An adaptive learning content processing method and system based on multi-agent collaboration

Abstract

The application belongs to the technical field of artificial intelligence, computer information processing and intelligent education, and discloses a self-adaptive learning content processing method and system based on multi-agent cooperation, which comprises the following steps: obtaining learning materials and learning requests of a user, establishing a session identifier and performing intention recognition on the learning request; obtaining examination and post public information, establishing an examination portrait and a post ability portrait, and analyzing and generating a structured content unit; distributing data and generating a question set according to the intention recognition result; outputting a question and receiving an answer result, outputting a correction result and extracting diagnosis information, generating a learning plan based on the diagnosis information; encapsulating each processing result into a record unit and storing; in the next round of learning, the generation strategy is adjusted according to the storage result to form a self-adaptive closed loop. Through the use of the multi-agent cooperation and asynchronous task distribution mode for different tasks, the application can reduce the system main thread blocking and improve the overall processing efficiency of complex learning tasks.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, computer information processing, and intelligent education technology, and in particular to an adaptive learning content processing method and system based on multi-agent collaboration. Background Technology

[0002] With the development of intelligent learning platforms, users are increasingly inputting various learning materials into the system, such as PDF documents, web pages, audio, video transcripts, and past exam papers, hoping that the system will automatically complete content parsing, knowledge extraction, question generation, answer assessment, and subsequent learning arrangements. However, existing technologies typically handle these steps in a fragmented manner or rely on a single model to complete them all at once, resulting in a lack of closed-loop coordination in the overall process and making it difficult to continuously and dynamically adjust subsequent learning content based on the user's actual responses.

[0003] Existing technologies typically suffer from the following defects and shortcomings: 1. Multimodal learning materials have large structural differences, and existing solutions often only complete simple text extraction, making it difficult to further form structured content units that can be reused for question setting, assessment, and planning.

[0004] 2. Question generation, answer grading, and learning plan adjustment are usually completed by independent modules. The lack of unified status coordination and result transmission between modules leads to a disconnect between the generated questions, grading results, and learning plans.

[0005] 3. Existing assessment systems mostly focus on giving scores or judging right or wrong answers, lacking a mechanism to extract knowledge points, weaknesses, and subsequent learning suggestions based on the content of the questions and the details of the answers.

[0006] 4. Learning plans are mostly statically generated results, making it difficult to dynamically adjust them based on the user's progress, learning feedback, quiz results, and remaining learning time.

[0007] 5. During multiple learning interactions, existing solutions typically lack session-oriented state management and version management, making it difficult to achieve a continuous learning loop across tasks and devices.

[0008] 6. When content parsing, question generation, grading, and plan generation are all processed sequentially and centrally, the system response latency is high and it is difficult to adapt to the processing requirements of different task types.

[0009] 7. Existing solutions typically struggle to comprehensively utilize information such as the exam syllabus, exam subjects, question structure, registration rules, exam deadlines, sample questions and papers, and scoring rules published on the official exam website, making it difficult to generate a fully authentic mock exam that matches the target exam structure.

[0010] 8. Existing learning platforms typically separate exam preparation from job competency preparation, lacking a mechanism to uniformly map publicly available internship job information, job responsibilities, and exam objectives into the same learning path, making it difficult to form a closed loop and personalized path for exam preparation and internship learning.

[0011] 9. Existing systems typically process all requests according to a fixed workflow when receiving user input, lacking the ability to automatically identify and categorize user learning intentions. User-input natural language text may contain various types of learning intentions, such as exam preparation planning, question generation, knowledge tutoring, and assessment. Existing solutions usually require users to specify the task type by clicking a specific button or selecting a menu, failing to automatically identify the user's true learning intentions and learning goal parameters (such as target exam name, exam deadline, expected learning intensity, etc.) from the user's natural language input. This results in low interaction efficiency and difficulty in handling complex requests containing multiple intentions.

[0012] In view of the shortcomings of the prior art, the actual technical problem to be solved by the present invention is not simply generating questions or outputting learning plans, but rather solving the following set of interrelated technical problems: 1. How to transform learning materials from different sources, such as PDFs, web pages, audio transcriptions, video transcriptions, and test papers, into structured content units that can be reused in subsequent question setting, assessment, and planning modules.

[0013] 2. How to transform content processing, question generation, answer grading, knowledge point diagnosis, and learning plan adjustment from isolated functional modules into a processing chain that can share intermediate results and collaborate continuously.

[0014] 3. How can the system not only provide questions and scores, but also identify the user's mastery of knowledge points, weaknesses, and subsequent learning suggestions based on the user's actual answers?

[0015] 4. How can the subsequent learning plan, question difficulty, question type distribution, and knowledge point coverage be automatically adjusted based on the previous round of answers, user feedback, and completion progress, thereby forming an adaptive learning loop?

[0016] 5. How to uniformly link and continuously manage learning materials, question results, assessment results, and learning plan versions in multi-round learning and cross-device scenarios.

[0017] 6. How to automatically construct a profile of the target exam using publicly available information from the official exam website, and generate a full-fledged mock exam that matches the target exam in terms of subject settings, question types, knowledge coverage, answering time, and output format.

[0018] 7. How to integrate publicly available internship job information or publicly available job competency requirements into the learning plan, combining job competency profiles with exam objectives, user learning performance, and historical feedback, to form an integrated and personalized path for further education, certification exams, job hunting, and internship preparation.

[0019] 8. How to automatically identify the learning intent type and learning target parameters from the natural language text input by the user, so that the system can automatically select the corresponding intelligent agent processing link based on the identification results, without the user having to manually specify the task type.

[0020] Therefore, how to provide an adaptive learning content processing method and system based on multi-agent collaboration is an urgent problem to be solved. Summary of the Invention

[0021] This invention provides an adaptive learning content processing method and system based on multi-agent collaboration to solve the problems mentioned above in the prior art.

[0022] According to a first aspect of the present invention, an adaptive learning content processing method based on multi-agent collaboration is provided.

[0023] In one embodiment, an adaptive learning content processing method based on multi-agent collaboration includes: Obtain the user's learning materials and learning requests, establish a session identifier associated with the user's learning requests, and perform intent recognition on the user's learning requests, outputting the intent recognition results; Obtain publicly available information on the target exam and the target position, and create exam profiles and job competency profiles respectively; analyze learning materials, publicly available exam information, and publicly available job information to generate structured content units; Based on the intent recognition results, structured content units, exam profiles, and job competency profiles are distributed to the corresponding processing flows; a set of questions is generated based on the structured content units, exam profiles, and job competency profiles. Output a set of questions to the user and receive the answers; output grading results and extract diagnostic information based on the answers; generate a learning plan based on the diagnostic information, exam profile, and job competency profile. The structured content units, question sets, grading results, diagnostic information, and learning plans are each encapsulated into a record unit, and each record unit is stored. In the next round of learning, the question generation strategy, exam preparation focus, and job skills training focus will be adjusted based on the stored results to form an adaptive learning loop.

[0024] According to a second aspect of the present invention, an adaptive learning content processing system based on multi-agent collaboration is provided.

