Interdisciplinary PBL teaching plan intelligent generation system based on knowledge graph and multi-Agent cooperation

Abstract

The invention discloses an interdisciplinary PBL teaching plan intelligent generation system based on knowledge graph and multi-Agent cooperation, and the system comprises an interdisciplinary knowledge base construction module which constructs knowledge graphs of mathematics, physics and computer subjects through semantic vectorization and entity alignment; the retrieval enhancement module is used for dynamically supplementing knowledge blind areas based on an RAG mechanism; the teaching demand analysis module is used for analyzing teacher input and generating a teaching target matrix; the structured modeling module is used for decomposing tasks and designing a teaching framework according to a three-dimensional target system (cognition, ability and subject literacy); the teaching plan generation module is used for generating subject contents through multi-Agent collaboration; and the teaching plan optimization module verifies interdisciplinary consistency and outputs a standardized teaching plan. According to the method, the knowledge graph and multi-Agent dynamic weight distribution are creatively fused, deep integration of interdisciplinary knowledge is realized, teaching plan generation efficiency and scientificity are remarkably improved, and intelligent lesson preparation support is provided for teachers.

Description

Technical Field

[0001] This invention belongs to the field of intelligent education technology, and in particular relates to an intelligent generation system for interdisciplinary PBL lesson plans based on knowledge graphs and multi-agent collaboration. Background Technology

[0002] In modern education, interdisciplinary project-based learning (PBL) is gaining increasing attention as an innovative teaching model. It emphasizes fostering students' active learning and critical thinking skills by solving real-world problems. However, designing and implementing interdisciplinary PBL lesson plans presents a challenge for teachers, requiring them not only to master knowledge from multiple disciplines but also to effectively integrate this knowledge into a coherent teaching plan. Furthermore, existing artificial intelligence tools often struggle to achieve deep integration of interdisciplinary knowledge when assisting in lesson plan design, resulting in lesson plans that lack systematicity and scientific rigor.

[0003] Knowledge graphs, as a structured semantic knowledge base, can store and represent relationships between entities, providing a possible solution for the integration of interdisciplinary knowledge. However, how to effectively combine knowledge graphs with PBL lesson plan generation to improve lesson plan quality and teachers' lesson preparation efficiency remains an unsolved problem. Summary of the Invention

[0004] This invention proposes an intelligent interdisciplinary PBL lesson plan generation system based on knowledge graphs and multi-agent collaboration to solve the problems existing in the above-mentioned prior art.

[0005] To achieve the above objectives, this invention provides an intelligent interdisciplinary PBL lesson plan generation system based on knowledge graphs and multi-agent collaboration, comprising:

[0006] The interdisciplinary knowledge base construction module is used to semantically vectorize and align entities of teaching content, and to construct an interdisciplinary knowledge graph.

[0007] The retrieval enhancement module is used to perform graph queries and vector searches on interdisciplinary knowledge graphs, and to enhance and supplement interdisciplinary knowledge based on context to obtain knowledge fragments;

[0008] The teaching needs analysis module is used to analyze the teaching needs input by teachers.

[0009] The structured modeling module designs lesson plans in a structured manner based on the cognitive, ability, and subject literacy dimensions of a three-dimensional goal system.

[0010] The lesson plan generation module is used to generate initial drafts of interdisciplinary PBL lesson plans based on teaching needs, knowledge fragments, and structured designs.

[0011] The lesson plan optimization module is used to optimize the initial draft of the structured lesson plan and generate the final lesson plan.

[0012] Optionally, the interdisciplinary knowledge base construction module specifically includes:

[0013] The knowledge extraction unit is used to extract structured knowledge from resources in mathematics, physics, and computer science.

[0014] The entity alignment unit uses parameter isolation training to align entities from heterogeneous data.

[0015] Storage units are used to store related knowledge points as a graph structure and create a vectorized index.

[0016] Optionally, the retrieval enhancement module includes:

[0017] When the graph query result is no match, vector retrieval is triggered to expand the query scope;

[0018] Enhanced knowledge fragments are generated by integrating graph query results with external text resources;

[0019] We use multi-dimensional verification to ensure that the output content is aligned with the curriculum standards.

