An individualized AI learning system and method based on learning condition diagnosis

Abstract

This invention relates to the field of learning assessment technology, specifically to a personalized AI learning system and method based on learning assessment. The method includes: pushing generated variation questions to students for targeted training, collecting students' performance on the variation questions, and updating students' mastery scores of corresponding knowledge points; dynamically adjusting the priority weight of each incorrect question in the error notebook based on the results of the variation questions, prioritizing the recommendation of repeatedly incorrect questions for further variation generation; training a variation generation model through a comparative learning framework to automatically generate semantically similar variation questions of the incorrect questions, and dynamically adjusting the priority of the error notebook based on students' performance on the variation questions, thereby realizing dynamic iteration of the error notebook and improving knowledge transfer ability, solving the problems of low learning efficiency and insufficient knowledge transfer ability caused by static error notebooks.

Description

Technical Field

[0001] This invention relates to the field of learning assessment technology, and in particular to a personalized AI learning system and method based on learning assessment. Background Technology

[0002] In online learning and intelligent education systems, personalized learning recommendations are a crucial means of improving learning efficiency. Existing technologies commonly employ personalized learning methods including knowledge graph-based learning path recommendations, item response theory-based ability assessments, and collaborative filtering-based exercise recommendations. Specifically, the system collects multi-dimensional data such as students' answer records, video viewing behavior, and test scores to construct a model of students' knowledge mastery, and then recommends corresponding learning content or practice questions based on this model. Among these, the error notebook function is widely used to record questions students answered incorrectly, allowing students to repeatedly review these questions to reinforce weak knowledge points. Some systems also utilize machine learning algorithms to categorize incorrect questions by knowledge point and adjust review priorities based on the frequency of incorrect answers. Furthermore, deep learning-based question generation technology is also beginning to be applied in the education field, automatically generating similar questions for students to practice based on existing questions.

[0003] However, in the aforementioned existing technologies, students often fall into the trap of memorizing questions when repeatedly practicing the same original incorrect problem. This leads to the possibility of making the same mistake again even if the answer is correct, when rephrased in a different way. Existing methods lack the ability to deeply explore the semantic structure of incorrect questions and generate variations, failing to capture the semantic differences between incorrect and correct answers. They also struggle to dynamically adjust the priority of the error notebook and generate targeted variation exercises based on students' performance. Therefore, there is an urgent need for a personalized learning method that can automatically generate variations of incorrect questions based on comparative learning and dynamically update the priority of the error notebook based on student training feedback. This would address the problems of low learning efficiency and insufficient knowledge transfer ability caused by static error notebooks. Summary of the Invention

[0004] The purpose of this invention is to provide a personalized AI learning system and method based on learning progress diagnosis, aiming to solve the technical problems of existing methods lacking the ability to deeply explore the semantic structure of wrong questions and generate variants, failing to capture the semantic differences between wrong questions and correct answers, and struggling to dynamically adjust the priority of wrong question notebooks and generate targeted variant exercises based on students' answer performance.

[0005] To achieve the above objectives, this invention employs a personalized AI learning method based on learning progress diagnosis, comprising the following steps: Collect data on students' incorrect answers during the learning process, and extract the knowledge point tags and error pattern characteristics for each incorrect answer; Based on the contrastive learning framework, positive and negative sample pairs are constructed by comparing incorrect and correct questions under the same knowledge point, and the variant generation model is trained to enable the model to capture the semantic differences between incorrect questions and correct answers. The target incorrect question is input into the trained variation generation model to generate several variation questions that are semantically similar to the incorrect question but different in form, and the difficulty coefficient of the variation questions is controlled. The generated variant questions are pushed to students for targeted training, and students' performance on the variant questions is collected to update their mastery scores of the corresponding knowledge points. Based on the answers to the variation questions, the priority weight of each wrong question in the error notebook is dynamically adjusted, and questions that are repeatedly wrong are given priority for further variation generation.

