A teacher individualized correction method and system based on a generative algorithm

Abstract

The application relates to a teacher personalized correction method and system based on a generative algorithm, and relates to the technical field of homework correction. The method comprises the following steps: collecting correction preference information and historical correction data of a teacher; generating a personalized correction model through machine learning based on the correction preference information and the historical correction data; receiving a homework file uploaded by a teacher end, preprocessing the homework file and extracting key information; obtaining homework score evaluation criteria according to a generative algorithm and the key information; receiving homework data uploaded by a student end; performing a correction operation on the homework data according to the generative algorithm, the personalized correction model and the homework score evaluation criteria, to obtain homework scores and personalized suggestions; and pushing the homework scores and the personalized suggestions to the student end and the teacher end. The application has the effects of improving homework correction efficiency and teaching quality and promoting the personalized development of students.

Description

Technical Field

[0001] This application relates to the field of homework correction technology, and in particular to a personalized correction method and system for teachers based on generative algorithms. Background Technology

[0002] With the advancement of educational informatization, homework grading is gradually shifting from purely manual methods to those utilizing electronic tools. From a technological perspective, the education sector has widely adopted electronic document processing and basic data analysis technologies, enabling teachers to assign and collect homework more conveniently.

[0003] Existing teacher grading methods primarily involve manual grading or grading on a homework grading system. Common homework grading systems typically consist of a homework management module, an answer matching module, and a result output module. In the homework management module, teachers can select assignments from the system's resource library to send to students, or upload assignments in a limited format (such as a document with a specific template). After students complete and submit their assignments, the answer matching module compares them against preset standard answers. For objective questions, it directly judges whether the answer is right or wrong; for subjective questions, it uses simple methods such as keyword matching to score. Finally, the result output module presents teachers and students with a relatively simple score and basic right / wrong judgment. For example, in grading Chinese compositions, points can only be deducted by identifying some common spelling and grammatical errors, and it is impossible to provide a comprehensive evaluation based on the article's theme, writing style, and other aspects.

[0004] Regarding the aforementioned technologies, existing systems lack personalized grading and feedback, cannot meet the needs of different teachers' teaching styles, and cannot provide students with learning suggestions tailored to their individual circumstances, thus affecting the improvement of teaching effectiveness. Summary of the Invention

[0005] To improve the efficiency of homework correction and teaching quality, and to promote students' personalized development, this application provides a teacher-based personalized correction method and system based on generative algorithms.

[0006] Firstly, this application provides a method and system for personalized teacher grading based on generative algorithms, employing the following technical solution:

[0007] A personalized teacher grading method based on generative algorithms includes:

[0008] Collect teachers' marking preferences and historical marking data;

[0009] Based on grading preference information and historical grading data, a personalized grading model is generated through machine learning.

[0010] Receive assignment files uploaded by teachers, preprocess the assignment files, and extract key information;

[0011] Based on the generative algorithm and key information, the evaluation criteria for assignment scores are obtained;

[0012] Receive homework data uploaded by students;

[0013] The homework data is graded based on generative algorithms, personalized grading models, and homework scoring criteria to obtain homework scores and personalized suggestions.

[0014] Homework scores and personalized suggestions are pushed to both students and teachers.

[0015] By adopting the above technical solutions, personalized grading models can be generated based on the collection of teachers' grading preferences and historical grading data. This enables multi-dimensional evaluation of different subjects, question types, and knowledge points, avoiding the shortcomings of mechanical comparison in traditional grading methods. Combining generative algorithms and key information, assignment scoring criteria are derived, and personalized scores and feedback are output in student assignment grading. This improves grading efficiency and accuracy, provides students with targeted improvement suggestions, offers teachers support for teaching decisions, enhances assignment grading efficiency and teaching quality, and promotes students' personalized development.

[0016] Optionally, the step of obtaining the evaluation criteria for the assignment score based on the generative algorithm and key information includes:

[0017] The key information is analyzed to obtain the analysis results, which include the subject category, question type characteristics, and target knowledge points of the homework questions;

[0018] Based on the analysis results, a generative algorithm is invoked to generate candidate reference answers;

[0019] The candidate reference answers are checked for semantic consistency, logical completeness, and knowledge point coverage to obtain a set of verified reference answers;

[0020] The assignment evaluation dimensions are constructed based on the reference answer set. The assignment evaluation dimensions include semantic similarity, logical and step completeness, and mastery of knowledge points.

[0021] Based on the evaluation dimensions of assignments and historical grading data, determine the weight allocation and grading thresholds for each evaluation dimension, and generate score mapping relationships and scoring rules.

[0022] Based on the score mapping relationship and scoring rules, and combined with the analysis results and reference answer set, the evaluation criteria for homework scoring are obtained.

[0023] By adopting the above technical solutions, a verified set of reference answers can be generated based on the analysis of key information. This allows for the construction of assignment evaluation dimensions that cover semantic similarity, logical and step completeness, and knowledge point mastery. Furthermore, by combining historical grading data to determine weights and scoring rules, a more scientific and reasonable assignment grading standard can be formed, thereby improving the accuracy and discriminative power of the grading results.

[0024] Optionally, the step of grading the assignment data according to the generative algorithm, personalized grading model, and assignment scoring criteria to obtain the assignment score includes:

[0025] Align the task data with the parsing results to obtain the task alignment result;

[0026] Based on the assignment scoring criteria, a generative algorithm is used to compare the assignment alignment results with the reference answer set semantically, step-by-step, and knowledge point coverage to obtain the original indicators for each assignment evaluation dimension.

[0027] Based on the original indicators and score mapping relationships, the initial scores of the assignment data in each assignment evaluation dimension are calculated.

[0028] Based on the initial score and scoring rules, the initial comprehensive score of the assignment data is calculated;

[0029] Input the initial score and initial composite score into the personalized grading model, and adjust the evaluation dimensions of the assignments.

[0030] Based on the revised evaluation dimensions, a comprehensive scoring operation is performed on the assignment data to obtain the assignment score.

[0031] By adopting the above technical solution, based on the alignment of assignment data and analysis results, and combined with the reference answer set for semantic comparison, step comparison, and knowledge point coverage comparison, multi-dimensional raw indicators can be obtained, avoiding the one-sidedness of traditional grading methods that rely solely on the final answer. Furthermore, through score mapping relationships and scoring rules, the raw indicators are quantified into initial scores and comprehensive scores, which are then input into a personalized grading model for correction, ensuring that the scoring results fully reflect the teacher's grading style and preferences. The final assignment scores not only possess objectivity and consistency but also reflect teachers' differentiated evaluation standards, thereby improving the accuracy and personalization of the grading results.

[0032] Optionally, the step of grading the assignment data based on generative algorithms, personalized grading models, and assignment scoring criteria to obtain personalized suggestions includes:

[0033] A difference analysis was conducted between the homework data and the reference answer set to identify the differences.

[0034] Map the differences to the target knowledge points to generate a knowledge point mastery profile;

[0035] Based on the knowledge point mastery profile, homework score, and revised homework evaluation dimensions, and combined with the grading style and expression preferences in the personalized grading model, the comment generation strategy and suggestion generation parameters are determined.

[0036] Generate personalized comments based on the comment generation strategy;

[0037] Based on the knowledge point mastery profile and suggested parameters, generate learning improvement suggestions;

[0038] Personalized feedback and learning improvement suggestions are combined to obtain personalized recommendations.

[0039] By adopting the above technical solutions, we can obtain a knowledge mastery profile based on differential analysis, and combine homework scores with a personalized grading model to generate personalized comments and targeted learning suggestions that match the teacher's style. This provides students with clear directions for improvement, while also enhancing the relevance and effectiveness of teacher feedback.

[0040] Optionally, scores for each assignment and personalized suggestions can be summarized by student and compiled into a learning record in chronological order.

[0041] Cluster analysis is performed on the learning records based on key information to obtain the performance sequence of the content dimensions contained in the key information.

[0042] Extract elements from personalized suggestions to obtain improvement element tags;

[0043] The improved element labels are associated with the performance sequence to form a set of learning performance features;

[0044] Based on the learning performance feature set, obtain the fluctuation index and determine whether the fluctuation index exceeds the preset fluctuation threshold and reaches the preset number of times within the preset observation window.