[0025] In one embodiment, an adaptive learning content processing system based on multi-agent collaboration includes: The input and session management module is used to obtain users' learning materials and learning requests, establish session identifiers associated with users' learning requests, and perform intent recognition on users' learning requests, and output the intent recognition results. The content analysis and profile building module is used to obtain publicly available information about the target exam and the target job, and to build exam profiles and job competency profiles respectively; it also analyzes learning materials, publicly available exam information, and publicly available job information to generate structured content units. The distribution and question generation module is used to distribute structured content units, exam profiles, and job competency profiles to the corresponding processing flows based on the intent recognition results; and to generate a set of questions based on the structured content units, exam profiles, and job competency profiles. The assessment, diagnosis, and plan generation module is used to output a set of questions to users and receive their answers; output grading results and extract diagnostic information based on the answers; and generate learning plans based on diagnostic information, exam profiles, and job competency profiles. The status management and persistence module is used to encapsulate structured content units, question sets, grading results, diagnostic information, and learning plans into record units, and store each record unit. The adaptive closed-loop control module is used to adjust the question generation strategy, exam preparation focus, and job skills training focus based on the stored results in the next round of learning, forming an adaptive learning closed loop.

[0026] According to a third aspect of the present invention, a computer device is provided.

[0027] In some embodiments, the computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method described above.

[0028] According to a fourth aspect of the present invention, a computer-readable storage medium is provided.

[0029] In one embodiment, a computer program is stored on the computer-readable storage medium, which, when executed by a processor, implements the steps of the above method.

[0030] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: 1) This invention integrates learning material processing, assessment, and subsequent planning into a unified closed loop through the collaborative execution of content processing, question generation, assessment, and learning plan agents, avoiding the problem of multiple independent modules being fragmented in existing technologies. Simultaneously, by employing agent-based collaboration and asynchronous task distribution for different tasks, it reduces main thread blocking and improves the overall processing efficiency of complex learning tasks. Furthermore, by recording session states, learning plan versions, and response results, it achieves continuity and traceability throughout multiple learning rounds, making it suitable for continuous learning scenarios across devices and stages.

[0031] 2) This invention can parse multimodal learning materials into structured content units, providing a unified data foundation for subsequent question generation, assessment, and learning plan generation, thereby improving content consistency across different tasks. After the user answers, it not only outputs the grading results but also further extracts the user's mastery of knowledge points and weaknesses, enabling the assessment results to directly drive adjustments to subsequent learning content and plans. Furthermore, this invention can dynamically adjust subsequent generation strategies based on factors such as the test paper blueprint, knowledge point coverage, accuracy rate, learning feedback, and remaining learning time, making the system output more aligned with the user's current mastery level and learning goals.

[0032] 3) This invention constructs an exam profile by incorporating publicly available information from the official exam website. This allows for the generation of realistic mock exams that are more consistent with the target exam in terms of subject settings, question structure, question quantity, time requirements, and knowledge coverage, thereby improving the relevance and credibility of mock training. Simultaneously, by incorporating publicly available internship information or job competency requirements to construct a job competency profile, the exam preparation path and job competency preparation can be planned in a unified manner, improving the practicality, closed-loop nature, and personalization of the learning path. Furthermore, this invention uses an intent recognition agent to automatically classify user natural language input and extract learning target parameters. This enables the system to automatically select the corresponding agent processing link based on the user's true learning intent, eliminating the need for the user to manually specify the task type and improving human-computer interaction efficiency. The extracted learning target parameters can be directly used as constraints for the generation of downstream agents, allowing the system output to more accurately match user needs.

[0033] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0034] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0035] Figure 1 This is a flowchart illustrating an adaptive learning content processing method based on multi-agent collaboration, according to an exemplary embodiment. Figure 2 This is a block diagram illustrating the principle of an adaptive learning content processing system based on multi-agent collaboration, according to an exemplary embodiment. Figure 3 This is a schematic diagram of the structure of a computer device according to an exemplary embodiment; Figure 4 This is a flowchart illustrating adaptive learning plan adjustment based on assessment results, according to an exemplary embodiment. Figure 5 This is a nine-node closed-loop workflow topology diagram of generating an intelligent agent based on a problem illustrated in an exemplary embodiment; Figure 6 This is a two-node directed acyclic graph workflow topology diagram of a tutoring agent, illustrated according to an exemplary embodiment. Figure 7 This is a flowchart illustrating a unified parsing pipeline for multimodal content, based on an exemplary embodiment. Figure 8 This is a flowchart illustrating user intent recognition and agent routing according to an exemplary embodiment. Detailed Implementation

[0036] The following description and accompanying drawings fully illustrate specific embodiments described herein to enable those skilled in the art to practice them. Some portions and features of certain embodiments may be included in or replace portions and features of other embodiments. The scope of the embodiments herein includes the entire scope of the claims and all available equivalents thereof. The various embodiments described herein are presented in a progressive manner, with each embodiment focusing on its differences from other embodiments; similar or identical parts between embodiments can be referred to interchangeably.

[0037] The modules in the apparatus or system of this application can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0038] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0039] Figure 1 An embodiment of an adaptive learning content processing method based on multi-agent collaboration according to the present invention is shown.

[0040] In this optional embodiment, the adaptive learning content processing method based on multi-agent collaboration includes: S101. Obtain the user's learning materials and learning requests, establish a session identifier associated with the user's learning requests, and perform intent recognition on the user's learning requests, and output the intent recognition results; S102. Obtain publicly available information about the target exam (i.e., information published on the official exam website) and publicly available job information for the target position, and establish exam profiles and job competency profiles respectively; analyze the learning materials, publicly available exam information, and publicly available job information to generate structured content units; S103. Based on the intent recognition results, distribute the structured content units, examination profiles, and job competency profiles to the corresponding processing flows; generate a question set based on the structured content units, examination profiles, and job competency profiles. S104. Output a set of questions to the user and receive the answer results; output the grading results and extract diagnostic information based on the answer results; generate a learning plan based on the diagnostic information, exam profile, and job competency profile. Specifically, the system outputs a set of questions or full-length mock exams to the user and receives the user's answers. The assessment agent module outputs grading results based on the question answers, explanations, and user responses. The diagnostic analysis module extracts knowledge point mastery, weaknesses, job competency gaps, and learning suggestions based on the grading results. The learning plan agent module generates learning plans, exam preparation paths, or internship learning paths based on the diagnostic information, exam profiles, and job competency profiles output by the diagnostic analysis module, or adjusts existing plans.

[0041] S105. Encapsulate the structured content units, question sets, grading results, diagnostic information, and learning plans into record units, and store each record unit. S106. In the next round of learning, adjust the question generation strategy, exam preparation focus, and job skills training focus based on the stored results to form an adaptive learning loop.

[0042] It needs to be explained that the process involves acquiring learning materials and user learning requests, and establishing a session identifier associated with the user's learning request; the intent recognition agent performs intent recognition on the natural language text in the user's learning request, and outputs the intent recognition result; the intent recognition result includes at least the intent type and confidence level, and may further include learning objective parameters extracted from the natural language text; the intent type includes at least one or more of the following: exam preparation planning intent, question generation intent, knowledge tutoring intent, assessment intent, and comprehensive learning intent; the learning objective parameters include at least one or more of the following: target exam name, exam preparation deadline, expected learning intensity, target question type, target number of questions, and target knowledge scope.

[0043] Obtain publicly available information related to the target exam and the target internship or job position, and create exam profiles and job competency profiles respectively. The publicly available information should include at least one or more of the following: exam syllabus, exam subjects, question structure, exam time, registration rules, sample questions and papers, and scoring rules published on the official exam website; and one or more of the following from the job posting page: job responsibilities, competency requirements, skill keywords, internship period, and recruitment process.

[0044] The content processing agent parses learning materials, publicly available exam information, and publicly available job postings to generate structured content units; the learning materials include at least one or more of the following: PDF documents, web page text, audio-transcribed text, video-transcribed text, or exam paper text.

