Teacher personalized correction method and system based on generative algorithm

Abstract

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

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of homework correction, in particular to a teacher personalized correction method and system based on generative algorithm. BACKGROUND

[0002] With the advancement of education informatization, homework correction gradually changes from pure manual mode to electronic tools. At the technical level, electronic document processing technology and basic data analysis technology have been widely used in the field of education, enabling teachers to more conveniently arrange and collect homework.

[0003] The existing teacher correction method is mainly manual correction or implemented on a homework correction system. The common homework correction system mainly consists of a homework management module, an answer matching module and a result output module. In the homework management module, teachers can select homework from the system resource library to publish to students or upload homework in limited formats (such as documents in specific templates). After students complete the homework and submit it, the answer matching module compares it with the preset standard answer. For objective questions, the answer is directly judged right or wrong; for subjective questions, the score is evaluated by simple methods such as keyword matching, and finally a relatively simple score and basic right or wrong judgment result is presented to the teacher and student in the result output module. For example, in Chinese composition correction, only some common wrong characters and grammatical errors can be identified to deduct points, and comprehensive evaluation cannot be made from the article idea and writing style.

[0004] For the related technology in the above, the existing system lacks personalized correction and feedback, cannot meet the needs of different teachers' teaching styles, and cannot provide students with learning suggestions that fit their individual circumstances, affecting the improvement of teaching effectiveness. SUMMARY

[0005] In order to improve the efficiency of homework correction and teaching quality and promote the personalized development of students, the present application provides a teacher personalized correction method and system based on generative algorithm.

[0006] In the first aspect, the present application provides a teacher personalized correction method and system based on generative algorithm, which adopts the following technical solution: A teacher personalized correction method based on generative algorithm, comprising: Collecting correction preference information and historical correction data of the teacher; Based on the correction preference information and the historical correction data, generating a personalized correction model through machine learning; Receiving homework files uploaded by the teacher end, preprocessing the homework files and extracting key information; According to the generative algorithm and the key information, obtaining homework score evaluation criteria; Receiving homework data uploaded by the student end; According to the generative algorithm and the personalized correction model, the homework score evaluation standard is used to correct the homework data, and homework scores and personalized suggestions are obtained; The homework scores and personalized suggestions are pushed to the student end and the teacher end.

[0007] By using the above technical solution, the personalized correction model can be generated based on the collection of teacher correction preference information and historical correction data, multi-dimensional evaluation of different subjects, question types and knowledge points can be realized, and the shortcomings of traditional correction methods such as mechanical comparison can be avoided. Combined with the generative algorithm and the key information, the homework score evaluation standard is obtained, and the personalized score and feedback are output in the student homework correction, which not only improves the correction efficiency and accuracy, but also provides targeted improvement suggestions for students and teaching decision support for teachers, improves the homework correction efficiency and teaching quality, and promotes the personalized development of students.

[0008] Optionally, the step of obtaining the homework score evaluation standard according to the generative algorithm and the key information comprises: The key information is analyzed to obtain an analysis result, and the analysis result includes the subject category, question type characteristics and target knowledge point of the homework question; Based on the analysis result, a generative algorithm is called to generate a candidate reference answer; The candidate reference answer is checked for semantic consistency, logical integrity and knowledge point coverage, and a set of checked reference answers is obtained; Based on the set of reference answers, homework evaluation dimensions are constructed, and the homework evaluation dimensions include semantic similarity, logical and step integrity and knowledge point mastery; According to the homework evaluation dimensions and the historical correction data, the weight distribution and the grading threshold of each homework evaluation dimension are determined, and the score mapping relationship and the scoring rule are generated; Based on the score mapping relationship and the scoring rule, and combined with the analysis result and the set of reference answers, the homework score evaluation standard is obtained.

[0009] By using the above technical solution, the set of checked reference answers can be generated based on the analysis of the key information, the homework evaluation dimensions covering semantic similarity, logical and step integrity and knowledge point mastery are constructed, and the weight and scoring rule are determined combined with the historical correction data, so that a more scientific and reasonable homework score evaluation standard is formed, and the accuracy and discrimination of the correction result are improved.

[0010] Optionally, the step of correcting the homework data according to the generative algorithm and the personalized correction model, the homework score evaluation standard, to obtain the homework score comprises: Align the homework data with the analysis result to obtain a homework alignment result; Based on the job score evaluation standard, the generative algorithm is called to perform semantic comparison, step comparison and knowledge point coverage comparison on the job alignment result and the reference answer set, to obtain the original index of each job evaluation dimension; According to the original index and the score mapping relationship, the initial score of the job data in each job evaluation dimension is calculated; According to the initial score and the scoring rule, the initial comprehensive score of the job data is calculated; The initial score and the initial comprehensive score are input into the personalized correction model, and the job evaluation dimension is corrected; Based on the corrected job evaluation dimension, the comprehensive scoring operation is performed on the job data to obtain the job score.

[0011] By adopting the above technical solution, on the basis of aligning the job data and the analysis result, the semantic comparison, step comparison and knowledge point coverage comparison are performed in combination with the reference answer set, so that the multi-dimensional original index is obtained, and the one-sidedness of the traditional correction method which only relies on the final answer is avoided. Further, by the score mapping relationship and the scoring rule, the original index is quantized into the initial score and the comprehensive score, and is input into the personalized correction model for correction, so that the scoring result can fully reflect the correction style and preference of the teacher. The final job score not only has objectivity and consistency, but also can reflect the differentiated evaluation standard of the teacher, thereby improving the accuracy and individualization degree of the correction result.

[0012] Optionally, the step of performing correction operation on the job data based on the generative algorithm and the personalized correction model, and the job score evaluation standard to obtain personalized suggestions comprises: Performing differential analysis on the job data and the reference answer set to obtain difference points; Mapping the difference points to target knowledge points to generate a knowledge point mastery portrait; According to the knowledge point mastery portrait, the job score, the corrected job evaluation dimension, and in combination with the correction style and expression preference in the personalized correction model, determining a comment generation strategy and a suggestion generation parameter; Generating personalized comments according to the comment generation strategy; Generating learning improvement suggestions according to the knowledge point mastery portrait and the suggestion generation parameter; Combining the personalized comments and the learning improvement suggestions to obtain personalized suggestions.

[0013] By adopting the above technical solution, the knowledge point mastery portrait can be obtained based on differential analysis, the job score and the personalized correction model are combined, the personalized comments and the targeted learning suggestions in line with the teacher's style are generated, so as to provide clear improvement direction for students, and improve the pertinence and effectiveness of teacher feedback.

[0014] Optionally, the scores of all previous assignments and the personalized suggestions are summarized according to the student dimension, and a learning record is formed in chronological order; The learning record is subjected to cluster analysis according to the key information, and a performance sequence oriented to the content dimension contained in the key information is obtained; The personalized suggestions are subjected to element extraction, and an improvement element label is obtained; The improvement element label is associated with the performance sequence to form a learning performance feature set; According to the learning performance feature set, a fluctuation index is obtained, and it is determined whether the fluctuation index exceeds a preset fluctuation threshold and reaches a preset number within a preset observation window; If yes, the content dimension corresponding to the key information is determined as a warning object; A stage training plan is generated for the warning object; The stage training plan is pushed to the student end and the teacher end.

[0015] By adopting the above technical solution, a learning record sorted by time in the student dimension can be formed, and a performance sequence of different content dimensions can be obtained based on key information clustering. The element extraction of personalized suggestions is associated with the performance sequence to form a learning performance feature set, so as to objectively calculate the fluctuation index of assignment scores and identify abnormal conditions. Further, the abnormal content dimension can be determined as a warning object, and a targeted stage training plan is generated and pushed to teachers and students to realize dynamic monitoring and precise intervention on the learning state of students.