[0020] Optionally, the vector retrieval adopts a dynamic weighting strategy, prioritizing the retrieval of external literature or textbook passages that are highly relevant to the teaching objectives.

[0021] Optionally, the teaching requirements analysis module includes:

[0022] The teaching topics and grade requirements input by the teacher are analyzed using natural language processing technology;

[0023] Generate structured query commands and output the teaching objective matrix and resource requirements analysis results.

[0024] Optionally, the structured modeling module includes:

[0025] The target decomposition unit is used to model teaching objectives in three dimensions according to Bloom's Taxonomy, the 4C competency framework, and subject core competencies.

[0026] The task orchestration unit is used to decompose interdisciplinary PBL projects into a three-level system of macro-level tasks, meso-level tasks, and micro-level tasks using a directed graph model.

[0027] The outcome specification unit is used to bind the outcome type and the evaluation quantity scale board to generate a structured outcome expression matrix.

[0028] Optionally, the lesson plan generation module includes:

[0029] Math Agent, used to generate mathematical concepts and formula derivations;

[0030] Physics Agent, used to design physics experiment schemes;

[0031] Computer agents are used to write programming examples;

[0032] The coordination unit is used to integrate the content generated by agents from various disciplines.

[0033] Optionally, the lesson plan optimization module includes:

[0034] Interdisciplinary conflicts were detected through dimensional consistency verification and experimental tolerance analysis.

[0035] The weighting parameters are dynamically adjusted based on the subject's contribution.

[0036] Output standardized lesson plans after conflict resolution.

[0037] Optionally, the system further includes:

[0038] The feedback iteration module is used to collect teacher evaluations and extract keywords;

[0039] The update module is used to update the knowledge graph association strength and Agent weight parameters based on the feedback results.

[0040] Compared with the prior art, the present invention has the following advantages and technical effects:

[0041] This invention utilizes knowledge graphs and multi-agent collaborative technology to intelligently generate interdisciplinary PBL lesson plans. This system should be able to understand teachers' teaching needs, automatically retrieve and integrate interdisciplinary knowledge, generate structured lesson plans, and provide optimization suggestions. Such a system will greatly improve the quality of lesson plan design, reduce teachers' preparation time, and simultaneously promote the deep integration of interdisciplinary knowledge and the cultivation of students' comprehensive abilities. Attached Figure Description

[0042] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0043] Figure 1 This is a schematic diagram of the system structure according to an embodiment of the present invention;

[0044] Figure 2 This is a structural abstraction and data modeling mapping diagram for an embodiment of the present invention;

[0045] Figure 3 This is a flowchart illustrating the interdisciplinary knowledge alignment and dynamic maintenance process according to an embodiment of the present invention.

[0046] Figure 4This is a diagram of the LEDVR production line according to an embodiment of the present invention. Detailed Implementation

[0047] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0048] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0049] Example 1

[0050] like Figure 1 As shown, this embodiment provides an intelligent interdisciplinary PBL lesson plan generation system based on knowledge graphs and multi-agent collaboration, including:

[0051] The interdisciplinary knowledge base construction module is used to semantically vectorize and align entities of teaching content, and to construct an interdisciplinary knowledge graph.

[0052] The retrieval enhancement module is used to perform graph queries and vector searches on interdisciplinary knowledge graphs, and to enhance and supplement interdisciplinary knowledge based on context to obtain knowledge fragments;

[0053] The teaching needs analysis module is used to analyze the teaching needs input by teachers.

[0054] The structured modeling module designs lesson plans in a structured manner based on the cognitive, ability, and subject literacy dimensions of a three-dimensional goal system.

[0055] The lesson plan generation module is used to generate initial drafts of interdisciplinary PBL lesson plans based on teaching needs, knowledge fragments, and structured designs.

[0056] The lesson plan optimization module is used to optimize the initial draft of the structured lesson plan and generate the final lesson plan.

[0057] Furthermore, the interdisciplinary knowledge base construction module specifically includes:

[0058] The knowledge extraction unit is used to extract structured knowledge from resources in mathematics, physics, and computer science.

[0059] The entity alignment unit uses parameter isolation training to align entities from heterogeneous data.

