A postgraduate examination knowledge point intensive training method based on large model wrong question accurate positioning

Abstract

The application discloses a postgraduate entrance examination knowledge point reinforcing training method based on a large model mistake question accurate positioning, and comprises the following steps: mistake question data acquisition and structured preprocessing; large driving mistake question knowledge point accurate positioning; multi-dimensional error attribution analysis; knowledge graph correlation expansion; personalized reinforcing training scheme generation; reinforcing training execution and effect iteration.The application has the beneficial effects that: through semantic level analysis of the postgraduate entrance examination field large model after fine tuning, the limitation of traditional keyword matching is broken through, and the corresponding subdivided knowledge points and cognitive levels of the deformation questions and comprehensive questions can be accurately identified; a multi-dimensional weight calculation model is adopted to distinguish the influence degree of different error types, realize the upgrading from knowing that mistakes are made to knowing why mistakes are made and where mistakes are made, and the training is more targeted; in combination with a difficulty self-adaptive algorithm and a mastery evaluation model, the question difficulty is dynamically adjusted according to the real-time learning state of a user, and in cooperation with the knowledge graph correlation expansion, the weak knowledge points are systematically reinforced.

Description

Technical Field

[0001] This invention relates to the fields of intelligent education and large language model technology, and in particular to a method for strengthening training of postgraduate entrance examination knowledge points based on accurate location of incorrect questions using a large model. Background Technology

[0002] With the increasingly fierce competition for postgraduate entrance examinations, efficient preparation has become a core need for candidates. Traditional postgraduate entrance examination preparation mainly relies on the "sea of ​​questions" approach, where candidates accumulate problem-solving experience through a large number of practice questions. However, this approach generally suffers from the following drawbacks: 1. Error analysis relies on manual intervention, resulting in a rough identification of knowledge points. It often only identifies the chapter level and cannot accurately pinpoint specific knowledge points and cognitive levels, leading to incomplete error correction. 2. The intensive training is highly homogenized, and the question bank pushes lack specificity. Students repeatedly practice the content they have mastered, but their weak knowledge points are not effectively reinforced. 3. The error attribution is too simplistic, failing to distinguish between different error types such as conceptual ambiguity, calculation errors, and misreading of the question, thus the training program cannot address the root causes of the errors. 4. Lack of a scientific forgetting and difficulty adjustment mechanism, unreasonable review rhythm, and low learning efficiency.

[0003] In recent years, intelligent question banks and online education systems have gradually incorporated knowledge point tagging systems. However, most existing technologies rely on keyword matching for knowledge point location, exhibiting weak semantic understanding capabilities and low accuracy in locating modified question stems and comprehensive question types. Furthermore, training schemes often rely on fixed rules, lacking adaptability and hindering truly personalized reinforcement. With the development of large language model technology, its powerful semantic understanding and knowledge reasoning capabilities offer new solutions for intelligent error analysis and personalized training. However, mature solutions for deeply applying large models to accurately locate errors and reinforce knowledge points in postgraduate entrance examinations remain relatively scarce.

[0004] Therefore, it is necessary to propose a method for strengthening postgraduate entrance examination knowledge points based on accurate location of incorrect questions using a large model to address the aforementioned technical issues. Summary of the Invention

[0005] This invention provides a method for intensive training of postgraduate entrance examination knowledge points based on accurate location of incorrect questions using a large model. It overcomes the shortcomings of existing technologies, such as inaccurate location of incorrect knowledge points, weak targeted training, and low learning efficiency. By using semantic parsing of a large model to achieve accurate mapping of incorrect questions to knowledge points, and combining error attribution quantification, knowledge graph expansion, and difficulty adaptive algorithms, a closed-loop personalized intensive training system is constructed, which significantly improves the efficiency of postgraduate entrance examination preparation.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A method for intensive training of postgraduate entrance examination knowledge points based on accurate location of incorrect answers using a large model includes the following steps: S1. Error Data Collection and Structured Preprocessing: Collect error data from users' postgraduate entrance examination practice, extract text information such as question stems, incorrect answers, and problem-solving processes, and perform noise reduction, word segmentation, and structured annotation. S2, Big Data-Driven Precise Positioning of Wrong Question Knowledge Points: Input structured wrong question data into a finely tuned big data model for the postgraduate entrance examination, and output the core knowledge points, subordinate knowledge points and cognitive level labels corresponding to the wrong questions through semantic understanding and knowledge point mapping; S3. Multi-dimensional error attribution analysis: Based on the characteristics of incorrect questions and the attributes of knowledge points, the causes of errors are classified and identified, the weight of each type of error factor is calculated, and an error attribution report is generated. S4. Knowledge Graph Association Expansion: Based on the core knowledge points located, association retrieval is performed in the pre-constructed postgraduate entrance examination subject knowledge graph to expand the knowledge points to include prerequisite knowledge points, subsequent knowledge points, and common error-prone knowledge points. S5. Personalized reinforcement training program generation: Combining error attribution results with the relationship between knowledge points, adaptively generate reinforcement training paths that include question difficulty gradients, training frequency, and training order. S6. Enhance training execution and effect iteration: Push corresponding training content, collect user answer data, evaluate the mastery of knowledge points in real time, and dynamically adjust subsequent training plans based on evaluation results to form a closed-loop iteration.

