Composition intelligent correction and group management method and system based on large language model

Abstract

The application relates to the field of natural language processing and intelligent education technology, and discloses an essay intelligent correction and group management method and system based on a large language model, wherein the essay intelligent correction and group management method based on the large language model comprises the following steps: adopting a large language model to perform multi-dimensional semantic deep analysis on original essay text; performing feature extraction and capability evaluation on comprehensive semantic analysis results; performing clustering analysis and similarity calculation on individualized student portraits to generate an optimized learning group configuration scheme; performing strategy matching processing to generate a differentiated correction strategy; comprehensively processing the differentiated correction strategy and the comprehensive semantic analysis results to generate a detailed correction feedback report; designing a cooperation scheme to organize a group cooperation learning activity scheme; performing data statistics and effect analysis to generate a system performance evaluation report and an optimization improvement scheme; and the application solves the problems of low efficiency, lack of individualized guidance and group management function of traditional essay correction.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing and intelligent education technology, and more specifically, to a method and system for intelligent essay grading and group management based on a large language model. Background Technology

[0002] With the deepening development of educational informatization, essay correction, as a crucial part of Chinese language teaching, faces numerous challenges due to its traditional manual correction model. In secondary school Chinese language teaching practice, teachers need to process a large number of student essays, with each essay requiring an average of 15-20 minutes of correction time, resulting in a heavy workload and high repetition. Furthermore, the subjective differences in grading standards among different teachers make it difficult to guarantee the consistency and fairness of the evaluation results. Traditional correction methods also suffer from low feedback quality; teachers often can only provide simple overall evaluations, making it difficult to offer detailed, personalized guidance and suggestions for each student's specific problems.

[0003] In recent years, with the rapid development of artificial intelligence technology, intelligent essay grading systems have gradually emerged, attempting to solve the problems of traditional grading through automation. Existing intelligent grading systems are mainly based on natural language processing technology, which can identify grammatical errors and structural problems to a certain extent, but still has significant shortcomings in deep semantic understanding, creative evaluation, and personalized feedback. These systems often employ rule-driven or shallow machine learning methods, making it difficult to understand the deeper meaning and creative expression of essays, resulting in a significant gap between the evaluation results and those of human grading.

[0004] Furthermore, existing technologies have relatively weak research capabilities in group management and collaborative learning. Traditional teaching management mainly relies on teachers' experience and intuition to group students, lacking scientific data support and intelligent decision-making mechanisms. Peer assessment and assistance among students also lack effective organization and guidance, making it difficult to realize the educational value of peer learning. Therefore, how to organically combine intelligent grading with group management to build a complete intelligent teaching ecosystem has become an important topic in current educational technology research. Summary of the Invention

[0005] This invention provides a method and system for intelligent essay grading and group management based on a large language model, which solves the technical problems in related technologies, such as the difficulty in providing detailed personalized guidance for each student's specific problems and the lack of scientific data support and intelligent decision-making mechanisms.

[0006] This invention provides a method for intelligent essay grading and group management based on a large language model, comprising the following steps: Obtain the original essay texts submitted by students, perform multi-dimensional semantic in-depth analysis of the original essay texts, and generate comprehensive semantic analysis results; Statistical analysis methods are used to extract features and assess abilities from the comprehensive semantic analysis results to construct personalized student profiles. Cluster analysis and similarity calculation are performed on personalized student profiles. Through ability level matching, learning needs analysis and complementarity assessment, an optimized learning group configuration scheme is generated. The learning group configuration scheme and personalized student profile are matched with strategies to generate differentiated grading strategies based on different ability levels and learning characteristics. By comprehensively processing differentiated correction strategies and comprehensive semantic analysis results through a large language model, detailed correction feedback reports are generated through problem identification, improvement suggestions, and personalized guidance. Based on an optimized learning group configuration scheme and detailed feedback reports, a collaborative learning scheme is designed, which organizes group collaborative learning activities through task allocation, peer review mechanism and collaborative guidance. The system performs statistical analysis and effectiveness analysis on the results of group collaborative learning activities, and generates system performance evaluation reports and optimization and improvement plans.

[0007] In a preferred embodiment, the step of generating comprehensive semantic analysis results includes: Structured text data is generated based on the original essay text. A large language model is used to compare standard grammar rules and contextual semantics to identify vocabulary collocation errors, syntactic structure problems, and improper use of punctuation marks, and a grammatical accuracy score is obtained. Based on grammatical accuracy scores and structured text data, a large language model is used to calculate semantic similarity between adjacent sentences, thematic consistency between paragraphs, and logical fluency of the whole text, to obtain a semantic coherence score. Based on semantic coherence scoring, a logical reasoning algorithm is used to evaluate the viewpoints, supporting arguments, and reasoning process, and to assess the completeness, persuasiveness, and logical consistency of the argument, resulting in a logical rigor score. Based on logical rigor scoring and original text content, an innovative evaluation algorithm is used to analyze the degree of expressive innovation in essays. By identifying novel viewpoints, unique expressions, creative metaphors, and original insights, an innovative expressive score is obtained.

[0008] In a preferred embodiment, the step of constructing a personalized student profile includes: Based on students' historical essay data, statistical analysis methods were used to extract writing style features, error pattern features, theme preference features, and ability change trend features to obtain historical feature vectors. Based on historical feature vectors and comprehensive semantic analysis results, a capability assessment model is used to process historical data through time-weighted averaging and then weighted and fused with current performance to obtain capability level indicators. Based on ability level indicators and historical feature vectors, clustering analysis algorithms are used to generate ability level labels, learning type labels, writing preference labels, and development potential labels to obtain student profiles.

[0009] In a preferred embodiment, the step of generating an optimized learning group configuration scheme includes: Based on all personalized student profiles, a multi-objective optimization theory is used to construct an objective function that comprehensively considers the similarity of abilities among members within a group, the differences between groups, and the balance of group size, thus obtaining the optimization objective. Based on actual teaching needs and system resource constraints, a constraint optimization method is used to set group size limits, ability distribution requirements, and teacher allocation constraints, resulting in a set of constraint conditions. Based on the set of optimization objectives and constraints, a genetic algorithm is used to search for the optimal group partitioning scheme through population evolution, selection, crossover and mutation operations to obtain the initial group configuration. Based on the initial group configuration and learning progress feedback, a dynamic adjustment algorithm is used to monitor the differences in learning performance within the group and trigger the reassignment of group members according to a preset threshold, resulting in a dynamically optimized group configuration scheme.

[0010] In a preferred embodiment, the step of generating a differentiated grading strategy includes: Based on group configuration and personalized student profiles, statistical analysis methods are used to calculate the average ability level, ability distribution characteristics, learning preference distribution and development potential of each group, and the group feature vector is obtained. Based on group feature vectors and a predefined strategy template library, a pattern matching algorithm is used to select basic reinforcement, capability enhancement, or innovative extension grading strategy templates to obtain an initial strategy scheme. Based on the initial strategy and specific student profiles, a strategy adjustment algorithm is used to analyze the differences between the individual profiles and the group average characteristics, and adjust the key areas of focus, the level of feedback detail, and the direction of guidance to obtain a personalized grading strategy. Based on personalized grading strategies and teaching schedules, a time-series programming algorithm is used to schedule grading time, feedback frequency, and follow-up plans, resulting in differentiated grading strategies.

[0011] In a preferred embodiment, the step of generating a detailed feedback report includes: Based on personalized correction strategies and semantic analysis results, a problem identification algorithm is used to locate specific problems in the essays and classify and label them according to severity, type and difficulty of improvement to obtain a problem list; Based on the problem list and language generation model, a hierarchical generation strategy is adopted to construct personalized feedback content that includes overall evaluation, specific problem correction, improvement suggestions, and affirmation of excellent expressions; Based on the text feedback content, quality assessment and optimization algorithms are used to check logical consistency, evaluate clarity of expression, verify guiding value, and adjust sentiment. Based on the optimized feedback content, visualization technology is used to generate an intuitive feedback display interface that includes layout templates, color annotations, chart elements, and interactive functions, resulting in a detailed feedback report.

[0012] In a preferred embodiment, the organizational group collaborative learning activity scheme includes: Based on group configuration and members' grading results, a task design algorithm is used to determine the type of collaborative task, match the task difficulty, design the task process, formulate evaluation criteria, and configure resource support according to the ability level and learning objectives of group members, so as to obtain a collaborative task plan. Based on the collaborative task scheme, a peer evaluation mechanism is established within the group using a peer evaluation algorithm. This involves formulating peer evaluation rules, designing evaluation dimensions, implementing anonymous allocation, setting cross-validation, integrating feedback information, and monitoring the quality of peer evaluation. Based on group characteristics and learning needs, a recommendation algorithm is used to recommend suitable learning resources to each group. This involves identifying learning needs, building a resource library, matching resource features, generating personalized recommendations, evaluating recommendation quality, and dynamically updating resources. Based on the implementation of collaborative learning activities and student feedback, an effectiveness evaluation method is used to analyze the effectiveness of collaborative learning. This involves collecting quantitative and qualitative data, calculating quantitative indicators, conducting qualitative analysis, performing comprehensive effectiveness evaluation, comparative analysis and verification, and generating improvement suggestions to obtain a collaborative learning activity plan.

[0013] In a preferred embodiment, the steps of generating the system performance evaluation report and optimization and improvement plan include: Based on learning outcome data and system operation indicators, a multi-dimensional analysis method is used to evaluate teaching effectiveness. By integrating data from multiple channels, analyzing student ability improvement, assessing system user satisfaction, analyzing teacher work efficiency, and evaluating teaching quality improvement, an effectiveness analysis report is generated. Based on the performance analysis report, a problem diagnosis algorithm is used to identify deficiencies and problems in the system operation. By identifying abnormal data, comparing performance benchmarks, analyzing user feedback, mining system logs, locating root causes, and prioritizing problems, the problem diagnosis results are obtained. Based on the problem diagnosis results, an optimization decision algorithm is used to formulate a system improvement plan. The optimization improvement plan is obtained through setting improvement goals, designing solutions, evaluating resource constraints, prioritizing solutions, analyzing risks, and formulating implementation plans. Based on the optimization and improvement plan, an adaptive optimization mechanism for the system is established using a continuous improvement approach. This mechanism involves feedback loop establishment, iterative optimization process, adaptive adjustment mechanism, version management control, effect tracking and evaluation, and knowledge accumulation and sedimentation, resulting in a system performance evaluation report.