[0060] Storage units are used to store related knowledge points as a graph structure and create a vectorized index.

[0061] In this embodiment, firstly, the LEDVR knowledge processing pipeline is used to semantically vectorize teaching content such as mathematical formulas, physical laws, and computer code using the BERT embedding model. Then, a multimodal index is built in the FAISS (Facebook AI SimilaritySearch) vector database to form an educational knowledge graph. This step involves multiple stages, including knowledge extraction, entity alignment, fixed extraction modules, parameter isolation training, and dynamic prompt generation. Next, an augmented reasoning framework is used to encode the knowledge graph subgraphs into soft prompts, and the RAG mechanism is combined to retrieve and supplement interdisciplinary content, providing a solid knowledge foundation for lesson plan generation.

[0062] Furthermore, the retrieval enhancement module includes:

[0063] When the graph query result is no match, vector retrieval is triggered to expand the query scope;

[0064] Enhanced knowledge fragments are generated by integrating graph query results with external text resources;

[0065] We use multi-dimensional verification to ensure that the output content is aligned with the curriculum standards.

[0066] Furthermore, the knowledge graph and RAG-enhanced reasoning are as follows:

[0067] This embodiment constructs a multi-source fusion retrieval system that integrates knowledge graphs (KG), soft prompts, and RAG mechanisms to achieve efficient, low-cost, and high-precision processing of complex knowledge graph question-answering tasks. This scheme draws upon and expands upon the "Retrieve-Embed-Reason" technical process, adding a collaborative driving strategy of soft and hard prompts, as well as an external knowledge retrieval assistance mechanism.

[0068] (1) Multimodal knowledge retrieval and structured semantic parsing:

[0069] The system uses semantic understanding models (such as BERT) to parse Agent requests, automatically predicts the length of multiple inference paths and key entities (anchor entities), and performs relation path retrieval in the knowledge graph based on these anchor entities. To improve retrieval quality and robustness, the system adopts a relation-centered retrieval paradigm and uses a constrained breadth-first search strategy to control the size and relevance of the inference graph.

[0070] (2) Text-knowledge collaborative representation learning:

[0071] The text embeddings of entity descriptions and ontology annotations are concatenated and input into a lightweight Transformer-based Knowledge Adapter, where structural and semantic information is further fused through a self-attention mechanism. Subsequently, the embeddings are mapped to the embedding space of a large language model via a projection module. The resulting "soft hints" are formally equivalent to the original tokens and are concatenated with carefully designed "hard hints" before being injected into the LLM. To quantify the effect, we define a training objective:

[0072]

[0073] Where p p It includes hard hints p h And soft prompts p s The input sequence, a t It is the t-th token in the output sequence. The training objective is to maximize the probability of generating the correct answer A for all samples in the dataset D, ensuring that the model can learn how to generate the correct answer based on the given input during training, without updating the parameters of the LLM itself.

[0074] (3) RAG retrieval inference optimization:

[0075] During the reasoning and generation phases, the system concatenates the aforementioned soft hints with pre-designed hard hints (including task descriptions, answer format constraints, etc.) according to a template and inputs them into a large language model (such as LLaMA-Instruct). To address blind spots in knowledge graph coverage, when the information in the reasoning subgraph is insufficient, the system automatically triggers the RAG module: using FAISS to retrieve semantically similar texts from pre-indexed textbooks and literature passages, injecting the search results as supplementary context into the generated hints, thereby improving the accuracy of answering cold-start and long-tail questions.

[0076] The RAG retrieval enhancement module dynamically supplements knowledge gaps through a multi-level triggering mechanism. When the system detects that the input contains entities not logged in, the generated confidence score is below the threshold, or an explicit instruction is received, it automatically activates the hybrid retrieval process, simultaneously querying structured knowledge bases (such as Wikipedia relationship tables) and unstructured documents (such as textbook and paper paragraphs). Semantic relevance is calculated using a cross-encoder, and entity coverage is combined for re-ranking. Finally, the result is determined by a gating weight α = σ(max...). sim The retrieval results (-0.5) × 0.8) are dynamically integrated with the original knowledge to alleviate the cold start problem and avoid interference from irrelevant information.

