A transaction-oriented generation and agent collaborative scheduling method for multi-modal teaching resources

Abstract

The application discloses a transactional generation and intelligent agent collaborative scheduling method for multi-modal teaching resources, and relates to the technical field of teaching resource management.The method comprises the following steps: constructing a transactional generation process, establishing an intelligent agent collaborative scheduling framework, and executing the transactional process and the intelligent agent collaborative scheduling.The application realizes the full-process standardization, automation and intelligentization of multi-modal teaching resource generation by constructing an atomized transactional generation process and a multi-agent collaborative scheduling framework, integrates fragmented generation tasks into a coherent process with transactional guarantee, ensures the integrity and consistency of resources, significantly improves the efficiency and system robustness of complex resource generation through multi-agent division of labor and cooperation and dynamic scheduling based on real-time feedback, and makes the generated teaching resources have precise education target alignment and personalized learning adaptability through deep integration of an education knowledge graph and a multi-dimensional evaluation mechanism.

Description

Technical Field

[0001] This invention relates to the field of teaching resource management technology, and in particular to a transactional generation and agent-based collaborative scheduling method for multimodal teaching resources. Background Technology

[0002] Teaching resources refer to all material and non-material elements that serve the teaching objectives during the teaching process, including textbooks, courseware, equipment, software, teachers, time, space, and even generative knowledge. The management and scheduling of teaching resources refers to the dynamic process of systematically identifying, organizing, allocating, integrating, and optimizing the allocation of these resources. Its core lies in accurately delivering the right resources to the teaching process at the right time and in the right way, based on teaching needs and through scientific planning and coordination mechanisms, to support the effective conduct of teaching activities and ultimately maximize teaching effectiveness and optimize resource utilization.

[0003] In the modern educational environment, the efficient scheduling and rational allocation of teaching resources (including teachers, classrooms, etc.) is a key task for school and educational institution management. With the continuous increase in the number of courses and teaching needs, how to arrange courses under limited resources and avoid problems such as teacher conflicts, resource waste and insufficient teaching space has become an important challenge for teaching management.

[0004] Based on existing technological findings, traditional teaching resource management and scheduling technologies often rely on manual intervention or simple tool-based approaches, resulting in fragmented generation processes and a lack of unified transaction management and fault tolerance mechanisms. This leads to a disconnect between the generation and integration of resources across different modalities, inefficient resource scheduling, and difficulty in ensuring the integration and unification of teaching resources across multiple dimensions such as cognitive science, technical quality, and user experience. Furthermore, it fails to achieve dynamic optimization based on real-time feedback and large-scale personalized resource adaptation. Therefore, this invention proposes a transaction-based generation and intelligent agent-based collaborative scheduling method for multimodal teaching resources to address the problems existing in the prior art. Summary of the Invention

[0005] To address the aforementioned problems, the present invention aims to propose a transactional generation and intelligent agent collaborative scheduling method for multimodal teaching resources, thereby solving the problems of long resource production cycles, inconsistent quality, and lack of intelligent collaboration and personalized support in traditional teaching resource management and scheduling technologies.

[0006] To achieve the objectives of this invention, the invention is implemented through the following technical solution: a transactional generation and agent-based collaborative scheduling method for multimodal teaching resources, comprising the following steps: Step 1: Abstract and encapsulate the process of generating multimodal teaching resources into a transactional task flow that includes five stages: requirements analysis, content design, multimodal material generation, integration optimization, and quality assessment. Predefine the input and output formats, execution constraints, and exception handling rules for each stage to form an atomic, standardized workflow template. Step 2: Deploy a multi-agent system that includes a demand analysis agent, a content generation agent, a cross-modal alignment agent, and a resource scheduling agent. Each agent synchronizes information and is aware of its state through a shared semantic context pool. The semantic context pool stores and updates task progress data, resource metadata, and educational rule knowledge in real time. Step 3: The resource scheduling agent dynamically decomposes the transactional task flow based on the logical dependencies between each stage and the capability tags of each agent, and initially allocates the task flow to the corresponding agents. During the task allocation process, the resource scheduling agent dynamically adjusts the task allocation and resource allocation strategies based on real-time monitoring of the stage output quality score and agent load status to ensure the continuity and consistency of the entire generation transaction, and finally outputs a composite multimodal teaching resource package that meets the educational objectives and quality standards.