[0014] In a preferred embodiment, the large language model adopts a pre-trained language model based on the Transformer architecture, which is trained on a large-scale Chinese essay corpus for domain-adaptive fine-tuning, and has the ability of grammatical analysis, semantic understanding, logical reasoning and innovation evaluation. The competency assessment model adopts a multi-layer neural network structure and obtains a quantitative assessment function of students' writing ability through training on historical data. It can process time-series data and output standardized competency indicators. The genetic algorithm uses real-number encoding to represent the group partitioning scheme, evaluates the quality of group configuration through the fitness function, and uses tournament selection, single-point crossover, and Gaussian mutation operations for population evolution.

[0015] This invention provides an intelligent essay grading and group management system based on a large language model, used to execute the aforementioned intelligent essay grading and group management method based on a large language model, including: The multi-dimensional semantic deep analysis module is used to obtain the original essay text submitted by students, perform multi-dimensional semantic deep analysis on the original essay text, and generate comprehensive semantic analysis results. The personalized student profile building module is used to extract features and assess abilities from the comprehensive semantic analysis results using statistical analysis methods to build personalized student profiles. The intelligent group division module is used to perform cluster analysis and similarity calculation on personalized student profiles, and generate an optimized learning group configuration scheme through ability level matching, learning needs analysis and complementarity assessment. The differentiated grading strategy generation module is used to perform strategy matching processing on learning group configuration schemes and personalized student profiles, and generate differentiated grading strategies based on different ability levels and learning characteristics. The intelligent correction feedback generation module is used to comprehensively process differentiated correction strategies and comprehensive semantic analysis results through a large language model, and generate detailed correction feedback reports through problem identification, improvement suggestions and personalized guidance; The group collaborative learning organization module is used to design collaborative learning plans based on optimized learning group configuration schemes and detailed feedback reports. It organizes group collaborative learning activities through task allocation, peer review mechanisms, and collaborative guidance. The teaching effectiveness evaluation and optimization module is used to perform data statistics and effect analysis on the execution results of group collaborative learning activities, and generate system performance evaluation reports and optimization and improvement plans.

[0016] The beneficial effects of this invention are as follows: By leveraging the powerful semantic understanding capabilities of a large language model, this invention comprehensively analyzes essays across four dimensions: grammatical accuracy, semantic coherence, logical rigor, and innovative expression. It not only identifies deep semantic issues that traditional systems struggle to detect but also accurately evaluates the essay's innovative value and expressive features. Compared to traditional automated grading systems, this invention improves scoring consistency and feedback detail, providing students with more professional and precise writing guidance. Furthermore, through the construction of personalized student profiles, the system can adjust grading strategies based on each student's characteristics and needs, achieving true individualized instruction and significantly enhancing the relevance and effectiveness of teaching.

[0017] By employing multi-objective optimization and genetic algorithms, students can be scientifically and rationally divided into learning groups with different ability levels, and a dynamic adjustment mechanism can be established to adapt to changes in students' abilities. Building upon group management, the system further organizes collaborative learning activities, including peer review and correction, problem discussions, and resource sharing, effectively leveraging the educational value of peer learning. This invention enhances students' collaborative abilities, learning enthusiasm, and overall satisfaction. Students' writing skills within the groups are improved compared to traditional teaching models, demonstrating the crucial role of group management mechanisms in improving teaching effectiveness. Attached Figure Description

[0018] Figure 1 This is a flowchart of the intelligent essay grading and group management method based on a large language model of the present invention; Figure 2 This is a module diagram of the intelligent essay grading and group management system based on a large language model of the present invention; Figure 3 This is a comparison chart of the processing efficiency of different methods of the present invention. Detailed Implementation

[0019] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0020] At least one embodiment of the present invention discloses a method for intelligent essay grading and group management based on a large language model, such as... Figure 1 As shown, it includes: Step 1: Obtain the original essay texts submitted by students, perform multi-dimensional semantic in-depth analysis on the original essay texts, and generate comprehensive semantic analysis results; Before collecting student essay data, the school clearly informs students and their legal guardians of the purpose of the data and obtains written authorization and consent; Anonymize the original essay texts, removing students' names, classes, and other personal identification information; Data transmission adopts encryption measures, such as TLS 1.3 encryption in existing technologies, and storage adopts AES-256 encryption, and access control and log auditing are implemented; The system follows the principle of minimum necessity and regularly cleans up data to ensure compliance with relevant regulations on the protection of minors.

[0021] Specifically, the following steps are included: Based on the original essay text, natural language processing techniques are used for word segmentation and syntactic analysis to obtain structured text data. This implementation method employs a pre-trained language model based on the Transformer architecture, which is trained on a large-scale Chinese essay corpus for domain-adaptive fine-tuning, enabling it to perform grammatical analysis, semantic understanding, logical reasoning, and innovation evaluation. The specific implementation steps are as follows: The Transformer model architecture is configured with a 12-layer Transformer encoder structure, each layer including a multi-head self-attention mechanism and a feedforward neural network. The hidden layer dimension is set to 768, the number of attention heads to 12, and the feedforward network dimension to 3072. BERT-base-Chinese is used as the pre-trained base model, with a vocabulary size of 21128 and a maximum sequence length of 512, supporting semantic understanding and analysis of long texts. Domain-adaptive fine-tuning training is performed using a domain corpus of 500,000 Chinese essays, covering different grades and genres. Pre-training is conducted using a masked language model task to learn the linguistic features of Chinese essays, with a learning rate of 2e-5, a batch size of 32, and 10 training epochs. The AdamW optimizer is used, with weight decay set to 0.01 to ensure model convergence stability. Multi-task joint training is implemented, designing four sub-tasks: grammatical error detection, semantic understanding, logical reasoning, and innovation evaluation. A multi-task learning framework is constructed, sharing Transformer encoder parameters, and a dedicated output layer and loss function are designed for each sub-task. The text is divided into several sections: **Failure Function:** A task weight balancing strategy is employed to ensure coordinated optimization of each task. **Character-Level Text Cleaning:** Each character in the original essay text is scanned, and non-standard characters, such as special symbols and garbled characters, are identified and removed. Redundant formatting tags, such as HTML tags and Markdown symbols, are removed, and character encoding is standardized to ensure text consistency. **Sentence Segmentation and Recognition:** Sentence boundaries are identified based on punctuation marks (periods, question marks, exclamation marks, etc.). Special cases, such as ellipses and quotations, are handled. Each identified sentence is numbered, and the total number of sentences is calculated and recorded. **Paragraph Structure Analysis:** Paragraph boundaries are identified based on line breaks and indentation. Consecutive sentences are categorized into corresponding paragraphs, each paragraph is assigned a unique identifier, and the total number of paragraphs and sentences contained within each paragraph is counted. **Hierarchical Structure Information Extraction:** Basic statistical information of the text is calculated, including the total number of words, sentences, and paragraphs. The length distribution of each sentence is analyzed, the average sentence length is calculated, and the logical hierarchy of the text, such as titles, body text, and conclusions, is identified. A tree-like structure representation of the text is constructed for subsequent analysis and processing.

[0022] Through the above steps, the original essay text is converted into a structured dataset containing multiple paragraphs. Each paragraph contains complete sentence information and location markers, providing standardized input data for subsequent semantic analysis.

[0023] Structured text data is generated from the original essay text, and error detection is performed using the grammatical analysis capabilities of a large language model to obtain a grammatical accuracy score. The specific implementation steps are as follows: Establish a standard grammar rule database, including lexical collocation rules and syntactic structure norms, and compile common grammatical error types such as subject-verb disagreement, tense errors, and misuse of prepositions. Establish a punctuation usage standard database, including standard usage rules for various punctuation marks such as commas, periods, question marks, exclamation marks, pauses, semicolons, colons, quotation marks, parentheses, and ellipses, and compile common punctuation error types such as comma overuse, missing periods, misuse of question marks, incorrect quotation mark pairings, and confusion between pauses and commas. Identify the appropriateness of punctuation usage based on contextual semantics, such as whether a period is correctly used at the end of a sentence, whether a question mark is used in an interrogative sentence, and whether pauses are correctly used for parallel components. The model assigns severity weights to different types of grammatical errors, such as commas or semicolons. It employs a bidirectional LSTM structure, consisting of an input layer, a bidirectional LSTM layer, and an output layer. The input layer receives a sequence of word vectors encoded by the aforementioned Transformer model, with a dimension of 768. The bidirectional LSTM layer contains 256 hidden units each in the forward and backward LSTM layers, capturing the contextual dependencies of words. The output layer uses a fully connected layer with a Softmax activation function to output the grammatical validity probability distribution of each word in the current context. The model is trained using a corpus of 100,000 essays labeled with grammatical errors. The model employs a cross-entropy loss function, the Adam optimizer, a learning rate of 0.001, a batch size of 64, and 50 training epochs. It identifies complex grammatical problems requiring contextual judgment, such as ambiguous pronoun references, semantic redundancy, and improper collocations. It iterates through each sentence in a structured text, performing part-of-speech tagging and syntactic analysis. The sentence structure is compared to standard grammar rules, identifying and marking grammatical errors, recording error types and locations. Punctuation usage is also checked, scanning the position and type of all punctuation marks in each sentence, comparing punctuation usage to standard specifications, and identifying improper punctuation. The system analyzes the matching degree between punctuation marks and contextual semantics, including issues such as missing, redundant, misused, and incorrectly placed punctuation marks. It determines whether punctuation marks accurately express semantic pauses and tone, and records the specific location, type, and correct usage suggestions for each punctuation error. Based on preset weighting coefficients, the severity score for each grammatical error is calculated. The weighting coefficient for vocabulary collocation errors is set at 0.8, for syntactic structure problems at 1.0, and for improper punctuation usage, the weighting coefficient is graded according to the degree of impact: 0.7 for missing or incorrect punctuation at the end of a sentence, 0.6 for punctuation within a sentence that severely affects comprehension, and 0 for general punctuation misuse.3. Considering the impact of errors on sentence comprehension, adjust the weighting allocation, cumulatively calculate multiple errors within the same sentence, and record the total number of errors and weighted error score for each sentence. Calculate the grammatical accuracy rate for each sentence using the formula: 1 minus (number of errors multiplied by error weight divided by sentence length). The number of errors includes the sum of vocabulary collocation errors, syntactic structure errors, and punctuation errors. The error weight is the cumulative value of each type of error weighted according to its respective weight coefficient. Punctuation errors are included in the overall error weight calculation according to the aforementioned tiered weight coefficients. Average the grammatical accuracy rates of all sentences to obtain a comprehensive grammatical accuracy score for the entire essay, generating a detailed grammatical error report, including error location, type, and improvement suggestions.

