Error diagnosis classification and intelligent review planning method based on multi-dimensional analysis

Abstract

The invention discloses an error diagnosis classification and intelligent review planning method based on multi-dimensional analysis, which is suitable for a personalized learning support system. The method comprises six steps of error collection, automatic labeling, error attribution, capability evaluation, knowledge point correlation analysis and dynamic review planning, and integrates OCR recognition, TF-IDF feature extraction, convolutional neural network classification, knowledge graph modeling and Q-learning intelligent strategies. The system can realize accurate classification and reason diagnosis of wrong questions based on answering behaviors and cognitive features of students, optimizes a review plan in combination with a forgetting curve and habit error, supports an online and offline mixed review mode, and significantly improves learning efficiency and personalized adaptation ability.

Description

Technical Field

[0001] The present invention belongs to the field of educational information processing technology, and specifically provides a method for diagnosing and classifying wrong questions and intelligent review planning based on multi-dimensional analysis. Background Art

[0002] With the development of educational informatization, the management of wrong questions has gradually shifted from traditional paper records to electronic management. Some online education platforms or intelligent learning systems have initially acquired functions such as wrong question recording, knowledge point classification, and review reminders. However, existing wrong question systems generally still have the following technical deficiencies:

[0003] First, existing methods for sorting out incorrect questions often rely on manual copying or image uploads. While some products have incorporated OCR technology, these methods only convert images into text and fail to automatically categorize and analyze incorrect questions. This results in low utilization of incorrect question data and coarse categorization, making it difficult to support systematic learning.

[0004] Secondly, existing methods for categorizing incorrect questions are mostly based on question types or superficial knowledge point labels. They lack in-depth understanding of the causes of students' errors and cannot distinguish between different types of cognitive errors (such as conceptual confusion, reasoning errors, and careless calculations). This makes it difficult for the system to provide review suggestions that match students' true weaknesses, affecting learning outcomes.

[0005] Thirdly, mainstream wrong question review systems mostly use fixed-cycle or user-defined review reminder mechanisms, without considering multi-dimensional factors such as the forgetting curve, students' individual memory characteristics, the difficulty of wrong questions and the type of errors. They cannot achieve dynamic and intelligent planning of review timing and content, resulting in inefficient review or ineffective repetition.

[0006] In addition, the existing system is relatively weak in assessing students' learning abilities. It usually only counts the number of errors or the accuracy rate, and lacks modeling and tracking of students' knowledge mastery structure and cognitive ability dimensions. It is difficult to assist students in establishing a systematic cognitive self-cognition model, and it is also difficult to provide teachers with in-depth teaching analysis support.

[0007] To sum up, the existing wrong question management system has technical defects to varying degrees in terms of wrong question data collection, error cause analysis, review path optimization and ability visualization modeling. There is an urgent need for a new wrong question analysis and review planning method that combines artificial intelligence technology, data modeling technology and cognitive psychology theory to achieve a more accurate, smarter and more personalized learning support system. Summary of the Invention

[0008] The purpose of the present invention is to provide a method for diagnosing and classifying wrong questions and intelligent review planning based on multi-dimensional analysis. The system of the present invention is superior to existing wrong question management and review technologies in many core indicators such as error diagnosis accuracy, knowledge point classification accuracy, knowledge retention rate after review, and student usage stickiness, and has significant educational effectiveness and application promotion value.

[0009] The technical solution adopted in the present invention is as follows:

[0010] A method for diagnosing and classifying wrong questions and intelligent review planning based on multi-dimensional analysis includes the following steps:

[0011] (1) Wrong question collection: Use smart terminals or mobile devices to capture wrong question images and use OCR recognition technology to extract wrong question text and accompanying image data;

[0012] (2) Wrong question annotation: Automatically annotate the extracted text and image data to generate subject information, knowledge points, and difficulty level attributes;

[0013] (3) Error diagnosis: The feature vector of the wrong question text and the student's answering behavior data are input into the error classification model based on the convolutional neural network, and the error type label is output;

[0014] (4) Ability assessment: Calculate scores under multiple ability dimensions based on student grade, unit test scores, wrong question attribution, and review records, and update them dynamically;

[0015] (5) Knowledge point relevance analysis: Using the improved Word2Vec model and knowledge graph, we construct knowledge point vectors and calculate their cosine similarity and co-occurrence frequency to evaluate the relevance between knowledge points;

[0016] (6) Dynamic review planning: Based on the student's personalized forgetting curve parameters, wrong question attributes and historical review records, an improved forgetting curve algorithm or Q-learning algorithm is used to generate a personalized review plan and recommended time.

[0017] The OCR recognition technology has an accuracy of no less than 98%, and the recognition process records metadata to assist in the time-series management of wrong question sorting.

[0018] The wrong question text feature vector is generated by the TF-IDF weighted algorithm, and the answering behavior data includes the operation time period, the duration of each operation and the number of review times.

[0019] The error classification model adopts the following algorithm: Error_Type = α×Semantic_Similarity+β×Operation_Sequence_Deviation+γ×Difficulty_Level+δ×Error_Pattern, where α, β, γ, and δ are weight factors determined through training.