[0007] Further improvements are made in the following steps: In step one, the requirements analysis stage is based on a pre-constructed subject knowledge graph. It analyzes the input teaching objectives to extract the knowledge point sequence, cognitive level and curriculum standard requirements. At the same time, it combines student profiles to obtain their cognitive level, learning style and modal preferences, and comprehensively generates a structured requirements vector that includes at least a knowledge point tag set, cognitive difficulty coefficient, modal priority list and interaction depth parameters. This vector serves as the core input for constraining all subsequent generation stages.

[0008] Further improvements are made in that: the subject knowledge graph is constructed by automatically extracting teaching concept entities, semantic relationships between entities, and educational attributes of entities from structured subject textbooks, curriculum standards, and academic literature using natural language processing technology; and the requirements analysis stage accurately determines the scope, hierarchical path, and cognitive attributes of the knowledge points involved by matching teaching objectives with the subject knowledge graph through subgraph matching.

[0009] A further improvement is made in the following: In step one, the multimodal material generation stage is completed by calling a modular AI generation component library. The AI ​​generation component library includes a text generation component based on a large language model, an image generation component based on a diffusion model, an audio generation component based on neural speech synthesis technology, and a video generation component based on an animation engine. Each generation component is parametrically called according to the structured demand vector and content design results output in the previous stage to generate corresponding text lesson plans and exercises, concept diagrams and schematic diagrams, audio explanations with emotional intonation, and dynamic demonstration video clips in parallel or serially.

[0010] A further improvement is that, in step one, the fusion optimization stage is performed by a cross-modal alignment agent, which includes extracting semantic feature vectors of different modal materials to calculate and optimize the semantic correlation between them, matching the timestamps of video streams and audio streams through a time series alignment algorithm to achieve audio-visual synchronization, and dynamically adjusting the presentation weight and priority of different modal content in the final resource package according to the user modal preferences obtained in the demand parsing stage, so as to enhance the overall semantic consistency and personalized adaptability of the resources.

[0011] Further improvements are made in the following aspects: In step one, the quality assessment stage adopts multi-dimensional assessment indicators. The educational effectiveness dimension is measured by analyzing the cognitive level covered by knowledge points and the data related to the assessment of cognitive load. The technical quality dimension is quantified by calculating the peak signal-to-noise ratio of images, video structural similarity, and average opinion score of speech. A satisfaction prediction model for the user experience dimension is constructed by analyzing simulated or real user interaction behavior data. When the comprehensive assessment score is lower than a preset threshold, a quality anomaly event is triggered to drive the system to execute the rollback and regeneration process.

[0012] Further improvements are made in the following aspects: In step one, the standardized workflow template supports dynamic expansion and customization. Through the provided visual configuration interface, users can add, delete, or modify task nodes in the generation process by dragging and dropping, and customize the dependency relationships and execution constraint rules between nodes. In addition, all version changes of workflow templates are managed through a version control system to support template reuse, rollback, and difference comparison.

[0013] A further improvement is made in step two, in which the resource scheduling agent adopts a cooperative strategy based on reinforcement learning for dynamic scheduling and conflict resolution. By constructing a reward function with the optimization objectives of resource generation quality and overall process efficiency, and training it using multi-agent reinforcement learning, the agent can autonomously decide on task allocation schemes based on the real-time system state and resolve potential resource competition or generation target conflicts between different agents.

[0014] A further improvement is made in the following: In step two, the multi-agent system uses an event-driven asynchronous communication mechanism for collaboration. Each agent is decoupled and connected through a message middleware to transmit task status changes, data readiness notifications, resource conflict alarms, and priority adjustment instructions by publishing and subscribing to events. The resource scheduling agent, as the core event listener and responder, triggers corresponding task migration, load balancing, or conflict coordination operations in real time according to the type of event received.

[0015] A further improvement is made in step three, where the dynamic adjustment of the task allocation strategy includes: When the monitoring detects that the phased output quality of an agent instance is consistently substandard or its load exceeds a critical value, the resource scheduling agent marks the current task on that instance and migrates it to a standby or idle agent instance of the same type. When a process blockage is detected due to an incomplete preceding task, a waiting mechanism is automatically triggered and a log is recorded. Execution will automatically resume once the preceding task is completed.