[0006] Among the steps involved in collecting students' incorrect answers during the learning process and extracting the knowledge point tags and error pattern features for each incorrect answer: The knowledge point tags are automatically labeled using a pre-constructed subject knowledge graph; The error pattern features include four types: calculation errors, conceptual confusion, question interpretation deviation, and logical jumps. These features are extracted from the answers to incorrect questions through rule matching or classification models.

[0007] In the step of constructing positive and negative sample pairs between incorrect and correct questions under the same knowledge point based on the contrastive learning framework, and training a variant generative model so that the model can capture the semantic differences between incorrect questions and correct answers: An encoder-decoder architecture is adopted. The encoder encodes the question text into a semantic vector, and the contrastive learning loss function reduces the distance between the incorrect question vector and the correct question vector, and increases the distance between the incorrect question vector and the random question vector. The decoder takes the error vector and difference vector as input and generates variant question text.

[0008] In the step of constructing positive and negative sample pairs between incorrect and correct questions under the same knowledge point based on the contrastive learning framework, and training a variant generative model so that the model can capture the semantic differences between incorrect questions and correct answers: The difference vector is a semantic difference encoding between the target incorrect question and the correct question. It is obtained by subtracting the incorrect question vector from the correct question vector and then mapping it to the latent space through a fully connected layer.

[0009] Among the steps, the target incorrect question is input into a trained variation generation model to generate several variation questions that are semantically similar to the incorrect question but different in form, and the difficulty coefficient of the variation questions is controlled: The difficulty of the variant questions can be controlled by adjusting the temperature parameters or Top-k sampling values ​​during the generation process. The higher the temperature, the greater the variation in the generated questions and the wider the range of difficulty fluctuations. At the same time, a difficulty range threshold is preset, and after generation, the difficulty assessment model is used to filter and retain variant questions whose difficulty coefficient is within the specified range.

[0010] Among the steps, the target incorrect question is input into a trained variation generation model to generate several variation questions that are semantically similar to the incorrect question but different in form, and the difficulty coefficient of the variation questions is controlled: The difficulty coefficient is calculated by weighting three features: the length of the question text, the number of knowledge points involved, and the average time taken to answer the question. The weight of each feature is determined by performing linear regression fitting on historical question data.

[0011] Among the steps, the generated variation questions are pushed to students for targeted training, students' performance on the variation questions is collected, and students' mastery scores of the corresponding knowledge points are updated as follows: A Bayesian knowledge tracing model is used to update the mastery score. The model input is the correct or incorrect result of the variant question, and the output is the updated mastery probability. This mastery score is used for dynamic adjustment of the difficulty of subsequent variation questions.

[0012] Among them, the step of dynamically adjusting the priority weight of each wrong question in the error notebook based on the answer results of the variant questions, and prioritizing the recommendation of repeatedly wrong questions for further variant generation: The priority weights are dynamically updated based on the average accuracy rate of the variant questions; the lower the average accuracy rate, the greater the weight increment. If you answer the variation question correctly three times in a row, remove the incorrect question from your error notebook.

[0013] The process includes the steps of sending generated variation questions to students for targeted practice, collecting students' answers to the variation questions, and updating students' scores on their mastery of the corresponding knowledge points: The students' mastery of each knowledge point is presented in the form of a radar chart or tree diagram, and the knowledge points that cannot be improved even after repeated variation training are highlighted in the chart, and teachers are advised to provide manual guidance.

[0014] This invention also provides a personalized AI learning system based on learning progress diagnosis, including a wrong question collection module, a comparative learning modeling module, a variation generation module, a targeted training module, and a wrong question notebook updating module; wherein: The error collection module collects students' error data during the learning process and extracts the knowledge point tags and error pattern features for each error. The contrastive learning modeling module: Based on the contrastive learning framework, it constructs positive and negative sample pairs between incorrect and correct questions under the same knowledge point, trains the variant generation model, and enables the model to capture the semantic differences between incorrect questions and correct answers; The variant generation module: inputs the target incorrect question into the trained variant generation model, generates several variant questions that are semantically similar to the incorrect question but different in form, and controls the difficulty coefficient of the variant questions; The targeted training module: pushes the generated variant questions to students for targeted training, collects students' answers to the variant questions, and updates students' mastery scores of the corresponding knowledge points. The error notebook update module dynamically adjusts the priority weight of each wrong question in the error notebook based on the answer results of the variant questions, and prioritizes recommending questions that are repeatedly wrong to generate variant versions again.

