy is 3.02; category The initial local conditional entropy is 4.05, and the later local conditional entropy is 2.63.

[0040] Step S52: Calculate global conditional entropy The global conditional entropy is obtained by weighting the proportion of each category of samples to the total number of samples in its window. ; in This represents the total number of samples in a single window. Substituting the data, we calculate the initial global conditional entropy. Later global conditional entropy .

[0041] Step S53: Calculation of the Aesthetic Education Immersion Effectiveness Index The efficiency index is calculated based on the relative decrease in global conditional entropy: ; Substituting the values, we get: ; Step S6: Evaluate the level output Preset evaluation level threshold: First threshold Second threshold The infiltration level is determined based on the comparison between the effectiveness index and the threshold: when... When, it is judged as highly infiltrated; when At that time, it was determined to be moderate infiltration; when At that time, it was determined to be insufficient immersion.

[0042] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0043] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A machine learning-based evaluation system for assessing the effectiveness of poetry, couplets, and cultural creative aesthetic education in basic education, characterized in that: It includes a data acquisition and window partitioning module, a text feature extraction module, a CD-CVAE training module, a contextual latent variable extraction and constrained clustering module, a conditional entropy estimation and performance index calculation module, and an evaluation level output module; The workflow of the evaluation system is as follows: Step S1: Collect the text data of the second line of students in multiple rounds of the embedded character couplet creation activity, record the timestamp and embedded character constraint label, and divide the creation sequence of each student into the initial time window and the later time window according to the teaching cycle. Step S2: Automatically extract the tonal sequence, part-of-speech sequence, and embedded character position mask for each second line of the poem, encode them into a metrical constraint vector, and at the same time use a pre-trained language model to extract the sentence-level semantic vector of the second line; Step S3: Construct a constraint-decoupling conditional variational autoencoder. The variational autoencoder includes an artistic conception posterior network, a metrical posterior network, a metrical prior network, and a generative network. It is trained using a loss function composed of reconstruction loss, artistic conception decoupling regularization term, and metrical encoding regularization term. The expression patterns related to metrical rules in the second couplet are encoded into the metrical latent variable, and the artistic conception innovation that is free from metrical constraints is separated into the artistic conception latent variable. Step S4: Fix the encoder parameters after training, input the semantic vectors of the lower couplet created by each student in the initial window and the later window into the posterior network of artistic conception, take the posterior mean as the latent variable of artistic conception, and divide the metrical constraint categories according to the similarity of the metrical constraint vectors to form the latent variable sample set of artistic conception under each constraint category in each window. Step S5: Use the K-nearest neighbor entropy estimation algorithm to calculate the local conditional entropy of the initial window and the later window under each metrical constraint category, and then use the sample proportion of each category as the weight to sum the global conditional entropy. Calculate the aesthetic education immersion effectiveness index based on the difference ratio between the initial global conditional entropy and the later global conditional entropy. Step S6: Based on the comparison results between the aesthetic education immersion effectiveness index and the preset threshold, automatically generate and output the evaluation level.

