A method and device for intelligent error data management

Abstract

This application relates to the field of computer technology, and in particular to an intelligent method and apparatus for managing incorrect question data. The method includes: acquiring students' historical incorrect question data and historical answer behavior data to construct a subject knowledge graph and a student cognitive ability profile; semantically encoding the incorrect question text using a pre-trained Transformer model; associating the incorrect questions with corresponding knowledge point nodes in the subject knowledge graph based on the encoded vector representation, and identifying the error type of the incorrect question; updating the student cognitive ability profile according to the attributes of the knowledge point nodes in the subject knowledge graph associated with the incorrect question and the identified error type, and dynamically updating the mastery weights of the associated knowledge point nodes in the subject knowledge graph; and optimizing the cognitive state through LEMMA-style reflective training based on the updated student cognitive ability profile and subject knowledge graph to generate a personalized learning path. This application helps to provide students with precise and personalized learning guidance.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to an intelligent method and apparatus for managing incorrect problem data. Background Technology

[0002] In the current process of digital education, error management is a key link for students to identify and fill gaps in their knowledge and improve learning efficiency. However, existing technologies still have many shortcomings. For example, traditional error management methods rely heavily on manual compilation and have a single data collection dimension; the association between errors and knowledge points is mostly based on keyword matching, which lacks semantic understanding accuracy and is prone to mismatching of knowledge points; and the modeling of students' cognitive abilities is static and cannot be updated in real time based on the latest error data, resulting in a lack of targeted learning paths.

[0003] Therefore, there is an urgent need for an intelligent error data management method that can provide students with precise and personalized learning guidance and improve their learning efficiency. Summary of the Invention

[0004] Therefore, it is necessary to provide an intelligent error data management method and device that can provide students with accurate and personalized learning guidance and improve their learning efficiency, addressing the aforementioned technical problems.

[0005] Firstly, this application provides an intelligent method for managing incorrect question data, the method comprising:

[0006] Acquire students' historical wrong answer data and historical answer behavior data, and construct a subject knowledge graph and a student cognitive ability profile based on the historical wrong answer data and historical answer behavior data;

[0007] The text of incorrect questions within a preset recent period is obtained through multiple collection methods, and the text of incorrect questions is semantically encoded using a pre-trained Transformer model;

[0008] Based on the encoded vector representation, incorrect questions are associated with the corresponding knowledge point nodes in the subject knowledge graph, and the error type corresponding to the incorrect questions is identified through a multi-label classification model.

[0009] Based on the knowledge point node attributes of the subject knowledge graph associated with the wrong questions and the identified error types, the student cognitive ability profile is updated, and based on the updated student cognitive ability profile, the mastery weight of the associated knowledge point nodes in the subject knowledge graph is dynamically updated.

[0010] Based on updated student cognitive ability profiles and subject knowledge graphs, cognitive states are optimized through LEMMA-style reflective training to generate personalized learning paths.

[0011] In one embodiment, the step of optimizing cognitive state and generating personalized learning paths based on updated student cognitive ability profiles and subject knowledge graphs through LEMMA-style reflective training includes:

[0012] Based on the updated student cognitive ability profile, the target knowledge points to be strengthened are identified, and based on the target knowledge points, prerequisite knowledge dependencies and historical high-frequency error types in the subject knowledge graph, customized error examples containing targeted errors are generated.

[0013] Based on the selected reflection mode and the customized error examples, error correction training is completed through LEMMA-style reflection training. After the student successfully corrects the error, the corresponding cognitive dimension parameters in the student's cognitive ability profile are modified based on the knowledge points already mastered, thereby optimizing the cognitive state. The reflection mode includes a step-by-step error correction mode and a holistic error correction mode.

[0014] Based on the updated student cognitive ability profile and subject knowledge graph, a personalized learning path is generated through a genetic algorithm.

[0015] In one embodiment, the step of associating incorrect questions with corresponding knowledge point nodes in the subject knowledge graph based on the encoded vector representation, and identifying the error type corresponding to the incorrect questions through a multi-label classification model, includes:

[0016] Based on the encoded vector representation, the similarity between the wrong question vector and the knowledge point node vector in the subject knowledge graph is calculated by the cosine similarity algorithm. Combined with the graph matching algorithm, the wrong questions are automatically associated with knowledge nodes with high similarity.

[0017] The multi-label classification model is invoked, and the vector representation and context information of the incorrect question are input to identify the error type corresponding to the incorrect question. The error type includes single error and compound error.

[0018] In one embodiment, updating the student's cognitive ability profile based on the attributes of knowledge point nodes in the subject knowledge graph associated with incorrect questions and the identified error types, and dynamically updating the mastery weights of associated knowledge point nodes in the subject knowledge graph based on the updated student cognitive ability profile, includes:

[0019] Extract the attributes of knowledge point nodes in the subject knowledge graph associated with incorrect questions. The attributes include the knowledge point difficulty coefficient, the prerequisite knowledge dependency relationship, and the knowledge point ID.

[0020] Based on the knowledge point node attributes of the subject knowledge graph associated with the wrong questions, and combined with the identified error types, the student's cognitive ability profile is updated according to preset rules.

[0021] Based on the updated student cognitive ability profile, the latest mastery weights of corresponding knowledge points are extracted, and the mastery weights of related knowledge point nodes in the subject knowledge graph are adjusted synchronously.

[0022] In one embodiment, the step of acquiring students' historical incorrect answer data and historical answer behavior data, and constructing a subject knowledge graph and a student cognitive ability profile based on the historical incorrect answer data and historical answer behavior data, includes:

[0023] Obtain students' historical error data and historical answer behavior data;

[0024] Based on the historical incorrect question data, the mapping relationship between questions and knowledge points is extracted, and a subject knowledge graph is constructed using graph structure modeling tools;

[0025] Based on the historical wrong question data and the historical answer behavior data, different cognitive dimension features are obtained through feature extraction and filtering. The different cognitive dimension features include at least knowledge mastery dimension features, thinking ability dimension features and comprehensive application dimension features.

[0026] For the different cognitive dimension features, an appropriate algorithm model is selected for calculation, the features are converted into quantitative results, and a student cognitive ability profile is constructed based on the quantitative results.