[0045] The agent orchestration module distributes structured content units, exam profiles, and job competency profiles to at least one of the following agents—question generation agent, assessment agent, tutoring agent, and learning plan agent—for processing based on the intent type and learning objective parameters in the intent recognition results. Specifically, when the intent type is exam preparation planning, the learning plan agent is scheduled first; when the intent type is question generation, the question generation agent is scheduled first, and the target question type, target number of questions, and target knowledge scope from the learning objective parameters are passed in as generation constraints; when the intent type is knowledge tutoring, the tutoring agent is scheduled first; when the intent type is assessment, the assessment agent is scheduled first; when the intent type is comprehensive learning intent or contains multiple intents, multiple agents are scheduled to work together in the order of content parsing, question generation, assessment, and plan generation.

[0046] It needs to be explained that the two-node directed acyclic graph workflow of the tutoring agent has the following topology: retrieval node - generation node - termination, as follows: Figure 6 As shown.

[0047] The workflow transmits data between nodes through a unified state object. The state object contains at least the following fields: a list of dialogue history messages, context text generated by retrieval enhancement, a list of retrieved original document fragments (each fragment contains content, metadata, and similarity scores), the current user's question, the answer generated by a pre-trained large language model (such as Deepseek V4), the number of query rewrites (maximum 2), the user identifier (for data isolation), the session identifier (for session isolation for vector retrieval), the retrieval mode identifier (supporting three modes: vector retrieval, hybrid retrieval, and full document retrieval), and a list of specified filenames (for limiting the retrieval scope).

[0048] The processing flow of the retrieval node is as follows: It receives user questions and selects the corresponding retrieval strategy based on the retrieval mode identifier; it converts the user questions into vector representations using a multilingual embedding model (e.g., calling the BGE-M3API); it performs semantic similarity searches in the vector database based on session isolation conditions, i.e., only on files submitted by the user and usage records, to obtain relevant document fragments; if a list of filenames is specified, it further filters the search results; it concatenates the retrieved document fragments into retrieval enhancement context text and writes it to the state object.

[0049] The processing flow of the generated node is as follows: read the enhanced generation context and dialogue history from the state object; construct a complete prompt containing system prompt words, context and user questions; call a pre-trained large language model (such as DeepseekV4) to generate an answer, supporting streaming output to the message queue for real-time display on the front end; and write the generated result to the state object.

[0050] The question-generating agent extracts knowledge points, test points, question type features, or test paper blueprint information based on structured content units, exam profiles, job competency profiles, and learning objective parameters, and generates a question set containing questions, answers, explanations, and question attribute information. When the question set is used for a mock exam, the question set is organized into a full-fledged mock exam that matches the target exam in terms of question type distribution, number of questions, answering time, knowledge coverage, and output format.

[0051] Specifically, such as Figure 4 As shown, after a user completes one round of testing, the assessment agent module outputs a question-by-question grading result. The diagnostic analysis module performs a correlation analysis between the grading results and the question content to identify knowledge points that are well-mastered, knowledge points that are not well-mastered, and suggested learning content for reinforcement. The learning plan agent module generates an adjusted learning plan based on diagnostic information, user feedback, progress, and remaining learning time. For example, if a user reports that the current learning pace is too fast, the learning plan agent module can extend the original seven-day learning plan to ten days; when the system identifies a low accuracy rate for a certain type of knowledge point, corresponding review arrangements and basic exercises can be added to subsequent plans.

[0052] The intent recognition results, structured content units, question sets, grading results, diagnostic information, and learning plans are associated with and persistently stored with the session identifiers for subsequent tasks to access. In subsequent learning rounds, based on the diagnostic information from the previous round, the learning plan adjustment results, and the historical intent recognition results, the learning content to be processed, question generation strategies, question difficulty, question type distribution, question quantity, exam preparation focus, job skills training focus, or plan arrangements are redefined, thereby forming an integrated adaptive closed loop of exam preparation, internship learning, and personalized growth.

[0053] Specifically, structured content units should include at least one or more of the following: text content, heading level, page number information, source information, chapter boundaries, session identifier, and content identifier. Exam profiles should include at least one or more of the following: exam name, exam subject, question type, score structure, exam duration, registration deadline, exam deadline, sample question characteristics, knowledge point scope, and scoring rules. Job competency profiles should include at least one or more of the following: job title, job responsibilities, skill requirements, competency keywords, internship period, work task category, and recruitment process information. The intelligent agent orchestration module submits content parsing tasks to the first asynchronous task queue and dialogue, learning plan generation, or assessment-related tasks to the second asynchronous task queue to improve the processing efficiency of different tasks. Question attribute information should include one or more of the following: question type, answer, explanation, knowledge point, difficulty level, and question source pattern.

[0054] Specifically, when the learning materials are test paper texts, the question generation agent first extracts the test paper blueprint information, and then constrains the question type distribution and output format according to the test paper blueprint information to generate a set of imitation questions. The result should include at least one or more of the following information: question number, correct / incorrect result, scoring result, question comments, and question-level answer records. Diagnostic information is obtained by correlating the question content with the grading results. Diagnostic information should include at least one or more of the following: main knowledge points, current mastery level, weak areas, and subsequent learning suggestions. The diagnostic information further includes job competency gap information, used to characterize the gap between the user's current ability and the target job competency profile. The learning plan, preparation path, or internship learning path should include at least one or more of the following: daily task arrangement, knowledge point review arrangement, mock exam arrangement, incorrect question review arrangement, job competency training tasks, project practice tasks, and phased goals. Persistent storage should include at least: storing the question set, learning plan, and version history by session identifier, and supporting cross-device synchronous access. Adaptive adjustment includes at least one of the following: 1) Adjust the coverage ratio of knowledge points in subsequent questions based on the students' mastery of those knowledge points. 2) Adjust the difficulty level of subsequent questions based on accuracy or areas of weakness. 3) Adjust the number of days, phase allocation, or review schedule in the learning plan based on user feedback. 4) Compress, expand, or rearrange the remaining learning content based on the progress made. 5) Adjust the focus of internship learning tasks and project practice tasks based on the gaps in job skills.

[0055] In this optional embodiment, when acquiring the user's learning materials and learning requests, establishing a session identifier associated with the user's learning requests, and performing intent recognition on the user's learning requests and outputting the intent recognition results, the system can receive the user's input learning requests and corresponding learning materials. The learning requests include natural language text, user identification information, and timestamp information. The learning requests are preprocessed to standardize and generate standardized request text. Based on the user identification information, timestamp information, and standardized request text, a unique initial session identifier is generated. A session context data structure corresponding to the initial session identifier (session identifier ID) is constructed, and an association mapping relationship is established between the initial session identifier and the learning materials and stored. The standardized request text is input into a preset prompt word template, and a structured prompt information containing task description, intent definition, and output format constraints is constructed. The structured prompt information is input into a pre-trained language model for semantic understanding and reasoning processing, generating a structured output result, which includes the intent type and corresponding confidence information. The pre-trained large language model parses the text materials provided by the user, extracts the learning target parameters corresponding to the intent type, and outputs the intent type, confidence, and learning target parameters.

[0056] Specifically, the system receives learning materials and user learning requests, and creates a session identifier corresponding to the learning request. The system receives the user's input learning request and corresponding learning materials, where the learning request includes natural language text, user identification information, and timestamp information. The learning request undergoes standardized preprocessing, including text cleaning, encoding standardization, and illegal character filtering, generating a standardized request text. Based on this, a unique initial session identifier ID is generated using the user identification information, timestamp information, and standardized request text. Further, a session context data structure corresponding to the session identifier ID is constructed. This session context data structure includes at least the user identifier, learning objective, historical interaction records, task status identifier, and current processing stage identifier. An association mapping relationship is established between the session identifier ID and the learning materials, and this association is stored in the session storage unit within the session management module to support context calls during subsequent multi-agent collaborative processing.