2. The machine learning-based evaluation system for the immersion effectiveness of poetry, couplets, and cultural creative aesthetic education in basic education, as described in claim 1, is characterized in that... The data acquisition and window partitioning module, text feature extraction module, mood latent variable extraction and constrained clustering module, conditional entropy estimation and efficiency index calculation module, and evaluation level output module are connected in sequence. The output end of the text feature extraction module is unidirectionally connected to the input end of the CD-CVAE training module, and the output end of the CD-CVAE training module is unidirectionally connected to the input end of the mood latent variable extraction and constrained clustering module. The data acquisition and window division module is used to automatically collect the second line text submitted by students in multiple rounds of the embedded character couplet creation activity, record the timestamp and embedded character constraint label, and divide the creation sequence of each student into an initial time window and a later time window according to the teaching cycle. The text feature extraction module is used to perform metrical analysis and semantic encoding on each second line of the couplet, outputting metrical constraint vectors and sentence-level semantic vectors, and transforming the unstructured couplet text into a numerical feature representation that can be processed by machine learning models. The CD-CVAE training module is used to construct and train a constraint-decoupled conditional variational autoencoder. Through joint optimization of reconstruction loss, mood decoupling regularization term and metrical encoding regularization term, it automatically learns to decompose the second line of the work into mood latent variables and metrical latent variables, thereby achieving decoupling of mood innovation features and metrical constraint features. The latent variable extraction and constraint clustering module is used to fix the encoder parameters after training, extract the latent variables of the artistic conception created by each student in the initial and later windows, and cluster them according to the similarity of the metrical constraint vectors to form a sample set of latent variables of artistic conception under each window and each constraint category for use in conditional entropy calculation. The conditional entropy estimation and efficacy index calculation module estimates the local conditional entropy under each constraint category in the early and later stages by using the K-nearest neighbor entropy estimation algorithm. The global conditional entropy is obtained by weighted summation. Then, the aesthetic education immersion efficacy index is calculated based on the entropy reduction rate to quantify the changes in students' artistic conception innovation ability under the constraints of metrical rules. The evaluation level output module is used to compare the effectiveness index with a preset threshold, automatically determine the immersion level, and output the evaluation results to the teaching management terminal to provide data support for personalized teaching.

3. The machine learning-based evaluation system for the immersion effectiveness of poetry, couplets, and cultural creative aesthetic education in basic education, as described in claim 1, is characterized in that... Step S1 includes the following steps: Step S11: Set the initial time window as the data set consisting of the first 4 creations of each student within a complete teaching cycle, and set the later time window as the data set consisting of the last 4 creations of each student within the same teaching cycle. Step S12: The embedding constraint tag is used to identify the list of specified characters required to be embedded in the second line in each creation task, as well as the specific position information of each specified character in the second line sentence.

4. The machine learning-based evaluation system for the immersion effectiveness of poetry, couplets, and cultural creative aesthetic education in basic education, as described in claim 1, is characterized in that... Step S2 includes the following steps: Step S21: Using an embedding layer, map the tone category to a tone embedding vector, map the part-of-speech category to a part-of-speech embedding vector, and map the position mask of the embedded character in the sentence to a position embedding vector. Concatenate the above embedding vectors to form a vector with dimension [missing information]. Regulation constraint vector ; Step S22: Extract sentence-level semantic vectors of the following text using the RoBERTa pre-trained language model. The semantic vector dimension is denoted as .

5. The machine learning-based evaluation system for the immersion effectiveness of poetry, couplets, and cultural creative aesthetic education in basic education, as described in claim 1, is characterized in that... In step S3, the network structure of the constraint decoupling conditional variational autoencoder includes the following sub-networks: Positive A posteriori network Only semantic vectors As input, output latent variables of mood / context. mean Sum of logarithmic variance ; Metrical A posteriori network semantic vectors and metrical constraint vector As input, output implicit metrical variables mean Sum of logarithmic variance ; Metrical A priori network Only using metrical constraint vectors As input, output implicit metrical variables prior mean Sum of logarithmic variance ; Generative Networks Implicit variables in artistic conception With prosody hidden variables Semantic vectors are reconstructed after concatenation and decoding. .

6. The machine learning-based evaluation system for the immersion effectiveness of poetry, couplets, and cultural creative aesthetic education in basic education, as described in claim 1, is characterized in that... In step S3, to simultaneously achieve semantic reconstruction, mood decoupling, and separation of prosody information, a decoupling variational loss function is used to train the constrained decoupling conditional variational autoencoder end-to-end. The loss function... The formula is as follows: The loss function is composed of the following terms: First item To reconstruct the loss term, mean squared error is used in the implementation. Alternating negative log-likelihood; this term enables generative networks to extract latent variables from context. With prosody hidden variables Accurately restore the original semantic vector ; Second item For decoupling the artistic conception, the regularization term is used, where The prior distribution is a standard normal distribution. Denotes KL divergence, a hyperparameter. This term controls the approximation strength between the posterior distribution of the latent mood variables and the uninformative prior distribution; it is used to remove the latent mood variables. The structured information contained therein related to metrical constraints allows the latent variables of artistic conception to encode only pure content innovation features; Third item Encode the regular term for the metrical pattern, where It is composed of a metrical prior network and metrical constraint vectors The prior distribution of the conditional output, hyperparameters This term controls the approximation strength between the posterior distribution of the latent metrical variables and the conditional prior distribution; it is used to encode the metrical constraint information into the latent metrical variables. In this process, the elements of artistic conception and metrical structure are decoupled.