[0045] If so, the content dimension corresponding to the key information under the volatility indicator will be identified as the warning target;

[0046] Generate phased training plans for the targets of the early warning;

[0047] The phased training plan will be pushed to both students and teachers.

[0048] By adopting the above technical solution, learning records can be generated in chronological order at the student level, and performance sequences for different content dimensions can be obtained by clustering based on key information. Combining personalized suggestion element extraction with the correlation of performance sequences, a learning performance feature set is formed, thereby objectively calculating fluctuation indicators of assignment scores and identifying anomalies. Furthermore, abnormal content dimensions can be identified as early warning targets, and targeted, phased training plans can be generated and pushed to teachers and students, achieving dynamic monitoring and precise intervention of students' learning status.

[0049] Optionally, the step of generating a phased training plan for the early warning target includes:

[0050] The warning level is determined based on the magnitude of the fluctuation index exceeding the threshold and the cumulative number of times the fluctuation index exceeds the preset fluctuation threshold within the preset observation window.

[0051] The training plan parameters are determined based on the warning level and the personalized correction model.

[0052] The practice tasks are determined based on the content dimensions corresponding to the key information and the warning targets;

[0053] The practice tasks are divided into stages according to the planning parameters, and stage goals and completion criteria are set.

[0054] Based on the phase goals and completion criteria, a phased training plan is obtained.

[0055] By employing the aforementioned technical solution, the warning level can be determined based on the threshold amplitude and cumulative frequency of fluctuation indicators, thereby classifying the severity of student learning abnormalities. Furthermore, by integrating a personalized grading model to adjust the training plan parameters, the training tasks are made to both conform to the warning level and align with teachers' grading preferences. Then, based on the key information and the content dimensions corresponding to the warning recipients, personalized practice tasks are determined and arranged in stages according to the training plan parameters, clearly defining stage goals and completion criteria. The resulting staged training plan can accurately match students' weaknesses and provide a gradual improvement path.

[0056] Optionally, the step of determining the training plan parameters based on the early warning level and the personalized correction model includes:

[0057] The basic parameter values ​​are determined based on the warning level. The basic parameter values ​​include the initial values ​​for the number of priority improvement items, the intensity of practice tasks, and the recommended learning time.

[0058] The basic parameter values ​​are adjusted based on the personalized grading model to obtain parameter correction results that conform to teachers' preferences;

[0059] The difficulty level, number of stages, workload of each stage, and check node time are determined based on the parameter correction results.

[0060] The training plan parameters are obtained based on the difficulty level, number of stages, stage task volume, and check node time.

[0061] By adopting the above technical solution, the basic parameter values ​​of the training task can be determined based on the early warning level, and then corrected in conjunction with a personalized grading model to better align with teachers' grading preferences. Based on the corrected parameter results, the difficulty level, number of stages, workload, and inspection node time are further determined, ultimately resulting in more accurate and personalized training plan parameters, enabling targeted learning intervention.

[0062] Secondly, this application provides a personalized teacher grading system based on generative algorithms, employing the following technical solution:

[0063] A personalized teacher grading system based on generative algorithms includes:

[0064] The acquisition module is used to acquire preference information, historical grading data, assignment files, and assignment data.

[0065] The memory is used to store the program for the aforementioned personalized teacher grading method based on generative algorithms;

[0066] The processor and memory can load and execute the program to implement the aforementioned personalized teacher grading method based on generative algorithms.

[0067] By adopting the above technical solutions, personalized grading models can be generated based on the collection of teachers' grading preferences and historical grading data. This enables multi-dimensional evaluation of different subjects, question types, and knowledge points, avoiding the shortcomings of mechanical comparison in traditional grading methods. Combining generative algorithms and key information, assignment scoring criteria are derived, and personalized scores and feedback are output in student assignment grading. This improves grading efficiency and accuracy, provides students with targeted improvement suggestions, offers teachers support for teaching decisions, enhances assignment grading efficiency and teaching quality, and promotes students' personalized development.

[0068] Thirdly, this application provides a smart terminal, which adopts the following technical solution:

[0069] A smart terminal includes a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and execute the method described in any of the above-mentioned embodiments.

[0070] Fourthly, this application provides a computer storage medium capable of storing corresponding programs, which facilitates improving homework grading efficiency and teaching quality, and promotes students' personalized development. The technical solution adopted is as follows:

[0071] A computer-readable storage medium storing a computer program capable of being loaded by a processor and executed for any of the aforementioned generative algorithm-based personalized teacher grading methods.

[0072] In summary, this application includes at least one of the following beneficial technical effects:

[0073] Based on the collection of teachers' grading preferences and historical grading data, this system can generate personalized grading models, enabling multi-dimensional evaluation of different subjects, question types, and knowledge points, avoiding the shortcomings of mechanical comparison in traditional grading methods. By combining generative algorithms and key information, it obtains homework scoring criteria and outputs personalized scores and feedback in student homework grading. This improves grading efficiency and accuracy, provides students with targeted improvement suggestions, offers teachers teaching decision support, enhances homework grading efficiency and teaching quality, and promotes students' personalized development.

[0074] It can generate a verified set of reference answers based on the analysis of key information, construct an assignment evaluation dimension that covers semantic similarity, logical and step completeness, and knowledge mastery, and determine weights and scoring rules by combining historical grading data, thereby forming a more scientific and reasonable assignment grading standard and improving the accuracy and discrimination of the grading results.

[0075] It can generate knowledge mastery profiles based on differential analysis, and combine homework scores with personalized grading models to generate personalized comments and targeted learning suggestions that match the teacher's style, thereby providing students with clear directions for improvement, while enhancing the relevance and effectiveness of teacher feedback. Attached Figure Description

[0076] Figure 1 This is a flowchart illustrating a personalized teacher grading method based on a generative algorithm, as described in an embodiment of this application.

[0077] Figure 2 This is a flowchart illustrating the steps in this application embodiment to obtain the evaluation criteria for job scores based on generative algorithms and key information.

[0078] Figure 3 This is a schematic diagram of the process for obtaining the job score in an embodiment of this application.

[0079] Figure 4 This is a schematic diagram of the process for obtaining personalized suggestions in the embodiments of this application.

[0080] Figure 5 This is a flowchart illustrating a phased training generation method in an embodiment of this application.

[0081] Figure 6 This is a flowchart illustrating the steps of generating a phased training plan for a warning target in an embodiment of this application.

[0082] Figure 7 This is a flowchart illustrating the steps of determining training plan parameters based on the warning level and the personalized correction model in this embodiment of the application. Detailed Implementation

[0083] To make the purpose, technical solution, and advantages of this application clearer, the following description is provided in conjunction with the appendix. Figure 1 - Appendix Figure 7 The present application will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the application.

[0084] This application discloses a personalized teacher grading method based on a generative algorithm. (Refer to...) Figure 1 Personalized teacher grading methods based on generative algorithms include:

[0085] Step S101: Collect teachers' marking preferences and historical marking data.

[0086] Grading preference information refers to the personalized tendencies shown by teachers when grading assignments, such as emphasizing theme, writing style, or structure in Chinese composition, or emphasizing the completeness of the proof or the correctness of the answer in math problems.

[0087] Historical grading data refers to the records of homework grading completed by teachers in the past, including the original student homework, teacher grading results, annotation marks, comment text, and corresponding timestamp information, which are used to reflect the teacher's actual grading habits.

[0088] Among them, grading preference information can be obtained through questionnaires, and historical grading data can be obtained by calling the homework grading database linked to the teacher's account or by accessing local storage files.

[0089] For example, when a teacher logs in to the teacher's client for the first time, they will complete the initial settings of their grading preferences through the interactive interface. This includes selecting a grading style (encouraging or strict) and key focus dimensions (theme, writing style, structure, logical completeness, etc.), and then uploading the settings to the server for storage. Simultaneously, the system will access the historical grading database linked to the teacher's account, retrieve and parse previously graded assignment data, extract and convert the original assignment text, grading records, comments, and annotations into structured data to form the initial dataset for the teacher's personalized grading habits.

