A knowledge graph and large model-based intelligent teaching three-dimensional closed-loop system

Abstract

The application discloses a kind of wisdom teaching three-dimensional closed loop system based on knowledge graph and big model, including data and resource layer, knowledge graph pedestal layer, intelligent engine layer, three-dimensional application service layer and interactive presentation layer.System automatically constructs and dynamically updates the attribute graph model knowledge graph covering course knowledge, ability target, post demand, practice resource;Through the multi-modal analysis of learning situation and knowledge tracking, precise teaching targeting push is realized;Fusion big model and knowledge graph semantic reasoning, dynamically plan personalized learning path for students and provide intelligent answering;Establish "knowledge-ability-resource-post" four-element association network, break through practice application and ability certification link.Three-dimensional business data backflow forms closed loop iteration, significantly improve the intelligent level of teaching and personalized learning effect.

Description

Technical Field

[0001] This invention belongs to the field of interdisciplinary technology of Educational Technology (EdTech) and Artificial Intelligence. Specifically, it relates to a three-dimensional closed-loop system for intelligent teaching based on knowledge graphs and large models. This system is used to achieve deep integration and dynamic closed loop of three dimensions: precise teaching, personalized learning, and practical application. It is applicable to scenarios such as online education platforms, smart campuses, and corporate training. Background Technology

[0002] In the field of educational technology, with the popularization of online education and the deepening of the concept of smart education, using technological means to achieve personalized teaching and precise intervention at a large scale has become a core development direction. Among them, knowledge graph technology, due to its excellent semantic association and structured representation capabilities, is widely regarded as a key infrastructure for building a new generation of intelligent teaching systems. Currently, the application of knowledge graphs in teaching mainly revolves around the visualization of course knowledge, personalized learning path recommendation, and learning analysis, aiming to solve problems in traditional teaching such as rigid resource organization, delayed learning feedback, and difficulty in accurately adapting teaching methods.

[0003] Existing technologies for constructing and applying instructional knowledge graphs generally suffer from the following problems:

[0004] 1. The knowledge graph construction process is highly dependent on manual labor, with a serious lack of automation and intelligence.

[0005] This is the most common and core flaw. Currently, the construction of many pedagogical knowledge graphs heavily relies on domain experts manually compiling knowledge points, defining relationships, and annotating teaching resources, or on crowdsourced annotation methods. This approach is inefficient, costly in terms of manpower and time, and struggles to quickly respond to the needs of building and updating multidisciplinary, large-scale teaching resource databases. While automated solutions exist for converting relational databases using ETL technology, their support for multi-source heterogeneous and distributed databases is limited, they struggle to handle complex semantic mappings, and they lack versatility.

[0006] 2. The semantic depth of knowledge associations is limited, making it difficult to support truly precise teaching and the cultivation of complex abilities.

[0007] (1) Coarse granularity and low accuracy of association: Existing resource annotation is often coarse-grained, making it difficult to achieve accurate matching at the knowledge point level. In the process of automated association, noise such as knowledge point abbreviations, aliases, abbreviations and even typos that are common in teaching texts (such as subtitles) leads to poor fault tolerance and low accuracy of entity recognition and linking methods based on full matching or traditional natural language processing models.

[0008] (2) Superficial understanding of logical relationships: Existing knowledge graphs mostly focus on surface connections, and fail to adequately depict the complex causal, progressive, and comparative deep logical relationships between knowledge points, as well as the internal path of transforming theoretical knowledge into practical skills. This makes it difficult for the graphs to support the development of students' systematic cognitive framework and complex problem-solving abilities, and students are prone to falling into point-based understanding.

[0009] (3) Lack of new knowledge discovery and dynamic update mechanism: The system relies heavily on the preset static knowledge point base. When new concepts or cutting-edge technologies not included in the knowledge base appear in the teaching resources, there is a lack of effective new word discovery and automatic association mechanism, forming a knowledge coverage blind spot and causing the recommendation to fail.

[0010] 3. There is a serious disconnect between theory and practice, and practical resources are detached from the overall framework.

[0011] Many knowledge graphs in various systems only include textbook knowledge points and exercise resources, excluding practical resources such as experimental datasets, algorithm models, enterprise cases, and job tasks. After completing theoretical learning, students struggle to obtain accurately matched and appropriately challenging practical task recommendations, resulting in a break in the learning-to-application transition.

[0012] 4. The intelligent learning companion has weak capabilities and lacks generative and contextualized interaction.

[0013] While the concept of a dual-drive approach combining data and knowledge has been proposed, the deep integration of data-driven and knowledge-guided learning remains superficial. How to deeply and effectively integrate and collaboratively reason with real-time, dynamic, and multimodal big data on learning behavior and static, structured domain knowledge graphs to generate truly adaptive, interpretable, and dynamically adjustable personalized learning companions remains a challenge in current technological practice. Existing recommendation systems are prone to problems such as static paths and single-dimensional profiles (over-reliance on knowledge mastery levels while neglecting multi-dimensional data such as emotions and cognitive processes), making it difficult for them to fulfill the role of a true "AI learning companion."

[0014] 5. Business functions are isolated and have not formed a closed-loop system that can be iteratively optimized.

[0015] Current technical solutions mostly focus on optimizing single functionalities (such as path recommendation and learning diagnosis), lacking a system architecture that organically couples the three core business functions of precise teaching on the teacher's side, personalized learning assistance on the student's side, and cultivation of practical application abilities. Data cannot be recycled across dimensions, and the system lacks the ability to evolve on its own. Summary of the Invention

[0016] To address the core pain points of existing knowledge graph-based teaching technologies, such as reliance on manual construction, shallow connections, performance limitations, and insufficient integration of data and knowledge, this invention proposes a three-dimensional closed-loop intelligent teaching system based on knowledge graphs and large-scale models. The core objective of this invention is to provide a closed-loop intelligent teaching system that uses a dynamic knowledge graph as a unified central hub, a generative large-scale language model as the core of intelligent interaction, and integrates precise teaching, personalized learning support, and practical application across three dimensions. The specific objectives of this invention are broken down as follows:

[0017] (i) Achieve full-process, automated knowledge graph construction and dynamic updates, reducing application thresholds and costs. Addressing the shortcomings of existing technologies, such as heavy reliance on human experts, low efficiency, high cost, difficulty in handling multi-source heterogeneous data, and challenges in updating and maintaining, this paper provides a technical solution to automatically extract knowledge entities, identify relationships, and construct a visual course knowledge graph from multi-source teaching data (such as relational databases, learning platform logs, and cutting-edge industry materials). This enables low-cost, high-efficiency knowledge graph construction and dynamic optimization.

[0018] (II) Constructing a Deeply Semantically Recognized Four-Element Relationship Network of "Knowledge-Ability-Resource-Job". Addressing the problems of existing technologies having coarse-grained relationships and shallow semantic depth, making it difficult to depict complex logical relationships between knowledge points (such as causality, progression, and application transfer), and lacking effective handling capabilities for textual noise (such as abbreviations and typos) and new knowledge, we design more advanced semantic understanding and association algorithms to deeply mine and formally represent the inherent semantic connections between knowledge points, as well as between theoretical knowledge nodes and job skills, practical cases, and experimental resources. This will construct a more tightly linked, logically clearer, and more comprehensive integrated theory-practice knowledge network.

[0019] (III) Enhance the system's scalability and real-time response capabilities to ensure stable performance in large-scale teaching scenarios. Addressing the challenges of computational and storage performance bottlenecks in querying, visualization, and real-time recommendation as the scale of knowledge graphs expands, optimize the underlying architecture and algorithms to ensure that the system maintains efficient graph retrieval, real-time learning analysis, personalized path calculation, and graphical rendering capabilities even when dealing with large-scale knowledge graphs with tens or even hundreds of millions of nodes and relationships.

[0020] (iv) Achieve deep integration of data-driven and knowledge-guided approaches to support a three-dimensional closed loop of precision teaching, personalized learning, and practical application. Break down the barriers between data and knowledge, and create a mechanism that uses a dynamically updated knowledge graph as the intelligent hub and deeply integrates real-time multimodal learning data, enabling: ① Precision teaching: Based on graph associations and real-time learning data, targeted delivery of teaching content and dynamic design of hybrid processes; ② Personalized learning: Based on the logical relationship between graph nodes and combined with real-time learning behavior data, dynamically generate truly personalized and adaptive learning paths for each student, and provide precise real-time Q&A and learning analysis dashboards; ③ Practical application: Utilize the graph to structure and organize course practice resource libraries (such as datasets and model libraries), and connect knowledge points with job tasks and practical cases to strengthen the transformation of learning into practical applications.

[0021] To achieve the above objectives, the present invention adopts the following technical solution:

[0022] This invention proposes a three-dimensional closed-loop intelligent teaching system based on knowledge graphs and large models, comprising:

[0023] The data and resource layer is used to collect and manage multi-source heterogeneous teaching data and resources, including structured and unstructured data.

[0024] The knowledge graph base layer, connected to the data and resource layer, is used to automatically construct and dynamically update a domain knowledge graph covering course knowledge, ability objectives, job requirements, and practical resources. The knowledge graph is stored in an attribute graph model and includes knowledge point entities, ability entities, job entities, practical resource entities, and their semantic relationships.

[0025] The intelligent engine layer includes:

[0026] The learning analysis engine is used to perform multimodal analysis of learning behavior and generate dynamic learner profiles and knowledge status diagnoses.

