Method and system for constructing a hybrid course knowledge graph based on a large language model

Abstract

The application relates to the technical field of knowledge graphs, and discloses a hybrid course knowledge graph construction method and system based on a large language model. First, multi-source heterogeneous teaching data such as teaching material resources and teaching generation resources are acquired, three types of core entities, namely knowledge points, resources and concepts, are extracted through natural language processing technology, and a mapping relationship between the entities is established. The extraction effect of the concept entities and the relationship is quantitatively evaluated by using preset evaluation indexes, the extraction process is dynamically optimized according to the result, and the association between the resource entities and the knowledge point entities is accurately identified by combining a large model based on concept attention and context learning to generate the association between the resources and the knowledge points. Finally, a knowledge graph is constructed by fusing all the entities and their relationships, and the dynamic updating of the graph is realized by generating new resources, knowledge points and adjusting the relationship based on the behaviors of teachers and students, so that the organization efficiency and the retrieval precision of the hybrid course resources are significantly improved, and personalized learning services can be effectively supported.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph technology, specifically to a method and system for constructing a hybrid course knowledge graph based on a large language model. Background Technology

[0002] Currently, blended learning has widely integrated various pre-set and dynamically generated teaching resources, including digital textbooks, online and offline teaching materials, course-related MOOCs, and knowledge communities, forming a multi-source heterogeneous data system encompassing textbook chapters, case studies, blog posts, and forum threads. This complex and diverse data structure significantly reduces the efficiency of information organization and precise retrieval, greatly hindering the optimization and upgrading of teaching and the deep implementation of personalized services.

[0003] Knowledge graphs, as an efficient resource organization tool, can provide strong support for the intelligent integration and structured management of learning resources in online teaching platforms. However, traditional knowledge graph construction methods have significant limitations: on the one hand, they often rely on experts to manually define ontologies or rules for information extraction, which is not only costly and inefficient, but also difficult to adapt to the core needs of blended courses, such as dynamic updates of teaching resources and deep integration of multi-source heterogeneous resources; on the other hand, they often ignore the impact of dynamic factors such as teaching-generated content and resource access behavior on the knowledge graph during the course teaching process, resulting in a knowledge graph that fails to truly reflect the essential characteristics of blended course knowledge and its structure constantly adapting and dynamically evolving with educational practice. Summary of the Invention

[0004] In view of this, to address the problems of traditional knowledge graph construction methods relying on expert definitions and neglecting dynamic factors in the teaching process, making it difficult to adapt to the dynamic updates of multi-source heterogeneous resources in blended courses and the evolution of knowledge structures with teaching practice, resulting in inefficient information organization and retrieval, and hindering teaching optimization and personalized services, this invention provides a method and system for constructing a blended course knowledge graph based on a large language model. This method enables the structured representation of knowledge points, heterogeneous resources, and generated resources involved in blended course teaching scenarios, and creatively develops an intelligent integration and structured management method for cross-platform heterogeneous resources with adaptability and scalability.

[0005] In a first aspect, this invention provides a method for constructing a hybrid course knowledge graph based on a large language model, including:

[0006] Acquire multi-source heterogeneous resource data in a blended learning environment, including textbook resources and teaching-generated resources;

[0007] Three types of entities are extracted from multi-source heterogeneous resource data using a natural language processing model: knowledge point entities, resource entities, and concept entities. Among them, knowledge point entities are extracted from the text data of textbook resources, and a mapping relationship between knowledge point entities and textbook resource entities is established.

[0008] Based on a pre-defined extraction model, conceptual entities and relationships between entities are extracted from multi-source heterogeneous resource data. The extraction effect is evaluated using pre-defined extraction effect evaluation indicators based on conceptual entities and relationships, and the extraction process is adjusted according to the evaluation results.

[0009] Based on a large model of concept attention and context learning, the relationship between resource entities and knowledge point entities is identified.

[0010] The constructed knowledge graph is updated, including generating new resources and knowledge points to update the graph, and adjusting the relationships between resource entities based on teachers' and students' recommendations and access behaviors.

[0011] This invention provides a method for constructing a knowledge graph for blended learning courses based on a large language model. By acquiring multi-source heterogeneous resource data and utilizing a natural language processing model to extract three types of entities—knowledge points, resources, and concepts—it achieves intelligent integration of blended learning course teaching resources. This method optimizes the extraction effect of entities and relationships through preset extraction models and evaluation indicators. Based on a large model of concept attention and contextual learning, it accurately identifies the relationships between generated resources and knowledge points, thereby constructing a high-quality knowledge graph. Furthermore, through a dynamic update mechanism (including generating new resources and adjusting relationships based on teacher-student behavior), the knowledge graph possesses adaptive evolution capabilities, continuously reflecting the evolution of knowledge in teaching practice. Compared to traditional methods relying on expert rules, this invention significantly reduces construction costs, enhances the dynamic adaptability of the knowledge graph, provides more precise resource organization and retrieval capabilities for blended learning, and effectively supports personalized learning services.

[0012] In one optional implementation, the preset concept entity and relation extraction performance evaluation indicators include: a concept entity extraction stability indicator, a concept entity extraction accuracy indicator, a concept entity relationship stability indicator, and a concept entity relationship accuracy indicator; wherein:

[0013] The concept entity extraction stability index is used to evaluate the stability of core concept extraction during multiple extractions of a large model. Its calculation incorporates weighting factors, sliding window mechanism, multi-dimensional concept extraction, nonlinear transformation, and external influencing factors.

[0014] The concept entity extraction accuracy index is used to evaluate the accuracy of large models in extracting concepts. It is calculated using Jaccard similarity and graph-based semantic perception similarity, and the two similarities are weighted and integrated.

[0015] The concept entity relationship stability index is used to evaluate the stability of concept relationship extraction during multiple extractions in a large model. Its calculation is based on the total number of times the large-scale language model performs multiple rounds of relationship extraction, the set of relationship knowledge extracted each time, and the global relationship space being the set of all observed relationship identifiers. It uses a multimodal embedding model to encode relationship values ​​and spatiotemporal features, combined with a temporal decay coefficient to calculate similarity. An adaptive concept relationship stability score is calculated for each relationship identifier, and information entropy is introduced to correct the relationship-level importance weight. The concept entity relationship stability index is obtained by combining the adaptive concept relationship stability score and the relationship-level importance weight using a dual-channel weighting method.

[0016] The accuracy index of relationships between concept entities is used to evaluate the accuracy of large models in extracting relationships between concepts. Its calculation employs a dynamic stability index, combined with a time decay factor and confidence level. In this embodiment, the concept entity extraction stability index addresses the issue of fluctuations in extraction results caused by changes in data distribution in traditional methods by introducing multi-dimensional optimizations such as weighting factors and sliding window mechanisms. The accuracy index combines Jaccard similarity and semantic-aware similarity to improve the semantic rationality of concept extraction. The relationship stability index uses multimodal embedding and spatiotemporal feature encoding to enhance the robustness of relationships in dynamic teaching scenarios. The relationship accuracy index ensures the timeliness of relationship extraction through the dynamic stability index (DSI) and time decay factor. The comprehensive application of these indicators makes the knowledge graph construction process quantifiable and optimizable, significantly improving the reliability of automated extraction and laying a high-quality data foundation for subsequent graph applications (such as intelligent recommendation and resource retrieval).