[0090] Step S102: Based on grading preference information and historical grading data, generate a personalized grading model through machine learning.

[0091] Personalized grading models are models trained using machine learning algorithms that simulate the grading styles and habits of specific teachers. These models use teachers' grading preferences as guiding parameters and historical grading data as training samples to learn teachers' specific scoring tendencies and comment styles across different subjects, question types, and grading dimensions.

[0092] The machine learning module is invoked, using grading preference information as input constraints and historical grading data as the training sample set to perform feature extraction and model training. Through iterative training, the model gradually fits the distribution of teachers' scores and the patterns of comment generation across different dimensions, and outputs a corresponding personalized grading model for scoring and feedback in subsequent homework grading processes.

[0093] In one feasible implementation, historical grading data is first characterized by converting teachers' scores, annotations, and written comments into numerical or labeled features. Then, a deep learning-based natural language processing model (such as the Transformer architecture) is used to semantically vectorize the comments. Simultaneously, teacher-defined preference parameters are incorporated as weighting factors into the loss function to guide the model to better align with teacher preferences during training. After training, the model can automatically adjust the weighting of different evaluation dimensions based on teacher habits, for example, emphasizing "logical completeness" or "creative expression."

[0094] Step S103: Receive the assignment file uploaded by the teacher, preprocess the assignment file and extract key information.

[0095] Homework files refer to electronic files containing student homework content uploaded by teachers through the teacher's terminal. File formats can include Word, PDF, images, etc.

[0096] Key information refers to the homework question information extracted by the system from the preprocessed homework file, including subject category, question type characteristics, and target knowledge points.

[0097] Assignment files are received via the teacher's upload interface, and their format is parsed and standardized. For text files, text extraction and paragraph segmentation are performed; for image files, OCR recognition and text reconstruction are performed. Then, the natural language processing module is invoked to perform semantic analysis of the assignment questions, automatically identifying the subject, question type, and target knowledge points, thereby generating key information for subsequent steps.

[0098] For example, in a Chinese composition scenario, the teacher uploads a PDF composition title file, which is then preprocessed into text, and the system extracts "subject category = Chinese", "question type feature = composition", and "target knowledge points = writing theme, language expression"; in a math proof scenario, the teacher uploads a Word file, and the system parses the question "Prove that ∠A = ∠B in triangle ABC", and identifies "subject category = math", "question type feature = proof", and "target knowledge points = geometric proof, properties of isosceles triangles".

[0099] Step S104: Based on the generative algorithm and key information, obtain the evaluation criteria for the assignment score.

[0100] Generative algorithms are a class of algorithms that can automatically generate new content based on existing data and model parameters. They are deep learning models based on the Transformer architecture (such as pre-trained language models), and their core mechanism is self-attention, which can model the semantic relationships between different parts of the input text to generate content that meets logical and semantic requirements. In this application, generative algorithms are mainly used to: generate candidate reference answers based on the key information of the assignment questions (subject category, question type characteristics, target knowledge points); generate assignment evaluation dimensions and scoring rules; and generate comments and learning suggestions in the personalized feedback stage.

[0101] The scoring criteria for assignments refer to the scoring rules system used in the grading process. This standard is generated by a generative algorithm combined with key information and covers the weighting, grading thresholds, and score mapping relationships of assignment evaluation dimensions such as semantic similarity, logical and step completeness, and mastery of knowledge points.

[0102] By analyzing the extracted key information, the subject category, question type, and target knowledge points of the assignment are determined. A generative algorithm is then used to generate candidate reference answers that match the key information. These reference answers are then validated for semantic consistency, logical completeness, and knowledge point coverage, resulting in a validated set of reference answers. Based on this set, assignment evaluation dimensions are constructed. Combined with teachers' historical grading data, the weights and grading thresholds of each evaluation dimension are dynamically adjusted to form the assignment scoring criteria.

[0103] The specific steps for obtaining the assignment score evaluation criteria based on the generative algorithm and key information can be found in the following reference. Figure 2 Example.

[0104] Step S105: Receive homework data uploaded by students.

[0105] Homework data refers to the homework files or answer data completed by students. The file format can include text, images, etc., and includes student identification and homework submission time information.

[0106] In one feasible implementation, an upload portal is set up on the student's end, allowing students to upload assignments by selecting local files or taking photos. Upon receiving assignment data, the system automatically adds the student ID, class ID, and submission timestamp to ensure data traceability. Uploaded text files are directly stored in the database; uploaded image files are converted into a processable text format using an OCR recognition module; for audio and video files (such as oral tests), the system uses a speech recognition or video transcription module to generate a transcript, which is then saved along with the original file. After all assignment data is stored, the system generates a unique data index number to ensure accurate location of the corresponding assignment data in subsequent grading processes.

[0107] For example, when a student uploads a photo of a math assignment, the system receives it and uses the OCR recognition module to extract the questions and answers. It then binds and stores the "Student ID=2023001", "Class ID=Grade 11 Class 3", and "Submission Time=September 10, 2025 20:15" as metadata along with the assignment data.

[0108] Step S106: Based on the generative algorithm, personalized grading model, and assignment scoring criteria, grade the assignment data to obtain the assignment score and personalized suggestions.

[0109] Homework scores are comprehensive scores calculated based on the evaluation criteria for each assignment, quantifying student data across various evaluation dimensions. These scores reflect the final scores that align with the teacher's grading preferences.

[0110] Personalized suggestions are improvement opinions generated based on students' learning progress by combining homework scores, knowledge point mastery profiles, and personalized grading models. These suggestions include personalized comments and learning improvement recommendations.

[0111] The specific steps for obtaining the assignment score can be found in the following reference. Figure 3 The steps in the embodiments will not be repeated here; the specific steps to obtain personalized suggestions can be found in [reference needed]. Figure 4 The steps described in the embodiments will not be repeated here.

[0112] Step S107: Push the assignment score and personalized suggestions to the student and teacher terminals.

[0113] After grading, the assignment scores and personalized suggestions are linked to student identity information and the data index number of the assignment data, generating corresponding data records. These records are then synchronized to both the student and teacher ends via a push notification mechanism. The student end primarily displays their individual assignment scores, personalized comments, and suggestions for improvement; the teacher end not only displays the data records of individual students' assignments but also provides statistical summaries and visual analysis of the overall class performance.

[0114] Reference Figure 2 The steps for obtaining the evaluation criteria for assignment scores based on generative algorithms and key information include:

[0115] Step S201: Analyze the key information to obtain the analysis results, which include the subject category, question type characteristics, and target knowledge points of the homework questions.

[0116] The analysis results refer to the analysis results of the homework questions obtained after semantic analysis and structured processing of key information. These mainly include the subject category, question type characteristics, and target knowledge points. Subject category: for example, Chinese, Mathematics, English, Physics. Question type characteristics: for example, multiple choice questions, fill-in-the-blank questions, problem-solving questions, and essays. Target knowledge points: refer to the specific teaching knowledge points corresponding to the questions, such as "linear functions," "Newton's second law," and "writing theme."

[0117] In one feasible implementation, when parsing key information, a classification model is first used to identify the subject of the question text, for example, using a convolutional neural network or Transformer classifier to determine that "this question belongs to mathematics". Then, rule matching and semantic discrimination algorithms are used to extract question type features, such as "if the question stem contains 'choose one correct answer,' it is determined to be a multiple-choice question"; "if the question stem contains 'write an essay...', it is determined to be an essay". Finally, the question text is input into a knowledge point mapping model. This model, based on a pre-trained semantic matching network and knowledge graph, aligns the keywords in the question stem with a predefined knowledge point database to obtain the target knowledge point. For example, if the question contains "y=kx+b", the target knowledge point is identified as "linear function".

[0118] Step S202: Based on the analysis results, call the generative algorithm to generate candidate reference answers.

[0119] Candidate reference answers refer to a set of multiple alternative answers generated by the generative algorithm based on the analysis results, which can be used for comparison during grading. These candidate reference answers not only include standard solutions, but can also cover diverse and reasonable solution paths and expressions.

[0120] Using the parsing results as input, a generative algorithm is invoked, with the subject category of the question and the target knowledge point as core constraints, to generate multiple candidate reference answers. In one feasible implementation, the system selects the corresponding generative algorithm model based on the subject category in the parsing results.