[0020] The ability assessment includes six dimensions: knowledge comprehension, standardized expression ability, logical reasoning, transfer application, test-taking strategies and data analysis ability.

[0021] The ability score calculation includes:

[0022] Unit ability score = basic score + test score + ∑ (review bonus) - ∑ (wrong question deduction),

[0023] Total score = the sum of the ability scores of each unit.

[0024] The calculation formula for the knowledge point relevance is:

[0025] Association_Score(A, B)=Cosine_Similarity(Vector(A), Vector(B))×Co_occurrence_Frequency(A, B)+λ×Knowledge_Graph_Link(A, B).

[0026] The dynamic review planning model is implemented using the following Q-learning algorithm:

[0027] Q(s, a)=(1-α)×Q(s, a)+α×[R(s, a)+γ×maxQ(s′, a′)],

[0028] Where s is the current state, a is the review action, R is the immediate reward, and γ is the discount factor.

[0029] The review suggestions include a list of wrong questions and recommended review time nodes, and the suggestions are personalized based on the Ebbinghaus memory curve.

[0030] Students can view their review plans or print review lists through terminal devices, supporting a mixed online and offline review method.

[0031] The beneficial effects of the present invention are as follows:

[0032] The present invention provides a method for diagnosing and classifying wrong questions and intelligent review planning based on multi-dimensional analysis, which significantly improves the intelligence level of the wrong question management system and the accuracy of learning guidance, and has the following beneficial effects:

[0033] First, by integrating OCR image recognition with metadata recording, this invention enables the automatic collection and structured processing of incorrect answers, avoiding the inefficiency, omissions, and difficulty in retrieval associated with traditional manual transcription and sorting of error books. OCR recognition accuracy exceeds 98%, and metadata such as the time of capture and the context of the answer are automatically recorded, providing reliable data support for subsequent error analysis and significantly improving the efficiency and quality of error compilation.

[0034] Secondly, the present invention introduces a convolutional neural network (CNN)-based error diagnosis model, which can integrate multiple factors such as the text features of the error question, student operation behavior, and question difficulty to output specific error type labels. Compared with the existing method of superficial classification based only on knowledge points or question types, the present invention can identify multiple error patterns such as "knowledge confusion type", "logical loophole type", "calculation deviation type", and "question error type", thus extending the error attribution analysis from results to processes, and from phenomena to in-depth cognitive mechanisms.

[0035] Furthermore, the present invention constructs a multi-dimensional ability assessment mechanism and proposes a personalized learning portrait model covering six major ability dimensions, including: knowledge comprehension, standard expression ability, logical reasoning, transfer application, test-taking strategy and data analysis ability. The ability dimension is set for all subjects and has good generalization adaptability. The system dynamically calculates the ability scores of each dimension and its changing trend over time by integrating information such as students' scores in unit tests, the attribution and type of wrong questions, and review behavior records (such as frequency and completion), constructs a visual ability portrait, assists students in identifying cognitive weaknesses, accurately locates learning shortcomings, and guides personalized learning path planning for strengthening and compensating weaknesses. Compared with the static feedback method in the prior art that only counts the number of wrong questions or the accuracy rate, the present invention can provide dynamic trend analysis and deep ability modeling in the cognitive dimension, which greatly improves the interpretability, accuracy and guiding significance of learning feedback.

[0036] To address the issue of rigid review plans and their neglect of individual differences, this invention leverages the Ebbinghaus forgetting curve and the Q-learning reinforcement learning algorithm. This approach dynamically generates review time and content recommendations based on factors such as the student's personalized memory model, the difficulty of the wrong questions, the type of error, and review frequency. Compared to fixed-period review methods, this model can push review tasks within the optimal memory retention window, significantly improving review effectiveness and memory retention.

[0037] In addition, the present invention uses Word2Vec to construct a knowledge point vector space. Through dimensions such as the semantic similarity and co-occurrence frequency between knowledge points and the structural relationship of the knowledge graph, it evaluates the intrinsic correlation between wrong questions and between wrong questions and knowledge points. It can realize the clustering, migration and induction of wrong questions across knowledge points, and enhance students' cognitive ability of the overall knowledge system.

[0038] In terms of incentive mechanism, the present invention designs a points and feedback system linked to ability improvement, such as rewards for entering wrong questions, rewards for review achievement, test progress feedback, etc., to stimulate students' learning initiative and system usage stickiness, and significantly improve the problem of lack of internal driving force in existing wrong question systems.

[0039] Data from randomized controlled experiments show that the system of the present invention is superior to existing wrong question management and review technologies in many core indicators such as error diagnosis accuracy, knowledge point classification accuracy, knowledge retention rate after review, and student usage stickiness, and has significant educational effectiveness and application promotion value. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a system architecture design diagram of the present invention;

[0041] Figure 2 Schematic diagram of the method flow of the present invention;

[0042] Figure 3 A schematic diagram of the method for marking wrong questions of the present invention;

[0043] Figure 4 A schematic diagram of the error diagnosis step of the present invention;

[0044] Figure 5 A schematic diagram of the capability assessment step of the present invention;

[0045] Figure 6 Schematic diagram of the knowledge point relevance analysis step of the present invention;

[0046] Figure 7 Schematic diagram of the method for dynamically reviewing planning steps of the present invention. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. Figure 1 and 2 The present invention provides a method for diagnosing and classifying wrong questions and intelligent review planning based on multi-dimensional analysis. It combines OCR recognition, deep learning, knowledge graph, behavioral data analysis and personalized learning engine to realize the full-process intelligent management of students' wrong questions from collection, analysis, attribution to review planning, with strong real-time performance, adaptability and learning efficiency.

