Dynamic learning planning teaching system and method based on multi-model agent cooperation

Abstract

The embodiment of the invention provides a dynamic learning planning teaching system and method based on multi-model agent cooperation, and relates to the technical field of intelligent education. According to the system, description information input by a user is obtained through a dynamic test question searching module, and a test question set is retrieved in a test question bank. And identifying wrong question data through a test paper dynamic analysis module, generating a structured wrong question bank through a wrong question mastering degree analysis module, and dynamically adjusting review interval parameters according to an interval repetition algorithm based on memory cards and answer data in the structured wrong question bank, so that a learning planning module generates learning task planning data. According to the system, a complete closed-loop teaching mode of question finding, wrong question, mastery degree analysis and question exercising can be constructed through a large language model, retrieval enhancement generation, YOLO image segmentation, an interval repetition algorithm and the like, dynamic test question searching, wrong question reason analysis, mastery degree tracking and personalized learning planning are achieved, and the accuracy of a recommendation result is improved.

Description

Technical Field

[0001] This application relates to the field of intelligent education technology, and in particular to a dynamic learning planning teaching system and method based on multi-model intelligent agent collaboration. Background Technology

[0002] A smart teaching system, also known as an intelligent teaching system, is a digital education platform that uses data-driven approaches and artificial intelligence (AI) analysis as its core to achieve a complete digital education process encompassing teaching, learning, management, and evaluation. A smart teaching system can integrate functions such as end-to-end data collection and intelligent analysis, personalized recommendations, interactive teaching, automated teaching assistance, and intelligent decision-making, helping teachers, students, and administrators achieve a scientific and intelligent teaching approach.

[0003] For example, a smart teaching system can be used for error management and learning planning. This system can extract similar practice questions from a database based on error data from assignments, quizzes, and exams, according to pre-defined relationships, and generate recommended error questions based on these similarities. Simultaneously, it can integrate error data to create personalized learning paths and generate learning planning information.

[0004] However, in terms of error management and learning planning, the intelligent teaching system fails to consider how students' mastery of incorrect questions changes over time, leading to a mismatch between recommended practice questions and actual needs, resulting in inaccurate error recommendation results. Furthermore, similar question recommendations are based solely on pre-defined database bindings, failing to provide personalized test recommendations. In terms of error mastery, the system lacks dynamic tracking and analysis, failing to provide precise learning guidance based on students' learning habits and weaknesses in specific knowledge points, thus resulting in inaccurate recommended learning planning information. Summary of the Invention

[0005] In view of this, embodiments of this application provide a dynamic learning planning teaching system and method based on multi-model intelligent agent collaboration to solve the problem of inaccurate recommendation results in the management of wrong questions and learning planning in intelligent teaching systems.

[0006] According to a first aspect of this application, a dynamic learning planning teaching system based on multi-model intelligent agent collaboration is provided, the system comprising: The dynamic test question search module is configured to obtain descriptive information input by the user and retrieve a set of test questions from the test question bank based on the semantic features of the descriptive information. The test paper dynamic analysis module is configured to acquire answer data uploaded by users based on the test question set, and to identify incorrect question data from the answer data. The incorrect question data includes a set of incorrect questions and the reason information for the incorrect questions corresponding to the set of incorrect questions. The module for analyzing the level of mastery of incorrect questions is configured to generate a structured incorrect question database based on the incorrect question data. The structured incorrect question database includes memory cards made for at least one incorrect question from a test paper. Based on the memory cards and the answer data, the review interval parameter is dynamically adjusted according to the interval repetition algorithm. The interval repetition algorithm is used to dynamically adjust the review interval time based on the forgetting curve. The learning planning module is configured to generate learning task plan data based on the structured error database and the review interval parameters.

[0007] In some embodiments, the dynamic test question search module is further configured to: Obtain candidate test question data, which includes multiple candidate test questions and their answer information; Semantic analysis is performed on the candidate test question data using a large language model to extract key information from the candidate test question data; the key information includes knowledge point information, test point information, difficulty level information, and question type information; Type labels are generated based on the key information, and the type labels are used to characterize the structured description of the candidate test questions; The type label is set to the candidate test questions, and the test question bank is constructed based on the candidate test questions with the type label set.

[0008] In some embodiments, the dynamic question search module, which constructs the question bank based on the candidate questions with the set type tags, is further configured to: Detailed text descriptions are generated by performing natural language processing on the candidate test question data; The candidate test questions are vectorized to obtain test question vectors; the test question vectors are used to capture the semantic similarity between test questions and form a structured corpus. The question bank is constructed based on the question vectors, the detailed text descriptions, and the candidate question data.

[0009] In some embodiments, the dynamic question search module is further configured to retrieve a set of questions from the question bank based on the semantic features of the description information. The description information is vectorized using an embedding model to obtain an input vector; The nearest neighbor algorithm is used to calculate the vector similarity between the input vector and multiple question vectors in the question bank; Target questions are extracted from the question bank based on the vector similarity, wherein the target questions are the candidate questions whose vector similarity is greater than or equal to the similarity threshold; The test question set is generated based on the target test questions.

[0010] In some embodiments, the test paper dynamic analysis module is further configured to identify incorrect answer data from the answer data. The answer data is segmented into multiple question regions; during the segmentation process, the detection head is dynamically adjusted to adapt to different test paper formats. Using YOLO object detection, the correct answer areas and incorrect answer areas are marked from multiple question areas; Based on the incorrect answer area, generate incorrect question tags, and use the incorrect question tags to mark the incorrect questions on the test paper corresponding to the incorrect answer area, to obtain the incorrect question set; The problem-solving agent extracts relevant information from the set of incorrect questions. This relevant information includes core knowledge points, test points, and summary information. The summary information includes a summary of the question and the solution approach. The error reason information is generated based on the relevant information and the answer data.