[0057] It needs to be explained that the intent recognition agent module performs intent recognition on the natural language text in the user's learning request. It acquires the normalized request text and inputs it into a preset prompt word template to construct structured prompt information containing task descriptions, intent definitions, and output format constraints. This structured prompt information is then input into a small-parameter pre-trained language model (e.g., Amazon Nova Micro) for semantic understanding and reasoning. The small-parameter pre-trained language model (e.g., Amazon Nova Micro) determines the user's intent based on contextual semantics (e.g., the user wants to generate a summary output) and generates a structured output result, which includes the intent type and corresponding confidence information. A pre-trained large language model (e.g., Deepseek V4) parses the text provided by the user, extracting learning target parameters corresponding to the intent type. These learning target parameters include at least the learning content scope, learning stage, difficulty level, and task type, thereby achieving intent recognition of the user's learning request and outputting the intent type, confidence level, and learning target parameters.

[0058] Specifically, when a user inputs natural language text through the front-end interaction layer (such as input in a chat box or voice-to-text input), the intent recognition agent module executes the following processing flow, such as... Figure 8 As shown: Step I1 (Text Preprocessing): Clean and standardize the user input text, including removing redundant whitespace characters, standardizing punctuation marks, and identifying and annotating entity information in the text (such as exam name, date, quantity, etc.).

[0059] Step I2 (Intent Classification): The preprocessed text, along with a predefined list of intent types and classification rules, is used to construct structured prompts. A pre-trained language model with a small parameter set (e.g., Amazon Nova Micro) is then invoked for inference, outputting the intent recognition result in JSON format. The intent recognition result includes the following fields: intent_type: The intent graph type, which can be one of exam_prep (exam preparation planning), question_gen (question generation), tutoring (knowledge tutoring), grading (assessment grading), or comprehensive (comprehensive learning).

[0060] confidence: The confidence level of the idea graph, with a value ranging from 0 to 1.

[0061] secondary_intents: A list of secondary intents that are populated when user input contains multiple intents.

[0062] Reasoning: Explanation of the reasoning basis for intent judgment in pre-trained large language models (such as Deepseek V4).

[0063] Step I3 (Learning Target Parameter Extraction): In the same or subsequent inference iterations of the pre-trained large language model (such as Deepseek V4), extract the learning target parameters from the user input text. The learning target parameters are structured JSON objects and may contain the following fields: target_exam: The name of the target exam (e.g., CFA Level 1, Certified Public Accountant, etc.).

[0064] deadline: the deadline for exam preparation or the exam date.

[0065] intensity: the desired intensity of learning (e.g., 2 hours per day, high intensity, etc.).

[0066] target_question_types: A list of expected question types (such as multiple choice, short answer, etc.).

[0067] target_question_count: The expected number of questions.

[0068] target_knowledge_scope: The scope of knowledge or list of chapters expected to be covered.

[0069] additional_requirements: Other special requirements proposed by the user.

[0070] When the user's input text does not contain a certain parameter, the corresponding field will be empty.

[0071] Step I4 (Intent Routing Decision): After receiving the intent recognition result, the agent orchestration module selects the corresponding agent scheduling strategy based on the intent_type field. When intent_type is exam_prep, the learning plan agent module is scheduled, and target_exam, deadline, and intensity are passed in as constraint parameters for plan generation.

[0072] When intent_type is question_gen, the question generation agent module is scheduled, and target_question_types, target_question_count, and target_knowledge_scope are passed in as generation constraints.

[0073] When intent_type is tutoring, the tutoring agent is scheduled and target_knowledge_scope is passed in as the retrieval scope constraint.

[0074] When intent_type is grading, schedule the evaluation agent module.

[0075] When intent_type is comprehensive or secondary_intents exist, multiple agents are scheduled in the order of content parsing, question generation, evaluation, and plan generation.

[0076] Step I5 (Persistence of Intent Results): The intent recognition result is associated with the session identifier and stored for reference in subsequent learning rounds, so as to achieve continuity of intent across rounds.

[0077] It should be explained that the following are typical examples of user intent recognition scenarios: Scenario 1: The user inputs "Help me create a 30-day study plan for CFA Level 1, studying 3 hours per day." The intent recognition result is: intent_type=exam_prep, target_exam=CFALevel1, deadline=30 days later, intensity=3 hours per day. The system prioritizes scheduling the study plan agent, which generates a 30-day study plan based on the CFA Level 1 exam profile and user constraints.

[0078] Scenario 2: The user inputs "Give me 20 multiple-choice questions based on this textbook, focusing on the fixed income chapter." The intent recognition result is: intent_type=question_gen, target_question_types=[multiple-choice questions], target_question_count=20, target_knowledge_scope=[fixed income]. The system schedules a question-generating agent, using the above parameters as generation constraints.

[0079] Scenario 3: The user inputs, "I want to review the content on derivative pricing and then do some practice questions to test myself." The intent recognition result is: intent_type=comprehensive, secondary_intents=[tutoring,question_gen], target_knowledge_scope=[derivative pricing]. The system first schedules a tutoring agent to explain the knowledge, and then schedules a question generation agent to generate practice questions.

[0080] In this optional embodiment, when distributing structured content units, exam profiles, and job competency profiles to the corresponding processing flows based on the intent recognition results, the corresponding processing flows can be scheduled according to the intent type and learning objective parameters. When the intent type is question generation intent, the structured content units are semantically parsed and knowledge units are divided to obtain a set of candidate knowledge points. The candidate knowledge point set is then filtered and weighted according to the exam profile and job competency profile to determine the target knowledge point set. When the intent type is exam preparation planning intent, the learning objective parameters are parsed to determine the learning cycle, target knowledge scope, and ability improvement goals. The learning content is prioritized according to the exam profile and job competency profile to generate a knowledge point learning sequence. Based on the learning cycle, the knowledge point learning sequence is allocated time and divided into stages to construct a learning plan framework containing daily or stage learning tasks. The corresponding learning activity arrangements and review strategies are generated according to the task type to form a structured exam preparation plan. When the intent type is comprehensive learning intent, multiple processing flows are scheduled sequentially according to task dependencies, and the processing results of each stage are written into the session context.

[0081] In this optional embodiment, when generating a question set based on structured content units, exam profiles, and job competency profiles, question generation constraint information can be constructed according to the difficulty level, question type requirements, and examination scope in the learning objective parameters. The target knowledge point set and the question generation constraint information are then input into a preset prompt word template and input into a pre-trained language model for question generation processing to obtain an initial question set. The initial question set is then subjected to format standardization and content consistency verification, and the questions are assembled according to a preset test paper structure template to generate the final question set.

