Child-oriented text grading method and system based on artificial intelligence

Abstract

The invention discloses a child-oriented text grading method and system based on artificial intelligence, and the method comprises the steps: carrying out the preprocessing of a to-be-evaluated text content, and generating a standardized text representation; extracting multi-dimensional feature vectors from the standardized text representation by utilizing a pre-training model, wherein the multi-dimensional feature vectors comprise theme, emotion and language complexity features; generating a content suitability score through a multi-layer perceptron based on the multi-dimensional feature vector; and using a multi-classification model to predict age classification according to the suitability score and the multi-dimensional feature vector. Compared with a traditional method, the method has the advantages that obscure information can be recognized, the safety and examination efficiency of child content are improved through automatic examination, and the method has remarkable innovativeness and social benefits.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence, and particularly relates to a text grading method and system for children based on artificial intelligence. Background Art

[0002] With the development of LLM technology, children's education and entertainment content are gradually dominated by personalized texts generated in real time, such as instantaneously created children's stories or interactive dialogues. However, traditional review methods are completely unable to adapt to this change. Manual review is inefficient and cannot handle the massive amounts of data generated in real time; while keyword filtering, due to the lack of context understanding ability, cannot identify implicit harmful information or judge the suitability of complex educational content. Therefore, the existing technology cannot protect children from potential risks, and there is an urgent need for innovative solutions. The traditional keyword filtering scheme has the following significant defects: Insufficient efficiency and coverage: Traditional manual review takes a long time, cannot handle the massive amounts of data generated in real time, and has a limited coverage, which can no longer meet the growing needs.

[0003] Consistency issues: Manual evaluation, due to differences in reviewers' experience and subjective biases, leads to inconsistent standards and it is difficult to ensure the reliability of the results.

[0004] Insufficient depth of analysis: Existing automated tools mostly rely on simple keyword matching and may overlook implicit harmful information. For example, the content may not contain explicit violation words, but convey negative emotions or values that do not conform to social norms through metaphor or narrative, having a potential negative impact on children's psychological development and behavior habits.

[0005] Misjudgment of complex content: For complex educational content suitable for children (such as science readings), keyword filtering may misjudge it as inappropriate due to the high language complexity, limiting children's access to beneficial information. Summary of the Invention

[0006] In view of the above problems, the present invention provides a text grading method and system for children based on artificial intelligence, aiming to provide an intelligent system that can comprehensively analyze the theme, sentiment tendency, language complexity and potential value impact of the text, and accurately grade according to the children's age group, so as to improve the safety, age-appropriateness and positive guidance of the content received by children.

[0007] According to the first aspect of the embodiments of the present disclosure, there is provided a text grading method for children based on artificial intelligence, the method comprising the following steps: Preprocess the text content to be evaluated to generate a standardized text representation, the preprocessing including text tokenization, stop word filtering and lemmatization; Generate the semantic representation of the text using a pre-trained model for the standardized text representation, and extract multi-dimensional feature vectors based on the semantic representation. The multi-dimensional feature vectors include topic classification, sentiment tendency, and language complexity features. Specifically, extracting multi-dimensional feature vectors from the standardized text representation using a pre-trained model includes: based on the input of the standardized text representation, using the pre-trained BERT model to obtain the multi-dimensional vector of the CLS token and obtain the topic classification feature vector; based on the input of the standardized text representation, using the pre-trained BERT model to encode, generate multi-dimensional semantic representations for each token to obtain the sentiment tendency confidence; based on the input of the standardized text representation, using the pre-trained BERT model to encode, generate multi-dimensional semantics for each token, add a bi-affine attention layer, predict the dependency relationship between each token, and quantify the language complexity to obtain the language complexity feature; Input the multi-dimensional feature vectors into a multi-layer perceptron. The multi-layer perceptron processes the non-linear mapping of the multi-dimensional feature vectors through the Sigmoid function to generate a content suitability score ranging from 0 to 100; Normalize the content suitability score, fuse it with the multi-dimensional feature vectors, and then input it into a Softmax classifier. Finally, the trained Softmax classifier outputs the prediction result of the age-appropriate classification label.

[0008] In some embodiments, extracting multi-dimensional feature vectors from the standardized text representation using a pre-trained model specifically includes: Based on the input of the standardized text representation, use the pre-trained or BERT model to obtain the 768-dimensional vector of the CLS token, which is used as the feature vector for topic classification and input into a support vector machine for topic classification; Based on the input of the standardized text representation, use the pre-trained BERT model to encode, generate 768-dimensional semantic representations for each token, add a fully connected layer, map the 768-dimensional semantic representations to the probability distribution of sentiment labels, and output the confidence through the softmax function; Based on the input of the standardized text representation, use the pre-trained BERT model to encode, generate 768-dimensional semantic representations for each token, add a bi-affine attention layer, predict the dependency relationship between tokens, generate a syntactic tree, and quantify the language complexity based on the topological analysis and information entropy calculation of the syntactic tree; Perform normalization processing on the extracted feature vectors to ensure the consistency of the dimensions of different features.

[0009] In some embodiments, the multi-dimensional feature vectors further include social value features, which are extracted through the graph embedding technology of the knowledge graph.

