Online homework correction method and system based on AI

Abstract

The invention discloses an AI-based online homework correction method and system, and the method comprises the steps: carrying out the feature extraction of text content, logic structure and language expression through a multi-dimensional feature extraction model according to the homework data submitted by a student; according to the multi-dimensional feature vectors, a subject knowledge base and logical reasoning rules are combined, and error detection and correction are conducted through a logical reasoning network; according to the homework feature vector, in combination with a teacher scoring standard and a student ability evaluation index, performing scoring and ability analysis by using a dynamic scoring and ability evaluation model; and generating a personalized feedback report by using a generative adversarial network according to a scoring result and a capability evaluation result. According to the embodiment of the invention, various errors in the homework can be efficiently and accurately identified and corrected, personalized feedback is provided, and the quality and effect of online education are improved.

Description

Technical Field

[0001] The present invention belongs to the field of AI technology, and particularly relates to an online homework correction method and system based on AI. Background Art

[0002] With the rapid development of information technology, online education has gradually become an important learning method. The submission and correction of online homework not only improve the flexibility of teaching but also facilitate the sharing of educational resources. However, traditional homework correction methods still mainly rely on manual grading, resulting in low correction efficiency, untimely feedback, and strong subjectivity. In addition, when correcting homework, teachers often have difficulty comprehensively and accurately identifying errors in students' text content, logical structure, and language expression, lacking a systematic evaluation mechanism, which limits the improvement of teaching effects.

[0003] With the rapid development of artificial intelligence technology, AI-based automated correction systems have gradually emerged. These systems can analyze the homework submitted by students through algorithms and provide timely feedback. However, existing AI correction methods often focus on single-aspect analysis, lack comprehensive consideration of multi-dimensional features, and still have deficiencies in the detection and correction of logical errors. Traditional scoring criteria also fail to effectively adapt to individual student differences, resulting in the accuracy and pertinence of scoring results being affected. Summary of the Invention

[0004] The purpose of the present invention is to provide an online homework correction method and system based on AI to solve the deficiencies in the prior art, which can efficiently and accurately identify and correct various errors in homework, provide personalized feedback, and improve the quality and effect of online education.

[0005] An embodiment of the present application provides an online homework correction method based on AI, and the method includes:

[0006] According to the homework data submitted by students, use a multi-dimensional feature extraction model to extract features from the text content, logical structure, and language expression. Among them, the multi-dimensional feature extraction model extracts semantic features from the three levels of vocabulary, sentence, and discourse through a hierarchical convolutional neural network and a self-attention mechanism, and generates a multi-dimensional feature vector;

[0007] According to the multi-dimensional feature vector, in combination with a subject knowledge base and logical reasoning rules, use a logical reasoning network for error detection and correction. Among them, the logical reasoning network uses a graph reasoning algorithm and a rule engine to identify logical errors in the homework and generate correction suggestions, and outputs a corrected homework feature vector;

[0008] According to the described assignment feature vector, in combination with the teacher's grading criteria and the student ability evaluation indicators, a dynamic grading and ability analysis are carried out using a dynamic grading and ability evaluation model. Among them, the dynamic grading and ability evaluation model dynamically adjusts the grading weights and ability evaluation parameters through multi-task learning and an adaptive weight allocation mechanism to ensure the accuracy of the grading results and the comprehensiveness of the ability evaluation results;

[0009] According to the grading results and the ability evaluation results, a personalized feedback report is generated using a generative adversarial network. Among them, the generative adversarial network generates multi-modal feedback content including text, charts, and speech through a multi-modal generator and a context-aware discriminator.

[0010] Optionally, based on the assignment data submitted by the student, a multi-dimensional feature extraction model is used to extract features from the text content, logical structure, and language expression. Among them, the multi-dimensional feature extraction model extracts semantic features from three levels of vocabulary, sentence, and discourse respectively through a hierarchical convolutional neural network and a self-attention mechanism, and generates a multi-dimensional feature vector, including:

[0011] The assignment data submitted by the student is tokenized, stop words are removed, and word vectorization is performed to obtain a feature representation at the vocabulary level;

[0012] According to the feature representation at the vocabulary level, a convolutional operation is performed using the first layer of the hierarchical convolutional neural network to extract local semantic features at the vocabulary level, and through a max-pooling operation, important features at the vocabulary level are screened out to generate a vocabulary feature vector;

[0013] According to the vocabulary feature vector, a convolutional operation is performed using the second layer of the hierarchical convolutional neural network to extract semantic features at the sentence level, and through a self-attention mechanism, the attention weights between words are calculated to dynamically adjust the contribution of the vocabulary features, generating a sentence feature vector;

[0014] According to the sentence feature vector, a convolutional operation is performed using the third layer of the hierarchical convolutional neural network to extract global semantic features at the discourse level, and through a self-attention mechanism, the attention weights between sentences are calculated to dynamically adjust the contribution of the sentence features, generating a discourse feature vector;

[0015] The vocabulary feature vector, the sentence feature vector, and the discourse feature vector are concatenated to generate a multi-dimensional feature vector. Among them, the hierarchical convolutional neural network includes: the first layer has a convolutional kernel size of 3 and is used to extract local semantic features at the vocabulary level; the second layer has a convolutional kernel size of 5 and is used to extract semantic features at the sentence level; the third layer has a convolutional kernel size of 7 and is used to extract global semantic features at the discourse level.

[0016] Optionally, based on the multi-dimensional feature vector, in combination with a subject knowledge base and logical reasoning rules, a logical reasoning network is used for error detection and correction. The logical reasoning network identifies logical errors in the homework and generates correction suggestions through a graph reasoning algorithm and a rule engine, and outputs a corrected homework feature vector, including:

[0017] Based on the multi-dimensional feature vector, knowledge matching is performed in combination with the subject knowledge base, and the matched subject knowledge is represented as a knowledge feature vector;

[0018] The multi-dimensional feature vector and the knowledge feature vector are input into the logical reasoning network. The logical reasoning network initializes the feature representations of nodes and edges through a graph reasoning algorithm and a rule engine, where each node represents a logical unit and each edge represents a logical relationship;

[0019] According to the graph reasoning algorithm, the nodes and edges in the logical reasoning network are traversed to detect logical errors, and the detected logical errors are represented as error feature vectors. For each logical unit, its matching degree with the subject knowledge base is calculated to identify unmatched logical units. For each logical relationship, its compliance degree with the rule engine is calculated to identify non-compliant logical relationships;

[0020] Based on the error feature vector, correction suggestions are generated using the rule engine. For unmatched logical units, replacement suggestions are generated, and for non-compliant logical relationships, adjustment suggestions are generated;

[0021] The correction suggestions are applied to the logical reasoning network to generate a corrected logical feature vector as the homework feature vector.

[0022] Optionally, based on the homework feature vector, in combination with the teacher's grading criteria and student ability evaluation indicators, a dynamic grading and ability evaluation model is used for grading and ability analysis. The dynamic grading and ability evaluation model dynamically adjusts the grading weights and ability evaluation parameters through multi-task learning and an adaptive weight allocation mechanism to ensure the accuracy of the grading results and the comprehensiveness of the ability evaluation results, including:

[0023] Load the grading rules according to the teacher's grading criteria and encode the grading rules as a grading weight vector;

[0024] Load the ability evaluation parameters according to the student ability evaluation indicators and encode the ability evaluation parameters as an ability weight vector;

[0025] The homework feature vector, the grading weight vector, and the ability weight vector are input into the multi-task learning model. The multi-task learning model initializes the weights of the grading task and the ability evaluation task through an adaptive weight allocation mechanism; where

[0026] Feature extraction is performed using the shared layer of the multi-task learning model based on the task feature vector. Among them, high-order feature representations are extracted through a fully connected network and a non-linear activation function;

[0027] Scoring is performed using the scoring task-specific layer based on the high-order feature representation and the scoring weight vector. Among them, the accuracy, integrity, and logic of the task are scored to generate a preliminary scoring result. According to the adaptive weight allocation mechanism, the weights of each scoring dimension are dynamically adjusted to generate a final scoring result;

[0028] Capability analysis is performed using the capability assessment task-specific layer based on the high-order feature representation and the capability weight vector. Among them, the knowledge mastery, logical thinking ability, and expression ability of the student are evaluated to generate a preliminary capability assessment result. According to the adaptive weight allocation mechanism, the weights of each capability assessment parameter are dynamically adjusted to generate a final capability assessment result.

[0029] Optionally, a personalized feedback report is generated using a generative adversarial network based on the scoring result and the capability assessment result. Among them, the generative adversarial network generates multi-modal feedback content including text, charts, and speech through a multi-modal generator and a context-aware discriminator, including:

[0030] The final scoring result and the final capability assessment result are input into the multi-modal generator, and the multi-modal generator generates text, chart, and speech feedback content respectively through a multi-task learning mechanism; among them,

[0031] Text feedback content is generated using the text generation module based on the final scoring result and the final capability assessment result. Among them, the final scoring result is summarized to generate a text evaluation, the final capability assessment result is explained to generate improvement suggestions, and the text feedback content is generated by combining the text evaluation and the improvement suggestions;

[0032] Chart feedback is generated using the chart generation module based on the final scoring result and the final capability assessment result. Among them, the final scoring result is visualized to generate a first chart, the final capability assessment result is visualized to generate a second chart, and the chart feedback content is generated by combining the first chart and the second chart;

[0033] Speech feedback content is generated using the speech generation module based on the final scoring result and the final capability assessment result. Among them, the text evaluation is synthesized into speech to generate a speech evaluation, the improvement suggestions are synthesized into speech to generate speech suggestions, and the speech feedback content is generated by combining the speech evaluation and the speech suggestions;

[0034] Input the text feedback content, chart feedback content, and voice feedback content into the context-aware discriminator. The context-aware discriminator performs consistency checking and optimization on the feedback content through a context-aware mechanism, and finally outputs multi-modal feedback content including text, charts, and voice.

[0035] Another embodiment of the present application provides an AI-based online homework correction system, and the system includes:

[0036] An extraction module, configured to extract features of text content, logical structure, and language expression from the homework data submitted by students by using a multi-dimensional feature extraction model. Among them, the multi-dimensional feature extraction model extracts semantic features from three levels of vocabulary, sentence, and discourse respectively through a hierarchical convolutional neural network and a self-attention mechanism, and generates a multi-dimensional feature vector;

[0037] A correction module, configured to perform error detection and correction by using a logical reasoning network according to the multi-dimensional feature vector, in combination with a subject knowledge base and logical reasoning rules. Among them, the logical reasoning network identifies logical errors in the homework and generates correction suggestions through a graph reasoning algorithm and a rule engine, and outputs a corrected homework feature vector;

[0038] An analysis module, configured to perform grading and ability analysis by using a dynamic grading and ability evaluation model according to the homework feature vector, in combination with a teacher grading standard and a student ability evaluation index. Among them, the dynamic grading and ability evaluation model dynamically adjusts the grading weight and ability evaluation parameters through multi-task learning and an adaptive weight allocation mechanism to ensure the accuracy of the grading result and the comprehensiveness of the ability evaluation result;

[0039] A generation module, configured to generate a personalized feedback report by using a generative adversarial network according to the grading result and the ability evaluation result. Among them, the generative adversarial network generates multi-modal feedback content including text, charts, and voice through a multi-modal generator and a context-aware discriminator.