[0007] The structured preprocessing in step S1 specifically includes: The text of incorrect questions is processed by a combination of OCR recognition and manual verification to extract fields such as the main body of the question, the question format, the incorrect options, the user's answer traces, and the error location markers. A word segmentation dictionary specifically for postgraduate entrance examination subjects is used for word segmentation and part-of-speech tagging to construct a standardized feature vector of incorrect questions.

[0008] Step S2, which involves precise localization of knowledge points in the large model, includes the calculation of knowledge point matching degree. The specific algorithm formula is as follows: Let the feature vector of the incorrect question text be... The feature vector of the i-th knowledge point in the knowledge point base is Then, the semantic matching degree Sim(T,Kᵢ) between the incorrect question and the i-th knowledge point is: ; Where w ⱼ The domain weight of the j-th feature word is determined by weighting the word frequency in the postgraduate entrance examination syllabus with the frequency of its appearance in past exam questions; Select knowledge points with a matching degree higher than a preset threshold as a candidate set, sort them in descending order of matching degree, and output the top-m core knowledge points and subordinate knowledge points.

[0009] Step S3, the multi-dimensional error attribution analysis, includes the error attribution weight calculation step. The specific algorithm formula is as follows: Let W be the attribution weight of the p-th error factor. p The calculation formula is: ; Where f p This is the normalized value of the frequency of this type of error in the user's historical incorrect questions; d i c represents the difficulty coefficient of the corresponding knowledge point in the postgraduate entrance examination syllabus, with a value range of 0-1; p α represents the impact coefficient of this type of error on the mastery of knowledge points; α, β, and γ are weighting coefficients, satisfying α+β+γ=1; Sort by weight value from high to low to identify dominant and secondary error types.

[0010] The construction method of the postgraduate entrance examination subject knowledge graph in step S4 is as follows: Using the official postgraduate entrance examination syllabus as the framework, a node system is constructed according to a three-level hierarchy of subject, chapter, and knowledge point. Node attributes include knowledge point name, difficulty level, frequency of examination, cognitive level, and percentage of past exam questions. Edge relationships include pre-dependency, subsequent extension, easily confused, and homologous parallel types. Through a large model, past exam questions and teaching materials are extracted to supplement the edge relationships and attribute values ​​of the knowledge graph.

[0011] The knowledge point association and expansion in step S4 are specifically as follows: Using core knowledge points as central nodes, a second-order traversal is performed in the knowledge graph to retrieve all prerequisite dependent nodes and easily confused nodes; based on the strength of the association between nodes and the user's historical knowledge, a set of related knowledge points that need to be supplemented and strengthened is selected.

[0012] Step S5, which generates a personalized reinforcement training scheme, includes a difficulty adaptive adjustment step. The specific algorithm formula is as follows: Let the target difficulty value of the next round of training questions for the k-th knowledge point be _____. The calculation formula is: ; Where D k The current training difficulty level is represented by a value between 0 and 1; P0 is the preset target accuracy threshold; P k This represents the user's historical accuracy rate in answering questions on this knowledge point; η is the learning rate adjustment factor, with a value ranging from 0.05 to 0.2. After calculating the target difficulty, questions with difficulty values ​​within ±0.1 are matched from the question bank to form a graded training question set.

[0013] The rule for generating reinforcement training paths in step S5 is as follows: The training units are arranged in the following order: supplementing prerequisite knowledge points, explaining core knowledge points in detail, identifying common mistakes, consolidating related questions, and improving comprehensive questions. The number of questions in each training unit is dynamically allocated according to the error attribution weight, with higher weight error types corresponding to a larger number of training questions.

[0014] Step S6, which involves real-time assessment of knowledge point mastery, includes a step of quantifying mastery. The specific algorithm formula is as follows: Let M be the degree of comprehensive mastery of a certain knowledge point, and the calculation formula is: ; Where A is the average accuracy rate of the last three answers to this knowledge point; N is the cumulative number of training sessions. The time interval since the last training session; τ is the training saturation coefficient; T is the forgetting time constant; For the weighting coefficients, satisfying ; When the mastery level M is higher than the preset threshold, the training frequency for that knowledge point is reduced; if it is lower than the threshold, the training frequency is increased and the difficulty of the questions is reduced.

[0015] The fine-tuning method for the large model in the postgraduate entrance examination field in step S2 is as follows: Based on a general large language model, the model uses past postgraduate entrance examination questions, syllabus analysis, authoritative teaching materials, and standard error sets as fine-tuning corpora. It constructs an error-knowledge point mapping instruction set using an instruction fine-tuning method, and trains the model to output standardized knowledge point labels and cognitive levels.

[0016] In step S6, the reinforcement training content is generated in a multimodal manner: The large model generates textual explanations, reasoning derivations, and error-prone point hints based on knowledge points and error types; it automatically matches corresponding formula derivation diagrams, logic framework diagrams, and problem-solving step animations based on knowledge point types; and it generates audio explanations for frequently missed knowledge points.