[0015] This invention discloses a personalized AI learning system and method based on learning progress diagnosis. It collects students' incorrect answer data during the learning process, extracts knowledge point tags and error pattern features for each incorrect answer, and, based on a contrastive learning framework, constructs positive and negative sample pairs between incorrect and correct answers for the same knowledge point to train a variation generation model. This model enables the model to capture the semantic differences between incorrect and correct answers. The target incorrect answer is input into the trained variation generation model to generate several variation questions that are semantically similar to the incorrect answer but different in form, and the difficulty level of the variation questions is controlled. The generated variation questions are pushed to students for targeted training, and students' performance on these variation questions is collected to update their mastery scores for the corresponding knowledge points. Based on the results of answering the variation questions, the priority weight of each incorrect answer in the error notebook is dynamically adjusted, prioritizing the recommendation of repeatedly incorrect questions for further variation generation. Therefore, by training a variation generation model using a comparative learning framework, the system can automatically extract semantic features from incorrect questions and generate variation questions that are semantically similar but different in form. This allows students to move beyond rote memorization of the original incorrect answers and truly understand the problem-solving logic behind the knowledge points, significantly improving their knowledge transfer ability. The difficulty coefficient of the variation questions is dynamically controlled through temperature parameters and a difficulty assessment model, ensuring that the generated questions are challenging yet within the students' current ability range, avoiding frustration or ineffective training. A Bayesian knowledge tracking model is used to update students' mastery scores of knowledge points in real time, and the priority weight of each incorrect question in the error notebook is dynamically adjusted based on the performance of answering variation questions. This allows the system to automatically focus on the weak knowledge points that students repeatedly make mistakes on, achieving dynamic iteration of the error notebook and optimization of personalized training paths. Simultaneously, the mastery of each knowledge point is presented through a visual radar chart or tree diagram, highlighting difficult-to-overcome weaknesses, facilitating teacher intervention and forming a human-computer collaborative teaching loop. This invention can be widely applied in online education, intelligent question banks, adaptive learning systems, and other scenarios, effectively improving students' learning efficiency and teachers' targeted teaching. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of the steps of the personalized AI learning method based on learning diagnosis of the present invention.

[0018] Figure 2 This is a flowchart of steps S100 of the present invention.

[0019] Figure 3 This is a flowchart of steps S200 of the present invention.

[0020] Figure 4 This is a flowchart of steps S300 of the present invention.

[0021] Figure 5 This is a flowchart of steps S400 of the present invention.

[0022] Figure 6 This is a flowchart of steps S500 of the present invention.

[0023] Figure 7 This is a structural diagram of the personalized AI learning system based on learning progress diagnosis of the present invention.

[0024] Figure 8 This is a schematic diagram of the electronic device of the present invention.

[0025] 601 - Error Collection Module, 602 - Comparative Learning Modeling Module, 603 - Variation Generation Module, 604 - Targeted Training Module, 605 - Error Notebook Update Module. Detailed Implementation

[0026] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0027] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms "a," "say," and "this" as used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms used herein refer to and / or include any or all possible combinations of one or more associated listed items.

[0028] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, words used herein may be interpreted as meaning "when..." or "when..." or "in response to determination".

[0029] Please see Figures 1-6 This invention provides a personalized AI learning method based on learning progress diagnosis, comprising the following steps: S100: Collect data on students' incorrect answers during the learning process, and extract the knowledge point tags and error pattern features for each incorrect answer; In this implementation, data on students' incorrect answers during the learning process is collected, and the knowledge point tags and error pattern features of each incorrect answer are extracted. Specifically, this includes: S101: The knowledge point tags are automatically labeled using a pre-constructed subject knowledge graph; S102: The error pattern features include four types: calculation error, conceptual confusion, question reading deviation, and logical jump. These features are extracted from the answers to incorrect questions through rule matching or classification models.