[0040] Another embodiment of the present application provides a storage medium, in which a computer program is stored. Among them, the computer program is set to execute the method described in any one of the above when running.

[0041] Another embodiment of the present application provides an electronic device, including a memory and a processor. A computer program is stored in the memory, and the processor is set to run the computer program to execute the method described in any one of the above.

[0042] Compared with the prior art, an AI-based online homework correction method provided by the present invention extracts features of text content, logical structure, and language expression using a multi-dimensional feature extraction model according to the homework data submitted by students; performs error detection and correction using a logical reasoning network based on the multi-dimensional feature vectors, in combination with a subject knowledge base and logical reasoning rules; performs scoring and ability analysis using a dynamic scoring and ability evaluation model based on the homework feature vectors, in combination with the teacher's scoring criteria and student ability evaluation indicators; and generates a personalized feedback report using a generative adversarial network according to the scoring results and ability evaluation results, so as to be able to efficiently and accurately identify and correct various errors in the homework, provide personalized feedback, and improve the quality and effect of online education. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 The hardware structure block diagram of a computer terminal for an AI-based online homework correction method provided by an embodiment of the present invention;

[0044] Figure 2 The flowchart of an AI-based online homework correction method provided by an embodiment of the present invention;

[0045] Figure 3 The structure diagram of an AI-based online homework correction system provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0046] The embodiments described below by referring to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as a limitation of the present invention.

[0047] An embodiment of the present invention first provides an AI-based online homework correction method, which can be applied to an electronic device, such as a computer terminal, specifically, an ordinary computer, etc.

[0048] The following takes running on a computer terminal as an example to describe it in detail. Figure 1 The hardware structure block diagram of a computer terminal for an AI-based online homework correction method provided by an embodiment of the present invention. As Figure 1 shown, the computer device includes a processor, a memory, and a network interface connected through a system bus. Among them, the memory may include a non-volatile storage medium and an internal memory.

[0049] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions, and when the program instructions are executed, the processor can execute any AI-based online homework correction method.

[0050] The processor is used to provide computing and control capabilities to support the operation of the entire computer device.

[0051] The internal memory provides an environment for the operation of a computer program in a non-volatile storage medium. When the computer program is executed by a processor, the processor can execute any one of the online homework grading methods based on AI.

[0052] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art can understand that Figure 1 the structure shown in is only a block diagram of some structures related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0053] It should be understood that the processor may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0054] See Figure 2 , an embodiment of the present invention provides an online homework grading method based on AI, which may include the following steps:

[0055] S201, according to the homework data submitted by students, use a multi-dimensional feature extraction model to extract features from the text content, logical structure and language expression. Among them, the multi-dimensional feature extraction model extracts semantic features from the three levels of vocabulary, sentence and discourse through a hierarchical convolutional neural network and a self-attention mechanism, and generates a multi-dimensional feature vector;

[0056] In this method step, the system extracts features from the text content, logical structure, and language expression of the students' submitted homework data using a multi-dimensional feature extraction model. Specifically, the adopted multi-dimensional feature extraction model combines a hierarchical convolutional neural network and a self-attention mechanism, enabling the model to conduct in-depth analysis from multiple levels. First, the hierarchical convolutional neural network extracts semantic features at the lexical level, capturing the local information of words. Then, the self-attention mechanism is used to model sentences, automatically adjusting the importance of each vocabulary. Finally, through another convolutional operation, discourse-level features are obtained, extracting the overall logical structure and language expression features. This series of feature extraction processes ultimately generates a multi-dimensional feature vector, representing the overall quality features of the students' homework, including multi-faceted evaluations of the semantic information, structural logic, and language expression of the text.

[0057] By using the multi-dimensional feature extraction method, the system can comprehensively and deeply analyze the students' homework, extracting multi-level feature information. This approach greatly improves the accuracy and comprehensiveness of homework grading, enabling the grading to not only correct surface errors but also conduct in-depth evaluations of the text's logical structure and language expression. This ability not only helps teachers quickly identify the weak links in students' learning but also provides a data basis for subsequent personalized feedback and improvement suggestions, promoting the continuous improvement of students' knowledge mastery and expression abilities. Through precise feature extraction, educational evaluation can be made more intelligent and personalized, creating a better support environment for students' learning and development.

[0058] Specifically, the homework data submitted by students can be tokenized, stop words removed, and word vectors generated to obtain feature representations at the lexical level.

[0059] In this step, the system first performs basic preprocessing on the text submitted by students, including tokenization, removal of stop words, and word vectorization. Tokenization divides the entire text into independent words, helping the system understand the text structure. Subsequently, stop words (such as meaningless words like "is", "of", etc.) are filtered out to focus on key substantive words. Finally, word embedding technology is used to convert each word into a high-dimensional vector for subsequent feature extraction.

[0060] Through this processing step, the system provides clean and structured input data for subsequent text analysis. Word vectorization can not only capture the semantic relationships between words but also provide a basis for subsequent convolutional and attention mechanisms. For example, in semantic analysis, the meaning of the word "bank" may vary in different contexts, and through vectorization, the system can better judge its context. Such processing enables the subsequent model to focus on meaningful features during the learning and evaluation processes, thereby improving the accuracy and efficiency of homework grading.

[0061] First, the system inputs the received student assignment text into the Natural Language Processing (NLP) module for preliminary processing. This module first performs word segmentation, dividing the entire text into independent words. For example, for the sentence "I went to the bank to deposit money today", the word segmentation module cuts it into "I", "today", "went", "bank", "deposit money". The result of word segmentation provides basic structured information for subsequent analysis, enabling each word to be individually recognized and processed.

[0062] Next, the system performs stop word removal, eliminating words with no practical meaning in the text. Stop words such as "of", "is", "in" etc. are often not helpful for text analysis and are thus removed. Continuing with the example "I went to the bank to deposit money today", after stop word removal, the remaining keywords are "went", "bank", "deposit money". This step ensures that subsequent feature extraction will focus on important content, thereby improving the effectiveness of the analysis.

[0063] Finally, the system uses word embedding technology to convert the remaining keywords into word vectors. Commonly used word embedding methods include Word2Vec and GloVe, etc. These methods map each word to a high-dimensional space, converting it into vector form. For example, "went" may be represented as a vector [0.1, 0.5, -0.3], while "bank" may be [0.7, 0.2, 0.4]. In this way, all words are converted into numerical form, laying the foundation for subsequent feature extraction.

[0064] Based on the word-level feature representation, perform a convolution operation using the first layer of the hierarchical convolutional neural network to extract local semantic features at the word level, and through the max-pooling operation, screen out important features at the word level to generate word feature vectors;

[0065] In this step, the system inputs the word-level feature vectors into the first layer of the hierarchical convolutional neural network. Through the convolution operation, the system can extract local semantic features of the words. This operation helps the model capture the potential correlations between words. At the same time, the max-pooling operation is used to select the most significant features from the convolutional feature map, significantly reducing the data dimension and retaining important information, thereby generating a compact word feature vector.

[0066] The generation of word feature vectors enables the model to perform feature learning more effectively and build text understanding around the keywords. By extracting local semantic features, the system can identify representative word combinations in the assignment, thereby improving the accuracy of subsequent analysis and judgment. For example, if the student uses words related to "finance" multiple times in the expression, the system will recognize the importance of this theme based on the word feature vectors. Ultimately, accurate feature extraction lays the foundation for error detection and logical analysis.

[0067] In this step, the system inputs the lexical feature vectors generated in the first step into the first layer of the hierarchical convolutional neural network. The size of the convolutional kernel in this layer is set to 3 (i.e., processing three adjacent words as a group), aiming to extract local semantic features at the lexical level. For example, for the word vectors of the input sequence "go", "bank", and "deposit money", the convolutional operation will perform convolutional calculations on this group of words to generate a series of new feature maps, expressing the local relationships and semantics between these words.

[0068] Next, the system performs max pooling operation to select the most prominent features from the convolutional feature maps. This process can be understood as retaining the most representative values from the feature maps. For example, if the feature map contains a series of values [0.6, 0.4, 0.8], through max pooling, the system will select 0.8 as the representative feature of this feature map. This feature screening not only reduces the dimensionality of the data but also highlights the information most valuable for text analysis, ensuring more efficient subsequent processing.

[0069] Finally, through the convolutional and pooling operations, the system generates a relatively compact lexical feature vector. For example, if the feature values after pooling are [0.8, 0.2, 0.5], this vector will reflect the key features related to the words, preparing for the subsequent sentence-level feature extraction. This process not only improves the efficiency of feature extraction but also provides a more accurate data basis for subsequent analysis work.

[0070] Based on the lexical feature vectors, convolutional operations are performed using the second layer of the hierarchical convolutional neural network to extract sentence-level semantic features, and through the self-attention mechanism, the attention weights between words are calculated to dynamically adjust the contribution of lexical features and generate sentence feature vectors;

[0071] In this step, the system inputs the lexical feature vectors output by the first layer into the second layer of the hierarchical convolutional neural network for further processing. Using a larger convolutional kernel (such as 5) can capture sentence-level semantic features. Through the self-attention mechanism, the system calculates the importance weights of each word in the sentence, thereby dynamically adjusting the contribution of each word to the sentence feature vector and enhancing the overall semantic understanding ability of the sentence.

[0072] The implementation of this step enables the system to more comprehensively understand the logical structure and meaning of the sentence. When the system can identify which words have the greatest impact on the sentence meaning, the finally generated sentence feature vector will be more accurate, thus providing a reliable basis for the detection and correction of logical errors. For example, in the sentence "I go to the bank to deposit money", the system can identify the relevance between "go" and "deposit money" and the semantics they jointly shape, making the overall sentence evaluation more precise.

[0073] The system inputs the lexical feature vectors generated in the previous step into the second layer of the hierarchical convolutional neural network, aiming to extract semantic features at the sentence level. In this layer, the convolutional kernel size is set to 5, aiming to capture the semantic relationships between words in the sentence. For example, for the sentence "I go to the bank to deposit money", the system performs convolution on adjacent five words during processing to extract their features.

[0074] After the convolution operation, the system will use the self-attention mechanism to calculate the attention weights between words. This step can help the model dynamically adjust the importance of words. For example, in a sentence, "deposit money" may be the core information and its attention weight will be relatively high, while "I" may be given a lower weight. The system will generate a weight vector, such as [0.1, 0.3, 0.5, 0.1], indicating the contribution degree of each word to the sentence meaning.

[0075] Finally, the system dynamically adjusts the lexical features according to these weights to generate a sentence feature vector. Assuming that the sentence features obtained after convolution calculation are [0.4, 0.6, 0.5], after weighted processing, the final sentence feature vector may be [0.5, 0.7, 0.6]. This vector not only synthesizes the semantic information of the sentence but also reflects the roles played by each word in the sentence, providing an important basis for subsequent discourse feature extraction.