[0027] In one embodiment, the intelligent error data management method further includes:

[0028] The incorrect question text, the encoded vector representation, the knowledge point nodes of the subject knowledge graph associated with the incorrect question, the identified error type, the student cognitive ability profile, and the generated personalized learning path are hashed to generate a unique data hash value.

[0029] By pre-setting a blockchain notarization contract, the data hash value is stored in the blockchain distributed ledger to ensure that the data cannot be tampered with.

[0030] In one embodiment, the step of acquiring the incorrect question text within a preset recent period through multiple acquisition methods, and performing semantic encoding on the incorrect question text using a pre-trained Transformer model, includes:

[0031] The text of incorrect questions within a preset recent period is obtained through multiple data collection methods;

[0032] Load subject-specific corpora to fine-tune the pre-trained Transformer model and optimize its semantic understanding ability for subject-specific texts;

[0033] The incorrect question text is segmented into multiple text fragments, and the text fragments are input into a fine-tuned pre-trained Transformer model. The semantic associations in the incorrect question text are captured through a multi-head attention mechanism to generate a fixed-dimensional semantic vector representation of the incorrect question.

[0034] Secondly, this application also provides an intelligent error data management device. The device includes:

[0035] The data acquisition module is used to acquire students' historical wrong answer data and historical answer behavior data, and to construct a subject knowledge graph and a student cognitive ability profile based on the historical wrong answer data and historical answer behavior data;

[0036] The semantic encoding module is used to acquire the text of incorrect questions within a preset recent period through various acquisition methods, and to perform semantic encoding on the text of incorrect questions through a pre-trained Transformer model;

[0037] The error type identification module is used to associate incorrect questions with the corresponding knowledge point nodes in the subject knowledge graph based on the encoded vector representation, and to identify the error type corresponding to the incorrect questions through a multi-label classification model;

[0038] The data dynamic update module is used to update the student's cognitive ability profile based on the knowledge point node attributes of the subject knowledge graph associated with the wrong questions and the identified error type, and to dynamically update the mastery weight of the associated knowledge point nodes in the subject knowledge graph based on the updated student cognitive ability profile.

[0039] The learning path generation module is used to generate personalized learning paths by optimizing cognitive states through LEMMA-style reflective training based on updated student cognitive ability profiles and subject knowledge graphs.

[0040] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to perform the following steps:

[0041] Acquire students' historical wrong answer data and historical answer behavior data, and construct a subject knowledge graph and a student cognitive ability profile based on the historical wrong answer data and historical answer behavior data;

[0042] The text of incorrect questions within a preset recent period is obtained through multiple collection methods, and the text of incorrect questions is semantically encoded using a pre-trained Transformer model;

[0043] Based on the encoded vector representation, incorrect questions are associated with the corresponding knowledge point nodes in the subject knowledge graph, and the error type corresponding to the incorrect questions is identified through a multi-label classification model.

[0044] Based on the knowledge point node attributes of the subject knowledge graph associated with the wrong questions and the identified error types, the student cognitive ability profile is updated, and based on the updated student cognitive ability profile, the mastery weight of the associated knowledge point nodes in the subject knowledge graph is dynamically updated.

[0045] Based on updated student cognitive ability profiles and subject knowledge graphs, cognitive states are optimized through LEMMA-style reflective training to generate personalized learning paths.

[0046] Fourthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, performs the following steps:

[0047] Acquire students' historical wrong answer data and historical answer behavior data, and construct a subject knowledge graph and a student cognitive ability profile based on the historical wrong answer data and historical answer behavior data;

[0048] The text of incorrect questions within a preset recent period is obtained through multiple collection methods, and the text of incorrect questions is semantically encoded using a pre-trained Transformer model;

[0049] Based on the encoded vector representation, incorrect questions are associated with the corresponding knowledge point nodes in the subject knowledge graph, and the error type corresponding to the incorrect questions is identified through a multi-label classification model.

[0050] Based on the knowledge point node attributes of the subject knowledge graph associated with the wrong questions and the identified error types, the student cognitive ability profile is updated, and based on the updated student cognitive ability profile, the mastery weight of the associated knowledge point nodes in the subject knowledge graph is dynamically updated.

[0051] Based on updated student cognitive ability profiles and subject knowledge graphs, cognitive states are optimized through LEMMA-style reflective training to generate personalized learning paths.

[0052] In summary, this application includes the following beneficial technical effects:

[0053] Constructing a subject knowledge graph provides a structured knowledge foundation for accurate association of incorrect questions and planning of learning paths. The generated student cognitive ability profile provides clear targets for personalized intervention, solving the problems of traditional incorrect question management lacking accurate data support and failing to adapt to individual student differences. By pre-training a Transformer model to semantically encode incorrect question texts, the semantic features of incorrect question texts can be accurately captured, providing a high-quality semantic foundation for subsequent incorrect question association and error identification. Based on the knowledge point node attributes of the subject knowledge graph associated with incorrect questions and the identified error types, the student cognitive ability profile is updated, ensuring that the profile accurately reflects the student's current learning status. The mastery weights of related knowledge point nodes in the subject knowledge graph are updated synchronously, enabling the subject knowledge graph to map the student's knowledge mastery level in real time, providing accurate dynamic data support for the generation of subsequent learning paths and ensuring the timeliness of learning guidance. Optimizing cognitive state through LEMMA-style reflective training helps students reconstruct erroneous cognitions, achieving accurate optimization of cognitive state and generating personalized learning paths. This solves the problems of homogenization and insufficient targeting in traditional learning paths, significantly improving the personalization and efficiency of learning. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating an intelligent error data management method in one embodiment;

[0055] Figure 2 This is a flowchart illustrating the intelligent error data management method in another embodiment;

[0056] Figure 3 This is a structural block diagram of an intelligent error data management device in one embodiment. Detailed Implementation

[0057] This invention provides an intelligent method and apparatus for managing incorrect question data.

[0058] The embodiments of the present invention will now be described in more detail with reference to the accompanying drawings. While some embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present disclosure. It should be understood that the accompanying drawings and embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0059] In the description of the embodiments disclosed in this invention, the term "comprising" and similar terms should be understood as open-ended inclusion, i.e., "including but not limited to". The term "based on" should be understood as "at least partially based on". The term "one embodiment" or "the embodiment" should be understood as "at least one embodiment". The terms "first", "second", etc., may refer to different or the same objects. Other explicit and implicit definitions may also be included below.