[0027] A personalized recommendation engine for path planning and resource matching based on knowledge graph semantic recommendations;

[0028] AIGC and intelligent interaction engine are used to provide intelligent Q&A, content generation and natural language interaction services;

[0029] The teaching evaluation and optimization engine is used to analyze teaching effectiveness, trigger early warnings, and optimize graph relationships and recommendation strategies based on feedback data.

[0030] The 3D application service layer includes:

[0031] The precision teaching module is used to achieve targeted delivery of teaching content and dynamic design of teaching processes based on knowledge graphs and learning analysis.

[0032] The personalized learning companion module is used to dynamically generate personalized learning paths for each student based on knowledge graphs and large models, and provide real-time intelligent Q&A.

[0033] The practical application module is used to organize practical resources based on knowledge graphs and connect knowledge points with job tasks and practical cases.

[0034] The interactive presentation layer is used to provide teachers, students, and administrators with a multi-terminal visual interactive interface;

[0035] The precision teaching module, personalized learning companion module, and practical application module achieve data interconnection and business linkage through the knowledge graph base layer, forming a three-dimensional closed loop of "teaching-learning companion-application".

[0036] Furthermore, the knowledge graph base layer includes:

[0037] The structured data conversion unit is used to automatically map relational teaching resource databases to attribute graph models. It generates mapping rules from relational models to graph models by analyzing the table structure and primary and foreign key metadata of the source database, thereby achieving batch automated semantic conversion.

[0038] The unstructured resource association unit is used to identify and link knowledge point entities in teaching texts. It uses a BERT pre-trained model combined with an LSTM-CRF sequence labeling model for entity recognition, and uses edit distance for fuzzy matching and cosine similarity for semantic disambiguation to link resources to corresponding knowledge point nodes in the knowledge graph.

[0039] The dynamic update unit is used to automatically adjust the correlation coefficients and weights between nodes in the knowledge graph based on learning behavior data, and to capture new knowledge and new cases from external data sources to trigger incremental updates.

[0040] Furthermore, the structured data conversion unit executes an intelligent attribute graph model mapping algorithm, including:

[0041] Collect metadata from the source database, including table-level information, field-level information, constraints and indexes, and data statistics.

[0042] Each table is comprehensively scored across multiple dimensions, and the table is mapped to a node, relationship, or attribute based on the score and business characteristics.

[0043] Based on foreign key naming patterns and characteristics of associated tables, semantically infer relationship types, including "contains", "prerequisite", "belongs to", "depends on", and "hierarchical containment";

[0044] Standardize fields and adapt them to data types, and generate index recommendations;

[0045] The mapping decision results are converted into Cypher scripts, which are used to build nodes, relationships, and indexes in the Neo4j graph database.

[0046] Furthermore, the personalized learning companion module includes:

[0047] The personalized learning path planning unit is used to generate the optimal learning sequence from the current knowledge state to the target knowledge state on the course knowledge graph by taking the student's personal dynamic knowledge graph as input, combining learning objectives and style preferences, and using graph search algorithms or reinforcement learning models. The learning sequence respects the prerequisite dependencies of knowledge points and is dynamically adjusted according to the student's latest performance.

[0048] The large model coupling interaction unit is used to deeply couple the AIGC engine with the knowledge graph. When a student asks a question, the question is parsed and the knowledge graph is used to evoke prior knowledge, analogous concepts, or extended applications, generating a contextualized and logical answer, and simultaneously pushing practical resources related to the current knowledge point.

[0049] Furthermore, the precision teaching module includes:

[0050] The learning progress diagnosis and visualization unit is used to map students' behavioral data on various platforms to the corresponding nodes of the knowledge graph in real time. Through knowledge tracking models and cluster analysis, it generates a heat map of the class's overall knowledge mastery and a dynamic knowledge and ability graph of individuals.

[0051] The targeted push unit is used to automatically match and push explanatory videos, case studies, and exercise resources from the resource library based on the teacher's selected weak knowledge points and relying on the semantic association network of the knowledge graph;

[0052] The real-time early warning unit is used to automatically send a real-time warning to the teacher when the error rate of answering a certain knowledge point exceeds a preset threshold during classroom interaction.

[0053] Furthermore, the practical application module includes:

[0054] The practical resource graphing unit is used to incorporate datasets, algorithm model libraries, experimental manuals, project task books, and enterprise cases as practical resource entities into the knowledge graph, and establish the "explanation / application" relationship between entities and theoretical knowledge points, the "cultivation" relationship with ability goals, and the "correspondence" relationship with job tasks, forming a four-element association network of "knowledge-ability-resource-job".

[0055] The practical task matching unit is used to automatically recommend relevant practical operations, datasets or simulation training based on knowledge graph associations after students have completed the learning of theoretical knowledge points. It is also used in project-based learning to intelligently decompose the required knowledge modules and ability units according to project objectives and recommend learning resource sequences.

[0056] The competency certification unit is used to collect data from the practical process and compare it with the preset competency standards in the knowledge graph. It automatically evaluates the standardization of practical operations and the efficiency of problem solving, generates a competency growth report, and issues digital badges that are bound to competency nodes.

[0057] Furthermore, it also includes a closed-loop optimization unit, which is used to feed back all the group learning data generated by the precision teaching module, the individual interaction data generated by the personalized learning module, and the process and result data generated by the practical application module to the data center, and use the feedback data to continuously optimize the weight of entity relationships in the knowledge graph, the accuracy of personalized recommendation strategies, and the quality of AIGC responses through machine learning algorithms.

[0058] Furthermore, the knowledge graph base layer is stored in a distributed graph database, and the intelligent engine layer and the 3D application service layer are deployed in a microservice architecture to support high-concurrency querying and rendering of large-scale graphs with tens of millions of nodes and relationships.

[0059] Furthermore, the teaching evaluation and optimization engine constructs an evaluation model that integrates knowledge graphs. By recording students' homework modification process, discussion thought process, and project practice logical path, it compares and analyzes these with standard solution paths in the knowledge graph to evaluate the accuracy of technology application, the innovation of solutions, and the level of ability transfer, thereby realizing the transformation from knowledge assessment to ability evaluation.

[0060] This invention also proposes a three-dimensional closed-loop intelligent teaching method based on any of the above-described systems, comprising the following steps:

[0061] Step 1: Collect multi-source heterogeneous teaching data and resources through the data and resource layer, and perform standardized management;

[0062] Step 2: Automatically construct and dynamically update a domain knowledge graph covering course knowledge, competency objectives, job requirements, and practical resources through a knowledge graph base layer;

[0063] Step 3: Use the learning analysis engine to perform multimodal analysis of learning behavior and generate learner profiles and knowledge status diagnoses;

[0064] Step 4: Through the precision teaching module, based on knowledge graphs and learning analysis, targeted delivery of teaching content and dynamic design of teaching processes are achieved;

[0065] Step 5: Through the personalized learning companion module, a personalized learning path is dynamically generated for each student based on knowledge graphs and large models, and real-time intelligent Q&A is provided.

[0066] Step 6: Through the practical application module, organize practical resources based on the knowledge graph and establish the connection links between knowledge points, job tasks, and practical cases;

[0067] Step 7: Feed the data generated in steps 4, 5, and 6 back to the data center, and optimize the knowledge graph and recommendation strategy through the teaching evaluation and optimization engine to form a three-dimensional closed-loop iteration.

[0068] Compared with the prior art, the present invention has the following beneficial effects:

[0069] Compared to existing technologies, this invention achieves significant beneficial effects through the deep integration of automated knowledge graph construction, deep semantic association, high-performance scalable architecture, and the deep fusion and system implementation of a three-dimensional closed loop of "precise teaching - personalized learning - practical application," specifically reflected in the following aspects:

[0070] (i) The efficiency and intelligence of knowledge graph construction have been significantly improved.

[0071] Traditional methods rely heavily on manual construction by domain experts, resulting in high costs, slow updates, and shallow connections. This invention achieves a fundamental breakthrough:

[0072] 1. Significantly improved efficiency and intelligence. Through multi-source heterogeneous data fusion technology, data from academic affairs, student affairs, and learning platforms are automatically cleaned, standardized, and semantically aligned, providing a high-quality data foundation for graph construction. The core lies in using ETL technology to automatically map relational databases to attribute graph models and batch import them into graph databases (such as Neo4j), achieving automated and semantic transformation and organization of existing structured data. Simultaneously, for unstructured teaching resources (such as video subtitles), entity recognition technology integrating BERT pre-trained models, bidirectional maximum matching algorithms, LSTM-CRF sequence annotation, and conditional random fields is employed. Combined with edit distance fuzzy matching and cosine similarity calculation, automated entity linking and association between teaching resources and knowledge point bases are achieved, significantly reducing the cost of manual annotation and association.

[0073] 2. Enhanced Semantic Relationship Depth and Cognitive Logic. It not only constructs basic connections between knowledge points (such as prerequisites and successors), but also, through natural language processing and relation extraction techniques, delves into and establishes complex logical relationships between knowledge points (such as causality, analogy, and parallelism) as well as cross-domain semantic links between "theory-practice-case studies." This upgrades the knowledge graph from a flat resource directory into a three-dimensional network that reflects the inherent cognitive logic and knowledge transfer paths of a discipline.