[0017] A method for reinforcing postgraduate entrance examination knowledge points based on precise location of incorrect answers using a large model, which also includes a cross-disciplinary knowledge point linkage step: When incorrect answers involve interdisciplinary knowledge points, the large-scale model identifies interdisciplinary related nodes, retrieves related knowledge points in the corresponding subject's knowledge graph, and generates an interdisciplinary joint training plan; it is suitable for training comprehensive question types that combine political science with current affairs and mathematics with professional courses in postgraduate entrance examinations. Compared with existing technologies, the advantages of this invention are: 1. By performing semantic-level analysis through a finely tuned large-scale model of the postgraduate entrance examination, the limitations of traditional keyword matching are broken through. It can accurately identify the subdivided knowledge points and cognitive levels corresponding to variant questions and comprehensive questions, and the positioning accuracy is significantly improved.

[0018] 2. Employing a multi-dimensional weight calculation model, it distinguishes the degree of impact of different error types, achieving an upgrade from knowing that a mistake is made to knowing why it is wrong and at what level the mistake occurs, resulting in more targeted training.

[0019] 3. By combining an adaptive difficulty algorithm with a mastery assessment model, the difficulty, number of questions, and training frequency are dynamically adjusted according to the user's real-time learning status. Combined with knowledge graph association expansion, this achieves systematic reinforcement of weak knowledge points.

[0020] 4. Construct a complete closed loop of error collection, precise location, attribution analysis, intensive training, effect evaluation, and solution iteration to form a continuously optimized personalized learning path and significantly improve the efficiency of postgraduate entrance examination preparation. Attached Figure Description

[0021] Figure 1 This is a flowchart of the postgraduate entrance examination knowledge point reinforcement training method of the present invention; Figure 2 This is a bar chart showing the comparison of training parameters in this invention. Figure 3 This is a bar chart comparing the number of knowledge graph associations and extensions in this invention. Detailed Implementation

[0022] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be further described below in conjunction with specific embodiments and accompanying drawings. However, the embodiments described are not intended to limit the present invention.

[0023] like Figures 1 to 3 As shown, a method for intensive training of postgraduate entrance examination knowledge points based on accurate location of incorrect questions using a large model is implemented as follows: Step S1: Collection and Structured Preprocessing of Error Data The system supports multiple methods for collecting incorrect questions, including photo upload, text input, and simultaneous answering of questions from a question bank. For incorrect questions in image format, OCR technology is used for text recognition, combined with a manual verification mechanism to correct recognition errors. The core fields extracted include: the original question stem, the question type, the user's incorrect answer, the standard correct answer, the user's handwritten solution process, and the location of the error.

[0024] The question stem and solution process are segmented using specialized word segmentation dictionaries for each subject in the postgraduate entrance examination. Stop words and meaningless symbols are filtered out, and keywords, formula symbols, and professional terminology features are extracted. Finally, each incorrect question is transformed into standardized structured data and feature vectors, providing standardized input for subsequent large-scale model processing.

[0025] Step S2: Precise Location of Error-Based Knowledge Points Driven by a Large Model This step uses a large language model fine-tuned for the postgraduate entrance examination as the core parsing engine. In the fine-tuning stage, a general large model is used as the base, and nearly 20 years of postgraduate entrance examination questions, official examination syllabus, mainstream authoritative teaching materials, and standardized error question sets are used as training data to construct an instruction dataset of error question stems, knowledge point labels and cognitive levels. The knowledge point mapping ability of the model is trained by supervised fine-tuning.

[0026] The core calculation process for knowledge point location is as follows: Let the feature vector of the incorrect question text be... The feature vector of the i-th knowledge point in the knowledge point base is Then, the semantic matching degree Sim(T,Kᵢ) between the incorrect question and the i-th knowledge point is: ; Where wⱼ is the domain weight of the j-th feature word, which is determined by the importance of the word in the postgraduate entrance examination syllabus and the frequency of its appearance in past exam questions. Core concept words have a higher weight than ordinary descriptive words.

[0027] The system presets a matching threshold of 0.6. All knowledge points exceeding this threshold are selected as a candidate set. The top-3 results are then sorted in descending order of matching degree and labeled as core knowledge points, first-level subordinate knowledge points, and second-level subordinate knowledge points, respectively. The system also outputs the cognitive level corresponding to each knowledge point, including four levels: memorization, understanding, application, and analysis.

[0028] Step S3: Multidimensional Error Attribution Analysis The system has a pre-defined error type system, including six major categories: conceptual ambiguity, misreading the question, calculation errors, misuse of methods, knowledge forgetting, and logical reasoning errors. The large model initially determines the error type based on the location of the error in the incorrect question, the characteristics of the incorrect answer, and the problem-solving process.

[0029] Furthermore, the impact of various errors is quantified using the attribution weighting formula: ; Where: f p This is the normalized value of the frequency of this type of error in the user's incorrect questions over the past 30 days; the higher the frequency, the greater the weight. d i The difficulty coefficient of the postgraduate entrance examination syllabus for the corresponding knowledge point is 0-1. The higher the difficulty of the knowledge point, the greater the weight of the corresponding conceptual error. C p The impact coefficient of this type of error on the mastery of knowledge points; for example, the impact coefficient of the type of error, such as conceptual ambiguity, is higher than that of calculation error. α, β, and γ are weighting coefficients, with default values ​​of 0.4, 0.3, and 0.3 respectively, which can be adjusted according to the characteristics of the subject.