[0121] If the subject category is mathematics, call the inference-based generative model based on the Transformer architecture to generate the standard solution and possible multi-step derivation paths;

[0122] If the subject category is Chinese, a language model based on natural language generation is invoked to generate sample answers for essays with different styles and themes. During the generation process, the target knowledge points are used as generation prompts, and the question stem is input into the model to obtain multiple independent candidate sample answers.

[0123] For example, in the mathematical proof problem "Prove that the base angles of an isosceles triangle are equal", the system, based on the analytical results (subject category = mathematics, problem type = proof problem, target knowledge point = properties of isosceles triangles), calls a generative reasoning model to generate two candidate reference answers:

[0124] Suggested answer A: Use congruent triangles to prove ∠A=∠B;

[0125] Suggested answer B: Use angle bisectors and symmetry to prove ∠A=∠B.

[0126] Step S203: Verify the semantic consistency, logical completeness, and knowledge point coverage of the candidate reference answers to obtain the verified reference answer set.

[0127] Semantic consistency is used to determine whether candidate reference answers semantically match the requirements of the assignment question, thus avoiding irrelevant answers.

[0128] Logical integrity is used to determine whether a candidate answer is complete and reasonable in terms of its reasoning chain, argumentation process, or expression structure.

[0129] Knowledge point coverage is used to determine whether the candidate reference answer covers the target knowledge points extracted from the analysis results.

[0130] The set of qualified answers retained after passing the above three checks will serve as the benchmark for grading subsequent assignments.

[0131] In one feasible implementation, the system performs three types of checks on the candidate reference answers:

[0132] Semantic consistency test: The semantic embedding model (such as BERT) is used to calculate the semantic similarity between the candidate reference answer and the question stem. If the similarity is lower than the preset threshold (such as 0.7), the semantic deviation is determined and the answer is removed.

[0133] Logical integrity check: For mathematical problem-solving, the step analysis algorithm is used to check whether the reasoning covers the necessary steps; for Chinese composition, the paragraph structure is analyzed to see if it includes the theme, argument, and conclusion.

[0134] Knowledge point coverage test: The matching between candidate reference answers and target knowledge points is compared using a knowledge graph. If major knowledge points are missing, the weight is reduced or the answer is removed.

[0135] Finally, the candidate answers that meet the three tests will be compiled into a verified reference answer set.

[0136] Step S204: Construct assignment evaluation dimensions based on the reference answer set. Assignment evaluation dimensions include semantic similarity, logical and step completeness, and mastery of knowledge points.

[0137] Semantic similarity is used to determine the semantic similarity between the answers to questions in homework data and the reference answers.

[0138] Logical and procedural completeness is used to judge the degree of completeness of the answer to a question in the task data in terms of the reasoning chain or expression structure.

[0139] The knowledge point mastery status is used to determine how well the answers to the questions in the homework data cover the target knowledge points in the analysis results.

[0140] First, semantic analysis is performed on the text content of the reference answer set to obtain semantic vectors, which are used to construct the semantic similarity evaluation dimension. Next, the content of the reference answer set is decomposed sequentially and structurally to form a hierarchical structure, which serves as the logical and step-by-step completeness evaluation dimension. Then, the knowledge points involved in the reference answer set are compared with the target knowledge points in the analysis results to form the knowledge point mastery evaluation dimension. The three dimensions—semantic similarity, logical and step-by-step completeness, and knowledge point mastery—are combined to form the assignment evaluation dimension.

[0141] Step S205: Based on the assignment evaluation dimensions and historical grading data, determine the weight allocation and grading thresholds for each assignment evaluation dimension, and generate score mapping relationships and scoring rules.

[0142] Weighting refers to setting the relative importance of different evaluation dimensions for different tasks, which is used to reflect the contribution of each dimension to the total score.

[0143] Grading thresholds refer to the numerical boundaries that divide homework performance into different level intervals (such as excellent, satisfactory, and unsatisfactory), used to determine the score level of homework data in each dimension.

[0144] The score mapping relationship refers to the correspondence between the original indicator values ​​of each task evaluation dimension and the specific scores.

[0145] The scoring rules are a quantitative scoring method composed of weight allocation and tier thresholds.

[0146] In one feasible implementation, firstly, statistical analysis of historical grading data is used to calculate the average distribution and variance of teachers across different dimensions, reflecting their scoring tendencies. For example, if there are significant differences in the distribution of scores on the logical completeness dimension, it indicates that this dimension has high discriminative power, and therefore, it is given greater weight. Secondly, based on the distribution patterns of scores in historical grading records, reasonable grading thresholds are set, such as "above 0.8 is excellent, 0.6–0.8 is satisfactory, and below 0.6 is unsatisfactory." Then, the original indicators of the assignment evaluation dimensions are mapped to the grading thresholds to generate a score mapping relationship; for example, semantic similarity = 0.75 is mapped to 8 points. Finally, the scores of each assignment evaluation dimension are weighted according to their respective weights to form a complete scoring rule.

[0147] For example, in a set of historical grading data, teachers paid high attention to the logical completeness dimension. Based on the historical grading data, the weight allocation was determined as follows: semantic similarity weight = 0.3, logical completeness weight = 0.5, and knowledge point mastery weight = 0.2. The grading thresholds were set as follows: semantic similarity ≥ 0.8 is considered excellent, 0.6–0.8 is considered satisfactory, and below 0.6 is considered unsatisfactory. According to this rule, a student's semantic similarity = 0.75 (8 points), logical completeness = 0.85 (9 points), and knowledge point mastery = 0.65 (7 points), for a total score of 0.3 × 8 + 0.5 × 9 + 0.2 × 7 = 8.3 points.

[0148] Step S206: Based on the score mapping relationship and scoring rules, and combined with the analysis results and reference answer set, obtain the assignment scoring criteria.

[0149] First, the subject categories, question types, and target knowledge points extracted from the analysis results are mapped to the answer content in the reference answer set to ensure that the scoring covers key assessment points. Then, using the score mapping relationship and scoring rules, each assignment evaluation dimension is transformed into a specific score calculation method. Finally, the quantitative results of each assignment evaluation dimension are integrated according to the weights in the scoring rules to generate a complete assignment scoring and evaluation standard.

[0150] For example, taking a question about the "law of conservation of energy" as an example, the analysis results show that the target knowledge point of the question is the conservation of energy and its applications. The reference answer set provides two verified standard answers: one based on theoretical derivation and the other based on experimental phenomena. In the assignment evaluation dimensions, logical and step-by-step completeness is assigned a weight of 0.4, knowledge point mastery is assigned a weight of 0.35, and semantic similarity is assigned a weight of 0.25. When a student's answer is basically complete in logical derivation but slightly omits knowledge points, a corresponding quantitative score will be given based on the score mapping relationship, and the final assignment scoring criteria will be formed by combining these scores to ensure that the scoring results both conform to the teacher's grading preferences and maintain objectivity and consistency.

[0151] Reference Figure 3 The steps for grading assignment data and obtaining assignment scores based on generative algorithms, personalized grading models, and assignment scoring criteria include:

[0152] Step S301: Align the job data with the parsing results to obtain the job alignment result.

[0153] Homework alignment results refer to the one-to-one matching results between the homework data submitted by students and the analysis results, clarifying the correspondence between students' answers under subject categories, question type characteristics, and target knowledge points, including matching based on question type characteristics, mapping based on target knowledge points, and constraints based on subject categories.

[0154] Among them, matching based on question type features: for example, multiple choice questions, fill-in-the-blank questions, and problem-solving questions. Different question types have different data formats, so it is necessary to match the presentation of the questions and answers in the students' homework data with the characteristics of the question types.

[0155] Mapping based on target knowledge points: The semantic content, key expressions, or calculation steps in the answers to questions in students' homework data are mapped to the target knowledge points in the analysis results to confirm the scope of knowledge points involved in the students' answers.

[0156] Subject category constraints: Further verify the logic and expression of the answers within the same subject category to make the alignment results more in line with the requirements of the specific subject.