[0017] In one optional implementation, the step of evaluating the extraction effect using preset concept entity and relation extraction effect evaluation indicators, and adjusting the extraction process based on the evaluation results, includes:

[0018] The concept entity extraction stability index and the concept entity extraction accuracy index are weighted and summed to obtain the comprehensive evaluation index of concept entity extraction.

[0019] The comprehensive index for evaluating concept relationship extraction is obtained by weighted summation of the stability index and the accuracy index of the relationship between concept entities.

[0020] If at least one of the comprehensive evaluation index for concept entity extraction and the comprehensive evaluation index for concept relationship extraction is lower than the corresponding preset experience threshold, the extraction model will be optimized by adjusting the prompts or providing supplementary examples.

[0021] This invention addresses the problem of the limitations of single-indicator evaluation by weighted integration of stability and accuracy metrics to generate a comprehensive evaluation index. The extraction model is then dynamically optimized based on thresholds, resolving the issue of bias in single-indicator evaluations. For example, focusing solely on accuracy may overlook the model's stability in long-term applications. By judging the extraction performance through the comprehensive index and adjusting prompts or supplementary examples when standards are not met, iterative optimization of the model is achieved. This feedback mechanism not only improves the automation level of knowledge graph construction but also reduces the need for manual intervention, making it particularly suitable for scenarios with dynamically updated resources in blended learning courses. Furthermore, this method provides a transferable evaluation framework for other knowledge extraction tasks based on large models, demonstrating broad applicability.

[0022] In one optional implementation, the large model based on concept attention and context learning identifies the association between resource entities and knowledge point entities, including:

[0023] By having experts divide the knowledge points in the textbooks, and using GraphRAG to extract text summaries and conceptual entities from textbook resources and teaching-generated resources, samples of textbook resources and teaching-generated resources are generated.

[0024] The dynamic similarity function is used to calculate the dynamic similarity between the teaching-generated resource samples and each sample in the set of textbook resources. Dynamic Top-K sampling is used to select positive and negative samples based on the dynamic similarity value at the current time to train the contrastive learning model.

[0025] By using a trained contrastive learning model, the similarity between the generated resource sample and each sample in the textbook resource set is calculated, and the most similar textbook resource demonstration samples are selected as auxiliary corpus for context learning.

[0026] Using the contextual learning strategy of concept attention, the conceptual entities, inter-conceptual relationships, and summaries extracted from the textbook resources, as well as the textbook resource demonstration samples output by the contrastive learning model, are used as auxiliary corpus, and the conceptual entities, inter-conceptual relationships, and summaries extracted from the teaching-generated resources are used as corpus to be classified.

[0027] Following the template of task prompts, auxiliary corpus, corpus to be classified, and output format, the prompts are constructed as input to the preset AI large model. The large model is required to output knowledge points corresponding to the resources, so as to realize the association between the generated resource entities and knowledge point entities.

[0028] This invention lays the foundation by dividing knowledge points by experts and extracting key information using GraphRAG. It combines dynamic similarity calculation and dynamic sampling strategies to train a contrastive learning model, accurately retrieves similar teaching materials as auxiliary corpus, and then uses a contextual learning strategy based on concept attention to construct a structured prompt input model. This effectively integrates domain knowledge from blended learning courses, improves the performance of the large model with few samples, and achieves accurate and dynamic identification of the relationship between generated resources and knowledge point entities. As a result, it efficiently integrates multi-source heterogeneous teaching resources, providing strong support for teaching optimization and personalized services.

[0029] In one optional implementation, the generation of new resource and knowledge point update graphs and the adjustment of relationships between resource entities based on teacher and student recommendations and access behaviors include:

[0030] The constructed knowledge graph is used to discover communities using a community discovery algorithm. The community discovery adopts a maximization objective function, which considers the spatiotemporal coupling coefficient and the time decay function, and the adaptive balance parameter is dynamically adjusted through deep learning training.

[0031] The community division results of the knowledge graph are obtained. For each community, the knowledge points, teaching materials, generated resources and concept construction prompts are input into the large model. The aggregated resources and generated knowledge points are output and updated to the knowledge graph.

[0032] Alternatively, based on teachers' and students' recommendations and access behaviors to resources, the relationships between resource-related knowledge nodes can be fine-tuned, including weighting relevant knowledge points according to teachers' and students' resource recommendations, and enhancing the semantic relationships between resource-related knowledge points based on students' resource access trajectories.

[0033] This invention proposes two update mechanisms for knowledge graphs: First, it generates new resources and knowledge points based on community discovery algorithms. Communities are dynamically divided using a spatiotemporal weighted objective function (combining spatial coupling coefficients and time decay functions), and aggregated content is generated using a large model, overcoming the limitations of traditional static graph updates. Second, it fine-tunes relationships based on teacher-student behavior. Knowledge point weights and semantic relationships are adjusted through recommendation and access data, making the graph more aligned with actual teaching needs. These two update methods improve the timeliness and practicality of knowledge graphs from the dimensions of content expansion and relationship optimization, respectively, and are particularly suitable for scenarios with rapid resource growth and frequent teacher-student interaction in blended learning courses.

[0034] In an optional implementation, the method further includes: visualization and application of the knowledge graph, specifically including:

[0035] A multi-level visualization architecture is constructed, with the course textbook as the first level, and chapters, sections, and specific knowledge points as the second to fourth levels, clearly showing the hierarchical structure and module division of course knowledge. The specific knowledge points in the fourth level are associated with corresponding textbook segments and teaching-generated resources, and a relationship is formed between the resources and knowledge points.

[0036] The conceptual entities in the knowledge graph are visualized and expanded, and the relationship between knowledge points, resources and conceptual entities is dynamically displayed in the form of nodes and edges. The specific relationship description between terms is displayed through the attribute information of the relationship edges.

[0037] To enable cross-platform collaborative applications, this includes embedding links to relevant teaching resources on xMOOC chapter pages, allowing users to jump to the cMOOC platform to view corresponding discussion content and materials.

[0038] This invention presents a course knowledge system intuitively through a multi-level architecture (textbook → chapter → section → knowledge point), and dynamically displays the relationships between resources and concepts. Its innovations lie in: 1) cross-platform integration (jumping between xMOOC and cMOOC), breaking down the barriers between structured textbooks and unstructured discussions; and 2) an interactive node-edge design that supports semantic focus and detailed relationship viewing, enhancing the user experience. This visualization solution not only helps teachers and students quickly locate knowledge threads but also provides an intuitive basis for teaching decisions (such as resource recommendations and learning path planning), significantly improving the practical value of knowledge graphs in blended learning.