[0016] Optionally, the step of generating a stage training plan for the warning object comprises: According to the super-threshold amplitude of the fluctuation index and the cumulative number of times that the fluctuation index exceeds the preset fluctuation threshold within the preset observation window, a warning level is determined; According to the warning level and the personalized correction model, a training plan parameter is determined; According to the key information and the content dimension corresponding to the warning object, an exercise task is determined; According to the plan parameter, the exercise task is arranged in stages, and stage goals and completion criteria are set; Based on the stage goals and completion criteria, a stage training plan is obtained.

[0017] By adopting the technical solution, the warning level can be determined based on the threshold amplitude and the cumulative number of the fluctuation index, so as to grade the severity of the student learning anomaly. Further, the training plan parameters are adjusted in combination with the personalized correction model, so that the training task meets the warning level and is consistent with the correction preference of the teacher. Then, the personalized practice task is determined according to the content dimension corresponding to the key information and the warning object, and is arranged in stages according to the training plan parameters, so as to clearly define the stage target and completion criterion. The stage training plan generated in this way can accurately match the weak links of the student and provide a gradual improvement path.

[0018] Optionally, the step of determining the training plan parameters according to the warning level and the personalized correction model comprises: determining the basic parameter value according to the warning level, the basic parameter value comprising the initial value of the number of priority improvement items, the intensity of the practice task, and the recommended learning duration; adjusting the basic parameter value according to the personalized correction model to obtain a parameter correction result consistent with the teacher's preference; determining the difficulty ladder, the number of stages, the stage task amount, and the inspection node time according to the parameter correction result; obtaining the training plan parameters according to the difficulty ladder, the number of stages, the stage task amount, and the inspection node time.

[0019] By adopting the technical solution, the basic parameter value of the training task can be determined based on the warning level, and the personalized correction model is combined for correction, so as to better meet the teacher's correction preference. The difficulty ladder, the number of stages, the task amount, and the inspection node time are further determined based on the corrected parameter correction result, and finally more accurate and personalized training plan parameters are obtained, so as to realize targeted learning intervention.

[0020] In a second aspect, the present application provides a teacher personalized correction system based on a generative algorithm, which adopts the following technical solution: A teacher personalized correction system based on a generative algorithm comprises: an acquisition module for acquiring preference information, historical correction data, assignment files, and assignment data; a storage for storing the program of the teacher personalized correction method based on a generative algorithm; a processor, the program in the storage can be loaded and executed by the processor and implement the teacher personalized correction method based on a generative algorithm.

[0021] By adopting the technical scheme, the personalized correction model can be generated on the basis of collecting teacher correction preference information and historical correction data, multi-dimensional evaluation of different subjects, question types and knowledge points can be realized, and the deficiency of mechanical comparison in the traditional correction mode can be avoided. The scoring evaluation criteria are obtained in combination with the generative algorithm and the key information, and the personalized score and feedback are output in the student homework correction, which improves the correction efficiency and accuracy, provides targeted improvement suggestions for students, provides teaching decision support for teachers, improves the homework correction efficiency and teaching quality, and promotes the personalized development of students.

[0022] In a third aspect, the present application provides an intelligent terminal, which adopts the technical scheme as follows: An intelligent terminal, comprising a memory and a processor, the memory storing a computer program capable of being loaded by the processor and executing the method of any one of the above.

[0023] In a fourth aspect, the present application provides a computer storage medium capable of storing a corresponding program, having the characteristics of facilitating the improvement of homework correction efficiency and teaching quality, and promoting the personalized development of students, and adopting the technical scheme as follows: A computer readable storage medium storing a computer program capable of being loaded by a processor and executing the teacher personalized correction method based on the generative algorithm.

[0024] In summary, the present application has at least one of the following beneficial technical effects: The personalized correction model can be generated on the basis of collecting teacher correction preference information and historical correction data, multi-dimensional evaluation of different subjects, question types and knowledge points can be realized, and the deficiency of mechanical comparison in the traditional correction mode can be avoided. The scoring evaluation criteria are obtained in combination with the generative algorithm and the key information, and the personalized score and feedback are output in the student homework correction, which improves the correction efficiency and accuracy, provides targeted improvement suggestions for students, provides teaching decision support for teachers, improves the homework correction efficiency and teaching quality, and promotes the personalized development of students. The verified reference answer set can be generated on the basis of analyzing the key information, the homework evaluation dimensions covering semantic similarity, logical and step integrity and knowledge point mastery are constructed, the weight and scoring rules are determined in combination with the historical correction data, so that more scientific and reasonable homework scoring evaluation criteria are formed, and the accuracy and discrimination of the correction result are improved. The knowledge point mastery portrait can be obtained based on the differential analysis, the homework score and the personalized correction model are combined, the personalized comments and targeted learning suggestions conforming to the teacher style are generated, so that the students are provided with clear improvement direction, and the targetedness and effectiveness of the teacher feedback are improved. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1is a flowchart of a teacher personalized correction method based on a generative algorithm in an embodiment of the present application.

[0026] Figure 2 is a flowchart of a step of obtaining a homework score evaluation standard according to a generative algorithm and key information in an embodiment of the present application.

[0027] Figure 3 is a flowchart of a step of obtaining a homework score in an embodiment of the present application.

[0028] Figure 4 is a flowchart of a step of obtaining a personalized suggestion in an embodiment of the present application.

[0029] Figure 5 is a flowchart of a step of obtaining a personalized suggestion in an embodiment of the present application.

[0030] Figure 6 is a flowchart of a step of obtaining a personalized suggestion in an embodiment of the present application.

[0031] Figure 7 is a flowchart of a step of obtaining a personalized suggestion in an embodiment of the present application. DETAILED DESCRIPTION

[0032] In order to make the objectives, technical solutions and advantages of the present application clearer, the following further describes the present application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely used to explain the present application and should not be used to limit the present application. Figure 1 - the accompanying drawings Figure 7 and embodiments. It should be understood that the specific embodiments described herein are merely used to explain the present application and should not be used to limit the present application.

[0033] An embodiment of the present application discloses a teacher personalized correction method based on a generative algorithm. Referring to Figure 1 , the teacher personalized correction method based on a generative algorithm comprises: Step S101: Collecting correction preference information and historical correction data of a teacher.

[0034] The correction preference information refers to the individualized tendency parameters shown by the teacher when correcting homework, such as emphasizing ideas, style or structure in Chinese composition, or emphasizing the integrity of proof or the correctness of answer in mathematics problems.

[0035] The historical correction data refers to the homework correction records completed by the teacher in the past, including student homework original text, teacher scoring results, annotation traces, comment texts and corresponding timestamp information, which are used to reflect the actual correction habits of the teacher.

[0036] The correction preference information can be obtained through a questionnaire survey, and the historical correction data can be obtained by calling a homework correction database bound to a teacher account or a local storage file.

[0037] For example, when a teacher logs in to the teacher end for the first time, the teacher will complete the initial setting of the correction preference information through the interactive interface, such as selecting the correction style (encouraging or strict), focusing on the dimension (idea, literary style, structure, logical integrity, etc.), and uploading the setting result to the server end storage. At the same time, the system will call the historical correction database bound to the teacher account to retrieve and analyze the past corrected homework data, extract the homework original text, scoring record, comment and annotation information and convert them into structured data to form an initial data set of the teacher's personalized correction habit.

[0038] Step S102: Based on the correction preference information and the historical correction data, a personalized correction model is generated through machine learning.

[0039] The personalized correction model refers to a model trained by a machine learning algorithm, which simulates the correction style and habit of a specific teacher. The model takes the teacher's correction preference information as a guide parameter and takes the historical correction data as a training sample, so as to learn the teacher's specific scoring tendency and comment style under different subjects, different types of questions and different scoring dimensions.

[0040] The machine learning module is called, the correction preference information is taken as an input constraint condition, the historical correction data is taken as a training sample set, and the feature extraction and model training operations are performed. Through iterative training, the teacher's scoring distribution and comment generation rule in different dimensions are gradually fitted, and the corresponding personalized correction model is output for subsequent scoring and feedback in the homework correction process.