[0030] After calculating the weight value of each type of error, sort them from high to low. The one with the highest weight is determined as the dominant error type, and the rest are secondary error types, generating a structured error attribution report.

[0031] Step S4: Knowledge Graph Association Expansion A pre-constructed knowledge graph of postgraduate entrance examination subjects is used, with Politics, English, Mathematics, and mainstream professional courses as independent sub-graphs. Each sub-graph is constructed with nodes at four levels: subject, chapter, module, and subdivided knowledge points. Node attributes include: knowledge point name, syllabus requirements, difficulty level, frequency of examination in previous years, frequently tested question types, and cognitive level. The edge relationships between nodes include: prerequisite dependencies, subsequent extensions, easily confused relationships, homologous parallel relationships, and interdisciplinary connections.

[0032] The knowledge graph was initially built manually based on the official syllabus. Subsequently, a large model was used to extract relationships from past exam questions and teaching materials to continuously expand the edge relationships and attribute data.

[0033] Once the core knowledge points are located, a second-degree adjacency search is performed in the graph using these knowledge points as the central node: First-level search: directly related prerequisite knowledge points and easily confused knowledge points; Secondary retrieval: Pre-requisite nodes for prior knowledge points, forming a knowledge completion chain.

[0034] Based on the correlation strength value and the user's historical knowledge, select 3-5 related knowledge points that need to be strengthened simultaneously, and form a set of knowledge points to be strengthened.

[0035] Step S5: Generation of Personalized Reinforcement Training Program The intensive training program includes four dimensions: training content, difficulty gradient, question allocation, and training order.

[0036] Difficulty adaptive adjustment is the core algorithm for solution generation, and the calculation formula is as follows: ; in: D k The current training difficulty is set to the default difficulty of the knowledge points. P0 is the preset target accuracy threshold, with a default value of 0.8; P k The user's historical accuracy rate on this knowledge point in the last 10 attempts. η is the learning rate adjustment factor, with a default value of 0.1, which controls the adjustment range of difficulty.

[0037] Calculate the target difficulty Then, questions with difficulty values ​​within ±0.1 are retrieved from the question bank and arranged into a gradient training question set from low to high difficulty.

[0038] The distribution of questions follows the principle of proportionality between error weights: the dominant error type accounts for 50% of the total number of questions, the secondary error types each account for 20%, and comprehensive improvement questions account for 10%.

[0039] The training sequence is arranged according to cognitive patterns: preliminary knowledge point completion questions, core knowledge point basic questions, common mistake identification questions, related variation questions, and comprehensive application questions, realizing a reinforcement path from simple to complex and step by step.

[0040] Step S6: Reinforce training execution and effect iteration The system pushes training content question by question according to the generated training plan, supporting multimodal presentation formats such as text explanations, formula derivation diagrams, and voice explanations. It collects data such as the user's answer results, time taken, and modification history for each question.

[0041] After each training unit is completed, the mastery of knowledge points is assessed using a comprehensive mastery model: ; in A represents the average accuracy rate of this knowledge point in the last three training sessions; N represents the cumulative number of training iterations for this knowledge point; τ is the saturation coefficient, which defaults to 10. Δt is the number of days since the last training; T is the forgetting time constant, which is set to 7 days by default; λ1, λ2, and λ3 are taken as 0.6, 0.2, and 0.2, respectively.

[0042] When the mastery level M ≥ 0.85, the knowledge point is considered temporarily up to standard, the training frequency is reduced, and the study mode is switched to interval review; when M < 0.6, the knowledge point is considered not up to standard, the proportion of basic questions is increased and the difficulty is reduced, and a second round of reinforcement is carried out.

[0043] The system updates the user's knowledge mastery profile daily and dynamically adjusts the reinforcement training plan for the next day, forming a continuous closed loop of error identification, reinforcement training, effect evaluation, and plan optimization until the knowledge point is mastered stably.

[0044] Interdisciplinary extended implementation examples For interdisciplinary question types in postgraduate entrance examinations, such as comprehensive current affairs questions and questions combining mathematics and professional courses, the large-scale model identifies multidisciplinary knowledge points within the questions and retrieves related knowledge points in the corresponding subject sub-maps to generate interdisciplinary joint training programs. For example, for an analytical question involving Marxist philosophy and current affairs materials, the model simultaneously locates Marxist philosophy knowledge points and current affairs hot topics, pushes corresponding reinforcement content to each, and then supplements it with interdisciplinary comprehensive application questions to achieve synergistic reinforcement of interdisciplinary knowledge points.

[0045] Example 1: This embodiment addresses calculation-related errors in the advanced mathematics module of the postgraduate entrance examination, implementing the reinforcement training method of this invention. The specific process is as follows: S1. Error data collection and structured preprocessing: Supports uploading paper-based error questions by photo. It adopts a combination of mathematical formula OCR recognition technology and manual verification mechanism to extract fields such as the main body of the question, the question format, the user's incorrect calculation steps, the incorrect result, and the standard solution process. It uses a special word segmentation dictionary for postgraduate entrance examination mathematics (including professional terms such as limits, integrals, differential equations, and LaTeX formula features) for word segmentation and part-of-speech tagging, and constructs a standardized error feature vector containing formula features and question type features.