[0039] Secondly, the present invention provides a hybrid course knowledge graph construction system based on a large language model, the system comprising:

[0040] The data acquisition module is used to acquire multi-source heterogeneous resource data in the blended learning environment, including textbook resources and teaching-generated resources;

[0041] The entity extraction module is used to extract three types of entities from multi-source heterogeneous resource data through a natural language processing model: knowledge point entities, resource entities, and concept entities. Among them, knowledge point entities are extracted from the text data of textbook resources, and a mapping relationship between knowledge point entities and textbook resource entities is established.

[0042] The concept entity and relation extraction module is used to extract concept entities and relationships between entities from multi-source heterogeneous resource data based on a preset extraction model, evaluate the extraction effect using preset concept entity and relation extraction effect evaluation indicators, and adjust the extraction process based on the evaluation results.

[0043] The resource entity and knowledge point entity association identification module is used to identify the association relationship between resource entities and knowledge point entities in large models based on concept attention and context learning.

[0044] The knowledge graph construction module is used to integrate the three types of entities and their relationships to construct a knowledge graph.

[0045] The knowledge graph update module is used to update the constructed knowledge graph, including generating new resources and knowledge points to update the graph, and adjusting the relationships between resource entities based on teachers' and students' recommendations and access behaviors.

[0046] Thirdly, the present invention provides a computer device, including: a memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, and the processor executing the computer instructions to perform the hybrid course knowledge graph construction method based on a large language model described in the first aspect or any corresponding embodiment.

[0047] Fourthly, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to execute the method for constructing a hybrid course knowledge graph based on a large language model as described in the first aspect or any corresponding embodiment above.

[0048] Fifthly, the present invention provides a computer program product, including computer instructions for causing a computer to execute the method for constructing a hybrid course knowledge graph based on a large language model as described in the first aspect or any corresponding embodiment above. Attached Figure Description

[0049] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0050] Figure 1 This is a flowchart illustrating a method for constructing a hybrid course knowledge graph based on a large language model according to an embodiment of the present invention.

[0051] Figure 2 This is a schematic diagram illustrating the knowledge representation of a three-layer ontology structure according to an embodiment of the present invention;

[0052] Figure 3 This is a schematic diagram of a large model framework for identifying the association between heterogeneous resources and knowledge points based on concept attention and context learning according to an embodiment of the present invention;

[0053] Figure 4 This is a schematic diagram illustrating how, according to an embodiment of the present invention, a construction prompt is used as input to a large model, requiring the large model to output knowledge points corresponding to generated resources.

[0054] Figure 5 This is a schematic diagram of the overall architecture of the graph according to an embodiment of the present invention;

[0055] Figure 6 This is a preliminary schematic diagram of the knowledge points according to an embodiment of the present invention;

[0056] Figure 7 This is a schematic diagram of concept-related and pointing resources in a knowledge graph according to an embodiment of the present invention;

[0057] Figure 8 This is a structural block diagram of a hybrid course knowledge graph construction system based on a large language model according to an embodiment of the present invention;

[0058] Figure 9 A schematic diagram of the hardware structure of a computer device according to an embodiment of the present invention. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] To address the problems of traditional knowledge graph construction methods, which rely on expert definitions, neglect dynamic factors in the teaching process, and are ill-suited to the dynamic updates of multi-source heterogeneous resources in blended courses and the evolution of knowledge structures with teaching practice, resulting in inefficient information organization and retrieval and hindering teaching optimization and personalized services, this embodiment provides a blended course knowledge graph construction method based on a large language model. Through a knowledge graph construction and dynamic optimization method driven by a large language model, intelligent integration, precise association, and adaptive evolution of teaching resources are achieved. Figure 1 This is a flowchart of a method for constructing a hybrid course knowledge graph based on a large language model according to an embodiment of the present invention. Figure 1 As shown, the process includes the following steps:

[0061] Step S1: Obtain multi-source heterogeneous resource data in the blended learning environment, including textbook resources and teaching-generated resources.

[0062] Specifically, the teaching materials resources acquired by this invention can be electronic textbooks, teaching materials prepared by teachers during the course design stage, and other resources associated with specific knowledge points. Generated resources can include content generated through teacher-student interaction during the course teaching process, student assignments, and resources continuously generated by course-related knowledge communities.

[0063] Step S2 involves extracting three types of entities from multi-source heterogeneous resource data using a natural language processing model: knowledge point entities, resource entities, and concept entities. Knowledge point entities are extracted from the text data of textbook resources, and a mapping relationship is established between knowledge point entities and textbook resource entities.

[0064] Specifically, this embodiment of the invention, based on a blended learning environment, extracts three types of entities from resource data using a natural language processing model, and constructs a three-layer ontology structure for knowledge representation, such as... Figure 2 As shown, it includes:

[0065] 1. Knowledge Point Layer: This layer includes textbook knowledge points defined by experts and automatically extracted knowledge points. It is the core framework of the knowledge graph and has a hierarchical structure.

[0066] 2. Resource Layer: This layer includes pre-defined resources, teaching-generated resources, and automatically generated aggregated resources. Textbook resources are fragments of electronic textbook text (such as sections cut from a PDF). Teaching-generated resources are content generated by learners during blended learning (such as online teacher-student interaction texts, offline dialogue materials, student assignments, etc.) and learning materials and generated resources from course-related MOOC platforms or knowledge communities (such as blog posts, threads, etc.). Aggregated resources are generated by using community discovery algorithms (such as Louvain) to divide textbook resources and teaching-generated resources into communities, and then using a large language model to reorganize the resources within each community to create new resources.

[0067] 3. Conceptual layer: The keywords, terms, entities, and other micro-level knowledge units in the resources form the basis of semantic links.

[0068] The entities and relations in the knowledge graph constructed based on the above knowledge representation method are shown in Table 1 below, and the relation matrix in the knowledge graph is shown in Table 2 below.

[0069] Table 1

[0070]

[0071] Table 2

[0072]

[0073]

[0074] Step S3: Extract conceptual entities and relationships between entities from multi-source heterogeneous resource data based on a preset extraction model. Evaluate the extraction effect using preset conceptual entity and relationship extraction effect evaluation indicators, and adjust the extraction process based on the evaluation results.

[0075] Specifically, GraphRAG is used to extract concepts and semantic relationships from textbook resources and generated resource texts. The effectiveness of concept entity and relationship extraction is evaluated using metrics such as concept entity extraction stability and concept entity extraction accuracy. The hierarchy between concept entities and textbook resource entities, and between concept entities and generated resource entities, is also stored. For example, the results of GraphRAG from the resource document with ID "60f7f1dec4872bf181e0b875ea8b22ef" are as follows:

[0076] {"source":"Resource Library","target":"Hundred Schools Thousand Courses Alliance","weight":9,"description":"Resource Library is the basic platform for the Hundred Schools Thousand Courses Alliance to achieve resource sharing","text_unit_ids":["60f7f1dec4872bf181e0b875ea8b22ef"],"id":"038a351551124e078d62d943d8a089f4","human_readable_id":"188","source_degree":19,"target_deg ree":7,"rank":26}

[0077] The extracted conceptual entities are "Resource Library" and "Hundred Schools Thousand Courses Alliance". The relationship between the extracted conceptual entities is that "Resource Library is the basic platform for Hundred Schools Thousand Courses Alliance to achieve resource sharing".