[0016] Further improvements include: in the scenario of exercise-based teaching resources, steps such as constructing knowledge point-level mastery and uncertainty, calculating conservative global cognitive health, and using steady-state gaps as constraints for exercise delivery are also included. in, For knowledge points The degree of conservative control, calculated by deducting the risk of overestimation due to estimation uncertainty from the original degree of control, is a core reference indicator for algorithmic decision-making. For local control, This is an uncertainty reduction factor used to quantify the negative impact of uncertainty on the confidence level forecast; the default value is 1.0. For knowledge points The state uncertainty reduction factor is used to measure the confidence level of the system in the current knowledge estimate; a larger value indicates a higher estimation bias. To safeguard the overall health of student learning, the overall learning level is calculated based on a weighted average of the weights of each knowledge point, thus avoiding overall judgment errors caused by prediction biases of a single knowledge point. For knowledge points The normalized weights reflect the importance of each knowledge point in the overall learning situation. The weights of all knowledge points are summed and fixed at 1. To preset the steady-state level of the teaching objectives, the system expects to achieve a baseline value for overall student learning, with a default value of 0.65. The learning progress steady-state gap is used to divide the algorithm's operating mode: a value greater than 0 activates the exercise supplementation mode, and a value less than or equal to 0 activates the learning progress maintenance mode.

[0017] Further improvement lies in: using a question-by-question value function for exercise recommendations. and time marginal decay factor It performs global optimization and can be switched and used in conjunction with rule-based scheduling strategies and reinforcement learning scheduling strategies.

[0018] Further improvements are made by pausing the current transaction after the resource scheduling agent issues a standardized query to the learner, writing the learner's selection result back to the semantic context pool, and resuming the execution of subsequent stages on the same transaction instance.

[0019] Further improvements include: storing subject planning data, question bank data, conversation task data, and chat display data in layers, with subject planning and question bank being the true source of the system, while chat display is not considered the true source of planning.

[0020] The beneficial effects of this invention are as follows: By constructing an atomic transaction generation process and a multi-agent collaborative scheduling framework, this invention achieves standardization, automation, and intelligence of the entire process of generating multimodal teaching resources. It can integrate fragmented generation tasks into a coherent process with transaction guarantees, ensuring the integrity and consistency of resources. Furthermore, through multi-agent division of labor and collaboration and dynamic scheduling based on real-time feedback, it significantly improves the efficiency and robustness of complex resource generation. At the same time, by deeply integrating educational knowledge graphs and multi-dimensional evaluation mechanisms, the generated teaching resources have both accurate alignment with educational goals and personalized learning adaptability, thereby effectively solving the problems of long production cycles, inconsistent quality, and lack of intelligent collaboration and personalized support in traditional resource production. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the transactional generation and agent-based collaborative scheduling method for multimodal teaching resources according to the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] It should be noted that the technical means not described in detail in the following embodiments are all conventional means in the art, are not the key points of the invention, and will not be elaborated upon.

[0024] Example 1 See Figure 1 This embodiment provides a transactional generation and agent-based collaborative scheduling method for multimodal teaching resources. The core of this method lies in making the resource generation process transactional (i.e., ensuring the atomicity, consistency, isolation, and durability of the process) and entrusting it to a multi-agent system for collaborative scheduling and execution. Specifically, it includes the following steps: Step 1: Construct a transactional generation process The process of generating multimodal teaching resources is abstracted and encapsulated into a transactional task flow comprising five stages: requirements analysis, content design, multimodal material generation, integration and optimization, and quality assessment. a. Requirements Analysis: This stage transforms vague teaching requirements into structured data. The core of this stage is converting unstructured input into precise, computer-processable instructions. The teaching objective input is natural language (e.g., "explain Newton's First Law and its applications"). The system uses natural language understanding technology, combined with a pre-constructed subject knowledge graph, to extract specific knowledge point sequences (e.g., ["Newton's First Law", "inertia", "relationship between force and motion"]), cognitive levels (e.g., "understanding" level), and mapped curriculum standard requirements (e.g., "corresponding to Article 3.2.1 of the 'Compulsory Education Physics Curriculum Standards'"). Simultaneously, the system queries student profiles to obtain the student's or student group's cognitive level (e.g., previous test scores), learning style (e.g., "visual learner"), and modal preferences (e.g., "prefers video > image > text"). Finally, this information is integrated to generate a structured requirements vector. This vector is a multi-dimensional data structure, for example: {“Knowledge Point Tag Set”: [“Buoyancy”, “Archimedes’ Principle”], “Cognitive Difficulty Coefficient”: 3, “Modal Priority List”: [“Video”, “Interactive Animation”, “Text”], “Interaction Depth Parameter”: {“Number of Exercises”: 5, “Variation Difficulty”: “Medium”}}.