[0077] Throughout the system training process, only the knowledge graph adapter parameters are updated, and the LLM is frozen, effectively reducing fine-tuning costs and preventing large models from forgetting. The dynamic cueing strategy module automatically adjusts the soft / hard cue weights based on validation set feedback to balance the contributions of structural and linguistic information.

[0078] Among them, vector retrieval adopts a dynamic weighting strategy, prioritizing the retrieval of external literature or textbook passages that are highly relevant to the teaching objectives.

[0079] Furthermore, the teaching requirements analysis module includes:

[0080] The teaching topics and grade requirements input by the teacher are analyzed using natural language processing technology;

[0081] Generate structured query commands and output the teaching objective matrix and resource requirements analysis results.

[0082] In this embodiment, the system parses the user's input teaching needs through a natural language interface, further optimizes the teaching adaptability using a dialogue guidance engine, extracts key information, and then the decision agent analyzes the project requirements and resource conditions, issuing task assignment instructions to the subject agents. The mathematics agent, physics agent, and computer agent each generate teaching content for their respective subjects. The mathematics agent is responsible for creating preliminary explanations of mathematical concepts and formula derivations, using the SymPy library to verify the correctness of the formula derivations; the physics agent is responsible for designing preliminary introductions to physics experiments and principles, calling the PhET simulation library to optimize experimental parameters; and the computer agent is responsible for writing preliminary programming tasks and code examples, checking the correctness of the code logic using an AST (Abstract Syntax Tree).

[0083] Furthermore, the structured modeling module includes:

[0084] The target decomposition unit is used to model teaching objectives in three dimensions according to Bloom's Taxonomy, the 4C competency framework, and subject core competencies.

[0085] The task orchestration unit is used to decompose interdisciplinary PBL projects into a three-level system of macro-level tasks, meso-level tasks, and micro-level tasks using a directed graph model.

[0086] The outcome specification unit is used to bind the outcome type and the evaluation quantity scale board to generate a structured outcome expression matrix.

[0087] Structured modeling for interdisciplinary lesson plans specifically includes:

[0088] (1) Structural abstraction;

[0089] The structural abstraction process based on PBL teaching theory follows the core logic of "problem-oriented - subject integration - ability cultivation", deconstructing traditional lesson plans into six core element modules.

[0090] Abstract teaching theory: A three-dimensional goal system is adopted, including a cognitive dimension (revised Bloom taxonomy), a competency dimension (4C competency framework: critical thinking, innovation, collaboration, and communication), and a subject-specific core competency dimension. This is formally represented as: G =<K(knowledge),S(skill),A(attitude)> Where K∈{memory, understanding, application, analysis, evaluation, creation}, A represents a set of subject-specific core competencies. The resulting teaching objectives can accurately pinpoint the synergies between different subject objectives, supporting the intelligent integration and dynamic matching of interdisciplinary tasks.

[0091] Project Theme Design: A theme generation model T = (C, R, P) is established, where C is an interdisciplinary concept network (constructed based on knowledge graph technology), R is a function evaluating the relevance of real-world problems, and P is the student interest matching degree. A multi-dimensional interest feature space is constructed based on explicit interests (questionnaire keywords) and implicit interests (past project participation behavior). S(u, r) calculates the matching degree between the student feature vector and the resource tag vector, and outputs the previous... One recommendation result. The method for calculating interest matching degree is as follows:

[0092]

[0093] Where u represents the target student, r represents the resource to be matched, and v u : Student's comprehensive interest vector (explicit and implicit weighted fusion), v r This serves as the feature vector for the resource. Referring to high-quality PBL case studies, the topic must meet the three principles of "authenticity, challenge, and openness," with genuine audience participation as a key evaluation indicator.

[0094] Task orchestration system: Employing a hierarchical task decomposition algorithm, interdisciplinary PBL projects are structurally deconstructed into a three-tiered task system that aligns with teaching logic and cognitive pace: macro-level tasks (overall project objectives) → meso-level tasks (stage deliverables) → micro-level tasks (specific operational steps). The task system is expressed as a directed graph model G = (V, E): V represents the set of task nodes, each carrying attributes including multidisciplinary knowledge tags (extracted from a knowledge graph), required resources and tools (such as development platforms and experimental setups), task type, and expected completion time; E represents the temporal dependencies and logical relationships between tasks, encompassing preconditions, parallel paths, and stage aggregation structures, used to support task scheduling and execution constraint resolution.