[0060] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 One embodiment of the intelligent error data management method in this invention includes:

[0061] S100 acquires students' historical wrong answer data and historical answer behavior data, and constructs a subject knowledge graph and a student cognitive ability profile based on the historical wrong answer data and historical answer behavior data.

[0062] Specifically, the process begins with a comprehensive collection of students' historical error data and answer behavior data. The historical error data includes core information such as the question text, subject, correct answer, and student's incorrect answer. The historical answer behavior data includes process data such as answering time, correction frequency, number of pauses / hesitations, and number of times prompts were given. This data provides multi-dimensional support for subsequent modeling. Based on this data, on the one hand, a scientific knowledge graph is constructed using graph structure modeling tools. This graph uses knowledge points as nodes and logical relationships such as prerequisites and extensions between knowledge points as edges. Nodes include attributes such as unique ID, difficulty level, and typical examples, clearly presenting the hierarchy and connections of the knowledge system. On the other hand, through feature extraction, a student cognitive ability profile is constructed, covering dimensions such as knowledge mastery, thinking ability, and comprehensive application, accurately depicting students' learning weaknesses and thinking characteristics.

[0063] In this embodiment, by collecting students' historical wrong question data and historical answer behavior data, the limitation of traditional wrong question management relying solely on single text data is overcome, providing multi-dimensional and full-scenario data support for modeling; a subject knowledge graph is constructed, clearly restoring the logical dependencies between knowledge points, while generating a student cognitive ability profile to accurately depict students' knowledge gaps and thinking characteristics.

[0064] S200 acquires the text of incorrect questions within a preset recent period through various collection methods, and performs semantic encoding on the text of incorrect questions using a pre-trained Transformer model.

[0065] Specifically, the preset recent period can be configured to 7, 15, or 30 days according to needs. The collection scope includes various types of incorrect questions marked by students within this period. Multiple collection methods are available, including OCR photo collection, which uses an OCR photo recognition component to photograph and recognize characters in paper-based incorrect questions, converting them into electronic text; handwriting analysis collection, which uses a handwriting analysis engine to convert incorrect question text written with a smart pen into standardized electronic text; and voice input collection, which integrates a voice input module and a natural language transcription algorithm to convert incorrect questions described by students into text data. Subsequently, a pre-trained Transformer model is used to semantically encode the incorrect question text, converting the text information into a fixed-dimensional vector representation.

[0066] In this embodiment, the text of incorrect questions is semantically encoded by a pre-trained Transformer model, which effectively solves the problems of single data collection dimension and semantic comprehension bias in traditional methods, and provides high-quality data support for accurate association of incorrect questions and identification of error types.

[0067] S300, based on the encoded vector representation, associates incorrect questions with the corresponding knowledge point nodes in the subject knowledge graph, and identifies the error type of the incorrect question through a multi-label classification model.

[0068] Specifically, based on the encoded vectors, similarity calculation and graph matching algorithms are used to automatically associate incorrect questions with the corresponding knowledge point nodes in the subject knowledge graph. At the same time, a multi-label classification model is called to identify the error type of the incorrect questions, including both single errors and compound errors with multiple errors superimposed, so as to achieve accurate classification of incorrect questions.

[0069] S400 updates the student's cognitive ability profile based on the knowledge point node attributes of the subject knowledge graph associated with the wrong questions and the identified error types. Based on the updated student cognitive ability profile, it dynamically updates the mastery weight of the associated knowledge point nodes in the subject knowledge graph.

[0070] Specifically, the attributes of the knowledge point nodes associated with incorrect questions are first extracted, including the knowledge point difficulty coefficient (value 0-1, reflecting the complexity and learning threshold of the knowledge point, such as ≤0.5 for basic concepts and ≥0.7 for comprehensive applications), the prerequisite knowledge dependency relationship (clarifying the basic premise of the knowledge point, such as the prerequisite knowledge of "Vieta's formulas" being "quadratic equations in one variable"), and the knowledge point ID. Combined with the identified error type, the core dimensions of the student's cognitive ability profile (including knowledge mastery dimension, thinking ability dimension, and comprehensive application dimension) are updated according to preset rules. Based on the updated cognitive ability profile, the latest mastery weight of the knowledge point is extracted, and the mastery weight of the associated nodes in the knowledge graph is adjusted synchronously, with a value range of 0-1, a single increase of ≤0.3, and a decrease of ≤0.2. At the same time, the subsequent associated nodes are pre-adjusted according to the knowledge point dependency relationship to ensure that the graph weight is consistent with the cognitive state logic.

[0071] In this embodiment, updating the student's cognitive ability profile can prevent the static profile from becoming disconnected from the student's real-time learning status; dynamically updating the mastery weights of related knowledge point nodes in the subject knowledge graph can provide accurate node weight references for learning planning and prevent the generated path from not matching the student's actual ability.

[0072] S500, based on updated student cognitive ability profiles and subject knowledge graphs, optimizes cognitive state through LEMMA-style reflective training and generates personalized learning paths.

[0073] Specifically, based on the updated cognitive ability profile and subject knowledge graph, LEMMA-style reflective training is used to specifically optimize students' cognitive state, fill knowledge gaps and thinking weaknesses, and generate learning paths that are tailored to students' individual needs. This achieves closed-loop management from error collection to learning improvement, significantly improving learning efficiency and effectiveness. LEMMA-style reflective training involves constructing targeted error data to guide learners to accurately locate errors, deeply reflect on their causes, and efficiently correct them, forming a complete learning reflection loop. This training method enables learners to clearly understand where they went wrong and how to correct it, not only solving immediate specific problems but also improving their self-reflection and error correction abilities.

[0074] In one embodiment, such as Figure 2 As shown, S500 includes:

[0075] S510, combined with the updated student cognitive ability profile, identifies the target knowledge points to be strengthened, and generates customized error examples containing targeted errors based on the target knowledge points, prerequisite knowledge dependencies and historical high-frequency error types in the subject knowledge graph.

[0076] S520, based on the selected reflection mode and customized error examples, completes error correction training through LEMMA-style reflection training. After the student successfully corrects the error, it adjusts the corresponding cognitive dimension parameters in the student's cognitive ability profile based on the knowledge points already mastered, thereby optimizing the cognitive state.

[0077] S530 generates personalized learning paths based on updated student cognitive ability profiles and subject knowledge graphs through a genetic algorithm.