[0030] The process involves extracting structured knowledge point labels and error pattern features from raw incorrect question data to provide input for subsequent comparative learning and variation generation. During the pre-construction of the subject knowledge graph, knowledge points in the curriculum standards are broken down into three levels of nodes (chapter-section-knowledge point). Each question is automatically labeled by matching keywords in the question stem to graph nodes; if a match fails, it is manually labeled once and the matching rules are updated. Error pattern features include: calculation errors are identified by detecting a deviation of less than 5% between the answer and the correct answer, and the formula must be correct; concept confusion is identified by comparing keywords of distractors in the answer options with common confused concepts of the corresponding knowledge point; question comprehension deviation is identified by extracting students' markings or the distribution of answer time (normal comprehension time is 3-5 seconds, deviation is less than 1 second or greater than 10 seconds); logical jumps are identified by the edit distance between the step jump sequence during the answer process and the standard solution steps. Regular expressions are used preferentially for rule matching. If new question types or new errors appear, they are collected as training samples. After accumulating 100 examples, a Naive Bayes classifier is retrained as a fallback model. The extracted tags and features are stored in JSON format. Each record contains the wrong question ID, knowledge point path, error type, and original answer text fragment.

[0031] S200: Based on the contrastive learning framework, positive and negative sample pairs are constructed by comparing incorrect and correct questions under the same knowledge point, and the variant generation model is trained to enable the model to capture the semantic differences between incorrect questions and correct answers. In this implementation, based on a contrastive learning framework, positive and negative sample pairs are constructed by comparing incorrect and correct answers for the same knowledge point. These pairs are then used to train a variant generation model, enabling the model to capture the semantic differences between incorrect and correct answers. Specifically, this includes: S201: It adopts an encoder-decoder architecture. The encoder encodes the question text into a semantic vector. The contrastive learning loss function makes the distance between the wrong question vector and the correct question vector closer and the distance between the wrong question vector and the random question vector farther. S202: The decoder takes the error vector and difference vector as input to generate variant question text; S203: The difference vector is the semantic difference encoding between the target wrong question and the correct question. It is obtained by subtracting the wrong question vector from the correct question vector and then mapping it to the latent space through a fully connected layer.

[0032] This study employs a contrastive learning approach to train a variation generation model capable of capturing the semantic differences between incorrect and correct questions. First, a training dataset is constructed: questions covering the same knowledge point are selected from historical answer records. Questions answered incorrectly by students are designated as incorrect samples, the question with the highest correct answer rate for the same knowledge point is designated as a positive sample, and questions from other knowledge points are randomly selected as negative samples. The encoder uses a pre-trained BERT model to encode the question text into a 768-dimensional semantic vector. During training, a contrastive loss function is used to gradually decrease the cosine distance between the incorrect question vector and the positive sample vector, and gradually increase the distance with the negative sample vector for the same knowledge point. An additional constraint loss is applied in the direction of the difference between the incorrect question vector and the positive sample vector to ensure that the difference vector stably converges to an interpretable semantic offset space. The difference vector is obtained by subtracting the positive sample vector from the incorrect question vector, and then mapped to a latent space through a two-layer fully connected network (384 hidden dimensions, 256 output dimensions). This latent space controls the modification direction during variation generation. The decoder employs a Transformer-based generative model, using a concatenated vector of incorrect questions and a difference vector as input to generate variant question text via autoregression. During training, a strategy combining teacher-mandated and planned sampling is used. Teacher-mandated sampling accelerates convergence in the first 100 rounds, while planned sampling is gradually increased to 0.3 in the last 50 rounds to enhance the model's generalization ability. After training, the model can output variant questions that are semantically similar but have different expressions based on the input incorrect questions. For example, "Xiaoming has 5 apples, gives 2 to Xiaohong, how many are left?" can be variantted as "Xiaohua has 7 oranges, gives 3 to Xiaoli, how many are left?".