[0025] This vector will serve as the "design blueprint" and constraint throughout all subsequent stages; The data sources for the aforementioned subject knowledge graph include structured subject textbooks (electronic versions), curriculum standard documents, and related academic papers. The construction process utilizes natural language processing technologies, such as pre-trained models like BERT and ERNIE for Named Entity Recognition (NER) to extract teaching concept entities (e.g., "electric field strength," "photosynthesis"), and relation extraction (RE) models to identify semantic relationships between entities (e.g., "is a type of A," "derives," "applies to"), and extract the educational attributes of entities (e.g., "grade level": "high school," "cognitive category": "concept"). During requirements analysis, the system performs subgraph matching between the teaching objective text and the knowledge graph, finding the most connected subgraph that best matches the teaching objective description. This accurately determines the scope of the knowledge points involved, their hierarchical or logical paths (e.g., predecessor-successor relationships), and the cognitive attributes of each node on the path, laying the foundation for subsequent personalized content generation. b. Content Design: Based on structured requirements, plan the content outline and presentation logic of the resources; c. Multimodal Material Generation: This stage utilizes AI tools to generate raw materials such as text, images, audio, and video. It is implemented through a pluggable, modular AI generation component library, where each component is an independent service. The text generation component receives the script outline output from the content design stage, calls large language models such as GPT-4 and Wenxin Yiyan, and generates narrative text with rich teaching logic and diverse exercises according to lesson plan templates such as "Import – Explanation – Example – Summary – Exercises." The image generation component receives text prompts containing scene descriptions (derived from the content design results), calls diffusion models such as StableDiffusion and DALL-E, and generates two-dimensional / images that accurately reflect scientific principles. The system includes a 3D schematic diagram; an audio generation component that uses neural speech synthesis (TTS) engines such as VITS and FastSpeech2 to synthesize generated text into narrated speech with specific timbre, speech rate, and emotional tone (e.g., raising the pitch when emphasizing key points); a video generation component that integrates 3D models and motion capture data, and uses animation engines such as Unity and Unreal to render and generate dynamic demonstration clips such as experimental operations; all components provide services through a unified RESTful API or gRPC interface, supporting parameterized calls based on relevant parameters (such as image resolution, speech rate, and video duration) in a structured requirement vector, thereby achieving efficient generation in parallel or in a dependent order; d. Fusion Optimization: Ensuring semantic and temporal alignment and coordination of materials from different modalities. This stage is handled by a cross-modal alignment agent. First, it uses cross-modal feature extractors such as CLIP and ImageBind to map text descriptions, images, and audio (which can be converted to text or have audio features directly extracted) to the same semantic vector space, and calculates cosine similarity as a quantitative indicator of semantic relevance. For material pairs with low relevance, it proposes optimization suggestions or regenerates them. Second, for audio-video synchronization, it employs time-series alignment algorithms such as Dynamic Time Warping (DTW) to automatically align the keyword time points of the narration with the timestamps of key action frames in the video demonstration. Finally, it introduces a dynamic weight adjustment mechanism. For example, for users with a "visual" preference, during the final synthesis, the presentation duration or interface weight of images and videos is increased by 20%, while text descriptions are appropriately simplified, thereby achieving deep personalized adaptation. e. Quality Assessment: The generated resource package undergoes multi-dimensional automated evaluation. This stage employs an automated, multi-dimensional evaluation system, including: Educational effectiveness dimension: By analyzing whether the generated content covers the Bloom's Taxonomy cognitive levels (memory, comprehension, application, etc.) comprehensively, and by combining simulated eye-tracking data (or actual collected data) to calculate learners' attentional distraction, the rationality of the content design is evaluated based on cognitive load theory. Technical quality dimensions: Peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) are used to evaluate the sharpness and fidelity of the generated images / videos; mean opinion score (MOS) algorithm or objective speech quality assessment metrics are used to evaluate the naturalness of the audio. User experience dimension: During the resource trial phase, collect user interaction behavior data (such as page dwell time, click heatmap, error rate of practice questions and number of repeated attempts) to build a satisfaction prediction model; The system assigns weights to indicators in each dimension and calculates a comprehensive evaluation score. When the score is lower than a preset threshold (e.g., 80 points), the system determines it as a quality anomaly event, automatically triggers the process to roll back to the corresponding stage and start regeneration, or notifies the administrator to intervene. Each stage predefines its input and output formats, execution constraints, and exception handling rules to form an atomic, standardized workflow template. Each stage is designed as an "atomic" sub-transaction, treating the operations within each stage as an indivisible whole: either all operations are executed successfully and produce predefined standard outputs, or the system rolls back to the state before the stage began due to failure, without generating dirty data in partial or intermediate states that would affect system consistency. In this embodiment, the resulting workflow template is a data structure or configuration file that clearly defines: stage goals (e.g., "generate narration audio"), input formats (e.g., "text script, JSON format"), output formats (e.g., "48kHz, 16-bit mono WAV audio file"), execution constraints (e.g., "must be triggered after the content design stage is completed", "generation time not exceeding 5 minutes"), and exception handling rules (e.g., "after generation timeout, retry 2 times or switch to a backup engine"). These templates enable the process to be automatically parsed and driven by the system. The standardized workflow template has high flexibility. The system provides a visual configuration interface (e.g., a web-based drag-and-drop canvas) that allows teachers or instructional design experts to add (e.g., add a "virtual reality (VR) scene generation" node), delete, or modify task nodes in the process without coding, and graphically set the dependencies between nodes (connections indicate the order) and execution constraint rules (e.g., conditional judgments). All modifications to the template are managed through the integration of version control systems such as Git, saving a complete template iteration history, supporting rollback of any version, comparison of differences between different versions, and