[0011] In some embodiments, the error mastery analysis module is further configured to: The system acquires incorrect question data from multiple data sources; the incorrect question data includes incorrect questions from multiple test papers. Memory association information is extracted from the incorrect question data, and the memory association information includes question stem information, answer information, explanation and error analysis information; Memory cards for the incorrect questions on the test paper are created based on the memory association information. The structured error database is generated by aggregating multiple memory cards.

[0012] In some embodiments, the error mastery analysis module is further configured to: Receive feedback information input by the user based on the memory card, the feedback information including a level record for characterizing mastery; Based on the feedback information, memory habit information is generated, which includes a first habit feature determined by the forgetting curve in the interval repetition algorithm and a second habit feature obtained by analyzing historical answer data through a large language model. The review interval parameter is dynamically adjusted based on the memory habit information, and the review interval parameter includes the review interval time; The mastery level of the incorrect questions on the test paper is set according to the feedback information and the review interval parameter; the mastery level is used to assess the learning progress by quantifying the degree of mastery based on the target level of memory half-life.

[0013] In some embodiments, the learning planning module is further configured to: Collect user browsing information and generate learning habit corpus based on the user browsing information; the user browsing information includes the frequency and time of the user browsing the memory cards; By analyzing the mastery curve and error reason information of each wrong question in the test paper, we can mine the data on users' mastery of knowledge points and their thinking defects. By combining the learning habit corpus, the knowledge point mastery corpus, and the thinking defect corpus, the learning task plan data within the preset period is generated; Receive dynamic feedback from users regarding the learning task plan data, and adjust the learning task plan data based on the dynamic feedback.

[0014] In some embodiments, the dynamic test question search module is further configured to: The resource generation agent is invoked, which is a resource generation tool that combines a large language model with a model context protocol. The resources are used to generate an intelligent agent that extracts contextual information from the description information; Modify the semantic features of the description information based on the context information; Retrieve a set of test questions from the test question bank based on the modified semantic features; Teaching resources are generated based on the set of test questions, and the teaching resources include lesson plan data and test paper data associated with the set of test questions.

[0015] According to a second aspect of this application, a dynamic learning planning teaching method based on multi-model agent collaboration is provided, applied to the system described in the first aspect, the method comprising: Obtain descriptive information input by the user, and retrieve a set of test questions from the test question bank based on the semantic features of the descriptive information; The system acquires answer data uploaded by users based on the question set, and identifies incorrect question data from the answer data. The incorrect question data includes a set of incorrect questions and information on the reasons for the incorrect questions corresponding to the set of incorrect questions. A structured error database is generated based on the error data, and the structured error database includes memory cards made for at least one error question from a test paper; Based on the memory cards and the answer data, the review interval parameters are dynamically adjusted according to the interval repetition algorithm; the interval repetition algorithm is used to dynamically adjust the review interval time based on the forgetting curve. Learning task plan data is generated based on the structured error database and the review interval parameters.

[0016] According to a third aspect of this application, a computer device is provided, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor executes the program to implement the above-described dynamic learning planning teaching method based on multi-model agent collaboration.

[0017] According to a fourth aspect of this application, a storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described dynamic learning planning teaching method based on multi-model agent collaboration.

[0018] By employing the above technical solutions, this application provides a dynamic learning planning teaching system and method based on multi-model intelligent agent collaboration. The system includes a dynamic question search module, a dynamic test paper analysis module, a wrong question mastery analysis module, and a learning planning module. Specifically, the dynamic question search module, upon obtaining user-input descriptive information, retrieves a set of questions from a question bank based on the semantic features of the descriptive information. The dynamic test paper analysis module identifies wrong question data based on answer data. The wrong question mastery analysis module generates a structured wrong question bank based on the wrong question data and dynamically adjusts the review interval parameters according to the interval repetition algorithm based on the memory cards and answer data in the structured wrong question bank. The learning planning module then generates learning task plan data based on the structured wrong question bank and the review interval parameters. The system can construct a complete closed loop of question finding, wrong question analysis, mastery analysis, and practice through technologies such as large language models, enhanced retrieval generation, YOLO image segmentation, and interval repetition algorithms, realizing dynamic question search, wrong question cause analysis, mastery tracking, and personalized learning planning, thereby improving the accuracy of recommendation results.

[0019] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0020] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 The overall architecture diagram of the dynamic learning planning teaching system based on multi-model intelligent agent collaboration provided in the embodiments of this application; Figure 2 A schematic diagram of the teaching method for dynamic learning planning based on multi-model agent collaboration provided in an embodiment of this application; Figure 3A schematic diagram of the structure of a dynamic learning planning teaching system based on multi-model agent collaboration provided in an embodiment of this application; Figure 4 This is a schematic diagram of the process for searching similar questions provided in an embodiment of this application; Figure 5 This is a schematic diagram of the process for generating teaching resources based on context information, provided in an embodiment of this application. Detailed Implementation

[0021] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.

[0022] To address the issue of inaccurate recommendations in error management and learning planning within intelligent teaching systems, this application provides a dynamic learning planning teaching system based on multi-model intelligent agent collaboration in some embodiments. This dynamic learning planning teaching system based on multi-model intelligent agent collaboration, also known as an intelligent teaching system or teaching system, is a full-chain digital education platform that utilizes data-driven approaches and artificial intelligence (AI) analysis to achieve comprehensive teaching, learning, management, and evaluation.

[0023] like Figure 1 As shown, the system includes a dynamic test question search module, a dynamic test paper analysis module, a wrong question mastery analysis module, and a learning planning module. By integrating functions such as full-link data collection and intelligent analysis, personalized recommendations, interactive teaching, automated teaching assistance, and intelligent decision-making, it assists teachers, students, and administrators in achieving a scientific and intelligent education approach.

[0024] The system includes several modules: a dynamic question search module to retrieve relevant question resources based on user input; a dynamic test paper analysis module to collect and analyze user answer data to identify incorrect answers; a question mastery analysis module to analyze incorrect answer data and dynamically adjust review interval parameters based on the analysis results; and a learning planning module to generate learning task plan data.