[0078] Specifically, firstly, the updated student cognitive ability profile is analyzed in depth to identify target knowledge points with low mastery weight and high error rates, representing the core weaknesses in students' current learning. Combining the specific attributes of these target knowledge points in the subject knowledge graph (such as difficulty level and knowledge hierarchy), prerequisite knowledge dependencies, and historically high-frequency error types (such as formula confusion in mathematics and conceptual misunderstanding in physics), customized error examples containing targeted errors are generated. These examples not only clearly indicate the error location but also label the error type and cause, ensuring a high degree of alignment between the training content and students' actual problems. Subsequently, two reflection modes (step-by-step error correction and overall error correction) are provided for students to choose from. The step-by-step error correction mode allows students to trace back along the problem-solving steps. After locating the first error node, the system provides an analysis of the error cause and a reference for the correct derivation of that step, helping students to systematically review the problem step by step. The overall error correction mode guides students to re-examine their problem-solving approach, providing only the initial question and core idea hints without directly giving the answer, cultivating students' holistic thinking and independent problem-solving abilities. After students complete error correction training based on the selected pattern and customized error examples, the system analyzes data such as error correction time, number of prompts, and error correction accuracy to determine the effectiveness of the correction. If the correction is successful, the system adjusts the corresponding cognitive dimension parameters in the student's cognitive ability profile based on the knowledge points already mastered. For example, it increases the weight of mastery of the knowledge point and decreases the intensity of the corresponding error type's mindset, achieving precise optimization of cognitive state. Finally, based on the updated student cognitive ability profile and subject knowledge graph, a genetic algorithm is used to generate personalized learning paths.

[0079] In one embodiment, based on the encoded vector representation, incorrect questions are associated with corresponding knowledge point nodes in the subject knowledge graph, and the error types corresponding to the incorrect questions are identified through a multi-label classification model, including:

[0080] Based on the encoded vector representation, the similarity between the wrong question vector and the knowledge point node vector in the subject knowledge graph is calculated by the cosine similarity algorithm. Combined with the graph matching algorithm, the wrong questions are automatically associated with knowledge nodes with high similarity. The multi-label classification model is called, and the input vector representation and wrong question context information are used to identify the error type corresponding to the wrong question.

[0081] Specifically, in the linking of incorrect questions with knowledge point nodes, the feature vectors of each knowledge point node in the subject knowledge graph are invoked. The cosine similarity algorithm is used to calculate the similarity between the incorrect question vector and the vectors of each knowledge point node in the subject knowledge graph. Candidate nodes with similarity higher than a preset threshold are selected, initially narrowing down the knowledge point range to which the incorrect question belongs. To further improve the accuracy of the linking, a graph matching algorithm is used. Information such as the logical dependencies of knowledge points in the subject knowledge graph, the subject to which the incorrect question belongs, and the question type are used to perform a secondary screening of candidate nodes. For example, if the candidate nodes for a geometry proof question include "congruent triangle criteria" and "solving fractional equations," irrelevant nodes can be excluded by combining the question type and the chapter classification of the knowledge graph, thus identifying a unique target knowledge point node. This achieves accurate linking of incorrect questions with knowledge points, avoiding the mis-linking problems caused by traditional keyword matching. In the error type identification stage, a multi-label classification model is invoked. The encoded error vector and the context information of the error (such as key information in the question stem and details of the wrong answer) are input into the model, and the model outputs the corresponding error type label. The error type is divided into single error (such as only calculation error) and compound error (such as calculation error + formula confusion), which effectively solves the problem that traditional classification is difficult to identify complex errors.

[0082] In one embodiment, the student's cognitive ability profile is updated based on the attributes of the knowledge point nodes in the subject knowledge graph associated with the incorrect questions and the identified error types. Based on the updated student cognitive ability profile, the mastery weights of the associated knowledge point nodes in the subject knowledge graph are dynamically updated, including:

[0083] Extract the attributes of knowledge point nodes in the subject knowledge graph associated with incorrect questions. The attributes include the difficulty coefficient of the knowledge point, the prerequisite knowledge dependency relationship, and the knowledge point ID. Based on the attributes of the knowledge point nodes in the subject knowledge graph associated with incorrect questions, and combined with the identified error type, update the student's cognitive ability profile according to preset rules. Based on the updated student cognitive ability profile, extract the latest mastery weight of the corresponding knowledge point, and adjust the mastery weight of the associated knowledge point nodes in the subject knowledge graph in sync.

[0084] Specifically, firstly, the core attributes of the knowledge nodes associated with incorrect answers are extracted, including the difficulty coefficient of the knowledge point, the dependency relationship of prerequisite knowledge, and the knowledge point ID. Then, combining the extracted node attributes with the identified error types, the student's cognitive ability profile is updated according to preset rules. For example, when a high-difficulty knowledge point is combined with a complex error, the corresponding knowledge mastery score is significantly reduced. Based on the updated student cognitive ability profile, the latest mastery weight of the corresponding knowledge point is extracted, and the mastery weight of the associated knowledge point nodes in the subject knowledge graph is adjusted simultaneously to ensure that the knowledge graph can reflect the student's knowledge mastery status in real time, providing accurate data support for the subsequent generation of learning paths.

[0085] In one embodiment, acquiring students' historical incorrect answer data and historical answer behavior data, and constructing a subject knowledge graph and a student cognitive ability profile based on the historical incorrect answer data and historical answer behavior data includes:

[0086] Acquire students' historical error data and historical answer behavior data; based on the historical error data, extract the mapping relationship between questions and knowledge points, and construct a subject knowledge graph using graph structure modeling tools; based on the historical error data and historical answer behavior data, obtain different cognitive dimension features through feature extraction and filtering, which include at least knowledge mastery dimension features, thinking ability dimension features, and comprehensive application dimension features; select appropriate algorithm models for different cognitive dimension features to perform calculations, transform the features into quantitative results, and construct student cognitive ability profiles based on the quantitative results.