cross-project reuse of excellent templates. This step standardizes the originally loose, manual or semi-automatic multimodal resource creation process into a controllable and traceable standardized process, in which the five stages are logically strictly progressive and have clear functional boundaries. Step 2: Establish an agent-based collaborative scheduling framework Deploy a multi-agent system comprising a demand analysis agent, a content generation agent, a cross-modal alignment agent, and a resource scheduling agent, including: Demand Analysis Agent: Specializes in analyzing teaching objectives and student profiles to output structured requirements; Content generation agent: responsible for content creation and planning, transforming requirements into specific content scripts and design diagrams; Cross-modal alignment agent: specifically designed to handle semantic associations and spatiotemporal synchronization issues between materials of different modalities; Resource scheduling agent: As the system's "scheduling center," it is responsible for task allocation, monitoring, and dynamic adjustment. Its core decision-making capability stems from a cooperative strategy based on reinforcement learning. The system constructs a multi-agent reinforcement learning (MARL) environment, where each generating and aligning agent is part of the "environment," and the resource scheduling agent acts as the learning subject. Its reward function is designed as follows: Reward = Resource quality score × 0.6 + Generation efficiency score × 0.4 Among them, the generation efficiency can take into account the overall task completion time. Through simulation training, the scheduling agent learns to make the optimal task allocation scheme in complex states (such as high load of some agents, conflicting goals of generation tasks) and can resolve conflicts. For example, when the requirements of "text conciseness" and "video detail" conflict, it learns a balancing strategy or prioritizes the main constraints in the demand vector. Each agent synchronizes information and perceives its state through a shared semantic context pool. This semantic context pool stores and updates task progress data, resource metadata, and educational rule knowledge in real time. It is a shared data area located in memory or distributed storage. It is not a simple message queue, but a structured context store that records and updates in real time: task status (e.g., "Material generation - in progress, progress 65%), resource metadata (e.g., "Image ID generated: IMG_001, resolution: 1920×1080, associated knowledge point: buoyancy"), user feedback (e.g., "Students have a low interaction rate with video clip A"), and educational rules loaded from the domain knowledge base (e.g., "The explanation of junior high school physics concepts should avoid using calculus expressions"). All agents perceive the global state and coordinate their actions by reading and writing to this context pool, achieving decoupled collaboration. To achieve efficient and decoupled collaboration, this embodiment of the multi-agent system adopts an event-driven asynchronous communication mechanism. The agents do not directly call each other, but instead connect to message middleware (message buses) such as MQTT, RedisPub / Sub, or Apache Kafka. Agents publish or subscribe to specific types of events on the bus. For example, after completing the content generation agent, it publishes a "data ready event" with a quality score. When two agents simultaneously request the same scarce GPU resource, the system publishes a "resource conflict event." The resource scheduling agent subscribes to all key events, acting as the core event listener and responder. Once it receives a corresponding event, it immediately triggers predefined response logic (e.g., if a conflict event is received, a load balancing algorithm is invoked to reallocate resources), greatly reducing system coupling. This step constructs a collaborative system composed of multiple specialized intelligent agents to replace the traditional centralized controller; Step 3: Execute transactional processes and coordinate the scheduling of intelligent agents. Based on the logical dependencies between the stages of the transactional task flow and the capability tags of each agent, the resource scheduling agent dynamically decomposes the transactional task flow and initially assigns it to the corresponding agents. The specific steps are as follows: When a new resource generation transaction is started, the resource scheduling agent first reads the definition of the "transactional generation process". Based on the logical dependencies between the five stages (such as "content design" must be done before "material generation" can be carried out) and its own maintained agent capability tag library (such as recording "agent A is good at generating chemical experiment videos"), it dynamically decomposes the entire process into a series of sub-tasks and matches and assigns each sub-task to the most suitable agent instance. During task allocation and execution, the resource scheduling agent dynamically adjusts task allocation and resource allocation strategies based on real-time monitoring of stage output quality scores and agent load status. Task allocation is not static; the resource scheduling agent continuously monitors two key indicators: stage output quality scores, generated by the quality assessment stage or self-assessment by each agent; and agent load status, including CPU / GPU utilization, memory usage, and task queue length. Based on this real-time feedback, the scheduling agent dynamically adjusts strategies, such as migrating tasks from agents with excessive load to idle nodes or reassigning tasks with consistently low quality scores to backup agents, ultimately outputting a composite multimodal teaching resource package that meets educational objectives and quality standards. In this embodiment, dynamic strategy adjustment is specifically reflected in fault recovery and process control, including: Quality / Load Monitoring and Migration: The system sets quality score thresholds and load thresholds (such as CPU utilization > 85%). Once triggered, the resource scheduling agent marks the tasks on the problematic instance as "to be optimized" and matches a backup agent instance with similar capabilities (which may be a cold backup or an elastically scaled instance in the cloud) from its capability tag library, and migrates the task and context over as a whole. Process blocking handling: The system maintains a task dependency graph. When it detects that the current task cannot be executed because the previous stage has not produced output, it automatically triggers a waiting mechanism, suspends the current task, and records detailed blocking logs (including blocking time, the ID of the previous task being waited for, etc.). Once the previous task completes, it automatically wakes up and resumes the execution of the current task, ensuring the automated progress of the process. Through the above mechanism, the continuity (the process is not interrupted by accidents) and consistency (the data status of each stage is consistent and the final resource package is complete) of the entire generation "transaction" are ensured. When all stages are successfully completed, the system packages and outputs a composite multimodal teaching resource package. This package is usually a compressed package containing associated text, image, audio, and video files, as well as a manifest file (such as XML or JSON format) describing their structure and metadata.