[0025] Based on the above modules, the dynamic learning planning teaching system based on multi-model intelligent agent collaboration can achieve dynamic question search, error analysis, mastery tracking, and personalized learning planning by executing a dynamic learning planning teaching method based on multi-model intelligent agent collaboration. The method can be applied to electronic devices with data processing capabilities. These electronic devices include, but are not limited to, computers, servers, mobile terminals, smart wearable devices, and industrial control machines. For ease of description, this application embodiment uses a teaching system as the executing subject of the method. It should be understood that the method can also be applied to other types of executing subjects, which are not shown one by one in this application embodiment. Figure 2 As shown, the method includes: S101. Obtain the descriptive information input by the user, and retrieve the set of test questions from the test question bank based on the semantic features of the descriptive information.

[0026] Dynamic question search, as one of the core functions of the system, aims to provide users with an efficient and accurate question retrieval experience through natural language processing and multimodal information retrieval technologies. The dynamic question search module can integrate large language models (LLM) and retrieval-augmented generation (RAG) technologies to realize a complete process from question processing to semantic search.

[0027] Therefore, the dynamic question search module can utilize Retrieval Enhancement Generation (RAG) technology to process user-input descriptive information. Through semantic analysis, it can retrieve question sets from the question bank based on the semantic features of the descriptive information, thus providing a precise question search function. The descriptive information can be text or image content; for example, users can enter keywords or upload images of exam papers to trigger the system to automatically match relevant questions, providing users with an efficient way to access resources.

[0028] In some embodiments, to retrieve a question set, the dynamic question search module first acquires candidate question data, which includes multiple candidate questions and their solutions. Then, a large language model is used to perform semantic analysis on the candidate question data to extract key information, including knowledge point information, test point information, difficulty level information, and question type information. Next, type labels are generated based on the key information to represent structured descriptions of the candidate questions. By assigning type labels to the candidate questions, a question bank is constructed based on the candidate questions with assigned type labels.

[0029] The dynamic question search module can perform LLM tagging on each question in the question bank. The tagging process can use a large language model to perform deep semantic analysis on the question content and answers, extracting key information, including knowledge points, test points, difficulty level, question type, etc.

[0030] For example, for a math word problem, LLM might identify tags such as "quadratic function," "maximum value problem," and "exam point for middle school entrance exam." These tags not only provide a structured description of the problem but also lay the foundation for subsequent semantic matching. Therefore, candidate problems can be tagged by assigning type tags to them.

[0031] After tagging, the system further performs natural language processing on each question to generate a detailed text description, which can cover the question stem, options, and explanations to ensure comprehensive information. Therefore, in some embodiments, the dynamic question search module can also generate detailed text descriptions by performing natural language processing on the candidate question data. The candidate questions are then vectorized to obtain question vectors, and a question bank is constructed based on the question vectors, detailed text descriptions, and candidate question data. The question vectors are used to capture the semantic similarity between questions and form a structured corpus.

[0032] When the dynamic question search module vectorizes tagged and described questions, it can utilize an embedding model to map the text content of each question into a high-dimensional vector space, resulting in question vectors. These question vectors can capture the semantic similarity between questions, forming a structured corpus. This structured corpus not only serves as the storage foundation for questions but also provides computational support for subsequent semantic searches. For example, two questions involving the concept of "trigonometric functions," even with different wording, will have similar vector representations in the space, facilitating retrieval.

[0033] Based on the constructed question bank and the descriptive information input by the user, the dynamic question search module can retrieve a set of questions from the question bank according to the semantic features of the descriptive information. In some embodiments, to retrieve the question set, the dynamic question search module can use an embedding model to perform vectorization processing on the descriptive information to obtain an input vector. Then, a nearest neighbor algorithm is used to calculate the vector similarity between the input vector and multiple question vectors in the question bank. Finally, target questions are extracted from the question bank based on the vector similarity, and a question set is generated based on the target questions. The target questions are candidate questions whose vector similarity is greater than or equal to a similarity threshold.

[0034] During the user input phase, the dynamic question search module supports multiple input methods, including text and images. For text input, users can directly enter keywords or question descriptions. The dynamic question search module can then vectorize the user input using an embedding model and perform a semantic-level search within the corpus. The search process can employ the nearest neighbor algorithm (Faiss) to return the most relevant question results based on vector similarity.

[0035] For image input, the dynamic question search module can preprocess the input image using LLM-based Optical Character Recognition (OCR) technologies such as deepseek-ocr. During preprocessing, the OCR engine, combined with LLM's semantic understanding capabilities, can accurately recognize the text content of exam papers or handwritten questions, thus converting the image into structured text. Subsequently, the aforementioned vectorization and semantic search process is applied to ensure that image-input questions can be efficiently matched with similar questions in the question bank. The semantic-level question search algorithm using RAG and OCR technologies significantly improves the processing capability for multimodal descriptive information, especially suitable for scenarios involving handwritten exam papers or scanned documents, thereby enhancing the system's practicality.

[0036] Through the above process, the dynamic question search module can achieve a complete closed loop from question semantic modeling to multimodal retrieval. Furthermore, thanks to the deep semantic understanding of LLM and the efficient computational power of the embedded model, it can meet different learning needs, achieving everything from basic knowledge retrieval to advanced question recommendation, thus improving the robustness of the teaching system.

[0037] S102. Obtain the answer data uploaded by the user based on the question set, and identify the wrong question data from the answer data.

[0038] The dynamic test paper analysis module is the backbone of the system's intelligent analysis, responsible for the automated processing and error analysis of user-uploaded test papers. Combining computer vision technology and large-scale language models, the module provides end-to-end support for test paper segmentation, error identification, and root cause analysis.