[0048] The method comprises the following steps:

[0049] (1) Wrong question collection: Use smart terminals or mobile devices to capture wrong question images, and use OCR recognition technology to extract the wrong question text and image data. The OCR recognition accuracy is not less than 98%, and metadata such as shooting time, user ID, and device information are recorded to assist in the time sequence management of wrong question sorting;

[0050] In actual application, the implementation process of the wrong question collection module is as follows:

[0051] 1) Image acquisition interface design: The system uses the app or smart terminal to call the camera to automatically identify the paper boundary, perform image cropping, automatic focusing and stabilization processing, and capture clear images of error questions;

[0052] 2) OCR Recognition Engine: The system uses commercial or proprietary OCR algorithms (such as CRNN+CTC architecture) to extract text from captured images, including question stems, options, formulas, and illustrations, and outputs structured content.

[0053] 3) Image preprocessing: Perform optimization processing such as edge detection, grayscale enhancement, angle correction, and image denoising before OCR to improve recognition accuracy;

[0054] 4) Metadata Collection: The system automatically and synchronously records metadata such as shooting time, user ID, terminal identifier, image hash code, etc., which is used for tracing, sorting, and deduplication of wrong questions;

[0055] 5) Data synchronization and caching: The collected content is uploaded to the cloud server through the API, supporting local caching and breakpoint retransmission mechanism to ensure data integrity.

[0056] (2) Marking of wrong questions: see Figure 3 , automatically annotate the extracted text and illustration data, generate subject information, knowledge points and difficulty level attributes, and construct a text feature vector, wherein the wrong question text feature vector is generated by the TF-IDF weighted algorithm;

[0057] The specific implementation is as follows:

[0058] 1) Text normalization: Normalize the question stem and option text after OCR recognition, including removing redundant spaces, standardizing punctuation, and encoding mathematical formulas in a unified manner;

[0059] 2) Subject Identification: After the OCR text is normalized, the system performs enhanced recognition using the following steps:

[0060] ① Preliminary screening (rule-based): Use regular rules and keyword libraries to quickly categorize subjects and question types. For example, "function," "chemical reaction," and "incorrect sentence" can be categorized as mathematics, chemistry, and Chinese, respectively. This allows for the initial identification of possible subject clusters and the marking of key phrase segments.

[0061] ② Deep Semantic Analysis (Transformer Nesting): Introducing a text classification model based on the Transformer architecture (such as BERT or RoBERTa). This model takes the question stem as input, automatically learns its contextual semantic features, and outputs subject probability distributions and question type labels (multiple-choice, fill-in-the-blank, proof, etc.) by fine-tuning the multi-label classification head. Training data includes manually annotated stem-subject-question type triplets to enhance generalization capabilities.

[0062] ③Contextual fusion enhancement (paragraph / question group analysis): When the input question is a paragraph or question group on the test paper page, the system will simultaneously process the question number, question group title and previous question content, and use a sliding window method to form context-enhanced input to improve classification accuracy in the context of multiple questions shared.

[0063] ④Confidence fusion mechanism: The final recognition result is determined by both rule judgment and model prediction. If the highest confidence subject output by the model matches the rule, the final credibility is improved; if there is a conflict, voting is performed based on the confidence weight or a prompt "low confidence requires manual review" is given;

[0064] 3) Knowledge Point Matching: The knowledge graph is compared with the keyword library to extract knowledge point labels that match the question stem's semantics. Semantic similarity is calculated and the most relevant matching result is presented. If multiple knowledge points exist, they are ranked by confidence score. To improve the matching accuracy between question stems and knowledge points, especially in scenarios with diverse expressions, interdisciplinary content, or complex semantic structures, the following combination of strategies is adopted:

[0065] ① Semantic encoding enhancement (Transformer nested dual-tower model): Build a dual-tower structure model based on Transformer (such as BERT, MacBERT, etc.): one tower input is the normalized question stem text; the other tower input is the standard knowledge point definition or knowledge point explanation description; the system encodes the question stem and knowledge point into semantic vectors respectively, performs cosine similarity calculation in the semantic space, and obtains the set of knowledge points with the best semantic matching.

[0066] ② Graph prior guidance: Combined with the structural information such as hierarchical relationships and dependency relationships provided by the knowledge graph, the semantic matching results are corrected by topological rules: if a high-matching result appears between adjacent nodes or parent-child nodes, its confidence weight is increased; if there is a "non-co-occurrence" structural restriction between the matching knowledge points (such as belonging to different grade modules), its score is lowered.