[0033] Among them, the difference vector is essentially a semantic deviation representation between incorrect questions and standard correct questions on the same knowledge point. Physically, it corresponds to the direction of cognitive deficiencies reflected by students' incorrect answers. Specifically, it covers the quantitative coding of dimensions such as misunderstanding of the question stem, misuse of concepts, and loopholes in problem-solving logic. Its position in the latent space directly points to the cognitive weaknesses that need to be corrected. It can accurately guide the variation generation model to modify the semantic elements in the questions that are prone to causing student errors while retaining the core knowledge points. The supervision signals during the training phase consist of three layers. The first layer is global supervision of the contrastive learning loss, which constrains the overall semantic direction of the difference vector by bringing the error vector closer to the positive sample vector and distancing it from the negative sample vector. The second layer is constraint loss supervision of the difference vector, which applies L2 regularization constraints in the direction of the difference between the error vector and the positive sample vector to ensure that the difference vector converges to a stable and interpretable semantic offset space and avoids meaningless random offsets. The third layer is cross-entropy loss supervision of the generation task, which uses high-quality variant questions with the same knowledge points labeled by humans as tags. The supervised decoder generates variant questions that meet the semantic similarity requirements based on the error vector and the difference vector, while inversely optimizing the encoding accuracy of the difference vector, so that the difference vector can accurately drive the model to generate targeted modified variant questions.

[0034] S300: Input the target incorrect question into the trained variation generation model to generate several variation questions that are semantically similar to the incorrect question but different in form, and control the difficulty coefficient of the variation questions; In this implementation, the target incorrect question is input into a trained variation generation model to generate several variation questions that are semantically similar to the incorrect question but different in form, and the difficulty level of the variation questions is controlled; specifically including: S301: The difficulty of the variant questions can be controlled by adjusting the temperature parameters or Top-k sampling values ​​during the generation process. The higher the temperature, the greater the variation in the generated questions and the wider the range of difficulty fluctuations. S302: Simultaneously, a difficulty range threshold is preset, and after generation, variant questions with difficulty coefficients within the specified range are filtered and retained through a difficulty assessment model; S303: The difficulty coefficient is calculated by weighting three features: the length of the question text, the number of knowledge points involved, and the average answering time. The weight of each feature is determined by performing linear regression fitting on historical question data.

[0035] The process involves using a trained variation generation model to generate multiple variation questions with controllable difficulty for the target incorrect question. During inference, a temperature parameter T=0.8 (default value) is set to control the randomness of generation; the higher the T, the flatter the word probability distribution and the greater the variation in generated text. Top-k sampling uses k=40, sampling only from the 40 words with the highest probability each time. Difficulty control is achieved through temperature adjustment: at low difficulty, T=0.4 generates variations with small changes and close to the original question; at high difficulty, T=1.2 generates larger variations that change quantifiers, scenarios, and numerical ranges. A preset difficulty range threshold of 0.3~0.7 (normalized value) is also used. After generation, a difficulty assessment model is called to calculate the difficulty coefficient of each variation question, and questions within this range are selected and retained. The difficulty assessment model is based on linear regression, with three features: question text length (number of characters, coefficient 0.2), number of knowledge points (statistics from the knowledge graph, coefficient 0.5), and historical average answer time (seconds, coefficient 0.3), with an intercept term of 1.2. Each coefficient was obtained by performing a least-squares fit on 500 questions (including those with known difficulty labels) from a historical question bank. The coefficient of determination R0 was calculated. 2 =0.86. For each incorrect question, 5 variation questions are generated by default. After difficulty filtering, 2-3 of these variation questions are selected and sent to the student.

[0036] S400: The generated variant questions are pushed to students for targeted training, students' performance on the variant questions is collected, and the students' mastery scores of the corresponding knowledge points are updated. In this implementation, the generated variation questions are pushed to students for targeted training, their performance on these variation questions is collected, and their scores on mastery of the corresponding knowledge points are updated. Specifically, this includes: S401: The mastery score is updated using a Bayesian knowledge tracing model. The model input is the correct or incorrect result of the variant question, and the output is the updated mastery probability. S402: This mastery score is used for dynamic adjustment of the difficulty of subsequent variation questions.