[0026] Example 2 This embodiment implements cognitive state modeling and collaborative scheduling of exercise-type teaching resources. Building upon the aforementioned transactional generation and agent-based collaborative scheduling of multimodal teaching resources, this embodiment further establishes a cognitive state representation and exercise selection mechanism that connects with the requirements analysis phase, specifically for the scenarios of pushing, practicing, and providing feedback on "exercise-type teaching resources." The following content serves as supplementary explanation of the specific implementation method, and its style remains consistent with the above embodiment.

[0027] a. The connection between the cognitive state layer and the structured demand vector Based on the student profiles obtained in the needs analysis phase, the mastery level of each knowledge point (KP) is represented as local mastery. With uncertainty Further define the degree of conservative control: in, For knowledge points The degree of conservative control, calculated by deducting the risk of overestimation due to estimation uncertainty from the original degree of control, is a core reference indicator for algorithmic decision-making. The function that maximizes the mastery level is used to constrain the lower limit of mastery, ensuring that the mastery level of knowledge points does not become negative after forgetting decay. For local control, This is an uncertainty reduction factor used to quantify the negative impact of uncertainty on the confidence level forecast; the default value is 1.0. For knowledge points The state uncertainty reduction factor is used to measure the confidence level of the system in the current knowledge estimate. The larger the value, the higher the estimation bias.

[0028] and overall cognitive health (weight) (Originated by normalizing the number of questions for each knowledge point in the question bank) in, To safeguard the overall health of student learning, the overall learning level is calculated based on a weighted average of the weights of each knowledge point, thus avoiding overall judgment errors caused by prediction biases of a single knowledge point. For knowledge points The normalized weight reflects the importance of the knowledge point in the overall learning situation, and the weights of all knowledge points are summed up and fixed at 1.

[0029] Given the steady-state level of the teaching objectives Define the steady-state gap: in, To preset the steady-state level of the teaching objectives, the system expects to achieve a baseline value for overall student learning, with a default value of 0.65. The learning progress steady-state gap is used to divide the algorithm's operating mode: a value greater than 0 activates the exercise supplementation mode, and a value less than or equal to 0 activates the learning progress maintenance mode.

[0030] In addition to being written into the semantic context pool, the aforementioned quantities, along with the knowledge point tag set, cognitive difficulty coefficient, and learning focus information such as subject / topic / chapter, constitute the extended components of the structured demand vector, which serve as constraint inputs for the resource scheduling agent when pushing exercises and planning paths. The learner's answers to each question are used to update the corresponding knowledge points. and This serves as a closed-loop feedback basis for the quality assessment stage and the educational effectiveness dimension.