[0078] In this embodiment of the invention, the concept extraction stability index is used to evaluate whether the large model can stably extract the core concepts of the course during multiple extraction processes. The smaller the index value, the higher the stability and convergence. The concept extraction accuracy index is used to evaluate whether the concepts extracted by the large model are accurate. The higher the index value, the more accurate the concept extraction. The concept relationship stability index is used to evaluate whether the large model can stably extract the concept relationships in the course during multiple extraction processes. The closer the index value is to 1, the more stable the concept relationship extraction. The concept relationship accuracy index is used to describe whether the large model's concept relationships for resources are reasonable and accurate. The higher the index value, the more accurate the concept relationship extraction. The calculation process of each index is explained in detail below:

[0079] (1) Extracting stability indicators for conceptual entities

[0080] set up: Given a collection of teaching materials, randomly select N resources from it. To generate a resource set, n resources are randomly selected from it, C iThe number of concepts extracted for the i-th concept entity (i = 1, 2, ..., x).

[0081] The stability calculation index of this invention has been upgraded based on the traditional stability formula as follows: ① Introducing a weighting factor: The importance of different time periods may be different, and a weighting factor w can be introduced for each time period t. t ① To reflect its importance or data quality. ② Incorporate a sliding window mechanism: Consider local stability over time. Let the window size be k, calculate the stability within each window, and then combine the results of all windows. ③ Multi-dimensional concept extraction: Divide concept extraction into multiple dimensions (such as semantics, syntax, topic, etc.), with the number of concepts in each dimension being... Then, calculate the comprehensive stability across multiple dimensions. ④ Nonlinear transformation: Perform a nonlinear transformation (such as an exponential or logarithmic transformation) on the stability index to enhance the complexity and adaptability of the formula. ⑤ Introduce external influencing factors: Incorporate the influence of external factors (such as concept activity) on stability, denoted as E. t And as an adjustment factor. The resource quantities m1 and m2 are randomly selected for each sampling, and the concept entity extraction stability index S is then used. dt The calculation method is as follows:

[0082]

[0083] Where: w t d is the weighting factor for time t, where T represents a predefined set of times and represents the range of values ​​for the time point t involved in the calculation; d is the number of dimensions for concept extraction. Let be the number of concepts in the j-th dimension at time t, and x be the total number of times the number of concepts are sampled at the same time t and the same dimension j. It is the mean of the j-th dimension at time t; E t λ is the external influencing factor of time t; λ is the adjustment parameter.

[0084] (2) Accuracy index of concept entity extraction

[0085] The concept extraction accuracy index P1 can be expressed as:

[0086]

[0087] in:

[0088]

[0089] in, The set of manually labeled concepts (gold standard); α is a variable parameter between 0 and 1, set according to expert experience; The set of concepts extracted for the k-th large model (k = 1, ..., x); Jaccard similarity is used to calculate the overlap between the two sets of concepts, with a value of [0, 1].

[0090]

[0091] Based on the semantic similarity GJaccard of graph structures, a graph convolutional network (GCN) is introduced to capture deep semantic relationships between concepts. ∩ represents a function that maps a set of concepts to a semantic graph structure. semantic and ∪ semantic This represents semantic-based graph structure intersection and merging operations, where node weights are determined by concept importance w. i The edge weight is determined by the strength of the semantic association between concepts.

[0092] Concept Entity Extraction Evaluation Comprehensive Index It can be represented as:

[0093]

[0094] Where α1∈[0,1] represents the weight, and max(C) is the maximum number of concepts extracted, used to eliminate dimensions. The comprehensive evaluation index for concept entity extraction improves the comprehensiveness of the evaluation by comprehensively considering the accuracy and stability of concept entity extraction. The weights are adjustable to support adaptation to different application scenarios.

[0095] (3) Stability index of relationships between conceptual entities

[0096] The stability index S2 of inter-concept relationships can be expressed as: Let M be the total number of rounds of relation extraction performed by the large-scale language model. Each extraction process is considered an independent knowledge discovery experiment. Let m be the set of relational knowledge obtained from the m-th extraction, where each element is a structured relation triple. Includes: a relation identifier r (composed of subject-object conceptual entity pairs); and a relation representation value. Spatiotemporal characteristics (Includes metadata such as extracted timestamps and input context hashes)

[0097] Define the global relation space This represents the set of all observed relation identifiers. A multimodal embedding model (emb(v,τ)) is used to simultaneously encode relation values ​​and spatiotemporal features, and an improved robustness measure is employed for similarity calculation.

[0098]

[0099] Where γ is the time-series decay coefficient, used to adjust the intensity of the influence of time factors on stability; τ u -τv The spatiotemporal distance is represented by a combination of timestamp distance and semantic distance from the context.

[0100] For each relation identifier Calculate the stability score of adaptive inter-concept relationships:

[0101]

[0102] The improvements in this formula include: ① Dynamic weights (Combining the extracted confidence weight w with the time interval). ② A non-linear adjustment factor (β>1) enhances the contribution of highly similar samples; ③ Effective comparison base (B r The sliding window counting method is used: only the most recent (K) valid matches are counted; Cr is the relationship stability; Br is the comparison cardinality, which is obtained by manually setting the sliding window count; sim is the similarity, which is calculated using the traditional cosine similarity method; vi and vj represent the vectors of the concepts.

[0103] Relationship-level importance weights are corrected using information entropy:

[0104]

[0105] Where H(r) is the Shannon entropy of the relation value distribution, and freq(r) represents the frequency;

[0106] The overall system's conceptual entity relationship stability index S2 is obtained using a dual-channel weighted calculation:

[0107]

[0108] in, Represents a set of highly deterministic relation spaces. A set of relational spaces representing uncertainty;

[0109] (4) Stability index of relationships between conceptual entities

[0110] The accuracy index P2 for relationships between concepts can be expressed as:

[0111]

[0112] The Dynamic Stability Index (DSI) is one of the key metrics for stability.

[0113]

[0114] Where: (λ) k =e -β·(k-1) ) is a time decay factor ((β>0)), which gives higher weight to recent results. It is the confidence level of the extraction result (γ is the smoothing parameter); This is a set of manually labeled relation names (gold standard). This is the set of relation names extracted from the k-th large model (k = 1, ..., M). Jaccard similarity is used to calculate the overlap between two sets of relation names, with values ​​in the range [0, 1]. A time decay factor is used to reflect the time sensitivity of the results. Confidence compensation is used to avoid spurious stability caused by low coverage and high similarity.

[0115] Conceptual Relationship Extraction Evaluation Comprehensive Index It can be represented as:

[0116]

[0117] Where α2∈[0,1] is the weight.