[0067] ③ Multi-channel information fusion: Based on semantic vector matching, the following auxiliary features are integrated:

[0068] Keyword hit rate: frequency / position statistics of knowledge point vocabulary in the question stem;

[0069] Historical co-occurrence frequency: the frequency with which this knowledge point appears with the current student or question type;

[0070] Question type preference weight: the prior probability that certain knowledge points tend to appear in a certain type of question.

[0071] ④Matching explanation generation module (optional):

[0072] The system can automatically output "matching explanation text", such as "This question contains function relationship graph transformation, which meets the characteristics of the knowledge point of 'function graph translation'", which is used for debugging and teacher confirmation to enhance the system's interpretability.

[0073] ⑤Confidence sorting mechanism:

[0074] The semantic similarity, graph structure distance, keyword hits and historical statistical results are normalized and weighted to generate the final knowledge point matching list, which is sorted in descending order of confidence and is the weight coefficient obtained through training or parameter adjustment.

[0075] 4) Difficulty level labeling: The system assigns a difficulty level (easy, medium, or difficult) based on parameters such as question type, keyword complexity, grade / exam type, and error rate of history questions using rules or machine learning models.

[0076] 5) Feature vector construction: Use the TF-IDF method to generate text feature vectors for the normalized question stem, while retaining word frequency statistics and position information, and send them as input to the subsequent diagnostic model.

[0077] (3) Error diagnosis: See Figure 4 The error text feature vector and student answer behavior data are input into the error classification model based on convolutional neural network (CNN) to output the error type label. The answer behavior data includes the operation time period, the duration of each operation, and the number of review times. The error classification model uses the following algorithm: Error_Type = α × Semantic_Similarity + β ×

[0078] peration_Sequence_Deviation+γ×Difficulty_Level+δ×

[0079] Error_Pattern+ε×Micro_Behavior_Index; Error_Type: error type label output, indicating the classification result after the model attributes the wrong question, such as conceptual error, question review error, habitual error, etc.; Semantic_Similarity: the semantic similarity between the wrong question text and the corresponding knowledge point (such as calculated based on TF-IDF or BERT), reflecting the degree of matching at the language level;

[0080] Problem_Sequence_Deviation: A measure of the step deviation between the student's answer path and the standard answer path, reflecting whether the problem-solving logic is correct; Difficulty_Level: The standard difficulty level label of the current question, usually derived from the question bank or manual annotation (such as easy, medium, and difficult); Error_Pattern: The typical error behavior type matched, such as symbol confusion, missed step calculation, format inversion, etc.; Micro_Behavior_Index: A quantitative indicator generated by the micro-behavior recognition module, which comprehensively considers detailed behavioral characteristics such as hesitation time, number of revisions, and handwriting abnormalities; α, β, γ, δ, ε: The weight coefficients corresponding to each influencing factor, learned from the training samples, indicating the importance of each factor in the comprehensive score.

[0081] A micro-behavior recognition module is introduced in this process: when students are answering, micro-behavior information such as writing trajectory, correction behavior, pause time, etc. is recorded synchronously, and micro-behavior features such as "hesitation index" and "correction density" are extracted to assist in determining the causes of errors.

[0082] The specific implementation is as follows:

[0083] 1) Model input construction: The TF-IDF text feature vector constructed in (2) and the behavioral data (operation duration, time period, review frequency, etc.) are jointly embedded into the model input layer;

[0084] 2) Micro-behavior recognition module: During the answering process, the system uses the following technical means to identify students' micro-behavior characteristics in real time:

[0085] Writing trajectory acquisition: The touch device records the spatiotemporal path of the handwriting, including the start and end of the stroke, the speed of movement, and the angle of the pen tip, to generate a continuous trajectory vector.

[0086] Revision behavior detection: Through handwriting overlap analysis, erase instruction recording, handwriting backtracking trajectory recognition, etc., the frequency of modifications per unit time is counted to form the "revision density" indicator;

[0087] Pause recognition and rhythm analysis: Count the time intervals between each step, character, or line of writing, identify long pauses, and calculate them into the "hesitation index";

[0088] Trajectory stability and nonlinearity assessment: Deducing the coherence and disorder of problem-solving thinking based on trajectory fluctuation frequency, handwriting jump distance, angular velocity changes, etc.

[0089] Feature extraction method: Use principal component analysis (PCA), sliding window clustering, or time series-based LSTM network to extract and encode micro-behavior features as the input basis of Micro_Behavior_Index.

[0090] 3) CNN diagnostic model structure: A multi-channel one-dimensional convolutional network is used. The first channel processes text features, the second channel processes behavioral data vectors, and the third channel processes micro-behavioral features. The outputs of each channel are concatenated and passed through a fully connected layer, outputting a multi-classification softmax result to form the Error_Type label set.

[0091] 4) Model training: The system uses supervised learning to train the parameter weights (α to ε) based on a dataset of annotated historical incorrect questions. The system uses a cross-entropy loss function for optimization to improve error classification accuracy.