[0031] b. Optimization of the value of each exercise in the recommended practice questions and marginal decay. In the sub-process of generating and pushing multimodal materials for exercises, the resource scheduling agent processes each candidate question in the question bank. Calculate the value score for each question and perform a single-step global optimization. The value function is defined as: in, To answer the question about the gain coefficient, we reuse the previously used parameter, with a default value of 0.18. The conditional expected probability of answering correctly: Based on the learner's current level of mastery and the difficulty of the question, predict the probability that the learner will answer the question correctly. For the title Its inherent difficulty parameters are pre-labeled in the question bank. To mitigate the marginal decay of practice questions over time, we avoid repeatedly pushing the same question within a short period and optimize the diversity of the push notifications. For knowledge points Due to the uncertainty in the assessment of student learning, the unified parameters mentioned earlier are reused. This represents the expected probability of answering correctly, estimated based on the current level of understanding and the difficulty of the question. This represents the cost weight for a single push notification, reusing the globally unified parameter mentioned earlier, with a default value of 1.0. The question-by-question time marginal decay factor is defined as: In the formula, This represents the current practice time sequence in the system, indicating the global time cycle index for this round of exercises. For the title The practice session number that was last pushed by the system. It is an exponential function with the natural constant e as the base. This is the time decay step size parameter, which controls how quickly the repeated question push is suppressed. The default value is 4. This is used to suppress repeated pushes of the same exercise within a short period without relying on a hard deduplication list, and to restore pushability after sufficient practice intervals or across days. When all candidate questions within the same study day... When the number of questions approaches zero, the question that has not been practiced for the longest time is selected as a backup for rotation to avoid scheduling stagnation.

[0032] In this embodiment, the question-by-question value selection strategy can be configured in parallel with the reinforcement learning-based collaborative scheduling strategy described in claim 8: the former focuses on maximizing the expected learning gains at the question level, while the latter focuses on multi-agent task allocation and process efficiency; the two share the same cognitive state layer and can switch between rule-based scheduling and value function scheduling according to the application scenario.

[0033] c. Human-machine collaborative consultation and transaction continuation During the execution of multi-step transactions by the resource scheduling agent, when it is necessary to solicit the learner's path selection, basic level, or willingness to help, the scheduling agent issues a standardized solicitation request, which is presented on the interactive interface with no less than two selectable options. This solicitation constitutes an atomic sub-step in the current transactional task flow, and its input and output formats and exception handling rules are incorporated into the aforementioned standardized workflow template for unified management.

[0034] After a request is issued, the current multi-step transaction is paused and the completed stage states are saved. Once the learner selects an option, the system writes that selection back to the semantic context pool as the formal output of that sub-step and resumes the scheduling of subsequent stages on the same transaction instance. The resource scheduling agent autonomously decides whether to continue publishing stage task plans, updating learning routes, pushing exercises, or calling the Q&A generation component. This mechanism echoes the "process blocking handling" in step three: the blocking reason is waiting for learner confirmation, and the resumption condition is that the request sub-step obtains valid output.

[0035] d. Layering of persistent data and transaction consistency To ensure the atomicity, consistency, and durability of the transactional generation process, this embodiment divides the runtime data into three categories according to its lifecycle: First, subject routes, chapter plans, question bank content, and learner cognitive states, etc., are system-level persistent data that are retained even after the application session ends, constituting the true source of planning and push notifications; Second, the phased task list in a single multi-step execution process resides only at the session task layer and becomes invalid when the current transaction ends or is explicitly cleared; Third, dialogue displays and verbal interaction records are retained at the chat display layer for review and communication, but do not serve as the true source of subject planning and question recommendation decisions. This layering semantically isolates "plans / question bank changes submitted to the system layer" from "conversation text used only for display," avoiding inconsistencies between the interface state and persistent planning.

[0036] e. Student Activity Perception and Active Intervention This embodiment performs lightweight perception of the learner's operational trajectory during exercise practice, including but not limited to: starting to answer, changing options, submitting results, receiving push notifications, etc., and counts the dwell time for each question. When the dwell time exceeds a preset threshold and no answer has been submitted, the resource scheduling agent can proactively inquire whether explanation or hints are needed through a standardized inquiry mechanism. The perception summary is written into the semantic context pool for the scheduling agent to reference in subsequent decisions, thereby reducing repeated inquiries about the same facts, and serving as a supplementary channel for the learner side in step three, "real-time monitoring based on stage output quality scores and agent load status".