[0087] Specifically, historical error data and answer behavior data of students are obtained through multiple channels. Based on the collected historical error data, the mapping relationship between questions and knowledge points is extracted through rule matching and manual verification. In the rule matching stage, algorithms such as keyword matching and question type matching are used to initially establish the mapping relationship of "question ID - knowledge point ID". In the manual verification stage, subject teachers sample and review the rule matching results to correct mismatches and omissions, ensuring the accuracy of the mapping relationship. Subsequently, a subject knowledge graph is constructed using graph structure modeling tools. At the same time, different cognitive dimension features are obtained from the historical error data and historical answer behavior data through feature extraction and filtering. These different cognitive dimension features include at least knowledge mastery dimension features, thinking ability dimension features, and comprehensive application dimension features. Knowledge mastery dimension features include the correct answer rate, error rate, and frequency of repeated errors for each knowledge point. Thinking ability dimension features include the sequence pattern of similar errors, the distribution of error types, and the distribution of hesitation time in answering questions. Comprehensive application ability dimension features include the score rate of cross-knowledge point questions, the completeness of answer steps, and the performance of knowledge point transfer and application. For different cognitive dimensions, appropriate algorithm models are selected for quantitative calculation. Specifically, the Bayesian Knowledge Traceability (BKT) model is used to calculate the probability of knowledge mastery; the FP-Growth sequence pattern mining algorithm is used to analyze the strength of mental set; and the Graph Neural Network (GNN) model is used to evaluate the level of comprehensive application ability. The characteristics of each dimension are transformed into quantifiable numerical or grade results, and these results are integrated to construct a comprehensive, accurate, and dynamically updated profile of students' cognitive abilities, clearly presenting students' knowledge mastery status, thinking characteristics, and ability levels.

[0088] In one embodiment, the intelligent error data management method further includes:

[0089] The system performs hash operations on the text of incorrect questions, the encoded vector representation, the knowledge point nodes of the subject knowledge graph associated with the incorrect questions, the identified error types, the student's cognitive ability profile, and the generated personalized learning path to generate a unique data hash value. Through a pre-set blockchain notarization contract, the data hash value is stored in the blockchain distributed ledger to ensure that the data is tamper-proof.

[0090] Specifically, firstly, the scope of core data requiring notarization is clearly defined, including the text of incorrect questions (including the question stem, incorrect answer, and correct answer), the encoded semantic vector representation of the incorrect questions, the knowledge point node information of the subject knowledge graph associated with the incorrect questions (such as knowledge point ID and name), the identified error type labels, the student's cognitive ability profile (including quantitative results of various dimensions), and the generated personalized learning path (including knowledge point sequence, learning resource configuration, and time duration suggestions), ensuring that the notarized data can fully reflect the core process and results of incorrect question management. Subsequently, a hash operation is performed on the above core data using secure hash algorithms such as SHA-256 to generate a unique data hash value. Hash algorithms are irreversible and unique; any minor modification to the original data will cause a significant change in the hash value. Therefore, this hash value can serve as a "digital fingerprint" of data integrity, effectively preventing data tampering. Finally, through a pre-defined blockchain notarization contract, the generated data hash value is stored in the blockchain distributed ledger. The blockchain adopts a decentralized storage architecture, and the data hash value is synchronously recorded on multiple nodes. There is no single control center, and no node can modify the data independently, ensuring the immutability and permanent storage of the data. Meanwhile, based on the hash value stored on the blockchain, a cross-platform credit authentication interface is provided. Other educational institutions or learning platforms can recognize the student's incorrect problem training results by verifying the validity of the hash value, which solves the problem that learning results are difficult to pass across platforms in traditional centralized storage.

[0091] In one embodiment, the text of incorrect questions within a preset recent period is obtained through multiple collection methods, and the semantic encoding of the incorrect question text is performed using a pre-trained Transformer model, including:

[0092] The system acquires incorrect question texts within a preset recent period through various acquisition methods; loads a subject-specific corpus to fine-tune the pre-trained Transformer model, optimizing its semantic understanding ability for subject-specific texts; segments the incorrect question texts into multiple text fragments and inputs these fragments into the fine-tuned pre-trained Transformer model, using a multi-head attention mechanism to capture semantic relationships within the incorrect question texts and generate fixed-dimensional semantic vector representations of the incorrect questions.

[0093] Specifically, the system first acquires the text of incorrect answers from the most recent period through various collection methods and loads a subject-specific corpus. This corpus contains targeted content such as subject-specific terminology, formula expressions, descriptions of typical question types, and standardized problem-solving steps. Based on this corpus, a pre-trained Transformer model (such as BERT or RoBERTa) is fine-tuned and trained. Through multiple rounds of iteration, the model parameters are optimized to enable the model to accurately capture the semantic features and logical connections of the subject-specific text. Subsequently, the collected and pre-processed text of incorrect answers is segmented according to the model's input requirements, breaking it down into multiple text fragments of appropriate length. Fragments that are too short are padded, and fragments that are too long are truncated according to semantic integrity to ensure that the input sequence meets the model's requirements. The processed text fragments are input into a fine-tuned pre-trained Transformer model. The model uses a multi-layer multi-head attention mechanism to comprehensively capture the semantic relationships within the text of incorrect questions (such as the logical relationship between the known conditions and the question in the question stem, and the semantic differences between incorrect and correct answers). Finally, it generates a fixed-dimensional semantic vector representation of the incorrect questions. This vector accurately contains the semantic information of the incorrect questions, providing high-quality feature input for subsequent association of incorrect questions with knowledge points and identification of error types.

[0094] In one embodiment, such as Figure 3 As shown, an intelligent error data management device is provided, comprising: a data acquisition module 10, a semantic encoding module 20, an error type identification module 30, a data dynamic update module 40, and a learning path generation module 50, wherein:

[0095] The data acquisition module 10 is used to acquire students' historical wrong answer data and historical answer behavior data, and to construct a subject knowledge graph and a student cognitive ability profile based on the historical wrong answer data and historical answer behavior data;

[0096] The semantic encoding module 20 is used to acquire the text of incorrect questions within a preset recent period through various acquisition methods, and to perform semantic encoding on the text of incorrect questions through a pre-trained Transformer model;

[0097] Error type identification module 30 is used to associate incorrect questions with corresponding knowledge point nodes in the subject knowledge graph based on the encoded vector representation, and to identify the error type corresponding to the incorrect questions through a multi-label classification model;

[0098] The data dynamic update module 40 is used to update the student's cognitive ability profile based on the knowledge point node attributes of the subject knowledge graph associated with the wrong questions and the identified error types, and to dynamically update the mastery weight of the associated knowledge point nodes in the subject knowledge graph based on the updated student cognitive ability profile.