[0041] In a feasible implementation, first, the historical correction data is processed to convert the scores, annotations and text comments given by the teacher into numerical or label features; then a natural language processing model based on deep learning (such as a Transformer architecture) is used to represent the comments in a semantic vector; at the same time, the preference parameters set by the teacher are introduced as weight factors into the loss function to guide the model to better fit the teacher's preferences during training. After training is completed, the model can automatically adjust the proportion of different evaluation dimensions according to the teacher's habit, for example, more emphasis on "logical integrity" or "creative expression".

[0042] Step S103: Receive the homework file uploaded by the teacher end, preprocess the homework file and extract key information.

[0043] The homework file refers to an electronic file containing student homework content uploaded by the teacher through the teacher end. The file format can include Word, PDF, picture, etc.

[0044] Key information refers to the job title information extracted by the system from the pre-processed job file, including subject category, question type characteristics and target knowledge points.

[0045] Through the upload interface of the teacher end, the job file is received and format analysis and standardization conversion are performed. For text files, text extraction and paragraph division are performed; for picture files, OCR recognition and text restoration are performed. Then, the natural language processing module is called to perform semantic analysis on the job title, automatically identify the subject category, question type characteristics and target knowledge points of the job, and generate key information for subsequent steps.

[0046] For example, in the Chinese composition scene, the teacher uploads a PDF format composition title file, which is converted to text through preprocessing, and extracts "subject category = Chinese", "question type characteristics = composition" and "target knowledge point = writing idea, language expression"; in the mathematical proof question scene, the teacher uploads a Word file, and the system parses the question "prove that ∠A = ∠B in triangle ABC", and identifies "subject category = mathematics", "question type characteristics = proof question" and "target knowledge point = geometric proof, properties of isosceles triangle".

[0047] Step S104: Obtain the job score evaluation standard according to the generative algorithm and the key information.

[0048] The generative algorithm is a class of algorithms that can automatically generate new content based on existing data and model parameters. It is a deep learning model based on the Transformer architecture (such as pre-trained language models), and its core mechanism is the self-attention mechanism (Self-Attention), which can model the semantic relationship between different parts of the input text, and generate content that meets the logical and semantic requirements. In this application, the generative algorithm is mainly used to: generate candidate reference answers according to the key information of the job title (subject category, question type characteristics, target knowledge points); generate job evaluation dimensions and scoring rules; generate comments and learning suggestions in the personalized suggestion feedback link.

[0049] The job score evaluation standard refers to the scoring rule system followed in the correction process. This standard is generated by the generative algorithm combined with the key information, and covers the weight distribution, grading threshold and score mapping relationship of the job evaluation dimensions such as semantic similarity, logical and step integrity, and knowledge point mastery.

[0050] The key information is analyzed to determine the subject category, question type characteristics and target knowledge points of the assignment. By calling the generative algorithm, candidate reference answers matching the key information are generated, and these reference answers are checked for semantic consistency, logical integrity and knowledge point coverage to obtain a set of checked reference answers. Based on the reference answer set, assignment evaluation dimensions are constructed, and the weights and threshold values of each evaluation dimension are dynamically adjusted in combination with the teacher's historical grading data to form the assignment score evaluation standard.

[0051] The specific steps of obtaining the assignment score evaluation standard according to the generative algorithm and the key information can refer to Figure 2 Embodiments.

[0052] Step S105: receiving assignment data uploaded by the student end.

[0053] The assignment data refers to the assignment file or answer data completed by the student, which can include text, pictures and other file formats, and has student identity identification and assignment submission time information.

[0054] In a feasible implementation, the student end sets an upload portal, and the student can choose to upload the assignment by selecting a local file or taking a photo. When the system receives the assignment data, it automatically adds the student ID, class ID and submission timestamp to ensure data traceability. For uploaded text files, they are directly stored in the database; for uploaded picture files, an OCR recognition module can be called to convert them into a processable text format; for audio and video files (such as oral test), a speech recognition or video transcription module is called to generate a script, which is saved together with the original file. After all the assignment data is stored, the system generates a unique data index number to ensure accurate positioning of the corresponding assignment data in the subsequent grading process.

[0055] For example, the student uploads a photo of a math assignment, and the system receives it and extracts the question and answer content using the OCR recognition module, and stores the "student ID = 2023001" "class ID = Class 3, Grade 1" "submission time = September 10, 2025 20:15" as metadata and assignment data.

[0056] Step S106: grading the assignment data according to the generative algorithm and the personalized grading model, and the assignment score evaluation standard to obtain the assignment score and personalized suggestions.

[0057] The assignment score is a comprehensive score obtained by quantitatively calculating the student's assignment data in each assignment evaluation dimension based on the assignment evaluation standard, which reflects the final score that meets the teacher's grading preferences.

[0058] The personalized suggestion is an improvement suggestion for the learning condition of the student, which is generated based on the operation score, the knowledge point mastering image and the personalized correction model, and includes personalized comments and learning improvement suggestions.

[0059] The specific steps of obtaining the operation score can refer to the steps in the embodiment, which will not be described here again. Figure 3 The specific steps of obtaining the personalized suggestion can refer to the steps in the embodiment, which will not be described here again. Figure 4 The specific steps of obtaining the personalized suggestion can refer to the steps in the embodiment, which will not be described here again.

[0060] Step S107: Push the operation score and the personalized suggestion to the student end and the teacher end.

[0061] After completing the correction operation, the operation score and the personalized suggestion are bound with the student identity information and the data index number of the operation data to generate corresponding data records, and are synchronized to the student end and the teacher end through a message pushing mechanism. The student end mainly displays the personal operation score, the personalized comments and the learning improvement suggestions; the teacher end not only displays the data records of the operation of a single student, but also provides statistical summary and visual analysis of the overall situation of the class.

[0062] Refer to Figure 2 , according to the generative algorithm and the key information, the steps of obtaining the operation score evaluation standard include: Step S201: Analyzing the key information to obtain an analysis result, the analysis result including a subject category, a question type feature and a target knowledge point of the operation question.

[0063] The analysis result refers to the analysis result of the operation question obtained after the semantic analysis and the structural processing of the key information, mainly including the subject category, the question type feature and the target knowledge point. The subject category is, for example, Chinese, mathematics, English and physics. The question type feature is, for example, a selection question, a fill-in-the-blank question, an answer question and an essay. The target knowledge point refers to the specific teaching knowledge point corresponding to the question, for example, "linear function", "Newton's second law" and "writing intention".

[0064] In a feasible implementation, when the key information is analyzed, the subject recognition is first performed on the question text by using a classification model, for example, the judgment of "the question belongs to mathematics" is performed by using a convolutional neural network or a Transformer classifier. Then, the question type feature is extracted by using a rule matching and semantic discrimination algorithm, for example, "if the stem contains'select one correct answer', the question is determined as a selection question"; "if the stem contains 'write an essay', the question is determined as an essay". Finally, the question text is input into a knowledge point mapping model, the model is based on a pre-trained semantic matching network and a knowledge graph, the stem key word is aligned with a pre-defined knowledge point library, and thus the target knowledge point is obtained. For example, when the question contains "y=kx+b", the target knowledge point is identified as "linear function".

[0065] Step S202: Based on the analysis result, call the generative algorithm to generate candidate reference answers.

[0066] The candidate reference answers refer to the multiple alternative answer sets generated by the generative algorithm based on the analysis result for comparison and use during grading. These candidate reference answers not only contain standard solutions, but also cover diversified reasonable answer paths and expression methods.

[0067] The analysis result is taken as an input condition to call the generative algorithm, with the subject category and target knowledge point of the question as core constraints to generate multiple candidate reference answers. In a feasible implementation, the system selects the corresponding generative algorithm model according to the subject category in the analysis result: If the subject category is mathematics, call the reasoning-based generative model based on the Transformer architecture to generate standard solutions and possible multi-step derivation paths; If the subject category is Chinese, call the language model based on natural language generation to generate reference answers for essays in different styles and ideas. In the generation process, the target knowledge point is used as a generation prompt, the stem is input into the model, and multiple candidate reference answers with independent content are obtained.