[0037] f. Q&A-type multimodal components and their fusion optimization integration For verbal Q&A requests such as "explain this question" or "give a hint," the content-generating agent or dedicated Q&A component will provide the answer based on the current exercise text, options, knowledge points, and the basis for the request. , The system generates explanatory or prompt text based on cognitive state data and presents it in the chat display layer. This type of output is a text modal supplement in the fusion and optimization phase and does not replace the persistent storage of the subject-specific path at the system layer. If the Q&A request originates from the aforementioned standardized inquiry sub-step, it will be prioritized for continuation within the same transaction instance. If it is initiated by the learner, the Q&A generation component can be directly invoked to complete the response.

[0038] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A transactional generation and agent-based collaborative scheduling method for multimodal teaching resources, characterized in that, Includes the following steps: Step 1: The process of generating multimodal teaching resources is abstracted into five stages: requirements analysis, content design, multimodal material generation, integration and optimization, and quality assessment. An atomized standardized workflow template is predefined for each stage. Step 2: Deploy a multi-agent system that includes a demand analysis agent, a content generation agent, a cross-modal alignment agent, and a resource scheduling agent. The agents collaborate through a shared semantic context pool. Step 3: The resource scheduling agent allocates tasks based on stage logical dependencies and agent capability tags, and dynamically adjusts the allocation strategy based on stage output quality scores and agent load status, ultimately outputting a multimodal teaching resource package.

2. The transactional generation and agent-based collaborative scheduling method for multimodal teaching resources according to claim 1, characterized in that: In step one, the requirement analysis stage analyzes the teaching objectives based on the subject knowledge graph and combines them with student profiles to generate a structured requirement vector containing a set of knowledge point tags, a cognitive difficulty coefficient, a modal priority list, and interaction depth parameters. This vector serves as the core input for constraining all subsequent generation stages.

3. The transactional generation and agent-based collaborative scheduling method for multimodal teaching resources according to claim 2, characterized in that: The subject knowledge graph is automatically constructed from structured textbooks and literature using natural language processing technology. In the needs analysis stage, the scope of knowledge points and cognitive attributes are determined by matching teaching objectives with subgraphs of the knowledge graph.

4. The transactional generation and agent-based collaborative scheduling method for multimodal teaching resources according to claim 1, characterized in that: In step one, the multimodal material generation stage is completed by calling a modular AI generation component library. The component library includes a text component based on a large language model, an image component based on a diffusion model, an audio component based on neural speech synthesis, and a video component based on an animation engine. Each component is parameterized according to structured requirements to generate corresponding materials.

5. The transactional generation and agent-based collaborative scheduling method for multimodal teaching resources according to claim 1, characterized in that: In step one, the fusion optimization stage is performed by a cross-modal alignment agent, which includes extracting multimodal semantic features to optimize the correlation, achieving audio-visual synchronization through a time series alignment algorithm, and dynamically adjusting the presentation weight and priority of each modality content according to the user's modal preferences.

6. The transactional generation and agent-based collaborative scheduling method for multimodal teaching resources according to claim 1, characterized in that: In step one, the quality assessment stage adopts multi-dimensional assessment indicators, including educational effectiveness, technology quality, and user experience. When the comprehensive assessment score is lower than a preset threshold, a quality anomaly event is triggered, driving the process to roll back and regenerate.

7. The transactional generation and agent-based collaborative scheduling method for multimodal teaching resources according to claim 1, characterized in that: In step one, the standardized workflow template can be dynamically expanded and customized through a visual configuration interface, allowing users to modify task nodes, dependencies, and constraint rules by dragging and dropping, and manage the template through a version control system.

8. The transactional generation and agent-based collaborative scheduling method for multimodal teaching resources according to claim 1, characterized in that: In step two, the resource scheduling agent uses a cooperative strategy based on reinforcement learning for dynamic scheduling and conflict resolution, and its reward function optimizes resource generation quality and process efficiency.

9. The transactional generation and agent-based collaborative scheduling method for multimodal teaching resources according to claim 1, characterized in that: In step two, the multi-agent system uses an event-driven asynchronous communication mechanism for collaboration. Each agent transmits events through a message middleware in a publish and subscribe manner. The resource scheduling agent listens for and responds to the events to trigger task migration, load balancing, or conflict coordination operations.

10. The transactional generation and agent-based collaborative scheduling method for multimodal teaching resources according to claim 1, characterized in that: In step three, the dynamic task allocation strategy includes: when the output quality of the agent instance is consistently substandard or the load exceeds the critical value, its tasks are migrated to a standby or idle instance; when the process is detected to be blocked due to the incomplete completion of the preceding tasks, a waiting mechanism is automatically triggered until the preceding tasks are ready and execution resumes.

Images