[0039] like Figure 3 As shown, after obtaining a set of test questions through retrieval, the dynamic test question search module can generate test paper data based on the search results and output the test paper data. For example, by arranging multiple test paper questions from the retrieved test paper set according to a preset test paper template, test paper data that can be displayed on the user's terminal device can be obtained. Then, by sending the test paper data to the terminal device, the user can answer the test paper and generate answer data.

[0040] At this time, the test paper dynamic analysis module can obtain the answer data uploaded by the user based on the test question set, and combine the You Only Look Once (YOLO) image recognition technology and LLM to identify the wrong question data from the answer data. The wrong question data includes the wrong question set and the wrong question reason information corresponding to the wrong question set.

[0041] The dynamic test paper analysis module can automatically segment test papers, identify incorrect questions, and conduct in-depth analysis of the reasons for errors. It uses image processing technology to extract specific questions and answers from the test paper, and combines this with a language model to determine the type of error, such as calculation errors or conceptual misunderstandings, providing targeted feedback for subsequent learning.

[0042] Therefore, in order to identify incorrect answer data, in some embodiments, when the test paper dynamic analysis module identifies incorrect answer data from the answer data, it can perform test paper question segmentation on the answer data to obtain multiple question regions. During the test paper question segmentation process, the detection head is dynamically adjusted to adapt to different test paper formats.

[0043] For example, the dynamic test paper analysis module preprocesses the uploaded test paper images and uses the DyHead algorithm to segment the test paper questions. DyHead is a highly efficient instance segmentation technology that can accurately identify each question area in the test paper and dynamically adjust the detection head to adapt to different test paper formats.

[0044] Then, using YOLO object detection, the correct answer area and the incorrect answer area are marked from multiple question areas. Then, incorrect question tags are generated based on the incorrect answer areas, and the incorrect question tags are used to mark the corresponding incorrect questions on the test paper, resulting in a set of incorrect questions.

[0045] After preprocessing the test paper images, the dynamic test paper analysis module uses the YOLO model to identify incorrect answer tags. Through YOLO object detection technology, the module can mark the correct and incorrect answer areas in the test paper and generate a unique identifier, i.e., an incorrect answer tag, for each incorrect answer. When marking the correct and incorrect answer areas, bounding box detection and classification functions are combined to ensure accurate differentiation between correct and incorrect questions. Then, the generated unique identifiers are used to label the incorrect answer areas with the corresponding incorrect questions, resulting in a set of incorrect questions.

[0046] After obtaining the set of incorrect questions, the test paper dynamic analysis module can use the question-answering agent to extract relevant information from the set of incorrect questions and generate information on the reasons for the incorrect questions based on the relevant information and answer data. The relevant information includes core knowledge point information, test point information, and summary information. The summary information includes a summary of the question's meaning and information on the problem-solving approach.

[0047] For example, the identified incorrect questions are sent to the question-solving AI for in-depth analysis. The question-solving AI is a data processing tool based on LLM technology, possessing natural language understanding and reasoning capabilities. For each incorrect question, the AI ​​extracts core knowledge points, such as "fraction operations" or "chemical equations," and key concepts, such as "classification and discussion" or "combining numerical and graphical methods," and generates a brief summary of the question. This summary may include a summary of the question's meaning and the solution approach.

[0048] Meanwhile, the problem-solving agent relies on the contextual understanding capabilities of LLM and combines the user's answer data with the corresponding answer situation to analyze the reasons for errors, such as "calculation errors", "conceptual confusion" and "missing solution steps", so as to identify the students' thinking bottlenecks in problem-solving in a targeted manner.

[0049] The dynamic test paper analysis module then outputs the analysis results in a structured format, providing targeted feedback for subsequent learning. For example, for a question answered incorrectly due to a "unit conversion error," the teaching system not only points out the error location but also recommends relevant knowledge points for review and similar practice questions. This feedback mechanism, linking incorrect question data with the dynamic test question search module, enhances the completeness of the learning loop. Furthermore, the dynamic test paper analysis module leverages the visual processing capabilities of DyHead and YOLO, as well as the semantic analysis capabilities of LLM, to support large-scale test paper processing and personalized error diagnosis.

[0050] S103. Generate a structured error database based on the error data, and dynamically adjust the review interval parameters based on memory cards and answer data according to the interval repetition algorithm.

[0051] After identifying error data, the test paper dynamic analysis module can transmit this data to the error mastery analysis module. The error mastery analysis module aims to scientifically assess the user's memory and understanding of incorrect questions and dynamically adjust review strategies. This module combines flashcard mechanisms, spaced repetition algorithms, and large-scale language models to achieve intelligent management of incorrect questions.

[0052] Therefore, the error mastery analysis module can first generate a structured error database based on the error data, wherein the structured error database includes memory cards created for at least one incorrect question from a test paper. In some embodiments, to generate the structured error database, the error mastery analysis module can acquire error data from multiple data sources. Since the error data includes incorrect questions from multiple test papers, memory association information can be extracted from the error data for each incorrect question from the test paper. This memory association information includes question stem information, answer information, explanation, and error analysis information. Then, memory cards for the incorrect test papers are created based on the memory association information, and the structured error database is generated by aggregating multiple memory cards.

[0053] For example, the error mastery analysis module aggregates all error data from the dynamic question search module and the dynamic test paper analysis module. This error data can include detailed information such as the knowledge points, test points, and specific reasons for the errors, forming a structured error database. Furthermore, each error is made into a memory card, containing information such as the question stem, answer, explanation, and error analysis, to facilitate users' browsing and review at any time.

[0054] The module for analyzing the mastery of incorrect questions can dynamically assess a user's understanding of those questions based on a flashcard mechanism and a spaced repetition algorithm based on a large-scale model. The spaced repetition algorithm dynamically adjusts the review interval based on the forgetting curve. Therefore, the module can dynamically adjust the review interval parameters based on flashcards and answer data, following the spaced repetition algorithm. By recording learners' answer history and review progress, the teaching system can intelligently adjust the review interval, optimize memory efficiency, and help users consolidate knowledge points.