[0068] For example, in the mathematical proof question "prove that the base angles of an isosceles triangle are equal", the system calls the generative reasoning model to generate two candidate reference answers based on the analysis result (subject category = mathematics, question type feature = proof question, target knowledge point = properties of isosceles triangle): Reference answer A: Prove ∠A = ∠B using congruent triangles; Reference answer B: Prove ∠A = ∠B using angle bisectors and symmetry.

[0069] Step S203: Check the semantic consistency, logical integrity, and knowledge point coverage of the candidate reference answers to obtain the set of verified reference answers.

[0070] Semantic consistency is used to determine whether the candidate reference answers are semantically matched with the requirements of the homework question, avoiding answering irrelevant questions.

[0071] Logical integrity is used to determine whether the candidate reference answers are complete and reasonable in terms of reasoning chain, argument process, or expression structure.

[0072] Knowledge point coverage is used to determine whether the candidate reference answers cover the target knowledge points extracted in the analysis result.

[0073] The qualified answer set that remains after the above three checks is the reference answer set, which serves as the basis for subsequent homework grading.

[0074] In a feasible implementation, the system performs three types of checks on the candidate reference answers: Semantic consistency check: Use semantic embedding models (such as BERT) to calculate the semantic similarity between the candidate reference answer and the stem. If the similarity is lower than the preset threshold (such as 0.7), it is determined that the semantics deviate, and the answer is excluded.

[0075] Logical integrity check: For mathematical answer questions, call step analysis algorithm to detect whether the reasoning covers the necessary steps; for Chinese composition, analyze whether the paragraph structure contains the idea, argument and conclusion part.

[0076] Knowledge point coverage test: Use knowledge graph to compare the matching of candidate reference answer and target knowledge point. If the main knowledge point is missing, reduce the weight or exclude it.

[0077] Finally, the candidate answers that meet the three tests are combined into the verified reference answer set.

[0078] Step S204: Based on the reference answer set, construct the homework evaluation dimension, which includes semantic similarity, logical and step integrity, and knowledge point mastery.

[0079] Semantic similarity is used to judge the closeness of the semantic level between the answer of the homework data and the reference answer.

[0080] Logical and step integrity is used to judge the completeness of the reasoning chain or expression structure of the answer of the homework data.

[0081] Knowledge point mastery is used to judge the coverage of the target knowledge point in the analysis result of the answer of the homework data.

[0082] First, the semantic analysis of the reference answer set is carried out, and the semantic vector is obtained, which is used to construct the semantic similarity evaluation dimension; then the content in the reference answer set is structured and decomposed in order, forming a hierarchical structure, which is used as the logical and step integrity evaluation dimension; the knowledge points involved in the reference answer set are compared with the target knowledge points in the analysis result, thus forming the knowledge point mastery evaluation dimension. The three dimensions of semantic similarity, logical and step integrity and knowledge point mastery are combined to form the homework evaluation dimension.

[0083] Step S205: According to the homework evaluation dimension and the historical correction data, determine the weight distribution and grading threshold of each homework evaluation dimension, and generate the score mapping relationship and scoring rules.

[0084] Weight distribution refers to setting the relative importance ratio between different homework evaluation dimensions, which is used to reflect the contribution of each dimension to the total score.

[0085] The grading threshold refers to the numerical boundary for dividing the performance of the work into different level intervals (e.g. excellent, qualified, unqualified), which is used to determine the score level of the work data in each dimension.

[0086] The score mapping relationship refers to the corresponding relationship between the original index value of each work evaluation dimension and the specific score.

[0087] The scoring rule is a quantitative scoring method composed of weight allocation and grading threshold.

[0088] In a feasible implementation, first, the average distribution and variance of the teacher in different dimensions are calculated through statistical analysis of historical correction data, which is used to reflect the scoring tendency of the teacher. For example, when the score distribution of the teacher in the logical integrity dimension is significantly different, it indicates that the dimension has high discrimination, and therefore a greater weight is given. Second, according to the distribution rule of the score in the historical correction record, a reasonable grading threshold is set, for example, "higher than 0.8 is excellent, 0.6-0.8 is qualified, and lower than 0.6 is unqualified". Then, the original index of the work evaluation dimension is corresponded with the grading threshold to generate the score mapping relationship, for example, semantic similarity = 0.75 is mapped to 8 points. Finally, the scores of each work evaluation dimension are weighted according to the weight to form a complete scoring rule.

[0089] For example, in a group of historical correction data, the teacher pays more attention to the logical integrity dimension. According to the historical correction data, the weight allocation is determined as follows: semantic similarity weight = 0.3, logical integrity weight = 0.5, and knowledge point mastery weight = 0.2. The grading threshold is set as follows: semantic similarity ≥ 0.8 is judged as excellent, 0.6-0.8 is judged as qualified, and lower than 0.6 is judged as unqualified. According to the rule, the semantic similarity of a student is 0.75 (8 points), the logical integrity is 0.85 (9 points), and the knowledge point mastery is 0.65 (7 points), and the comprehensive score is 0.3 × 8 + 0.5 × 9 + 0.2 × 7 = 8.3 points.

[0090] Step S206: Based on the score mapping relationship and the scoring rule, and combined with the analysis result and the reference answer set, the work scoring evaluation standard is obtained.

[0091] First, the subject category, question type feature, and target knowledge point extracted from the analysis result are corresponded with the answer content in the reference answer set to ensure that the scoring can cover the key examination points. Then, the score mapping relationship and the scoring rule are used to convert each work evaluation dimension into a specific score calculation method. Finally, the quantitative results of each work evaluation dimension are integrated according to the weight in the scoring rule to generate a complete work scoring evaluation standard.

[0092] Exemplary, taking a question involving the "law of conservation of energy" as an example, the analysis result shows that the target knowledge point of the question is the conservation of energy and its application. The reference answer set provides two verified standard answers: one is mainly based on theoretical derivation, and the other is mainly based on experimental phenomena. In the homework evaluation dimension, the logic and step integrity is given a weight of 0.4, the knowledge point mastery is given a weight of 0.35, and the semantic similarity is given a weight of 0.25. When the student's answer is basically complete in logical derivation, but slightly incomplete in knowledge point coverage, the corresponding quantitative score will be given based on the score mapping relationship, and the final homework score evaluation standard will be formed to ensure that the scoring result not only meets the teacher's grading preferences, but also remains objective and consistent.

[0093] Reference Figure 3 According to the generative algorithm and the personalized grading model, the homework score evaluation standard is used to grade the homework data, and the homework score is obtained. Step S301: Align the homework data with the analysis result to obtain the homework alignment result.

[0094] The homework alignment result refers to the matching result of the student's submitted homework data and the analysis result one by one, which clearly shows the corresponding relationship of the student's answer under the subject category, the question type characteristic and the target knowledge point, including matching based on question type characteristics, mapping based on target knowledge points, and constraints based on subject categories.

[0095] Among them, the matching based on question type characteristics: for example, multiple choice questions, fill-in-the-blank questions, and answer questions, different question types have different data formats, and the consistency of the form of the student's homework data and the question type characteristics needs to be matched.

[0096] Mapping based on target knowledge points: the semantic content, key expression or calculation steps in the student's homework data are matched with the target knowledge points in the analysis result, and the knowledge point range involved in the student's answer is confirmed.

[0097] The constraints based on the subject category: further check the logic and expression form of the answer under the same subject category, so that the alignment result is more in line with the requirements of a specific subject.

[0098] Exemplary, for example, the question analysis result is: subject category "mathematics", question type characteristic "calculation question", target knowledge point "solving linear equation"; the student's homework data is "x=2".

[0099] First, confirm that "x=2" belongs to the question type characteristic of "calculation question"; Then compare "x=2" with the target knowledge point "solving linear equation" to confirm that it covers the knowledge point; Under the constraint of the subject category being mathematics, it is confirmed that the answer conforms to the logic expression of the subject.

[0100] The final generated homework alignment result is: "Problem - Solving Linear Equations - Student Answer x = 2".