[0099] The learning path generation module 50 is used to optimize cognitive state and generate personalized learning paths based on updated student cognitive ability profiles and subject knowledge graphs through LEMMA-style reflective training.

[0100] In one embodiment, the learning path generation module 50 is further configured to combine the updated student cognitive ability profile to determine the target knowledge points to be strengthened, and generate customized error examples containing targeted errors based on the target knowledge points, prerequisite knowledge dependencies, and historical high-frequency error types in the subject knowledge graph; according to the selected reflection mode and customized error examples, error correction training is completed through LEMMA-style reflection training, and after the student successfully corrects the error, the corresponding cognitive dimension parameters in the student cognitive ability profile are corrected based on the mastered knowledge points to achieve cognitive state optimization, wherein the reflection mode includes a step-by-step error correction mode and a holistic error correction mode; based on the updated student cognitive ability profile and subject knowledge graph, a personalized learning path is generated through a genetic algorithm.

[0101] In one embodiment, the error type identification module 30 is also used to calculate the similarity between the error question vector and the knowledge point node vector in the subject knowledge graph based on the encoded vector representation using the cosine similarity algorithm, and automatically associate the error question with the knowledge node with high similarity using the graph matching algorithm; call the multi-label classification model, input the vector representation and the error question context information, and identify the error type corresponding to the error question, including single error and compound error.

[0102] In one embodiment, the data dynamic update module 40 is further used to extract the knowledge point node attributes of the subject knowledge graph associated with the wrong questions. The attributes include the knowledge point difficulty coefficient, the prerequisite knowledge dependency relationship, and the knowledge point ID. Based on the knowledge point node attributes of the subject knowledge graph associated with the wrong questions, and combined with the identified error type, the student's cognitive ability profile is updated according to preset rules. Based on the updated student cognitive ability profile, the latest mastery weight of the corresponding knowledge point is extracted, and the mastery weight of the associated knowledge point nodes in the subject knowledge graph is adjusted synchronously.

[0103] In one embodiment, the data acquisition module 10 is further used to acquire students' historical wrong question data and historical answer behavior data; based on the historical wrong question data, the mapping relationship between questions and knowledge points is extracted, and a subject knowledge graph is constructed using a graph structure modeling tool; based on the historical wrong question data and historical answer behavior data, different cognitive dimension features are obtained through feature extraction and filtering, and the different cognitive dimension features include at least knowledge mastery dimension features, thinking ability dimension features, and comprehensive application dimension features; for different cognitive dimension features, an appropriate algorithm model is selected for calculation, the features are converted into quantitative results, and a student cognitive ability profile is constructed based on the quantitative results.

[0104] In one embodiment, the intelligent error data management device further includes a blockchain storage module, which performs hash operations on the error text, the encoded vector representation, the knowledge point nodes of the subject knowledge graph associated with the error, the identified error type, the student's cognitive ability profile, and the generated personalized learning path to generate a unique data hash value; and stores the data hash value in the blockchain distributed ledger through a preset blockchain notarization contract to ensure that the data is tamper-proof.

[0105] In one embodiment, the semantic encoding module 20 is further configured to acquire the text of incorrect questions within a preset recent period through various acquisition methods; load a subject-specific corpus to fine-tune the pre-trained Transformer model and optimize the Transformer model's semantic understanding ability for subject-specific texts; segment the incorrect question text into multiple text fragments and input the text fragments into the fine-tuned pre-trained Transformer model, capture the semantic associations in the incorrect question text through a multi-head attention mechanism, and generate a fixed-dimensional semantic vector representation of the incorrect questions.

[0106] In one embodiment, this application discloses a computer device including a memory and a processor. The memory stores a computer program that can run on the processor. When the processor loads the computer program, it executes an intelligent error data management method according to the above embodiment.

[0107] In one embodiment, this application discloses a computer-readable storage medium storing a computer program, wherein when the computer program is loaded by a processor, it executes an intelligent error data management method according to the above embodiment.