[0157] For example, the question analysis results are: subject category "mathematics", question type "calculation problem", target knowledge point "solving a linear equation in one variable"; student's homework data is "x=2".

[0158] First, confirm that "x=2" belongs to the category of "calculation problems";

[0159] Then compare "x=2" with the target knowledge point "solving linear equations in one variable" to confirm that it covers this knowledge point;

[0160] Under the constraint of the subject category being mathematics, confirm that the answer conforms to the logical expression of that subject.

[0161] The final generated homework alignment result is: "Calculation problem - Solving a linear equation in one variable - Student answer x=2".

[0162] Step S302: Based on the assignment scoring criteria, a generative algorithm is called to perform semantic comparison, step comparison, and knowledge point coverage comparison between the assignment alignment results and the reference answer set to obtain the original indicators for each assignment evaluation dimension.

[0163] The raw metrics refer to the quantitative values ​​obtained under the dimensions of task evaluation (such as semantic similarity, logical and step completeness, and knowledge mastery), which serve as the input for calculating the initial score in subsequent steps.

[0164] Semantic comparison refers to using generative algorithms to semantically match the answers in the homework alignment results with the reference answers, calculate the similarity in the content of the two, and obtain a semantic similarity index.

[0165] Step-by-step comparison refers to comparing the consistency between the reasoning process of the answer in the task alignment results and the logical chain of the reference answer, and obtaining a logical and step-by-step integrity index.

[0166] The knowledge point coverage comparison determines whether the answers in the homework alignment results cover the target knowledge points covered by the reference answer set.

[0167] In one feasible implementation, textual features are first extracted from the answers in the assignment alignment results, and semantic comparison is performed with the reference answer set. Cosine similarity, edit distance, or context similarity based on generative algorithms are used to calculate semantic similarity indices. Then, the solution step sequence of the reference answer is compared with the reasoning process of the student's answer, and the number of missing and incorrect steps is counted to obtain a logic and step completeness index. Further, the knowledge points involved in the student's answer are matched one-to-one with the target knowledge points extracted from the analysis results, and the proportion of covered, missing, and incorrectly covered knowledge points is counted to obtain a knowledge point mastery index. Finally, according to the assignment scoring criteria, the above semantic similarity index, logic and step completeness index, and knowledge point mastery index are mapped to preset grading thresholds to form the original indices for each assignment evaluation dimension.

[0168] For example, the reference answer was "First subtract 3 from both sides of the equation, then divide by 2, and we get x=4"; the student's answer was "We get x=4". In the semantic similarity comparison, the student's answer and the reference answer were consistent, with a similarity of 0.95. In the logical and step completeness comparison, the student's answer lacked a derivation process, resulting in a logical completeness score of 0.5. Regarding knowledge point mastery, the student's answer covered the knowledge points of the final solution, but did not cover the "subtraction" and "division" operations in the solution steps, with a coverage rate of 0.7. The final raw indicators were: semantic similarity = 0.95, logical and step completeness = 0.5, and knowledge point mastery = 0.7.

[0169] Step S303: Calculate the initial scores of the task data in each task evaluation dimension based on the original indicators and score mapping relationship.

[0170] First, the original indicators such as semantic similarity, logical and step completeness, and knowledge mastery are substituted into the score mapping relationship. Based on the grading threshold, the original indicators are transformed into standardized score intervals, and the initial score of the assignment data in the assignment evaluation dimension is calculated.

[0171] A student's homework answers were highly consistent with the reference answers in terms of semantic expression, with a raw semantic similarity score of 0.88; one logical step was missing, with a raw score of 0.6; and two of the three knowledge points in the reference answers were covered, with a raw score of 0.67. After score mapping and conversion according to the scoring rules, the semantic similarity score was 92, the logical and step completeness score was 70, and the knowledge point mastery score was 75, forming the student's initial scores in each dimension.

[0172] Step S304: Calculate the initial comprehensive score of the assignment data based on the initial score and scoring rules.

[0173] The initial composite score refers to the overall score obtained by weighting and integrating the initial scores of each task evaluation dimension according to the scoring rules.

[0174] For example, in the scoring rules, the weight of the semantic similarity dimension is 0.4, the weight of the logical completeness dimension is 0.35, and the weight of knowledge point mastery is 0.25. If a student's initial scores in semantic similarity, logical completeness, and knowledge point mastery are 92, 70, and 75 respectively, then their initial composite score can be calculated using the following formula:

[0175] Initial composite score = 80 × 0.4 + 70 × 0.35 + 90 × 0.25 = 83.80.

[0176] Step S305: Input the initial score and initial comprehensive score into the personalized grading model, and correct the evaluation dimensions of the assignment.

[0177] After obtaining the initial score and initial composite score, these results are input into the personalized grading model. The personalized grading model first analyzes the weighting of the initial composite score, identifying the degree of influence of each assignment evaluation dimension in the composite score formation process. Based on this, the personalized grading model adjusts the weighting of the assignment evaluation dimensions according to the teacher's grading preferences. For example, if the teacher prefers to emphasize logical process, the model will increase the weight of the logic and step completeness dimension; if the teacher values ​​knowledge point coverage more, the model will increase the weight of the knowledge point mastery dimension.

[0178] The student's initial scores across the three dimensions were: semantic similarity 92, logical completeness 70, and knowledge mastery 75. Their initial composite score, calculated based on the original weight allocation (semantic similarity 0.4, logical completeness 0.35, knowledge mastery 0.25), was 83.8. When the initial score and composite score were input into the personalized grading model, the model first analyzed the weight allocation of these three dimensions; then, considering the teacher's grading preferences, it adjusted the weight of the logical and step-by-step completeness dimension from 0.35 to 0.45, the weight of the semantic similarity dimension from 0.4 to 0.3, and kept the knowledge mastery dimension unchanged at 0.25.

[0179] Step S306: Based on the revised evaluation dimensions, perform a comprehensive scoring operation on the assignment data to obtain the assignment score.

[0180] The comprehensive scoring operation refers to the process of scoring according to the weight allocation and scoring rules of the revised evaluation dimensions of the assignment.

[0181] For example, a student's initial scores on semantic similarity, logical completeness, and knowledge mastery are 92, 70, and 75, respectively. After calculation in step S304, the initial comprehensive score is 83.8. In step S305, the personalized grading model increases the weight of the logical completeness dimension from 0.35 to 0.45, decreases the weight of the semantic similarity dimension from 0.4 to 0.3, and keeps the weight of the knowledge mastery dimension unchanged at 0.25. In this step, the comprehensive score is recalculated based on the revised weight allocation.

[0182] Homework score = 92 × 0.3 + 70 × 0.45 + 75 × 0.25 = 81.5.

[0183] Reference Figure 4 The steps for grading assignment data and obtaining personalized suggestions based on generative algorithms, personalized grading models, and assignment scoring criteria include:

[0184] Step S401: Perform a differential analysis on the homework data and the reference answer set to identify the differences.

[0185] Differentiation analysis refers to the process of using generative algorithms to perform semantic and step-by-step differentiation analysis on students' homework answers and reference answers.

[0186] Differences refer to the differences between students' homework answers and the reference answer set in terms of semantic expression and problem-solving steps.

[0187] Step S402: Map the differences to the target knowledge points to generate a knowledge point mastery profile.

[0188] A knowledge point mastery profile refers to the degree to which students have mastered each target knowledge point. By mapping differences to corresponding knowledge points, a structured mastery status label is formed.

[0189] Map the points of difference to the target knowledge points. If a knowledge point has many points of difference, it is marked as "not mastered" or "insufficiently mastered" in the knowledge point mastery profile; if there are no points of difference, it is marked as "mastered" in the knowledge point mastery profile. By mapping all the points of difference, a knowledge point mastery profile covering the target knowledge point can be obtained.

[0190] Step S403: Based on the knowledge point mastery profile, homework score, and revised homework evaluation dimensions, and combined with the grading style and expression preferences in the personalized grading model, determine the comment generation strategy and suggestion generation parameters.

[0191] Grading style and expression preference refer to the teacher's personal habits stored in the personalized grading model, such as "rigorous grading", "encouraging feedback", "concise expression" or "detailed analysis".