[0055] Therefore, in some embodiments, the error-answering mastery analysis module can receive feedback information input by the user based on the memory cards. This feedback information includes a level record representing the user's mastery level. For example, when using the memory cards, the user can click on options such as "Remember," "Very vague," "Forgot," or "Completely not" to send feedback information to the error-answering mastery analysis module based on their own mastery level.

[0056] Then, based on the feedback information, memory habit information is generated, and review interval parameters such as review interval time are dynamically adjusted based on the memory habit information. The memory habit information includes a first habit feature determined based on the forgetting curve in the spaced repetition algorithm and a second habit feature obtained by analyzing historical answer data through a large language model.

[0057] For example, after receiving feedback from users based on flashcard input, the teaching system can record this feedback. This allows the error analysis module to combine spaced repetition with a large language model to analyze different students' memory habits, thereby generating memory habit information. Specifically, the spaced repetition algorithm dynamically adjusts the review interval based on the Ebbinghaus forgetting curve, while LLM (Learning Learning Model) personalizes the interval parameters by analyzing students' historical answer data. That is, for students with strong memories, the teaching system can extend the review interval; while for easily forgotten knowledge points, the system can shorten the interval to strengthen memory.

[0058] The module then sets the mastery level of the incorrect questions on the test paper based on feedback information and the aforementioned review interval parameters. This mastery level is used to assess learning progress by quantifying the degree of mastery, based on a target level determined by memory half-life. For example, the module dynamically categorizes incorrect questions into different mastery levels, such as "beginner," "introductory," "familiar," "proficient," "master," "completely mastered," etc., based on student feedback and review effectiveness.

[0059] The mastery level classification can be based on target levels of memory half-life, quantifying the user's mastery level to accurately assess learning progress. For example, upgrading a wrong question from "beginner" to "proficient" requires the user to answer it correctly multiple times and reach the preset memory consolidation standard. The wrong question mastery analysis module will also adjust subsequent review plans according to the level to ensure that the user's memory of each wrong question achieves long-term retention.

[0060] As can be seen, the error analysis module, through the combination of memory cards and intelligent algorithms, achieves a shift from passive recording to proactive optimization. Thanks to LLM's deep modeling of user behavior, it significantly enhances personalized adjustment capabilities, assisting users in reviewing their mistakes.

[0061] S104. Generate learning task plan data based on the structured error database and review interval parameters.

[0062] The error-based question mastery analysis module generates mastery assessment results and dynamically adjusts the review interval parameters using an interval repetition algorithm. These results and parameters are then passed to the learning planning module. The learning planning module is the core of the intelligent learning support system, responsible for generating personalized learning tasks based on user learning data. It utilizes a large-scale language model and multi-dimensional analysis to develop a scientific learning plan.

[0063] The learning planning module comprehensively analyzes user learning data, including records of incorrect answers and mastery assessment results, to generate personalized learning task plans. These plans recommend specific learning content and practice questions based on the user's learning progress and weaknesses. Therefore, the learning planning module is configured to generate learning task plan data based on a structured error database and review interval parameters.

[0064] In some embodiments, the learning planning module can first collect user browsing information and generate learning habit corpus based on this information. The user browsing information includes the frequency and time of browsing memory cards. Then, by analyzing the mastery curve and error reasons information of each test paper's incorrect answers, the module mines user knowledge point mastery corpus and thinking deficiency corpus. Finally, by combining the learning habit corpus, knowledge point mastery corpus, and thinking deficiency corpus, learning task plan data for a preset period is generated.

[0065] The learning planning module analyzes users' learning habits, i.e., the information they browse, based on the frequency and completion time of their flashcard browsing. For example, if a user frequently reviews math problems late at night, the system may infer that the user's nighttime learning efficiency is higher and adjust task release times accordingly. Simultaneously, completion time reflects a user's focus level; therefore, the learning planning module combines this data to optimize task allocation, ensuring that the learning pace and habits align with the learning plan.

[0066] The learning planning module can also analyze the mastery curve of each incorrect question and the specific reasons for the error, revealing the user's knowledge mastery and thinking deficiencies. This knowledge mastery and thinking deficiencies can be used as corpus data for model identification and application. For example, if the mastery curve for a question caused by a "logical reasoning error" remains stagnant for a long period, the teaching system may identify the user's weakness in "logical thinking." Based on this, the learning planning module can recommend targeted practice questions or video tutorials.

[0067] The learning planning module then integrates all the data to generate a learning task plan for the following week. The learning task plan includes daily goals (such as reviewing 10 wrong questions), key knowledge points (such as "function properties"), and recommended resources (such as online course links).

[0068] After generating learning task plan data, the teaching system can also receive dynamic feedback from users on the learning task plan data and adjust the learning task plan data based on the dynamic feedback. For example, the learning planning module can rely on the comprehensive reasoning ability of LLM to dynamically adjust the learning task plan based on student feedback, ensuring that the difficulty of the tasks in the learning task plan matches the user's learning progress, thereby providing scientific and personalized learning guidance information.

[0069] It is evident that the learning planning module can achieve closed-loop support from data to action through multi-dimensional data analysis and dynamic adjustment. The recommendation information generated by the learning planning module can be used to intelligently assist users in learning.

[0070] By applying the technical solutions of the above embodiments, the dynamic learning planning teaching system based on multi-model intelligent agent collaboration described in the above embodiments can retrieve a set of questions from the question bank based on the semantic features of the user-input description information through a dynamic question search module. Then, the dynamic test paper analysis module identifies incorrect question data based on the answer data, enabling the incorrect question mastery analysis module to generate a structured incorrect question bank based on the incorrect question data. Based on the memory cards and answer data in the structured incorrect question bank, the review interval parameter is dynamically adjusted according to the interval repetition algorithm. Finally, the learning planning module generates learning task plan data based on the structured incorrect question bank and the review interval parameter. The system can construct a complete closed loop of question finding, incorrect question analysis, mastery analysis, and practice through technologies such as large language models, enhanced retrieval generation, YOLO image segmentation, and interval repetition algorithms, realizing dynamic question search, incorrect question cause analysis, mastery tracking, and personalized learning planning, thereby improving the accuracy of recommendation results.