[0095] Student Activity Flow: Based on the "Inquiry-Collaboration-Reflection" cyclical mechanism, a five-stage student activity flow for interdisciplinary PBL teaching is designed: ① Problem Situation Experience, ② Inquiry and Research, ③ Solution Design and Implementation, ④ Testing, Evaluation, and Revision, ⑤ Results Presentation and Reflection. The entire process incorporates an AI scaffolding strategy, including concept link recommendations, strategy transfer suggestions, and path tracking and control, ensuring the logical consistency of the task chain and the controllability of the cognitive path. This dynamically supports the interdisciplinary PBL teaching process, achieving intelligent guidance and dynamic support throughout the entire process.

[0096] Standardized outcome formats: This system constructs a structured outcome expression matrix to support the generation of diverse and assessable outcomes in interdisciplinary PBL tasks, ensuring a clear correspondence between outcome expression and teaching objectives and task attributes.

[0097] The formal definition is as follows:

[0098] S={(s i ,R i )|i=1,2,…,n}

[0099] Where s i This represents the i-th type of structured deliverable (such as programming implementation, modeling scheme, experimental design document, simulation verification results, etc.), which is intelligently generated by the system from the deliverable type library based on the task graph node attributes and subject knowledge tags; R i The Rubric template, which is an evaluation metric bound to this type of achievement, includes multi-dimensional indicators such as algorithmic logic, modeling rationality, interdisciplinary relevance, and problem-solving completeness, and adopts a hierarchical weight design.

[0100] The generation and recommendation of outcome types are entirely based on the knowledge tags, skill requirements, and associated teaching objectives of the node tasks in the task system G=(V,E).<K,S,A> The system dynamically schedules outputs. It automatically selects suitable outcome templates based on entity attributes in the knowledge graph (such as the subject of the task and the required cognitive level). Simultaneously, the system integrates an AI-assisted toolchain (such as automatic structure suggestions, content consistency checks, and Rubric correspondence verification) to achieve efficient integration of standardized outcome structures with subsequent evaluation systems. This provides a highly adaptable and quantifiable support system for interdisciplinary teaching activities.

[0101] Evaluation system construction: The evaluation system adopts a hybrid evaluation model A = αA p +βA f +γA s Among them, A p For process-oriented evaluation, it mainly relies on indirect behavioral data such as task submission status, self-evaluation and peer evaluation results, and stage completion rate; A f For formative assessment, focus is placed on the quality of mid-term outcomes and their alignment with teaching objectives; As For summative evaluation, this reflects the completeness of the final outcome and the level of interdisciplinary integration. Coefficients α, β, and γ are weighting parameters that can be flexibly adjusted according to teaching strategies. Evaluation data sources include human scoring, Rubric mapping results, and some traceable task process data, ensuring a scientific evaluation of the entire PBL cycle even under data-constrained conditions.

[0102] (2) Data modeling;

[0103] To achieve standardization and computability of lesson plan data, a hierarchical data modeling approach is adopted:

[0104] Metamodel Layer: Based on the CreativeWork type extension of Schema.org, the knowledge graph defines the lesson plan metadata framework, including core attributes (title, author, subject, applicable grade, class hours, etc.) and management attributes (version, license, modification history, etc.). Semantic annotation is implemented using JSON-LD format to enhance data discoverability and interoperability.

[0105] Conceptual Model Layer: The Pydantic library is used to define a strictly typed lesson plan data class structure. The teaching objective class uses enumeration types to constrain cognitive levels (Bloom levels) and ability types, and regular expressions are used to verify the standardization of competency descriptions. The project context class stores real problem backgrounds, audience information, and interdisciplinary connections in a structured manner, and establishes entity links with external knowledge graphs (such as Wikidata). The task network class uses a graph structure to represent task decomposition relationships, and each node contains attributes such as preconditions, success criteria, and resource links. The assessment rubric class defines multi-dimensional assessment indicators and scoring standards, and supports a mixed representation of qualitative descriptions and quantitative indicators.