[0076] Based on the sentence feature vector, perform a convolution operation using the third layer of the hierarchical convolutional neural network to extract the global semantic features at the discourse level, and through the self-attention mechanism, calculate the attention weights between sentences to dynamically adjust the contribution of sentence features and generate a discourse feature vector;

[0077] In this step, the system inputs the sentence feature vector into the third layer of the hierarchical convolutional neural network to extract the global semantic features at the discourse level. Using a larger convolutional kernel (such as 7) and the self-attention mechanism, the system can capture the interrelationships between sentences within the discourse. In this way, the model can not only analyze the content of individual sentences but also understand their positions and roles in the whole article, generating a comprehensive discourse feature vector.

[0078] The extraction of discourse features is of crucial significance, which enables the system to evaluate the overall logic and theme of students' homework. Once the generated discourse feature vector can reflect the structural clarity and the continuity of the core ideas of the homework, the accuracy of grading and personalized feedback will also be improved accordingly. For example, if a student's homework shows strong hierarchical and logical order, the system can give positive feedback through the discourse feature vector to promote the student's further improvement in writing. <>

[0079] In this step, the system inputs the sentence feature vectors generated in the second step into the third layer of the hierarchical convolutional neural network. This layer is designed to use a relatively large convolutional kernel (e.g., 7) with the aim of capturing global semantic features at the discourse level. This operation will help to understand the logical relationships between sentences and the overall theme. For example, if the processed text consists of multiple sentences, the system will, through convolutional calculations, extract the role played by each sentence in the discourse.

[0080] Next, the system will apply the self-attention mechanism to weight between sentences in order to calculate the importance of each sentence in the discourse. This mechanism allows the system to dynamically adjust the contribution degree of sentences. For example, in a descriptive text, the introductory sentence at the beginning and the summary sentence at the end may be given higher weights. The finally generated sentence attention weight vector, such as [0.2, 0.3, 0.5], reflects the relative importance of each sentence.

[0081] Finally, based on these weights, the system dynamically adjusts the sentence feature vectors to form the discourse feature vector. Suppose the discourse features obtained during the convolutional process are [0.6, 0.4, 0.5], after being adjusted by the self-attention mechanism, the finally generated discourse feature vector may be [0.7, 0.5, 0.6]. This vector will comprehensively reflect the overall logic and semantic structure of the assignment, ensuring that subsequent analysis can take into account both details and the overall picture, providing deeper feedback for teachers.

[0082] The lexical feature vectors, sentence feature vectors, and discourse feature vectors are concatenated to generate a multi-dimensional feature vector. Among them, the hierarchical convolutional neural network includes: the first layer with a convolutional kernel size of 3 for extracting local semantic features at the lexical level; the second layer with a convolutional kernel size of 5 for extracting semantic features at the sentence level; the third layer with a convolutional kernel size of 7 for extracting global semantic features at the discourse level.

[0083] In this crucial step, the system concatenates the lexical feature vectors, sentence feature vectors, and discourse feature vectors generated by the first layer, second layer, and third layer to form a multi-dimensional feature vector. This multi-dimensional feature vector synthesizes features at multiple levels of the text, from the use of words in detail to the structure of sentences and then to the logical fluency of the overall discourse, providing rich input information for subsequent analysis.

[0084] Through the concatenation operation, the system can effectively integrate feature information from different levels, enabling the subsequent model to not only focus on features at a specific level during evaluation, but comprehensively consider information at all levels, ensuring the comprehensiveness and accuracy of assignment marking. For example, the system can, while identifying the clarity of expression in a certain sentence, also evaluate the appropriateness of word selection and the logic of the overall structure, thus providing more comprehensive feedback for teachers.

[0085] In this final step, the system concatenates the lexical feature vectors, sentence feature vectors, and passage feature vectors generated in the previous steps to generate a new multi-dimensional feature vector. Suppose the generated lexical feature vector is [0.9, 0.1, 0.5], the sentence feature vector is [0.7, 0.3, 0.6], and the passage feature vector is [0.8, 0.5, 0.4]. The system integrates these three vectors in sequence to form a new feature vector, such as [0.9, 0.1, 0.5, 0.7, 0.3, 0.6, 0.8, 0.5, 0.4].

[0086] This concatenation operation ensures the comprehensiveness of information, enabling the subsequent model to obtain data support at multiple levels in one analysis. The multi-dimensional feature vector not only integrates the detailed information of vocabulary but also synthesizes the context logic of sentences and passages, making the evaluation of students' homework more comprehensive. For example, this feature vector can help the system grasp the theme, structure, and fluency of the entire homework.

[0087] In addition, during the implementation process, the system also ensures the effectiveness and consistency of the convolutional kernel settings of each layer in the hierarchical convolutional neural network. The size of the convolutional kernel in the first layer is 3 for extracting local features at the lexical level, the size of the convolutional kernel in the second layer is 5 for focusing on sentence-level features, and the size of the convolutional kernel in the third layer is 7 for concentrating on global features at the passage level. This design enables the model to capture its unique features at different levels, ensuring the comprehensiveness and systematicness of the generation process of the multi-dimensional feature vector and laying a solid foundation for subsequent analysis and feedback.

[0088] S202, according to the multi-dimensional feature vector, combined with the subject knowledge base and logical reasoning rules, use the logical reasoning network to perform error detection and correction, where the logical reasoning network identifies logical errors in the homework and generates correction suggestions through graph reasoning algorithms and rule engines, and outputs the corrected homework feature vector;

[0089] According to the multi-dimensional feature vector, the system combines the subject knowledge base and logical reasoning rules and uses the logical reasoning network to perform error detection and correction. This process aims to identify logical errors in students' homework through intelligent means and generate corresponding correction suggestions. The logical reasoning network is constructed based on graph reasoning algorithms and rule engines, enabling the system to effectively perform logical analysis on the received homework data. Through the feature representation of nodes and edges, the system can intuitively present the relationships between various logical units in the homework, thereby identifying inconsistencies and errors. This function of error detection and correction enables the system not only to identify surface errors but also to deeply analyze the logical structure and provide comprehensive feedback.

[0090] The introduction of the logical reasoning network has effectively improved the intelligence level of homework grading, enabling the evaluation to not only stay on the scoring criteria but also delve into the level of logical reasoning. This process allows teachers and students to more clearly see the logical flaws in the homework, thereby promoting their improvement at the thinking level. Especially in some complex subject contents, students may have problems with poor connection or logical leaps, and the system can promptly identify these issues and provide targeted correction suggestions. The timeliness and accuracy of this feedback will greatly help students avoid repeating the same mistakes in future learning and improve their logical thinking ability.

[0091] Specifically, according to the multi-dimensional feature vector, knowledge matching can be carried out in combination with the subject knowledge base, and the matched subject knowledge is represented as a knowledge feature vector.

[0092] In this step, the system first matches the generated multi-dimensional feature vector with the subject knowledge base. The subject knowledge base contains information such as concepts, rules, and theories related to a specific subject, which are the basis for constructing logical reasoning. The system extracts and analyzes keywords and phrases in the homework through natural language processing technology, and then searches for corresponding knowledge points in the knowledge base for matching. Through this process, the system can generate a knowledge feature vector, reflecting the correlation between the homework and subject knowledge.

[0093] The importance of this step lies in that through knowledge matching, the system can identify which knowledge points are applied in the homework and which knowledge points may be omitted. This identification provides the basis for subsequent logical reasoning, ensuring that when the system conducts error detection, it can make judgments based on real subject knowledge. This not only makes the error detection more accurate but also helps students understand which knowledge points are correctly or incorrectly applied in their homework, thus promoting the comprehensiveness of their learning.

[0094] The system first matches the multi-dimensional feature vector (composed of vocabulary, sentence, and passage features extracted from the student's homework) with the subject knowledge base. The subject knowledge base is a large database containing rich educational resources, which includes core concepts, theorems, facts, example questions, and their corresponding explanations of the relevant subject. For example, if the student's homework involves physics, the knowledge base should contain basic principles such as Newton's laws and conservation of energy. The system will use natural language processing technology to analyze the keywords in the multi-dimensional feature vector and search for knowledge points associated with these keywords in the knowledge base.

[0095] When the matching process starts, the system first extracts keywords from the multi-dimensional feature vector, such as "acceleration", "force", etc. Then, the system uses text similarity algorithms (such as cosine similarity or Jaccard similarity) to match each keyword and identify the corresponding knowledge features. For example, if the system identifies the keyword "acceleration", the knowledge base may return formulas and laws related to acceleration, such as "Newton's second law F = m * a". This matching is ensured to be selected only when the relevance is higher than the set value by setting a threshold, which can improve the accuracy of generating the knowledge feature vector.

[0096] Finally, the system represents the matched knowledge as a knowledge feature vector. This feature vector not only contains the representation of the identified knowledge points, but also converts relevant additional information, such as definitions, formulas, and examples, into numerical forms for subsequent reasoning. These knowledge feature vectors synthesize the important knowledge points in the student's homework and provide a reliable basis for subsequent methods. For example, assuming that the matched knowledge points include "Newton's laws" and related formulas, the finally generated knowledge feature vector may be [0.7, 0.8, 0.5], where each dimension corresponds to the weight of a knowledge point.

[0097] Input the multi-dimensional feature vector and the knowledge feature vector into the logical reasoning network. The logical reasoning network initializes the feature representations of nodes and edges through graph reasoning algorithms and rule engines, where each node represents a logical unit and each edge represents a logical relationship.

[0098] In this step, the system inputs the previously generated multi-dimensional feature vector and the knowledge feature vector into the logical reasoning network together. The design of the logical reasoning network is based on graph reasoning algorithms. By establishing the feature representations of nodes and edges, the system can effectively simulate the relationships between logical units (such as propositions, sentences, etc.) in the homework. Each node here represents a logical element in the homework, and the edge represents the logical relationship between them. For example, in the student's argument, a certain node may represent a conclusion, and the edge pointing to this conclusion represents the premise supporting this conclusion.

[0099] This structured representation method greatly improves the efficiency and clarity of logical analysis. In the form of a graph, the system can more intuitively identify the logical structure of the homework and discover potential logical errors. For example, if the supporting edge of a certain node is missing, this usually indicates that the conclusion has not been reasonably argued, and thus there may be a logical flaw. By identifying these logical relationships, the system can specifically perform error detection and correction in subsequent steps and provide accurate feedback to the student.

[0100] The system inputs the generated multi-dimensional feature vector and knowledge feature vector into the logical reasoning network simultaneously. The logical reasoning network uses graph reasoning algorithms to represent the logical structure of the assignment in the form of a graph. In this graph, each node represents a logical unit in the assignment, such as a proposition or conclusion, while the edges represent the relationships between the logical units, such as causal relationships or support relationships. In this way, the logical reasoning network can clearly display the logical framework of the assignment, providing an intuitive basis for subsequent error detection.

[0101] During the initialization process of the nodes, the system performs feature representation for each logical unit. Suppose a student's assignment mentions "an object accelerates under the action of an external force". The system will split this sentence into multiple logical nodes, such as "object", "external force", "accelerate", etc., and assign a preliminary feature vector to each node to represent its relevance and features. These features can include the type of the node (such as proposition, concept, etc.), the frequency of occurrence, and the matching degree with subject knowledge. For example, the feature vector of the node "external force" may be represented as [0.9, 0.2, 0.8], corresponding to the weights and frequencies of the relevant knowledge points in the knowledge base respectively.