[0071] In some embodiments, as a refinement and extension of the specific implementation of the above embodiments, and to fully illustrate the specific implementation process of this embodiment, some embodiments of this application also provide a dynamic learning planning teaching method based on multi-model intelligent agent collaboration. This method allows the incorrect question data identified by the dynamic test paper analysis module to be linked with the dynamic test question search module through a multimodal similar question search algorithm to find similar questions for the user to practice further, thereby achieving a closed-loop learning system for consolidating incorrect questions. Figure 4 As shown, the method includes: S201. Extract similar question features from incorrect question data; S202. By integrating the features and descriptive information of similar test questions, similar question description information is generated; S203. Retrieve a set of similar questions from the question bank based on the semantic features of the description information of similar questions.

[0072] After identifying incorrect question data, the dynamic test paper analysis module can pass this data to the dynamic question search module. The dynamic question search module then extracts similar question features from the incorrect question data. This is done by executing a multimodal similar question search algorithm.

[0073] Similarly, the teaching system can process incorrect question data in both image and text formats. For image data, the system uses LLM-optimized OCR technology for preprocessing. By combining LLM's semantic understanding capabilities, it can more accurately identify complex content such as mathematical formulas and chemical equations in handwritten test papers or scanned documents. For example, for a handwritten question containing "sin(x)+cos(x)=1", OCR not only extracts the text but also determines from the context that it relates to trigonometric functions. After extracting the text, it undergoes structured transformation for subsequent matching of similar questions. Text input can directly enter the RAG processing flow, using RAG technology to retrieve a massive number of questions from the question bank and combining it with an embedding model to generate a candidate result set, ensuring comprehensive semantic coverage.

[0074] After extracting the content of the incorrect question data, the next step is feature extraction and similarity calculation. In this stage, similar question description information can be generated by fusing similar question features and descriptive information, and a set of similar questions can be retrieved from the question bank based on the semantic features of the similar question description information.

[0075] For the images and text of incorrect answers, the teaching system can generate high-dimensional vector representations respectively. The image vectors are generated by extracting visual features, such as font layout and graphic elements, through a convolutional neural network (CNN), while the text vectors are generated by a pre-trained embedding model based on knowledge points, test points, and reasons for errors.

[0076] Then, a Bayesian weighted calculation method is used to comprehensively evaluate similarity based on multi-dimensional features. To this end, a Bayesian model can be constructed, which can dynamically adjust weights based on the specific attributes of incorrect questions. For example, if the error in a question is a "unit conversion error," the system will increase the scoring weight of questions related to "unit conversion"; if it involves "geometric proofs," proof-type questions will be matched first. This weighted mechanism can improve the relevance of the recommendation results to the learning needs.

[0077] Furthermore, the multimodal similar question search algorithm also considers the difficulty level and question type consistency of the questions, avoiding the recommendation of overly simple or completely irrelevant questions. For example, for a physics question containing both graphs and text, the system can simultaneously analyze the graphical features and formula content to find matching practice questions. Therefore, the dynamic adjustment characteristic of the multimodal similar question search algorithm makes it highly adaptable, enabling it to continuously optimize the recommendation strategy based on student feedback.

[0078] By applying the technical solutions of the above embodiments, the dynamic learning planning teaching method based on multi-model agent collaboration described in the above embodiments can provide users with highly relevant and targeted similar practice questions based on multimodal similar question search algorithms. By integrating multiple data formats such as images and text, and combining retrieval augmented generation (RAG) technology with statistical models, efficient cross-modal matching is achieved.

[0079] In some embodiments, as a refinement and extension of the specific implementation of the above embodiments, and to fully illustrate the specific implementation process of this embodiment, some embodiments of this application also provide a dynamic learning planning teaching method based on multi-model agent collaboration. The difference between this method and the above embodiments is that after the learning planning module generates personalized learning planning data, it can combine the Model Context Protocol (MCP) test paper generation agent with the user's real-time feedback to form a complete closed-loop optimization process. For example... Figure 5 As shown, the method includes: S301, Call resources to generate an intelligent agent; S302. Use resource-generated intelligent agents to extract contextual information from descriptive information; S303. Modify the semantic features of the description information based on context information; S304. Retrieve the test question set from the test question bank based on the modified semantic features; S305. Generate teaching resources based on the set of test questions.

[0080] In order to form a complete closed-loop optimization process by combining contextual information with the user's real-time task feedback, the dynamic test question search module can call the resource generation agent after the learning planning module generates learning task plan data. The resource generation agent is a resource generation tool that combines a large language model with a model context protocol.

[0081] The resource generation agent based on MCP can serve as a core intelligent tool in teaching systems. It aims to automatically generate high-quality teaching resources such as lesson plans or test papers by combining Large Language Models (LLMs) with the MCP protocol. The resource generation agent can retrieve content from a question bank through dynamic context management and generate structured, personalized teaching resources according to user needs.

[0082] For example, the functions of the resource generation agent can include curriculum standards requirements, learning objectives, scenario introduction, self-directed learning, pre-class assessment, collaborative inquiry, in-class testing, post-class consolidation, and question recommendation. It also relies on the MCP protocol to achieve seamless collaboration between modules, including context modeling, content generation, and optimization, to suit diverse scenarios for teachers' lesson preparation and students' personalized learning.

[0083] After generating an intelligent agent by invoking resources, the agent can be used to extract contextual information from the descriptive information and modify its semantic features based on this contextual information. This contextual information includes learning task plan data and task feedback data input by the user in response to the learning task plan data.