[0106] Constraint Rule Layer: Business logic constraints are defined using JSON Schema, ensuring completeness, consistency, and reasonableness. For example, the total required class hours for each task should not exceed the total allocated class hours, and the matching degree between cognitive objectives and assessment methods is verified. Schematron rules are used to implement cross-field logical validation, such as ensuring that higher-order cognitive objectives are paired with corresponding deep learning activity designs. Structural abstraction and data modeling mapping are as follows: Figure 2 .

[0107] (3) Interdisciplinary knowledge alignment and dynamic maintenance;

[0108] Interdisciplinary knowledge alignment and dynamic maintenance process, such as Figure 3 :

[0109] The BERTopic algorithm is used to model terms across disciplines, with a subject-specific dictionary added as a constraint to achieve hierarchical clustering of terms. In the cross-disciplinary association stage, an improved TransE model is used to project entities from each discipline onto a unified vector space. Comparative experiments are conducted to find a performance balance point between 128 and 768 dimensions. A multi-head attention mechanism is used to calculate concept similarity.

[0110]

[0111] Where d k This is the vector dimension, used to adjust numerical stability. Consideration is given to dynamically adjusting the threshold range based on subject combinations. When the cosine similarity of concepts from different subjects exceeds a certain threshold, the system automatically establishes a concept mapping relationship; associations with low confidence are manually reviewed.

[0112] The dynamic maintenance module deploys a gated graph convolutional network (GGCN) architecture. When a new teaching topic is added, the subgraph expansion algorithm automatically associates nodes from multiple disciplines. The system continuously updates the cross-disciplinary relationship matrix to reflect changes in the strength of associations between knowledge points.

[0113] The system employs a Pro-Optimization (PPO) algorithm to automatically plan the optimal teaching path, comprehensively considering 22 key factors such as knowledge relevance, student cognitive load, and equipment resources. The reward function balances four key indicators: knowledge coverage, cognitive coherence, equipment utilization, and student interest index. With knowledge completeness, logical coherence, resource utilization, and learning interest as its core optimization objectives, the system automatically designs diverse teaching plans through an intelligent course generator.

[0114] Furthermore, the lesson plan generation module includes:

[0115] Math Agent, used to generate mathematical concepts and formula derivations;

[0116] Physics Agent, used to design physics experiment schemes;

[0117] Computer agents are used to write programming examples;

[0118] The coordination unit is used to integrate the content generated by agents from various disciplines.

[0119] The design and implementation of a multi-agent collaborative task chain are as follows:

[0120] During the construction and execution of the task chain, the system adopts a state machine-based pipeline architecture and introduces a closed-loop feedback mechanism to ensure process controllability and iterative optimization. The task chain is modeled using the Neo4j graph database, and its structure is based on the core path of "teaching objectives - task modules - resource dependencies - evaluation methods", defining the state transition logic of "to be generated → structure integration → cross-disciplinary verification → encapsulated output".

[0121] (1) Decision Agent: Analysis of teaching needs and issuance of instructions;

[0122] The semantic intent parsing employs an optimized few-shot prompt template recognition mechanism. By embedding typical input examples such as teaching topics, grade requirements, and ability objectives, a subject intent recognition model with generalization capabilities is constructed. Based on the prompt templates, the system performs multiple rounds of semantic discrimination and subject classification on the input task, automatically generating a subject weight matrix. Determine the dominance and level of participation of each agent in the task chain. Dynamically adjust the sparsity constraint parameters of prompts based on user query history and feedback behavior to optimize the accuracy and contextual adaptability of intent parsing.

[0123] (2) Subject Agent: Knowledge Generation and Lesson Plan Filling;

[0124] The three core subject agents (mathematics, physics, and computer science) receive task instructions from the decision agent, generate and populate knowledge modules, and after fusion and sorting, are versioned and encapsulated through the knowledge graph semantic proofreading and structured modeling modules to form a lesson plan draft.