[0102] At the same time, the system assigns weights to each edge connecting the nodes to represent the strength of the relationships between these logical units. For example, if the logical unit "external force" supports the argument of "accelerate", the weight of the edge will be higher; conversely, if the relationship between a certain logical unit is weak, the weight of the edge will be correspondingly lower. This process enables the structure of the graph to not only display the relationships between the logical units but also reflect the strength of these relationships, providing a clear basis for subsequent logical reasoning and error finding.

[0103] According to the graph reasoning algorithm, traverse the nodes and edges in the logical reasoning network to detect logical errors, and represent the detected logical errors as error feature vectors. Among them, for each logical unit, calculate its matching degree with the subject knowledge base to identify the unmatched logical units, and for each logical relationship, calculate its compliance with the rule engine to identify the non-compliant logical relationships;

[0104] In this step, the system applies the graph reasoning algorithm to traverse the nodes and edges in the logical reasoning network one by one to identify the logical errors in the assignment. By calculating the matching degree of each node with the subject knowledge base, the system can determine whether a certain logical unit meets the subject requirements. For example, if a node corresponds to a logical claim and the associated knowledge features do not support this claim, the system will mark it as unmatched. At the same time, the system also checks the compliance of the edges to ensure that the logical relationships comply with the predefined logical rules.

[0105] This process is a crucial step in ensuring the logical rigor of assignments. Through systematic automated logical checks, teachers can more easily identify and focus on logical errors when grading assignments, thus saving time and improving efficiency. For students, timely logical feedback can help them recognize their thinking loopholes, and then improve their thinking patterns and enhance their logical reasoning abilities. This error detection process that combines subject knowledge and logical rules will greatly enhance students' learning effects.

[0106] The system uses graph reasoning algorithms to traverse all nodes and edges in the logical reasoning network to identify logical errors in the assignment. To this end, the system will compare each logical unit one by one and evaluate its matching degree with the subject knowledge base. Suppose there is a logical unit in the network that is "everyone should abide by the law". The system will compare this unit with the definitions and principles regarding law compliance in the knowledge base. If the content cited by this logical unit is not supported in the knowledge base, then this logical unit will be marked as unmatched.

[0107] For each edge, the system will analyze the compliance degree of the logical relationship. For example, if an edge represents "because of abiding by the law, so one should be protected", and the logical unit "abiding by the law" fails to find corresponding supporting evidence in the knowledge base, then the compliance degree of this edge will also be marked as low. Subsequently, the system generates an error feature vector for each logical unit and logical relationship marked as low, recording its corresponding matching degree and compliance degree. This error feature vector will serve as the basic data for generating correction suggestions in subsequent steps.

[0108] To optimize the accuracy of error detection, the system may also consider the context relationship between logical units. For example, if a certain logical unit appears multiple times in the entire assignment but expresses contradictory views in different contexts, the system will further deepen the identification of logical errors through context analysis. At this time, the system will form a comprehensive error feature vector for each logical unit, identifying the lack of logic and consistency, providing detailed basis for subsequent steps.

[0109] According to the error feature vector, use the rule engine to generate correction suggestions. Among them, for unmatched logical units, generate replacement suggestions, and for non-compliant logical relationships, generate adjustment suggestions;

[0110] In this step, the system uses the error feature vector to generate correction suggestions. Based on the previous logical error detection results, the rule engine will analyze each unmatched logical unit and logical relationship and generate corresponding correction suggestions. For those logical units identified as unmatched, the system will put forward specific replacement suggestions; for non-compliant logical relationships, the system will suggest how to make adjustments to improve the logical rationality. For example, if a certain logical unit expresses a view without support, the system will recommend relevant supporting arguments or theories.

[0111] This process enables the system not only to stay at error recognition but also to provide students with specific and actionable improvement suggestions. The personalization and pertinence of this feedback ensure that when students revise their assignments, they can make targeted improvements, truly achieving the internalization and application of knowledge. The generation of such correction suggestions not only improves the quality of students' assignments but also provides teachers with in-depth insights into students' comprehension abilities, facilitating differentiated guidance for different students.

[0112] The system inputs data into the rule engine using the generated error feature vectors to generate correction suggestions. Based on the analysis logic of the rule engine, it will provide specific alternative suggestions for each mismatched logical unit according to the identified logical errors. For example, if the logical unit "The existence of law guarantees people's freedom" is marked as mismatched, the rule engine will recommend replacing it with "The existence of law protects people's rights" according to the supporting theories in the knowledge base, thereby enhancing the logical consistency and accuracy.

[0113] For non-conforming logical relationships, the system will also generate adjustment suggestions. For example, assume that the logical relationship represented by an edge is "Because the law exists, so people abide by the law". If this causal relationship is not supported in the knowledge base, the rule engine may suggest changing it to "Because people abide by the law, therefore the law is maintained". This adjustment not only makes the logical relationship more rigorous but also ensures that the connections between logical units are more in line with the actual situation.

[0114] When generating correction suggestions, the system will also assign a credibility score to each suggestion, reflecting the logical support degree of the suggestion and the effectiveness of implementing the suggestion. Assume that the rule engine generates multiple replacement suggestions for a certain logical unit. The system will give priority to recommending the suggestion with the highest credibility score to the user to ensure that the corrected logical relationship is clear and easy to understand. In addition, these correction suggestions will be recorded for forming the subsequent updated logical feature vectors.

[0115] Apply the correction suggestions to the logical inference network to generate the corrected logical feature vectors as the assignment feature vectors.

[0116] In this step, the system applies the correction suggestions generated in step 4 to the logical inference network. Specifically, the system will adjust the logical units and their corresponding logical relationships according to the identified logical errors and the generated correction suggestions. This process includes modifying the content of nodes or adding new supporting arguments, as well as adjusting the edges between nodes to strengthen the logical relationship. This operation not only makes the structure of the logical inference network more reasonable but also effectively updates the logical feature vectors to reflect the modified assignment status.

[0117] The importance of this step lies in that by applying the correction suggestions, the system can generate more accurate and consistent logical feature vectors, thus improving the quality of the assignment. This process not only helps students understand and correct their own logical errors, but also provides a clear and accurate evaluation basis for teachers, thereby ensuring the effectiveness of the review results. The corrected logical feature vectors, as part of the assignment feature vectors, can more comprehensively reflect students' understanding and expression abilities, laying a solid foundation for subsequent grading and feedback.

[0118] The system applies the generated correction suggestions to the logical reasoning network to generate corrected logical feature vectors. Specifically, the system will identify the logical units that need to be modified and update their contents and the weights of the edges. Taking a specific example, if the logical unit "The law should safeguard citizens' rights and interests", which was previously marked as "mismatched", is suggested to be replaced with "The law is a tool for protecting citizens' rights and interests", the system will directly update this node in the logical reasoning network.

[0119] During the update process, the system will not only replace the content of the logical unit, but also re-evaluate the edges with other logical units according to the new content and update the weights of the edges. This process ensures the overall consistency and coherence of the logical reasoning network. For example, if the replaced logical unit needs to be connected to multiple supporting arguments, the system will adjust the weights of the edges according to the new support relationships to reflect this change in logical support.

[0120] Finally, after the above adjustments, the system will generate corrected logical feature vectors, which, as part of the assignment feature vectors, are used for subsequent grading and feedback. These corrected logical feature vectors will contain the assignment data with enhanced logic, ensuring that all information can be effectively utilized during the logical reasoning and error detection processes. This feature vector not only provides an accurate reflection of students' assignments for teachers, but also provides specific references and guidance for students' subsequent improvement, making the entire evaluation process more comprehensive and in-depth.

[0121] S203, according to the assignment feature vectors, in combination with the teacher's grading criteria and the student ability evaluation indicators, use the dynamic grading and ability evaluation model to conduct grading and ability analysis, wherein the dynamic grading and ability evaluation model dynamically adjusts the grading weights and ability evaluation parameters through multi-task learning and an adaptive weight allocation mechanism to ensure the accuracy of the grading results and the comprehensiveness of the ability evaluation results;

[0122] According to the assignment feature vector, combined with the teacher's grading criteria and the student ability assessment indicators, a dynamic grading and ability assessment model is used for grading and ability analysis. This process first requires the teacher to define corresponding evaluation dimensions according to the grading criteria for a specific course, such as accuracy, integrity, and logic. The student ability assessment indicators may include knowledge mastery, logical thinking ability, and expression ability. These criteria and indicators will be transformed into specific grading rules and assessment parameters, forming two vectors, which are used for grading and ability analysis respectively. Subsequently, the dynamic grading and ability assessment model will combine the assignment feature vector, the grading weight vector, and the ability weight vector, and use an adaptive weight allocation mechanism to dynamically adjust the grading dimensions and ability assessment parameters to ensure the fairness and comprehensiveness of grading and assessment.

[0123] The implementation of this step effectively guarantees the accuracy of grading and the comprehensiveness of ability assessment. By combining the teacher's subjective criteria with the students' actual performance and dynamically adjusting the weights of grading and ability assessment, it can truly reflect the real progress and logical thinking ability of each student in the learning process. Such dynamic assessment can not only timely detect the difficulties encountered by students in the learning process, but also provide targeted teaching suggestions for teachers to help them adjust teaching strategies and improve teaching quality. At the same time, students can also be aware of their own strengths and weaknesses through clear assessment feedback, so as to make corresponding improvements in subsequent learning.

[0124] Specifically, the grading rules can be loaded according to the teacher's grading criteria, and the grading rules are encoded as a grading weight vector;

[0125] In the first step, the system will load the corresponding grading rules according to the grading criteria formulated by the teacher. Teachers usually set grading dimensions according to the specific requirements of the assignment. For example, for a paper, it may include the accuracy of the content, the rigor of the logic, the expression ability of the language, etc. Through the refined analysis of these grading dimensions, the system transforms the grading criteria into specific grading rules and encodes them in the form of a vector. For example, the accuracy of the content may account for 40% of the total score, the rigor of the logic accounts for 30%, and the expression ability of the language accounts for 30%. The corresponding grading weight vector will be [0.4, 0.3, 0.3].

[0126] This process provides a clear benchmark for subsequent dynamic grading, ensuring the standardization and fairness of grading. By clearly expressing the teacher's grading criteria in a vectorized manner, the system can clearly follow these criteria during the grading process, thereby improving the transparency and consistency of grading. For teachers, this can reduce subjective biases in the grading process, while for students, it provides clear expectations to help them understand the aspects that need to be focused on in the assignment, so as to make targeted improvements.

[0127] The system first needs to collect detailed information about the grading criteria from teachers. Teachers may formulate a series of grading dimensions based on course objectives and assignment types. For example, they may grade the accuracy, logic, and completeness of an assignment. When collecting the criteria, the system can use online questionnaires or interviews to enable teachers to clearly express the importance and specific requirements of each grading dimension. For example, for science assignments, teachers may emphasize that accuracy (such as the correctness of formulas and the precision of experimental data) accounts for 40% of the weight, while logic and completeness each account for 30%. This information will be integrated into the system.