[0084] As the core framework of the intelligent agent, MCP is responsible for managing the contextual interactions between the LLM and various modules of the system (such as the dynamic question search module). Through dynamic context modeling, MCP ensures that the LLM can efficiently collaborate with the question bank data based on real-time input parameters (such as grade, subject, knowledge point, difficulty level, etc.).

[0085] For example, when a user specifies "Grade 8 Math - Fraction Operations - Medium Difficulty", MCP will activate the dynamic question search module, retrieve relevant questions from the question bank using RAG technology, and pass the results to LLM for structured processing.

[0086] MCP also allows the system to dynamically adjust the context based on task requirements. For example, it can prioritize knowledge points like "addition and subtraction of fractions with unlike denominators" or "calculation problems" to ensure the relevance of the generated content. MCP can use Function Call technology to specifically invoke LLM to execute predefined functions. For example, it can extract objectives based on educational standards, such as "mastering fraction addition and subtraction" to generate curriculum requirements; clarify learning objectives, such as "completing the addition of fractions with unlike denominators"; design scenarios such as "supermarket discount calculation" to introduce the topic; provide self-study guides to assist users in independent learning; generate test questions to implement pre-class assessment; design discussion tasks to facilitate collaborative inquiry; create practice questions for in-class testing; recommend homework for after-class reinforcement; and provide extended questions for question recommendation.

[0087] Then, based on the modified semantic features, a set of test questions is retrieved from the test question bank, and teaching resources are generated based on the test question set. These teaching resources include lesson plan data and test paper data associated with the test question set. The content generation process relies on the dynamic question search capability of the MCP protocol. As a key component of MCP collaboration, the dynamic question search module can use RAG technology to vectorize user input and perform semantic-level searches within the test question bank.

[0088] For example, for "Grade 8 Math - Fraction Operations," the system might retrieve hundreds of relevant questions. The MCP protocol, through context management, ensures that the search results match the generated task, allowing the LLM to generate preliminary lesson plans or test drafts based on the MCP plan. In the draft, the "In-Class Assessment" section might contain 5 multiple-choice questions and 3 calculation questions, all dynamically obtained from the question bank. MCP also supports multi-concept mapping, balancing the knowledge point coverage of the test (such as "addition and subtraction with the same denominator" and "finding a common denominator") and the distribution of question types (such as 30% multiple-choice questions, 50% calculation questions, and 20% application questions) to reduce content duplication or omissions.

[0089] Meanwhile, the system calls upon external knowledge bases to verify the accuracy of test questions and adjusts the difficulty according to curriculum standards. For example, if the curriculum standard requires "understanding fraction simplification," the MCP will guide the LLM to prioritize questions that include simplification steps.

[0090] After initial generation, LLM refines the scenario based on context feedback from the MCP protocol. For example, if the "supermarket shopping" scenario is deemed too simple, MCP can dynamically adjust the context and replace it with a "project budget" scenario to increase the difficulty.

[0091] Because Function Call technology supports multiple iterations, if users report insufficient test questions, MCP will re-trigger the dynamic question search module to add 5-10 questions. This leverages MCP's real-time collaboration capabilities and LLM's generative potential to ensure that the output content is both demanding and diverse.

[0092] In addition, MCP also supports cross-module data integration. For example, it can combine data from the module that analyzes students' mastery of incorrect questions to prioritize recommending questions that address students' weak knowledge points.

[0093] Finally, the intelligent agent outputs the final product in a structured format. Lesson plans or test papers can be generated in PDF format, including a title, learning objectives, a list of questions, answer explanations, and usage suggestions; they can also be generated in an interactive format, allowing users to preview and adjust them online. For example, a Grade 8 math lesson plan might include modules such as "Curriculum Standards: Master the four arithmetic operations with fractions," "Learning Objectives: Complete addition and subtraction calculations with unlike denominators," and "Scenario Introduction: Supermarket discount calculations," as well as 10 practice questions retrieved from a question bank. After output, teachers or students can provide feedback, and the MCP protocol will adjust the context based on the feedback to generate an optimized version, such as increasing the difficulty or changing the questions.

[0094] By applying the technical solutions of the above embodiments, the dynamic learning planning teaching method based on multi-model agent collaboration described in the above embodiments can combine the dynamic collaboration of MCP with the generative capabilities of LLM. The resource-generating agent based on MCP can not only quickly generate teaching resources, but also ensure the scientific accuracy and relevance of the content through dynamic question searching. Therefore, in practical applications, it can support diverse teaching needs, enabling the generation of complete unit lesson plans for teachers or personalized error practice sets for students. Utilizing the context management efficiency of MCP, the retrieval capabilities of RAG, and the generative potential of LLM, the cost of teaching preparation is significantly reduced, while simultaneously improving the quality of generated educational resources.

[0095] This application also provides a computer device, specifically a personal computer, server, network device, etc. The computer device includes a bus, processor, memory, and communication interface, and may also include input / output interfaces and a display device. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the computer device stores location information. The network interface of the computer device is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the various method embodiments.

[0096] Those skilled in the art will understand that the structure of the computer device described above is only a partial structure related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. A specific computer device may include more or fewer components, or combine certain components, or have different component arrangements.

[0097] In one embodiment, a computer-readable storage medium is also provided, which may be non-volatile or volatile, and a computer program is stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0098] In one embodiment, a computer program product is also provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0099] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0100] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods.

[0101] Any references to memory, database, or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc.

[0102] Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can take many forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0103] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchain. The processors involved in the embodiments provided in this application may be, but are not limited to, general-purpose processors, graphics processors, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc.