[0046] S2. The large-scale model-driven precise positioning of incorrect questions and knowledge points involves inputting structured incorrect questions into a domain-specific large-scale model fine-tuned from the postgraduate entrance examination mathematics corpus. The matching value between the incorrect questions and the knowledge point database is calculated using a semantic matching formula: the matching degree between the feature vector of the incorrect question and the feature vector of the knowledge point of L'Hôpital's rule for finding limits is 0.82, which is higher than the preset threshold of 0.6. The final output core knowledge point is L'Hôpital's rule for finding limits, the first-level subordinate knowledge point is equivalent infinitesimal substitution, the second-level subordinate knowledge point is the four arithmetic rules of limits, and the corresponding cognitive level is application level.

[0047] S3. Multi-dimensional error attribution analysis: Based on the large model identification, the error type corresponding to this question is calculation error or method misuse. Substituting these values ​​into the attribution weight formula, the normalized frequency value f for the calculation error category is calculated. p =0.7, Difficulty level of knowledge point d i =0.75, influence coefficient c p =0.6; Method misuse class f p =0.4, d i =0.75, c p =0.8; weighting coefficients are set to α=0.4, β=0.3, and γ=0.3. The final calculation shows that the computational error category has a higher weight and is determined to be the dominant error type, while the method misuse category is a secondary error type, generating a structured attribution report.

[0048] S4. Knowledge graph association and expansion: Taking L'Hôpital's rule for finding limits as the central node, a second traversal is performed on the knowledge graph of postgraduate entrance examination mathematics: the first traversal retrieves the definition of limits and the easily confused knowledge point of Taylor expansion for finding limits; the second traversal retrieves the set of equivalent infinitesimal formulas for the prerequisite knowledge points, and finally selects 4 related knowledge points to form a set of reinforced knowledge points.

[0049] S5. Personalized reinforcement training plan generation, using a difficulty adaptive algorithm to calculate the target difficulty: current training difficulty D. k =0.7, target accuracy threshold P0=0.8, user historical accuracy P k =0.5, learning rate η=0.1, the target difficulty D is calculated. k=0.73, and questions with a difficulty deviation of ±0.1 are matched from the question bank to form a graded question set. The question distribution follows the principle of proportionality between error weight: calculation errors account for 50% of the question types, misuse of methods accounts for 20% of the question types, basic concept questions account for 20%, and comprehensive improvement questions account for 10%. The training path is arranged in the following order: completion of prerequisite knowledge points, basic calculation of core knowledge points, identification of common error methods, variation questions of the same source, and comprehensive application questions.

[0050] S6. Strengthen training execution and iterative effects, pushing training content question by question. Multimodal content includes animated steps for formula derivation, textual breakdown of problem-solving approaches, and voice prompts for common mistakes. After each training unit is completed, the overall mastery is calculated using the following formula: average accuracy over the last three training sessions A=0.6, cumulative training sessions N=5, interval since last training Δt=1 day, saturation coefficient τ=10, forgetting constant T=7, and weighting coefficients λ1=0.6, λ2=0.2, λ3=0.2. The calculated mastery M=0.58, below the 0.6 threshold, automatically increases the proportion of basic calculation questions and reduces their difficulty, initiating a second round of reinforcement; once the mastery reaches ≥0.85, it transitions to spaced repetition mode.

[0051] Example 2: This embodiment addresses semantic-related incorrect questions in the postgraduate entrance examination's English reading comprehension module, implementing the reinforcement training method of this invention. The specific process is as follows: S1. Error data collection and structured preprocessing: Supports two collection methods: synchronous question bank answering and text input. Extracts fields such as reading comprehension stems, incorrect options, original text location sentences, and user error misunderstanding annotations. Uses a core vocabulary dictionary for postgraduate entrance examination English for word segmentation, annotates semantic fields, parts of speech, and synonym replacement features, filters meaningless function words, and constructs text semantic feature vectors.

[0052] S2. The large-scale model-driven precise positioning of incorrect knowledge points involves inputting structured incorrect questions into a domain-specific large-scale model that has been fine-tuned based on the postgraduate entrance examination English corpus. After calculating the semantic matching degree, the core knowledge point output is the reasoning and judgment question-solving logic, the first-level subordinate knowledge point is the semantic analysis of long and difficult sentences, the second-level subordinate knowledge point is the recognition of emotional attitude words, and the corresponding cognitive level is the analysis level.

[0053] S3. Multi-dimensional error attribution analysis: Based on the large model identification, the error types are classified as question misunderstanding and conceptual ambiguity (vocabulary). Substituting these into the attribution weight formula, the normalized frequency value f for question misunderstanding is calculated. p =0.6, Difficulty coefficient of knowledge point d i =0.6, influence coefficient c p =0.7; fuzzy concept class f p =0.5, d i =0.6, c p=0.75; weighting coefficients are α=0.4, β=0.3, and γ=0.3. The final determination is that misunderstanding the question is the dominant error type, while unclear vocabulary and concepts are the secondary error types.