[0108] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for intelligent error data management, characterized in that, include: Acquire students' historical wrong answer data and historical answer behavior data, and construct a subject knowledge graph and a student cognitive ability profile based on the historical wrong answer data and historical answer behavior data; The text of incorrect questions within a preset recent period is obtained through multiple collection methods, and the text of incorrect questions is semantically encoded using a pre-trained Transformer model; Based on the encoded vector representation, incorrect questions are associated with the corresponding knowledge point nodes in the subject knowledge graph, and the error type corresponding to the incorrect questions is identified through a multi-label classification model. Based on the knowledge point node attributes of the subject knowledge graph associated with the wrong questions and the identified error types, the student cognitive ability profile is updated, and based on the updated student cognitive ability profile, the mastery weight of the associated knowledge point nodes in the subject knowledge graph is dynamically updated. Based on updated student cognitive ability profiles and subject knowledge graphs, cognitive states are optimized through LEMMA-style reflective training to generate personalized learning paths; The method of associating incorrect questions with corresponding knowledge point nodes in the subject knowledge graph based on the encoded vector representation, and identifying the error type of the incorrect question through a multi-label classification model, includes: Based on the encoded vector representation, the similarity between the wrong question vector and the knowledge point node vector in the subject knowledge graph is calculated by the cosine similarity algorithm. Combined with the graph matching algorithm, the wrong questions are automatically associated with knowledge nodes with high similarity. The multi-label classification model is invoked, and the vector representation and context information of the incorrect question are input to identify the error type corresponding to the incorrect question. The error type includes single error and compound error. The process of automatically associating incorrect questions with highly similar knowledge nodes based on the encoded vector representation, using a cosine similarity algorithm to calculate the similarity between the incorrect question vector and the knowledge node vectors in the subject knowledge graph, and combining this with a graph matching algorithm, includes: In the process of associating incorrect questions with knowledge node vectors, the feature vectors of each knowledge node in the subject knowledge graph are called, and the cosine similarity algorithm is used to calculate the similarity between the incorrect question vector and the knowledge node vectors in the subject knowledge graph. Candidate associated nodes with similarity higher than a preset threshold are selected to initially determine the knowledge point range to which the incorrect question belongs. To further improve the accuracy of the association, the graph matching algorithm is used to perform a second screening of the candidate associated nodes using the logical dependencies of knowledge points in the subject knowledge graph, the subject to which the incorrect question belongs, and the question type information. This achieves accurate association between incorrect questions and knowledge points and avoids the misassociation problem caused by traditional keyword matching. The process of acquiring students' historical incorrect answers and historical answering behavior data, and constructing a subject knowledge graph and a student cognitive ability profile based on the historical incorrect answers and historical answering behavior data, includes: Obtain students' historical error data and historical answer behavior data; Based on the historical incorrect question data, the mapping relationship between questions and knowledge points is extracted, and a subject knowledge graph is constructed using graph structure modeling tools; Based on the historical wrong question data and the historical answer behavior data, different cognitive dimension features are obtained through feature extraction and filtering. The different cognitive dimension features include at least knowledge mastery dimension features, thinking ability dimension features and comprehensive application dimension features. For the different cognitive dimension features, an appropriate algorithm model is selected for calculation, the features are converted into quantitative results, and a student cognitive ability profile is constructed based on the quantitative results. The knowledge mastery dimension features include the correct answer rate, error rate, and frequency of repeated errors for each knowledge point; the thinking ability dimension features include the sequence pattern of similar errors, the distribution of error types, and the distribution of hesitation time in answering questions; the comprehensive application ability dimension features include the score rate of cross-knowledge point questions, the completeness of answering steps, and the performance of knowledge point transfer and application. The process of selecting appropriate algorithm models for calculation based on the different cognitive dimension features, converting the features into quantitative results, and constructing a student cognitive ability profile based on the quantitative results includes: selecting appropriate algorithm models for quantitative calculation based on different cognitive dimension features; using a Bayesian knowledge tracing model to calculate the probability of knowledge mastery; analyzing the strength of mindset through the FP-Growth sequence pattern mining algorithm; using a graph neural network model to evaluate the level of comprehensive application ability; and integrating the quantifiable numerical or level results converted from the features of each dimension to construct a comprehensive, accurate, and dynamically updated student cognitive ability profile that clearly presents the student's knowledge mastery status, thinking characteristics, and ability level. The method of optimizing cognitive state and generating personalized learning paths based on updated student cognitive ability profiles and subject knowledge graphs through LEMMA-style reflective training includes: Based on the updated student cognitive ability profile, the target knowledge points to be strengthened are identified, and based on the target knowledge points, prerequisite knowledge dependencies and historical high-frequency error types in the subject knowledge graph, customized error examples containing targeted errors are generated. Based on the selected reflection mode and the customized error examples, error correction training is completed through LEMMA-style reflection training. After the student successfully corrects the error, the corresponding cognitive dimension parameters in the student's cognitive ability profile are modified based on the knowledge points already mastered, thereby optimizing the cognitive state. The reflection mode includes a step-by-step error correction mode and a holistic error correction mode. Based on the updated student cognitive ability profile and subject knowledge graph, a personalized learning path is generated through a genetic algorithm; The LEMMA-style reflective training involves constructing targeted error data to guide learners to accurately locate errors, deeply reflect on their causes, and efficiently correct them, forming a complete learning and reflection loop. This allows learners to clearly understand where they went wrong and how to correct it, not only solving immediate problems but also enhancing their self-reflection and error correction abilities. The step-by-step error correction mode allows students to backtrack along the problem-solving steps. After locating the first error node, the system provides an analysis of the error cause and a reference for the correct derivation of the corresponding steps, helping students to sort out the problem step by step. The overall error correction mode guides students to rethink their problem-solving approach, providing only the initial question and core idea hints without directly giving the answer, thus cultivating students' global thinking and independent problem-solving ability. After students complete error correction training based on the selected mode and the customized error examples, the system analyzes the error correction time, the number of hints relied upon, and the error correction accuracy data to determine whether the error correction is effective. If the error correction is successful, the system adjusts the corresponding cognitive dimension parameters in the student's cognitive ability profile based on the knowledge points already mastered. The step of acquiring incorrect question texts within a preset recent period through multiple collection methods, and performing semantic encoding on the incorrect question texts using a pre-trained Transformer model, includes: The text of incorrect questions within a preset recent period is obtained through multiple data collection methods; Load subject-specific corpora to fine-tune the pre-trained Transformer model and optimize its semantic understanding ability for subject-specific texts; The incorrect question text is segmented into multiple text fragments, and the text fragments are input into a fine-tuned pre-trained Transformer model. The semantic associations in the incorrect question text are captured through a multi-head attention mechanism to generate a fixed-dimensional semantic vector representation of the incorrect question. The subject-specific corpus contains subject-specific terminology, formulas, descriptions of typical question types, and standardized problem-solving steps. The pre-trained Transformer model is fine-tuned based on the subject-specific corpus, and the model parameters are optimized through multiple rounds of iteration, enabling the model to accurately capture the semantic features and logical connections of subject texts.

2. The intelligent error data management method according to claim 1, characterized in that, The process of updating the student's cognitive ability profile based on the attributes of knowledge point nodes in the subject knowledge graph associated with incorrect answers and the identified error types, and then dynamically updating the mastery weights of associated knowledge point nodes in the subject knowledge graph based on the updated student cognitive ability profile, includes: Extract the attributes of knowledge point nodes in the subject knowledge graph associated with incorrect questions. The attributes include the knowledge point difficulty coefficient, the prerequisite knowledge dependency relationship, and the knowledge point ID. Based on the knowledge point node attributes of the subject knowledge graph associated with the wrong questions, and combined with the identified error types, the student's cognitive ability profile is updated according to preset rules. Based on the updated student cognitive ability profile, the latest mastery weights of corresponding knowledge points are extracted, and the mastery weights of related knowledge point nodes in the subject knowledge graph are adjusted synchronously.

3. The intelligent error data management method according to claim 1, characterized in that, Also includes: The incorrect question text, the encoded vector representation, the knowledge point nodes of the subject knowledge graph associated with the incorrect question, the identified error type, the student cognitive ability profile, and the generated personalized learning path are hashed to generate a unique data hash value. By pre-setting a blockchain notarization contract, the data hash value is stored in the blockchain distributed ledger to ensure that the data cannot be tampered with.