[0010] In some embodiments, a content suitability score is generated by a multi-layer perceptron based on a multi-dimensional feature vector, and the specific expression is: , where represents the content suitability score, represents the Sigmoid function, is the weight matrix of the neural network, x represents the multi-dimensional feature vector, represents the bias vector, and ReLU represents the rectified linear unit function.

[0011] In some embodiments, a classification model is used to predict the age classification based on the suitability score and the feature vector, specifically including: Fusing the suitability score and the feature vector into an input vector; Outputting the probability distribution of each age classification through a Softmax classifier; Selecting the age classification with the highest probability as the prediction result.

[0012] In some embodiments, the content suitability scoring process further includes: Pre-training the MLP model and optimizing the parameters using a labeled dataset, where the labeled dataset includes suitable and unsuitable children's text samples.

[0013] In some embodiments, the age classification further includes: Pre-defining the cognitive and psychological characteristics of children's age groups, including vocabulary comprehension ability and emotional acceptance range; Training the Softmax classifier with historical data.

[0014] According to the second aspect of the embodiments of the present disclosure, a text grading system for children based on artificial intelligence is provided, and the system includes: A preprocessing module for preprocessing the text content to be evaluated to generate a standardized text representation, where the preprocessing includes text tokenization, stop word filtering, and word form normalization; A feature extraction module, which is used to generate a semantic representation of the text by using a pre-trained model for the standardized text representation, and extract a multi-dimensional feature vector based on the semantic representation, where the multi-dimensional feature vector includes topic classification, sentiment tendency, and language complexity features; specifically, extracting the multi-dimensional feature vector by using the pre-trained model for the standardized text representation includes: based on the input of the standardized text representation, using the pre-trained BERT model to obtain the multi-dimensional vector of the CLS token and obtain the topic classification feature vector; based on the input of the standardized text representation, using the pre-trained BERT model to encode, generate a multi-dimensional semantic representation for each token to obtain the sentiment tendency confidence; based on the input of the standardized text representation, using the pre-trained BERT model to encode, generate a multi-dimensional semantics for each token, add a bi-affine attention layer, predict the dependency relationship between each token, and quantify the language complexity to obtain the language complexity feature; A content suitability scoring module, which is used to input the multi-dimensional feature vector into a multi-layer perceptron, and the multi-layer perceptron processes the non-linear mapping of the multi-dimensional feature vector through a Sigmoid function to generate a content suitability score from 0 to 100; An age grading module, which is used to normalize the content suitability score, fuse it with the multi-dimensional feature vector, and then input it into a Softmax classifier, and finally the trained Softmax classifier outputs the prediction result of the age-appropriate grading label.

[0015] According to the third aspect of the embodiments of the present disclosure, an electronic device is provided, including a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the program, the steps of the above-mentioned text grading method for children based on artificial intelligence are implemented.

[0016] According to the fourth aspect of the embodiments of the present disclosure, a non-temporary computer-readable storage medium is provided, on which computer instructions are stored, and when the instructions are executed by a processor, the steps of the above-mentioned text grading method for children based on artificial intelligence are implemented.

[0017] A text grading method, system, electronic device, and storage medium for children based on artificial intelligence provided by the embodiments of the present disclosure utilize multi-dimensional feature extraction and deep learning technologies to achieve efficient and accurate automated review. The system includes a text feature extraction, suitability scoring, and age grading module, equipped with a processor and a memory, analyzes the text topic, sentiment, and language complexity through natural language processing and machine learning, generates a suitability score, and predicts the age-appropriate grading. Compared with traditional methods, the present system can identify hidden bad information, improve the security and review efficiency of children's content through automated review, and has significant innovation and social benefits. Specifically, the beneficial effects of the present invention include: Multi-dimensional analysis: By integrating various features such as theme, sentiment, language complexity, and values, it goes beyond traditional keyword matching methods to ensure comprehensive and in-depth evaluation.

[0018] Intelligent identification of implicit information: Through semantic understanding and context analysis, it can identify implicit negative content or negative tendencies, improving the accuracy of review.

[0019] Efficient automation: It realizes real-time evaluation and grading of text content, reduces manual intervention, and improves the review efficiency.

[0020] Precise age grading: By combining the cognitive and psychological needs of children at different ages, it outputs age-appropriate grading results, which is highly targeted.

[0021] Transparent feedback: It provides a detailed evaluation report, facilitating content creators to optimize the text, and at the same time supporting the effective supervision of parents and regulatory agencies.

[0022] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present disclosure. Brief Description of the Drawings

[0023] The accompanying drawings here are incorporated into the specification and form a part of this specification, showing embodiments in accordance with the present invention, and are used together with the specification to explain the principles of the present invention. Figure 1 It is a flowchart of a text grading method for children based on artificial intelligence in an embodiment of the present invention; Figure 2 It is a schematic diagram of the preprocessing process for the text content to be evaluated in an embodiment of the present invention; Figure 3 It is a schematic diagram of the structure of a text grading system for children based on artificial intelligence in an embodiment of the present invention; Figure 4 It is a schematic diagram of an electronic device in an embodiment of the present invention. Detailed Description of the Embodiments

[0024] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It can be understood that the specific embodiments described here are only used to explain the present invention, rather than limiting the present invention. Additionally, it should be noted that for the sake of description, only parts related to the present invention are shown in the accompanying drawings, rather than all the structures.