[0118] The comprehensive evaluation index for concept relationship extraction improves the overall comprehensiveness of the evaluation by taking into account both the accuracy and stability of concept relationship extraction. The weights are adjustable to adapt to different application scenarios.

[0119] Furthermore, the extraction process is adjusted based on the evaluation results, specifically including: if the comprehensive evaluation index for conceptual entity extraction is used... Comprehensive indicators for extracting and evaluating conceptual relationships If at least one of the indicators falls below the corresponding preset experience threshold, the extraction model is optimized by adjusting the prompts or adding supplementary examples. For example, the GraphRAG Prompt can be fine-tuned by adding correct or incorrect examples to the Prompt, or the instruction template of the input large model can be modified. This could involve refining the entity type description (e.g., specifying that textbook knowledge points must include definitions, attributes, and examples), or adding constraints (e.g., excluding terms from non-educational fields) to improve the targeting of extraction, until the calculated indicator exceeds the experience threshold.

[0120] This invention addresses the problem of accumulated errors caused by the lack of dynamic optimization in traditional methods by adjusting the extraction process based on evaluation results to form a closed loop of "evaluation-feedback-adjustment". It is particularly suitable for the dynamic changes of blended learning resources. It uses threshold-triggered targeted optimization (such as adjusting only the relation extraction module) to avoid the resource consumption of global retraining, thus improving efficiency. The prompts and supplementary examples are explicit interventions, which facilitate tracking the optimization process and meet the requirements for the credibility of knowledge graphs in teaching scenarios.

[0121] Step S4: Based on the large model of concept attention and context learning, identify the relationship between resource entities and knowledge point entities.

[0122] Specifically, this invention proposes a framework for identifying the association between multi-source heterogeneous resources and knowledge points based on concept attention and contextual learning in a large-scale model. For example... Figure 3 As shown, the framework comprises three parts: input processing, similarity comparison, and resource knowledge point classification based on a large model. The input processing part extracts concepts from textbook resources and teaching-generated resources, providing attention targets for subsequent models. The similarity comparison part constructs a contrastive learning model to achieve similar resource retrieval. The resource knowledge point classification based on a large model is used to identify knowledge points corresponding to the target resource. Specifically:

[0123] 1. Input processing: Experts divide the knowledge points in the textbook, and use GraphRAG (which combines knowledge graph (Graph) and retrieval augmentation generation (RAG) methods to improve the accuracy and interpretability of large language models (LLM) in knowledge-intensive tasks) to extract text summaries and concept entities from textbook resources and teaching-generated resources.

[0124] 2. Similar sample comparison: A contrastive learning model calculates and uses similar textbook resource demonstration samples as aids to improve the performance of few-shot learning. This involves three steps:

[0125] 1) Construction of spatiotemporally dynamically weighted positive and negative samples. A dynamic similarity function is used to calculate the dynamic similarity between textbook resources and generated resource samples, constructing positive and negative sample pairs with spatiotemporal evolution characteristics to train an adaptive contrastive learning model. Specifically, this includes:

[0126] (1) Dynamic similarity voting mechanism

[0127] For each input, generate a resource sample (G). t (t represents the time step), calculate its relationship with the set of teaching material resources (D). t The multimodal dynamic similarity index for all samples in the dataset. The dynamic similarity function is defined as: S dynamic (G t D t,i )=α(t)·S cos (G t D t,i )+β(t)·S jac (G t D t,i )+γ(t)·S lsa (G t D t,i )

[0128] Wherein: the time-varying weights α(t), β(t), and γ(t) are all dynamically adjusted through a time decay function:

[0129] α(t)=α0·e -λt,β(t)=β0·(1-e -μt ),

[0130] Among them, λ, μ, and τ are all adjustable hyperparameters used to realize the dynamic transfer of semantic, conceptual, and latent semantic weights.

[0131] Dynamic summary cosine similarity S cos Calculated using the following formula:

[0132]

[0133] Among them, RoBERTa t (·) represents the version of the RoBERTa model trained incrementally over time, with the embedding indicating dynamic updates.

[0134] Dynamic concept Jaccard similarity S jac Calculated using the following formula:

[0135]

[0136] Among them, t update Let T be the last update time of the concept set, T be the concept decay cycle parameter, and A be the concept set. t B t,i All are collections of a certain time period.

[0137] Dynamic LSA similarity S lsa Calculated using the following formula:

[0138]

[0139] Among them, LSA t Based on the corpus update within the sliding window, Relevance(t) is the time decay factor.

[0140] (2) Adaptive positive and negative sample selection: Dynamic Top-K sampling is performed based on the current S value. dynamic Values ​​are selected to determine positive and negative samples. Ultimately, this forms the dataset used for the contrastive learning model.

[0141] Positive samples are: (TopK) pos ={D t,i |rank(S) dynamic (G t D t,i ))≤K})

[0142] Negative samples: (TopK) neg ={D t,i |rank(S) dynamic (G t D t,i))≥|D t |-K})

[0143] Where K is the upper limit of the number of positive samples, and |Dt| represents the total number of sets.

[0144] Introducing a caching mechanism to retain historically high-value samples enhances spatiotemporal continuity:

[0145] D t =D current ∪{D t-Δ |S dynamic (G t-Δ D t-Δ )>θ}

[0146] Where θ represents the threshold parameter, which is set manually.

[0147] The above process has the following beneficial effects:

[0148] 1. Time-varying weighting mechanism: Adjust the contribution ratio of multimodal similarity through decay function and dynamic hyperparameter to adapt to data distribution drift.

[0149] 2. Incremental model updates: Embedded models (e.g., RoBERTa / LSA) and concept sets are trained incrementally over time to avoid the obsolescence problem of static models.

[0150] 3. Spatiotemporal caching strategy: Combining real-time data with historical high-value samples to resolve the contradiction between short-term fluctuations and long-term dependence.

[0151] 2) Contrastive Learning Model Training. This embodiment of the invention trains a contrastive learning model based on the supervised simCSE framework. The model learns the semantic representation of samples by minimizing the distance between generated resource samples and positive samples and maximizing the distance between generated resource samples and negative samples. A normalized cross-entropy loss function is used to optimize the model parameters, enabling the model to distinguish between similar and dissimilar samples. The normalized cross-entropy loss function...

[0152]

[0153] Where N is the number of sample pairs in a batch. and It is the cosine similarity between positive and negative sample pairs, and τ is the temperature hyperparameter.

[0154] 3) Similar Sample Retrieval. The trained contrastive learning model is used to calculate the similarity between the generated resource sample and each sample in the textbook resource set. Several textbook resource example samples with the highest similarity are selected as supplementary corpus for context learning, thereby improving the model's performance in predicting knowledge points from the generated resource samples.