[0054] S4. Knowledge Graph Association Expansion: Taking the reasoning and judgment question-solving logic as the central node, a second traversal is performed on the postgraduate entrance examination English knowledge graph: the first traversal retrieves the prerequisite knowledge points question stem positioning techniques, easily confused knowledge point detail questions and reasoning question question analysis; the second traversal retrieves the prerequisite knowledge points synonym substitution rules, and finally selects 3 related knowledge points to form a reinforcement set.

[0055] S5. Personalized reinforcement training plan generation, with adaptive difficulty calculation: Current training difficulty D k =0.6, target accuracy P0=0.8, user historical accuracy P k =0.55, learning rate η=0.1, the target difficulty D is calculated. k =0.625, and the reading comprehension questions with a difficulty deviation of ±0.1 form a graded question set. Question distribution: 50% for question analysis training, 20% for vocabulary and semantics, 20% for breaking down long and difficult sentences, and 10% for comprehensive reading questions; the training path is arranged in the following order: location skills completion, special training in question stem analysis, identification of option traps, consolidation of cognate reading, and improvement of comprehensive reasoning.

[0056] S6. Strengthen training execution and iterative effects by pushing training content in a multimodal format, including animated breakdowns of complex sentence structures, audio explanations of core vocabulary, and textual analysis of incorrect answer choices. Mastery is calculated after each unit: the average accuracy rate of the last three sessions A=0.65, the cumulative number of training sessions N=4, Δt=2 days, and substituting into the formula yields a mastery level M=0.63, which is within the critical range for achieving the target. The system maintains the current difficulty level and adds one more training session focusing on common mistakes. Once the mastery level is achieved, the training frequency is reduced, and content is pushed according to a spaced review pattern.

[0057] Example 3: This embodiment addresses cross-module errors in the data structures and operating systems section of the postgraduate entrance examination (subject 408), implementing the reinforcement training method of this invention. The specific process is as follows: S1. Error data collection and structured preprocessing: Collect users' comprehensive errors across modules, extract fields such as question stem, pseudocode answer, error logic nodes, and standard problem-solving ideas; use a computer science word segmentation dictionary (including data structure and operating system terminology) for word segmentation, label cross-module related features, and construct a structured error vector containing code features and logical features.

[0058] S2, a large model-driven system for precise location of incorrect knowledge points, inputs a domain-wide large model fine-tuned from a 408-subject corpus, identifies cross-module related nodes, calculates semantic matching degree for each, and finally outputs two core knowledge points: process scheduling algorithm (operating system) and queue application (data structure), with time complexity analysis as the subordinate knowledge point. The cognitive level covers both application and analysis levels, triggering a cross-module linkage mechanism.

[0059] S3. Multi-dimensional error attribution analysis identifies error types as conceptual confusion and logical reasoning errors; substituting these into the weighting formula calculates: Conceptual confusion type f p =0.65, average difficulty of knowledge points d i =0.8, influence coefficient c p =0.8; Logical reasoning error class f p =0.5, d i =0.8, c p =0.85; weighting coefficients are α=0.4, β=0.3, and γ=0.3. The final determination is that conceptual confusion is the dominant error type, and logical reasoning error is the secondary error type.

[0060] S4. Knowledge graph association and expansion: a second traversal is performed in the data structure and operating system subgraphs respectively. On the operating system side, the prerequisite knowledge points are expanded to include basic concepts of processes and easily confused knowledge points to distinguish between job scheduling and process scheduling. On the data structure side, the prerequisite knowledge points are expanded to include basic queue operations and application scenarios of stacks of the same origin. Finally, 5 related knowledge points are selected to form a cross-module reinforcement set.

[0061] S5. Personalized reinforcement training plan generation, with adaptive difficulty calculation: Current training difficulty D k =0.75, target accuracy P0=0.8, user historical accuracy P k =0.45, learning rate η=0.12, the target difficulty D is calculated. k The difficulty level is approximately 0.79, and questions with a difficulty deviation of ±0.1 form a graded question set. Question distribution: 50% for concept identification questions, 20% for logical reasoning questions, 20% for code implementation questions, and 10% for cross-module comprehensive questions; the training path is arranged in the following order: single-module pre-concept completion, single-module basic training, easily confused concept identification, single-module variation consolidation, and cross-module comprehensive improvement.

[0062] S6. Enhance training execution and iterative results by pushing training content in a multimodal format, including animations of process scheduling, line-by-line pseudocode analysis, and knowledge point comparison tables. Calculate overall mastery after each unit, assessing mastery of two core knowledge points separately: if mastery of a single knowledge point is below the threshold, supplementary basic training for that module is provided; if the accuracy rate of cross-module comprehensive questions is low, increase the proportion of cross-module collaborative questions to form a closed-loop iterative cross-module joint training.