[0024] Through the above steps, the system can accurately identify various grammatical problems in the essay and provide a quantitative score for grammatical accuracy, providing precise data support for subsequent personalized guidance.

[0025] Based on grammatical accuracy scores and structured text data, the semantic understanding capabilities of a large language model are used to analyze the logical coherence of the text, resulting in a semantic coherence score. The specific implementation steps are as follows: Inter-sentence coherence analysis calculates the semantic similarity between adjacent sentences. Sentence vector encoding technology is used to convert each sentence into a high-dimensional vector representation. The cosine similarity between adjacent sentence vectors is calculated to quantify the degree of semantic similarity; a similarity value closer to 1 indicates a stronger semantic relevance. The similarity value is linearly mapped to a range of 0 to 100 points as the base score for semantic similarity. Logical relationships between sentences are analyzed, identifying causal, adversative, and progressive relationships, among other logical types. The use of conjunctions is examined; adjacent sentence pairs using appropriate conjunctions receive an additional 10 points on the base score; sentence pairs without conjunctions but with naturally connected semantics retain their base score; and sentence pairs with inappropriate conjunctions or confused logical relationships deduct 15 points from the base score. The analysis also detects the main themes between sentences. Topic consistency is assessed by analyzing the topic distribution differences between adjacent sentences using a topic model and calculating the KL divergence value. A KL divergence value less than 0.3 is considered topic consistent and receives full marks. A KL divergence value between 0.3 and 0.6 is considered partially topic-related and deducts points linearly, specifically by subtracting the difference between the divergence value and 0.3 from 100 points. A KL divergence value greater than 0.6 is considered a serious topic jump and deducts 30 points. The effectiveness of transition words and conjunctions is evaluated. The semantic similarity score, logical relationship score, and topic consistency score are weighted and averaged according to a weight ratio of 50%, 30%, and 20% to obtain the coherence score for each pair of adjacent sentences. Then, the arithmetic mean of the scores of all adjacent sentence pairs in the entire text is calculated to obtain the final inter-sentence coherence score.

[0026] Inter-paragraph coherence analysis extracts the theme and core viewpoint of each paragraph. A key sentence extraction algorithm identifies the central sentence of each paragraph as the core viewpoint, and word frequency statistics are used to extract thematic keywords, constructing a thematic feature representation of each paragraph. The logical order and structural arrangement between paragraphs are analyzed, identifying the functional type of each paragraph, including introductory, argumentative, transitional, and concluding paragraphs. The paragraph arrangement is checked to ensure it conforms to common writing structure patterns. Articles conforming to standard structural patterns receive 80 points for basic structure; those with reasonable but non-standard structures receive 70 points; and those with chaotic and illogical paragraph arrangements receive 50 points. Thematic continuity and development between paragraphs are examined by calculating the cosine similarity of thematic features between adjacent paragraphs. A similarity score between 0.3 and 0.7 is considered a reasonable continuation of the theme, demonstrating both continuity and development, and is awarded a full score of 100. A similarity score greater than 0.7 is considered a repetitive theme with redundant content and a lack of originality, resulting in a deduction of 20 points. A similarity score less than 0.3 is considered a theme jump with excessive abruptness and a lack of coherence, resulting in a deduction of 25 points. The naturalness and rationality of paragraph transitions are assessed by checking for transitional sentences at the beginning and end of paragraphs. Paragraphs with clear transitional sentences receive an additional 15 points, while those without obvious transitional sentences but with natural logic receive no deduction. Abrupt transitions result in a deduction of 10 points. The scores for structural arrangement, theme continuation, and naturalness of transitions are weighted and combined at a ratio of 40%, 40%, and 20%, respectively, to obtain the inter-paragraph coherence score.

[0027] The entire text is analyzed for coherence, identifying its overall structural framework. A structural pattern recognition algorithm is used to analyze the sequence of paragraph functions, comparing it against eight predefined standard writing structure templates, including common structural types such as general-specific-general, parallel, progressive, and contrastive structures. Articles that fully conform to a standard structure receive a score of 90, those that partially conform receive 70, and those without a clear structural framework receive 50. The text also analyzes the argumentative logic and the development of viewpoints, extracting the central argument and sub-arguments, constructing a hierarchical tree diagram of the relationships between arguments, and checking for clear logical relationships, including supporting, parallel, and progressive relationships. Articles with complete and clearly structured argumentation receive a score of 100, while those with only basic logical relationships receive a score of 50. A score of 80 is given for clarity but minor issues, and 60 is given for logical inconsistencies. The text then examines the coherence between the introduction, body, and conclusion, extracting ten keywords from each of the opening and closing paragraphs. The overlap ratio is calculated by dividing the number of overlapping keywords by the total number of keywords. A ratio greater than 0.4 indicates good coherence and thematic consistency, earning a full score of 100. A ratio between 0.2 and 0.4 indicates average coherence, earning 80 points. A ratio less than 0.2 indicates a lack of coherence and a disconnect between the opening and closing themes, earning 60 points. Finally, the text is evaluated for logical flow and overall consistency. The structural framework score, logical argumentation score, and coherence score are weighted at 35%, 45%, and 20% respectively to obtain the overall coherence score.

[0028] Data preprocessing and standardization were performed on the coherence scores across three dimensions: sentence-to-sentence, paragraph-to-paragraph, and full-text. A standard score conversion method was used to calculate the deviation of the raw scores from historical averages. Historical averages were determined based on statistical data from 100,000 sample essays, with historical averages of 72.5 for sentence-to-sentence coherence, 75.3 for paragraph-to-paragraph coherence, and 78.1 for full-text coherence. The standard deviations for each dimension were recorded as 12.3, 11.8, and 10.5, respectively. The standardized score was obtained by subtracting the corresponding historical average from the raw score and then dividing by the standard deviation. This standardized score reflects the current essay's position relative to the historical average. An S-curve mapping function was used to convert the standardized scores to a range of 0 to 1. The mapping process ensured that a standardized score of 0 corresponded to 0.5; a larger positive deviation resulted in a mapping value closer to 1, and a larger negative deviation resulted in a mapping value closer to 0. Finally, the mapping result was multiplied by 100 to convert it into a percentage score.

[0029] The weighted fusion calculation sets weight coefficients for sentence-to-sentence, paragraph-to-paragraph, and overall coherence based on different writing styles and evaluation needs. For narrative essays, the weight ratios are set to 40%, 30%, and 30%, because narrative essays emphasize smooth transitions between sentences. For argumentative essays, the weight ratios are set to 30%, 30%, and 40%, because argumentative essays emphasize the integrity of the overall argumentative structure. For expository essays, the weight ratios are set to 35%, 35%, and 30%, reflecting the balanced requirement for coherence at all levels in expository essays, ensuring that the sum of the three weight coefficients equals 100%, and maintaining the rationality of the score. The scores for sentence-to-sentence coherence, paragraph-to-paragraph coherence, and overall coherence are multiplied by their respective weight coefficients and then summed to obtain the final comprehensive semantic coherence score, ranging from zero to one hundred points.

[0030] Through the above steps, the system can comprehensively evaluate the semantic coherence of the essay from multiple levels and provide students with targeted suggestions for improving the logical structure.

[0031] Based on semantic coherence scoring, a logical reasoning algorithm is used to analyze the logical relationship between arguments and evidence to obtain a logical rigor score. The specific implementation steps are as follows: Argumentation unit identification involves scanning text content, identifying main viewpoints and argumentative paragraphs, extracting the core arguments in each argumentation unit, identifying the specific arguments and evidence supporting each argument, and analyzing the hierarchical relationships and logical structure between argumentation units.

[0032] Logical relationship strength analysis assesses the degree of logical connection between arguments and evidence. Semantic relevance calculation is used to extract keyword vectors from arguments and evidence, and the cosine similarity between these vectors is calculated. A similarity greater than 0.6 is considered strong relevance and receives 90 points; a similarity between 0.4 and 0.6 is considered moderate relevance and receives 70 points; and a similarity less than 0.4 is considered weak relevance and receives 50 points. The relevance and persuasiveness of evidence are analyzed, identifying evidence types including factual evidence, data evidence, theoretical evidence, and case evidence. Evidence using specific facts or data receives a high persuasiveness score of 90 points; evidence using theories or cases receives a moderate persuasiveness score of 75 points; and evidence using personal opinions or vague expressions receives a low persuasiveness score of 60 points. The rationality and rigor of the reasoning process are examined. The algorithm uses logical rules to identify each step in the reasoning chain, checking for a necessary connection between premises and conclusion. Full marks are awarded for complete and logically necessary reasoning steps; 15 points are deducted for reasonably reasonable but abrupt leaps in reasoning steps; and 30 points are deducted for obvious flaws in reasoning. The algorithm also identifies logical loopholes and circular arguments, checking for instances of circular arguments (i.e., using arguments to prove evidence or restating arguments with evidence), and for logical fallacies such as substitution of concepts, generalization, and causal inversion. 20 points are deducted for each logical loophole, and 25 points for each circular argument. The algorithm then calculates a weighted average of the relevance of arguments to evidence, the persuasiveness of arguments, and the rationality of reasoning (30%, 35%, and 35% respectively), subtracting the deductions for logical problems to obtain the logical relationship strength value.

[0033] The argument quality assessment evaluates the credibility and authority of the arguments, identifying the types of sources, including authoritative publications, academic research findings, official statistics, expert opinions, news reports, and personal experience. Arguments from authoritative institutions or academic research are awarded a high credibility score of 95 points; those from official data or expert opinions are awarded a relatively high credibility score of 85 points; those from news reports are awarded a moderate credibility score of 70 points; and those from personal experience or without a cited source are awarded a low credibility score of 60 points. The assessment also analyzes the sufficiency and representativeness of the evidence, counting the number of arguments supporting each point. Three or more arguments are considered sufficient (100 points); two arguments are considered substantially sufficient (80 points); and one argument is considered insufficient (60 points). The assessment also checks whether the arguments cover different angles or types, ensuring they cover three or more different... The evaluation criteria are as follows: A strong representativeness of the angle of view earns an extra 10 points; a view covering only a single angle is considered weakly representative and receives no extra points. The completeness and comprehensiveness of the argument are examined, analyzing whether the argument structure includes the three complete stages of presenting the argument, developing the argument, and summarizing the viewpoint. A complete structure earns 90 points; a structure lacking a summary earns 75 points; and a structure containing only the argument and simple evidence earns 60 points. The handling of counterarguments is considered, checking whether possible opposing opinions or different viewpoints are mentioned in the text. Actively presenting counterarguments and providing effective responses earns an extra 15 points; mentioning counterarguments but failing to adequately address them earns an extra 8 points; and completely omitting counterarguments earns no extra points. The score for the quality of the argument is calculated by weighting the scores for the credibility of the evidence, the sufficiency of the evidence, and the completeness of the argument according to a weighting ratio of 40%, 35%, and 25%, respectively. This weighted average, along with the extra points for representativeness and handling of counterarguments, yields the final score, with a maximum of 100 points.