[0074] 3. Dynamic Update and Self-Optimization Capabilities. The system possesses a continuous data-driven update mechanism. By analyzing students' learning behavior data (learning trajectories, interactions, tests), and utilizing multimodal and semantic analysis, it dynamically infers students' knowledge mastery status and learning needs, and updates their individual learning graphs accordingly. Simultaneously, the system can connect to external industry databases, policy and regulatory libraries, etc., and automatically detect new knowledge and cases using generative artificial intelligence and other technologies, triggering the knowledge graph update process. This ensures the timeliness and cutting-edge nature of the teaching content, making the knowledge graph a continuously growing knowledge ecosystem.

[0075] (II) Significant Effects of Precision Teaching

[0076] Based on the aforementioned dynamic and deeply interconnected knowledge graph, the precision of teaching on the teacher's end has achieved a qualitative leap:

[0077] 1. More Precise Learning Diagnosis and Profiling. The system integrates multimodal data from online learning platforms, classroom interactions, and homework tests. Utilizing techniques such as cluster analysis (e.g., K-means), association rule mining, and machine learning, it constructs dynamic learner profiles encompassing multiple dimensions, including knowledge level, behavioral characteristics, interests, and emotional state. Combined with knowledge graph-based knowledge status diagnosis, it clearly and visually presents the overall class and each student's mastery of various knowledge points (e.g., generating mastery heatmaps and individual dynamic knowledge graphs), identifying common weaknesses and individual differences, thus transforming the process from vague perception to precise measurement.

[0078] 2. More Timely Teaching Intervention and Resource Delivery. The system establishes a full-process (pre-class, in-class, and post-class) assessment and real-time early warning mechanism. For example, during class, when the error rate of real-time answers exceeds a preset threshold (e.g., 25%), the system can automatically issue an early warning to the teacher. Based on the visualized "learning progress dashboard" (e.g., knowledge mastery radar chart, performance trend chart) provided by the system, teachers can immediately adjust the teaching pace, conduct targeted explanations, or initiate group peer learning. After class, based on the diagnostic results, the system can automatically and accurately push targeted practice questions, micro-lessons, or extended case studies to students or groups from a resource library deeply linked to the knowledge graph, achieving targeted teaching that "fills in the gaps in understanding."

[0079] 3. More scientific and comprehensive teaching evaluation dimensions. Moving beyond single-score assessment, the system utilizes the relationships within a knowledge graph to evaluate students' overall understanding and ability to apply knowledge. For example, by analyzing the knowledge points students use in solving complex problems, their logical reasoning, and the innovativeness of their assignments, the system can evaluate higher-order thinking skills and knowledge application abilities.

[0080] (III) Significant Effects of Personalized Learning Companionship

[0081] For students, the system provides a truly personalized learning support environment, acting as an "AI learning companion":

[0082] 1. Dynamic Personalized Learning Path Planning. Based on the analysis of students' real-time knowledge status (knowledge tracking results), learning objectives, and style preferences, the system uses a knowledge graph as a navigation map and leverages graph search algorithms (such as...) Algorithms or reinforcement learning models are used to dynamically generate and continuously optimize personalized learning paths. These paths not only follow the logic of prior knowledge acquisition but also dynamically adjust based on student learning feedback, avoiding excessive cognitive load and achieving personalized adaptive learning.

[0083] 2. Intelligent Accompaniment of Learning Resources and Support. Employing a graph-based semantic recommendation strategy, the system accurately matches learning resources to students. For example, it prioritizes recommending videos and charts for visual learners, and suggests relevant basic theories or advanced application cases based on their weaknesses. More importantly, the system can integrate generative artificial intelligence (AIGC) as an intelligent teaching assistant, providing real-time Q&A, problem-solving guidance, automatic homework grading, and generative feedback, achieving personalized academic support 24 / 7.

[0084] 3. Intelligent Discovery and Construction of Learning Communities. By analyzing learners' knowledge graph status, interest tags, and communication behaviors, the system can intelligently identify and recommend learning partners with similar learning goals or complementary knowledge structures, promoting collaborative learning and experience sharing, and building online learning communities.

[0085] (iv) Significant effects in practical application dimension

[0086] The core of this invention addresses the pain point of the disconnect between theory and practice, and builds a bridge for the transformation of knowledge into ability:

[0087] 1. Graphical Connection Between Practical Resources and Theoretical Knowledge. Practical resources such as experimental projects, engineering cases, and internship tasks are also incorporated as entities into the knowledge graph, and are explicitly and structurally linked to relevant theoretical knowledge nodes, skill requirements, and job competencies. Students can intuitively see what practical cases support a certain theory and what knowledge combinations are required for a certain skill, thus bridging the visual path from theory to practice.

[0088] 2. Personalized Intelligent Matching of Practical Tasks. Based on students' professional direction, knowledge graph mastery, and career interest profile, the system can intelligently match and recommend practical tasks or project topics of moderate difficulty and high relevance from the practical resource graph, making practical training more targeted and challenging.

[0089] 3. Competency-based learning outcome certification. After students complete the learning path and practical tasks, their process and outcome data (such as the clusters of knowledge points mastered, the complexity of the completed practical projects, and the types of problems solved) are recorded and mapped onto a competency model. Based on this, the system can issue fine-grained digital badges or competency certificates, objectively and reliably reflecting the specific skills and comprehensive abilities they have acquired, rather than just the score of a single course, thus strongly supporting employment and lifelong learning.

[0090] (v) Significant Effects of System Performance and Architecture

[0091] 1. Improved processing performance and response efficiency. By employing a distributed graph database to store the knowledge graph and combining it with a microservice architecture, the system can effectively support the management and querying of tens of millions of nodes and relationships, meeting the needs of large-scale concurrent access. In the front-end visual interaction, batch processing optimization (such as setting the optimal batch size) and targeted performance tuning ensure the smoothness of large-scale graph rendering and operations.

[0092] 2. Enhanced Architectural Flexibility and Ecosystem Openness. The five-layer decoupled architecture allows each layer to be upgraded and expanded independently. Through standard API interfaces, the system can flexibly connect to new data sources (such as new teaching tools and industry databases), new intelligent algorithm engines (such as updated AI models), and diverse terminal applications, possessing good ecosystem inclusiveness and future adaptability, and can continuously evolve to meet the ever-changing educational technology needs. Attached Figure Description

[0093] Figure 1 This is a schematic diagram of the architecture of a three-dimensional closed-loop intelligent teaching system based on knowledge graphs and large models, according to an embodiment of the present invention.

[0094] Figure 2 This is a schematic diagram of the intelligent attribute graph model mapping algorithm provided in an embodiment of the present invention;

[0095] Figure 3 This is a flowchart illustrating a three-dimensional closed-loop method for intelligent teaching based on knowledge graphs and large models, according to an embodiment of the present invention. Detailed Implementation

[0096] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments:

[0097] This invention proposes a technical solution for a three-dimensional closed-loop smart teaching system that uses knowledge graphs as a unified foundation and deeply integrates big data, artificial intelligence, distributed computing, and data security technologies to systematically realize "precise teaching - personalized learning support - practical application." This invention aims to solve core problems such as loose organization of teaching resources, inadequate student assessment, and a disconnect between theory and practice. By constructing a self-driven, scalable, and highly responsive smart teaching operating system, it specifically addresses the four core requirements of the invention's purpose.

[0098] like Figure 1 As shown, the present invention provides a three-dimensional closed-loop intelligent teaching system based on knowledge graphs and large models, comprising:

[0099] (I) Overall Technical Architecture

[0100] The core of this invention is to construct a smart teaching system that is knowledge-guided, data-driven, human-machine collaborative, and iteratively closed-loop. Its overall technical architecture follows a modular and service-oriented design philosophy, divided into five layers, aiming to provide solid technical support for a three-dimensional teaching closed loop.

[0101] Data and Resource Layer: This layer is used to collect and manage multi-source, heterogeneous teaching data and resources, including both structured and unstructured data. Specifically, as the system foundation, this layer aggregates and manages multi-source, heterogeneous teaching data and resources, including structured and unstructured data from online teaching platforms (such as Chaoxing and Yu Classroom), academic affairs systems, practical platforms, and IoT devices, as well as teaching resources such as course materials, videos, question banks, case studies, datasets, and algorithm models. Through unified data governance and ETL pipelines, it achieves data standardization, cleaning, and quality control, providing a reliable data source for upper-layer applications.

[0102] Knowledge Graph Foundation Layer: Connected to the data and resource layer, this layer is used to automatically construct and dynamically update a domain knowledge graph covering course knowledge, competency objectives, job requirements, and practical resources. The knowledge graph is stored using an attribute graph model and includes knowledge point entities, competency entities, job entities, practical resource entities, and their semantic relationships. This layer is the core innovation hub of the system. Employing automated construction and dynamic updating technologies, it extracts, associates, and integrates semantic information from the data and resource layer to construct a domain knowledge graph covering course knowledge, competency objectives, job requirements, and practical resources. This graph is not only a static knowledge structure graph but also a dynamic semantic network connecting teaching, learning, application, and assessment, providing a unified logical framework for precise and personalized services.

[0103] Intelligent Engine Layer: Encapsulates core algorithms and models, serving as the system's "brain." It mainly includes:

[0104] Learning Analysis Engine: Based on machine learning and data mining, it performs multimodal analysis of learning behavior and generates dynamic learner profiles and knowledge status diagnoses.

[0105] Personalized recommendation engine: Used to employ graph-based semantic recommendations to provide intelligent decision-making for path planning and resource matching.