[0125] During the lesson plan generation phase, agents from each subject actively participate in generating teaching tasks and designing questions based on ontology embedding and prior knowledge graphs. The mathematics agent automatically generates challenging theoretical guidance and data processing tasks according to students' grade level and teaching objectives; the physics agent retrieves PhET scenario templates based on core concepts and recommends suitable inquiry experiment paths; and the computer science agent transforms the required programming thinking into executable tasks and designs structured pseudocode examples.

[0126] During the verification and evaluation phase, the mathematics agent uses SymPy for formula verification and algorithm design, the physics agent integrates PhET simulation and principle visualization, and the computer science agent performs AST syntax tree checking and computational thinking transformation. The subject-specific agents achieve knowledge collaboration through a shared weighted association network G = (V, E, w), where edge weights are dynamically updated: w = α·Sim 语义 +(1-α)·CurriculumWeight 课标 .

[0127] (3) Coordination Agent: Lesson plan evaluation, improvement and version management;

[0128] In the process of generating lesson plans, the coordinating agent first receives the initial draft lesson plans from the subject agents and extracts key information using text parsing and information extraction techniques. By collecting teacher feedback, sentiment analysis and topic extraction are used to process the opinions, and algorithms are then applied for targeted optimization. Next, using the Neo4j graph database and semantic matching technology, the lesson plan content is compared with the curriculum standards. Quantitative analysis is performed using Python libraries, and the entropy weight method is used to evaluate the weight of each subject. The formula is as follows:

[0129] Entropy calculation:

[0130] Weight calculation:

[0131] (4) Multi-Agent Conflict Management;

[0132] Standardized communication protocols and information sharing: A unified communication protocol (such as JSON-LD format) is adopted, and core event protocols (such as KnowledgeRequest and ValidationTrigger) are defined to enable information sharing and collaborative work among agents. A global response timeout mechanism is set to improve the overall flexibility and maintainability of the system.

[0133] Dynamic Task Routing and Conflict Detection: A dynamic task routing strategy is introduced to adjust task allocation in real time based on task priority, resource availability, and the current state of the Agent. The system continuously monitors task execution status and promptly detects and identifies potential conflicts, such as resource contention or task overlap.

[0134] Conflict resolution and task reallocation: Adjust the execution order of tasks based on their urgency and importance to ensure that critical tasks are completed first. When resources are limited, dynamically adjust resource allocation to ensure that all tasks can proceed smoothly. For complex or conflicting tasks, split or merge them to simplify the task structure and reduce points of conflict.

[0135] Furthermore, the lesson plan optimization module includes:

[0136] Interdisciplinary conflicts were detected through dimensional consistency verification and experimental tolerance analysis.

[0137] The weighting parameters are dynamically adjusted based on the subject's contribution.

[0138] Output standardized lesson plans after conflict resolution.

[0139] This embodiment constructs a multi-dimensional conflict detection module, encompassing a three-layer verification system: formula dimensional consistency verification, experimental parameter tolerance analysis, and code logic verification. It innovatively designs a subject contribution vector, achieving dynamic conflict resolution through real-time edge weight updates. The system's weight allocation optimizes the priority of each subject parameter in real-time based on the conflict type (such as the cumulative effect of mathematical approximation calculations and physical measurement errors), ultimately outputting an arbitration scheme containing correction suggestions. When parameter contradictions are detected, the system automatically generates a joint diagnostic report, effectively alleviating the pain point of low efficiency in manual coordination.

[0140] Furthermore, the system also includes:

[0141] The feedback iteration module is used to collect teacher evaluations and extract keywords;

[0142] The update module is used to update the knowledge graph association strength and Agent weight parameters based on the feedback results.

[0143] Furthermore, this system uses LangChain's LEDVR (Loader-Embedder-Divider-VectorStore-Retriever) pipeline as its core to build a multidisciplinary knowledge engine, achieving full automation of the data loading and intelligent retrieval process. The LEDVR pipeline, as shown... Figure 4 As shown.