[0082] Specifically, the content processing agent module parses learning materials, publicly available exam information, and publicly available job postings to obtain structured content units. The agent orchestration module then schedules the corresponding agents based on the output intent type and learning objective parameters. When the intent type is "question generation," the agent orchestration module schedules the question generation agent module to execute the processing flow. The structured content units are semantically parsed and divided into knowledge units to obtain a set of candidate knowledge points. These candidate knowledge points are then filtered and ranked by weight based on exam profiles and job competency profiles to determine the target knowledge point set. Question generation constraints are constructed based on the difficulty level, question type requirements, and examination scope in the learning objective parameters. The target knowledge point set and question generation constraints are then input into a preset prompt word template and fed into a pre-trained large language model (such as Deepseek V4) for question generation, resulting in an initial question set. The initial question set is further formatted and its content is verified for consistency. The questions are then assembled according to a preset test paper structure template to generate a question set or a full-fledged mock test paper. When the intent type is exam preparation planning, the agent orchestration module prioritizes scheduling the learning plan agent module to execute the processing flow. The learning objective parameters are analyzed to determine the learning cycle, target knowledge scope, and ability improvement goals. Learning content is prioritized based on exam profiles and job competency profiles, generating a knowledge point learning sequence. Based on the learning cycle, the knowledge point learning sequence is allocated time and divided into stages, constructing a learning plan framework that includes daily or stage-based learning tasks. Corresponding learning activity arrangements and review strategies are generated based on task types, forming a structured exam preparation plan. When the intent type is a comprehensive learning intent, the agent orchestration module, based on a preset multi-agent collaboration strategy, sequentially schedules multiple agent modules to process the learning request in stages according to task dependencies, and writes the processing results of each stage into the session context to achieve multi-agent collaborative processing.

[0083] It should be explained that the multimodal content parsing pipeline of the content processing intelligent agent module adopts a four-stage processing flow: format recognition, dedicated parser, unified output, and vectorized storage. Figure 7 As shown: Phase 1 (Format Recognition): Identify the format category of the input data based on the file extension and MIME type.

[0084] The second stage (dedicated parser processing): The PDF document undergoes layout analysis through a document intelligent parsing engine. First, page-level parsing is performed, breaking the document down into multiple page images or layout objects. Then, layout structure detection is performed on each page, identifying different regions within the page based on a pre-trained layout analysis model (e.g., using MinerU). The API is used for parsing, and the regions are categorized and labeled to obtain layout elements such as text paragraph regions, table regions, and image regions. Further text extraction processing is performed on the text paragraph regions, including character recognition and paragraph reorganization, to obtain continuous text content. Table regions are parsed to analyze the table structure, identify row and column structures, and extract cell text information. Image regions are extracted or OCR is used to recognize text information. Based on this, the extracted results are sorted and reorganized according to the position and hierarchy of each region on the page, generating a structured content representation containing text paragraphs, table data, and image descriptions. Audio tracks are extracted from video files and transcribed into timestamped text using a speech recognition engine. Online video links prioritize obtaining subtitle files provided by the platform (such as MOOCs); if no subtitles are available, the audio is downloaded and transcribed into speech. Presentations render each slide as an image and then perform OCR text recognition. Audio files are directly transcribed into text using a speech recognition engine.

[0085] The third stage (unified output): The dedicated parser uses OCR text recognition and speech transcription technologies to convert multimodal data into text. Then, the text results are converted into structured text in Markdown format, preserving format information such as heading levels, paragraph separators, and list structures. The text is then divided into blocks, and each text block is accompanied by metadata (source file, page number or timestamp, block number, etc.).

[0086] The fourth stage (vectorized storage): Text blocks are converted into high-dimensional vector representations through a multilingual embedding model, document vectors and their metadata are stored in a vector database, and indexes are built according to user and session dimensions to achieve data isolation and semantic retrieval.

[0087] In this optional embodiment, after constructing question generation constraint information based on the difficulty level, question type requirements, and examination scope in the learning objective parameters, and inputting the target knowledge point set and question generation constraint information into a preset prompt word template, the information is then input into a pre-trained language model for question generation processing. When the initial question set is obtained, the parsed content of the user-uploaded file can be read from the object storage service or cache, the parsed content can be organized into a text block sequence and written into the content text block field of the state object; an inference call is performed on the input material, and the classification decision result is output based on the file extension matching processing mode provided by the user; the classification decision result includes pure extraction The study materials include four modes: demonstration mode, past exam imitation mode, syllabus alignment mode, and knowledge point alignment mode. In the syllabus alignment and past exam imitation modes, word segmentation and semantic analysis are performed on the past exam text and syllabus text to identify candidate knowledge point sets. These candidate knowledge points are then matched with a pre-defined knowledge point dictionary or knowledge structure to determine standardized knowledge point names. Based on the past exam text, the frequency of each knowledge point's appearance in the questions, its question type distribution, and its score percentage are statistically analyzed. An importance score is calculated for each knowledge point to obtain a weight value representing its importance. In the knowledge point alignment mode, the study material text is divided into blocks according to a pre-defined length to obtain multiple... The system generates text blocks and assigns a block identifier to each text block. For each text block, it constructs a prompt word template containing knowledge point definitions, output format constraints, and example descriptions. The text block and the prompt word template are combined to form a structured input, which is then input into a pre-trained large language model for semantic parsing and knowledge point extraction, resulting in a knowledge point list for the corresponding text block. Based on the type of user input data, a corresponding knowledge point extraction strategy is used to generate a target knowledge point set. The knowledge point lists output from all text blocks are then merged, and duplicate or semantically similar knowledge points are deduplicated and merged using semantic similarity calculation to obtain the target knowledge point set. The system first identifies a set of knowledge points and generates a question-generating plan based on the target set of knowledge points and the target number of questions. This plan allocates the number of questions, question types, and difficulty distribution for each knowledge point and assigns the knowledge points to multiple question-generating sub-tasks. Based on the question-generating plan, multiple processing units are concurrently invoked to execute the question generation process. Each processing unit receives a corresponding task object, constructs question generation prompts based on the task object, and calls a pre-trained large language model to generate a batch of questions at once. The generated results are output in a structured data format, including the question stem, options, answer, explanation, relevant knowledge point, difficulty, and score fields. This process yields the initial question set.

[0088] In this optional embodiment, when formatting and verifying the consistency of the initial question set, and assembling the questions according to a preset test paper structure template to generate the final question set, an independently configured pre-trained large language model instance can be used to solve each question independently, and the solution results are written to the solution verification result field of the state object; the solution results are compared with the generated answers to calculate the self-solved accuracy rate, and a quality evaluation is generated for each question, written to the scoring result and scoring summary fields of the state object; based on the scoring result and scoring summary fields, a comprehensive quality review is conducted on the generated results, checking whether the number of questions is consistent with the corresponding real exam questions, whether the self-solved accuracy rate meets the standard, and whether the question type distribution is reasonable; in the knowledge point alignment mode, the knowledge point coverage rate is additionally checked, and when the coverage rate is lower than the preset threshold and the current review retry counter is zero, it will be necessary to retry. The generation flag is set to true, the retry counter is incremented, and the pre-trained large language model is called again to process the questions. If the question generation fails after one retry, a warning message is added before allowing the question to proceed. The question set is then parsed in a structured manner, with each question broken down into the stem, options, answer, explanation, related knowledge point, difficulty, and score fields. Each field is formatted according to a preset output template to generate a standardized question data structure. In syllabus alignment mode, the pre-built complete set of syllabus knowledge points is obtained. Based on the knowledge point identifier associated with each question, the complete set of syllabus knowledge points is matched to obtain the set of covered knowledge points, and the set of uncovered knowledge points and the coverage ratio are calculated. The coverage ratio, the list of covered knowledge points, and the list of uncovered knowledge points are encapsulated in a structured manner and written together with the standardized question data structure into the final output field of the state object to generate the final question set.

[0089] It should be explained that the question generation agent module uses a nine-node closed-loop directed acyclic graph workflow to generate the complete exam paper. The workflow topology is as follows: parsing node - classification node - knowledge point extraction node - deep knowledge point extraction node - generation node - problem-solving node - scoring node - review node - output node. The review node can trigger a conditional backtracking to the generation node, forming a quality assurance closed loop.