[0155] 3. Based on an AI large-scale model, resource knowledge points are classified. This embodiment of the invention utilizes a contextual learning strategy based on concept attention. It uses concept entities, inter-concept relationships, and summaries extracted from textbook resources, along with example samples of textbook resources output by the contrastive learning model, as auxiliary corpus. It also uses concept entities, inter-concept relationships, and summaries extracted from teaching-generated resources as corpus to be classified. The format is as follows: Figure 4 As shown, following the template of AI large model task prompts + auxiliary corpus + corpus to be classified + output format, the prompts are constructed as input to the AI ​​large model. The large model is required to output knowledge points corresponding to the resources, so as to achieve a precise association between the generated resources and the knowledge points in the textbook.

[0156] Step S5: Integrate the three types of entities and their relationships to construct a knowledge graph.

[0157] The embodiments of the present invention integrate the resource entities, knowledge point entities, concept entities and their corresponding relationships extracted in the aforementioned stages to complete the construction of a knowledge graph.

[0158] Step S6 updates the constructed knowledge graph, including generating new resource and knowledge point updates to the graph, and adjusting the relationship weights between resource entities based on teachers' and students' recommendation and access behaviors.

[0159] Specifically, this embodiment of the invention combines systems science methods with the text generation capabilities of large-scale models. A community discovery algorithm aggregates knowledge points, resources, and concepts in a knowledge graph to generate an aggregated resource Prompt. The constructed Prompts are then input into a large-scale model in batches, leveraging the model's text generation capabilities to generate new resources and knowledge points. This process comprises two stages: community discovery and resource and knowledge point generation.

[0160] (1) In the community discovery stage, a dynamic evolution community discovery mechanism is added, which is constructed from the existing knowledge graph, and the spatiotemporally weighted knowledge graph G(t)=(V,E(t),W s (t),W t (t) where: V is the knowledge entity node; set E(t) is the time-varying edge set; W s (t) is the spatial correlation weight matrix; W t (t) is the time decay weight matrix, A a,b Is there an edge between entity i and entity j (1 for yes, 0 for no, or edge weight)? a δ(c) represents the degree of entity a (i.e., the number of connected edges or the sum of edge weights). a ,c b The value is 1 if a and b belong to the same club, and 0 otherwise.

[0161] The community found that the maximization objective function used takes into account the spatiotemporal coupling coefficient and the time decay function, and the adaptive balancing parameter is dynamically adjusted through deep learning training, and is calculated by the following formula:

[0162]

[0163] Its essence is to optimize the tightness of connections within the community and differentiate the connections between the community and external entities, including:

[0164] The normalization factor is represented by t(t), which usually refers to the number of edges in the network at time t or related statistics.

[0165] Used to compare "actual edge αA" a,b (t)” and “Random expected edge (by node degree k)” a (t)k b (t) calculation), α and β are adaptive balance parameters, which are dynamically adjusted through training with deep learning such as LSTM, so that the model can dynamically adjust the balance between "internal connection strengthening" and "external connection inhibition" according to the data, and adapt to complex scenarios.

[0166] Φ(s a ,s b ) represents the spatiotemporal coupling coefficient:

[0167]

[0168] s a ,s b This refers to different entities or entity vectors, whose spatial locations s are characterized by a Gaussian function. a ,s b The closer the distance, the stronger the coupling.

[0169] Γ(Δt) is the time decay function:

[0170]

[0171] It reflects the timeliness of the connection, t current -t origin When the time difference is λ, which controls the decay rate (λ<1), the influence of historical connections decreases over time.

[0172] (2) In the resource and knowledge point generation stage, for any community, the knowledge points, teaching materials, generated resources, and concepts in that community are taken as the model input corpus, concatenated to construct prompt words, and input into the large model to output aggregated resources and generated knowledge points. The aggregated resources and generated knowledge points output by the large model will be integrated into the existing graph according to the corresponding relationships in the "relationship matrix in the knowledge graph" mentioned above.

[0173] This invention also fine-tunes the relationships between resource-related knowledge nodes based on the recommendation and access behavior of teachers and learners on the course teaching platform. This makes the knowledge graph adaptable to teaching scenarios and learner groups, ensuring that the knowledge graph reflects important and trending knowledge and resources that teachers and students are interested in, and aligns with learners' cognitive processes and needs. Specific operations include teacher and student resource recommendations (such as "likes"), weighting resource-related knowledge points, tracking student resource access patterns, and adjusting the semantic relationships between resource-related knowledge points.

[0174] The method provided in this embodiment of the invention also includes: visualization and application of knowledge graphs, specifically including:

[0175] (1) Construct a multi-level visualization architecture, with the course textbook as the first level, and chapters, sections and specific knowledge points as the second to fourth levels, clearly showing the hierarchical structure and module division of course knowledge. The specific knowledge points in the fourth level are associated with the corresponding textbook segments and teaching-generated resources, and a related relationship is formed between the resources and the knowledge points.

[0176] (2) The embodiments of the present invention use a four-level hierarchical structure of "course textbook - chapter - subsection - knowledge point" as a framework, which can intuitively show the logical progression of knowledge. This allows students to quickly locate the position of a certain knowledge point in the overall knowledge system, avoid isolated learning, and form a complete understanding "from macro to micro" (for example, in the course "Computer Networks", it can be clearly seen that "TCP / IP protocol" belongs to the subsection "network layer", and "network layer" belongs to the chapter "network architecture", which is ultimately related to the core theme of the textbook).

[0177] (2) Visualize the conceptual entities in the knowledge graph, dynamically display the relationship between knowledge points, resources and conceptual entities in the form of nodes and edges, and display the specific relationship description between terms through the attribute information of the relationship edges.

[0178] This invention visualizes hidden connections across levels and modules by showing nodes (knowledge points, resources, conceptual entities) and relational edges. For example, the knowledge point of "derivative" in "calculus" (level 4) can be connected to the concept of "limit" (same-level connection), the case of "instantaneous velocity in physics" (resource connection), and "Chapter 1 of Advanced Mathematics" (upper-level chapter connection) through an edge. The attribute information of the relational edges (such as "derived from", "applied to", "extended from") can clarify the logic between terms (such as the "inverse operation" relationship between "derivative" and "differential"), helping students understand the intrinsic connections between knowledge points rather than rote memorization.

[0179] (3) Enable cross-platform collaborative applications, including: embedding links to relevant teaching resources on the chapter pages of xMOOC, supporting jumps to the cMOOC platform to view the corresponding discussion content and materials.

[0180] xMOOC refers to an online course model centered on traditional structured teaching, emphasizing the systematic and standardized nature of knowledge transmission. cMOOC, based on Connectivism, emphasizes interaction and co-construction of knowledge among learners. This invention connects structured textbooks (xMOOC) and unstructured generated content (cMOOC) through a knowledge graph, breaking down the barriers between the two models and achieving a seamless transition from systematic learning to extended discussion. For example, when a student is learning a chapter on "Principles of Economics" in xMOOC, they can directly jump to cMOOC to view discussion posts on that knowledge point (such as "The Deviation of Supply and Demand Models in the Actual Market") and case analyses shared by users, deepening their understanding of the knowledge. Teachers can identify students' learning difficulties and adjust their teaching strategies by using cross-platform data (such as the correlation between the learning time of a knowledge point in xMOOC and the popularity of related discussions in cMOOC).