[0034] The logical rigor is comprehensively calculated by dividing the logical relationship strength value and the argument quality score of each argument unit by 100 for normalization. The two normalized values ​​are then multiplied to obtain the comprehensive score for that argument unit, reflecting its overall performance in both logical relationship and argument quality dimensions. The comprehensive scores of all argument units are summed and divided by the total number of argument units to obtain the average comprehensive score. This average score is then multiplied by 100 to convert it to a percentage, yielding the average logical rigor score for the entire text. Considering the differences in the importance of argument units, a weighting coefficient of 1.2 is assigned to argument units related to the central argument, and a weighting coefficient of 1.2 is assigned to argument units related to the main sub-arguments. Each argument unit is assigned a weight coefficient of 1.0, and the argument units for secondary arguments are assigned a weight coefficient of 0.8. The comprehensive score of each argument unit is multiplied by its corresponding weight coefficient, summed, and divided by the total weight coefficient to obtain the weighted average logical rigor score. A detailed logical analysis report is generated, listing the logical relationship strength value, argument quality score, and comprehensive score for each argument unit. Argument units with scores below 70 are marked as weak links, and specific problem types such as insufficient evidence, logical jumps, and low credibility of evidence are pointed out for weak links. Corresponding improvement directions are provided, such as supplementing authoritative evidence, improving reasoning steps, and adding handling of refutation viewpoints.

[0035] Through the above steps, the system can deeply analyze the argumentation logic of the essay, identify problems in the reasoning process, and provide students with specific guidance to improve their logical thinking skills.

[0036] Based on logical rigor scoring and the original text content, an innovation assessment algorithm is used to analyze the degree of expressive innovation in essays, resulting in an innovative expressiveness score. The innovation assessment algorithm, based on natural language processing technology, objectively measures the innovative characteristics of essays by constructing multi-dimensional computable indicators. Specifically, it assesses the innovative value of essays by identifying novel viewpoints, unique expressions, creative metaphors, and original insights. The specific implementation steps are as follows: Lexical uniqueness analysis involves constructing a standard vocabulary database and a high-frequency vocabulary list as benchmarks, scanning all words in the essay, identifying low-frequency words and technical terms, calculating the inverse document frequency of words, assessing the rarity of words, and statistically analyzing the frequency and distribution of unique words to obtain a lexical uniqueness score. Sentence structure innovation assessment involves analyzing the grammatical structure and sentence type of sentences, identifying special sentence structures such as compound sentences, inverted sentences, and rhetorical questions, statistically analyzing the frequency and variability of different sentence structures, assessing the novelty and expressive effect of sentence combinations, and calculating a sentence structure innovation score. Rhetorical originality detection involves identifying rhetorical devices such as metaphor, personification, and parallelism used in the text, analyzing the frequency and degree of innovation of these devices, assessing the originality and artistic effect of rhetorical expression, and checking the relationship between rhetorical devices and the content. The matching degree of content yields a rhetorical originality score; the judgment of viewpoint novelty involves extracting the core viewpoints and argumentative angles from the essay, comparing and analyzing them with a common viewpoint database, evaluating the uniqueness and depth of the viewpoints, analyzing their innovative value and inspirational value, and calculating the viewpoint novelty score; data preprocessing and comprehensive scoring involve using IDF weighting to process the vocabulary uniqueness score to highlight the value of rare words, using One-hot encoding to standardize different sentence structure types, using label encoding to quantify the diversity of rhetorical techniques, and using Min-Max normalization to ensure a consistent score range for the viewpoint novelty score. The standardized scores of the four dimensions are then weighted and integrated to generate a comprehensive score for innovative expression and a detailed analysis report.

[0037] Through the above steps, the system can comprehensively evaluate students' creative expression ability in their writing, identify highlights and areas for improvement, and provide targeted guidance for cultivating students' innovative thinking and expression skills.

[0038] The output is a comprehensive semantic analysis result, including scores for four dimensions: grammatical accuracy, semantic coherence, logical rigor, and innovative expressiveness, which serve as the basis for subsequent personalized analysis.

[0039] Furthermore, a weighted fusion method can be used to calculate the comprehensive score. By adjusting the weights of different dimensions, the evaluation needs of different types of essays can be adapted, thereby improving the relevance and accuracy of the scoring.

[0040] Step 2: Use statistical analysis methods to extract features and assess abilities from the comprehensive semantic analysis results to construct personalized student profiles; Specifically, the following steps are included: Based on students' historical essay data, statistical analysis methods are used to extract students' writing feature patterns, resulting in historical feature vectors. The specific implementation steps are as follows: Data collection and organization involves gathering all past student essay data, including content, scoring records, and revision history. The data is organized chronologically to establish a complete learning trajectory. Basic information for each essay, such as writing time, topic type, and word count, is extracted. Outliers and missing values ​​are removed to ensure data quality. Writing style analysis analyzes students' vocabulary preferences and language expression habits, statistically analyzes the frequency and trends of sentence structure usage, identifies rhetorical device usage patterns, extracts structural features and patterns, and forms style feature vectors. Common error identification involves statistically analyzing the frequency of errors in grammar, spelling, and punctuation, analyzing the distribution patterns and trends of error types, identifying students' weaknesses and common mistakes, and establishing personalized error pattern profiles to provide a basis for targeted guidance. Analysis of areas of expertise involves analyzing students' performance differences in essays on different topics, identifying students' strengths in writing areas and themes, assessing their depth of understanding and expression of different topics, and establishing a mapping relationship between topic preferences and abilities. Assessment of ability improvement trends involves analyzing the time trajectory of students' writing abilities, identifying key nodes and influencing factors for ability improvement, evaluating learning effectiveness and improvement speed, and predicting future development potential and room for improvement.

[0041] Through the above steps, the system can comprehensively extract students' historical writing characteristics, including style features, error patterns, theme preferences, and improvement trends, providing data support for personalized teaching.

[0042] Combining the semantic analysis results of the current essays with historical feature vectors, an ability assessment model is used to calculate the students' current writing ability level. The specific implementation steps are as follows: This approach integrates multi-dimensional ability indicators, standardizing the current essay's scores for grammatical accuracy, semantic coherence, logical rigor, and creative expression. It then adjusts the weight of the current score by incorporating ability trend information from historical feature vectors. Considering students' learning stages and individual characteristics, personalized assessment standards are set, and a comprehensive ability indicator is calculated through weighted fusion. Time decay weighting assigns different weight coefficients to historical essay data based on their time proximity, assigning higher weights to recent essays to reflect students' current true level and lower weights to earlier essays to avoid interference from outdated information. An exponential decay function is used to calculate the time weights, ensuring the timeliness of the assessment. Standardized comparative analysis establishes a benchmark database of abilities for students of the same grade and level. Students' ability indicators are compared with the benchmark data to calculate their relative ranking and percentile in each dimension, identifying their strengths and weaknesses. Multi-layer perceptron network fusion constructs a multi-layer perceptron neural network model, inputting the current score and historical features. Using a multi-layer neural network structure, a quantitative assessment function for students' writing ability is obtained through training on historical data, capable of processing time-series data and outputting standardized... The ability index network structure is designed as follows: input layer (64-dimensional), first hidden layer (128-dimensional), second hidden layer (64-dimensional), third hidden layer (32-dimensional), and output layer (4-dimensional). The input layer receives the four-dimensional score of the current essay and historical feature vectors, which are then input into the network after standardization. The first hidden layer uses the ReLU activation function to learn the nonlinear combination relationship of input features. The second hidden layer uses the Tanh activation function to further extract high-level abstract features. The third hidden layer uses the ReLU activation function for feature compression and key information extraction. The output layer uses the Sigmoid activation function to output the standardized ability index in four dimensions. The mean squared error loss function is used, and the Adam optimizer is used for parameter updates. The learning rate is set to 0.001, the batch size is 64, and the number of training epochs is 100. Dropout technology (dropout rate of 0.2) is used to prevent overfitting and improve the model's generalization ability. Through the nonlinear transformation of the hidden layers, complex ability association patterns are captured. The backpropagation algorithm is used to optimize the network parameters to improve prediction accuracy and output the student's comprehensive writing ability level assessment result, i.e., the ability level index.

[0043] Through the above steps, the system can accurately assess students' current writing ability level, providing a reliable basis for subsequent personalized teaching and group management.

[0044] Based on current ability level indicators and historical feature vectors, a clustering analysis algorithm is used to generate personalized tags for students, resulting in student profiles. The specific implementation steps are as follows: Feature vector construction involves integrating current ability level scores from various dimensions, extracting key features from historical feature vectors, standardizing and normalizing all features to construct complete student feature vectors, and preparing for cluster analysis. Clustering algorithm selection and parameter setting involve choosing K-means clustering as the primary analysis method, determining the number of clusters based on the total number of students and teaching needs, setting the convergence criteria and iteration count, and initializing cluster centers to ensure algorithm stability. Ability level label generation involves classifying students into levels based on their comprehensive ability scores, setting four levels: Excellent, Good, Average, and Needs Improvement, taking into account the students' grade level and learning stage characteristics to generate... Personalized ability level tags; learning type tag identification, analyzing students' learning habits and cognitive characteristics to identify visual, auditory, and hands-on learning types, and assigning the most suitable learning type tag to each student based on their writing performance and historical data; writing preference tag extraction, analyzing students' performance in different topics and genres to identify students' writing areas of expertise and style preferences, taking into account students' interests and knowledge background to generate personalized writing preference tags; development potential tag assessment, analyzing students' development potential based on their historical progress trajectory, taking into account students' learning attitudes and efforts, assessing students' room for improvement in different dimensions, and generating tags reflecting students' development potential.

[0045] Through the above steps, the system can generate personalized tags for each student, including dimensions such as ability level, learning type, writing preferences, and development potential, forming a complete student profile.