[0025] Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts depict the steps as sequential processes, many of the steps can be implemented in parallel, concurrently, or simultaneously. In addition, the order of the steps can be rearranged. The process can be terminated when its operations are completed, but it can also have additional steps not included in the figures. The process can correspond to a method, function, procedure, subroutine, subprogram, and so on.

[0026] Embodiments of the present invention are directed to a text grading method, system, electronic device, and storage medium for children based on artificial intelligence, and provide the following embodiments: A text grading method for children based on artificial intelligence, as Figure 1 shown, the method includes the following steps: S1. Preprocess the text content to be evaluated to generate a standardized text representation. The preprocessing includes text tokenization, stop word filtering, and word form normalization; S2. Use a pre-trained model to generate a semantic representation of the text for the standardized text representation, and extract multi-dimensional feature vectors based on the semantic representation. The multi-dimensional feature vectors include topic classification, sentiment tendency, and language complexity features. Specifically, extracting multi-dimensional feature vectors using a pre-trained model for the standardized text representation includes: based on the input of the standardized text representation, using the pre-trained BERT model to obtain multi-dimensional vectors of the CLS token and obtain the topic classification feature vector; based on the input of the standardized text representation, using the pre-trained BERT model for encoding, generating multi-dimensional semantic representations for each token to obtain the sentiment tendency confidence; based on the input of the standardized text representation, using the pre-trained BERT model for encoding, generating multi-dimensional semantics for each token, adding a bi-affine attention layer, predicting the dependency relationship between each token, and quantifying the language complexity to obtain the language complexity feature; S3. Input the multi-dimensional feature vectors into a multi-layer perceptron. The multi-layer perceptron processes the non-linear mapping of the multi-dimensional feature vectors through the Sigmoid function to generate a content suitability score from 0 to 100; S4. Normalize the content suitability score, fuse it with the multi-dimensional feature vectors, and input it into a Softmax classifier. Finally, the trained Softmax classifier outputs the prediction result of the age-appropriate grading label.

[0027] Among them, the preprocessing in S1 includes tokenization, stop word filtering, and word form normalization based on a statistical model. In a preferred embodiment, it also includes sentence splitting and part-of-speech tagging for the text. The specific implementation steps are as Figure 2 shown, including: S11. Tokenization: Split the text into word units; S12. Stop word filtering: Remove common words that have no meaning; S13. Word normalization: Unify the forms of words; S14. Syntactic analysis: Generate syntactic structures.

[0028] Furthermore, perform Chinese word segmentation (based on the conditional random field CRF model), stop word removal, and word normalization operations on the input text. The word segmentation model is optimized using the maximum entropy criterion, and the transition probability is calculated by the following formula: , where represents the transition probability, exp represents the exponential function, ∑k represents the summation over the index k, is the weight parameter, is the feature function, represents the summation over all possible summations.

[0029] In S2, multi-dimensional feature vectors are extracted from the standardized text representation using a pre-trained model, specifically including: Based on the input of the standardized text representation, use the pre-trained BERT model to obtain the 768-dimensional vector of the CLS token as the feature vector for topic classification and input it into a support vector machine for topic classification; Specifically, topic classification combines traditional statistical methods and deep semantic representation techniques to construct a high-dimensional topic classification model. The semantic representation of the pre-trained BERT (Bidirectional Encoder Representations) model is preferred, and the multi-class classification task is completed through a support vector machine (SVM).

[0030] More preferably, load the pre-trained BERT model, input the segmented text sequence, and obtain the 768-dimensional vector of the CLS token as the feature vector for topic classification; input the feature vector into the SVM, and the kernel function is the radial basis function (RBF): , where represents the similarity measure; represents the feature vector extracted by BERT; represents the squared Euclidean distance; represents the parameter that controls the influence of the distance; exp represents the exponential function.

[0031] The optimization objective is: , where w represents the weight vector; b represents the bias term; represents the squared norm of the weight; C represents the regularization parameter; represents the slack variable; represents the true label; represents the distance from the sample to the hyperplane; n represents the total number of samples.

[0032] Based on the input of the standardized text representation, a pre-trained BERT model is used for encoding to generate a 768-dimensional semantic representation for each token. A fully connected layer is added to map the 768-dimensional semantic representation to the probability distribution of sentiment labels and output the confidence through the softmax function. Specifically, a BERT model combined with a self-attention mechanism is used to construct a sentiment classification model to quantify the sentiment polarity and intensity. The specific implementation steps include: Text encoding: Use a pre-trained BERT model to encode the input text and generate a 768-dimensional embedding vector for each token.

[0033] CLS token output: Extract the CLS token vector (768-dimensional) output by the BERT model as the semantic representation of the entire text.

[0034] Sentiment classification: Add a fully connected layer to the CLS token output to map to the probability distribution of sentiment labels (such as positive, negative, neutral) and output the confidence through the softmax function.

[0035] Model fine-tuning: Fine-tune the BERT model on the sentiment analysis task and use a labeled sentiment dataset for children to optimize the model parameters to adapt to the language characteristics and sentiment expression methods of children's texts.