[0144] Using LangChain's DocumentLoaders module in conjunction with LangSmithLoader, teaching resources can be loaded from data sources across multiple disciplines. Vectorization is performed through embedded models, followed by semantic segmentation, and the results are ultimately stored in a vector database to support efficient semantic retrieval. The Agents module allows the construction of intelligent agents across multiple disciplines, configuring specific Tool modules. The AgentExecutor module coordinates the agents, optimizing lesson plan content based on reinforcement learning strategies (such as proximal policy optimization algorithms) to ensure content quality and teaching effectiveness. The front-end can use the React framework, and the back-end uses the FastAPI framework, combined with the LangServe module, deploying the constructed chain as a RESTful API to achieve efficient communication between the front-end and back-end.

[0145] To achieve a self-governing closed loop, the system uses LangSmith to record generation logs and user feedback, combined with improved Git version control to track subject contributions, dynamically adjusting agent fine-tuning parameters and knowledge graph edge weights, forming a self-evolving mechanism of "generation-verification-optimization-output". Ultimately, through a knowledge engine, multi-agent collaboration, and continuous learning, an efficient, scalable, and secure intelligent lesson plan generation system is constructed.

[0146] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A cross-disciplinary PBL lesson plan intelligent generation system based on knowledge graphs and multi-agent collaboration, characterized in that, include: The interdisciplinary knowledge base construction module is used to semantically vectorize and align entities of teaching content, and to construct an interdisciplinary knowledge graph. The retrieval enhancement module is used to perform graph queries and vector searches on interdisciplinary knowledge graphs, and to enhance and supplement interdisciplinary knowledge based on context to obtain knowledge fragments; The teaching needs analysis module is used to analyze the teaching needs input by teachers. The structured modeling module designs lesson plans in a structured manner based on the cognitive, ability, and subject literacy dimensions of a three-dimensional goal system. The lesson plan generation module is used to generate initial drafts of interdisciplinary PBL lesson plans based on teaching needs, knowledge fragments, and structured designs. The lesson plan optimization module is used to optimize the initial draft of the structured lesson plan and generate the final lesson plan.

2. The system according to claim 1, characterized in that, The interdisciplinary knowledge base construction module specifically includes: The knowledge extraction unit is used to extract structured knowledge from resources in mathematics, physics, and computer science. The entity alignment unit uses parameter isolation training to align entities from heterogeneous data. Storage units are used to store related knowledge points as a graph structure and create a vectorized index.

3. The system according to claim 1, characterized in that, The retrieval enhancement module includes: When the graph query result is no match, vector retrieval is triggered to expand the query scope; Enhanced knowledge fragments are generated by integrating graph query results with external text resources; We use multi-dimensional verification to ensure that the output content is aligned with the curriculum standards.

4. The system according to claim 3, characterized in that, The vector retrieval employs a dynamic weighting strategy, prioritizing the retrieval of external literature or textbook passages that are highly relevant to the teaching objectives.

5. The system according to claim 1, characterized in that, The teaching needs analysis module includes: The teaching topics and grade requirements input by the teacher are analyzed using natural language processing technology; Generate structured query commands and output the teaching objective matrix and resource requirements analysis results.

6. The system according to claim 1, characterized in that, The structured modeling module includes: The target decomposition unit is used to model teaching objectives in three dimensions according to Bloom's Taxonomy, the 4C competency framework, and subject core competencies. The task orchestration unit is used to decompose interdisciplinary PBL projects into a three-level system of macro-level tasks, meso-level tasks, and micro-level tasks using a directed graph model. The outcome specification unit is used to bind the outcome type and the evaluation quantity scale board to generate a structured outcome expression matrix.

7. The system according to claim 1, characterized in that, The lesson plan generation module includes: Math Agent, used to generate mathematical concepts and formula derivations; Physics Agent, used to design physics experiment schemes; Computer agents are used to write programming examples; The coordination unit is used to integrate the content generated by agents from various disciplines.

8. The system according to claim 1, characterized in that, The lesson plan optimization module includes: Interdisciplinary conflicts were detected through dimensional consistency verification and experimental tolerance analysis. The weighting parameters are dynamically adjusted based on the subject's contribution. Output standardized lesson plans after conflict resolution.

9. The system according to claim 1, characterized in that, Also includes: The feedback iteration module is used to collect teacher evaluations and extract keywords; The update module is used to update the knowledge graph association strength and Agent weight parameters based on the feedback results.

Images