[0063] Based on the above Examples 1-3, the following table 1 is derived:

[0064] Based on Table 1 above, the following conclusions can be drawn: The core framework of this invention, which consists of "error collection - large model localization - multi-dimensional attribution - knowledge graph expansion - adaptive training - closed-loop iteration", can be adapted to different subject scenarios such as postgraduate entrance examination mathematics, English, and professional courses. It can be quickly implemented and has high versatility. It only requires adjusting the word segmentation dictionary, fine-tuning the corpus, knowledge graph node attributes and algorithm parameters according to the characteristics of the subject.

[0065] Training resources need to be strongly linked to error types and adapted to different subjects. The dominant error types differ significantly: in science and calculation-related subjects, errors are mainly calculation mistakes and misuse of methods; training resources should focus on calculation practice and method analysis. In humanities and language-related subjects, errors are mainly due to misreading the question and semantic ambiguity; training should emphasize question reading and semantic decomposition. In interdisciplinary subjects, errors are mainly due to conceptual confusion and logical errors; training should strengthen conceptual analysis and cross-module linkage. A strategy of dynamically allocating the number of questions according to error weight can effectively improve the targeting of training and avoid ineffective practice.

[0066] Knowledge graph construction needs to match the knowledge structure of disciplines. For highly logical subjects like mathematics, the knowledge dependency chains are long, and knowledge graphs should deepen the second-order traversal of pre-dependent relationships to ensure the completeness of knowledge completion. For language-related subjects like English, the knowledge connections are more horizontal and easily confused; knowledge graphs should focus on expanding easily confused, cognate, and parallel relationships. For cross-module integrated subjects, a multi-subgraph linkage architecture needs to be constructed to support cross-domain related retrieval. Only differentiated graph construction strategies can maximize the value of knowledge point expansion.

[0067] Adaptive algorithm parameters need to be optimized according to subject characteristics. For subjects with higher knowledge point integration and greater difficulty fluctuations, the learning rate η for difficulty adaptation can be appropriately increased to ensure the sensitivity of difficulty adjustment; while for subjects with a gentle difficulty gradient, a conventional learning rate is sufficient to maintain stability. Calibrating parameters in conjunction with subject characteristics allows difficulty adjustments to better match the actual learning pace of test-takers, avoiding the impact of sudden changes in difficulty on training effectiveness.

[0068] Multimodal content adaptation to subject-specific expressions can improve training effectiveness. Matching corresponding multimodal formats to the knowledge expression characteristics of different subjects (science with formula animations, humanities with voice explanations, and engineering with process animations) can lower the threshold for understanding knowledge points and enhance the memory effect of correcting mistakes. This is an important optimization direction to improve the user's learning experience and training efficiency.

[0069] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application also intends to include these modifications and variations.

Claims

1. A method for intensive training of postgraduate entrance examination knowledge points based on precise location of incorrect questions using a large model, characterized by: Includes the following steps: S1. Error Data Collection and Structured Preprocessing: Collect error data from users' postgraduate entrance examination practice, extract text information such as question stems, incorrect answers, and problem-solving processes, and perform noise reduction, word segmentation, and structured annotation. S2, Big Data-Driven Precise Positioning of Wrong Question Knowledge Points: Input structured wrong question data into a finely tuned big data model for the postgraduate entrance examination, and output the core knowledge points, subordinate knowledge points and cognitive level labels corresponding to the wrong questions through semantic understanding and knowledge point mapping; S3. Multi-dimensional error attribution analysis: Based on the characteristics of incorrect questions and the attributes of knowledge points, the causes of errors are classified and identified, the weight of each type of error factor is calculated, and an error attribution report is generated. S4. Knowledge Graph Association Expansion: Based on the core knowledge points located, association retrieval is performed in the pre-constructed postgraduate entrance examination subject knowledge graph to expand the knowledge points to include prerequisite knowledge points, subsequent knowledge points, and common error-prone knowledge points. S5. Personalized reinforcement training program generation: Combining error attribution results with the relationship between knowledge points, adaptively generate reinforcement training paths that include question difficulty gradients, training frequency, and training order. S6. Enhance training execution and effect iteration: Push corresponding training content, collect user answer data, evaluate the mastery of knowledge points in real time, and dynamically adjust subsequent training plans based on evaluation results to form a closed-loop iteration.

2. The method for intensive training of postgraduate entrance examination knowledge points based on accurate location of incorrect questions using a large model, as described in claim 1, is characterized in that: The structured preprocessing in step S1 specifically includes: The text of incorrect questions is processed by a combination of OCR recognition and manual verification to extract fields such as the main body of the question, the question format, the incorrect options, the user's answer traces, and the error location markers. A word segmentation dictionary specifically for postgraduate entrance examination subjects is used for word segmentation and part-of-speech tagging to construct a standardized feature vector of incorrect questions.

3. The method for intensive training of postgraduate entrance examination knowledge points based on accurate location of incorrect questions using a large model, as described in claim 1, is characterized in that: Step S2, which involves precise localization of knowledge points in the large model, includes the calculation of knowledge point matching degree. The specific algorithm formula is as follows: Let the feature vector of the incorrect question text be... The feature vector of the i-th knowledge point in the knowledge point base is Then, the semantic matching degree Sim(T,Kᵢ) between the incorrect question and the i-th knowledge point is: ； Where w ⱼ The domain weight of the j-th feature word is determined by weighting the word frequency in the postgraduate entrance examination syllabus with the frequency of its appearance in past exam questions; Select knowledge points with a matching degree higher than a preset threshold as a candidate set, sort them in descending order of matching degree, and output the top-m core knowledge points and subordinate knowledge points.