[0092] Furthermore, the system incorporates a "cognitive intention inversion mechanism": by analyzing students' answering steps, revision behavior, and problem-solving paths, it determines whether the correct cognitive structure underlying their errors exists. For example, if the initial solution steps align with the standard path but deviate midway, this can be identified as "correct understanding but operational deviation." Conversely, if the first step is incorrect, this is considered a "cognitive error." This mechanism, acting as an intermediate layer of information, assists in optimizing error type classification and knowledge point identification logic, enhancing the model's interpretability.

[0093] The specific implementation of the cognitive intention inversion mechanism is as follows:

[0094] Answer path reconstruction: The system restores the student's problem-solving path sequence based on the real-time input data (handwriting trajectory or click steps) in the digital answer interface, including: the starting point of the problem-solving; the logical order of the intermediate reasoning steps; the formulas or sub-conclusions written in each step; and the modification points and interruption points that occurred during the problem-solving process.

[0095] Standard Path Comparison Library: The system maintains a "standard solution path model" corresponding to each problem, which is preset by teaching experts or automatically generated. It includes a standard step sequence and reasoning path diagram. A graph structure (Directed Acyclic Graph) can be used to represent the standard logical flow.

[0096] Path similarity calculation: The system uses the following two methods to perform path matching and difference evaluation:

[0097] Edit distance algorithm: measures the difference in operations (addition, deletion, replacement, wrong order) between the student path and the standard path;

[0098] Path alignment scoring: The time-weighted local dynamic time warping (DTW) method is introduced to calculate the structural alignment and obtain indicators such as "initial consistency", "mid-segment deviation", and "overall similarity".

[0099] Behavior classification rule definition: If the alignment degree of the starting step is high, and the subsequent gradually deviates, the system generates the label: "cognitive correctness but execution deviation"; if the starting deviates or is completely inconsistent with the standard path, the system generates the label: "cognitive root error"; if the path is similar as a whole but there is a structural interruption (such as skipping steps, mispositioned insertion), the system generates the label: "break in thought coherence"; if the path is consistent as a whole, and the final result is wrong, combined with micro-behavior parameters (such as high revision density), it is labeled as "careless error type".

[0100] Information cross-validation: combined with micro-behavior indicators (such as pause time distribution, revision point density) and behavior inversion results, the system forms the final "cognitive intention judgment label", and inputs it as an intermediate variable into the error classification model and the ability model, optimizing the accuracy of error attribution and the interpretability of the model.

[0101] (4) Ability evaluation: see Figure 5 , according to the student's grade, unit test score, error question attribution and review record, the scores in multiple ability dimensions are calculated and dynamically updated, and the ability evaluation includes six dimensions: knowledge understanding, normative expression ability, logical reasoning, transfer application, test strategy and data analysis ability.

[0102] The meanings of each dimension are as follows:

[0103] Knowledge understanding: reflects the students' mastery of basic concepts, formulas, principles in the teaching material, and is the basis for the formation of subsequent abilities;

[0104] Normative expression: measures the students' ability to use subject terminology, symbols and structured answer expression, covering answer format, unit writing, essay structure, etc.;

[0105] Logical reasoning: evaluates the students' ability to gradually deduce conclusions based on conditions, suitable for proof, deduction, analytical problem solving, etc.;

[0106] Transfer application: embodies the students' ability to apply learned knowledge to variant situations, interdisciplinary scenarios or real-life problems;

[0107] Test strategy: identifies students' strategic behavior in the exam environment, including question selection order, time allocation, question reading skills, etc.;

[0108] Data analysis: focuses on students' understanding and use of charts, data, statistical results, widely used in science and information courses.

[0109] The ability score is calculated as follows:

[0110] Unit ability score = basic score + test score + ∑ (review bonus) - ∑ (error question deduction),

[0111] Total score = sum of unit ability scores.

[0112] Wherein, the base score: the preset ability score, set according to the student's grade, course stage, as a reference starting point; Test score: the student's performance contribution part in the current unit test; Review bonus: the bonus reward obtained by the student after completing the corresponding ability dimension related review task, encouraging effective learning behavior; Mistake deduction: the deduction corresponding to the student's error in this ability dimension, the deduction value is adjusted according to the difficulty and type of the mistake; Each unit ability score: the score of different units in a certain ability dimension; Total score: cumulative ability score of multiple units, used to comprehensively judge the overall performance of the student in this ability dimension.

[0113] The specific implementation is as follows:

[0114] 1) Ability dimension mapping: each mistake is mapped to one or more ability dimensions in the annotation process, and the ability dimension attribution is determined according to the knowledge point associated with the mistake and its cognitive requirements;

[0115] 2) Base score setting: the system initializes the base score of each ability dimension according to the grade and course stage, which is used to represent the average level of students in this grade;

[0116] 3) Test score calculation: after each unit test, the system maps the test scores to the corresponding ability dimensions according to the knowledge point attribution, and uses proportional allocation method to weight and allocate the scores;

[0117] 4) Review bonus mechanism: the student completes the review task (such as re-exercise, variant exercise, knowledge point explanation reading, etc.) for the mistake of a certain ability dimension, and the system gives corresponding bonus according to the task completion degree and the difficulty of the question;

[0118] 5) Mistake deduction mechanism: the system sets the mistake deduction weight according to the student's input mistake, combined with the ability dimension and difficulty level of the question, and executes dynamic deduction in the corresponding ability dimension;

[0119] 6) Dynamic adjustment and trend modeling: the system supports updating the ability score over time, and builds a trend chart to show the change trajectory of the student in each ability dimension, supporting teacher or parent intervention.