[0192] The evaluation generation strategy refers to the overall logic used when generating personalized evaluations, including different focuses such as emphasizing strengths, pointing out weaknesses, and making suggestions for improvement.

[0193] Suggestion generation parameters refer to the specific set of parameters required when generating learning improvement suggestions, such as which knowledge points to suggest improvements for, the level of detail in the suggestions, and the tone of expression.

[0194] First, based on the assignment scores and revised evaluation dimensions, the strengths and weaknesses of students in both overall performance and specific dimensions are identified. Then, these results are combined with a knowledge mastery profile to pinpoint students' weaknesses. Next, a personalized grading model is invoked, and the focus of the generated comments is determined based on the teacher's grading style (e.g., "emphasis on logic" or "encouragement of expression") and expression preferences (e.g., "concise and intuitive" or "detailed and specific"). For example, for teachers who emphasize logic, the comment generation strategy will tend to provide a more detailed explanation of the completeness of the problem-solving steps and highlight logic-related content in the suggested parameters.

[0195] For example, when analyzing a student's math homework, the student's score was 78 points, with a significantly low score in the logic and step completeness dimension (65 points) and a high score in the semantic similarity dimension (90 points). Simultaneously, the knowledge point mastery profile showed insufficient mastery of "discriminative reasoning." If the teacher's grading style leans towards "rigorous logic" and their expression preference is "objective and detailed," the following strategies will be generated: Comment generation strategy: Focus on providing feedback on the student's deficiencies in logical steps and offer rigorous comments based on the score; Suggestion generation parameters: Generate supplementary practice suggestions for the "discriminative reasoning" knowledge point and provide detailed reasoning process hints.

[0196] Step S404: Generate personalized comments based on the comment generation strategy.

[0197] Personalized comments refer to customized written feedback generated by combining students' homework scores, their mastery of knowledge points, and teachers' grading style and expression preferences.

[0198] By invoking a generative algorithm and using the comment generation strategy as input control conditions, corresponding comment content is generated. For example, when the teacher's style emphasizes logic, the comment will point out the advantages and disadvantages of the problem-solving steps in more detail; when the teacher's style is more encouraging, more positive language will be used.

[0199] For example, in a math homework scenario, a student scored 78 points, with low scores in the logical reasoning and step completeness dimensions, and insufficient mastery of the "discriminative calculation" knowledge point. The comment generation strategy is "rigorous logic and objective detail." Therefore, the system calls a generative algorithm to generate the following comment:

[0200] "Overall, this assignment was completed well, and the student was able to correctly understand most of the questions. However, there were errors in the discriminant calculation during the problem-solving process, resulting in incomplete logical reasoning chains. It is recommended that you focus on practicing the application of the discriminant formula during your review and pay attention to the completeness of each step of the reasoning."

[0201] Step S405: Based on the knowledge point mastery profile and suggested parameters, generate learning improvement suggestions.

[0202] Learning improvement suggestions refer to targeted learning guidance opinions generated based on the weak points in students' knowledge revealed in their assignments, combined with the teacher's grading style and expression preferences.

[0203] Using a knowledge point mastery profile as input, the system identifies areas of knowledge deficiency that students demonstrate in their assignments. Combined with suggested parameters, the system determines the direction and presentation format of the output for improvement. For example, if the teacher's preference is to "emphasize knowledge point supplementation," the generated suggestions will focus on reinforcing specific knowledge points; if the teacher's preference is "encouraging guidance," the suggestions will include positive feedback and highlight learning habits that students can improve.

[0204] Step S406: Combine personalized comments and learning improvement suggestions to obtain personalized recommendations.

[0205] By combining personalized feedback and learning improvement suggestions, a personalized recommendation is obtained.

[0206] This application provides a staged training generation method, referring to... Figure 5 The method includes:

[0207] Step S501: Summarize the scores of each assignment and personalized suggestions by student, and form a learning record in chronological order.

[0208] The student dimension refers to the dimension that retrieves homework data and personalized suggestions related to a single student, using that student as an index.

[0209] The system retrieves students' assignment scores and personalized suggestions within a specified time period from historical records, and integrates these scores and suggestions in chronological order to form a learning record.

[0210] For example, a student completes 10 math assignments in a semester. The scores and personalized suggestions for each assignment are summarized, and then these summarized scores and personalized suggestions are arranged in chronological order to generate the student's math-related learning records.

[0211] Step S502: Perform cluster analysis on the learning records based on the key information to obtain the performance sequence of the content dimensions contained in the key information.

[0212] Content dimensions refer to dimensions determined based on the specific content within key information, such as "mathematics," "proof problems," and "quadratic functions." Each dimension represents a focus area within the key information.

[0213] A performance sequence refers to a time series formed by connecting multiple assignments performed by the same student on a specific content dimension in chronological order. This performance sequence must include at least the score for each assignment and personalized suggestions.

[0214] First, learning records are categorized based on key information, such as subject type, question type, or target knowledge point. For the same student, learning records with consistent key information are grouped into the same content dimension cluster. Then, the learning records in each cluster are arranged in chronological order, thus transforming discrete records into a coherent sequence of presentations.

[0215] For example, in a math learning scenario, student A's learning record contains multiple assignments related to the key information "target knowledge point = quadratic function". First, these assignments are identified as belonging to the same content dimension cluster; then, they are arranged chronologically to generate a performance sequence: (2025-03-01, 72), (2025-03-15, 78), (2025-04-02, 65), (2025-04-20, 83). This sequence visually reflects the fluctuations and trends in the student's assignment scores on the "quadratic function" target knowledge point.

[0216] Step S503: Extract elements from personalized suggestions to obtain improvement element labels.

[0217] Element extraction refers to the process of parsing personalized suggestions into structured labels using natural language models.

[0218] Improved element labels refer to the structured results obtained after element extraction, through synonym normalization and dictionary mapping, which are used to reflect the key points of students' learning improvement.

[0219] Personalized suggestions are parsed using a generative algorithm, extracting relevant elements according to a predefined field structure (such as JSONSchema or function call format). The extracted results undergo synonym normalization and dictionary mapping to ensure that similar expressions receive consistent label representations. When the algorithm lacks confidence, the system initiates a fallback mechanism, such as re-invoking the generative algorithm, using rule matching as a fallback, or submitting for manual confirmation when necessary.

[0220] Step S504: Associate the improved element labels with the performance sequence to form a set of learning performance features.

[0221] The learning performance feature set refers to the comprehensive feature set formed by aligning the improvement element labels with the performance sequence according to the time series and then associating them.

[0222] For example, in a mathematical scenario, student A's performance sequence under the dimension of "target knowledge point = quadratic function" is: (target knowledge point = quadratic function, 2025-03-01, 72) (target knowledge point = quadratic function, 2025-04-02, 65). The improvement element labels are "weak discriminative calculation" and "insufficient logical integrity". In this step, these improvement element labels are aligned with the corresponding time nodes; for example, "weak discriminative calculation" is added to 2025-03-01, and "insufficient logical integrity" is added to 2025-04-02. The final learning performance feature set is:

[0223] (Target knowledge point = quadratic function, 2025-03-01, 72, [weak in discriminant calculation])

[0224] (Target knowledge point = quadratic function, 2025-04-02, 65, [Insufficient logical completeness])

[0225] Step S505: Based on the learning performance feature set, obtain the fluctuation index and determine whether the fluctuation index exceeds the preset fluctuation threshold and reaches the preset number of times within the preset observation window.

[0226] The volatility index refers to the fluctuation of assignment scores over time, used to characterize the stability of students' scores on a specific content dimension. It is derived by extracting assignment score sequences for a specific content dimension from a set of learning performance features, and then quantifying the fluctuations in scores over time by calculating the difference between two consecutive scores, the sliding standard deviation, or the absolute deviation of the median, thus obtaining the volatility index.

[0227] The preset fluctuation threshold is a preset constant that can be adjusted according to actual needs.

[0228] The preset observation window is a preset time constant, which can be adjusted according to actual needs.

[0229] The preset number of attempts is a fixed constant that can be adjusted according to actual needs.

[0230] Step S506: If so, the content dimension corresponding to the key information under the volatility indicator is determined as the warning object.