[0101] Step S302: Based on the homework scoring evaluation criteria, call the generative algorithm to perform semantic comparison, step comparison, and knowledge point coverage comparison between the homework alignment result and the reference answer set, to obtain the original indicators of each homework evaluation dimension.

[0102] The original indicator refers to the quantitative value obtained under the homework evaluation dimension (such as semantic similarity, logical and step integrity, and knowledge point mastery), which is used as input for the initial score calculation in the subsequent steps.

[0103] Semantic comparison refers to using the generative algorithm to perform semantic matching between the answer in the homework alignment result and the reference answer, calculating the similarity between the two in terms of content expression, and obtaining the semantic similarity indicator.

[0104] Step comparison refers to comparing the consistency of the reasoning process of the answer in the homework alignment result and the logical chain of the reference answer, to obtain the logical and step integrity indicator.

[0105] Knowledge point coverage comparison compares whether the answer in the homework alignment result covers the target knowledge points involved in the reference answer set.

[0106] In one possible implementation, first, the text features of the answer in the homework alignment result are extracted and compared with the reference answer set for semantic comparison, using cosine similarity, edit distance, or context similarity calculation methods based on the generative algorithm to obtain the semantic similarity indicator. Subsequently, the sequence of problem-solving steps of the reference answer is compared with the reasoning process of the student answer, and the number of missing steps and error steps is counted to obtain the logical and step integrity indicator. Further, the knowledge points involved in the student answer are corresponded one by one with the target knowledge points extracted from the analysis result, and the proportions of covered, missing, and incorrectly covered knowledge points are counted to obtain the knowledge point mastery indicator. Finally, according to the homework scoring evaluation criteria, the above-mentioned semantic similarity indicator, logical and step integrity indicator, and knowledge point mastery indicator are compared with the pre-set grading threshold to form the original indicators of each homework evaluation dimension.

[0107] For example, the reference answer is "first subtract 3 from both sides of the equation, then divide by 2, and solve for x = 4"; the student answer is "solve for x = 4". In the semantic similarity comparison link, the student answer and the reference answer are consistent in the final result, the similarity is 0.95; in the logic and step integrity comparison link, the student answer lacks the derivation process, the logic integrity score is 0.5; in the knowledge point mastery situation, the student answer covers the final solution knowledge point, but does not cover the "subtraction operation" and "division operation" in the problem solving steps, the coverage rate is 0.7. The final obtained original indicators are: semantic similarity = 0.95, logic and step integrity = 0.5, knowledge point mastery situation = 0.7.

[0108] Step S303: According to the original indicators, the score mapping relationship is calculated to obtain the initial score of the homework data in each homework evaluation dimension.

[0109] First, the semantic similarity, logic and step integrity, knowledge point mastery and other original indicators are respectively substituted into the score mapping relationship. According to the grading threshold, the original indicators are converted into a standardized score interval, and the initial score of the homework data in the homework evaluation dimension is calculated.

[0110] The answer to the homework data of a student is highly consistent with the reference answer in semantic expression, the semantic similarity original indicator is 0.88; one step of logic step is omitted, the original indicator is 0.6; two of the three knowledge points in the reference answer are covered, the original indicator is 0.67. After score mapping and scoring rule conversion, the semantic similarity dimension score is 92 points, the logic and step integrity score is 70 points, the knowledge point mastery score is 75 points, forming the initial score of the student in each dimension.

[0111] Step S304: According to the initial score and the scoring rule, the initial comprehensive score of the homework data is calculated.

[0112] The initial comprehensive score refers to the total score obtained by weighting and integrating according to the scoring rule based on the initial score of each homework evaluation dimension.

[0113] For example, in the scoring rule, the weight of the semantic similarity dimension is 0.4, the weight of the logic integrity dimension is 0.35, and the weight of the knowledge point mastery is 0.25. If the student's initial scores in semantic similarity, logic integrity and knowledge point mastery are 92, 70 and 75 respectively, then the initial comprehensive score can be calculated according to the following formula: Initial comprehensive score = 80x0.4 + 70x0.35 + 90x0.25 = 83.80.

[0114] Step S305: Input the initial score and the initial comprehensive score into the individualized correction model, and modify the homework evaluation dimension.

[0115] After obtaining the initial scores and the initial comprehensive score, the results are input into the personalized grading model. The personalized grading model first analyzes the weight distribution of the initial comprehensive score, identifying the influence of each assignment evaluation dimension in the formation of the comprehensive score. On this basis, the personalized grading model corrects the weight distribution of the assignment evaluation dimensions according to the teacher's grading preferences. For example, if the teacher prefers to focus on logical processes, the model will increase the weight of the logical and step integrity dimension; if the teacher places more emphasis on knowledge point coverage, the model will increase the weight of the knowledge point mastery dimension.

[0116] The student's initial scores in the three dimensions are: semantic similarity 92 points, logical integrity 70 points, and knowledge point mastery 75 points. The initial comprehensive score is calculated according to the original weight distribution (semantic similarity 0.4, logical integrity 0.35, knowledge point mastery 0.25) as 83.8 points. When the initial scores and the comprehensive score are input into the personalized grading model, the model first analyzes the weight distribution of the three dimensions; then, combined with the teacher's grading preferences, the weight of the logical and step integrity dimension is adjusted from 0.35 to 0.45, the weight of the semantic similarity dimension is adjusted from 0.4 to 0.3, and the weight of the knowledge point mastery dimension remains 0.25 unchanged.

[0117] Step S306: Based on the corrected assignment evaluation dimensions, perform a comprehensive scoring operation on the assignment data to obtain the assignment score.

[0118] The comprehensive scoring operation refers to the process of scoring according to the weight distribution and scoring rules of the corrected assignment evaluation dimensions.

[0119] For example, a student's initial scores in semantic similarity, logical integrity, and knowledge point mastery are 92, 70, and 75 respectively. After step S304, the initial comprehensive score is 83.8. In step S305, the personalized grading model increases the weight of the logical integrity dimension from 0.35 to 0.45, reduces the weight of the semantic similarity dimension from 0.4 to 0.3, and keeps the weight of the knowledge point dimension unchanged at 0.25. In this step, the comprehensive score is recalculated according to the corrected weight distribution: Assignment score = 92x0.3 + 70x0.45 + 75x0.25 = 81.5.

[0120] Reference Figure 4 According to the generative algorithm and the personalized grading model, the assignment scoring criteria, the steps of the personalized grading operation include: Step S401: Perform differential analysis on the assignment data and the reference answer set to obtain the difference points.

[0121] The differentiated analysis refers to the process of generating the semantic and step-level differentiated analysis of the question answers of the student's homework data and the reference answers by the generative algorithm.

[0122] The differentiated points refer to the differences between the question answers of the student's homework data and the reference answer set in the semantic expression and problem-solving step dimensions.

[0123] Step S402: Map the differentiated points to the target knowledge points to generate the knowledge point mastery image.

[0124] The knowledge point mastery image refers to the mastery degree of the student at each target knowledge point. By mapping the differentiated points to the corresponding knowledge points, a structured mastery state label is formed.

[0125] Map the differentiated points to the target knowledge points. If a knowledge point corresponds to more differentiated points, mark it as "not mastered" or "insufficiently mastered" in the knowledge point mastery image; if there are no differentiated points, mark it as "mastered" in the knowledge point mastery image. Through the mapping of all differentiated points, the knowledge point mastery image covering the target knowledge points can be obtained.

[0126] Step S403: According to the knowledge point mastery image, the homework score, the revised homework evaluation dimension, and combined with the correction style and expression preference in the individualized correction model, determine the comment generation strategy and suggestion generation parameters.

[0127] The correction style and expression preference refer to the personal habits of the teacher stored in the individualized correction model, such as "strict correction", "encouraging feedback", "concise expression", or "detailed analysis".

[0128] The comment generation strategy refers to the overall logic used when generating individualized comments, including emphasizing strengths, pointing out weaknesses, and making improvement suggestions, etc.

[0129] The suggestion generation parameters refer to the specific parameter set required when generating learning improvement suggestions, such as which knowledge points to suggest for improvement, the level of detail of the suggestions, and the expression tone, etc.