[0128] Next, the system will analyze the grading criteria input by teachers to determine the weight of each dimension. At this time, the system can adopt statistical analysis methods to quantify the collected grading criteria and calculate the specific proportion of each dimension. To ensure the transparency of this process, the system can generate a visual chart of the grading weights to show the proportions of each grading dimension. This visual tool not only helps teachers verify the input grading rules but also serves as a reference for subsequent evaluations. For example, teachers can see from the chart that accuracy accounts for 40%, and logic and completeness each account for 30%, thus confirming the grading rules.

[0129] Finally, the system will encode these grading rules into a vector. This vector will serve as the basis for the subsequent grading process. Inside the system, this weight vector may be represented as [0.4, 0.3, 0.3], where each value represents the corresponding grading dimension. For ease of subsequent use, the system will store this weight vector in the database to ensure that it can be quickly retrieved and used during subsequent grading.

[0130] Load the ability evaluation parameters according to the student ability evaluation indicators and encode the ability evaluation parameters into an ability weight vector;

[0131] The system will load the ability evaluation parameters according to the student ability evaluation indicators. The indicators for student ability evaluation may include knowledge mastery, logical thinking ability, and expression ability, etc. Teachers need to set the goals and standards for these ability evaluations according to the specific course and the learning situation of students. These ability evaluation indicators will be converted into specific parameters and also encoded in the form of a vector. For example, if knowledge mastery accounts for 50% of the total evaluation, logical thinking ability accounts for 30%, and expression ability accounts for 20%, the corresponding ability weight vector will be [0.5, 0.3, 0.2].

[0132] By representing the student ability assessment criteria in vector form, the system can clearly consider the ability growth and deficiencies of each student during the grading process. This provides teachers with a more comprehensive basis to help them better formulate personalized teaching plans for effective differentiated instruction. At the same time, students can also identify their deficiencies in logical thinking, knowledge mastery, or expression ability through understanding their own ability assessments and thus make targeted improvements.

[0133] The system first needs to clarify the specific indicators for student ability assessment. These ability assessment indicators usually include knowledge mastery, logical thinking ability, and expression ability, etc. When formulating these indicators, teachers will consider students' performance in class, the quality of homework completion, and exam scores. For example, a teacher may consider that knowledge mastery accounts for 50% of the weight, logical thinking ability accounts for 30%, and expression ability accounts for 20%. At this stage, the system can gather teachers' insights and opinions on each ability dimension through questionnaires or teacher meetings to ensure that the collected indicators truly reflect students' comprehensive abilities.

[0134] Next, the system will analyze the collected ability assessment criteria and quantify each assessment indicator. This usually involves scoring students' performance in different ability dimensions and finally determining the proportion of each dimension in the total assessment. To provide more intuitive data, the system can generate charts or reports showing the specific weights of each ability dimension. For example, it can show that the weight of knowledge mastery is 50%, the weight of logical thinking ability is 30%, and the weight of expression ability is 20%. This not only facilitates teachers' review but also helps them adjust the weights of the indicators in a timely manner.

[0135] Finally, through the analysis of the ability assessment parameters, the system will generate an ability weight vector for subsequent ability analysis. This weight vector can be expressed as [0.5, 0.3, 0.2], and each value corresponds to the proportion of its corresponding ability. The system stores this parameter vector in the database so that it can be quickly called during ability assessment and ensure the coherence of the entire set of ability analysis processes.

[0136] Input the homework feature vector, the grading weight vector, and the ability weight vector into the multi-task learning model, and the multi-task learning model initializes the weights of the grading task and the ability assessment task through an adaptive weight allocation mechanism;

[0137] The system inputs the job feature vector, scoring weight vector, and ability weight vector into a multi-task learning model. The multi-task learning model is designed to handle multiple tasks simultaneously, including the scoring task and the ability assessment task. The system uses an adaptive weight allocation mechanism to initialize the weights of each task. By analyzing the relationships between different tasks, it reasonably allocates the weights of the scoring and assessment tasks on the premise of ensuring efficient and accurate output results. Specifically, the multi-task learning model extracts features through a shared layer, enabling the scoring and ability assessment tasks to benefit from the same feature information.

[0138] The implementation of this step can effectively improve the accuracy and reliability of scoring and ability assessment. By using multi-task learning, different tasks can promote each other. By leveraging the shared underlying features, the system can have a more comprehensive understanding of each task. At the same time, the dynamic weight allocation mechanism ensures that in different situations, the results of scoring and ability assessment are more in line with the actual situation. For example, in some cases, a student's logical ability may be more important, and the system can adjust the weight allocation accordingly to reflect the real learning situation.

[0139] The system inputs the job feature vector, scoring weight vector, and ability weight vector into a designed multi-task learning model to handle the scoring and ability assessment tasks simultaneously. The structure of this model usually includes a shared layer and specific task layers. Among them, the job feature vector contains multi-dimensional information demonstrated by students in their homework, such as the fluency and logic of language expression. By analyzing these features, the system can identify the performance characteristics of students, giving full play to the learning ability of the model.

[0140] To initialize the weights of the scoring task and the ability assessment task, the multi-task learning model dynamically adjusts through an adaptive weight allocation mechanism. In this process, the system evaluates the importance of each task based on the performance of the input features. For example, assume a student performs excellently in logic but is slightly lacking in expression ability. The model will correspondingly increase the weight of logic in scoring. In this way, the system can ensure the accuracy of scoring in different situations, making the ability assessment of each student more in line with the actual situation.

[0141] In addition, during the implementation process, the system also uses historical data for training to ensure that the multi-task learning model can accurately reflect the relationships between different tasks during initialization. This means that through past performance data, the system can learn the mutual influences between various scoring dimensions and ability assessment indicators, and then optimize the connection of tasks. The structural design of sharing information among different tasks helps improve the comprehensive accuracy and generalization ability of the model, enabling it to perform stably in different scenarios.

[0142] Among them, according to the job feature vector, the shared layer of the multi-task learning model is used for feature extraction. Among them, through the fully connected network and the non-linear activation function, high-order feature representations are extracted;

[0143] The system uses the shared layer of the multi-task learning model to extract high-order features. The purpose of the design of the shared layer is to make full use of the information in the job feature vector, and more complex high-order feature representations are extracted through the fully connected network and the non-linear activation function. The fully connected network linearly combines the input features to increase the interaction between features, while the non-linear activation function introduces non-linearity, enabling the model to capture more complex feature patterns. For example, the ReLU (Rectified Linear Unit) activation function can effectively improve the non-linear ability of the model, making the expression of high-order features more flexible and rich.

[0144] The process of extracting high-order feature representations lays a solid foundation for subsequent scoring and ability analysis. By performing high-order transformation on the features, the system can fully explore the potential patterns and relationships in the data, making the results of subsequent scoring and ability analysis more accurate. In addition, this process enhances the learning ability of the model, enabling it to make reasonable responses when facing different types of assignments and evaluations, providing double support for both teachers and students.

[0145] The system uses the shared layer of the multi-task learning model to perform high-order feature extraction on the input job feature vector. The design of the shared layer aims to make full use of the shared information between different tasks to improve the generalization ability of the model. First, the system processes the job feature vector through the fully connected network to achieve linear combination of features. The fully connected network connects multiple layers of neurons to perform weighted summation on the input features, generating a more complex feature combination, enabling the model to capture deeper information.

[0146] Next, the system applies a non-linear activation function, such as ReLU (Rectified Linear Unit) or Sigmoid, to the output of the fully connected network, aiming to introduce non-linearity. This process enables the model to better fit the complex patterns in the data, thus enhancing the expression ability of high-order features. For example, assume that a student performs particularly well in logical reasoning. Through this process, the system will be able to more clearly identify the superiority of the student's logical thinking ability and emphasize it in subsequent analysis.

[0147] Finally, the extracted high-order features will be used for subsequent scoring and ability analysis tasks. Through the high-order features, the system will enhance the understanding of the overall performance of students, enabling a more accurate reflection of the strengths and weaknesses of each student in subsequent scoring and ability assessment. This feature extraction process is not only the basis for scoring and assessment but also provides data support for subsequent personalized feedback, ensuring that the system can provide precise guidance according to the specific situation of students.

[0148] According to the high-order feature representation and the scoring weight vector, the scoring task-specific layer is used for scoring, where the accuracy, integrity, and logic of the assignment are scored to generate a preliminary scoring result, and according to the adaptive weight allocation mechanism, the weights of each scoring dimension are dynamically adjusted to generate the final scoring result;

[0149] The system will use the high-order feature representation and the scoring weight vector extracted to score the assignment using the scoring task-specific layer. The scoring task-specific layer focuses on feature analysis for scoring dimensions such as accuracy, integrity, and logic. The system will score based on the input features through a specific evaluation model to generate a preliminary scoring result. These scores are calculated not only based on the weights of the scoring dimensions but also in combination with the adaptive weight allocation mechanism to ensure that the weights of each dimension are appropriate in different situations, thus more accurately reflecting the actual performance of students.

[0150] This process is a crucial link in ensuring the fairness and reliability of the scoring results. By dynamically adjusting the weights of each scoring dimension, the system can better adapt to the specific performance of students and timely adjust the degree of emphasis on different dimensions. For example, if a student shows strong expression ability but lacks logic, the scoring system can correspondingly reduce the weight of logic in the overall scoring to ensure that the final score is more authentic and personalized. This not only improves the quality of scoring but also provides targeted feedback information for teachers to help them make adjustments in teaching.

[0151] The system combines the high-order feature representation with the scoring weight vector and scores the assignment through the scoring task-specific layer. This layer specifically analyzes and evaluates the scoring dimensions (accuracy, integrity, and logic). The system uses the information in the high-order feature representation to independently score each scoring dimension of the assignment and generates a preliminary scoring result. For example, when evaluating the logic of the assignment, the system will give corresponding scores based on the effectiveness and coherence of the student's argument in the assignment.

[0152] With the generation of the preliminary scores, the system will use an adaptive weight allocation mechanism to dynamically adjust the weights of each scoring dimension. For example, if the system finds that a certain student's assignment is particularly outstanding in terms of logic at this time, but mediocre in terms of accuracy, the system may correspondingly reduce the weight of accuracy in the total score and emphasize the influence of logic. This dynamic adjustment ensures that the final score can more accurately reflect the comprehensive ability performance of the student, rather than simply relying on fixed scoring rules.

[0153] Finally, after the above calculations, the system will generate a comprehensive score, which is fed back to the teacher and the student. This score is not only a quantitative assessment of the student's performance, but also an important basis for subsequent educational guidance. Teachers can formulate teaching plans according to the scores of the students to help students improve in their weak links. At the same time, students can also clarify their learning directions through these feedbacks, and then adjust their learning strategies to enhance their learning effects.

[0154] Based on the high-order feature representation and the ability weight vector, the ability assessment task-specific layer is used for ability analysis. Among them, the knowledge mastery, logical thinking ability and expression ability of the student are evaluated to generate a preliminary ability assessment result. According to the adaptive weight allocation mechanism, the weights of each ability assessment parameter are dynamically adjusted to generate the final ability assessment result.