[0036] Based on the input of the standardized text representation, a pre-trained BERT model is used for encoding to generate a 768-dimensional semantic representation for each token. A bi-affine attention layer is added to predict the dependency relationship between tokens, generate a syntactic tree, and quantify the language complexity based on the topological analysis and information entropy calculation of the syntactic tree. Specifically, combining dependency parsing and information entropy to quantify the syntactic structure complexity and lexical diversity. The specific implementation steps include: Syntactic analysis: Use a BERT-based dependency parsing model to generate a syntactic tree. The specific steps are as follows: Use a pre-trained BERT model to encode the text and obtain a 768-dimensional semantic representation for each token; Add a bi-affine attention layer (Biaffine Attention) to the BERT output to predict the dependency relationship between tokens and generate a syntactic tree.

[0037] Syntactic complexity, including: tree depth: calculate the maximum depth (tree_depth) of the syntactic tree; branching factor: calculate the average branching factor of the syntactic tree ; Comprehensive syntactic complexity: , where β is the weighted coefficient of the branching factor.

[0038] Lexical entropy: Calculate the lexical entropy of the text, denoted as vocab_entropy, which is used to measure the lexical diversity.

[0039] Complexity fusion: Generate the final language complexity score through weighted summation. The formula is as follows:

[0040] where w1 and w2 are the weight coefficients of syntactic complexity and lexical entropy respectively.

[0041] Preferably, the multi-dimensional feature vector further includes social value features, and the social value features are extracted by the graph embedding technology of the knowledge graph.

[0042] Specifically, based on the knowledge graph and graph embedding technology (TransE), combined with the sentiment dictionary, extract the value orientation. The specific implementation steps include: Knowledge graph construction: Include positive and negative value entities and their relationships.

[0043] Graph embedding: , where min means that the optimization goal is to minimize the value of the entire function; means summing over all correct triples in the knowledge graph for summation; represents the embedding vectors of the head entity, relation, and tail entity, which are learned through the TransE model; represents the wrong head entity and tail entity generated by negative sampling; represents the embedding vector distance of the correct triple, calculated using the squared L2 norm, and this value should be small; represents the embedding vector distance of the wrong triple, also calculated using the squared L2 norm, and this value should be large; represents the margin, which is a hyperparameter (such as 0.1 or 1) used to control the difference in distance between the correct triple and the wrong triple.

[0044] Text representation: Use BERT to generate , and the knowledge graph entities are embedded through TransE ;

[0045] Similarity calculation: ;

[0046] where sim represents the cosine similarity score, ranging from -1 to 1. The closer the value is to 1, the more similar the semantics of the text and the entity are; The BERT embedding vectors representing the text, capturing the semantic information of the text; The TransE embedding vectors representing the knowledge graph entities, representing positive or negative values; Represents the dot product of two vectors, measuring the similarity of their directions; Represents the L2 norm of the text embedding vector and the entity embedding vector, used for normalization.

[0047] Tendency judgment: Determine the value tendency according to the threshold.

[0048] It should be noted that the extracted feature vectors are normalized to ensure the consistency of the dimensions of features in different dimensions. That is, the feature vectors output by each sub-module are normalized and then weighted and fused to generate the feature vector x = [T, E, C, V] 。 Among them, T represents the topic classification features of the text, such as the probability distribution or classification label of the text belonging to categories such as education, entertainment, violence, etc. These features are usually generated by the pre-trained BERT model and the support vector machine (SVM), helping the system understand the main content and type of the text. E represents the sentiment tendency features of the text, such as sentiment polarity (positive, negative, neutral) and its intensity. These features are generated by the BERT model combined with the self-attention mechanism, used to evaluate the impact of the text on children's emotions. C represents the quantification index of the language complexity of the text, such as the complexity score calculated based on the depth of the syntactic tree, the branching factor, and the lexical entropy. These features are generated by the dependency syntactic analysis and information entropy calculation, helping the system judge whether the language difficulty of the text is suitable for children. V represents the social value tendency features of the text, such as the positive or negative value tendency extracted by the knowledge graph and the graph embedding technology (TransE). These features are used to evaluate whether the text conveys values that conform to social norms and children's educational needs. These features are generated by the four sub-modules of the text feature extraction module respectively and integrated into a comprehensive feature vector x , as the input of the suitability scoring module and the age grading module. Through this multi-dimensional feature extraction, the system can comprehensively analyze the text content to ensure the accuracy and scientificity of the evaluation.

[0049] In S3, the content suitability score is generated through a multi-layer perceptron based on the multi-dimensional feature vector, and the specific expression is: , where represents the content suitability score, and the score range is 0 - 100. The higher the score, the more suitable the content is for children; x represents the feature vector, including topic classification, sentiment tendency, and language complexity, etc., : the weight matrix of the neural network; represents the bias vector; ReLU(z) = max(0, z) represents the rectified linear unit function; : The Sigmoid function maps the output to the range of 0 to 1.

[0050] The content suitability scoring process also includes: pre-training the MLP model and optimizing the parameters using an annotated dataset, which includes suitable and unsuitable children's text samples. In the inference stage, the feature vector is input in real time, and the suitability score is output.