[0130] First, according to the homework score and the revised homework evaluation dimension, identify the strengths and weaknesses of the student in the overall performance and specific dimensions; then, combine these results with the knowledge point mastery image to locate the weak links in the student's knowledge mastery. Next, call the individualized correction model and determine the focus when generating comments according to the teacher's correction style (such as "emphasis on logic" and "encouraging expression") and expression preference (such as "concise and intuitive" or "detailed and specific"). For example, for a teacher who emphasizes logic, the comment generation strategy will tend to provide more detailed explanations of the completeness of the problem-solving steps, and highlight the content related to logic training in the suggestion generation parameters.

[0131] Exemplarily, in the analysis of a student's math homework, it is concluded that the student's homework score is 78, with the logic and step integrity dimension being significantly low (65) and the semantic similarity dimension being relatively high (90); at the same time, the knowledge point mastery image shows that the mastery of "discriminant calculation" is insufficient. If the teacher's correction style is biased towards "strict logic", and the expression preference is "objective and detailed", the following strategy will be generated: comment generation strategy: focus on feedback on the student's defects in logic steps, and make strict comments combined with the score; suggestion generation parameter: generate supplementary exercises for the "discriminant calculation" knowledge point, and provide detailed reasoning process prompts.

[0132] Step S404: generating a personalized comment according to the comment generation strategy.

[0133] The personalized comment refers to the customized textual feedback generated in combination with the student's homework score, knowledge point mastery, and the teacher's correction style and expression preference.

[0134] By calling the generative algorithm, the comment generation strategy is used as an input control condition to generate the corresponding comment content. For example, when the teacher's style emphasizes logic, the advantages and disadvantages of the problem solving steps will be pointed out in more detail in the comment; when the teacher is biased towards encouragement, more positive language will be used.

[0135] Exemplarily, in the math homework scenario, a student's homework score is 78, the logic and step integrity dimension is low, and the "discriminant calculation" knowledge point mastery is insufficient. The comment generation strategy is "strict logic, objective and detailed". Therefore, the system calls the generative algorithm to generate the following comment: "This homework is generally well done, and most of the problem intentions can be correctly understood. However, there are errors in discriminant calculation in the problem solving process, which leads to incomplete logic reasoning chain. It is suggested to focus on training the application of discriminant formula and pay attention to the integrity of each step of reasoning when reviewing." Step S405: generating learning improvement suggestions according to the knowledge point mastery image and suggestion generation parameters.

[0136] The learning improvement suggestion refers to the targeted learning guidance generated in combination with the teacher's correction style and expression preference according to the weak links in the knowledge points exposed by the student in the homework.

[0137] Taking the knowledge point mastery image as the input basis, the knowledge point defect area exhibited by the student in the homework is identified; then, combined with the suggestion generation parameter, the improvement direction and presentation form of the output are determined. For example, if the teacher's preference is "emphasis on knowledge point supplementation", the generated suggestion will focus more on strengthening specific knowledge points; if the teacher's preference is "encouraging guidance", the suggestion will contain positive feedback and point out the student's improvable learning habits.

[0138] Step S406: combine the personalized comments and learning improvement suggestions to obtain personalized suggestions.

[0139] In obtaining the personalized comments and learning improvement suggestions, the two are combined to obtain personalized suggestions.

[0140] Embodiments of the present application provide a phased training generation method, referring to Figure 5 The method comprises: Step S501: aggregate the scores of previous assignments and personalized suggestions according to student dimensions, and form learning records in chronological order.

[0141] The student dimension refers to a dimension that indexes a single student and obtains assignment data and personalized suggestions related to the student.

[0142] The assignment scores and personalized suggestions of the student within a specified time are obtained from the historical records, and the assignment scores and personalized suggestions of all previous assignments are integrated in chronological order to form learning records.

[0143] For example, a student has completed 10 math assignments in a semester, and the scores of each assignment and personalized suggestions are aggregated, and then the aggregated assignment scores and personalized suggestions are arranged in chronological order to generate the student's math-related learning records.

[0144] Step S502: cluster analysis of learning records according to key information to obtain performance sequences oriented to content dimensions contained in the key information.

[0145] The content dimension refers to a dimension determined according to specific content in the key information, such as "mathematics", "proof questions", "quadratic functions", etc., and each dimension represents a focus of the key information.

[0146] The performance sequence refers to the performance of a student on multiple assignments in a content dimension, connected in chronological order to form a time sequence. The performance sequence at least includes the score of each assignment and the personalized suggestion.

[0147] First, the learning records are classified according to subject categories, question type characteristics or target knowledge points according to the key information as the clustering basis. For the same student, learning records with consistent key information are clustered into the same content dimension cluster. Then, the learning records in each cluster are arranged in chronological order, so that discrete records are converted into a coherent performance sequence.

[0148] Exemplarily, in a mathematics learning scenario, student A's learning record has multiple homework data related to the "target knowledge point = quadratic function" contained in the key information. First, it is identified that these homeworks belong to the same content dimension cluster; then, they are arranged in chronological order to generate a performance sequence: (2025-03-01, 72), (2025-03-15, 78), (2025-04-02, 65), (2025-04-20, 83). This sequence intuitively reflects the fluctuations and trends of the student's homework scores on the "quadratic function" target knowledge point.

[0149] Step S503: Element extraction is performed on the personalized suggestions to obtain improvement element labels.

[0150] Element extraction refers to the process of parsing personalized suggestions into structured labels by a natural language model.

[0151] The improvement element label refers to the structured result obtained after element extraction, which is obtained through synonym normalization and dictionary mapping, and is used to reflect the key points of students' learning improvement.

[0152] By using a generative algorithm to parse the personalized suggestions, relevant elements are extracted according to the preset field structure (such as JSONSchema or function call format). The extraction result is subjected to synonym normalization and dictionary mapping to ensure that similar expressions are represented by uniform labels. When the algorithm is not confident, the system will start a fallback mechanism, such as calling the generative algorithm again, using rule matching to cover up, or submitting manual confirmation when necessary.

[0153] Step S504: Associate the improvement element labels with the performance sequence to form a learning performance feature set.

[0154] The learning performance feature set refers to the comprehensive feature set formed by associating the improvement element labels with the performance sequence after aligning them in chronological order.

[0155] Exemplarily, in a mathematics scenario, student A's performance sequence in the "target knowledge point = quadratic function" dimension 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 discriminant calculation" and "insufficient logical integrity". In this step, these improvement element labels are aligned with the corresponding time nodes, such as adding "weak discriminant calculation" on 2025-03-01 and "insufficient logical integrity" on 2025-04-02. The final learning performance feature set is: (target knowledge point = quadratic function, 2025-03-01, 72, [weak discriminant calculation]) (Objective knowledge point = quadratic function, 2025-04-02, 65, [Insufficient logical integrity]) Step S505: According to 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 within the preset observation window.

[0156] The fluctuation index refers to the fluctuation of the homework score in the time sequence, which is used to describe the score stability of the student in a certain content dimension. The homework score sequence under a certain content dimension is extracted from the learning performance feature set, and then the fluctuation of the score in the time sequence is quantified by calculating the difference between adjacent scores, sliding standard deviation or median absolute deviation, etc. to obtain the fluctuation index.

[0157] The preset fluctuation threshold is a preset constant, which can be adjusted according to actual needs.

[0158] The preset observation window is a preset time constant, which can be adjusted according to actual needs.

[0159] The preset number is a preset constant, which can be adjusted according to actual needs.

[0160] Step S506: If yes, determine the content dimension corresponding to the key information corresponding to the fluctuation index as the early warning object.

[0161] In another aspect, if the fluctuation index does not exceed the preset fluctuation threshold and reaches the preset number within the preset observation window, no processing is performed.

[0162] The early warning object refers to the content dimension marked for special attention due to abnormal fluctuation index.