[0106] AIGC and Intelligent Interaction Engine: Integrating generative artificial intelligence to provide natural language interaction services such as intelligent Q&A, content generation, and homework grading.

[0107] Teaching evaluation and optimization engine: used to analyze teaching effectiveness in real time, trigger alerts, and optimize graph relationships and recommendation strategies based on feedback data.

[0108] The three-dimensional application service layer directly addresses the three business dimensions of "precision teaching," "personalized learning support," and "practical application," providing specific functional modules and service interfaces. This includes:

[0109] The precision teaching module is used to achieve targeted delivery of teaching content and dynamic design of teaching processes based on knowledge graphs and learning analysis.

[0110] The personalized learning companion module is used to dynamically generate personalized learning paths for each student based on knowledge graphs and large models, and provide real-time intelligent Q&A.

[0111] The practical application module is used to organize practical resources based on knowledge graphs and to connect knowledge points with job tasks and practical cases.

[0112] The services across different dimensions are not isolated, but rather interconnected through a knowledge graph foundation layer to achieve data exchange and business collaboration, collectively forming a closed loop for teaching.

[0113] Interactive Presentation Layer: This layer provides teachers, students, and administrators with a multi-terminal, visually-based interactive interface. It includes a learning data dashboard for teachers, precise lesson preparation and intervention tools, a personalized learning space for students, an AI learning companion interface, and an operational environment for practical applications. Advanced graphics rendering and interactive technologies are employed to ensure a superior user experience.

[0114] The precision teaching module, personalized learning companion module, and practical application module achieve data interoperability and business linkage through the knowledge graph base layer, forming a three-dimensional closed loop of "teaching-learning companion-application". This architecture achieves loosely coupled communication between layers through a service bus and API gateway, and integrates supporting technologies such as high-concurrency processing, distributed storage, and comprehensive security protection. This ensures the stability, performance, and security of the system in large-scale application scenarios.

[0115] (II) Specific Implementation of the Knowledge Graph Base Layer

[0116] Preferably, the knowledge graph base layer includes:

[0117] The structured data conversion unit is used to automatically map relational teaching resource databases to attribute graph models. It generates mapping rules from relational models to graph models by analyzing the table structure and primary and foreign key metadata of the source database, thereby achieving batch automated semantic conversion.

[0118] The unstructured resource association unit is used to identify and link knowledge point entities in teaching texts. It uses a BERT pre-trained model combined with an LSTM-CRF sequence labeling model for entity recognition, and uses edit distance for fuzzy matching and cosine similarity for semantic disambiguation to link resources to corresponding knowledge point nodes in the knowledge graph.

[0119] The dynamic update unit is used to automatically adjust the correlation coefficients and weights between nodes in the knowledge graph based on learning behavior data, and to capture new knowledge and new cases from external data sources to trigger incremental updates.

[0120] Furthermore, the structured data conversion unit executes an intelligent attribute graph model mapping algorithm. The core objective of this algorithm is to automate and batch map relational teaching resource databases (such as MySQL and PostgreSQL) to attribute graph models (taking Neo4j as an example), eliminating the need for manual table-by-table rule definition. Simultaneously, it adapts to the association needs of various entities such as knowledge points, courses, resources, and positions in teaching scenarios, achieving semantic data enhancement and significantly reducing the manual cost of knowledge graph construction. Detailed technical steps include:

[0121] 1) Database access and full metadata collection

[0122] Connect to the source database via JDBC / ODBC interface to collect complete metadata in batches, covering core information in the teaching scenario:

[0123] Table-level information: table name, business comments (such as "course schedule" or "knowledge point table"), number of data rows, and storage engine.

[0124] Field-level information: field name, data type, whether it is nullable, default value, comments (such as "parent ID of knowledge point" or "course difficulty"), character set.

[0125] Constraints and Indexes: Primary Key (including composite primary keys), Foreign Key (including source table, source field, target table, and target field), Unique Constraints, Index Types and Fields.

[0126] Data statistics: Sample 1000 rows of core data, analyze the distribution of field values ​​(such as the "easy / medium / difficult" distribution of the "difficulty" field), selectivity (proportion of different values), and identify enumeration types and high-frequency values.

[0127] 2) Deep metadata analysis and business semantic extraction

[0128] The collected metadata is organized and analyzed to uncover teaching-related business logic:

[0129] Clarify relationships between tables: identify "one-to-many" (e.g., course knowledge points), "many-to-many" (e.g., student courses), and "self-associations" (e.g., knowledge point parent knowledge point).

[0130] Business keyword extraction: Extract teaching-related keywords (such as "course", "knowledge point", "teacher", "position", "grade") from table names, field names, and comments to lay the groundwork for subsequent semantic mapping.

[0131] Data feature identification: label enumerated fields (such as "resource type" and "learning status"), related fields (such as "course_id" and "parent_knowledge_id"), and core business fields (such as "knowledge point name" and "course objective").

[0132] 3) Core decision-making process: Pattern analysis and mapping rule formulation

[0133] This is the core of the algorithm, which automatically determines the mapping method for "table entities," "field attributes," and "relationships" through a series of judgment rules tailored to teaching scenarios:

[0134] <1> Intelligent Entity Type Judgment

[0135] By using a multi-dimensional comprehensive scoring system, it is determined whether each table should be mapped to a "node," "relationship," or "attribute" in an attribute graph, thus avoiding redundancy caused by fixed rules.

[0136] Scoring dimensions (weights): whether it contains a primary key (0.3), number of times it is referenced by other tables (0.4), number of fields (0.1), whether it contains core teaching keywords (0.15), and percentage of data rows (0.05).

[0137] Decision-making rules:

[0138] Rating ≥ 0.6: Mapped to nodes (such as “Course Schedule”, “Knowledge Point Table”, “Job Table”, these tables have independent business meaning and rich fields).

[0139] For tables with a score < 0.6 and that are purely relational (containing only foreign key fields and no other business fields): map them to a relation (such as a pure course selection relational table for "Student Courses").

[0140] For scores < 0.6 and few fields with no independent business meaning (such as configuration tables or enumeration tables): map them to the attributes of other nodes (such as the "type" attribute of the "Teaching Resources" node being directly embedded in the "Resource Type Configuration Table").

[0141] <2> Semantic inference of relation types

[0142] To avoid generating meaningless "RELATED" relationships, generate concrete relationship names based on teaching business scenarios:

[0143] Basic Relationship Generation: Based on the foreign key naming pattern (e.g., "course_id" corresponds to "course", "parent_knowledge_id" corresponds to "parent knowledge point"), generate relationship names such as "contains", "prerequisites", and "belongs to".

[0144] Multiplicity of relationships: Distinguish between "one-to-one" (e.g., user personal configuration) and "one-to-many" (e.g., course knowledge points) relationships by checking if the foreign key contains a unique constraint; identify "many-to-many" relationships (e.g., student courses, linked through the course selection table) by checking if the foreign key contains a unique constraint.

[0145] Special relationship handling: Generate relationships such as "dependency" and "hierarchical containment" from self-related tables (such as "parent_id" in the knowledge point table); generate distinctive relationship names such as "creator ID" and "auditor ID" from multiple foreign keys in the same pair of tables (such as "creator ID" and "auditor ID" in the order table).

[0146] <3> Attribute mapping and index optimization

[0147] Field standardization: Standardize field name format (e.g., unify "create_time" and "created_at" to "createdAt"), and remove fields without business meaning (e.g., "deletion mark" and "redundant code").

[0148] Data type adaptation: Convert relational database data types to attribute graph compatible types (such as DATETIME to timestamp, JSON to MAP, enumeration encoding to semantic value).

[0149] Intelligent index recommendation: Primary key fields are automatically indexed with unique constraints; high-selectivity fields (such as "knowledge point name" and "user ID") are recommended to be indexed; low-selectivity fields (such as "gender" and "resource type") are not recommended to be indexed; if foreign key fields need to be retained as attributes, indexes are created synchronously.

[0150] Foreign key field handling: Foreign keys that have already been mapped to relationships will not be retained as node attributes again (to avoid data redundancy); if users need fast queries, they can be configured to retain them as attributes and create indexes.

[0151] <4> Adaptive processing of related tables

[0152] For many-to-many relationship tables commonly found in teaching scenarios (such as course selection tables and resource relationship tables), handle them flexibly according to the actual situation:

[0153] A purely relational table (containing only two foreign keys and no other fields) directly maps to the relationship between two nodes (such as the "Course Selection" relationship between "Student" and "Course"), without generating intermediate nodes.

[0154] Attributed relational tables (including foreign keys and additional business fields): mapped to relations with attributes (e.g., if the course selection table contains "score" and "semester", then the "course selection" relation will have the attributes score and semester).

[0155] Related tables with independent business significance (such as "order table" and "project task table") are still mapped as nodes and are associated with other nodes through relationships (such as the "association" relationship between the "order" node and the "student" and "course" nodes).

[0156] 4) Mapping rule implementation and script generation

[0157] The above decision results are then transformed into a directly executable Cypher script, containing complete property graph construction logic:

[0158] Node script: Defines node tags (e.g., course: knowledge point), core attributes, and unique constraints (e.g., CREATE CONSTRAINT FOR (c:course) REQUIRE c.id IS UNIQUE).

[0159] Relationship script: Defines the relationship type (such as containment, prerequisite), the type of associated node, and relationship attributes.

[0160] Indexing script: Generates indexing statements based on the recommendation results (e.g., CREATE INDEX FOR (k: knowledge point) ON (k.name)).