[0051] Specifically, the dynamic suitability score non-linearly fuses the feature vector through a multi-layer perceptron (MLP) architecture to generate a suitability score (0 - 100). The specific implementation process includes: Feature embedding: , where We is the weight matrix of the fully connected layer, be is the bias vector, and x is the input comprehensive feature vector; Single-head self-attention: (where Q = WQ⋅e, K = WK⋅e, V = WV⋅e are the linear transformation results of the query, key, and value respectively, and d is the feature dimension); Feed-forward network (FFN): , where W 1 is the weight matrix of the first layer, b 1 is the bias vector); Output: , where W 2 is the weight matrix of the second layer, b 2 is the bias vector, and σ is the Sigmoid function.

[0052] In S4, the classification model is used to predict the age classification based on the suitability score and the feature vector, specifically including: Fusing the suitability score and the feature vector into an input vector; Outputting the probability distribution of each age classification through a Softmax classifier; Selecting the age classification with the highest probability as the prediction result.

[0053] Furthermore, the age classification also includes: Pre-defining the cognitive and psychological characteristics of children's age groups, including vocabulary comprehension ability and emotional acceptance range; Training the Softmax classifier through historical data.

[0054] Specifically, the age classification is based on a multi-classification model (such as a Softmax classifier), combines the suitability score and the feature vector, and predicts the appropriate age classification. The model is trained through cross-entropy loss and an L2 regularization term is introduced to prevent overfitting. The implementation steps include: (1)Feature Fusion: The inputs to the age classification module are the suitability score and the feature vector. First, the suitability score (Score) and the feature vector (x) are fused to form a unified input vector , and the specific operations are as follows: Normalize the suitability score Score (ranging from [0, 100]) to the interval [0, 1], that is , to ensure consistency with the dimension of the feature vector.

[0055] Concatenate the normalized suitability score with the feature vector x (including multi-dimensional features such as topic classification, sentiment tendency, language complexity, and value assessment) to obtain the fused input vector .

[0056] (2)Classification Prediction: The fused input vector is fed into the Softmax classifier to output the probability distribution of each age classification. The specific process is as follows: The Softmax classifier first performs a linear transformation on the input vector to calculate the scores (logits) for each age classification: z = Wx′ + b where W is the weight matrix, b is the bias vector, and z is the score vector for each age classification, with the dimension equal to the number of age classifications K (such as 0 - 3 years old, 3 - 6 years old, 7 - 12 years old, etc.).

[0057] Subsequently, the scores z are converted into a probability distribution through the Softmax function: , where P(y = k∣x′) represents the probability that the input x′ is classified into the k-th age classification, and K is the total number of age classification categories.

[0058] (3)Prediction Result: According to the probability distribution output by the Softmax classifier, select the age classification with the highest probability as the final prediction result: , where is the predicted age suitability classification label.

[0059] (4)Model Training: The parameters (W and b) of the Softmax classifier are optimized through supervised learning, and the training objective is to minimize the cross-entropy loss function: , where: N is the number of training samples; is the true label of the i-th sample (using one-hot encoding, if the sample belongs to the k-th class, then , otherwise 0); is the probability predicted by the model. To prevent the model from overfitting, an L2 regularization term is added to the loss function: , where is a hyperparameter of the regularization strength, determined by cross-validation, is the sum of squares of the L2 norm of the weight W. During the training process, the Adam optimizer is used to update the model parameters, and a learning rate decay strategy is introduced (e.g., the initial learning rate is set to 0.001 and decays by 10% every several rounds) to improve the model convergence speed and stability.

[0060] (5) Inference stage: In practical applications, for new children's text content, the system first generates a feature vector x and a suitability score Score through the text feature extraction module and the suitability scoring module, and then generates an input vector according to the above feature fusion steps . Through the trained Softmax classifier, the probability distribution of each age classification is quickly calculated, and the optimal age classification is output .

[0061] (6) Prediction process: The probability distribution of each age classification is output through the Softmax classifier, and the age classification with the highest probability is selected as the prediction result. The specific steps include: fusing the suitability score and the feature vector into a comprehensive feature vector, inputting it into the Softmax classifier, generating the probability distribution and taking the maximum value.

[0062] Another embodiment is used to illustrate a text grading system for children based on artificial intelligence. Refer to Figure 3 , the system 300 includes: A preprocessing module 310 for preprocessing the text content to be evaluated and generating a standardized text representation. The preprocessing includes text tokenization, stop word filtering, and word form normalization; A feature extraction module 320 for generating a semantic representation of the text using a pre-trained model for the standardized text representation, and extracting a multi-dimensional feature vector based on the semantic representation. The multi-dimensional feature vector includes topic classification, sentiment tendency, and language complexity features; specifically extracting the multi-dimensional feature vector using the pre-trained model for the standardized text representation includes: based on the input of the standardized text representation, using the pre-trained BERT model to obtain the multi-dimensional vector of the CLS token and obtain the topic classification feature vector; based on the input of the standardized text representation, using the pre-trained BERT model for encoding, generating a multi-dimensional semantic representation for each token to obtain the sentiment tendency confidence; based on the input of the standardized text representation, using the pre-trained BERT model for encoding, generating a multi-dimensional semantics for each token, adding a bi-affine attention layer, predicting the dependency relationship between each token, and quantifying the language complexity to obtain the language complexity feature; A content suitability scoring module 330 is configured to input the multi-dimensional feature vector into a multi-layer perceptron. The multi-layer perceptron processes the non-linear mapping of the multi-dimensional feature vector through a Sigmoid function to generate a content suitability score ranging from 0 to 100. An age grading module 340 is configured to normalize the content suitability score, fuse it with the multi-dimensional feature vector, and then input the result into a Softmax classifier. Finally, the trained Softmax classifier outputs a predicted result of the age grading label.