[0163] After determining the early warning object, a set of meta information will be further attached to the early warning object, including student identification, corresponding content dimension, fluctuation index, preset fluctuation threshold, preset observation window, and cumulative trigger number. Among them, the early warning object not only marks "which dimension needs intervention", but also carries the data basis for "why the early warning is triggered".

[0164] Step S507: Generate a phased training plan for the early warning object.

[0165] The phased training plan refers to a plan scheme with stage division and target setting generated according to the specific circumstances of the early warning object.

[0166] Among them, the specific steps of generating a phased training plan for the early warning object can refer to the steps in the Figure 6 embodiment, which will not be described here.

[0167] Step S508: Push the phased training plan to the student end and the teacher end.

[0168] The student identifier is included in the periodic training plan, and after obtaining the periodic training plan, the periodic training plan is pushed to the student end and the teacher end corresponding to the student identifier according to the student identifier.

[0169] With reference to Figure 6 The step of generating the periodic training plan for the early warning object includes: Step S601: determining a warning level according to a super-threshold amplitude of the fluctuation index and a cumulative number of times that the fluctuation index exceeds a preset fluctuation threshold in a preset observation window.

[0170] The super-threshold amplitude refers to a value of the fluctuation index that exceeds the preset fluctuation threshold. The value of the fluctuation index corresponding to the preset fluctuation threshold can be obtained by subtraction.

[0171] The cumulative number of times refers to the number of times that the fluctuation index exceeds the preset fluctuation threshold in the preset observation window.

[0172] The warning level refers to a severity level of the fluctuation index when it exceeds the preset fluctuation threshold.

[0173] The super-threshold amplitude-cumulative number of times reference table is set to obtain the warning level. The severity level includes mild warning, moderate warning, and severe warning.

[0174] Step S602: determining a training plan parameter according to the warning level and the individualized correction model.

[0175] The training plan parameter refers to a configuration element when the periodic training plan is generated, and at least includes an initial value of the number of priority improvement items, the intensity of the practice task, and the recommended learning duration.

[0176] The specific steps of determining the training plan parameter according to the warning level and the individualized correction model can refer to the steps in the Figure 7 embodiment.

[0177] Step S603: determining a practice task according to the key information and the content dimension corresponding to the early warning object.

[0178] The practice task refers to a training content unit generated for the early warning object.

[0179] In a feasible implementation, the system determines the practice task through the following process: The candidate question set corresponding to the content dimension is called from the question bank; the questions are filtered according to the specific dimension of the early warning object, for example, "target knowledge point = quadratic function" only retains the questions related to the function relationship; the number of questions of appropriate quantity and difficulty is selected according to the priority improvement item number and task intensity in the training plan parameter; for the case of insufficient question bank, the generative algorithm is called to automatically generate the question content that meets the content dimension. For example: if the early warning object is a certain knowledge point (such as "quadratic function"), questions related to the knowledge point are preferentially selected; if the early warning object is a certain question type feature (such as "proof question"), multi-level difficulty exercises under this question type are generated.

[0180] Step S604: The exercise tasks are arranged in stages according to the plan parameters, and stage targets and completion criteria are set.

[0181] The stage arrangement refers to dividing the determined exercise tasks into several stages, and each stage corresponds to a specific training target.

[0182] The stage target is the learning level or mastery that the student is expected to achieve on the corresponding content dimension in each stage.

[0183] The completion criterion is a quantitative standard for determining whether the student has achieved the stage target.

[0184] In a feasible implementation, if the training plan parameter is set as "number of stages = 3, difficulty ladder = gradually increasing", the exercise tasks are divided into three stages: Stage one: allocate 30% of the basic exercise questions, the stage target is "master core knowledge points"; the completion criterion is "correct rate ≥ 70%"; Stage two: allocate 40% of the medium difficulty questions, the stage target is "significantly improve the logical integrity"; the completion criterion is "logical integrity score ≥ 0.75"; Stage three: allocate 30% of the high difficulty questions or comprehensive questions, the stage target is "can independently complete multi-step reasoning or cross-knowledge point application"; the completion criterion is "comprehensive score ≥ 80 points".

[0185] Step S605: Obtain the stage training plan based on the stage target and the completion criterion.

[0186] The exercise tasks, stage targets and completion criteria of each stage are integrated to form an executable stage training plan, including: the task list of each stage (number, question type, difficulty); stage target (knowledge points or ability level to be mastered); completion criterion (correct rate, score threshold, completion rate, etc.); the connection logic between stages (the completion of the criterion of the previous stage can enter the next stage).

[0187] Reference Figure 7The step of determining the training plan parameters according to the early warning level and the personalized correction model comprises: Step S701: determining the basic parameter values according to the early warning level, the basic parameter values comprising initial values of the number of priority improvement items, the exercise task intensity, and the recommended learning duration.

[0188] The basic parameter values refer to the initial values of the training plan parameters determined according to the early warning level.

[0189] The number of priority improvement items refers to the number of task items that need to be improved in the training plan.

[0190] The exercise task intensity refers to the combination level of the number and difficulty of the training tasks, for example, low intensity (small amount, basic questions), medium intensity (moderate amount, containing medium difficulty questions), and high intensity (large amount, containing high difficulty questions).

[0191] The recommended learning duration is the initial learning time required for the training according to the early warning level.

[0192] In a feasible implementation, the system presets the corresponding relationship between the early warning level and the basic parameters: Mild early warning: the number of priority improvement items = 1, the exercise task intensity = low, and the recommended learning duration = 30 minutes; Moderate early warning: the number of priority improvement items = 2, the exercise task intensity = medium, and the recommended learning duration = 60 minutes; Severe early warning: the number of priority improvement items = 3-4, the exercise task intensity = high, and the recommended learning duration = 90 minutes or more.

[0193] Step S702: adjusting the basic parameter values according to the personalized correction model to obtain a parameter correction result conforming to the teacher's preference.

[0194] The parameter correction result refers to the training plan parameters obtained by adjusting the basic parameter values through the personalized correction model combined with the teacher's correction preference information.

[0195] Taking the basic parameter values as input, the teacher's correction preference in the personalized correction model is adjusted. If the teacher tends to focus on logic and step integrity, the proportion of reasoning or process questions will be increased under the same level, and the learning duration will be increased accordingly. If the teacher prefers to focus on knowledge point coverage, the number of exercise tasks or the breadth of covered knowledge points will be increased. If the teacher emphasizes learning habits or expression style, the training period will be extended and the single task intensity will be reduced to ensure repeated practice and detailed feedback. The final output parameter correction result is more in line with the teacher's individualized teaching goals.

[0196] In a feasible implementation, the base parameters of student B are as follows: the number of priority improvement items = 2, the exercise task intensity = medium, and the recommended learning duration = 60 minutes. The teacher prefers to focus on logical integrity, and the personalized correction model corrects the parameters as follows: the number of priority improvement items remains unchanged; the exercise task intensity is increased from “medium” to “medium+”, that is, one more logical reasoning exercise is added; and the recommended learning duration is adjusted from 60 minutes to 75 minutes to meet the requirement of the teacher emphasizing reasoning training. The corrected parameter results are as follows: the number of items = 2, the exercise task intensity = medium+, and the learning duration = 75 minutes.

[0197] Step S703: Determine the difficulty ladder, the number of stages, the stage task quantity, and the check node time according to the parameter correction results.

[0198] The difficulty ladder refers to the progressive design of the difficulty of training tasks in different stages.

[0199] The number of stages refers to the division of training tasks into several stages for step-by-step completion.

[0200] The stage task quantity refers to the specific number of questions or training tasks that need to be completed in each stage.

[0201] The check node time refers to the evaluation time point set at the end of each stage to verify whether the student meets the completion criteria.

[0202] The parameter correction results are obtained based on the base parameter values, which include the initial values of the number of improvement items, the exercise task intensity, and the recommended learning duration.