[0120] The above model can be linked with the cognitive intention inversion result, for example: if a mistake is determined as "cognitive correct but operation error", its deduction will be partially offset in the ability assessment, to reflect the true level of the student's cognitive structure, and enhance the fairness and explanatory power of the assessment.

[0121] Wherein, the general weight association matrix structure of the six-dimensional ability model and the knowledge point is:

[0122]

[0123] Matrix generation methods include one or more of the following:

[0124] Expert scoring method: manually set by subject experts based on the syllabus and cognitive requirements;

[0125] Data-driven approach: Based on large-scale wrong question data and ability assessment history, training is carried out through techniques such as factor analysis and LDA;

[0126] Hybrid optimization method: Initial scoring is done by experts, and dynamic fine-tuning is performed later based on student performance feedback.

[0127] (5) Knowledge point correlation analysis: see Figure 6 This system is based on the semantic and structural dual modeling method driven by wrong questions, integrating the improved semantic embedding algorithm, knowledge graph modeling and robustness verification mechanism to achieve accurate attribution, expanded deduction and high-confidence recommendation of the knowledge points involved in the wrong questions.

[0128] The specific implementation includes the following steps:

[0129] 1) Semantic Embedding Modeling: Use large-scale subject corpora to train an improved Word2Vec or BERT model, perform high-dimensional vector representation of all knowledge points, construct a knowledge point semantic space, and support cosine similarity calculation;

[0130] 2) Structural Knowledge Graph Modeling: Construct a knowledge graph covering the curriculum system, where nodes represent knowledge points and edges represent logical relationships between knowledge points (including pre-dependencies, inclusions, derivations, or equivalence relationships), serving as a structural semantic reference.

[0131] 3) Perturbation robustness identification mechanism: Multiple semantic perturbation samples are constructed for each incorrect question stem, including:

[0132] Insertion class (insert interference words such as "sometimes", "a certain type", "among them");

[0133] Sequence perturbation (disrupting the word order while retaining the main stem);

[0134] Masking class (masking keywords and constructing MASK samples);

[0135] Calculate the cosine similarity between the original question stem and its perturbation version and all knowledge point vectors. If a knowledge point still maintains a high match in the perturbation sample, it is identified as a robust target knowledge point and its confidence weight is increased.

[0136] 4) Wrong Question-Knowledge Point Confidence Modeling: Construct a mapping set between wrong questions and high-confidence knowledge points, expressed as:

[0137] K match (Q i )={(K j ,Pj )}; where Q i is the i-th wrong question, K j To match knowledge points, P j Assign confidence to the match.

[0138] 5) Extended correlation analysis mechanism driven by wrong questions: for each core knowledge point K j , retrieve its upstream knowledge points in the knowledge graph (pre-concept), downstream knowledge points (Migration application point) and equivalent node K eq ; Assign weights to the corresponding structure types (w_up, w_down, w_eq); True or False Questions Q i With each potential extended knowledge point K m Calculate its relative expansion score:

[0139] Several extended knowledge points with high Related_Score scores are included in the review recommendations to form a three-order semantic-structural mapping chain of "wrong questions - core knowledge points - highly related knowledge points".

[0140] 6) Knowledge point association score calculation: Define the comprehensive association between two knowledge points A and B: Association_Score(A, B) = Cosine_Similarity(Vector(A), Vector(B)) × Co_occurrence_Frequency(A, B) + λ × Knowledge_Graph_Link(A, B);

[0141] in:

[0142] Vector(A), Vector(B): semantic vector representation of knowledge points A and B;

[0143] Cosine_Similarity(...): semantic similarity between vectors;

[0144] Co_occurrence_Frequency(A, B): The frequency with which knowledge points A and B co-occur in wrong history questions or textbook content;

[0145] Knowledge_Graph_Link(A, B): whether there is a direct structural connection between A and B in the knowledge graph (1 means yes, 0 means no);

[0146] λ: The weight adjustment factor of the structural relationship in the overall score.

[0147] Through this analysis mechanism, the system not only identifies the core knowledge points responsible for incorrect questions, but also explores potential upstream forgetting points, downstream migration points, and equivalent concepts, enabling dynamic extension and strategic optimization of review paths. This mechanism further enhances the system's ability to understand and intervene in students' overall knowledge mastery, improving personalized teaching effectiveness.

[0148] (6) Dynamic review planning: see Figure 7 ,Combined with students' personalized forgetting curve parameters, wrong question attributes and historical review records, an improved forgetting curve algorithm or Q-learning algorithm is used to generate personalized review plans and recommended time.