[0063] In addition to the above modules, the system 300 may further include other components. However, since these components are not related to the content of the embodiments of the present disclosure, their illustrations and descriptions are omitted here.

[0064] For other specific working processes of the text grading system 300 for children based on artificial intelligence, reference may be made to the description of the embodiments of the text grading method for children based on artificial intelligence above, and details are not repeated here.

[0065] Another embodiment is used to illustrate that the system of the present invention can also be implemented by means of Figure 4 the architecture of the computing device shown. Figure 4 The architecture of the computing device is shown. As Figure 4 shown, a computer system 410, a system bus 430, one or more CPUs 440, an input / output 420, a memory 450, etc. The memory 450 can store various data or files used for computer processing and / or communication, as well as program instructions executed by the CPU, including the program instructions of the text grading method for children based on artificial intelligence in the embodiments. Figure 4 The architecture shown is only exemplary. When implementing different devices, one or more components in Figure 4 are adjusted according to actual needs. The memory 450, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and modules, such as the program instructions / modules corresponding to the text grading method for children based on artificial intelligence in the embodiments of the present invention (for example, the preprocessing module 310, the feature extraction module 320, the content suitability scoring module 330, and the age grading module 340 in the text grading system 300 for children based on artificial intelligence). One or more CPUs 440 execute various functional applications and data processing of the system of the present invention by running the software programs, instructions, and modules stored in the memory 450, that is, implementing the above-mentioned text grading method for children based on artificial intelligence, and the method includes the following steps: Preprocess the text content to be evaluated to generate a standardized text representation. The preprocessing includes text tokenization, stop word filtering, and lemmatization. Generate the semantic representation of the text using a pre-trained model for the standardized text representation, and extract multi-dimensional feature vectors based on the semantic representation. The multi-dimensional feature vectors include topic classification, sentiment tendency, and language complexity features. Specifically, extracting multi-dimensional feature vectors from the standardized text representation using a pre-trained model includes: based on the input of the standardized text representation, using the pre-trained BERT model to obtain the multi-dimensional vector of the CLS token and obtain the topic classification feature vector; based on the input of the standardized text representation, using the pre-trained BERT model to encode, generate multi-dimensional semantic representations for each token to obtain the sentiment tendency confidence; based on the input of the standardized text representation, using the pre-trained BERT model to encode, generate multi-dimensional semantics for each token, add a bi-affine attention layer, predict the dependency relationship between each token, and quantify the language complexity to obtain the language complexity feature; Input the multi-dimensional feature vectors into a multi-layer perceptron. The multi-layer perceptron processes the non-linear mapping of the multi-dimensional feature vectors through the Sigmoid function to generate a content suitability score from 0 to 100; Normalize the content suitability score, fuse it with the multi-dimensional feature vectors, and then input them into a Softmax classifier. Finally, the trained Softmax classifier outputs the prediction result of the age-appropriate classification label.

[0066] Of course, for the server provided by the embodiments of the present invention, its processor is not limited to performing the method operations as described above, and can also perform related operations in the text grading method for children based on artificial intelligence provided by any embodiment of the present invention.

[0067] The memory 450 may mainly include a program storage area and a data storage area. Among them, the program storage area may store an operating system and application programs required for at least one function; the data storage area may store data created according to the use of the terminal, etc. In addition, the memory 450 may include high-speed random access memory, and may also include non-volatile memory, such as at least one magnetic disk storage device, flash memory device, or other non-volatile solid-state storage devices. In some instances, the memory 450 may further include a memory remotely set relative to one or more CPUs 440, and these remote memories may be connected to the device through a network. Examples of the above network include but are not limited to the Internet, enterprise intranet, local area network, mobile communication network, and their combinations.

[0068] The input / output 420 can be used to receive input digital or character information, and generate key signal inputs related to the user settings and function control of the device. The input / output 420 may also include a display device such as a display screen.

[0069] An embodiment of the present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the text grading method for children based on artificial intelligence described in the above embodiments. The computer-readable storage medium of the embodiment of the present invention can adopt any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the above. More specific examples (non-exhaustive list) of the computer-readable storage medium include: an electrical connection having one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this document, the computer-readable storage medium can be any tangible medium that contains or stores a program, and this program can be used by or in combination with an instruction execution system, apparatus, or device.

[0070] The computer-readable signal medium can include a data signal propagated in a baseband or as part of a carrier wave, which carries the computer-readable program code. Such a propagated data signal can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. The computer-readable signal medium can also be any computer-readable medium other than the computer-readable storage medium, and this computer-readable medium can send, propagate, or transmit a program for use by or in combination with an instruction execution system, apparatus, or device.