[0090] The workflow transmits data between nodes through a unified typed state object. The state object must contain at least the following fields: a list of dialogue history messages, user identifier, session identifier, task identifier, a list of content text blocks, a list of knowledge points, a list of original questions, a list of solution verification results, a scoring result, a review response, self-solved accuracy rate, a requirement matching flag, a quality pass flag, a regeneration requirement flag, a review retry counter, and the final output result. The specific processing logic of each node is as follows: Figure 5 As shown: Step E1 (parse node): Read the parsed content of the user-uploaded file from the object storage service or cache, organize the content into a sequence of text blocks and write it to the content text block field of the state object.

[0091] Step E2 (Classification Node): The Nova Lite model performs an inference call on the input material, matches the processing pattern based on the file extension provided by the user, and outputs the classification decision result in JSON format. The classification decision includes four generation modes: Pure prompt mode: Users only provide text descriptions and do not upload files.

[0092] Real Exam Simulation Mode: Users upload one or more past exam papers.

[0093] Exam Syllabus Alignment Mode: Users upload exam syllabi.

[0094] Knowledge point alignment mode: Users upload learning materials (not past exam questions, not exam syllabus).

[0095] The category node simultaneously outputs the number of target questions, the role label of each file (past exam questions / syllabus / study materials), and the explanation of the reasoning basis.

[0096] Step E3 (Knowledge Point Extraction Node): For the syllabus alignment mode and the past exam question imitation mode, the past exam question text and the syllabus text are processed by word segmentation and semantic analysis to identify a set of candidate knowledge points. These candidate knowledge points are then matched with a pre-set knowledge point dictionary or knowledge structure to determine standardized knowledge point names. Subsequently, based on the past exam question text, the frequency of each knowledge point appearing in the questions, the distribution of question types, and the percentage of marks are statistically analyzed. An importance score is calculated for each knowledge point, including a frequency weight based on its appearance frequency and a contribution weight based on the question's mark. Specifically: 1) Frequency weighting based on occurrence frequency This is used to characterize the frequency of occurrence of a knowledge point in the test question set, and it is calculated as follows: ,in, Representing knowledge points Number of times it appears in all test questions This indicates the total number of knowledge points.

[0097] 2) Contribution weight based on question score This is used to characterize the contribution of the knowledge point to the total exam score, and it is calculated as follows: ;in, Indicates the knowledge points involved A collection of questions, Indicates the topic The score.

[0098] The multi-source weights of each knowledge point are further summed to obtain initial weight values, and then normalized to map the weight values ​​of each knowledge point to a preset range. These weight values ​​represent the importance of the corresponding knowledge point in the current exam task. The larger the value, the more important the knowledge point is in the current exam task; when When the time is specified, it indicates that the knowledge point is marginal or low-frequency test content. This indicator is used to support priority control in the subsequent question generation and test paper structure configuration process; when in other modes, this node directly transmits the status without performing knowledge point extraction and weight calculation processing.

[0099] Step E4 (Deep Knowledge Point Extraction Node): Executed only in knowledge point alignment mode. The learning material text is divided into blocks according to a preset length, resulting in multiple text blocks, and each text block is assigned a block identifier. Then, a prompt word template containing knowledge point definitions, output format constraints, and example descriptions is constructed for each text block. The text block and prompt word template are combined to form structured input, which is then input into a pre-trained large language model (such as Deepseek V4) for semantic parsing and knowledge point extraction. The output is a list of knowledge points for the corresponding text block, where each knowledge point is represented in a structured form, including at least the knowledge point name and its contextual description. Different knowledge point extraction strategies are used to generate the target knowledge point set depending on the type of user input material, specifically including the following modes: When the user does not upload a file and only enters a description of the exam type, the system obtains the description of the exam type and additional requirements entered by the user, and constructs a prompt message containing the exam type, question objectives and output constraints. The prompt message is then input into a pre-trained large language model for knowledge point reasoning processing. Exam-related information is obtained through web search, and the pre-trained large language model (such as Deepseek V4) generates a set of target knowledge points based on the semantics of the exam type.

[0100] When a user uploads past exam papers or mock exam papers, the text of the past exam papers is extracted and the content is merged. The merged text is then truncated. The past exam paper text is input into a pre-trained large language model for semantic parsing and knowledge point extraction, resulting in a set of knowledge points corresponding to the past exam papers. At the same time, the pre-trained large language model (such as Deepseek V4) further outputs a set of core knowledge point identifiers as a set of mandatory knowledge points to be covered. The number of mandatory knowledge points to be covered is limited based on the target number of questions.

[0101] When a user uploads an exam syllabus file, the syllabus text is input into a pre-trained large language model (such as Deepseek V4) to extract knowledge points, resulting in a set of exam syllabus knowledge points.

[0102] When a user uploads only textbooks or learning materials, the textbook text is divided into blocks to generate multiple text blocks; then, a multi-threaded concurrent approach is used to call a pre-trained large language model to extract knowledge points from each target text block.

[0103] Further processing involves merging the knowledge point lists output from all text blocks. This is done by deduplicating and merging duplicate or semantically similar knowledge points using string matching or semantic similarity calculations. The semantic distance between each knowledge point is then calculated. Assuming there are... Each knowledge point will be similar. Combining knowledge points There are several clusters (the number of knowledge points in each cluster varies, determined by semantic distance, among which...) The target knowledge point set is obtained; the target knowledge point set is written into the knowledge point target field of the status object according to the preset data structure, wherein the status object includes the current processing stage identifier and knowledge point related data fields for subsequent processing nodes to call.

[0104] Step E5 (Generation Node): After obtaining the knowledge point extraction results, the question generation node executes the question generation process. Based on the knowledge point list and the target number of questions, it generates a question generation execution plan, allocates the number of questions, question type requirements, and difficulty distribution corresponding to each knowledge point, and assigns the knowledge points to multiple question generation sub-tasks. The concurrent question generation node concurrently calls multiple workers to perform question generation processing based on the question generation execution plan. Each worker receives a corresponding task object, which includes at least the target knowledge point, corresponding number of questions, question type requirements, difficulty distribution, and style indication information. Each worker constructs question generation prompts based on the task object and calls the pre-trained large language model to generate the corresponding batch of questions at once. The generation results are output in a structured data format, including at least the question stem, options, answer, explanation, knowledge point, difficulty, and score fields.

[0105] Step E6 (Solving Node): Using an independently configured pre-trained large language model (such as Deepseek V4) instance (with a different model role configuration than the generating node), solve each problem independently and write the solution results to the solving verification result field of the state object. This node uses a different model instance than the generating node to ensure the independence of solving verification and the effectiveness of cross-validation.

[0106] Step E7 (Scoring Node): Compare the solution results from the problem-solving node with the standard answers given by the generation node using a pre-trained large language model (such as Deepseek V4), calculate the self-solving accuracy (i.e., the consistency between the solution results and the analysis given by the generation node), and generate a quality evaluation for each question, writing the scoring results and scoring summary fields into the state object.

[0107] Step E8 (Review Node): Perform a comprehensive quality review on the generated results. Checks include: whether the number of questions matches the corresponding actual exam questions, whether the accuracy rate of self-solved questions meets the standard (questions with errors in self-solved questions will be corrected), and whether the question type distribution is reasonable. In knowledge point alignment mode, an additional check of knowledge point coverage is performed. When the coverage rate is lower than the preset threshold (80%) and the current review retry counter is 0, the "need to regenerate" flag is set to true, and the retry counter is incremented by 1. The workflow then backtracks to the generation node via a conditional edge to re-execute steps E5 to E8. If the result still fails to meet the standard after one retry, a warning message is added before proceeding to the output node. The conditional edge's judgment logic is: read the "need to regenerate" flag from the state object; if true, route to the generation node; otherwise, route to the output node.