[0149] The specific implementation is as follows:

[0150] 1) State modeling: Define the current student learning state s, including dimensions such as knowledge point mastery (based on ability model scores), wrong question labels, historical review intervals, cognitive load state (described later), and review response performance;

[0151] 2) Construction of optional action set a: including options such as "review immediately", "postpone 1 day", "postpone 3 days", and "add variation practice";

[0152] 3) Definition of instant reward R: The system calculates instant rewards based on feedback information such as the student's test accuracy rate, answering speed, and whether they actively initiate review after each review;

[0153] 4) Strategy Learning: Use the following Q-learning algorithm to update and review the decision-making strategy:

[0154] Q(s, a)=(1-α)×Q(s, a)+α×[R(s, a)+γ×maxQ(s′, a′)]

[0155] Where: Q(s, a): the long-term expected return value of performing action a in the current state s; α: learning rate, which controls the update intensity of new experience to the original experience; γ: discount factor, which controls the weight of future returns in the current state; s′: the new state after performing action a; a′: the action with the largest Q value among all optional actions in the new state.

[0156] 5) Cognitive Load Prediction Mechanism: The system predicts a student's current cognitive load index based on the following data: cumulative review time over the past three days; average pause duration per answer; micro-behavioral fatigue indicators (such as the number of consecutive revisions and the rate of operational conflicts); and the number of frequently incorrectly assigned knowledge points. When cognitive load exceeds a threshold, the system automatically adjusts its strategy: reducing the recommendation intensity for low-priority incorrectly assigned questions or shifting review tasks to later periods of high cognitive load.

[0157] 6) Ebbinghaus memory curve fusion: The system records the students' first learning and previous review times, estimates the forgetting risk based on the forgetting model, and integrates it with the Q value results to finally determine the review time point.

[0158] 7) Habitual Error Reinforcement Intervention Mechanism: The system continuously clusters and analyzes students' wrong answer records, identifies recurring structural error types (such as unit confusion and skipped steps), marks them as "habitual errors", and automatically increases the weight of such questions in the review plan, adding multiple repetitions, variant question tests and error explanation prompts until the error is eliminated or the frequency decreases.

[0159] 8) Plan output and execution feedback: The system generates a list of executable review tasks every day. Users can choose to review online, print exercises or test mode. The system tracks the execution status and results and continues to update the Q value.

[0160] In summary, this module realizes closed-loop control of the entire process from "individual forgetting modeling" to "review behavior feedback learning" to "dynamic time regulation and habitual error intervention", significantly improving the personalized matching and effect response of the review strategy.

[0161] The experimental comparison data are summarized as follows:

[0162] Experiment 1: Comparison of the accuracy of wrong question diagnosis

[0163] Sample: 90 students in the second grade of a middle school were selected (average mathematics score ±1σ interval);

[0164] Grouping: Control group (n=45): using traditional electronic error book (supporting only OCR recognition + manual annotation); Experimental group (n=45): using this system (multimodal diagnosis + ability model association);

[0165] Test data: A test paper containing 10 typical, easy-to-make mistakes questions (covering arithmetic, geometry, statistics, etc.) is uniformly distributed. Students are required to complete the test on their device and upload the wrong answers.

[0166] index Control group (traditional method) Experimental group (this system) Improvement Error type recognition accuracy 72.4％±8.1％ 91.7％±5.3％ +26.7％* Knowledge point association accuracy 68.9％±7.6％ 89.2％±4.8％ +29.5％* Review suggestions are efficient 61.3％±9.2％ 83.5％±6.4％ +36.2％*

[0167] Note: indicates statistically significant difference at p < 0.01.

[0168] Experiment 2: Review Efficiency Comparison (Knowledge Retention Rate)

[0169] Sample: 80 first-year high school students from a certain high school were selected (with a 20% fluctuation around the median math score);

[0170] Grouping: Control group (n=40): using fixed cycle review method (once a day × 7 days); Experimental group (n=40): using dynamic programming review of this system (based on Ebbinghaus curve);

[0171] Testing process: Pre-test: a unified unit test (maximum 100 points); Intervention: the two groups reviewed the wrong questions according to the plan; Post-test: a test on the same content was conducted 30 days later;

[0172] index Control group (traditional method) Experimental group (this system) Improvement Short-term memory retention rate 42.7％±10.3％ 68.4％±8.9％ +60.2％* Long-term memory retention rate 28.9％±7.5％ 54.3％±9.1％ +87.9％* Review time efficiency ratio 1.0x 1.8x +80％*

[0173] Note: indicates extremely significant difference with p < 0.001.

[0174] Experiment 3: Learning Motivation and Continuous Use Effect

[0175] Sample: 120 fifth-grade students from a primary school (with similar scores on the Mathematics Interest Scale) were selected;

[0176] Grouping: Control group (n=60): Use the wrong question management system without incentive mechanism; Experimental group (n=60): Use this system (including ability model + point reward);

[0177] Testing period: 8 consecutive weeks of tracking;

[0178] index Control group (traditional method) Experimental group (this system) Improvement Average number of wrong questions submitted per week 8.2±3.1 questions 14.7±4.5 questions +79.3％* Continuous use weeks 3.8±1.6 weeks 7.4±1.2 weeks +94.7％* Capability model dimension coverage 3.2±1.4 dimensions 5.8±0.9 dimensions +81.3％*

[0179] Note: indicates extremely significant difference with p < 0.001.