[0104] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0105] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A dynamic learning planning teaching system based on multi-model agent cooperation, characterized in that, The system includes: The dynamic test question search module is configured to obtain descriptive information input by the user and retrieve a set of test questions from the test question bank based on the semantic features of the descriptive information. The test paper dynamic analysis module is configured to acquire answer data uploaded by users based on the test question set, and to identify incorrect question data from the answer data. The incorrect question data includes a set of incorrect questions and the reason information for the incorrect questions corresponding to the set of incorrect questions. The module for analyzing the level of mastery of incorrect questions is configured to generate a structured incorrect question database based on the incorrect question data. The structured incorrect question database includes memory cards made for at least one incorrect question from a test paper. Based on the memory cards and the answer data, the review interval parameter is dynamically adjusted according to the interval repetition algorithm. The interval repetition algorithm is used to dynamically adjust the review interval time based on the forgetting curve. The learning planning module is configured to generate learning task plan data based on the structured error database and the review interval parameters.

2. The system of claim 1, wherein, The dynamic test question search module is also configured as follows: Obtain candidate test question data, which includes multiple candidate test questions and their answer information; Semantic analysis is performed on the candidate test question data using a large language model to extract key information from the candidate test question data; the key information includes knowledge point information, test point information, difficulty level information, and question type information; Type labels are generated based on the key information, and the type labels are used to characterize the structured description of the candidate test questions; The type label is set to the candidate test questions, and the test question bank is constructed based on the candidate test questions with the type label set.

3. The system of claim 2, wherein, The dynamic question search module is configured to construct the question bank based on the candidate questions with the set type tags, and is further configured to: Detailed text descriptions are generated by performing natural language processing on the candidate test question data; The candidate test questions are vectorized to obtain test question vectors; the test question vectors are used to capture the semantic similarity between test questions and form a structured corpus. The question bank is constructed based on the question vectors, the detailed text descriptions, and the candidate question data.

4. The system of claim 3, wherein, The dynamic test question search module performs a retrieval of a set of test questions from the test question bank based on the semantic features of the description information, and is further configured to: The description information is vectorized using an embedding model to obtain an input vector; The nearest neighbor algorithm is used to calculate the vector similarity between the input vector and multiple question vectors in the question bank; Target questions are extracted from the question bank based on the vector similarity, wherein the target questions are the candidate questions whose vector similarity is greater than or equal to the similarity threshold; The test question set is generated based on the target test questions.

5. The system of claim 1, wherein, The test paper dynamic analysis module is configured to identify incorrect answer data from the answer data and is also configured to: The answer data is segmented into multiple question regions; during the segmentation process, the detection head is dynamically adjusted to adapt to different test paper formats. Using YOLO object detection, the correct answer areas and incorrect answer areas are marked from multiple question areas; Based on the incorrect answer area, generate incorrect question tags, and use the incorrect question tags to mark the incorrect questions on the test paper corresponding to the incorrect answer area, to obtain the incorrect question set; The problem-solving agent extracts relevant information from the set of incorrect questions. This relevant information includes core knowledge points, test points, and summary information. The summary information includes a summary of the question and the solution approach. The error reason information is generated based on the relevant information and the answer data.

6. The system of claim 1, wherein, The module for analyzing the level of mastery of incorrect questions is also configured as follows: The system acquires incorrect question data from multiple data sources; the incorrect question data includes incorrect questions from multiple test papers. Memory association information is extracted from the incorrect question data, and the memory association information includes question stem information, answer information, explanation and error analysis information; Memory cards for the incorrect questions on the test paper are created based on the memory association information. The structured error database is generated by aggregating multiple memory cards.

7. The system of claim 6, wherein, The module for analyzing the level of mastery of incorrect questions is also configured as follows: Receive feedback information input by the user based on the memory card, the feedback information including a level record for characterizing mastery; Based on the feedback information, memory habit information is generated, which includes a first habit feature determined by the forgetting curve in the interval repetition algorithm and a second habit feature obtained by analyzing historical answer data through a large language model. The review interval parameter is dynamically adjusted based on the memory habit information, and the review interval parameter includes the review interval time; The mastery level of the incorrect questions on the test paper is set according to the feedback information and the review interval parameter; the mastery level is used to assess the learning progress by quantifying the degree of mastery based on the target level of memory half-life.

8. The system of claim 1, wherein, The learning planning module is also configured as follows: Collect user browsing information and generate learning habit corpus based on the user browsing information; the user browsing information includes the frequency and time of the user browsing the memory cards; By analyzing the mastery curve and error reason information of each wrong question in the test paper, we can mine the data on users' mastery of knowledge points and their thinking defects. By combining the learning habit corpus, the knowledge point mastery corpus, and the thinking defect corpus, the learning task plan data within the preset period is generated; Receive dynamic feedback from users regarding the learning task plan data, and adjust the learning task plan data based on the dynamic feedback.

9. The system of claim 1, wherein, The dynamic test question search module is also configured as follows: The resource generation agent is invoked, which is a resource generation tool that combines a large language model with a model context protocol. The resource is used to generate an intelligent agent to extract context information from the description information. The context information includes learning task plan data and task feedback data input by the user in response to the learning task plan data. Modify the semantic features of the description information based on the context information; Retrieve a set of test questions from the test question bank based on the modified semantic features; Teaching resources are generated based on the set of test questions, and the teaching resources include lesson plan data and test paper data associated with the set of test questions.

10. A dynamic learning planning teaching method based on multi-model agent cooperation, characterized in that, Applied to the system according to any one of claims 1-9, the method comprises: Obtain descriptive information input by the user, and retrieve a set of test questions from the test question bank based on the semantic features of the descriptive information; The system acquires answer data uploaded by users based on the question set, and identifies incorrect question data from the answer data. The incorrect question data includes a set of incorrect questions and information on the reasons for the incorrect questions corresponding to the set of incorrect questions. A structured error database is generated based on the error data, and the structured error database includes memory cards made for at least one error question from a test paper; Based on the memory cards and the answer data, the review interval parameters are dynamically adjusted according to the interval repetition algorithm; the interval repetition algorithm is used to dynamically adjust the review interval time based on the forgetting curve. Learning task plan data is generated based on the structured error database and the review interval parameters.

Images