[0181] In one embodiment, such as Figure 5 As shown in the diagram, the overall architecture of the course "Principles of Distance Education" clearly demonstrates the hierarchical structure and module division of knowledge. The diagram reveals that the course knowledge points are divided into four levels: Level 1: "Principles of Distance Education," which is the example textbook used in this study. Level 2: Chapters. Level 3: Subsections within chapters. Level 4: Specific knowledge points. For example, starting from the Level 1 knowledge point "Principles of Distance Education," the structure extends to the Level 2 "The Development Context of Online Education," then to the Level 3 "The Three Stages of Online Education Development," and finally reaches the Level 4 specific knowledge point "The Open Sharing Stage," forming a complete hierarchical relationship of knowledge points.

[0182] like Figure 6 As shown, taking the preliminary expansion diagram of the knowledge point "System Connectivity" (Level 4) as an example, it includes textbook excerpts and related blog-generated resources. These resources are interconnected with the knowledge points. For example... Figure 7 As shown, taking the concept of "district / county-level national primary and secondary schools" as an example, we can observe the "PART-OF" relationship between this entity and the resources it points to in the knowledge graph. Through this graph, we can see the connections between "Ministry of Education" and other terms and resources, and the specific relationship descriptions between these terms can be viewed through the description attribute of the edge "RELATED".

[0183] This embodiment also provides a hybrid course knowledge graph construction system based on a large language model. This system is used to implement the above embodiments and preferred embodiments, and details already described will not be repeated. As used below, the term "module" can be a combination of software and / or hardware that performs a predetermined function. Although the system described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0184] This embodiment provides a hybrid course knowledge graph construction system based on a large language model, such as... Figure 8 As shown, it includes:

[0185] The data acquisition module 81 is used to acquire multi-source heterogeneous resource data in the blended learning environment, including textbook resources and teaching-generated resources;

[0186] The entity extraction module 82 is used to extract three types of entities from multi-source heterogeneous resource data through a natural language processing model: knowledge point entities, resource entities, and concept entities. Among them, knowledge point entities are extracted from the text data of textbook resources, and a mapping relationship between knowledge point entities and textbook resource entities is established.

[0187] The concept entity and relation extraction module 83 is used to extract concept entities and relationships between entities from multi-source heterogeneous resource data based on a preset extraction model, evaluate the extraction effect using preset concept entity and relation extraction effect evaluation indicators, and adjust the extraction process based on the evaluation results.

[0188] The resource entity and knowledge point entity association identification module 84 adopts a large model of concept attention and context learning to identify the association relationship between resource entities and knowledge point entities;

[0189] Graph construction module 85 is used to integrate the three types of entities and their relationships to construct a knowledge graph;

[0190] The knowledge graph update module 86 is used to update the constructed knowledge graph, including generating new resources and knowledge points to update the graph, and adjusting the relationships between resource entities based on teachers' and students' recommendations and access behaviors.

[0191] In some optional implementations, a knowledge graph visualization and application module is also included, specifically including:

[0192] Multi-level visualization building units are used to construct a multi-level visualization architecture. The course textbook is the first level, and chapters, sections, and specific knowledge points are divided into the second to fourth levels, clearly showing the hierarchical structure and module division of course knowledge. The specific knowledge points in the fourth level are associated with corresponding textbook segments and teaching-generated resources, and a relationship is formed between the resources and knowledge points.

[0193] The visualization unit is used to visualize and expand the conceptual entities in the knowledge graph. It dynamically displays the relationship between knowledge points, resources and conceptual entities in the form of nodes and edges, and displays the specific relationship description between terms through the attribute information of the relationship edges.

[0194] The cross-platform linkage unit is used to enable cross-platform collaborative applications, including: embedding links to relevant teaching-generated resources on xMOOC chapter pages, supporting jumps to the cMOOC platform to view corresponding discussion content and materials.

[0195] Further functional descriptions of the above modules and units are the same as those in the corresponding embodiments described above, and will not be repeated here.

[0196] The hybrid course knowledge graph construction system based on a large language model in this embodiment is presented in the form of functional units. Here, a unit refers to an ASIC (Application Specific Integrated Circuit), a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0197] This invention also provides a computer device having the above-described features. Figure 8 The system shown is a hybrid course knowledge graph construction system based on a large language model.

[0198] Please see Figure 9 , Figure 9 This is a schematic diagram of the structure of a computer device provided in an optional embodiment of the present invention, such as... Figure 9 As shown, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interfaces). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 9 Take a processor 10 as an example.

[0199] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GDA), or any combination thereof.

[0200] The memory 20 stores instructions executable by at least one processor 10 to cause at least one processor 10 to perform the method shown in the above embodiments.

[0201] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0202] The memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0203] The computer device also includes a communication interface 30 for communicating with other devices or communication networks.

[0204] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods shown in the above embodiments.

[0205] A portion of this invention can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to the invention through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0206] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A method for constructing a knowledge graph for a blended learning curriculum based on a large language model, characterized in that, include: Acquire multi-source heterogeneous resource data in a blended learning environment, including textbook resources and teaching-generated resources; Three types of entities are extracted from multi-source heterogeneous resource data using a natural language processing model: knowledge point entities, resource entities, and concept entities. Among them, knowledge point entities are extracted from the text data of textbook resources, and a mapping relationship between knowledge point entities and textbook resource entities is established. Based on a pre-defined extraction model, conceptual entities and relationships between entities are extracted from multi-source heterogeneous resource data. The extraction effect is evaluated using pre-defined extraction effect evaluation indicators based on conceptual entities and relationships, and the extraction process is adjusted according to the evaluation results. A large model based on concept attention and context learning identifies the relationships between resource entities and knowledge point entities, including: By having experts divide the knowledge points in the textbooks, and using GraphRAG to extract text summaries and conceptual entities from textbook resources and teaching-generated resources, samples of textbook resources and teaching-generated resources are generated. The dynamic similarity function is used to calculate the dynamic similarity between the teaching-generated resource samples and each sample in the set of textbook resources. Dynamic Top-K sampling is used to select positive and negative samples based on the dynamic similarity value at the current time to train the contrastive learning model. By using a trained contrastive learning model, the similarity between the generated resource sample and each sample in the textbook resource set is calculated, and the most similar textbook resource demonstration samples are selected as auxiliary corpus for context learning. Using the contextual learning strategy of concept attention, the conceptual entities, inter-conceptual relationships, and summaries extracted from the textbook resources, as well as the textbook resource demonstration samples output by the contrastive learning model, are used as auxiliary corpus, and the conceptual entities, inter-conceptual relationships, and summaries extracted from the teaching-generated resources are used as corpus to be classified. Following the template of task prompts + auxiliary corpus + corpus to be classified + output format, the prompts are constructed as input to the preset AI large model. The large model is required to output knowledge points corresponding to the resources, so as to realize the association between the generated resource entities and knowledge point entities. A knowledge graph is constructed by integrating the three types of entities and their relationships. The constructed knowledge graph is updated, including generating new resources and knowledge points to update the graph, and adjusting the relationships between resource entities based on teachers' and students' recommendations and access behaviors.