4. The method for intensive training of postgraduate entrance examination knowledge points based on accurate location of incorrect questions using a large model, as described in claim 1, is characterized in that: Step S3, the multi-dimensional error attribution analysis, includes the error attribution weight calculation step. The specific algorithm formula is as follows: Let W be the attribution weight of the p-th error factor. p The calculation formula is: ； Where f p This is the normalized value of the frequency of this type of error in the user's historical incorrect questions; d i c represents the difficulty coefficient of the corresponding knowledge point in the postgraduate entrance examination syllabus, with a value range of 0-1; p α represents the impact coefficient of this type of error on the mastery of knowledge points; α, β, and γ are weighting coefficients, satisfying α+β+γ=1; Sort by weight value from high to low to identify dominant and secondary error types.

5. The method for intensive training of postgraduate entrance examination knowledge points based on accurate location of incorrect questions using a large model, as described in claim 1, is characterized in that: The construction method of the postgraduate entrance examination subject knowledge graph in step S4 is as follows: Using the official postgraduate entrance examination syllabus as the framework, a node system is constructed according to a three-level hierarchy of subject, chapter, and knowledge point. Node attributes include knowledge point name, difficulty level, frequency of examination, cognitive level, and percentage of past exam questions. Edge relationships include pre-dependency, subsequent extension, easily confused, and homologous parallel types. Through a large model, past exam questions and teaching materials are extracted to supplement the edge relationships and attribute values ​​of the knowledge graph.

6. The method for intensive training of postgraduate entrance examination knowledge points based on accurate location of incorrect questions using a large model, as described in claim 5, is characterized in that: The knowledge point association and expansion in step S4 are specifically as follows: Using core knowledge points as central nodes, a second-order traversal is performed in the knowledge graph to retrieve all prerequisite dependent nodes and easily confused nodes; based on the strength of the association between nodes and the user's historical knowledge, a set of related knowledge points that need to be supplemented and strengthened is selected.

7. The method for intensive training of postgraduate entrance examination knowledge points based on accurate location of incorrect questions using a large model, as described in claim 1, is characterized in that: Step S5, which generates a personalized reinforcement training scheme, includes a difficulty adaptive adjustment step. The specific algorithm formula is as follows: Let the target difficulty value of the next round of training questions for the k-th knowledge point be _____. The calculation formula is: ； Where D k The current training difficulty level is represented by a value between 0 and 1; P0 is the preset target accuracy threshold; P k This represents the user's historical accuracy rate in answering questions on this knowledge point; η is the learning rate adjustment factor, with a value ranging from 0.05 to 0.

2. After calculating the target difficulty, questions with difficulty values ​​within ±0.1 are matched from the question bank to form a graded training question set.

8. The method for intensive training of postgraduate entrance examination knowledge points based on accurate location of incorrect questions using a large model, as described in claim 1, is characterized in that: The rule for generating reinforcement training paths in step S5 is as follows: The training units are arranged in the following order: supplementing prerequisite knowledge points, explaining core knowledge points in detail, identifying common mistakes, consolidating related questions, and improving comprehensive questions. The number of questions in each training unit is dynamically allocated according to the error attribution weight, with higher weight error types corresponding to a larger number of training questions.

9. The method for intensive training of postgraduate entrance examination knowledge points based on accurate location of incorrect questions using a large model, as described in claim 1, is characterized in that: Step S6, which involves real-time assessment of knowledge point mastery, includes a step of quantifying mastery. The specific algorithm formula is as follows: Let M be the degree of comprehensive mastery of a certain knowledge point, and the calculation formula is: ； Where A is the average accuracy rate of the last three answers to this knowledge point; N is the cumulative number of training sessions. The time interval since the last training session; τ is the training saturation coefficient; T is the forgetting time constant; For the weighting coefficients, satisfying ; When the mastery level M is higher than the preset threshold, the training frequency for that knowledge point is reduced; if it is lower than the threshold, the training frequency is increased and the difficulty of the questions is reduced.

10. The method for intensive training of postgraduate entrance examination knowledge points based on accurate location of incorrect questions using a large model, as described in claim 1, is characterized in that: The fine-tuning method for the large model in the postgraduate entrance examination field in step S2 is as follows: Based on a general large language model, this study uses past postgraduate entrance examination questions, syllabus analysis, authoritative teaching materials, and standard error sets as fine-tuning corpora. An instruction set mapping error-knowledge point is constructed using an instruction fine-tuning method, and the model is trained to output standardized knowledge point labels and cognitive levels. In step S6, the reinforcement training content is generated in a multimodal format. The large model generates textual explanations, reasoning derivations, and error-prone point hints based on knowledge points and error types; it automatically matches corresponding formula derivation diagrams, logic framework diagrams, and problem-solving step animations based on knowledge point types; and it generates audio explanations for frequently missed knowledge points.

Images