7. The machine learning-based evaluation system for the immersion effectiveness of poetry, couplets, and cultural creative aesthetic education in basic education, as described in claim 1, is characterized in that... In step S4, the extraction of latent variables of artistic conception and the classification of metrical constraints include the following steps: Step S41: Create the semantic vector of the next line for each creation in the initial and later windows. Input the values ​​into the trained posterior network of mood and take the posterior mean. As a latent variable in the artistic conception of corresponding creations The dimension of the latent variable of artistic conception is denoted as ; Step S42: Cluster all created metrical constraint vectors, grouping similar metrical constraint vectors into the same metrical constraint category, and labeling the category as... All categories constitute a set ; Step S43: Summarize the results according to the initial window and the later window to form a sample set of latent variables of artistic conception under each metrical constraint category in each window.

8. The machine learning-based evaluation system for the immersion effectiveness of poetry, couplets, and cultural creative aesthetic education in basic education, as described in claim 1, is characterized in that... In step S5, the calculation process of the aesthetic education immersion effectiveness index includes the following steps: Step S51: To quantify the uncertainty of students' expression of artistic conception under specific metrical constraints, the K-nearest neighbor entropy estimation algorithm is used to calculate the local conditional entropy under each constraint category; for each metrical constraint category... Let the initial window contain the latent variable sample set of mood under this category. The initial local conditional entropy is estimated using the following formula: In the formula, For the dimension of hidden variables in artistic conception, Category within the initial window The number of samples, For the first The sample to the first The Euclidean distance between the nearest neighbors, with constant terms canceled out in the difference calculation between subsequent stages; The same method is used to calculate the same metrical constraint category within the later window. Local conditional entropy ; Step S52: To obtain a comprehensive measure reflecting the overall uncertainty of each constraint category, the local conditional entropy is weighted and summed using the proportion of each category's sample size to the total number of samples in its respective stage. This yields the initial global conditional entropy and the later global conditional entropy. The weighting formula for the global conditional entropy is as follows: In the formula, For the current stage of metrical constraints The number of samples, This represents the total number of samples in the current stage. This is the estimated value of the local conditional entropy for the corresponding stage in step S51; Step S53: Based on the relationship between the changes in global conditional entropy in the initial and later stages, define the aesthetic education immersion efficacy index. The aesthetic education immersion efficacy index reflects the relative decrease in the uncertainty of students' expression of artistic conception under the same metrical constraints, thereby quantifying the effect of aesthetic education immersion on improving innovation ability under constraints. The formula is as follows: In the formula, This represents the absolute decrease in global conditional entropy. As the initial global conditional entropy baseline, the ratio form makes the aesthetic education immersion effectiveness index comparable among different students and with different constraint complexities.

9. A machine learning-based evaluation system for the immersion effectiveness of poetry, couplets, and cultural creative aesthetic education in basic education, as described in claim 1, is characterized in that... In step S6, the rule for generating the evaluation level is as follows: When the effectiveness index of aesthetic education immersion When the value is greater than or equal to the first preset threshold, it is determined to be highly immersive; When the effectiveness index of aesthetic education immersion If the threshold value is greater than or equal to the second preset threshold and less than the first preset threshold, it is determined to be moderate infiltration. When the effectiveness index of aesthetic education immersion If the value is less than the second preset threshold, it is determined to be insufficient immersion.

Images