[0108] Step E9 (Output Node): The final question list is structured and parsed, breaking down each question into fields such as stem, options, answer, explanation, and associated knowledge point identifiers. Each field is then formatted according to a preset output template, including field completion, encoding standardization, and anomaly data validation, generating a standardized question data structure. In syllabus alignment mode, a pre-built complete set of syllabus knowledge points is obtained, and the entire set is matched against the associated knowledge point identifiers in each question to obtain the set of covered knowledge points. Further, set operations are used to calculate the set of uncovered knowledge points, and the coverage ratio is calculated based on the quantitative relationship between the covered knowledge point set and the complete syllabus knowledge point set, where the coverage ratio is the ratio of the number of covered knowledge points to the total number of syllabus knowledge points. Based on this, the coverage ratio, the list of covered knowledge points, and the list of uncovered knowledge points are structurally encapsulated and written into the final output field of the state object along with the standardized question data structure for the caller to read and display.

[0109] In this optional embodiment, when encapsulating structured content units, question sets, grading results, diagnostic information, and learning plans into record units and storing each record unit, a version identifier can be generated for the current round of learning process based on the corresponding session context data structure obtained from the session identifier. The structured content units, question sets, grading results, diagnostic information, and learning plans are respectively structured and encapsulated to form record units containing data type identifiers, generation timestamps, and source identifiers. Using the session identifier as the primary key and the version identifier as the subkey, a multi-type data association index is constructed in the session context data structure. Each record unit is written into its corresponding index node, and reference relationships and time sequence relationships are established between each record unit to characterize the evolution process of processing results at different stages within the same session. The session identifier, version identifier, and corresponding record units are stored according to a preset data storage structure. The data storage structure includes key-value pair or document structure storage formats to support fast retrieval based on the session identifier, historical backtracking based on the version identifier, and data consistency maintenance across multiple rounds of learning processes.

[0110] Specifically, the session and version management module obtains the corresponding session context data structure based on the session identifier and generates a version identifier for the current processing round. It then structurally encapsulates the structured content units, question sets, grading results, diagnostic information, and learning plans to form record units containing data type identifiers, generation timestamps, and source module identifiers. Using the session identifier as the primary key and the version identifier as the subkey, it constructs a multi-type data association index in the session context data structure, writes each record unit into its corresponding index node, and establishes reference relationships and time sequence relationships between record units to characterize the evolution of processing results at different stages within the same session. Furthermore, the persistent storage module stores the session identifier, version identifier, and corresponding record units according to a preset data storage structure. This data storage structure includes key-value pair or document-based storage formats to support fast retrieval based on session identifiers, historical backtracking based on version identifiers, and data consistency maintenance across multiple learning rounds.

[0111] Figure 2 An embodiment of an adaptive learning content processing system based on multi-agent collaboration according to the present invention is shown.

[0112] In this optional embodiment, the adaptive learning content processing system based on multi-agent collaboration includes: The input and session management module 201 is used to obtain the user's learning materials and learning requests, establish a session identifier associated with the user's learning requests, and perform intent recognition on the user's learning requests, and output the intent recognition results. The content analysis and profile building module 202 is used to obtain publicly available information about the target exam and the target position, and to build exam profiles and job competency profiles respectively; it analyzes learning materials, publicly available exam information and publicly available position information to generate structured content units. The distribution and question generation module 203 is used to distribute structured content units, examination profiles, and job competency profiles to the corresponding processing flows based on the intent recognition results; and to generate a question set based on the structured content units, examination profiles, and job competency profiles. The assessment diagnosis and plan generation module 204 is used to output a set of questions to the user and receive the answer results; output the grading results and extract the diagnostic information based on the answer results; and generate a learning plan based on the diagnostic information, the exam profile, and the job competency profile. The state management and persistence module 205 is used to encapsulate structured content units, question sets, grading results, diagnostic information and learning plans into record units, and store each record unit; The adaptive closed-loop control module 206 is used to adjust the question generation strategy, exam preparation focus, and job skills training focus based on the stored results in the next round of learning, forming an adaptive learning closed loop.

[0113] It should be explained that the input and conversation management module 201 includes a data receiving module and an intent recognition agent module. The content parsing and profile building module 202 includes a content processing agent module. The distribution and question generation module 203 includes an agent orchestration module and a question generation agent module. The assessment diagnosis and plan generation module 204 includes an assessment agent module, a diagnostic analysis module, and a learning plan agent module. The state management and persistence module 205 includes a conversation and version management module and a persistent storage module. The adaptive closed-loop control module 206 includes a result output module.

[0114] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 3 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory 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 stores static and dynamic information data. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the above method embodiments.

[0115] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the computer device to which the present invention is applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0116] In addition, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0117] In addition, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0118] Those skilled in the art will understand that all or part of the processes in the methods of 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, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0119] This invention is not limited to the structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this invention is limited only by the appended claims.

Claims

1. A method for adaptive learning content processing based on multi-agent collaboration, characterized in that, include: Obtain the user's learning materials and learning requests, and establish a session identifier associated with the user's learning requests; It also performs intent recognition on user learning requests and outputs the intent recognition results; Obtain publicly available information on the target exam and the target position, and create exam profiles and job competency profiles respectively. The learning materials, publicly available exam information, and publicly available job postings are analyzed to generate structured content units; Based on the intent recognition results, the structured content units, exam profiles, and job competency profiles are distributed to the corresponding processing flows. Based on structured content units, exam profiles, and job competency profiles, a set of questions is generated; Output a set of questions to the user and receive the answers; output grading results and extract diagnostic information based on the answers; generate a learning plan based on the diagnostic information, exam profile, and job competency profile. The structured content units, question sets, grading results, diagnostic information, and learning plans are each encapsulated into a record unit, and each record unit is stored. In the next round of learning, the question generation strategy, exam preparation focus, and job skills training focus will be adjusted based on the stored results to form an adaptive learning loop.

2. The adaptive learning content processing method based on multi-agent collaboration according to claim 1, characterized in that, The process involves acquiring the user's learning materials and learning requests, and establishing a session identifier associated with the user's learning requests. It also performs intent recognition on the user's learning request and outputs the intent recognition results, including: It receives learning requests and corresponding learning materials input by users. The learning requests include natural language text, user identification information, and timestamp information. The learning requests are preprocessed to standardize them and generate standardized request text. A unique initial session identifier is generated based on user identification information, timestamp information, and normalized request text; Construct a session context data structure corresponding to the initial session identifier, establish an association mapping relationship between the initial session identifier and the learning materials, and store it. The standardized request text is input into a preset prompt word template, and a structured prompt message containing task description, intent definition and output format constraints is constructed. The structured prompts are input into a pre-trained language model for semantic understanding and reasoning, generating structured output results, which include intent type and corresponding confidence information. The pre-trained large language model parses the text data provided by the user, extracts the learning target parameters corresponding to the intent type, and outputs the intent type, confidence level, and learning target parameters.