[0155] The system combines the high-order feature representation with the ability weight vector and uses the ability assessment task-specific layer to analyze the student's ability. This layer is specifically designed to evaluate multiple dimensions such as knowledge mastery, logical thinking ability and expression ability. Based on the input high-order features, the system scores each ability dimension separately to generate a preliminary ability assessment result. Through the adaptive weight allocation mechanism, the system can dynamically adjust the weights of each assessment parameter according to the student's performance in different abilities, ensuring that the final assessment result can comprehensively and accurately reflect the student's ability level.

[0156] This ability analysis process provides teachers with a comprehensive basis for assessing students' abilities, enabling them to effectively understand the learning status of students. Thus, teachers can formulate targeted tutoring plans based on these analysis results to help students improve in their weak links. At the same time, students can also obtain feedback from this, understand their strengths and weaknesses, and then set more appropriate learning goals to promote the improvement of learning effects as a whole.

[0157] The system will rely on high-order feature representations and ability weight vectors to analyze students' abilities through task-specific layers for ability assessment. This layer focuses on evaluating students' performance in multiple dimensions such as knowledge mastery, logical thinking ability, and expression ability. The system will independently score each ability based on the high-order features demonstrated by students in their assignments to obtain preliminary ability assessment results. For example, when evaluating knowledge mastery, the system will refer to students' understanding of course content, the correct application of relevant concepts, and the corresponding examples provided in the assignments.

[0158] Once the preliminary ability assessment results are generated, the system will use an adaptive weight allocation mechanism to dynamically adjust the weights of each evaluation parameter. This mechanism aims to ensure that in different situations, the assessment results can more accurately reflect students' actual abilities. For example, if a student performs outstandingly in logical thinking ability, the system may increase the weight of this dimension while correspondingly reducing the weights of other dimensions to ensure that the final result is more realistic and in line with educational goals.

[0159] Finally, the system will synthesize the final assessment results of each ability dimension and feedback them to teachers and students. These assessment results not only help teachers understand students' performance in different abilities but also provide clear improvement directions for students. Teachers can formulate personalized teaching plans based on these assessment results to help students remedy their deficiencies in certain abilities. And students can, through the analysis of these results, recognize their strengths and weaknesses, thereby more effectively planning their learning paths to improve their overall abilities.

[0160] S204, according to the scoring results and ability assessment results, use a generative adversarial network to generate a personalized feedback report, where the generative adversarial network generates multi-modal feedback content including text, charts, and speech through a multi-modal generator and a context-aware discriminator.

[0161] According to the scoring results and ability assessment results, use a generative adversarial network to generate a personalized feedback report. The generative adversarial network includes a multi-modal generator and a context-aware discriminator. The multi-modal generator automatically generates multi-modal feedback content containing text, charts, and speech by analyzing the final scoring results and ability assessment results. The task of this generator is to create valuable feedback based on the input information to help students understand the meaning of their learning outcomes and how to further improve. The context-aware discriminator is responsible for checking the consistency and quality of the generated feedback content to ensure the accuracy and logic of the content when received by students, so as to guarantee the effectiveness of the final feedback.

[0162] The implementation of this step will greatly enhance the personalization and interactivity of students' homework feedback. In traditional grading methods, the feedback content is often single and difficult to comprehensively reflect students' learning situations. Through generative adversarial networks, students can obtain richer feedback, including text descriptions, chart visualizations, and voice explanations, which helps students more intuitively understand the results. For example, charts can help students identify the trends of their grades, and voice feedback diversifies the information transmission method, improving students' acceptance and comprehension. In addition, this multi-modal feedback mechanism also provides teachers with more specific bases for teaching adjustments and can provide practical guidance according to the needs of different students.

[0163] Specifically, the final grading result and the final ability assessment result can be input into the multi-modal generator, which generates text, chart, and voice feedback content respectively through a multi-task learning mechanism;

[0164] At this stage, the system will pass the final grading result and the ability assessment result as inputs to the multi-modal generator. The multi-modal generator adopts a multi-task learning mechanism and can process multiple types of outputs simultaneously. Through this mechanism, the system can not only generate detailed text feedback but also generate chart and voice feedback, providing comprehensive assessment information for students. The design of the generator enables it to extract diverse feedback from the same input, ensuring the consistency and relevance of all feedback in terms of content.

[0165] The core of this process lies in the diversity and accessibility of information. Rich feedback forms help students better understand their learning effects. By presenting the grading result and the ability assessment result in multiple dimensions, students can analyze their strengths and weaknesses from different perspectives, which helps in formulating more specific improvement plans. In addition, this diverse feedback form can improve students' learning interest and participation.

[0166] The system first organizes the final grading result (such as specific scores for logic, accuracy, expression ability, etc.) and the final ability assessment result (such as knowledge mastery, logical thinking ability) to ensure the integrity and accuracy of the data. This data organization process includes summarizing the scores for each grading dimension and classifying the levels of ability assessment. For example, the grading result may include "the logic score is 85 points, and the accuracy score is 70 points", while the ability assessment result may indicate that "the logical thinking ability is at a medium level and the knowledge mastery is good". This information will be passed as an input to the multi-modal generator.

[0167] Next, the multimodal generator utilizes its multi-task learning mechanism to process information and generate feedback content. At this stage, the generator will process all input information in parallel and generate text, charts, and voice feedback content respectively. For example, for the input of "logical score is 85 points", the generator may generate a text summary such as "Your logical reasoning ability is excellent". At the same time, it may also generate a chart to visually display the distribution of various scores. For example, a bar chart is used to show the scores of different rating dimensions. Finally, the generator will also synthesize a voice feedback to improve the diversity and interactivity of information transmission.

[0168] Finally, the generated feedback content will be output in a structured manner, including text descriptions, chart files, and voice audio files. For example, the system will label the corresponding evaluation dimensions for each feedback form, making it clear at a glance for teachers and students. This can ensure that students can quickly find the information they need when referring to the feedback, promoting personalized and targeted learning. This way of generating multimodal feedback not only improves the quality of feedback but also provides students with a diverse learning experience.

[0169] Among them, according to the final scoring result and the final ability assessment result, the text generation module is used to generate text feedback content. Among them, the final scoring result is summarized to generate a text evaluation, and the final ability assessment result is explained to generate improvement suggestions. The text feedback content is generated by combining the text evaluation and the improvement suggestions;

[0170] In this step, the system focuses on using the text generation module to create specific text feedback. The text feedback includes summarizing and explaining the scoring result and ability assessment result of the student's assignment, which is detailed and targeted enough. For example, the system will analyze the scoring result, generate a summary of the student's performance in different scoring dimensions, and improvement suggestions based on the ability assessment result.

[0171] The generated text feedback aims to help students identify their strengths and weaknesses and provide clear directions for improvement. For example, clear summaries and suggestions can enable students to have a clear goal in subsequent learning, thus effectively improving learning effects. This customized text feedback also provides important information for teachers to help them adjust teaching plans to meet the learning needs of students.

[0172] In this step, the system will utilize the text generation module to analyze and process the previous input of the final scoring results and ability assessment results. First, the system will conduct a detailed assessment summary of the final scoring results. For example, if a student's logical score is 85 points and the accuracy score is 70 points, the system will generate a text evaluation like "Your logical reasoning ability score is 85 points, which is very excellent, but the score in terms of accuracy is relatively low, only 70 points, indicating that you may have errors when applying knowledge in practice." This information aims to help students identify their strengths and weaknesses and make improvements based on this.

[0173] Secondly, the system will also analyze the ability assessment results and provide specific improvement suggestions according to the performance in each ability dimension. If the ability assessment shows that "the knowledge mastery level is good", the system may generate a statement such as "Consider participating in more extracurricular trainings and reviews to improve the mastery of knowledge, especially in the fields of logical reasoning and problem-solving." These suggestions are all based on real data analysis and reflect the specific areas where students urgently need to improve in the learning process.

[0174] Finally, the system combines the above text evaluation and improvement suggestions to generate a complete text feedback. For example, the output text feedback may be: "Your logical reasoning ability is outstanding, but there are certain deficiencies in accuracy. It is recommended to strengthen the review of relevant knowledge to improve the overall comprehension ability." Such text feedback is not only simple and clear but also can effectively guide the students' next learning direction. Through this process, the system provides personalized and detailed learning feedback to help them make more targeted self-improvements in future learning.

[0175] According to the final scoring results and final ability assessment results, use the chart generation module to generate chart feedback, where the final scoring results are visualized to generate the first chart, the final ability assessment results are visualized to generate the second chart, and the chart feedback content is generated by combining the first chart and the second chart;

[0176] In this stage, the system uses the chart generation module to create visual feedback, aiming to display the final scoring results and ability assessment results in the form of charts. The design of the chart feedback enables students to quickly understand their performance in an intuitive way, including the specific scores in each scoring dimension and ability assessment.

[0177] The chart feedback can not only help students quickly identify the strengths and weaknesses of their performance but also enable them to obtain information visually and enhance their analytical ability. The visualization of the charts makes the information easier to understand. Especially for students who are not so proficient in data processing, charts can attract more attention and provide more help than text. In addition, this form of feedback is convenient for teachers to quickly present the overall performance of students in class, improving teaching efficiency.

[0178] In this step, the system will pay special attention to the visualization of data and display the final scoring results and ability assessment results through the chart generation module. First, by analyzing the final scoring results, the system generates the first chart, usually in the form of a bar chart or a line chart. For example, assuming that a student's scores in logical thinking, accuracy, and expression ability are 85 points, 70 points, and 75 points respectively, the system will display these three scores in the form of a bar chart, enabling the student to intuitively compare the differences in the performance of various abilities.

[0179] Next, the system will visualize the final ability assessment results to generate the second chart. For example, the system may use a pie chart to show the student's mastery in different knowledge areas (such as logical thinking ability, knowledge mastery, and expression ability). Assuming the ability assessment results are "medium logical thinking ability, good knowledge mastery, and excellent expression ability", the system will display the proportions of these data through a pie chart to help the student understand the share of each area in the overall ability.

[0180] Finally, the system will combine these two charts to generate a comprehensive chart feedback content. By integrating the charts with the analysis results, the system will output a visual report with text explanations for the student to download or view. This kind of feedback not only enhances the intuitiveness of information but also strengthens the student's understanding of their own learning status, enabling them to evaluate their learning performance from multiple perspectives and then make effective improvements.

[0181] According to the final scoring results and the final ability assessment results, use the voice generation module to generate voice feedback content, where the text evaluation is synthesized into voice to generate a voice evaluation, the improvement suggestions are synthesized into voice to generate voice suggestions, and the voice evaluation and improvement suggestions are combined to generate voice feedback content;

[0182] In this stage, the system converts the previously generated text feedback content into voice through the voice generation module to provide the student with an auditory feedback experience. The system first synthesizes the text evaluation into voice to generate a voice evaluation, and then, for the improvement suggestions, it also synthesizes them into voice to form voice suggestions. Finally, the system combines the two to generate a complete voice feedback content, which is convenient for the student to study and reflect in any environment.

[0183] The introduction of voice feedback provides students with a flexible learning method, enabling them to still obtain feedback information when they are busy or need to rest. This not only improves the acceptability of feedback but also deepens the student's memory of the feedback content. In addition, voice feedback provides another way for students with visual fatigue to receive learning information, enhancing the learning experience.