[0161] Data import script: Generates LOADCSV or JDBC import statements, supporting batch reading of relational database data and writing to attribute graphs.

[0162] 5) Batch data migration and conflict handling

[0163] Distributed parallel processing: Using frameworks such as Spark, data is read in parallel by partitioning tables to avoid performance bottlenecks caused by reading the entire table.

[0164] Conflict and exception handling:

[0165] Duplicate data: Remove duplicates by node unique ID, and retain the record with the most complete attributes for duplicate relationships.

[0166] Foreign key missing: If a knowledge point does not have a corresponding parent knowledge point ID, it is temporarily stored as an independent node and marked as "not associated". It will be automatically associated after the data is completed.

[0167] Data type incompatibility: Automatic truncation of excessively long strings, conversion of illegal format data (such as converting text-type dates to standard timestamps), and synchronous recording of exception logs for verification.

[0168] Batch import: Use Neo4j's Bolt protocol or the LOADCSV tool to write to the property graph database in batches, avoiding the inefficiency of inserting one row at a time.

[0169] 6) Confidence assessment and human-machine collaboration optimization

[0170] Rule confidence score: Each mapping rule is scored (0~1 points). The scoring criteria include: the clarity of foreign key constraints, the clarity of field semantics, the matching degree of data distribution with business, and whether it conforms to the preset teaching business rules.

[0171] Human-machine collaborative verification: Rules with a confidence level < 0.7 are marked as "pending review" and a visual report (including table structure, recommended mapping method, and points of contention) is generated for teachers or administrators to review; it supports adjusting mapping rules through a simple configuration interface (such as modifying relation names and adding attribute mappings) without refactoring the algorithm.

[0172] Rule iteration feedback: Manually adjusted rules will be synchronized to the algorithm knowledge base to optimize the accuracy of automatic mapping in subsequent similar scenarios.

[0173] 7) Subsequent adaptation and optimization

[0174] Index optimization: Dynamically adjust indexing strategies based on high-frequency queries in teaching scenarios (such as "knowledge point resource association" and "course job matching") to improve query response speed.

[0175] Scenario-based adaptation: To meet the teaching needs of multiple disciplines, it supports a custom business keyword library (such as adding entity keywords such as "experiment", "case", and "certificate") to adapt to the entity association characteristics of different disciplines.

[0176] The overall process of the intelligent attribute graph model mapping algorithm is as follows: Figure 2 As shown. This entire process requires no manual intervention: from database access to attribute graph output, core steps are completed automatically, with only low-confidence rules requiring manual review, significantly improving efficiency. It is well-suited to teaching scenarios: it focuses on adapting to the association needs of entities such as knowledge points, courses, resources, and job positions, supporting flexible handling of many-to-many and self-association tables. It is highly fault-tolerant: it has clear handling mechanisms for common problems such as missing data and type incompatibility, ensuring the integrity of data migration. It is iteratively optimizable: it supports manual adjustment of rules and feedback to the algorithm, adapting to the mapping needs of teaching data of different subjects and scales.

[0177] (III) Specific Implementation of the Personalized Learning Companion Module

[0178] Preferably, the personalized learning companion module includes:

[0179] The personalized learning path planning unit takes a student's personal dynamic knowledge graph as input, combines learning objectives and style preferences, and applies graph search algorithms (such as...). The system uses either a graph search algorithm or a reinforcement learning model to generate the optimal learning sequence from the current knowledge state to the target knowledge state on the course knowledge graph. This learning sequence respects the pre-requirement dependencies of knowledge points and dynamically adjusts based on the student's latest performance. Specifically, using the student's personal dynamic knowledge graph as input, combined with their learning goals and style preferences, the system employs graph search algorithms or reinforcement learning models to find the optimal learning sequence from the current state to the target state on the course knowledge graph. This path strictly respects the pre-requirement dependencies of knowledge points and dynamically avoids content already mastered. The learning path is not static; after each step the student completes, the system reassesses their knowledge state based on their latest performance and adjusts the subsequent path accordingly, achieving generative and adaptive recommendation through a "learn one step, push one step" approach.

[0180] The large-scale model coupling interaction unit is used to deeply couple the AIGC engine with the knowledge graph. When a student asks a question, the system analyzes the question and uses the knowledge graph to elicit prior knowledge, analogical concepts, or extended applications, generating a contextualized and logical answer. Simultaneously, it pushes practical resources related to the current knowledge point. Specifically, the intelligent learning companion integrating the AIGC engine utilizes the knowledge graph for semantic understanding and reasoning. When a student asks a question, the learning companion not only analyzes the question itself but also automatically evokes prior knowledge, analogical concepts, or extended applications through graph connections, providing a logical and systematic answer. During the Q&A or learning process, the system, based on the context of the current interaction (the knowledge points involved), pushes relevant micro-lecture videos, reference documents, typical code examples, etc., from the graph-connected resource network in real time, achieving "learn and use immediately" resource support.

[0181] In addition, the personalized learning module also supports in-depth learning analysis and social insights: it analyzes data such as the depth of student interaction with the system (e.g., the quality of questions and the breadth of resource exploration) and task persistence, and combines it with sentiment computing (e.g., analyzing attitude tendencies from text) to present the characteristics of students' learning engagement and thinking processes on the teacher's dashboard; by analyzing students' interaction and discussion data on knowledge graph nodes, it uses social network analysis technology to automatically discover learning groups with similar interests or complementary knowledge structures, promoting collaborative learning and experience sharing.

[0182] (iv) Specific implementation of the precision teaching module

[0183] Preferably, the precision teaching module includes:

[0184] The learning progress diagnosis and visualization unit maps students' behavioral data across various platforms to corresponding nodes in the knowledge graph in real time. Through knowledge tracking models and cluster analysis, it generates a heatmap of overall class knowledge mastery and an individual dynamic knowledge ability graph. Specifically, the system maps students' behavioral data (answering questions, watching videos, interacting) across platforms to corresponding nodes in the knowledge graph in real time. Through knowledge tracking models and cluster analysis, it dynamically generates two types of visualization dashboards: a heatmap of overall class knowledge mastery (intuitively displaying the average mastery level, common weaknesses, and performance distribution of the whole class across various knowledge points) and an individual dynamic knowledge ability graph (generating a personal knowledge graph for each student, clearly marking their mastery status of each knowledge point, and accurately pinpointing the root causes of their ability weaknesses).

[0185] The targeted recommendation unit automatically matches and recommends explanatory videos, case studies, and exercises from the resource library based on the semantic association network of the knowledge graph, according to the teacher's selected weak knowledge points. Teachers can select specific weak knowledge points based on the learning progress dashboard, and the system automatically and accurately matches and recommends diverse resources such as explanatory videos, in-depth case studies, and consolidation exercises from the resource library to achieve targeted breakthroughs. At the same time, the knowledge graph provides a logical framework for the hybrid process of "online self-exploration - offline in-depth explanation - after-class practice iteration". The system can recommend self-study knowledge sequences for students based on the predecessor and successor relationships of the graph, and the teacher's offline explanations target common problems revealed by the graph. After-class practice tasks are intelligently associated with application-oriented nodes in the graph.

[0186] The real-time early warning unit automatically sends a warning to the teacher during classroom interactions when the error rate for a particular knowledge point exceeds a preset threshold (e.g., 25%), prompting immediate intervention and explanation. Furthermore, it includes multi-dimensional process evaluation: moving beyond single scores, it constructs an evaluation model integrating a knowledge graph, recording the revision process of student assignments, the thought process behind discussions, and the logical path of project practice. Through comparative analysis with standard solution paths in the knowledge graph, it assesses the accuracy of technical application, the innovation of solutions, and the level of ability transfer, achieving a transformation from knowledge assessment to ability evaluation.

[0187] (v) Specific implementation of the practical application module

[0188] Preferably, the practical application module includes:

[0189] The Practical Resource Graph Unit is used to incorporate datasets, algorithm model libraries, lab manuals, project task books, and enterprise cases as practical resource entities into the knowledge graph. It establishes relationships between these entities and theoretical knowledge points ("explanation / application"), competency goals ("cultivation"), and job tasks ("correspondence"), forming a four-element network of "knowledge-ability-resource-job". Specifically, typical datasets (such as image sets), algorithm model libraries, software tools, lab manuals, project task books, and real-world enterprise cases required for the course are incorporated into the knowledge graph as entities. The graph not only defines the attributes of the practical resources themselves (such as difficulty, applicable scenarios, and technology stack), but more importantly, it establishes relationships between them and theoretical knowledge points ("explanation / application"), competency goals ("cultivation"), and job tasks ("correspondence"), forming a three-dimensional, traceable map of "learning-application transformation".

[0190] The practical task matching unit automatically recommends relevant practical operations, datasets, or simulation training based on knowledge graph associations after students complete theoretical knowledge learning. In project-based learning, it intelligently breaks down the required knowledge modules and skill units according to project objectives and recommends learning resource sequences. Specifically, after learning a theoretical knowledge point, the system can recommend or directly jump to relevant practical operations, accompanying datasets, or simulation training with a single click based on knowledge graph associations. When students are developing comprehensive projects, the system can analyze project objectives, intelligently break down the required knowledge modules and skill units based on the knowledge graph, and recommend corresponding learning resources, practical tools, and reference case sequences, providing intelligent scaffolding for complex practices.