[0046] Based on student profiles and historical development trends, time series analysis is used to predict students' ability improvement paths, yielding prediction results. The specific implementation steps are as follows: Historical trend data processing involves collecting students' past ability assessment data and learning performance records, organizing the data chronologically, constructing a complete time series of ability development, identifying outliers and missing values, cleaning the data, and extracting key trend features and periodic patterns. Time series model construction utilizes the ARIMA model as the primary time series analysis method, determining model parameters through autocorrelation and partial autocorrelation functions, considering the impact of seasonal factors and external interventions, and constructing a predictive model suitable for individual student characteristics. Multi-dimensional ability prediction is performed, forecasting grammatical accuracy, semantic coherence, logical rigor, and creative expression, considering the interrelationships and correlations between dimensions, and combining student performance with... The system generates learning plans and teaching arrangements, producing predictive curves for ability improvement across various dimensions; it evaluates the effectiveness of interventions by analyzing the historical effects of different teaching methods and learning strategies, assessing the impact of personalized tutoring and group learning, considering changes in students' learning motivation and participation, and predicting the trajectory of ability development under various interventions; it calculates confidence intervals by calculating the uncertainty of predictions based on the volatility of historical data, setting prediction intervals at different confidence levels, considering the influence of model errors and external factors, and providing a reliability assessment for the prediction results; and it establishes a dynamic adjustment mechanism by establishing a real-time update mechanism for prediction results, adjusting the prediction model based on new learning data, optimizing the parameter settings of the prediction algorithm, and ensuring the accuracy and timeliness of the prediction results.

[0047] Through the above steps, the system can accurately predict students' ability improvement paths, providing a scientific basis for personalized teaching and learning planning.

[0048] It outputs personalized student writing ability profiles and current ability level assessment results, providing data support for group segmentation and differentiated teaching.

[0049] Furthermore, deep learning methods can be used to build more complex student modeling systems, which learn students' implicit features and complex patterns through neural networks, thereby providing more accurate personalized assessment and prediction results.

[0050] Step 3: Perform cluster analysis and similarity calculation on personalized student profiles, and generate an optimized learning group configuration scheme through ability level matching, learning needs analysis and complementarity assessment. Specifically, the following steps are included: Based on all personalized student profiles, a multi-objective optimization theory is used to construct the objective function for group partitioning, resulting in the optimization objective. The specific implementation steps are as follows: The similarity index within a group calculates the similarity between student ability profiles within the same group, using Euclidean distance to measure the similarity of student feature vectors. It also calculates the average similarity between all student pairs within the group and maps the similarity values ​​to a standard interval using Min-Max normalization. The inter-group difference index assesses the degree of feature difference between different groups, performing Z-score standardization on each feature dimension and using Mahalanobis distance to measure inter-group differences, ensuring significant feature differentiation between different groups. Finally, the group size balance analysis analyzes the distribution of member numbers in each group and calculates the group size balance. The variance and standard deviation of the group size are used to perform a logarithmic transformation on the group size to reduce the influence of extreme values ​​and assess the balance of the group size distribution. Weight coefficients are set according to teaching objectives and actual needs, balancing the importance of group cohesion and inter-group differentiation, considering factors such as teaching resource allocation and management convenience, and ensuring the rationality and interpretability of the weight coefficients. A multi-objective optimization function is constructed by weighting and integrating the three indicators of intra-group similarity, inter-group differences, and size balance to build a comprehensive group partitioning optimization objective function, setting optimization directions and constraints to provide clear optimization objectives for subsequent algorithm solutions.

[0051] Through the above steps, the system can establish a scientific and reasonable objective function for group partitioning, providing a quantitative optimization standard for intelligent group partitioning.

[0052] During the group partitioning process, reasonable constraints need to be set to ensure the practicality and operability of the partitioning results. The specific implementation steps are as follows: Group size constraints are set to determine the minimum and maximum number of members in each group based on teaching management needs. A reasonable range is set considering teacher guidance capabilities and classroom management efficiency to ensure each group has enough members for effective interaction, avoiding overly large groups that lead to management difficulties or undersized groups that negatively impact learning outcomes. Ability distribution constraints ensure that each group includes students of varying ability levels, avoiding extremes of over-concentration or over-dispersion of abilities. A reasonable range of ability differences within groups is set to ensure both complementarity and challenge. Learning preference balance constraints consider differences in students' learning styles and writing preferences, ensuring diversity and complementarity of learning preferences within groups and preventing excessive concentration of a particular learning preference in a single group, promoting mutual learning among students with different learning styles. Gender ratio and social factors constraints consider a reasonable gender ratio distribution, taking into account students' social relationships and willingness to cooperate, avoiding potentially unfavorable combinations that could affect learning outcomes, and promoting a positive group learning atmosphere. A dynamic adjustment constraint mechanism sets trigger conditions and frequency limits for group adjustments to ensure relative stability of group divisions and establishes a mechanism for handling abnormal situations, ensuring the flexibility and adaptability of the constraints.

[0053] By setting these constraints, the system can optimize the group division effect while ensuring that the division results meet actual teaching needs and management requirements.

[0054] Based on the objective function and constraints, a genetic algorithm is used to optimize the group partitioning and obtain the initial group configuration. The specific implementation steps are as follows: Population initialization: The population size is set at 100 individuals. Each individual is represented by a real number encoding to indicate the group partitioning scheme. The fitness function is used to evaluate the quality of the group configuration. Encoding scheme: Each student is represented by a real number in the range [0,1], which is mapped to a specific group through interval partitioning. Initialization strategy: The initial population is generated using a uniform random distribution to ensure that each group has a reasonable initial number of members. The initial population is also generated randomly to ensure that each individual meets the basic constraints. The feasibility of the initial population is checked and corrected to ensure that the group size is within the preset range. Fitness function design: A comprehensive... The fitness evaluation function quantifies and scores intra-group similarity, inter-group differences, and size balance. A constraint violation penalty is set to penalize individuals that do not meet the constraints, ensuring the fitness function accurately reflects the quality of group partitioning. The selection operation employs a tournament selection strategy for individual selection, and uses tournament selection, single-point crossover, and Gaussian mutation operations for population evolution. The tournament size is set to 5 individuals, randomly selected from the population to compete. Individuals with the best fitness values ​​are selected to advance to the next generation, with the individual with the highest fitness winning. This maintains population diversity and optimization direction, avoiding premature convergence. Crossover is implemented using a single-point crossover method, where genes are exchanged at random locations on chromosomes. The crossover probability is set to 0.8 to ensure that most individuals participate. Crossover point selection: A crossover point is randomly chosen within the coding length range. Gene exchange: The gene fragment after the crossover point is exchanged between two parent individuals, ensuring that the crossovered individuals still meet the constraints. Individuals that violate the constraints are repaired. The crossover operation generates a new group partitioning scheme, increasing population diversity. Mutation is handled using a Gaussian mutation strategy, with a mutation probability set to 0.1 for each gene locus. Independent mutation judgment is performed. Gaussian mutation: random noise following a Gaussian distribution is added to the original gene value. Mutation intensity: the standard deviation is set to 0.1 to control the magnitude of the mutation. Boundary handling: ensure that the gene value after mutation is still within the range of [0,1]. Local adjustments are made to individuals to increase population diversity, prevent the algorithm from getting stuck in local optima, ensure that the mutated individuals meet the constraints, and maintain the feasibility of group partitioning. Iterative evolution is used to solve the problem. A 200-generation iterative evolution process is performed. In each generation, selection, crossover, and mutation operations are performed. The fitness changes of the best individual are tracked, and the individual with the highest fitness is selected as the optimal group partitioning scheme.

[0055] Through iterative optimization using a genetic algorithm, the system can find the optimal group partitioning configuration that satisfies the requirements of multi-objective optimization.

[0056] Based on the initial group configuration and learning progress feedback, a dynamic adjustment algorithm is used to achieve dynamic optimization management of the groups, resulting in a dynamically optimized group configuration scheme. The learning progress feedback is generated by continuously collecting and analyzing students' essay correction data, including dynamic changes in students' various ability indicators, differences in learning effectiveness among group members, and trends in student ability improvement. The specific implementation steps are as follows: Learning progress feedback data collection establishes a learning progress tracking mechanism. After each essay is graded, the system automatically records students' current ability assessment data, compares this data with historical data to calculate the magnitude of ability changes, statistically analyzes the ability distribution of all members within the group and calculates the group's ability variance, and analyzes the trend of student ability changes across multiple essays to determine whether it is increasing, stable, or decreasing. This generates learning progress feedback information containing ability change data, group difference data, and trend data. Change monitoring and evaluation continuously monitor changes in students' writing abilities based on the learning progress feedback information. It analyzes the updating and evolution trends of student profile data, identifies students with significant ability changes (i.e., those whose ability indicators change by more than fifteen points or whose ability ranking changes by more than three places within the group), assesses the impact of these students' ability changes on the current group configuration, and calculates the ability within the group before and after adjustments. The degree of impact is quantified by the changes in variance and inter-group variability; adjustment trigger conditions are determined by setting thresholds and conditions for group adjustments, initiating adjustments when student ability changes exceed preset thresholds, considering overall group stability and adjustment costs to avoid the impact of frequent adjustments on learning continuity; local optimization algorithms are executed, using local search algorithms to adjust group members, prioritizing adjustments with minimal impact, maintaining the stability of most group structures, and ensuring that the adjusted configuration still meets constraints; adjustment scheme verification evaluates the impact of the adjustment scheme on group quality, verifies whether the new configuration is superior to the original configuration, checks the rationality and operability of the adjustment, and ensures that the adjustment meets teaching management requirements; configuration updates and records are implemented, updating group configurations and member affiliation information, recording adjustment history and effect evaluation data to provide a reference for subsequent dynamic adjustments.

[0057] Through a dynamic adjustment mechanism, the system can adapt to changes in students' abilities, maintaining the optimality and practicality of group configuration.

[0058] The system outputs optimized learning group configuration schemes and group characteristic descriptions, providing a foundation for the subsequent development of differentiated teaching strategies.

[0059] Furthermore, reinforcement learning methods can be used to optimize group adjustment strategies. By interacting with the environment, the optimal adjustment timing and magnitude can be learned, thereby achieving more intelligent dynamic group management.

[0060] Step 4: Perform strategy matching processing on the learning group configuration scheme and personalized student profiles, and generate differentiated grading strategies based on different ability levels and learning characteristics; Specifically, the following steps are included: Based on group configuration and personalized student profiles, statistical analysis methods are used to calculate the overall characteristics of each group, resulting in group feature vectors. The specific implementation steps are as follows: Data collection and organization: Collect competency profiles of all members within the group, organize the characteristic values ​​of each member across different dimensions, and statistically analyze the basic information and learning status of group members to ensure data completeness and accuracy; Average competency level calculation: Calculate the average competency level of the group across various competency dimensions, analyze the overall writing level distribution of the group, identify the group's strengths, and determine the group's overall competency positioning; Weakness identification: Statistically analyze the common problems and weaknesses of group members, analyze the frequency and severity of these problems, identify areas requiring focused improvement, and determine the main directions for improvement; Learning preference distribution analysis: Statistically analyze the distribution of different learning styles within the group, analyze members' writing preferences and habits, identify the group's dominant learning patterns, and assess the diversity of learning preferences; Development potential assessment: Analyze the competency improvement trends of group members, assess the overall development potential of the group, predict the group's possible performance under different strategies, and provide potential references for strategy formulation.