[0184] The system converts the previously generated text feedback into voice feedback. First, the system performs speech synthesis on the text evaluation. For example, assuming the text evaluation is "Your logical reasoning performance is good, and the score is 85 points", the system will use natural language processing technology to generate the corresponding voice content. Through high-quality speech synthesis, students can hear this feedback. In this way, students can focus more on the content when listening to the feedback instead of reading it word by word.

[0185] Then, the system performs speech synthesis on the improvement suggestions. Such as "Please try to enhance accuracy in the next assignment, which can be improved by doing more exercises", the system will perform speech synthesis on this part of the content to generate voice suggestions. In this way, students can not only hear the specific improvement suggestions but also feel an interactive experience during the learning process, adding an affinity to the feedback.

[0186] Finally, the system combines the entire voice evaluation and voice suggestions into a complete voice feedback. This voice feedback may form a coherent audio file, covering the summary of the student's score and personalized improvement suggestions. Such voice feedback content is convenient for students to listen to in different situations, such as during breaks, exercise, or any other circumstances to obtain learning feedback, enhancing the flexibility and convenience of learning.

[0187] Input the text feedback content, chart feedback content, and voice feedback content into the context-aware discriminator. The context-aware discriminator, through the context-aware mechanism, conducts consistency checks and optimizations on the feedback content, and finally outputs multi-modal feedback content including text, charts, and voice.

[0188] In the last step, the system passes the previously generated text feedback, chart feedback, and voice feedback content to the context-aware discriminator for consistency checks and content optimizations. The context-aware discriminator will analyze the relationships between these feedback contents to ensure that the information is logical at all levels and ensure that each part of the feedback plays a role in the overall evaluation.

[0189] This process guarantees the overall quality of the feedback and helps to eliminate possible conflicts or inconsistencies between the feedback contents. Through this consistency check, the system can improve the reliability and effectiveness of the feedback report, making the finally output feedback content more professional and accurate, thereby enhancing the learning experience of students. Good feedback also encourages students to draw on the suggestions in future assignments and make positive progress.

[0190] In the last step, the system will integrate the previously generated text feedback, chart feedback, and voice feedback content and input it into the context-aware discriminator for consistency checking and content optimization. First, the discriminator will analyze various types of feedback content to ensure their consistency in content and theme. For example, if the text feedback emphasizes the outstanding performance of students in logical reasoning, then the corresponding logical score in the chart should also show a high score to avoid contradictions and confusion in information.

[0191] Secondly, the discriminator will check the language and expression in the feedback content. The system will perform semantic analysis on the text, chart, and voice feedback to ensure the consistency of all feedback. For example, if the text feedback suggests that students improve accuracy, then this suggestion should be clearly conveyed in the voice feedback, and the chart should also be able to intuitively reflect the deficiency in the accuracy score. This ensures that the feedback not only has a rich variety of forms but also forms a joint force in content, guiding students to obtain clear and consistent learning guidance.

[0192] Finally, after consistency detection and necessary content optimization, the system will finally output the integrated multi-modal feedback content, covering text, charts, and voice. This comprehensive feedback will be presented in a structured manner, enabling students and teachers to easily understand and use it. The final feedback content not only provides a comprehensive assessment but also stimulates students' learning interest through multi-modal means, helping them make continuous progress in subsequent learning.

[0193] It can be seen that according to the homework data submitted by students, a multi-dimensional feature extraction model is used to extract features from the text content, logical structure, and language expression; according to the multi-dimensional feature vectors, combined with the subject knowledge base and logical reasoning rules, a logical reasoning network is used for error detection and correction; according to the homework feature vectors, combined with the teacher's grading criteria and students' ability assessment indicators, a dynamic grading and ability assessment model is used for grading and ability analysis; according to the grading results and ability assessment results, a generative adversarial network is used to generate personalized feedback reports, so as to be able to efficiently and accurately identify and correct various errors in the homework, provide personalized feedback, and improve the quality and effect of online education.

[0194] Another embodiment of the present invention provides an AI-based online homework correction system. Refer to Figure 3 , the system may include:

[0195] An extraction module 301, configured to extract features from the text content, logical structure, and language expression by using a multi-dimensional feature extraction model according to the homework data submitted by students, wherein the multi-dimensional feature extraction model extracts semantic features from three levels of vocabulary, sentence, and discourse through a hierarchical convolutional neural network and a self-attention mechanism, and generates multi-dimensional feature vectors;

[0196] The correction module 302 is configured to perform error detection and correction using a logical reasoning network based on the multi-dimensional feature vector, in combination with a subject knowledge base and logical reasoning rules. The logical reasoning network identifies logical errors in the assignment and generates correction suggestions through a graph reasoning algorithm and a rule engine, and outputs a corrected assignment feature vector.

[0197] The analysis module 303 is configured to perform grading and ability analysis using a dynamic grading and ability evaluation model based on the assignment feature vector, in combination with a teacher grading standard and a student ability evaluation index. The dynamic grading and ability evaluation model dynamically adjusts grading weights and ability evaluation parameters through multi-task learning and an adaptive weight allocation mechanism to ensure the accuracy of the grading result and the comprehensiveness of the ability evaluation result.

[0198] The generation module 304 is configured to generate a personalized feedback report using a generative adversarial network based on the grading result and the ability evaluation result. The generative adversarial network generates multi-modal feedback content including text, charts, and speech through a multi-modal generator and a context-aware discriminator.

[0199] It can be seen that based on the assignment data submitted by students, a multi-dimensional feature extraction model is used to extract features from the text content, logical structure, and language expression; based on the multi-dimensional feature vector, in combination with a subject knowledge base and logical reasoning rules, a logical reasoning network is used to perform error detection and correction; based on the assignment feature vector, in combination with a teacher grading standard and a student ability evaluation index, a dynamic grading and ability evaluation model is used to perform grading and ability analysis; based on the grading result and the ability evaluation result, a generative adversarial network is used to generate a personalized feedback report, thereby enabling efficient and accurate identification and correction of various errors in the assignment, providing personalized feedback, and improving the quality and effect of online education.

[0200] An embodiment of the present invention further provides a storage medium storing a computer program, where the computer program is configured to execute the steps in any one of the above method embodiments when running.

[0201] Specifically, in this embodiment, the above storage medium may be configured to store a computer program for performing the following steps:

[0202] S201, based on the assignment data submitted by students, use a multi-dimensional feature extraction model to extract features from the text content, logical structure, and language expression. The multi-dimensional feature extraction model extracts semantic features from three levels of vocabulary, sentence, and discourse through a hierarchical convolutional neural network and a self-attention mechanism, and generates a multi-dimensional feature vector.

[0203] S202. Based on the multi-dimensional feature vector, in combination with the subject knowledge base and logical reasoning rules, use a logical reasoning network to perform error detection and correction. Among them, the logical reasoning network uses a graph reasoning algorithm and a rule engine to identify logical errors in the homework and generate correction suggestions, and outputs a corrected homework feature vector.

[0204] S203. Based on the homework feature vector, in combination with the teacher's grading criteria and student ability evaluation indicators, use a dynamic grading and ability evaluation model to perform grading and ability analysis. Among them, the dynamic grading and ability evaluation model uses multi-task learning and an adaptive weight allocation mechanism to dynamically adjust the grading weights and ability evaluation parameters to ensure the accuracy of the grading results and the comprehensiveness of the ability evaluation results.

[0205] S204. Based on the grading results and ability evaluation results, use a generative adversarial network to generate a personalized feedback report. Among them, the generative adversarial network uses a multi-modal generator and a context-aware discriminator to generate multi-modal feedback content including text, charts, and voice.

[0206] It can be seen that based on the homework data submitted by students, use a multi-dimensional feature extraction model to extract features from the text content, logical structure, and language expression; based on the multi-dimensional feature vector, in combination with the subject knowledge base and logical reasoning rules, use a logical reasoning network to perform error detection and correction; based on the homework feature vector, in combination with the teacher's grading criteria and student ability evaluation indicators, use a dynamic grading and ability evaluation model to perform grading and ability analysis; based on the grading results and ability evaluation results, use a generative adversarial network to generate a personalized feedback report, so as to be able to efficiently and accurately identify and correct various errors in the homework, provide personalized feedback, and improve the quality and effect of online education.

[0207] An embodiment of the present invention also provides an electronic device, including a memory and a processor. A computer program is stored in the memory, and the processor is configured to run the computer program to execute the steps in any one of the above method embodiments.

[0208] Specifically, the above electronic device may further include a transmission device and an input / output device. Among them, the transmission device is connected to the above processor, and the input / output device is connected to the above processor.

[0209] Specifically, in this embodiment, the above processor may be configured to execute the following steps through a computer program:

[0210] S201. According to the assignment data submitted by students, use a multi-dimensional feature extraction model to extract features from the text content, logical structure, and language expression. Among them, the multi-dimensional feature extraction model extracts semantic features from three levels: vocabulary, sentence, and discourse through a hierarchical convolutional neural network and a self-attention mechanism, and generates multi-dimensional feature vectors.

[0211] S202. According to the multi-dimensional feature vectors, combine the subject knowledge base and logical reasoning rules, and use a logical reasoning network for error detection and correction. Among them, the logical reasoning network identifies logical errors in the assignment and generates correction suggestions through a graph reasoning algorithm and a rule engine, and outputs the corrected assignment feature vectors.

[0212] S203. According to the assignment feature vectors, combine the teacher's grading criteria and student ability evaluation indicators, and use a dynamic grading and ability evaluation model for grading and ability analysis. Among them, the dynamic grading and ability evaluation model dynamically adjusts the grading weights and ability evaluation parameters through multi-task learning and an adaptive weight allocation mechanism to ensure the accuracy of the grading results and the comprehensiveness of the ability evaluation results.

[0213] S204. According to the grading results and ability evaluation results, use a generative adversarial network to generate a personalized feedback report. Among them, the generative adversarial network generates multi-modal feedback content including text, charts, and speech through a multi-modal generator and a context-aware discriminator.

[0214] It can be seen that according to the assignment data submitted by students, use a multi-dimensional feature extraction model to extract features from the text content, logical structure, and language expression; according to the multi-dimensional feature vectors, combine the subject knowledge base and logical reasoning rules, and use a logical reasoning network for error detection and correction; according to the assignment feature vectors, combine the teacher's grading criteria and student ability evaluation indicators, and use a dynamic grading and ability evaluation model for grading and ability analysis; according to the grading results and ability evaluation results, use a generative adversarial network to generate a personalized feedback report, so as to be able to efficiently and accurately identify and correct various errors in the assignment, provide personalized feedback, and improve the quality and effect of online education.

[0215] The above has detailed the structure, features, and function effects of the present invention according to the illustrated embodiments. The above is only the preferred embodiment of the present invention, but the present invention is not limited to the scope defined by the drawings. Any changes made according to the concept of the present invention, or equivalent embodiments modified into equivalent changes, still within the spirit covered by the description and drawings, shall be within the protection scope of the present invention.