[0191] The competency certification unit collects practical process data and compares it with preset competency standards in the knowledge graph. It automatically assesses the standardization of practical operations and problem-solving efficiency, generates competency growth reports, and awards digital badges linked to competency nodes. Specifically, when students operate in virtual experiments or programming environments, their code submission records, debugging steps, running results, operation logs, and other process data are captured by the system. The system compares this practical data with preset competency achievement standards and best practice paths in the knowledge graph, automatically assesses students' practical operation standardization, problem-solving efficiency, and innovation capabilities, and generates competency growth reports. Based on the assessment results, digital badges are automatically awarded to students who have reached specific competency levels. These badges are linked to competency nodes and practical evidence chains in the knowledge graph, providing credible competency credentials for employment.

[0192] (vi) Closed-loop optimization unit

[0193] Preferably, the system further includes a closed-loop optimization unit, used to feed back all group learning data generated by the precision teaching module, individual interaction data generated by the personalized learning companion module, and process and result data generated by the practical application module to the data center. This feedback data is then used to continuously optimize the weights of entity relationships in the knowledge graph, the accuracy of personalized recommendation strategies, and the quality of AIGC responses through machine learning algorithms. Specifically, this involves a three-dimensional data linkage feedback loop: all group learning data generated by precision teaching, individual interaction data generated by personalized learning companion, and process and result data generated by practical application are fed back to the data center. Using this feedback data, machine learning algorithms continuously optimize the weights of entity relationships in the knowledge graph, the accuracy of personalized recommendation strategies, and the response quality of intelligent learning companions, forming a self-evolving closed loop of "data-driven optimization, optimizing and improving services."

[0194] (vii) High concurrency and scalability guarantees

[0195] Preferably, the knowledge graph base layer uses a distributed graph database for storage, and the intelligent engine layer and 3D application service layer are deployed using a microservice architecture to support high-concurrency querying and rendering of large-scale graphs with tens of millions of nodes and relationships. Specifically, for large-scale graphs with tens of millions or even hundreds of millions of nodes and relationships, a distributed graph database or graph computing engine is used to optimize the performance of core operations such as querying and recommendation to support high-concurrency access. Each functional module is split into independent microservices, supporting elastic scaling to ensure the stability and smoothness of the system when a large number of users are online simultaneously. In the front-end visual interaction, batch processing optimization (such as setting the optimal batch size) and targeted performance tuning ensure the smoothness of large-scale graph rendering and operation.

[0196] (viii) Competency-oriented assessment model

[0197] Preferably, the teaching evaluation and optimization engine constructs an evaluation model that integrates knowledge graphs. By recording students' homework modification process, discussion thought process, and project practice logical path, it compares and analyzes these with standard solution paths in the knowledge graph to evaluate the accuracy of technology application, the innovation of solutions, and the level of ability transfer, thereby realizing the transformation from knowledge assessment to ability evaluation.

[0198] Based on the above embodiments, such as Figure 3 As shown, the present invention also provides a three-dimensional closed-loop intelligent teaching method based on the above system, comprising the following steps:

[0199] S1: Collect multi-source heterogeneous teaching data and resources through the data and resource layer, and carry out standardized governance;

[0200] S2: Automatically construct and dynamically update a domain knowledge graph covering course knowledge, competency objectives, job requirements, and practical resources through a knowledge graph base layer;

[0201] S3: Through the learning analysis engine, multimodal analysis of learning behavior is performed to generate learner profiles and knowledge status diagnoses;

[0202] S4: Through the precision teaching module, based on knowledge graphs and learning analysis, targeted delivery of teaching content and dynamic design of teaching processes are achieved;

[0203] S5: Through the personalized learning companion module, based on knowledge graphs and large models, a personalized learning path is dynamically generated for each student and real-time intelligent Q&A is provided.

[0204] S6: Through the practical application module, practical resources are organized based on knowledge graphs and the links between knowledge points, job tasks, and practical cases are established.

[0205] S7: The data generated by S4, S5, and S6 are fed back to the data center, and the knowledge graph and recommendation strategy are optimized through the teaching evaluation and optimization engine to form a three-dimensional closed-loop iteration.

[0206] Specifically, S1 includes:

[0207] (1) Unified data standards and governance: Establish school-level data standards and follow the principle of "one source of data" to thoroughly investigate and govern data from different platforms and in different formats, such as student affairs, academic affairs, campus card, and learning behavior.

[0208] (2) Distributed ETL processing: A distributed ETL pipeline is built using computing frameworks such as Spark and Storm. Through rule engines and data cleaning algorithms (such as clustering, deduplication, and missing value imputation), multi-source data is extracted, transformed, cleaned and integrated to output high-quality and standardized data assets.

[0209] Specifically, S2 includes:

[0210] For relational teaching resource databases, we developed an intelligent attribute graph model mapping algorithm to automatically analyze the table structure, primary and foreign keys, and other metadata of the source database, and generate mapping rules from the relational model to the graph model (Neo4j) to achieve batch, automated semantic upgrade of massive resource data.

[0211] For unstructured resources such as video subtitles and document text, a sequence labeling model combining BERT pre-trained model and LSTM-CRF is adopted, and a bidirectional maximum matching algorithm is introduced to enhance boundary recognition using an external dictionary. This accurately extracts knowledge point entities from the teaching text, and then performs preliminary screening of candidate entities through fuzzy matching using edit distance. Semantic matching degree is calculated using cosine similarity, and entities mentioned in the resource text are automatically and accurately linked to corresponding nodes in the knowledge point base, establishing a fine-grained association between "resources and knowledge points".

[0212] The system continuously collects students' learning trajectories, interaction data, and assessment results. Using multimodal analysis tools, it dynamically infers the mastery of knowledge points and the effectiveness of resource associations, automatically adjusting the correlation coefficients and weights between knowledge points in the knowledge graph. It establishes connections with external data sources such as industry databases, technical forums, open-source code repositories, and policy and regulatory databases, automatically capturing new knowledge, new cases, and new technology standards using information extraction and change detection technologies, triggering incremental updates to the knowledge graph. When new concepts not included in the knowledge base appear in teaching resources, it identifies potential new entities through semantic analysis, assesses confidence levels, and prompts the administrator for review, adding them as new nodes to the knowledge graph.

[0213] Specifically, S3 includes:

[0214] Based on machine learning and data mining, multimodal analysis of learning behavior is performed to generate dynamic learner profiles and knowledge status diagnoses.

[0215] (1) Learning data fusion analysis: The system maps students' behavioral data (answering questions, watching, interacting) on ​​various platforms to the corresponding nodes of the knowledge graph in real time;

[0216] (2) Class and Individual Knowledge Profiles: Two types of visual dashboards are dynamically generated through knowledge tracing models and cluster analysis:

[0217] Class-wide knowledge mastery heatmap: Visually displays the class's average mastery of each knowledge point, common weaknesses, and performance distribution;

[0218] Individual dynamic knowledge and ability map: Generate a personal knowledge map for each student, clearly marking their mastery status of each knowledge point (proficient, weak, not learned), and revealing the mastery relationship chain between knowledge points to accurately locate the root cause of ability shortcomings.

[0219] Specifically, S4 includes:

[0220] Based on the graph's correlations and real-time learning progress, the system enables targeted delivery of teaching content and dynamic design of blended learning processes. Teachers can select specific weak knowledge points from the learning progress dashboard, and the system automatically matches and pushes relevant resources from the resource library. The system can also recommend self-study knowledge sequences for students based on the graph's predecessor and successor relationships. Offline teacher explanations address common problems revealed by the graph. After-class practice tasks are intelligently linked to application-oriented nodes in the graph. During classroom interaction, when the error rate of a knowledge point exceeds a preset threshold (e.g., 25%), a real-time alert is automatically sent to the teacher, prompting immediate intervention and explanation.

[0221] Specifically, S5 includes:

[0222] Path generation algorithm: Taking the student's personal dynamic knowledge graph as input, combined with their learning goals and style preferences, and using graph search algorithms (such as...) Alternatively, a reinforcement learning model can be used to find the optimal learning sequence from the current state to the target state on the course knowledge graph, strictly respecting the prerequisite dependencies of knowledge points and dynamically avoiding already mastered content.

[0223] Generative recommendation and adjustment: After each step a student completes, the system reassesses their knowledge status based on their latest performance and adjusts the subsequent path in real time, achieving generative and adaptive recommendation.

[0224] Contextualized knowledge interaction: The intelligent learning companion, which integrates the AIGC engine, uses knowledge graphs for semantic understanding and reasoning. When students ask questions, the learning companion not only analyzes the question itself, but also automatically derives prior knowledge, analogical concepts, or extended applications through graph associations, providing logical and systematic answers.

[0225] Resource-accompanied recommendation: During Q&A or learning, the system pushes relevant micro-lecture videos, reference documents, typical code examples, etc. in real time from the resource network associated with the graph based on the current interaction context (the knowledge points involved), realizing the resource accompaniment that can be used immediately after learning.

[0226] Specifically, S6 includes:

[0227] Multimodal resource integration: Typical datasets (such as image sets), algorithm model libraries, software tools, experimental manuals, project task books, real-world enterprise cases, and other practical resources required for the course are incorporated into the knowledge graph as entities.

[0228] Establish a network linking "knowledge-ability-resources-jobs": In the graph, define the attributes of practical resources themselves (such as difficulty, applicable scenarios, and technology stack), and establish their relationship with theoretical knowledge points ("explanation / application"), ability goals ("cultivation"), and job tasks in the enterprise. This forms a three-dimensional, traceable map of learning and application transformation.