[0071] The program code contained on the storage medium can be transmitted by any suitable medium, including but not limited to wireless, wire, optical cable, RF, etc., or any suitable combination of the above.

[0072] In addition, for the other specific working processes of a non-transitory computer-readable storage medium, reference can be made to the description of the embodiments of the above text grading method for children based on artificial intelligence, and details will not be repeated here.

[0073] To verify the effectiveness of the method, system, electronic device, and storage medium of the present invention, verification experiments were conducted. Multiple rounds of experiments were carried out on a diverse annotated dataset containing 500 publicly available children's books and 500 online-generated contents. The dataset covers various themes such as education, entertainment, adventure, science fiction, history, literature, etc., ensuring the robustness of the model across different content types. The dataset was annotated by child education experts according to Piaget's cognitive development theory, and the annotation content includes artificial suitability scores (0 - 100 points) and age-appropriate grading (0 - 3 years old, 3 - 6 years old, 7 - 12 years old, etc.). The experiments adopted five-fold cross-validation and introduced multiple evaluation metrics and baseline methods for comparison to enhance the scientificity and persuasiveness of the verification.

[0074] The experimental settings include: (1) Dataset division: Divided into a training set, a validation set, and a test set in the ratio of 8:1:1.

[0075] (2) The model includes: topic classification, sentiment analysis, suitability scoring, and age-appropriate grading.

[0076] (3) Baseline: The performance of human experts is used as a benchmark.

[0077] The evaluation metrics include: (1) Topic classification: Accuracy; (2) Sentiment analysis: F1 score; (3) Suitability scoring: Pearson correlation coefficient; (4) Age-appropriate grading: Consistency (Accuracy); Experimental results: Table 1 shows the comparison data (accuracy comparison) between the present method and the human benchmark. After adjustment, the performance of the present method is slightly higher than the human benchmark.

[0078] Table 1 Comparison Data between the Present Method and the Human Benchmark

[0079] The experimental results show that on the 1,000-piece dataset, the present method outperforms the baseline method in tasks such as topic classification, sentiment analysis, suitability scoring, and age-appropriate grading, verifying its technical advantages and practicality.

[0080] The present invention realizes the comprehensive evaluation and age-appropriate grading of children's text content through multi-dimensional feature extraction, dynamic fusion of deep neural networks, and refined grading of probability models. Its main technical innovation points include: Multi-level feature fusion: Combining statistical, semantic, and knowledge-driven methods to capture the deep characteristics of the text.

[0081] Dynamic scoring mechanism: Using the self-attention mechanism to adapt to the complex interaction relationships of different texts.

[0082] Cognitive science constraints: Incorporate the laws of child development to enhance the scientific nature of grading.

[0083] The above content is only one of the specific implementation manners of the technical solution of the present invention. Its detailed description aims to explain the technical principles, algorithm architectures and application scenarios of the present invention, rather than a comprehensive limitation of the protection scope of the present invention. After those skilled in the art master the technical ideas and method frameworks disclosed by the present invention, they can, without departing from the core spirit and technical purpose of the present invention, combine their own professional knowledge and application requirements to diversify the adjustment, optimization or equivalent replacement of the technical details, parameter configurations, algorithm combinations or processing flows in the above embodiments. Any variation forms derived from the technical solution of the present invention, including but not limited to the improvement of feature extraction strategies, the variant design of suitability scoring models, and the alternative implementation of age grading algorithms, shall be regarded as the natural extension and coverage of the technical protection scope of the present invention. Therefore, the actual protection scope of the present invention shall be subject to the technical features clearly defined in the claims and their equivalent scopes, rather than being limited to the specific embodiments described herein. Through the complex integration of multi-dimensional feature extraction and deep learning technologies, the present invention provides an innovative solution for the evaluation and grading of text content for children. Its technical value and application potential have significant expandability and universality in this field. It can be widely applied to fields such as children's digital reading materials, online education platforms, and game content reviews, providing strong technical support for protecting children from the potential harms of inappropriate content.

[0084] In this article, the term "including", "comprising" or any other variant thereof is intended to cover non-exclusive inclusion, so that a step, method including a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such step, method.

[0085] The above content is a further detailed description of the present invention in combination with specific preferred implementation manners, and it cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention pertains, without departing from the concept of the present invention, several simple deductions or replacements can still be made, and all should be regarded as belonging to the protection scope of the present invention.