[0061] By analyzing group characteristics, the system can gain a comprehensive understanding of the overall characteristics of each group, providing accurate basic data for subsequent strategy matching.

[0062] Based on group feature vectors and a predefined strategy template library, a pattern matching algorithm is used to select a suitable grading strategy template to obtain an initial strategy scheme. The specific implementation steps are as follows: The strategy template library is constructed by establishing different types of strategy templates, including basic reinforcement, capability enhancement, and innovation expansion. For each template, the applicable group feature range and conditions are defined, and the core elements and adjustable parameters of the template are set to ensure the completeness and coverage of the template library. Feature similarity calculation involves calculating the matching degree between the group feature vector and each strategy template, using weighted Euclidean distance or cosine similarity for similarity evaluation, considering the importance weights of different feature dimensions, and generating a matching score between the group and each template. Optimal template selection involves choosing the most suitable strategy template based on the matching score, considering the applicability and feasibility of the template, handling cases where multiple template scores are close, and ensuring the rationality and effectiveness of the selection results. Initial strategy scheme generation involves generating an initial correction strategy based on the selected strategy template, setting the basic framework and key directions of the strategy, determining the areas of focus and evaluation criteria for correction, and providing a basic scheme for subsequent personalized adjustments.

[0063] Through strategy template matching, the system can select the most suitable basic framework for grading strategies for groups with different characteristics.

[0064] Based on the initial strategy and the individual student profile, a strategy adjustment algorithm is used to personalize the grading strategy. The specific implementation steps are as follows: Individual difference analysis calculates the degree of difference between individual student profiles and group average characteristics, uses multi-dimensional distance calculation methods to assess individual specificity, identifies students' unique performance in each ability dimension, and determines the necessity and intensity of personalized adjustments. Problem domain priority ranking analyzes the comprehensive semantic analysis results to identify students' main problem domains, ranks them according to the severity and difficulty of improvement, considers students' ability base and acceptance level, and determines the focus of correction and guidance priority. Personalized adjustment of strategy parameters adjusts the feedback detail coefficient according to individual differences, reallocates the weight allocation of guidance focus, selects improvement suggestion types suitable for student characteristics, and adjusts the style and expression of correction language. Strategy consistency verification verifies the internal logical consistency of the strategy through a rule engine, checks the rationality and feasibility of personalized adjustments, ensures that the strategy conforms to teaching logic and actual needs, and verifies the operability and effectiveness of the strategy. Personalized strategy generation integrates all adjustment parameters to generate the final personalized strategy, ensuring that the strategy reflects both group characteristics and individual needs, generates specific correction guidance principles and operational points, and provides clear strategy guidance for correction feedback generation.

[0065] By adjusting personalized strategies, the system can develop differentiated grading strategies for each student that both align with the characteristics of the group and reflect individual needs.

[0066] Based on personalized grading strategies and historical teaching effectiveness data, an effectiveness prediction model is used to verify the expected effectiveness of the strategies, resulting in strategy evaluation results. The specific implementation steps are as follows: Historical data collection and organization involves gathering historical teaching effectiveness data from similar student groups, organizing learning outcomes and improvements under different strategies, analyzing successful and unsuccessful cases of strategy implementation, and establishing reference benchmarks and evaluation standards for strategy effectiveness. Effectiveness prediction model construction involves building a strategy effectiveness prediction model based on historical data, considering factors such as student characteristics, strategy type, and implementation conditions. Machine learning methods are used to train the prediction model, and the accuracy and reliability of the model are verified. Strategy effectiveness simulation prediction involves inputting personalized strategies into the prediction model to simulate student performance under the strategy, assessing the improvement effect of the strategy on different ability dimensions, and analyzing the risks and uncertainties of strategy implementation. Strategy scientific validity assessment involves checking the theoretical basis and logical rationality of the strategy, evaluating its operability and implementation difficulty, analyzing the degree of matching between the strategy and student characteristics, and ensuring that the strategy conforms to the laws of education and teaching. A comprehensive judgment on strategy effectiveness is made by combining the prediction results and the scientific validity assessment to form a final judgment, quantifying the expected effect of the strategy, identifying the advantages and potential problems of the strategy, and providing suggestions for further optimization of the strategy.

[0067] By validating the effectiveness of the strategy, the system can ensure that the generated personalized grading strategy is scientific and effective.

[0068] It outputs differentiated grading strategies tailored to the characteristics of specific students and groups, providing guiding principles for the generation of intelligent grading feedback.

[0069] Furthermore, a collaborative mechanism for strategy generation can be established using a multi-agent system. By collaborating with different expert agents, more comprehensive and professional correction strategies can be generated, thereby improving the quality and applicability of the strategies.

[0070] Step 5: The differentiated correction strategy and comprehensive semantic analysis results are processed through a large language model. A detailed correction feedback report is generated through problem identification, improvement suggestions and personalized guidance. Specifically, the following steps are included: Based on semantic analysis results and personalized correction strategies, a problem identification algorithm is used to locate specific problems in the essays and classify and label them according to severity, type, and difficulty of improvement. The specific implementation steps are as follows: Semantic problem scanning: Based on semantic analysis results, the system scans for semantic-level problems in essays, such as grammatical errors, inappropriate word choice, and logical inconsistencies, and establishes a preliminary problem list. Strategy-oriented screening: Based on the key areas of focus of personalized correction strategies, the preliminary problem list is screened and prioritized to ensure that problem identification aligns with students' individual needs. Problem severity assessment: Using a rule engine and machine learning model, identified problems are graded according to their impact on essay quality, into three levels: severe, moderate, and minor. Problem type classification: Identified problems are classified into different types such as grammatical, logical, expressive, and structural problems to facilitate targeted subsequent processing. Improvement difficulty labeling: Based on the complexity of the problem and the student's ability level, each problem is labeled with an improvement difficulty level, guiding the level of detail and method of feedback.

[0071] Based on the problem identification results and the language generation model, a hierarchical generation strategy is adopted to construct personalized feedback content, including overall evaluation, specific problem correction, improvement suggestions, and affirmation of excellent expressions. The specific implementation steps are as follows: The feedback framework is constructed based on student profiles and a list of questions. Appropriate feedback templates are selected to build a framework structure that includes overall evaluation, problem analysis, improvement suggestions, and positive reinforcement. Overall evaluation generation utilizes a large language model to analyze the overall quality of the essays, generating comprehensive evaluation content covering aspects such as theme expression, structural organization, and language use. Specific problem correction involves generating detailed explanations, error analysis, and guidance on correct expression for each identified problem. Personalized improvement suggestions are generated based on students' abilities and learning preferences, including specific revision methods, practice suggestions, and recommended reference resources. Positive reinforcement of excellent expressions is achieved by identifying highlights and excellent expressions in the essays, generating positive feedback to enhance students' writing confidence and learning motivation. Finally, feedback language is optimized by adjusting the language expression of the feedback content according to students' age characteristics and comprehension abilities, ensuring that the feedback is easy to understand and accept.

[0072] Based on the generated feedback content, quality assessment and optimization algorithms are used to improve the accuracy and readability of the feedback, ensuring that the feedback quality meets the expected standards. The specific implementation steps are as follows: The feedback process includes several key steps: Logical consistency check: Using logical reasoning algorithms to examine the logical relationships within the feedback content, ensuring that the evaluation conclusions align with the specific problem analysis and avoiding contradictory statements; Clarity assessment: Analyzing the clarity of language in the feedback content using natural language processing technology, identifying vague expressions, redundant information, and difficult sentence structures, and making corresponding optimizations; Guiding value verification: Evaluating the practicality and operability of the feedback content, ensuring that improvement suggestions are specific and clear, enabling students to effectively revise their essays based on the feedback; Emotional tone adjustment: Analyzing the emotional tone of the feedback content, ensuring that critical opinions are appropriately expressed, encouraging content is positive, and the overall feedback is constructive and motivating; Personalization suitability test: Verifying the degree of matching between the feedback content and the student's individual characteristics, ensuring that the feedback method and content depth are appropriate for the student's cognitive level and comprehension ability.

[0073] Based on the optimized feedback content, visualization technology is used to generate an intuitive feedback display interface, enhancing the comprehensibility and appeal of the feedback. The specific implementation steps are as follows: Layout template selection: Based on the structure of the feedback content and students' usage habits, select an appropriate page layout template to ensure clear information hierarchy and highlight key points; Color-coded labeling design: Use different colors to label different types of questions and evaluations, such as red for errors, yellow for suggestions, and green for advantages, to help students quickly identify key information; Chart element generation: Transform quantitative analysis results into intuitive chart formats, such as ability radar charts, progress trend charts, and problem distribution pie charts, to enhance the visualization of data; Interactive function design: Design interactive functions such as click to expand details, hover to display prompts, and links to related resources to improve user experience and information acquisition efficiency; Multimedia content integration: Integrate multimedia content such as audio explanations, video examples, and animation demonstrations as needed to enrich feedback formats and improve learning outcomes; Responsive adaptation: Ensure that the feedback interface displays well on different devices and screen sizes, providing a consistent user experience.

[0074] It outputs personalized, detailed feedback reports, providing specific guidance and suggestions for improving students' writing.

[0075] Furthermore, interactive feedback systems can be created using dialogic generation methods, allowing students to ask questions and gain a deeper understanding of the feedback content, thereby providing a better learning experience and guidance.

[0076] Step 6: Based on the optimized learning group configuration scheme and detailed feedback reports, design a collaboration scheme, and organize group collaborative learning activities through task allocation, peer review mechanism and collaboration guidance. Specifically, the following steps are included: Based on group configuration and members' feedback, a task design algorithm is used to create suitable collaborative learning tasks, including peer review, problem discussion, sharing of writing skills, and collaborative creation. The specific implementation steps are as follows: The process involves several key steps: First, the task type is determined based on the group members' skill levels and learning objectives. Suitable collaborative tasks include peer review, group discussions, collaborative writing, and experience sharing. Second, the difficulty level is matched to the overall group skill level and individual differences, designing collaborative tasks of appropriate difficulty to ensure effective participation and benefit for each member. Third, a detailed task execution process is designed, including task allocation, time allocation, communication methods, and results presentation, ensuring an orderly collaborative process. Fourth, clear evaluation criteria and incentive mechanisms are established, including participation evaluation, contribution quality assessment, and collaboration effectiveness evaluation, to promote active participation. Finally, necessary learning resources and technical support, such as reference materials, discussion platforms, and collaborative tools, are provided to ensure the smooth implementation of the tasks.