[0229] Access to practical resources on demand: After students learn a certain theoretical knowledge point, the system can recommend or directly jump to related practical operations, supporting datasets or simulation training with one click based on the map association.

[0230] Intelligent scaffolding for project-based learning: When students develop comprehensive projects, the system can analyze the project objectives, intelligently break down the required knowledge modules and ability units based on the graph, and recommend corresponding learning resources, practical tools and reference case sequences.

[0231] The Practical Results Feedback and Competency Certification System captures students' code submission records, debugging steps, running results, operation logs, and other process data in virtual experiments or programming environments. It compares these data with the preset competency achievement standards and best practice paths in the graph, automatically evaluates the standardization of practical operations, problem-solving efficiency, and innovation capabilities, generates competency growth reports, and issues digital badges linked to competency nodes.

[0232] Specifically, S7 includes:

[0233] Three-dimensional data linkage feedback: The group learning data generated by precision teaching, the individual interaction data generated by personalized learning, and the process and result data generated by practical application are all fed back to the data center.

[0234] Continuous optimization of knowledge graphs and algorithms: Utilizing this feedback data, machine learning algorithms are used to continuously optimize the weights of entity relationships in the knowledge graph, the accuracy of personalized recommendation strategies, and the response quality of intelligent learning companions, forming a self-evolving closed loop of "data-driven optimization, and service improvement through optimization".

[0235] To enable those skilled in the art to fully understand and implement the intelligent teaching and three-dimensional closed-loop system based on knowledge graphs and large models proposed in this invention, the specific implementation method will be described below in conjunction with the system architecture and core workflow. After system deployment, its core operation is manifested in a collaborative workflow across three dimensions: "precise teaching," "personalized learning support," and "practical application." Each workflow uses a knowledge graph as a unified foundation and achieves closed-loop operation through specific module calls and data flow.

[0236] I. System Deployment and Initialization

[0237] In practical implementation, the first step is to complete the deployment of the system's hardware and software environment and data initialization to lay the foundation for three-dimensional closed-loop operation.

[0238] 1. Environment Deployment

[0239] (1) Server-side: Deployed using a microservice architecture based on Spring Boot. Neo4j is selected as the graph database service for storing and querying the core knowledge graph; MySQL is selected as the relational database for storing user information, system logs, resource metadata, etc.

[0240] (2) Data processing layer: Deploy the Apache Spark computing framework to perform distributed cleaning, transformation and fusion computing on batch learning behavior data from multiple platforms (such as Chaoxing, Yu Classroom and the One-Card System).

[0241] (3) AI model services: Deploy entity recognition and linking models (based on BERT pre-trained models and LSTM-CRF sequence labeling models), recommendation algorithm engines, AIGC services (such as integrating large language model APIs), and provide services to the outside world in the form of RESTful API or gRPC interface.

[0242] (4) Front-end application: The management backend and student / teacher portal are developed using the Vue.js framework combined with the ElementUI component library. The core of the knowledge graph visualization uses the D3.js library to implement force-directed graph rendering.

[0243] 2. Knowledge Graph Initialization and Construction

[0244] (1) Structured Data Import: For existing structured data such as course outlines, textbook catalogs, and exercise banks, an ETL (Extract-Transform-Load) process is implemented. The source data table structure is analyzed, mapping rules are formulated, and relational data is imported into the Neo4j graph database in batches according to the attribute graph model, forming initial knowledge point nodes and basic relationships such as "belongs to" and "prerequisites". Specifically, the aforementioned intelligent attribute graph model mapping algorithm is adopted, including a complete set of steps such as metadata collection, deep analysis, pattern analysis, mapping rule formulation, script generation, data migration, and conflict handling.

[0245] (2) Unstructured Resource Association: An automated association pipeline is initiated for unstructured resources such as course videos, documents, and case studies. First, the deployed NLP model is used to perform entity recognition on the resource text (such as video subtitles) to extract knowledge point mentions; then, fuzzy matching is performed using edit distance (Levenshtein Distance) to generate a candidate entity list; finally, semantic disambiguation is performed by calculating cosine similarity to link resources to the most relevant knowledge point entities, thereby automatically establishing the "teaching resource-knowledge point" association edge.

[0246] II. Core Workflow 1: Precision Teaching Implementation Process

[0247] This workflow primarily serves teachers, enabling data-driven, targeted teaching and competency-based assessment. The specific implementation steps are as follows:

[0248] 1. Real-time access and fusion of multi-source learning data

[0249] The system continuously collects data from multiple heterogeneous data sources, including online learning platforms, classroom interaction tools, and academic affairs systems, through pre-built API interfaces. It then uses deployed Spark jobs to clean (deduplicate and handle missing values) and standardize (unify time format and grade conversion standards) the data, performs semantic alignment according to school-level data standards, and stores the data in a data lake.

[0250] 2. Generation of student learning profiles and knowledge heatmaps

[0251] (1) Based on the fused data, the system uses cluster analysis (such as K-means) and association rule mining to dynamically generate learner profiles for the class as a whole and for each student, and labels their learning type (such as "system consolidation type" and "curiosity exploration type").

[0252] (2) At the same time, the system calls the Knowledge Tracking model to analyze students' historical answer records and interaction time on each knowledge point, and dynamically diagnoses their mastery level (e.g., proficient, weak, not learned). The results are rendered into a class knowledge mastery heat map and individual dynamic knowledge graphs by D3.js, and presented intuitively on the teacher's dashboard.

[0253] 3. Targeted teaching content delivery and blended workflow execution

[0254] (1) Teachers locate common weak knowledge points in the class based on the heat map (such as "backpropagation of convolutional neural networks"). The system automatically recommends resource packages such as micro-lecture videos, classic cases, and special exercises that are strongly related to the knowledge point based on the correlation of the knowledge graph.

[0255] (2) Teachers design and implement a blended learning process: Before class, recommended resources are pushed to students through the system for independent exploration; during class, weak points are explained in detail and real-time feedback is collected using classroom interaction tools; after class, a comprehensive practical task based on the knowledge point is released, requiring students to complete iteratively.

[0256] 4. Dynamic evaluation of competency-based teaching effectiveness

[0257] (1) The system automatically evaluates homework and practical projects. For objective questions, the system directly compares the answers; for subjective questions and coding assignments, the system calls the AIGC service for semantic understanding and intelligent grading, and provides error analysis and improvement suggestions.

[0258] (2) The assessment not only focuses on the final answer, but also records the process data such as the time spent solving problems, the revision process, and the behavior of seeking help. Combining the relationship between knowledge points and ability objectives in the knowledge graph, the system generates a multi-dimensional assessment report to evaluate the students' ability growth from "knowledge understanding" to "technology application" and even "innovation transfer", rather than a single score.

[0259] III. Core Workflow Two: Personalized Learning Companion Implementation Process

[0260] This workflow primarily serves students, providing personalized adaptive learning support. The specific implementation steps are as follows:

[0261] 1. Dynamic Programming of Personalized Learning Paths

[0262] After a student logs in, the system first assesses their latest knowledge status profile (from knowledge tracking) and preset learning objectives. On the course knowledge graph, starting from the current state and ending at the target knowledge point, it uses graph search algorithms (such as...) An algorithm or reinforcement learning model is used to calculate an optimal learning path in real time. This path strictly adheres to the priority relationships of knowledge points and takes into account cognitive load and individual learning style preferences.

[0263] 2. AIGC-driven intelligent Q&A and feedback

[0264] (1) When students encounter questions in the learning path, they can ask AI teaching assistants. The system combines the context of the question with the current learning knowledge point, calls the AIGC service to generate explanatory answers that match the student's cognitive level, and can recommend prior knowledge or analogous cases based on the knowledge graph to achieve in-depth Q&A.

[0265] (2) After students complete the exercises, the system provides immediate analysis. For incorrect questions, it not only provides the correct answer, but also analyzes the student's answering process to pinpoint the specific missing or misunderstood nodes in the knowledge graph and pushes targeted remedial learning resources.

[0266] 3. Learning Community Discovery and Incentives

[0267] (1) The system continuously analyzes the learning behavior sequence of all students, uses the semantic similarity calculation of the knowledge graph to find learners with similar knowledge gaps or interest preferences, and automatically forms or recommends online learning groups.

[0268] (2) Teachers can view the learning situation big data analysis dashboard, which integrates multi-dimensional indicators such as the overall progress of the class, individual engagement, and community activity, making it easier for teachers to provide macro-level supervision and implement targeted incentives.

[0269] IV. Core Workflow 3: Practical Application Implementation Process

[0270] This workflow focuses on streamlining the "learning-to-application" process to facilitate the transfer of knowledge to skills. The specific implementation steps are as follows:

[0271] 1. The graphical construction of the course practice resource database

[0272] Based on job requirements, we collect and develop practical resources such as typical datasets, algorithm model libraries, project task books, and industry case studies. Using a knowledge graph as a framework, we treat these resources as new node types and semantically associate them with theoretical knowledge points. For example, we associate the knowledge point of "image enhancement algorithm" with "foggy driving image dataset" and "CLAHE enhancement model code library".

[0273] 2. Application of the "Knowledge-Skills-Resources-Job" Four-Element Relationship

[0274] (1) When a student learns a certain theoretical knowledge point, the system can display all the practical resources associated with it with one click, as well as the ability goals trained by these resources (such as "image preprocessing ability"), and further associate them with potential positions that require this ability (such as "computer vision algorithm engineer").