Claims

1. A text grading method for children based on artificial intelligence, characterized in that: The method comprises the following steps: Preprocessing the text content to be evaluated to generate a standardized text representation, wherein the preprocessing includes text segmentation, stop word filtering, and word form normalization; The standardized text representation is represented by using a pre-trained model to generate a semantic representation of the text, and a multi-dimensional feature vector is extracted based on the semantic representation, wherein the multi-dimensional feature vector includes topic classification, sentiment tendency and language complexity features; the standardized text representation is represented by using a pre-trained model to extract a multi-dimensional feature vector, which specifically includes: based on the input of the standardized text representation, a pre-trained BERT model is used to obtain a multi-dimensional vector of the CLS tag and obtain a topic classification feature vector; based on the input of the standardized text representation, a pre-trained BERT model is used for encoding, and a multi-dimensional semantic representation is generated for each token to obtain sentiment tendency confidence; based on the input of the standardized text representation, a pre-trained BERT model is used for encoding, and a multi-dimensional semantic representation is generated for each token, a dual affine attention layer is added, the dependency relationship between each token is predicted, and the language complexity is quantified to obtain the language complexity features; Inputting the multi-dimensional feature vector into a multi-layer perceptron, wherein the multi-layer perceptron processes the non-linear mapping of the multi-dimensional feature vector through a Sigmoid function to generate a content suitability score from 0 to 100; The content suitability score is normalized, fused with the multi-dimensional feature vector and input into the Softmax classifier. Finally, the trained Softmax classifier outputs the age-appropriate classification label prediction result.

2. The method for text grading for children based on artificial intelligence according to claim 1, characterized in that: The pre-trained model is used to extract multi-dimensional feature vectors from the standardized text representation, including: Based on the input of the standardized text representation, a pre-trained BERT model is used to obtain a 768-dimensional vector of the CLS tag, which is used as a feature vector for topic classification and input into a support vector machine. The support vector machine performs topic classification based on a radial basis function. Based on the input of standardized text representation, the pre-trained BERT model is used for encoding to generate a 768-dimensional semantic representation for each token. A fully connected layer is added to map the 768-dimensional semantic representation to the probability distribution of the sentiment label, and the confidence is output through the softmax function. Based on the input of standardized text representation, the pre-trained BERT model is used for encoding, a 768-dimensional semantic representation is generated for each token, a dual affine attention layer is added, the dependency relationship between tokens is predicted, a syntactic tree is generated, and the language complexity is quantified based on the topological analysis and information entropy calculation of the syntactic tree; The extracted feature vectors are normalized to ensure that the dimensions of features in different dimensions are consistent.

3. The method for text grading for children based on artificial intelligence according to claim 1, characterized in that: The multi-dimensional feature vector also includes social value features, which are extracted through graph embedding technology of knowledge graph.

4. The method for text grading for children based on artificial intelligence according to claim 1, characterized in that: The content suitability score is generated by a multi-layer perceptron based on the multi-dimensional feature vector. The specific expression is: ， in represents the content suitability score, represents the Sigmoid function, is the weight matrix of the neural network, x represents the multi-dimensional feature vector, ReLU represents the rectified linear unit function.

5. The method for text grading for children based on artificial intelligence according to claim 1, characterized in that: The classification model is used to predict age classification based on the suitability score and multi-dimensional feature vector, including: The suitability score and the feature vector are fused into an input vector; Output the probability distribution of each age level through the Softmax classifier; The age classification with the highest probability is selected as the prediction result.

6. The method for text grading for children based on artificial intelligence according to claim 4, characterized in that: The content suitability scoring process also includes: A multilayer perceptron is pre-trained and its parameters are optimized using a labeled dataset that includes appropriate and inappropriate children's text samples.

7. The method for text grading for children based on artificial intelligence according to claim 5, characterized in that: The age ratings also include: Predefine cognitive and psychological characteristics of children of different ages, including vocabulary comprehension and emotional receptivity; Train the Softmax classifier using historical data.

8. A text grading system for children based on artificial intelligence, characterized in that: The system comprises: A preprocessing module, used to preprocess the text content to be evaluated to generate a standardized text representation, wherein the preprocessing includes text segmentation, stop word filtering, and word form normalization; A feature extraction module is used to generate a semantic representation of the text using a pre-trained model for the standardized text representation, and extract a multi-dimensional feature vector based on the semantic representation, wherein the multi-dimensional feature vector includes topic classification, sentiment tendency and language complexity features; the method of extracting a multi-dimensional feature vector from the standardized text representation using a pre-trained model specifically includes: based on the input of the standardized text representation, using a pre-trained BERT model to obtain a multi-dimensional vector of the CLS tag and obtain a topic classification feature vector; based on the input of the standardized text representation, using a pre-trained BERT model for encoding, generating a multi-dimensional semantic representation for each token to obtain sentiment tendency confidence; based on the input of the standardized text representation, using a pre-trained BERT model for encoding, generating a multi-dimensional semantic for each token, adding a dual affine attention layer, predicting the dependency between each token, and quantifying the language complexity to obtain the language complexity features; A content suitability scoring module, used for inputting the multi-dimensional feature vector into a multi-layer perceptron, wherein the multi-layer perceptron processes the non-linear mapping of the multi-dimensional feature vector through a Sigmoid function to generate a content suitability score from 0 to 100; The age classification module is used to normalize the content suitability score, fuse it with the multi-dimensional feature vector, and then input it into the Softmax classifier. Finally, the trained Softmax classifier outputs the age classification label prediction result.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the artificial intelligence-based text grading method for children as described in any one of claims 1 to 7 are implemented.

10. A non-transitory computer-readable storage medium having computer instructions stored thereon, characterized in that: When the instructions are executed by the processor, the steps of the artificial intelligence-based text grading method for children as described in any one of claims 1 to 7 are implemented.

Images