[0037] The system dynamically updates students' mastery scores for knowledge points based on their correct and incorrect answers to variation questions. During model initialization, the prior mastery probability for each knowledge point is set to 0.3, with guessing and error parameters obtained from historical data (default guessing probability 0.2, error probability 0.1). After each variation question is pushed, the system records the student's answer (1 for correct, 0 for incorrect). The model updates the mastery probability using the following recursive formula: first, calculate the probability that the student answers the question correctly at their current mastery level; then, calculate the new mastery probability after answering using Bayes' theorem. The updated mastery probability serves as the latest mastery score for that knowledge point, ranging from 0 to 1. After completing three variation questions on the same knowledge point, if the mastery score increases twice consecutively and exceeds 0.7, the target difficulty range is automatically increased by 0.1 when generating the next variation question; if the mastery score is below 0.4 and decreases twice consecutively, the target difficulty range is decreased by 0.1, making the variation questions more closely reflect the student's current level. Once a student's mastery score exceeds 0.85, the knowledge point is marked as "mastered," and no further variation questions will be generated for that knowledge point unless the student makes a mistake again in the comprehensive test. All mastery score changes are stored in a time-series database for generating learning trend charts.

[0038] S500: Based on the answers to the variation questions, dynamically adjust the priority weight of each wrong question in the error notebook, and prioritize recommending questions that are repeatedly wrong for further variation generation.

[0039] In this implementation, based on the answers to the variation questions, the priority weight of each incorrect question in the error notebook is dynamically adjusted, prioritizing the generation of variations for questions that are repeatedly answered incorrectly. Specifically, this includes: S501: Priority weights are dynamically updated based on the average accuracy rate of the variant questions. The lower the average accuracy rate, the greater the weight increment. S502: If a variant question is answered correctly three times in a row, the incorrect question will be removed from the error notebook.

[0040] The system dynamically adjusts the priority weight of each incorrect question in the error log based on students' performance on variation questions, prioritizing questions that are repeatedly answered incorrectly for secondary variation generation. Each incorrect question has an initial weight of 1.0. After each variation question is answered, its correctness or incorrectness is recorded. During batch processing, the average accuracy rate of the corresponding variation question set for that incorrect question is calculated (the number of correct answers divided by the total number of variation questions generated for that incorrect question). The weight update rule is as follows: if the average accuracy rate is below 0.3, the weight increases by 0.5; if it is between 0.3 and 0.6, the weight increases by 0.2; and if it is above 0.6, the weight remains unchanged. The updated weights are used to sort the error log. Each time, the system prioritizes the top 5 incorrect questions with the highest weights for the next round of variation generation. A decay factor is also set: after each iteration, the weight of all incorrect questions is multiplied by 0.95 to prevent old incorrect questions from occupying high positions for extended periods. When a student answers three consecutive variations of a given incorrect question correctly (regardless of the batch generated, in chronological order), it is determined that the student has essentially mastered the knowledge point involved in that incorrect question. The system then permanently removes the incorrect question from the error log and records the final weight and mastery score at the time of removal for subsequent analysis model calibration. If the student makes the same mistake again on a new question after removal, the incorrect question is added back to the error log with an initial weight of 0.8.

[0041] Corresponding to the aforementioned embodiments of personalized AI learning methods based on learning progress diagnosis, this application also provides embodiments of personalized AI learning systems based on learning progress diagnosis.

[0042] Figure 7 This is a block diagram illustrating a personalized AI learning system based on learning progress diagnosis, according to an exemplary embodiment. (Refer to...) Figure 7 The system may include: a wrong question collection module 601, a comparative learning modeling module 602, a variation generation module 603, a targeted training module 604, and a wrong question notebook update module 605, wherein: The error collection module 601 collects students' error data during the learning process and extracts the knowledge point tags and error pattern features of each error. The contrastive learning modeling module 602: Based on the contrastive learning framework, constructs positive and negative sample pairs between incorrect and correct questions under the same knowledge point, trains the variant generation model, and enables the model to capture the semantic differences between incorrect questions and correct answers. The variant generation module 603: inputs the target wrong question into the trained variant generation model, generates several variant questions that are semantically similar to the wrong question but different in form, and controls the difficulty coefficient of the variant questions; The targeted training module 604: pushes the generated variant questions to students for targeted training, collects students' answers to the variant questions, and updates students' mastery scores of the corresponding knowledge points. The error notebook update module 605 dynamically adjusts the priority weight of each error in the error notebook based on the answer results of the variant questions, and prioritizes recommending repeatedly wrong questions for re-generation of variant questions.