[0231] On the other hand, if the volatility index does not exceed the preset volatility threshold and reaches the preset number of times within the preset observation window, no action will be taken.

[0232] The warning targets refer to the content dimensions that are marked as key areas of concern due to abnormal fluctuation indicators.

[0233] After identifying the target of the warning, a set of metadata will be attached to it, including student identifier, corresponding content dimension, fluctuation index, preset fluctuation threshold, preset observation window, and cumulative number of triggers. The warning target not only indicates "which dimension requires intervention" but also includes the data basis for "why the warning was triggered".

[0234] Step S507: Generate a phased training plan for the warning target.

[0235] A phased training plan refers to a plan scheme with phased divisions and target settings generated based on the specific circumstances of the warning target.

[0236] The specific steps for generating a phased training plan for the early warning targets can be found in [reference needed]. Figure 6 The steps described in the embodiments will not be repeated here.

[0237] Step S508: Push the phased training plan to the student and teacher terminals.

[0238] The phased training plan includes a student identifier. After the phased training plan is obtained, it is pushed to the student and teacher terminals corresponding to the student identifier.

[0239] Reference Figure 6 The steps for generating a phased training plan for the early warning target include:

[0240] Step S601: Determine the warning level based on the threshold exceedance of the fluctuation index and the cumulative number of times the fluctuation index exceeds the preset fluctuation threshold within the preset observation window.

[0241] The threshold amplitude refers to the value of the fluctuation indicator that exceeds the preset fluctuation threshold. It can be obtained by subtracting the value of the fluctuation indicator from the preset fluctuation threshold.

[0242] The cumulative number of times refers to the number of times the fluctuation index exceeds the preset fluctuation threshold within the preset observation window.

[0243] The warning level refers to the severity level of the volatility indicator when it exceeds the preset volatility threshold.

[0244] The warning level is determined by setting up a table to compare the threshold exceedance with the cumulative number of occurrences. The severity levels include mild warning, moderate warning, and severe warning.

[0245] Step S602: Determine the training plan parameters based on the warning level and the personalized correction model.

[0246] Training plan parameters refer to the configuration elements when generating a phased training plan, including at least the initial values ​​for the number of priority improvement items, the intensity of practice tasks, and the recommended learning duration.

[0247] The specific steps for determining training plan parameters based on the warning level and the personalized correction model can be found in [reference needed]. Figure 7 The steps in the embodiments.

[0248] Step S603: Determine the practice task based on the content dimensions corresponding to the key information and the warning object.

[0249] The practice task refers to the training content unit generated for the warning target.

[0250] In one feasible implementation, the system determines the practice task through the following process:

[0251] The system retrieves a candidate question set corresponding to the content dimension from the question bank; filters questions based on the specific dimension of the warning object, for example, if "target knowledge point = quadratic function", only questions involving function relationships are retained; selects an appropriate number and difficulty of questions based on the priority improvement items and task intensity in the training plan parameters; if the question bank is insufficient, it calls a generative algorithm to automatically generate question content that matches the content dimension. For example: if the warning object is a certain knowledge point (such as "quadratic function"), questions involving that knowledge point are selected first; if the warning object is a certain question type feature (such as "proof question"), multi-level difficulty practice tasks under that question type are generated.

[0252] Step S604: Arrange the practice tasks in stages according to the planning parameters, and set stage goals and completion criteria.

[0253] Phased programming refers to dividing a defined practice task into several phases, with each phase corresponding to a specific training objective.

[0254] The stage goals are the expected learning level or mastery of students in the corresponding content dimensions at each stage.

[0255] Completion criteria are quantitative standards used to determine whether students have achieved their stage goals.

[0256] In one feasible implementation, if the training plan parameters are set to "Number of stages = 3, Difficulty level = Increasing stepwise", then the practice task is divided into three stages:

[0257] Phase 1: Allocate 30% of the questions to basic practice questions. The goal of this phase is to "master the core knowledge points". The criterion for completion is "accuracy rate ≥ 70%".

[0258] Phase Two: Allocate 40% of the questions to medium difficulty, with the goal of "significantly improving logical integrity"; the completion criterion is "logical integrity score ≥ 0.75".

[0259] Phase 3: Allocate 30% of the questions to high-difficulty or comprehensive questions. The goal of this phase is to "be able to independently complete multi-step reasoning or cross-knowledge point application"; the completion criterion is "overall score ≥ 80 points".

[0260] Step S605: Based on the stage goals and completion criteria, obtain the stage training plan.

[0261] The practice tasks, stage goals, and completion criteria for each stage are integrated to form an executable, phased training plan, which includes: a task list for each stage (quantity, question type, and difficulty); stage goals (knowledge points or ability levels to be mastered); completion criteria (accuracy rate, score threshold, completion rate, etc.); and the logical connection between stages (the next stage can only be entered after the criteria of the previous stage are met).

[0262] Reference Figure 7 The steps for determining training plan parameters based on the warning level and the personalized correction model include:

[0263] Step S701: Determine the values ​​of basic parameters based on the warning level. The initial values ​​of basic parameters include the number of priority improvement items, the intensity of practice tasks, and the recommended learning time.

[0264] The basic parameter values ​​refer to the initial values ​​of the training plan parameters determined based on the warning level.

[0265] The number of priority improvement items refers to the number of task items that need to be prioritized for improvement in the training plan.

[0266] The intensity of a practice task refers to the combination level of the quantity and difficulty of the training task, such as low intensity (a small number of basic questions), medium intensity (a moderate number of questions, including medium-difficulty questions), and high intensity (a large number of questions, including high-difficulty questions).

[0267] The recommended learning duration is the initial learning time required for this training, calculated based on the warning level.

[0268] In one feasible implementation, the system pre-determines the correspondence between warning levels and basic parameters:

[0269] Mild warning: Number of items to be improved = 1, intensity of practice task = low, recommended study time = 30 minutes;

[0270] Moderate warning: Priority number of items to improve = 2, exercise task intensity = medium, recommended study time = 60 minutes;

[0271] Serious warning: Prioritize improving 3-4 items, increase the intensity of practice tasks to high, and recommend a study time of 90 minutes or more.

[0272] Step S702: Adjust the values ​​of the basic parameters according to the personalized grading model to obtain parameter correction results that meet the teacher's preferences.

[0273] The parameter correction result refers to the training plan parameters obtained after adjusting the basic parameter values ​​by combining the personalized grading model with the teacher's grading preference information.

[0274] The system takes basic parameter values ​​as input and adjusts them based on teacher grading preferences within a personalized grading model. If a teacher prioritizes logic and step-by-step completeness, the proportion of reasoning or procedural questions at the same grade level will be increased, along with correspondingly longer study time. If a teacher prioritizes knowledge point coverage, the number of practice tasks will be increased or the breadth of knowledge points covered will be expanded. If a teacher emphasizes learning habits or expression style, the training cycle will be extended and the intensity of each task reduced to ensure repeated practice and detailed feedback. The final parameter adjustment results better align with the teacher's personalized teaching objectives.

[0275] In one feasible implementation, student B's basic parameters are: number of priority improvement items = 2, practice task intensity = medium, and suggested study time = 60 minutes. The teacher prefers to emphasize logical integrity, and the personalized grading model adjusts the parameters as follows: the number of priority improvement items remains unchanged; the practice task intensity is increased from "medium" to "medium+", i.e., an additional logical reasoning exercise is added; and the suggested study time is adjusted from 60 minutes to 75 minutes to meet the teacher's emphasis on reasoning training. The adjusted parameters are: number of items = 2, practice task intensity = medium+, and study time = 75 minutes.

[0276] Step S703: Determine the difficulty level, number of stages, task load of each stage, and check node time based on the parameter correction results.

[0277] Difficulty ladder refers to the progressive design of training tasks at different stages.

[0278] The number of stages refers to dividing the training task into several stages to complete it step by step.

[0279] The phase workload refers to the specific number of questions or training tasks that need to be completed in each phase.

[0280] The checkpoint time refers to the assessment time point set at the end of each stage, used to verify whether students have met the completion criteria.