[0203] In a feasible implementation, the difficulty ladder is mainly based on the corrected exercise task intensity, with low intensity corresponding to two layers of basic-application, medium intensity corresponding to three layers of basic-advanced-comprehensive, and high intensity expanding to four layers of basic-advanced-enhanced-breakthrough. The number of stages is mainly based on the corrected number of improvement items, with fewer items divided into two stages, moderate items divided into three stages, and more items expanded to four stages. The stage task quantity is allocated according to the difficulty ladder and the number of stages, with a relatively large allocation ratio in the basic stage and a relatively reduced number of enhanced or breakthrough stages. The check node time is based on the corrected learning duration, with the total duration evenly divided or weighted divided according to the number of stages, and a check node is set at the end of each stage.

[0204] Step S704: Obtain the training plan parameters according to the difficulty ladder, the number of stages, the stage task quantity, and the check node time.

[0205] The difficulty ladder, the number of stages, the stage task quantity, and the check node time are combined to obtain the training plan parameters.

[0206] In a feasible implementation, the parameter correction result of the student D is: the number of improved entries = 12, the task intensity = medium, and the learning duration = 80 minutes. It is determined that the difficulty ladder = 3 layers (basic - advanced - comprehensive), the number of stages = 3, the stage task amount = [5, 4, 3], and the examination node time = [25min, 55min, 80min]. The above is integrated into the training plan parameters: {difficulty ladder = 3 layers, number of stages = 3, stage task amount = [5, 4, 3], examination node time = [25, 55, 80] minutes}.

[0207] Based on the same inventive concept, an embodiment of the present application provides a teacher personalized correction system based on a generative algorithm, comprising: An acquisition module is configured to acquire preference information, historical correction data, assignment files, and assignment data. A memory is configured to store a program of the teacher personalized correction method based on the generative algorithm. A processor, the program in the memory can be loaded and executed by the processor and implement the teacher personalized correction method based on the generative algorithm.

[0208] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above functional modules is exemplified, and in actual application, the above functions can be completed by different functional modules according to needs, that is, the internal structure of the device is 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 refer to the corresponding process in the foregoing method embodiments, which will not be repeated here.

[0209] An embodiment of the present application provides a computer readable storage medium, which stores a computer program capable of being loaded and executed by a processor to execute a teacher personalized correction method based on a generative algorithm.

[0210] The computer storage medium includes, for example, a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0211] Based on the same inventive concept, an embodiment of the present application provides an intelligent terminal, comprising a memory and a processor, and the memory stores a computer program capable of being loaded and executed by the processor to execute a teacher personalized correction method based on a generative algorithm.

[0212] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above functional modules is taken as an example, and in actual application, the above functions can be completed by different functional modules according to needs, that is, the internal structure of the device is 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 refer to the corresponding process in the foregoing method embodiments, which will not be described here.

[0213] The above are preferred embodiments of the present application, and are not intended to limit the protection scope of the present application. Any feature disclosed in the specification (including the abstract and drawings) can be replaced by other equivalent or similar features unless specifically described. That is, each feature is only an example of a series of equivalent or similar features unless specifically described.

Claims

1. A teacher-personalized correction method based on a generative algorithm, characterized by: include: Collect teachers' grading preference information and historical grading data; Generate a personalized grading model through machine learning based on grading preference information and historical grading data; Receive homework files uploaded by the teacher, pre-process the homework files and extract key information; Based on the generative algorithm and key information, the scoring criteria for the assignment are obtained; Receive homework data uploaded by students; Correct homework data based on generative algorithms, personalized correction models, and homework scoring criteria to obtain homework scores and personalized recommendations; Push assignment scores and personalized recommendations to students and teachers.

2. A teacher-specific correction method based on a generative algorithm according to claim 1, characterized in that: The steps of obtaining the assignment scoring criteria based on the generative algorithm and key information include: Analyze the key information and obtain the analysis results, which include the subject category, question type characteristics and target knowledge points of the homework questions; Based on the parsing results, the generative algorithm is called to generate candidate reference answers; Verify the semantic consistency, logical integrity, and knowledge point coverage of candidate reference answers to obtain a verified reference answer set; Construct homework evaluation dimensions based on the reference answer set, including semantic similarity, logic and step completeness, and knowledge point mastery; Based on the homework evaluation dimensions and historical correction data, determine the weight distribution and grading thresholds for each homework 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 assignment scoring criteria are obtained.

3. A teacher-specific correction method based on a generative algorithm according to claim 2, characterized in that: The step of correcting the homework data according to the generative algorithm, the personalized correction model, and the homework score evaluation criteria to obtain the homework score includes: Align the job data with the parsing results to obtain the job alignment result; Based on the assignment scoring criteria, a generative algorithm is used to compare the assignment alignment results with the reference answer set in terms of semantics, steps, and knowledge point coverage, obtaining the original indicators for each assignment evaluation dimension. Based on the original indicator and score mapping relationship, the initial score of the operation data in each operation evaluation dimension is calculated; According to the initial score and scoring rules, the initial comprehensive score of the operation data is calculated; Input the initial score and initial comprehensive score into the personalized grading model and revise the homework evaluation dimensions; Based on the revised homework evaluation dimensions, the homework data is comprehensively scored to obtain the homework score.

4. A teacher-specific correction method based on a generative algorithm according to claim 3, characterized in that: The step of performing a correction operation on the homework data according to the generative algorithm, the personalized correction model, and the homework scoring evaluation criteria to obtain personalized suggestions includes: Conduct differential analysis on the homework data and the reference answer set to obtain the difference points; Map the difference points to the target knowledge points to generate a knowledge point mastery profile; Based on the knowledge point mastery profile, homework scores, and revised homework evaluation dimensions, 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 reviews based on review generation strategies; Generate learning improvement suggestions based on the mastery of knowledge points and suggested parameters; Combine personalized comments with learning improvement suggestions to get personalized recommendations.

5. The method for personalized teacher correction based on a generative algorithm according to claim 1, characterized in that: The method further comprises: Summarize all previous homework scores and personalized suggestions by student dimension, and form a learning record in chronological order; Conduct cluster analysis on learning records based on key information to obtain performance sequences oriented towards the content dimensions contained in the key information; Extract elements from personalized recommendations to obtain improved element labels; Associating the improvement factor labels with the performance sequences to form a learning performance feature set; Obtain a fluctuation index based on the learning performance feature set, and determine whether the fluctuation index exceeds a preset fluctuation threshold and reaches a preset number of times within a preset observation window; If so, the content dimension corresponding to the key information corresponding to the fluctuation indicator is determined as the warning object; Generate phased training plans for warning targets; Push phased training plans to students and teachers.

6. A teacher-specific correction method based on a generative algorithm according to claim 5, characterized in that: The step of generating a phased training plan for the warning target includes: 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; Determine training plan parameters based on warning levels and personalized correction models; Determine the practice tasks based on the content dimensions corresponding to the key information and warning objects; Arrange the practice tasks in stages according to the planned parameters, set stage goals and completion criteria; Based on the stage goals and completion criteria, a stage-by-stage training plan is obtained.

7. A teacher-specific correction method based on a generative algorithm according to claim 6, characterized in that: The step of determining the training plan parameters according to the warning level and the personalized correction model includes: Determine the basic parameter values ​​based on the warning level. These include the number of priority improvement items, the intensity of practice tasks, and the initial values ​​of the recommended learning time. Adjust the basic parameter values ​​according to the personalized correction model to obtain the parameter correction results that meet the teacher's preferences; Determine the difficulty level, number of stages, amount of tasks for each stage, and checkpoint time based on the parameter modification results; The training plan parameters are obtained based on the difficulty level, number of stages, amount of tasks in each stage, and check node time.

8. A teacher personalized correction system based on generative algorithm, characterized by: The system is used to execute the teacher personalized correction method based on a generative algorithm according to any one of claims 1 to 7, comprising: The acquisition module is used to obtain preference information, historical correction data, homework files, and homework data; A memory for storing a program of the teacher's personalized correction method based on a generative algorithm; The processor and the program in the memory can be loaded and executed by the processor to implement the teacher personalized correction method based on the generative algorithm.

9. An intelligent terminal, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and execute the method according to any one of claims 1 to 7.

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

Images