4. An intelligent error data management device, characterized in that, include: The data acquisition module is used to acquire students' historical wrong answer data and historical answer behavior data, and to construct a subject knowledge graph and a student cognitive ability profile based on the historical wrong answer data and historical answer behavior data; The semantic encoding module is used to acquire the text of incorrect questions within a preset recent period through various acquisition methods, and to perform semantic encoding on the text of incorrect questions through a pre-trained Transformer model; The error type identification module is used to associate incorrect questions with the corresponding knowledge point nodes in the subject knowledge graph based on the encoded vector representation, and to identify the error type corresponding to the incorrect questions through a multi-label classification model; The data dynamic update module is used to update the student's cognitive ability profile based on the knowledge point node attributes of the subject knowledge graph associated with the wrong questions and the identified error type, and to dynamically update the mastery weight of the associated knowledge point nodes in the subject knowledge graph based on the updated student cognitive ability profile. The learning path generation module is used to optimize cognitive state and generate personalized learning paths based on updated student cognitive ability profiles and subject knowledge graphs through LEMMA-style reflective training. The error type identification module is also used to calculate the similarity between the error vector and the knowledge node vector in the subject knowledge graph based on the encoded vector representation using the cosine similarity algorithm, and automatically associate the error with the knowledge node with high similarity using the graph matching algorithm. The multi-label classification model is invoked, and the vector representation and context information of the incorrect question are input to identify the error type corresponding to the incorrect question. The error type includes single error and compound error. The process of automatically associating incorrect questions with highly similar knowledge nodes based on the encoded vector representation, using a cosine similarity algorithm to calculate the similarity between the incorrect question vector and the knowledge node vectors in the subject knowledge graph, and combining this with a graph matching algorithm, includes: In the process of associating incorrect questions with knowledge node vectors, the feature vectors of each knowledge node in the subject knowledge graph are called, and the cosine similarity algorithm is used to calculate the similarity between the incorrect question vector and the knowledge node vectors in the subject knowledge graph. Candidate associated nodes with similarity higher than a preset threshold are selected to initially determine the knowledge point range to which the incorrect question belongs. To further improve the accuracy of the association, the graph matching algorithm is used to perform a second screening of the candidate associated nodes using the logical dependencies of knowledge points in the subject knowledge graph, the subject to which the incorrect question belongs, and the question type information. This achieves accurate association between incorrect questions and knowledge points and avoids the misassociation problem caused by traditional keyword matching. The data acquisition module is also used to acquire students' historical wrong question data and historical answer behavior data; based on the historical wrong question data, the mapping relationship between questions and knowledge points is extracted, and a subject knowledge graph is constructed using a graph structure modeling tool; according to the historical wrong question data and the historical answer behavior data, different cognitive dimension features are obtained through feature extraction and filtering, and the different cognitive dimension features include at least knowledge mastery dimension features, thinking ability dimension features, and comprehensive application dimension features; for the different cognitive dimension features, an appropriate algorithm model is selected for calculation, the features are converted into quantitative results, and a student cognitive ability profile is constructed based on the quantitative results; The knowledge mastery dimension features include the correct answer rate, error rate, and frequency of repeated errors for each knowledge point; the thinking ability dimension features include the sequence pattern of similar errors, the distribution of error types, and the distribution of hesitation time in answering questions; the comprehensive application ability dimension features include the score rate of cross-knowledge point questions, the completeness of answering steps, and the performance of knowledge point transfer and application. The process of selecting appropriate algorithm models for calculation based on the different cognitive dimension features, converting the features into quantitative results, and constructing a student cognitive ability profile based on the quantitative results includes: selecting appropriate algorithm models for quantitative calculation based on different cognitive dimension features; using a Bayesian knowledge tracing model to calculate the probability of knowledge mastery; analyzing the strength of mindset through the FP-Growth sequence pattern mining algorithm; using a graph neural network model to evaluate the level of comprehensive application ability; and integrating the quantifiable numerical or level results converted from the features of each dimension to construct a comprehensive, accurate, and dynamically updated student cognitive ability profile that clearly presents the student's knowledge mastery status, thinking characteristics, and ability level. The learning path generation module is also used to combine the updated student cognitive ability profile to determine the target knowledge points to be strengthened, and generate customized error examples containing targeted errors based on the target knowledge points, prerequisite knowledge dependencies, and historical high-frequency error types in the subject knowledge graph; according to the selected reflection mode and the customized error examples, error correction training is completed through LEMMA-style reflection training, and after the student successfully corrects the error, the corresponding cognitive dimension parameters in the student cognitive ability profile are corrected based on the mastered knowledge points to achieve cognitive state optimization, wherein the reflection mode includes a step-by-step error correction mode and a holistic error correction mode; based on the updated student cognitive ability profile and subject knowledge graph, a personalized learning path is generated through a genetic algorithm; The LEMMA-style reflective training involves constructing targeted error data to guide learners to accurately locate errors, deeply reflect on their causes, and efficiently correct them, forming a complete learning and reflection loop. This allows learners to clearly understand where they went wrong and how to correct it, not only solving immediate problems but also enhancing their self-reflection and error correction abilities. The step-by-step error correction mode allows students to backtrack along the problem-solving steps. After locating the first error node, the system provides an analysis of the error cause and a reference for the correct derivation of the corresponding steps, helping students to sort out the problem step by step. The overall error correction mode guides students to rethink their problem-solving approach, providing only the initial question and core idea hints without directly giving the answer, thus cultivating students' global thinking and independent problem-solving ability. After students complete error correction training based on the selected mode and the customized error examples, the system analyzes the error correction time, the number of hints relied upon, and the error correction accuracy data to determine whether the error correction is effective. If the error correction is successful, the system adjusts the corresponding cognitive dimension parameters in the student's cognitive ability profile based on the knowledge points already mastered. The semantic encoding module is also used to acquire the text of incorrect questions within a preset recent period through various acquisition methods; load a subject-specific corpus to fine-tune the pre-trained Transformer model and optimize the Transformer model's semantic understanding ability for subject-specific texts; segment the text of incorrect questions into multiple text fragments and input the text fragments into the fine-tuned pre-trained Transformer model, and capture the semantic associations in the text of incorrect questions through a multi-head attention mechanism to generate a fixed-dimensional semantic vector representation of incorrect questions; The subject-specific corpus contains subject-specific terminology, formulas, descriptions of typical question types, and standardized problem-solving steps. The pre-trained Transformer model is fine-tuned based on the subject-specific corpus, and the model parameters are optimized through multiple rounds of iteration, enabling the model to accurately capture the semantic features and logical connections of subject texts.

5. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 3.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 3.

Images