[0077] Based on the collaborative task scheme, a peer evaluation mechanism is established within the group using a peer review algorithm. This mechanism promotes mutual learning among students through anonymous peer review, cross-review, and group discussion. The specific implementation steps are as follows: The peer review process is structured as follows: Clear rules and guiding principles are established, including evaluation criteria, procedures, anonymity mechanisms, and fairness, to ensure a standardized and orderly process. A multi-dimensional evaluation system is designed, covering content quality, language expression, logical structure, and innovation, guiding students to conduct comprehensive and in-depth evaluations. An anonymous assignment mechanism uses a random anonymous assignment algorithm to ensure each student's work receives evaluations from multiple peers while protecting student privacy and minimizing the impact of interpersonal relationships. Cross-validation is implemented, allowing the same work to be evaluated by multiple evaluators, improving the reliability and objectivity of the results through comparative analysis. Feedback is integrated and analyzed to identify common problems and differing viewpoints, providing comprehensive improvement suggestions for the evaluated student. A quality monitoring mechanism is established to identify and address issues such as malicious or perfunctory evaluations, ensuring the effectiveness and constructiveness of the peer review activity.

[0078] Based on group characteristics and learning needs, a recommendation algorithm is used to recommend suitable learning resources to each group, including writing materials, skills guidance, excellent model essays, and practice question banks. The specific implementation steps are as follows: Needs analysis and identification: Based on the ability level, learning preferences, and weaknesses of group members, analyze the specific learning needs and resource requirements of each group; Resource library construction: Establish a rich learning resource library, including various types of learning resources such as categorized writing materials, skill guidance documents, excellent sample essays, practice question banks, and video tutorials; Feature matching algorithm: Use content similarity calculation and collaborative filtering algorithms to match group features with resource features to identify the most suitable learning resources; Personalized recommendation generation: Based on the matching results, generate personalized resource recommendation lists for different groups to ensure that the recommended content highly matches the group's needs; Recommendation quality evaluation: Evaluate the quality and applicability of recommended resources through user feedback and usage effect data, and continuously optimize the recommendation algorithm and resource library; Dynamic update mechanism: Establish a dynamic update mechanism for resource recommendations, and adjust and update the recommended content in a timely manner according to the group's learning progress and changes in needs.

[0079] Based on the implementation of collaborative learning activities and student feedback, an effectiveness evaluation method is used to analyze the results of collaborative learning. The value of collaborative learning is evaluated through a combination of quantitative indicators and qualitative analysis. The specific implementation steps are as follows: Data collection and organization involves gathering various data related to collaborative learning activities, including quantitative data such as participation statistics, task completion status, peer assessment quality, and learning outcome presentations, as well as qualitative data such as student feedback and teacher observations. Quantitative indicator calculation involves calculating key quantitative evaluation indicators, such as average participation, task completion rate, peer assessment accuracy, learning progress, and collaboration efficiency, to form an objective data foundation. Qualitative analysis and processing involves content analysis and theme extraction of qualitative data such as student feedback, teacher observations, and learning experiences to identify the strengths and weaknesses of collaborative learning. Comprehensive effect evaluation involves integrating quantitative indicators and qualitative analysis results to assess the overall effectiveness of collaborative learning from multiple dimensions, including learning outcomes, participation experience, skill improvement, and collaboration ability. Comparative analysis and verification involves comparing and analyzing the collaborative learning method with traditional teaching methods or other collaborative models to verify its effectiveness and superiority. Improvement suggestion generation involves identifying problems and areas for improvement in collaborative learning based on the evaluation results, and generating specific optimization suggestions and improvement plans.

[0080] Output group collaborative learning activity plans and learning effect evaluation results to provide data support for teaching optimization.

[0081] Furthermore, gamified learning methods can be used to enhance the fun and engagement of collaborative learning, and mechanisms such as points rewards, level upgrades, and team competitions can be used to stimulate students' learning motivation and enthusiasm for collaboration.

[0082] Step 7: Perform data statistics and effect analysis on the execution results of the group collaborative learning activity plan, and generate a system performance evaluation report and optimization and improvement plan; Specifically, the following steps are included: Based on learning outcome data and system performance indicators, a multi-dimensional analysis method is used to evaluate teaching effectiveness, analyzing aspects such as improvement in students' writing skills, system user satisfaction, changes in teacher work efficiency, and the degree of improvement in teaching quality. The specific implementation steps are as follows: Data integration and processing integrates data from multiple channels, including learning outcome assessments, system operation logs, and user feedback surveys, to establish a unified foundation for data analysis. Student ability improvement analysis compares and contrasts changes in students' writing skills before and after using the system, assessing improvements in grammatical accuracy, logical clarity, and expressive richness. System user satisfaction assessment analyzes student and teacher satisfaction levels regarding system functionality, interface design, and user experience based on user surveys and usage behavior data. Teacher work efficiency analysis evaluates the system's impact on teachers' grading work, including time savings, reduced workload, and improved teaching quality. Teaching quality improvement assessment evaluates the system's overall improvement in teaching quality from the perspectives of achieving teaching objectives, student participation, and knowledge mastery. A comprehensive effectiveness report is generated by synthesizing the analysis results from all dimensions, providing data support for system optimization.

[0083] Based on the performance analysis report, a problem diagnosis algorithm is used to identify deficiencies and problems in the system's operation. Data mining and anomaly detection techniques are employed to discover the system's weaknesses and areas for improvement. The specific implementation steps are as follows: Anomaly identification employs statistical analysis and machine learning methods to identify abnormal patterns and deviations from normal ranges in system operation data, pinpointing potential problem areas. Performance benchmark comparison compares the system's current performance metrics with preset benchmark standards to identify substandard functional modules and performance bottlenecks. User feedback analysis delves into user feedback and complaints to identify pain points in user experience and deficiencies in system functionality. System log mining utilizes log data mining techniques to analyze error patterns, performance bottlenecks, and abnormal resource usage during system operation. Root cause analysis locates the root cause of problems using causal analysis methods, distinguishing between surface phenomena and deep-seated issues to ensure accurate problem diagnosis. Problem prioritization prioritizes identified problems based on their impact, severity, and difficulty of resolution, providing guidance for subsequent optimization.

[0084] Based on the problem diagnosis results, an optimization decision-making algorithm is used to formulate a system improvement plan, including algorithm parameter adjustment, functional module improvement, user experience optimization, and system architecture upgrade. The specific implementation steps are as follows: Improvement goals include: setting clear optimization objectives and expected effects based on problem diagnosis results and system development needs to ensure the correctness of the improvement direction; designing solutions for each identified problem, including technical solutions, implementation steps, resource requirements, and timelines; assessing available human, material, and time constraints to ensure the feasibility and implementability of the optimization solutions; prioritizing solutions based on the urgency of the problem, difficulty of resolution, resource requirements, and expected effects; analyzing risks and side effects of each optimization solution and developing corresponding risk control and contingency plans; and developing a detailed implementation plan, including phase division, milestone settings, division of responsibilities, and progress monitoring mechanisms.

[0085] Based on the optimization and improvement scheme, an adaptive optimization mechanism for the system is established using a continuous improvement method. This mechanism achieves continuous performance improvement through feedback loops and iterative optimization. The specific implementation steps are as follows: A feedback loop is established to monitor system performance and collect user feedback in a cyclical manner, ensuring timely identification of new problems and improvement needs. An iterative optimization process is developed, including standardized steps such as problem identification, solution design, implementation verification, and effect evaluation, forming a closed-loop management system. An adaptive adjustment mechanism is designed to automatically optimize key parameters based on usage and environmental changes. A robust version management and rollback mechanism is established to ensure the controllability and security of the system optimization process. A long-term effect tracking and evaluation mechanism is established to continuously monitor the actual effectiveness and sustainability of optimization measures. Finally, knowledge is accumulated and preserved, summarizing lessons learned and best practices from the optimization process to form a system optimization knowledge base to guide future improvements.

[0086] Output system performance evaluation reports and optimization and improvement plans to provide guidance for the continuous development and improvement of the system.

[0087] The essay intelligent grading and group management system based on a large language model is used to execute the aforementioned essay intelligent grading and group management method based on a large language model, such as... Figure 2 As shown, it includes: The multi-dimensional semantic deep analysis module is used to obtain the original essay text submitted by students, perform multi-dimensional semantic deep analysis on the original essay text, and generate comprehensive semantic analysis results. The personalized student profile building module is used to extract features and assess abilities from the comprehensive semantic analysis results using statistical analysis methods to build personalized student profiles. The intelligent group division module is used to perform cluster analysis and similarity calculation on personalized student profiles, and generate an optimized learning group configuration scheme through ability level matching, learning needs analysis and complementarity assessment. The differentiated grading strategy generation module is used to perform strategy matching processing on learning group configuration schemes and personalized student profiles, and generate differentiated grading strategies based on different ability levels and learning characteristics. The intelligent correction feedback generation module is used to comprehensively process differentiated correction strategies and comprehensive semantic analysis results through a large language model, and generate detailed correction feedback reports through problem identification, improvement suggestions and personalized guidance; The group collaborative learning organization module is used to design collaborative learning plans based on optimized learning group configuration schemes and detailed feedback reports. It organizes group collaborative learning activities through task allocation, peer review mechanisms, and collaborative guidance. The teaching effectiveness evaluation and optimization module is used to perform data statistics and effect analysis on the execution results of group collaborative learning activities, and generate system performance evaluation reports and optimization and improvement plans.

[0088] In one embodiment of the present invention, a specific example is provided: A 180-day field test was conducted at a key middle school. During the test, an intelligent essay grading and group management system based on a large language model was deployed, covering 36 classes from grades 7 to 9. The test area covered the entire middle school teaching area, with a total of 1,620 students and 18 Chinese language teachers participating. During the system's operation, more than 12,000 essays were processed, 36,000 personalized feedback reports were generated, and 480 group collaborative learning activities were organized.

[0089] Table 1 shows an example of semantic analysis data from student essays: Table 1: Example of semantic analysis data for student essays;

[0090] The group configuration and adjustment data examples are shown in Table 2: Table 2: Example of group configuration and adjustment data;

[0091] like Figure 3 As shown, this invention demonstrates its significant advantages in processing efficiency. Compared to the traditional 18.5 minutes for manual grading, this invention reduces the processing time for a single essay to 2.8 minutes, representing an 85% increase in efficiency. Compared to the traditional 8.2 minutes of automated grading systems, this invention improves processing speed by 66%, primarily due to the optimized design of its layered asynchronous processing architecture and intelligent task scheduling mechanism.