[0043] In this embodiment, the error collection module 601 collects students' error data during the learning process and extracts the knowledge point tags and error pattern features of each error; the comparative learning modeling module 602, based on the comparative learning framework, constructs positive and negative sample pairs between error questions and correct questions under the same knowledge point, trains the variation generation model, and enables the model to capture the semantic differences between error questions and correct answers; the variation generation module 603 inputs the target error question into the trained variation generation model to generate several variation questions that are semantically similar to the error question but different in form, and controls the difficulty coefficient of the variation questions; the targeted training module 604 pushes the generated variation questions to students for targeted training, collects students' answer performance on the variation questions, and updates students' mastery scores of the corresponding knowledge points; the error notebook update module 605 dynamically adjusts the priority weight of each error question in the error notebook according to the answer results of the variation questions, and prioritizes recommending repeatedly wrong questions for further variation generation.

[0044] Regarding the system in the above embodiments, the specific ways in which each module performs operations have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0045] For the system embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0046] Accordingly, this application also provides an electronic device, including: one or more processors; a memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the personalized AI learning method based on learning diagnosis as described above. Figure 8 The diagram shown is a hardware structure diagram of any device with data processing capabilities, used in an embodiment of the present invention to provide a personalized AI learning system based on learning assessment. (Except for...) Figure 8In addition to the processor, memory, and network interface shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0047] Accordingly, this application also provides a computer-readable storage medium storing computer instructions, which, when executed by a processor, implement the personalized AI learning method based on learning diagnosis as described above. The computer-readable storage medium can be an internal storage unit of any data-processing device as described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data-processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data-processing device, and can also be used to temporarily store data that has been output or will be output.

[0048] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0049] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A personalized AI learning method based on learning situation diagnosis, characterized in that, Includes the following steps: Collect data on students' incorrect answers during the learning process, and extract the knowledge point tags and error pattern characteristics for each incorrect answer; Based on the contrastive learning framework, positive and negative sample pairs are constructed by comparing incorrect and correct questions under the same knowledge point, and the variant generation model is trained to enable the model to capture the semantic differences between incorrect questions and correct answers. The target incorrect question is input into the trained variation generation model to generate several variation questions that are semantically similar to the incorrect question but different in form, and the difficulty coefficient of the variation questions is controlled. The generated variant questions are pushed to students for targeted training, and students' performance on the variant questions is collected to update their mastery scores of the corresponding knowledge points. Based on the answers to the variation questions, the priority weight of each wrong question in the error notebook is dynamically adjusted, and questions that are repeatedly wrong are given priority for further variation generation.

2. The AI learning method based on the diagnosis of learning situation according to claim 1, wherein, In the steps of collecting students' incorrect answers during the learning process and extracting the knowledge point tags and error pattern features for each incorrect answer: The knowledge point tags are automatically labeled using a pre-constructed subject knowledge graph; The error pattern features include four types: calculation errors, conceptual confusion, question interpretation deviation, and logical jumps. These features are extracted from the answers to incorrect questions through rule matching or classification models.

3. The personalized AI learning method based on learning assessment as described in claim 1, characterized in that, In the step of constructing positive and negative sample pairs between incorrect and correct questions for the same knowledge point based on a contrastive learning framework, and training a variant generative model to enable the model to capture the semantic differences between incorrect and correct answers: An encoder-decoder architecture is adopted. The encoder encodes the question text into a semantic vector, and the contrastive learning loss function reduces the distance between the incorrect question vector and the correct question vector, and increases the distance between the incorrect question vector and the random question vector. The decoder takes the error vector and difference vector as input and generates variant question text.