Claims

1. An AI-based online homework correction method, characterized in that, The method includes: Based on the homework data submitted by students, use a multi-dimensional feature extraction model to extract features from the text content, logical structure, and language expression. Among them, the multi-dimensional feature extraction model extracts semantic features from three levels of vocabulary, sentence, and discourse through a hierarchical convolutional neural network and a self-attention mechanism, and generates a multi-dimensional feature vector. Based on the multi-dimensional feature vector, combined with the subject knowledge base and logical reasoning rules, use a logical reasoning network for error detection and correction. Among them, the logical reasoning network identifies logical errors in the homework and generates correction suggestions through a graph reasoning algorithm and a rule engine, and outputs a corrected homework feature vector. Based on the homework feature vector, combined with the teacher's grading criteria and student ability evaluation indicators, use a dynamic grading and ability evaluation model for grading and ability analysis. Among them, the dynamic grading and ability evaluation model dynamically adjusts the grading weights and ability evaluation parameters through multi-task learning and an adaptive weight allocation mechanism to ensure the accuracy of the grading results and the comprehensiveness of the ability evaluation results. Based on the grading results and ability evaluation results, use a generative adversarial network to generate a personalized feedback report. Among them, the generative adversarial network generates multi-modal feedback content including text, charts, and speech through a multi-modal generator and a context-aware discriminator.

2. The method according to claim 1, wherein The step of based on the homework data submitted by students, using a multi-dimensional feature extraction model to extract features from the text content, logical structure, and language expression. Among them, the multi-dimensional feature extraction model extracts semantic features from three levels of vocabulary, sentence, and discourse through a hierarchical convolutional neural network and a self-attention mechanism, and generates a multi-dimensional feature vector, includes: Perform word segmentation, stop word removal, and word vectorization on the homework data submitted by students to obtain a feature representation at the vocabulary level. Based on the feature representation at the vocabulary level, use the first layer of the hierarchical convolutional neural network for convolution operation to extract local semantic features at the vocabulary level, and through max pooling operation, screen out important features at the vocabulary level to generate a vocabulary feature vector. Based on the vocabulary feature vector, use the second layer of the hierarchical convolutional neural network for convolution operation to extract semantic features at the sentence level, and through the self-attention mechanism, calculate the attention weights between words to dynamically adjust the contribution of vocabulary features and generate a sentence feature vector. Based on the sentence feature vector, use the third layer of the hierarchical convolutional neural network for convolution operation to extract global semantic features at the discourse level, and through the self-attention mechanism, calculate the attention weights between sentences to dynamically adjust the contribution of sentence features and generate a discourse feature vector. Concatenate the vocabulary feature vector, sentence feature vector, and discourse feature vector to generate a multi-dimensional feature vector. Among them, the hierarchical convolutional neural network includes: the first layer has a convolutional kernel size of 3 for extracting local semantic features at the vocabulary level; the second layer has a convolutional kernel size of 5 for extracting semantic features at the sentence level; the third layer has a convolutional kernel size of 7 for extracting global semantic features at the discourse level.

3. The method according to claim 2, characterized in that Based on the multi-dimensional feature vector, combined with the subject knowledge base and logical reasoning rules, a logical reasoning network is used for error detection and correction. Among them, the logical reasoning network uses graph reasoning algorithms and rule engines to identify logical errors in the homework and generate correction suggestions, and outputs the corrected homework feature vector, including: Based on the multi-dimensional feature vector, knowledge matching is carried out in combination with the subject knowledge base, and the matched subject knowledge is represented as a knowledge feature vector; The multi-dimensional feature vector and the knowledge feature vector are input into the logical reasoning network. The logical reasoning network initializes the feature representations of nodes and edges through graph reasoning algorithms and rule engines. Among them, each node represents a logical unit, and each edge represents a logical relationship; According to the graph reasoning algorithm, traverse the nodes and edges in the logical reasoning network to detect logical errors, and represent the detected logical errors as error feature vectors. Among them, for each logical unit, calculate its matching degree with the subject knowledge base to identify unmatched logical units. For each logical relationship, calculate its compliance with the rule engine to identify non-compliant logical relationships; According to the error feature vector, use the rule engine to generate correction suggestions. Among them, for unmatched logical units, generate replacement suggestions, and for non-compliant logical relationships, generate adjustment suggestions; Apply the correction suggestions to the logical reasoning network to generate the corrected logical feature vector as the homework feature vector.

4. The method according to claim 3, wherein Based on the homework feature vector, combined with the teacher's grading criteria and student ability evaluation indicators, a dynamic grading and ability evaluation model is used for grading and ability analysis. Among them, the dynamic grading and ability evaluation model dynamically adjusts the grading weights and ability evaluation parameters through multi-task learning and adaptive weight allocation mechanisms to ensure the accuracy of the grading results and the comprehensiveness of the ability evaluation results, including: Load the grading rules according to the teacher's grading criteria and encode the grading rules as a grading weight vector; Load the ability evaluation parameters according to the student ability evaluation indicators and encode the ability evaluation parameters as an ability weight vector; Input the homework feature vector, grading weight vector, and ability weight vector into the multi-task learning model. The multi-task learning model initializes the weights of the grading task and the ability evaluation task through an adaptive weight allocation mechanism. Among them, Based on the homework feature vector, use the shared layer of the multi-task learning model to extract features, and extract high-order feature representations through a fully connected network and a non-linear activation function; Based on the high-order feature representation and the grading weight vector, use the grading task-specific layer to grade, and grade the accuracy, integrity, and logic of the homework to generate a preliminary grading result. According to the adaptive weight allocation mechanism, dynamically adjust the weights of each grading dimension to generate a final grading result; Based on the high-order feature representation and the ability weight vector, use the ability evaluation task-specific layer to conduct ability analysis, and evaluate the student's knowledge mastery, logical thinking ability, and expression ability to generate a preliminary ability evaluation result. According to the adaptive weight allocation mechanism, dynamically adjust the weights of each ability evaluation parameter to generate a final ability evaluation result.

5. The method according to claim 4, wherein Based on the scoring results and ability assessment results, a personalized feedback report is generated using a generative adversarial network. The generative adversarial network generates multimodal feedback content including text, charts, and speech through a multimodal generator and a context-aware discriminator, including: The final scoring results and final ability assessment results are input into the multimodal generator. The multimodal generator generates text, chart, and speech feedback content respectively through a multi-task learning mechanism. Among them, Based on the final scoring results and final ability assessment results, a text generation module is used to generate text feedback content. Among them, the final scoring results are summarized to generate a text evaluation, the final ability assessment results are explained to generate improvement suggestions, and the text evaluation and improvement suggestions are combined to generate text feedback content; Based on the final scoring results and final ability assessment results, a chart generation module is used to generate chart feedback. Among them, the final scoring results are visualized to generate a first chart, the final ability assessment results are visualized to generate a second chart, and the first chart and the second chart are combined to generate chart feedback content; Based on the final scoring results and final ability assessment results, a speech generation module is used to generate speech feedback content. Among them, the text evaluation is synthesized into speech to generate a speech evaluation, the improvement suggestions are synthesized into speech to generate speech suggestions, and the speech evaluation and improvement suggestions are combined to generate speech feedback content; The text feedback content, chart feedback content, and speech feedback content are input into the context-aware discriminator. The context-aware discriminator performs consistency checking and optimization on the feedback content through a context-aware mechanism, and finally outputs multimodal feedback content including text, charts, and speech.

6. An AI-based online homework grading system, characterized in that, The system includes: An extraction module for extracting features of text content, logical structure, and language expression from the homework data submitted by students using a multi-dimensional feature extraction model. The multi-dimensional feature extraction model extracts semantic features from three levels of vocabulary, sentence, and discourse respectively through a hierarchical convolutional neural network and a self-attention mechanism, and generates a multi-dimensional feature vector; A correction module for detecting and correcting errors using a logical reasoning network based on the multi-dimensional feature vector in combination with a subject knowledge base and logical reasoning rules. The logical reasoning network identifies logical errors in the homework and generates correction suggestions through a graph reasoning algorithm and a rule engine, and outputs a corrected homework feature vector; An analysis module for performing scoring and ability analysis using a dynamic scoring and ability assessment model based on the homework feature vector in combination with a teacher scoring standard and a student ability assessment index. The dynamic scoring and ability assessment model dynamically adjusts the scoring weight and ability assessment parameters through a multi-task learning and adaptive weight allocation mechanism to ensure the accuracy of the scoring results and the comprehensiveness of the ability assessment results; A generation module for generating a personalized feedback report using a generative adversarial network based on the scoring results and ability assessment results. The generative adversarial network generates multimodal feedback content including text, charts, and speech through a multimodal generator and a context-aware discriminator.

7. The system according to claim 6, characterized in that, The extraction module is specifically used for: Tokenize, remove stop words, and perform word vectorization on the homework data submitted by students to obtain feature representations at the word level; Based on the feature representations at the word level, perform convolution operations using the first layer of the hierarchical convolutional neural network to extract local semantic features at the word level, and through max pooling operations, filter out important features at the word level to generate word feature vectors; Based on the word feature vectors, perform convolution operations using the second layer of the hierarchical convolutional neural network to extract semantic features at the sentence level, and through the self-attention mechanism, calculate the attention weights between words to dynamically adjust the contributions of word features and generate sentence feature vectors; Based on the sentence feature vectors, perform convolution operations using the third layer of the hierarchical convolutional neural network to extract global semantic features at the discourse level, and through the self-attention mechanism, calculate the attention weights between sentences to dynamically adjust the contributions of sentence features and generate discourse feature vectors; Concatenate the word feature vectors, sentence feature vectors, and discourse feature vectors to generate multi-dimensional feature vectors. Among them, the hierarchical convolutional neural network includes: the first layer has a convolutional kernel size of 3 for extracting local semantic features at the word level; the second layer has a convolutional kernel size of 5 for extracting semantic features at the sentence level; the third layer has a convolutional kernel size of 7 for extracting global semantic features at the discourse level.

8. The system according to claim 7, wherein The correction module is specifically used for: Based on the multi-dimensional feature vectors, perform knowledge matching in combination with the subject knowledge base, and represent the matched subject knowledge as knowledge feature vectors; Input the multi-dimensional feature vectors and knowledge feature vectors into the logical reasoning network. The logical reasoning network initializes the feature representations of nodes and edges through graph reasoning algorithms and rule engines, where each node represents a logical unit and each edge represents a logical relationship; According to the graph reasoning algorithm, traverse the nodes and edges in the logical reasoning network to detect logical errors, and represent the detected logical errors as error feature vectors. Among them, for each logical unit, calculate its matching degree with the subject knowledge base to identify unmatched logical units, and for each logical relationship, calculate its compliance with the rule engine to identify non-compliant logical relationships; Based on the error feature vectors, use the rule engine to generate correction suggestions. Among them, for unmatched logical units, generate replacement suggestions, and for non-compliant logical relationships, generate adjustment suggestions; Apply the correction suggestions to the logical reasoning network to generate corrected logical feature vectors as homework feature vectors.

9. A storage medium, characterized in that, The computer program is stored in the storage medium, where the computer program is set to execute the method described in any one of claims 1-5 when running.

10. An electronic device, comprising a memory and a processor, characterized in that, The computer program is stored in the memory, and the processor is set to run the computer program to execute the method described in any one of claims 1-5.

Images