[0180] In summary, by integrating micro-behavior recognition, cognitive intent inversion, anti-disturbance knowledge point attribution, cognitive load regulation, and habit bias mapping mechanisms, the present invention has achieved several unexpected beneficial technical effects compared to the existing technologies:

[0181] First, in terms of the depth of attribution of wrong questions, the present invention introduces the micro-behavior recognition and cognitive intention inversion mechanism for the first time, which can further distinguish seemingly identical errors into "cognitive errors" and "execution deviations", greatly improving the system's ability to understand students' true cognitive structure. It has high explanatory and personalized guidance significance, and is a breakthrough in the traditional static error classification model.

[0182] Secondly, in terms of the robustness of knowledge point matching, a disturbance-resistant knowledge point identification method is proposed. Even in the face of changes in question stem wording and complex language structure, the core knowledge points can still be locked through disturbance stability assessment, achieving accurate identification in semantically ambiguous scenarios, significantly improving the fault tolerance of traditional keyword matching models.

[0183] Thirdly, at the level of dynamic review recommendation, the present invention integrates Q-learning intelligent strategy with cognitive load prediction to dynamically adjust the review content and frequency. It not only arranges the best review time according to the student's forgetting status, but also identifies the current learning pressure status, thereby avoiding cognitive overload and ensuring review efficiency.

[0184] In addition, the present invention constructs a six-dimensional ability model assessment and introduces cognitive bias labels into the ability adjustment mechanism, making the ability evaluation more fair, dynamic and growth-oriented, which is conducive to the formation of long-term learning portraits.

[0185] Finally, the system as a whole forms a closed-loop structure of the entire process of collection-analysis-attribution-diagnosis-evaluation-intervention-feedback, breaking through the static linear mode of "storing errors-checking errors-passive review" of the traditional system, and realizing a truly intelligent and adaptive learning intervention platform with the comprehensive value of high education promotion and personalized tutoring.

[0186] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for diagnosing and classifying wrong questions and intelligent review planning based on multi-dimensional analysis, characterized in that: The following steps are involved: (1) Wrong question collection: Use smart terminals or mobile devices to capture wrong question images and use OCR recognition technology to extract wrong question text and accompanying image data; (2) Wrong question annotation: Automatically annotate the extracted text and image data to generate subject information, knowledge points, and difficulty level attributes; (3) Error diagnosis: The feature vector of the wrong question text and the student's answering behavior data are input into the error classification model based on the convolutional neural network, and the error type label is output; (4) Ability assessment: Calculate scores under multiple ability dimensions based on student grade, unit test scores, wrong question attribution, and review records, and update them dynamically; (5) Knowledge point relevance analysis: Using the improved Word2Vec model and knowledge graph, we construct knowledge point vectors and calculate their cosine similarity and co-occurrence frequency to evaluate the relevance between wrong questions and related knowledge points, as well as between knowledge points themselves. (6) Dynamic review planning: Based on the student's personalized forgetting curve parameters, wrong question attributes and historical review records, an improved forgetting curve algorithm or Q-learning algorithm is used to generate a personalized review plan and recommended time.

2. The method according to claim 1, wherein the recognition accuracy of the OCR recognition technology is not less than 98%, and the recognition process records metadata to assist in the time sequence management of wrong question sorting.

3. The method according to claim 1, wherein the wrong question text feature vector is generated by a TF-IDF weighted algorithm, and the answering behavior data includes an operation time period, a duration of each operation, and a review number indicator.

4. The method according to claim 1, wherein the error classification model adopts the following algorithm: Error_Type = α×Semantic_Similarity+β×Operation_Sequence_Deviation+γ×Difficulty_Level+δ×Error_Pattern, wherein α, β, γ, and δ are weight factors determined through training.

5. The method according to claim 1, wherein the ability assessment includes six dimensions: knowledge comprehension, standard expression ability, logical reasoning, transfer application, test-taking strategy and data analysis ability.

6. The method of claim 1 , wherein the capability score calculation comprises: Unit ability score = basic score + test score + ∑ (review bonus) - ∑ (wrong question deduction), Total score = the sum of the ability scores of each unit.

7. The method according to claim 1, wherein the calculation formula of the knowledge point relevance is: Association_Score(A, B)=Cosine_Similarity(Vector(A), Vector(B))×Co_occurrence_Frequency(A, B)+λ×Knowledge_Graph_Link(A, B).

8. The method according to claim 1, wherein the dynamic review planning model is implemented using the following Q-learning algorithm: Q(s, a)=(1-α)×Q(s, a)+α×[R(s, a)+γ×maxQ(s′, a′)], Where s is the current state, a is the review action, R is the immediate reward, and γ is the discount factor.

9. The method according to claim 1, wherein the review suggestion includes a list of wrong questions and a suggested review time node, and the suggestion is personalized based on the Ebbinghaus memory curve.

10. The method according to claim 1, wherein students can view the review plan or print the review checklist through the terminal device, supporting a hybrid review mode of online and offline.

Images