4. The personalized AI learning method based on learning assessment as described in claim 1, characterized in that, In the step of constructing positive and negative sample pairs between incorrect and correct questions for the same knowledge point based on a contrastive learning framework, and training a variant generative model to enable the model to capture the semantic differences between incorrect and correct answers: The difference vector is a semantic difference encoding between the target incorrect question and the correct question. It is obtained by subtracting the incorrect question vector from the correct question vector and then mapping it to the latent space through a fully connected layer.

5. The personalized AI learning method based on learning assessment as described in claim 1, characterized in that, In the steps of inputting the target incorrect question into the trained variation generation model, generating several variation questions that are semantically similar to the incorrect question but different in form, and controlling the difficulty level of the variation questions: The difficulty of the variant questions can be controlled by adjusting the temperature parameters or Top-k sampling values ​​during the generation process. The higher the temperature, the greater the variation in the generated questions and the wider the range of difficulty fluctuations. At the same time, a difficulty range threshold is preset, and after generation, the difficulty assessment model is used to filter and retain variant questions whose difficulty coefficient is within the specified range.

6. The personalized AI learning method based on learning assessment as described in claim 1, characterized in that, In the steps of inputting the target incorrect question into the trained variation generation model, generating several variation questions that are semantically similar to the incorrect question but different in form, and controlling the difficulty level of the variation questions: The difficulty coefficient is calculated by weighting three features: the length of the question text, the number of knowledge points involved, and the average time taken to answer the question. The weight of each feature is determined by performing linear regression fitting on historical question data.

7. The personalized AI learning method based on learning assessment as described in claim 1, characterized in that, In the steps of pushing the generated variation questions to students for targeted practice, collecting students' answers to the variation questions, and updating students' mastery scores of the corresponding knowledge points: A Bayesian knowledge tracing model is used to update the mastery score. The model input is the correct or incorrect result of the variant question, and the output is the updated mastery probability. This mastery score is used for dynamic adjustment of the difficulty of subsequent variation questions.

8. The personalized AI learning method based on learning assessment as described in claim 1, characterized in that, In the step of dynamically adjusting the priority weight of each wrong question in the error notebook based on the answer results of the variation questions, and prioritizing the recommendation of repeatedly wrong questions for further variation generation: The priority weights are dynamically updated based on the average accuracy rate of the variant questions; the lower the average accuracy rate, the greater the weight increment. If you answer the variation question correctly three times in a row, remove the incorrect question from your error notebook.

9. The personalized AI learning method based on learning assessment as described in claim 1, characterized in that, After the steps of pushing the generated variation questions to students for targeted practice, collecting students' answers to the variation questions, and updating students' mastery scores of the corresponding knowledge points are completed: The students' mastery of each knowledge point is presented in the form of a radar chart or tree diagram, and the knowledge points that cannot be improved even after repeated variation training are highlighted in the chart, and teachers are advised to provide manual guidance.

10. A personalized AI learning system based on learning assessment, employing the personalized AI learning method based on learning assessment as described in claim 1, characterized in that, It includes a module for collecting incorrect answers, a module for comparative learning and modeling, a module for generating variations, a module for targeted training, and a module for updating the incorrect answer notebook; among which: The error collection module collects students' error data during the learning process and extracts the knowledge point tags and error pattern features for each error. The contrastive learning modeling module: Based on the contrastive learning framework, it constructs positive and negative sample pairs between incorrect and correct questions under the same knowledge point, trains the variant generation model, and enables the model to capture the semantic differences between incorrect questions and correct answers; The variant generation module: inputs the target incorrect question into the trained variant generation model, generates several variant questions that are semantically similar to the incorrect question but different in form, and controls the difficulty coefficient of the variant questions; The targeted training module: pushes the generated variant questions to students for targeted training, collects students' answers to the variant questions, and updates students' mastery scores of the corresponding knowledge points. The error notebook update module dynamically adjusts the priority weight of each wrong question in the error notebook based on the answer results of the variant questions, and prioritizes recommending questions that are repeatedly wrong to generate variant versions again.

Images