3. The adaptive learning content processing method based on multi-agent collaboration according to claim 1, characterized in that, The step of distributing structured content units, exam profiles, and job competency profiles to the corresponding processing flows based on the intent recognition results includes: The corresponding processing flow is scheduled according to the intent type and learning target parameters; When the intent type is question generation intent, the structured content unit is semantically parsed and knowledge unit is divided to obtain a set of candidate knowledge points. The candidate knowledge point set is then filtered and weighted by combining the exam profile and job competency profile to determine the target knowledge point set. When the intent type is exam preparation planning intent, the learning objective parameters are analyzed to determine the learning cycle, target knowledge scope and ability improvement goals, and the learning content is prioritized in combination with the exam profile and job competency profile to generate a knowledge point learning sequence. Based on the learning cycle, the learning sequence of knowledge points is allocated time and divided into stages, a learning plan framework containing daily or stage learning tasks is constructed, and corresponding learning activity arrangements and review strategies are generated in combination with task types to form a structured exam preparation plan. When the intent type is a comprehensive learning intent, multiple processing flows are scheduled sequentially according to task dependencies, and the processing results of each stage are written into the session context.

4. The adaptive learning content processing method based on multi-agent collaboration according to claim 1, characterized in that, The question set generated based on structured content units, exam profiles, and job competency profiles includes: Based on the difficulty level, question type requirements, and examination scope in the learning objective parameters, question generation constraint information is constructed. The target knowledge point set and question generation constraint information are then input into a preset prompt word template and input into a pre-trained language model for question generation processing to obtain an initial question set. The initial question set is formatted and its content is verified for consistency. The questions are then assembled according to a pre-set test paper structure template to generate the final question set.

5. The adaptive learning content processing method based on multi-agent collaboration according to claim 4, characterized in that, The process involves constructing question generation constraints based on the difficulty level, question type requirements, and examination scope in the learning objective parameters. The target knowledge point set and the question generation constraints are then input into a preset prompt word template and fed into a pre-trained language model for question generation, resulting in an initial question set including: Read the parsed content of the user-uploaded file from the object storage service or cache, organize the parsed content into a sequence of text blocks and write it to the content text block field of the state object; The system performs a reasoning call on the input material, matches the processing pattern based on the file extension provided by the user, and outputs the classification decision results. The classification decision results include pure prompt mode, real question imitation mode, syllabus alignment mode, and knowledge point alignment mode. In the syllabus alignment mode and the real exam imitation mode, the real exam text and the syllabus text are processed by word segmentation and semantic analysis to identify the candidate knowledge point set, and the candidate knowledge points are matched with the preset knowledge point dictionary or knowledge structure to determine the standardized knowledge point name; Based on the actual exam questions, we statistically analyzed the frequency of each knowledge point in the questions, the distribution of question types, and the percentage of marks. We then calculated the importance score for each knowledge point to obtain a weight value that represents the importance of the knowledge point. In the knowledge point alignment mode, the learning material text is divided into blocks according to a preset length to obtain multiple text blocks, and a block identifier is assigned to each text block; a prompt word template containing knowledge point definition, output format constraints and example description is constructed for each text block, and the text block and the prompt word template of example description are combined to form a structured input, which is input into a pre-trained large language model for semantic parsing and knowledge point extraction processing to obtain the knowledge point list of the corresponding text block; Based on the different types of user input materials, a corresponding knowledge point extraction strategy is adopted to generate a target knowledge point set. The knowledge point lists output by all text blocks are merged. The semantic similarity calculation method is used to remove duplicates or semantically similar knowledge points and merge them to obtain the target knowledge point set. Based on the target set of knowledge points and the target number of questions, a question-generating execution plan is generated, which allocates the number of questions, question type requirements and difficulty distribution for each knowledge point, and assigns the knowledge points to multiple question-generating sub-tasks; Based on the question generation execution plan, multiple processing units are concurrently invoked to perform question generation processing. Each processing unit receives the corresponding task object, constructs question generation prompts based on the task object, and calls the pre-trained large language model to generate the corresponding batch of questions at once. The generated results are output in a structured data format, including the question stem, options, answer, explanation, related knowledge point, difficulty, and score fields; thus obtaining the initial question set.

6. The adaptive learning content processing method based on multi-agent collaboration according to claim 4, characterized in that, The process of standardizing the format and verifying the consistency of the content of the initial question set, and assembling the questions according to the preset test paper structure template to generate the final question set includes: Each question is solved independently using a pre-trained large language model instance with independent configuration, and the solution results are written to the solution verification result field of the state object. The solution results are compared with the generated answers to calculate the accuracy of the self-solved questions. A quality evaluation is generated for each question and written into the scoring results and scoring summary fields of the status object. Based on the scoring results and scoring summary fields, a comprehensive quality review is conducted on the generated results. The checks include whether the number of questions is consistent with the corresponding real exam questions, whether the accuracy of self-solved questions meets the standard, and whether the distribution of question types is reasonable. In the knowledge point alignment mode, the knowledge point coverage rate is checked in addition. When the coverage rate is lower than the preset threshold and the current review retry counter is zero, the regeneration flag is set to true and the retry counter is incremented. The pre-trained large language model is called again to process the questions. If the standard is still not met after one retry, a warning message is added and the result is released. The question set is structured and parsed, and each question is broken down into the question stem, options, answer, explanation, related knowledge point, difficulty and score fields. The format of each field is standardized according to the preset output template to generate a standardized question data structure. In the syllabus alignment mode, a pre-constructed complete set of syllabus knowledge points is obtained. Based on the knowledge point identifiers associated with each question, the complete set of syllabus knowledge points is matched to obtain the set of covered knowledge points, and the set of uncovered knowledge points and the coverage ratio are calculated. The coverage ratio, the list of covered knowledge points, and the list of uncovered knowledge points are encapsulated in a structured manner, and written together with the standardized question data structure into the final output field of the state object to generate the final question set.

7. The adaptive learning content processing method based on multi-agent collaboration according to claim 1, characterized in that, The process of encapsulating structured content units, question sets, grading results, diagnostic information, and learning plans into recording units, and storing each recording unit, includes: Based on the session identifier, the corresponding session context data structure is obtained to generate a version identifier for the current round of learning. The structured content units, question sets, grading results, diagnostic information, and learning plans are encapsulated in a structured manner to form record units containing data type identifiers, generation timestamps, and source identifiers; Using the session identifier as the primary key and the version identifier as the subkey, construct a related index for multiple data types in the session context data structure; Each record unit is written to its corresponding index node, and the reference relationship and time sequence relationship between each record unit are established to characterize the evolution process of the processing results at different stages under the same session; According to the preset data storage structure, the session identifier, version identifier and corresponding record unit are stored. The data storage structure includes key-value pair or document structure storage to support fast retrieval based on session identifier, historical backtracking based on version identifier and data consistency maintenance in multi-round learning process.

8. An adaptive learning content processing system based on multi-agent collaboration, characterized in that, include: The input and session management module is used to obtain users' learning materials and learning requests, and to establish session identifiers associated with users' learning requests; It also performs intent recognition on user learning requests and outputs the intent recognition results; The content analysis and profile building module is used to obtain publicly available information about the target exam and the target position, and to build exam profiles and position competency profiles respectively. The learning materials, publicly available exam information, and publicly available job postings are analyzed to generate structured content units; The distribution and question generation module is used to distribute structured content units, exam profiles, and job competency profiles to the corresponding processing flows based on the intent recognition results. Based on structured content units, exam profiles, and job competency profiles, a set of questions is generated; The assessment, diagnosis, and plan generation module is used to output a set of questions to users and receive their answers; output grading results and extract diagnostic information based on the answers; and generate learning plans based on diagnostic information, exam profiles, and job competency profiles. The status management and persistence module is used to encapsulate structured content units, question sets, grading results, diagnostic information, and learning plans into record units, and store each record unit. The adaptive closed-loop control module is used to adjust the question generation strategy, exam preparation focus, and job skills training focus based on the stored results in the next round of learning, forming an adaptive learning closed loop.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images