[0092] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for intelligent essay grading and group management based on a large language model, characterized in that, Includes the following steps: The system obtains the original essay texts submitted by students, performs multi-dimensional semantic in-depth analysis of the original essay texts, including scoring in four dimensions: grammatical accuracy, semantic coherence, logical rigor, and innovative expression, and generates comprehensive semantic analysis results. Statistical analysis methods are used to extract features and assess abilities from the comprehensive semantic analysis results to construct personalized student profiles. Cluster analysis and similarity calculation are performed on personalized student profiles. Through ability level matching, learning needs analysis and complementarity assessment, an optimized learning group configuration scheme is generated. The learning group configuration scheme and personalized student profile are matched with strategies to generate differentiated grading strategies based on different ability levels and learning characteristics. By comprehensively processing differentiated correction strategies and comprehensive semantic analysis results through a large language model, detailed correction feedback reports are generated through problem identification, improvement suggestions, and personalized guidance. Based on an optimized learning group configuration scheme and detailed feedback reports, a collaborative learning scheme is designed, which organizes group collaborative learning activities through task allocation, peer review mechanism and collaborative guidance. The system performs statistical analysis and effectiveness analysis on the execution results of group collaborative learning activities, and generates system performance evaluation reports and optimization and improvement plans. The steps for generating comprehensive semantic analysis results include: Structured text data is generated based on the original essay text. A large language model is used to compare standard grammar rules and contextual semantics to identify vocabulary collocation errors, syntactic structure problems, and improper use of punctuation marks, and a grammatical accuracy score is obtained. Based on grammatical accuracy scores and structured text data, a large language model is used to calculate semantic similarity between adjacent sentences, thematic consistency between paragraphs, and logical fluency of the whole text, to obtain a semantic coherence score. Based on semantic coherence scoring, a logical reasoning algorithm is used to analyze viewpoints, supporting arguments, and reasoning processes, and to evaluate the completeness, persuasiveness, and logical consistency of the argument, resulting in a logical rigor score. Based on logical rigor scoring and original text content, an innovative evaluation algorithm is used to analyze the degree of expressive innovation in essays. By identifying novel viewpoints, unique expressions, creative metaphors, and original insights, an innovative expressive score is obtained.

2. The intelligent essay grading and group management method based on a large language model according to claim 1, characterized in that, The steps for constructing personalized student profiles include: Based on students' historical essay data, statistical analysis methods were used to extract writing style features, error pattern features, theme preference features, and ability change trend features to obtain historical feature vectors. Based on historical feature vectors and comprehensive semantic analysis results, a capability assessment model is used to process historical data through time-weighted averaging and then weighted and fused with current performance to obtain capability level indicators. Based on ability level indicators and historical feature vectors, clustering analysis algorithms are used to generate ability level labels, learning type labels, writing preference labels, and development potential labels, resulting in personalized student profiles.

3. The intelligent essay grading and group management method based on a large language model according to claim 1, characterized in that, The steps for generating the optimized learning group configuration scheme include: Based on all personalized student profiles, a multi-objective optimization theory is used to construct an objective function that comprehensively considers the similarity of abilities among members within a group, the differences between groups, and the balance of group size, thus obtaining the optimization objective. Based on actual teaching needs and system resource constraints, a constraint optimization method is used to set group size limits, ability distribution requirements, and teacher allocation constraints, resulting in a set of constraint conditions. Based on the set of optimization objectives and constraints, a genetic algorithm is used to search for the optimal group partitioning scheme through population evolution, selection, crossover and mutation operations to obtain the initial group configuration. Based on the initial group configuration and learning progress feedback, a dynamic adjustment algorithm is used to monitor the differences in learning performance within the group and trigger the reassignment of group members according to a preset threshold, resulting in a dynamically optimized group configuration scheme.

4. The intelligent essay grading and group management method based on a large language model according to claim 1, characterized in that, The steps for generating the differentiated correction strategy include: Based on group configuration and personalized student profiles, statistical analysis methods are used to calculate the average ability level, ability distribution characteristics, learning preference distribution and development potential of each group, and the group feature vector is obtained. Based on group feature vectors and a predefined strategy template library, a pattern matching algorithm is used to select basic reinforcement, capability enhancement, or innovative extension grading strategy templates to obtain an initial strategy scheme. Based on the initial strategy and specific student profiles, a strategy adjustment algorithm is used to analyze the differences between the individual profiles and the group average characteristics, and adjust the key areas of focus, the level of feedback detail, and the direction of guidance to obtain a personalized grading strategy. Based on personalized grading strategies and teaching schedules, a time-series programming algorithm is used to schedule grading time, feedback frequency, and follow-up plans, resulting in differentiated grading strategies.

5. The intelligent essay grading and group management method based on a large language model according to claim 1, characterized in that, The steps for generating a detailed feedback report include: Based on personalized correction strategies and semantic analysis results, a problem identification algorithm is used to locate specific problems in the essays and classify and label them according to severity, type and difficulty of improvement to obtain a problem list; Based on the problem list and language generation model, a hierarchical generation strategy is adopted to construct personalized feedback content that includes overall evaluation, specific problem correction, improvement suggestions, and affirmation of excellent expressions; Based on the generated feedback content, quality assessment and optimization algorithms are used to check logical consistency, evaluate clarity of expression, verify guiding value, and adjust sentiment, thereby generating optimized feedback content. Based on the optimized feedback content, visualization technology is used to generate an intuitive feedback display interface that includes layout templates, color annotations, chart elements, and interactive functions, resulting in a detailed feedback report.

6. The intelligent essay grading and group management method based on a large language model according to claim 1, characterized in that, The proposed collaborative learning activity plan for group organizations includes: Based on group configuration and members' grading results, a task design algorithm is used to determine the type of collaborative task, match the task difficulty, design the task process, formulate evaluation criteria, and configure resource support according to the ability level and learning objectives of group members, so as to obtain a collaborative task plan. Based on the collaborative task scheme, a peer evaluation mechanism is established within the group using a peer evaluation algorithm. This involves formulating peer evaluation rules, designing evaluation dimensions, implementing anonymous allocation, setting cross-validation, integrating feedback information, and monitoring the quality of peer evaluation. Based on group characteristics and learning needs, a recommendation algorithm is used to recommend suitable learning resources to each group. This involves identifying learning needs, building a resource library, matching resource features, generating personalized recommendations, evaluating recommendation quality, and dynamically updating resources. Based on the implementation of collaborative learning activities and student feedback, an effectiveness evaluation method is used to analyze the effectiveness of collaborative learning. This involves collecting quantitative and qualitative data, calculating quantitative indicators, conducting qualitative analysis, performing comprehensive effectiveness evaluation, comparative analysis and verification, and generating improvement suggestions to obtain a collaborative learning activity plan.

7. The intelligent essay grading and group management method based on a large language model according to claim 1, characterized in that, The steps for generating the system performance evaluation report and optimization and improvement plan include: Based on learning outcome data and system operation indicators, a multi-dimensional analysis method is used to evaluate teaching effectiveness. By integrating data from multiple channels, analyzing student ability improvement, assessing system user satisfaction, analyzing teacher work efficiency, and evaluating teaching quality improvement, an effectiveness analysis report is generated. Based on the performance analysis report, a problem diagnosis algorithm is used to identify deficiencies and problems in the system operation. By identifying abnormal data, comparing performance benchmarks, analyzing user feedback, mining system logs, locating root causes, and prioritizing problems, the problem diagnosis results are obtained. Based on the problem diagnosis results, an optimization decision algorithm is used to formulate a system improvement plan. The optimization improvement plan is obtained through setting improvement goals, designing solutions, evaluating resource constraints, prioritizing solutions, analyzing risks, and formulating implementation plans. Based on the optimization and improvement plan, an adaptive optimization mechanism for the system is established using a continuous improvement approach. This mechanism involves feedback loop establishment, iterative optimization process, adaptive adjustment mechanism, version management control, effect tracking and evaluation, and knowledge accumulation and sedimentation, resulting in a system performance evaluation report.

8. The intelligent essay grading and group management method based on a large language model according to claim 1, characterized in that, The process of integrating differentiated grading strategies and comprehensive semantic analysis results through a large language model includes: the large language model is a pre-trained language model based on the Transformer architecture, which is trained on a large-scale Chinese essay corpus for domain-adaptive fine-tuning and has the ability of grammatical analysis, semantic understanding, logical reasoning and innovation evaluation. The competency assessment model adopts a multi-layer neural network structure and obtains a quantitative assessment function of students' writing ability through training on historical data. It can process time-series data and output standardized competency indicators. The genetic algorithm uses real-number encoding to represent the group partitioning scheme, evaluates the quality of group configuration through the fitness function, and uses tournament selection, single-point crossover, and Gaussian mutation operations for population evolution.

9. An intelligent essay grading and group management system based on a large language model, characterized in that: The method for implementing the intelligent essay grading and group management based on a large language model as described in any one of claims 1-8 includes: The multi-dimensional semantic deep analysis module is used to obtain the original essay text submitted by students, perform multi-dimensional semantic deep analysis on the original essay text, including scoring in four dimensions: grammatical accuracy, semantic coherence, logical rigor and innovative expression, and generate comprehensive semantic analysis results. The personalized student profile building module is used to extract features and assess abilities from the comprehensive semantic analysis results using statistical analysis methods to build personalized student profiles. The intelligent group division module is used to perform cluster analysis and similarity calculation on personalized student profiles, and generate an optimized learning group configuration scheme through ability level matching, learning needs analysis and complementarity assessment. The differentiated grading strategy generation module is used to perform strategy matching processing on learning group configuration schemes and personalized student profiles, and generate differentiated grading strategies based on different ability levels and learning characteristics. The intelligent correction feedback generation module is used to comprehensively process differentiated correction strategies and comprehensive semantic analysis results through a large language model, and generate detailed correction feedback reports through problem identification, improvement suggestions and personalized guidance; The group collaborative learning organization module is used to design collaborative learning plans based on optimized learning group configuration schemes and detailed feedback reports. It organizes group collaborative learning activities through task allocation, peer review mechanisms, and collaborative guidance. The teaching effectiveness evaluation and optimization module is used to perform data statistics and effect analysis on the execution results of group collaborative learning activities, and generate system performance evaluation reports and optimization and improvement plans.

Images