[0281] The parameter correction results are obtained by correcting the basic parameter values, which include the initial values ​​of the number of improvement items, the intensity of the practice task, and the recommended learning time.

[0282] In one feasible embodiment, the difficulty level is based primarily on the revised practice task intensity: low intensity corresponds to two levels (basic to application), medium intensity corresponds to three levels (basic to advanced to comprehensive), and high intensity expands to four levels (basic to advanced to improvement to breakthrough). The number of stages is primarily referenced by the revised number of improvement items; fewer items result in two stages, a moderate number in three stages, and a large number in four stages. The task load per stage is allocated according to the difficulty level and the number of stages, with a relatively larger allocation to the basic stage and a relatively smaller allocation to the improvement or breakthrough stages. The checkpoint time is based on the revised learning time, with the total time evenly or weightedly divided according to the number of stages, and a checkpoint is established at the end of each stage.

[0283] Step S704: Obtain the training plan parameters based on the difficulty level, number of stages, stage task volume, and check node time.

[0284] The training plan parameters are obtained by combining the difficulty level, the number of stages, the amount of tasks in each stage, and the time for checking the nodes.

[0285] In one feasible implementation, student D's parameter correction results are: number of improvement items = 12, task intensity = medium, learning time = 80 minutes. The difficulty level is determined to be 3 levels (basic—advanced—comprehensive), number of stages = 3, stage task volume = [5, 4, 3], and checkpoint time = [25 min, 55 min, 80 min]. The above content is integrated into the training plan parameters: {difficulty level = 3 levels, number of stages = 3, stage task volume = [5, 4, 3], checkpoint time = [25, 55, 80] minutes}.

[0286] Based on the same inventive concept, embodiments of this application provide a personalized teacher grading system based on generative algorithms, including:

[0287] The acquisition module is used to acquire preference information, historical grading data, assignment files, and assignment data.

[0288] The memory is used to store the program for the aforementioned personalized teacher grading method based on generative algorithms;

[0289] The processor and memory can load and execute the program to implement the aforementioned personalized teacher grading method based on generative algorithms.

[0290] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. The specific working process of the system, device, and unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0291] This application provides a computer-readable storage medium storing a computer program that can be loaded by a processor and executed as a teacher-personalized grading method based on a generative algorithm.

[0292] Computer storage media include, for example, USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media that can store program code.

[0293] Based on the same inventive concept, embodiments of this application provide a smart terminal, including a memory and a processor, wherein the memory stores a computer program that can be loaded and executed by the processor for a personalized teacher grading method based on a generative algorithm.

[0294] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. The specific working process of the system, device, and unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0295] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A personalized teacher grading method based on generative algorithms, characterized in that, include: Collect teachers' marking preferences and historical marking data; Based on grading preference information and historical grading data, a personalized grading model is generated through machine learning. Receive assignment files uploaded by teachers, preprocess the assignment files, and extract key information; Based on the generative algorithm and key information, the evaluation criteria for assignment scores are obtained; Receive homework data uploaded by students; The homework data is graded based on generative algorithms, personalized grading models, and homework scoring criteria to obtain homework scores and personalized suggestions. Homework scores and personalized suggestions are pushed to both students and teachers.

2. The personalized teacher grading method based on generative algorithms according to claim 1, characterized in that, The steps for obtaining the evaluation criteria for assignment scores based on the generative algorithm and key information include: The key information is analyzed to obtain the analysis results, which include the subject category, question type characteristics, and target knowledge points of the homework questions; Based on the analysis results, a generative algorithm is invoked to generate candidate reference answers; The candidate reference answers are checked for semantic consistency, logical completeness, and knowledge point coverage to obtain a set of verified reference answers; The assignment evaluation dimensions are constructed based on the reference answer set. The assignment evaluation dimensions include semantic similarity, logical and step completeness, and mastery of knowledge points. Based on the evaluation dimensions of assignments and historical grading data, determine the weight allocation and grading thresholds for each evaluation dimension, and generate score mapping relationships and scoring rules. Based on the score mapping relationship and scoring rules, and combined with the analysis results and reference answer set, the evaluation criteria for homework scoring are obtained.

3. The personalized teacher grading method based on generative algorithms according to claim 2, characterized in that, The steps of grading homework data based on generative algorithms, personalized grading models, and homework scoring criteria to obtain homework scores include: Align the task data with the parsing results to obtain the task alignment result; Based on the assignment scoring criteria, a generative algorithm is used to compare the assignment alignment results with the reference answer set semantically, step-by-step, and knowledge point coverage to obtain the original indicators for each assignment evaluation dimension. Based on the original indicators and score mapping relationships, the initial scores of the assignment data in each assignment evaluation dimension are calculated. Based on the initial score and scoring rules, the initial comprehensive score of the assignment data is calculated; Input the initial score and initial composite score into the personalized grading model, and adjust the evaluation dimensions of the assignments. Based on the revised evaluation dimensions, a comprehensive scoring operation is performed on the assignment data to obtain the assignment score.

4. The personalized teacher grading method based on generative algorithms according to claim 3, characterized in that, The steps of grading homework data based on generative algorithms, personalized grading models, and homework scoring criteria to obtain personalized suggestions include: A difference analysis was conducted between the homework data and the reference answer set to identify the differences. Map the differences to the target knowledge points to generate a knowledge point mastery profile; Based on the knowledge point mastery profile, homework score, and revised homework evaluation dimensions, and combined with the grading style and expression preferences in the personalized grading model, the comment generation strategy and suggestion generation parameters are determined. Generate personalized comments based on the comment generation strategy; Based on the knowledge point mastery profile and suggested parameters, generate learning improvement suggestions; Personalized feedback and learning improvement suggestions are combined to obtain personalized recommendations.

5. The personalized teacher grading method based on generative algorithms according to claim 1, characterized in that, The method further includes: Summarize the scores of each assignment and personalized suggestions from each student, and compile the learning records in chronological order. Cluster analysis is performed on the learning records based on key information to obtain the performance sequence of the content dimensions contained in the key information. Extract elements from personalized suggestions to obtain improvement element tags; The improved element labels are associated with the performance sequence to form a set of learning performance features; Based on the learning performance feature set, obtain the fluctuation index and determine whether the fluctuation index exceeds the preset fluctuation threshold and reaches the preset number of times within the preset observation window. If so, the content dimension corresponding to the key information under the volatility indicator will be identified as the warning target; Generate phased training plans for the targets of the early warning; The phased training plan will be pushed to both students and teachers.

6. The personalized teacher grading method based on generative algorithms according to claim 5, characterized in that, The steps for generating a phased training plan for the early warning target include: The warning level is determined based on the magnitude of the fluctuation index exceeding the threshold and the cumulative number of times the fluctuation index exceeds the preset fluctuation threshold within the preset observation window. The training plan parameters are determined based on the warning level and the personalized correction model. The practice tasks are determined based on the content dimensions corresponding to the key information and the warning targets; The practice tasks are divided into stages according to the planning parameters, and stage goals and completion criteria are set. Based on the phase goals and completion criteria, a phased training plan is obtained.

7. The personalized teacher grading method based on generative algorithms according to claim 6, characterized in that, The steps for determining training plan parameters based on the early warning level and the personalized correction model include: The basic parameter values ​​are determined based on the warning level. The basic parameter values ​​include the initial values ​​for the number of priority improvement items, the intensity of practice tasks, and the recommended learning time. The basic parameter values ​​are adjusted based on the personalized grading model to obtain parameter correction results that conform to teachers' preferences; The difficulty level, number of stages, workload of each stage, and check node time are determined based on the parameter correction results. The training plan parameters are obtained based on the difficulty level, number of stages, stage task volume, and check node time.

8. A personalized teacher grading system based on generative algorithms, characterized in that, The system is used to execute the teacher-personalized grading method based on generative algorithms as described in any one of claims 1 to 7, including: The acquisition module is used to acquire preference information, historical grading data, assignment files, and assignment data. A memory for storing the program of the teacher-personalized grading method based on generative algorithms; The processor and memory can load and execute the program to implement the personalized teacher grading method based on generative algorithms.

9. A smart terminal, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer program is stored that can be loaded by a processor and execute the method as described in any one of claims 1 to 7.

Images