[0275] (2) After a student completes a practical project (such as “Traffic Sign Detection Based on YOLO”), their project results, code quality, and performance in solving practical problems will be evaluated and converted into digital badges or competency labels, which will be permanently recorded in their personal learning map to form a verifiable competency profile.

[0276] 3. Closed-loop data feedback and system iteration

[0277] (1) The massive amounts of process and outcome data generated by the above three workflows (such as resource utilization effect, path completion rate, and project achievement) are continuously fed back to the system.

[0278] (2) The system uses this data to dynamically optimize the knowledge graph: for example, automatically adjust the correlation coefficient between knowledge points; discover and add new knowledge point associations; and eliminate resources with poor performance.

[0279] At the same time, this data is also used to train and optimize recommendation algorithms, knowledge tracking models, and the response quality of AIGC teaching assistants, thereby achieving adaptive evolution and continuous iteration of the entire "teaching-learning-application" three-dimensional closed-loop system.

[0280] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several improvements and modifications without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A three-dimensional closed-loop intelligent teaching system based on knowledge graphs and large models, characterized in that, include: The data and resource layer is used to collect and manage multi-source heterogeneous teaching data and resources, including structured and unstructured data. The knowledge graph base layer, connected to the data and resource layer, is used to automatically construct and dynamically update a domain knowledge graph covering course knowledge, ability objectives, job requirements, and practical resources. The knowledge graph is stored in an attribute graph model and includes knowledge point entities, ability entities, job entities, practical resource entities, and their semantic relationships. The intelligent engine layer includes: The learning analysis engine is used to perform multimodal analysis of learning behavior and generate dynamic learner profiles and knowledge status diagnoses. A personalized recommendation engine for path planning and resource matching based on knowledge graph semantic recommendations; AIGC and intelligent interaction engine are used to provide intelligent Q&A, content generation and natural language interaction services; The teaching evaluation and optimization engine is used to analyze teaching effectiveness, trigger early warnings, and optimize graph relationships and recommendation strategies based on feedback data. The 3D application service layer includes: The precision teaching module is used to achieve targeted delivery of teaching content and dynamic design of teaching processes based on knowledge graphs and learning analysis. The personalized learning companion module is used to dynamically generate personalized learning paths for each student based on knowledge graphs and large models, and provide real-time intelligent Q&A. The practical application module is used to organize practical resources based on knowledge graphs and to connect knowledge points with job tasks and practical cases. The interactive presentation layer is used to provide teachers, students, and administrators with a multi-terminal visual interactive interface; The precision teaching module, personalized learning companion module, and practical application module achieve data interconnection and business linkage through the knowledge graph base layer, forming a three-dimensional closed loop of "teaching-learning companion-application".

2. The intelligent teaching three-dimensional closed-loop system based on knowledge graphs and large models according to claim 1, characterized in that, The knowledge graph base layer includes: The structured data conversion unit is used to automatically map relational teaching resource databases to attribute graph models. It generates mapping rules from relational models to graph models by analyzing the table structure and primary and foreign key metadata of the source database, thereby achieving batch automated semantic conversion. The unstructured resource association unit is used to identify and link knowledge point entities in teaching texts. It uses a BERT pre-trained model combined with an LSTM-CRF sequence labeling model for entity recognition, and uses edit distance for fuzzy matching and cosine similarity for semantic disambiguation to link resources to corresponding knowledge point nodes in the knowledge graph. The dynamic update unit is used to automatically adjust the correlation coefficients and weights between nodes in the knowledge graph based on learning behavior data, and to capture new knowledge and new cases from external data sources to trigger incremental updates.

3. The intelligent teaching three-dimensional closed-loop system based on knowledge graphs and large models according to claim 2, characterized in that, The structured data conversion unit executes an intelligent attribute graph model mapping algorithm, including: Collect metadata from the source database, including table-level information, field-level information, constraints and indexes, and data statistics. Each table is scored across multiple dimensions, and the table is mapped to a node, relationship, or attribute based on the score and business characteristics. Based on foreign key naming patterns and characteristics of associated tables, semantically infer relation types, including "contains", "prerequisite", "belongs to", "depends on", and "hierarchical containment"; Standardize fields and adapt them to data types, and generate index recommendations; The mapping decision results are converted into Cypher scripts, which are used to build nodes, relationships, and indexes in the Neo4j graph database.

4. The intelligent teaching three-dimensional closed-loop system based on knowledge graphs and large models according to claim 1, characterized in that, The personalized learning companion module includes: The personalized learning path planning unit is used to generate the optimal learning sequence from the current knowledge state to the target knowledge state on the course knowledge graph by taking the student's personal dynamic knowledge graph as input, combining learning objectives and style preferences, and using graph search algorithms or reinforcement learning models. The learning sequence respects the prerequisite dependencies of knowledge points and is dynamically adjusted according to the student's latest performance. The large model coupling interaction unit is used to deeply couple the AIGC engine with the knowledge graph. When a student asks a question, the question is parsed and the knowledge graph is used to evoke prior knowledge, analogous concepts, or extended applications, generating a contextualized and logical answer, and simultaneously pushing practical resources related to the current knowledge point.

5. The intelligent teaching three-dimensional closed-loop system based on knowledge graphs and large models according to claim 1, characterized in that, The precision teaching module includes: The learning progress diagnosis and visualization unit is used to map students' behavioral data on various platforms to the corresponding nodes of the knowledge graph in real time. Through knowledge tracking models and cluster analysis, it generates a heat map of the class's overall knowledge mastery and a dynamic knowledge and ability graph of individuals. The targeted push unit is used to automatically match and push explanatory videos, case studies, and exercise resources from the resource library based on the teacher's selected weak knowledge points and relying on the semantic association network of the knowledge graph; The real-time warning unit is used to automatically send a real-time warning to the teacher when the error rate of answering a certain knowledge point exceeds a preset threshold during classroom interaction.

6. The intelligent teaching three-dimensional closed-loop system based on knowledge graphs and large models according to claim 1, characterized in that, The practical application module includes: The practical resource graphing unit is used to incorporate datasets, algorithm model libraries, experimental manuals, project task books, and enterprise cases as practical resource entities into the knowledge graph, and establish the "explanation / application" relationship between entities and theoretical knowledge points, the "cultivation" relationship with ability goals, and the "correspondence" relationship with job tasks, forming a four-element association network of "knowledge-ability-resource-job". The practical task matching unit is used to automatically recommend relevant practical operations, datasets or simulation training based on knowledge graph associations after students have completed the learning of theoretical knowledge points. It is also used in project-based learning to intelligently decompose the required knowledge modules and ability units according to project objectives and recommend learning resource sequences. The competency certification unit is used to collect data from the practical process and compare it with the preset competency standards in the knowledge graph. It automatically evaluates the standardization of practical operations and the efficiency of problem solving, generates a competency growth report, and issues digital badges that are bound to competency nodes.

7. The intelligent teaching three-dimensional closed-loop system based on knowledge graphs and large models according to claim 1, characterized in that, It also includes a closed-loop optimization unit, which is used to feed back all the group learning data generated by the precision teaching module, the individual interaction data generated by the personalized learning module, and the process and result data generated by the practical application module to the data center, and use the feedback data to continuously optimize the weight of entity relationships in the knowledge graph, the accuracy of personalized recommendation strategies, and the quality of AIGC responses through machine learning algorithms.

8. The intelligent teaching three-dimensional closed-loop system based on knowledge graphs and large models according to claim 1, characterized in that, The knowledge graph base layer uses a distributed graph database for storage, while the intelligent engine layer and the 3D application service layer are deployed using a microservice architecture to support high-concurrency querying and rendering of large-scale graphs with tens of millions of nodes and relationships.

9. A three-dimensional closed-loop intelligent teaching system based on knowledge graphs and large models according to claim 1, characterized in that, The teaching evaluation and optimization engine constructs an evaluation model that integrates knowledge graphs. By recording students' homework modification process, discussion thought process, and project practice logical path, it compares and analyzes these with standard solution paths in the knowledge graph to evaluate the accuracy of technology application, the innovation of solutions, and the level of ability transfer, thus realizing the transformation from knowledge assessment to ability evaluation.

10. A three-dimensional closed-loop intelligent teaching method based on the system described in any one of claims 1 to 9, characterized in that, Includes the following steps: Step 1: Collect multi-source heterogeneous teaching data and resources through the data and resource layer, and perform standardized management; Step 2: Automatically construct and dynamically update a domain knowledge graph covering course knowledge, competency objectives, job requirements, and practical resources through a knowledge graph base layer; Step 3: Use the learning analysis engine to perform multimodal analysis of learning behavior and generate learner profiles and knowledge status diagnoses; Step 4: Through the precision teaching module, based on knowledge graphs and learning analysis, targeted delivery of teaching content and dynamic design of teaching processes are achieved; Step 5: Through the personalized learning companion module, a personalized learning path is dynamically generated for each student based on knowledge graphs and large models, and real-time intelligent Q&A is provided. Step 6: Through the practical application module, organize practical resources based on the knowledge graph and establish the connection links between knowledge points, job tasks, and practical cases; Step 7: Feed the data generated in steps 4, 5, and 6 back to the data center, and optimize the knowledge graph and recommendation strategy through the teaching evaluation and optimization engine to form a three-dimensional closed-loop iteration.

Images