2. The method according to claim 1, characterized in that, The preset evaluation indicators for concept entity and relation extraction include: concept entity extraction stability index, concept entity extraction accuracy index, concept entity relationship stability index, and concept entity relationship accuracy index; wherein: The concept entity extraction stability index is used to evaluate the stability of core concept extraction during multiple extractions of a large model. Its calculation incorporates weighting factors, sliding window mechanism, multi-dimensional concept extraction, nonlinear transformation, and external influencing factors. The concept entity extraction accuracy index is used to evaluate the accuracy of large models in extracting concepts. It is calculated using Jaccard similarity and graph-based semantic perception similarity, and the two similarities are weighted and integrated. The concept entity relationship stability index is used to evaluate the stability of concept relationship extraction during multiple extractions in a large model. Its calculation is based on the total number of times the large-scale language model performs multiple rounds of relationship extraction, the set of relationship knowledge extracted each time, and the global relationship space being the set of all observed relationship identifiers. It uses a multimodal embedding model to encode relationship values ​​and spatiotemporal features, combined with a temporal decay coefficient to calculate similarity. An adaptive concept relationship stability score is calculated for each relationship identifier, and information entropy is introduced to correct the relationship-level importance weight. The concept entity relationship stability index is obtained by combining the adaptive concept relationship stability score and the relationship-level importance weight using a dual-channel weighting method. The accuracy index of the relationship between conceptual entities is used to evaluate the accuracy of the large model in extracting the relationship between concepts. It is calculated using a dynamic stability index, combined with a time decay factor and confidence level.

3. The method according to claim 2, characterized in that, The process of evaluating the extraction effect using preset concept entity and relation extraction effect evaluation indicators, and adjusting the extraction process based on the evaluation results, includes: The concept entity extraction stability index and the concept entity extraction accuracy index are weighted and summed to obtain the comprehensive evaluation index of concept entity extraction. The stability index and the accuracy index of the relationship between conceptual entities are weighted and summed to obtain the comprehensive index for the evaluation of conceptual relationship extraction. If at least one of the comprehensive evaluation index for concept entity extraction and the comprehensive evaluation index for concept relationship extraction is lower than the corresponding preset experience threshold, the extraction model will be optimized by adjusting the prompts or providing supplementary examples.

4. The method according to claim 1, characterized in that, The generation of new resource and knowledge point update graphs, and the adjustment of relationships between resource entities based on teachers' and students' resource recommendation and access behavior, include: The constructed knowledge graph is used to discover communities using a community discovery algorithm. The community discovery adopts a maximization objective function, which considers the spatiotemporal coupling coefficient and the time decay function, and the adaptive balance parameter is dynamically adjusted through deep learning training. The community division results of the knowledge graph are obtained. For each community, the knowledge points, teaching materials, generated resources and concept construction prompts are input into the large model, and the aggregated resources and generated knowledge points are output and updated to the knowledge graph. Alternatively, based on teachers' and students' recommendations and access behaviors to resources, the relationships between resource-related knowledge nodes can be fine-tuned, including weighting relevant knowledge points according to teachers' and students' resource recommendations, and enhancing the semantic relationships between resource-related knowledge points based on students' resource access trajectories.

5. The method according to claim 1, characterized in that, Also includes: The visualization and application of knowledge graphs specifically include: A multi-level visualization architecture is constructed, with the course textbook as the first level, and chapters, sections, and specific knowledge points as the second to fourth levels, clearly showing the hierarchical structure and module division of course knowledge. The specific knowledge points in the fourth level are associated with corresponding textbook segments and teaching-generated resources, and a relationship is formed between the resources and knowledge points. The conceptual entities in the knowledge graph are visualized and expanded, and the relationship between knowledge points, resources and conceptual entities is dynamically displayed in the form of nodes and edges. The specific relationship description between terms is displayed through the attribute information of the relationship edges. To enable cross-platform collaborative applications, this includes embedding links to relevant teaching resources on xMOOC chapter pages, allowing users to jump to the cMOOC platform to view corresponding discussion content and materials.

6. A hybrid course knowledge graph construction system based on a large language model, characterized in that, include: The data acquisition module is used to acquire multi-source heterogeneous resource data in the blended learning environment, including textbook resources and teaching-generated resources; The entity extraction module is used to extract three types of entities from multi-source heterogeneous resource data through a natural language processing model: knowledge point entities, resource entities, and concept entities. Among them, knowledge point entities are extracted from the text data of textbook resources, and a mapping relationship between knowledge point entities and textbook resource entities is established. The concept entity and relation extraction module is used to extract concept entities and relationships between entities from multi-source heterogeneous resource data based on a preset extraction model, evaluate the extraction effect using preset concept entity and relation extraction effect evaluation indicators, and adjust the extraction process based on the evaluation results. The resource entity and knowledge point entity association recognition module is used by a large model based on concept attention and context learning to identify the association between resource entities and knowledge point entities. This includes: segmenting textbook knowledge points by experts; extracting text summaries and concept entities from textbook resources and teaching-generated resources using GraphRAG to generate samples of textbook resources and teaching-generated resources; calculating the dynamic similarity between the teaching-generated resource samples and each sample in the textbook resource set based on a dynamic similarity function; and selecting positive and negative samples based on the current dynamic similarity value through dynamic Top-K sampling to train a contrastive learning model; and using the trained contrastive learning model to calculate the relationship between the generated resource samples and the textbook resource set. The similarity of each sample is calculated, and the most similar textbook resource demonstration samples are selected as auxiliary corpus for context learning. Using the concept attention context learning strategy, the concept entities, inter-concept relationships, and summaries extracted from the textbook resources, as well as the textbook resource demonstration samples output by the contrastive learning model, are used as auxiliary corpus, and the concept entities, inter-concept relationships, and summaries extracted from the teaching generated resources are used as the corpus to be classified. Following the template of large model task prompt + auxiliary corpus + corpus to be classified + output format, the prompt is constructed as the input of the preset AI large model. The large model is required to output knowledge points corresponding to the resources, realizing the association between generated resource entities and knowledge point entities. The knowledge graph construction module is used to integrate the three types of entities and their relationships to construct a knowledge graph. The knowledge graph update module is used to update the constructed knowledge graph, including generating new resources and knowledge points to update the graph, and adjusting the relationships between resource entities based on teachers' and students' recommendations and access behaviors.

7. A computer device, characterized in that, include: The system includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to perform the hybrid course knowledge graph construction method based on a large language model as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the hybrid course knowledge graph construction method based on a large language model as described in any one of claims 1 to 5.

9. A computer program product, characterized in that, It includes computer instructions for causing a computer to execute the method for constructing a hybrid course knowledge graph based on a large language model, as described in any one of claims 1 to 5.

Images