A method and system for generating digital courses based on a large language model

Abstract

This invention discloses a digital course generation method and system based on a large language model. By closely integrating the acquisition of course materials, chapter segmentation and semantic analysis based on the large language model, and the construction of a multimodal index, it automatically maps multi-source heterogeneous information in teaching materials into a clear chapter hierarchy structure, and rapidly generates a structured model on this basis, thereby improving the organization efficiency and processing consistency of course content. Subsequently, by extracting semantic features from the structured model and integrating chapter hierarchy information to construct a knowledge graph, it achieves efficient transformation from raw teaching materials to a knowledge network that can be intelligently retrieved and visualized. Finally, the generated digital human image and its expressions and teaching actions are seamlessly integrated with the 3D teaching scene, structured model, and knowledge graph to form a complete digital course, significantly shortening the overall process time from material preparation to course delivery.

Description

Technical Field

[0001] This invention relates to the field of digital course technology, and in particular to a method and system for generating digital courses based on a large language model. Background Technology

[0002] Currently, with the application of big data and artificial intelligence technologies, digital textbooks and teaching platforms can achieve personalized learning and intelligent recommendations, improving learning efficiency and effectiveness. Simultaneously, the integration of virtual reality (VR) and augmented reality (AR) technologies makes online teaching content more vivid and enhances immersive interactive experiences, aiding students' understanding of complex knowledge. Furthermore, the analysis of learning behavior data can construct adaptive learning paths, allowing for timely adjustments to teaching strategies and promoting individualized instruction. Generative artificial intelligence significantly reduces the burden on teachers in content creation and automatic script generation, achieving intelligent content production. The remote and scalable nature of digital courses effectively meets the learning needs of remote areas or during specific time periods, promoting a more equitable distribution of educational resources.

[0003] While existing digital course generation methods can improve content production efficiency to some extent, they also have several shortcomings: First, the generation process relies heavily on fixed templates and preset rules, lacking personalized customization for different learners and failing to meet diverse learning needs. Second, automated content production often ignores instructional design principles, resulting in inconsistent quality of generated courseware and question banks, making it difficult to guarantee depth and accuracy. Third, these methods generally lack real-time teacher-student interaction and emotional feedback mechanisms, leading to a relatively one-way learning experience and impacting student engagement. Furthermore, the underlying technical architecture and algorithm implementation are often highly complex, requiring significant investment in hardware and maintenance costs, hindering widespread adoption. Additionally, the knowledge update cycle for generated content is long, making knowledge bases prone to information lag and obsolescence. Moreover, data privacy and copyright protection lack unified standards, potentially posing security and compliance risks in large-scale applications. Finally, the lack of unified standards and interfaces between different platforms hinders cross-system sharing and compatibility of course content, increasing the complexity of subsequent maintenance and upgrades. Summary of the Invention

[0004] This invention provides a method and system for generating digital courses based on a large language model, in order to improve the flexibility and quality of digital course generation.

[0005] To address the aforementioned technical problems, embodiments of the present invention provide a digital course generation method based on a large language model, comprising:

[0006] The teaching materials are acquired, and the teaching materials are split into chapters and semantically analyzed based on a large language model to obtain the chapter-level information of the teaching materials. A multimodal index is then constructed based on the chapter-level information to generate a course structured model.

[0007] Based on the large language model, feature extraction is performed on the course structure model to obtain semantic feature vectors, and a course knowledge graph is constructed based on the semantic feature vectors and the chapter level information.

[0008] A digital human image is generated based on 3D facial modeling technology, and a digital human expression sequence and a teaching action sequence are generated based on the digital human image, preset voice data, the course structure model, and a preset teaching action library.

[0009] Based on the digital human image, the digital human expression sequence, and the teaching action sequence, the digital human is embedded in a 3D teaching scene model, and a digital course is synthesized based on the 3D teaching scene model, the course structure model, and the course knowledge graph.

[0010] This invention tightly integrates course material acquisition, chapter segmentation and semantic analysis based on a large language model, and multimodal index construction. It automatically maps multi-source heterogeneous information in teaching materials into a clear chapter hierarchy structure, and rapidly generates a structured model on this basis, thereby improving the efficiency of course content organization and processing consistency. Subsequently, by extracting semantic features from the structured model and integrating chapter hierarchy information to construct a knowledge graph, it achieves efficient transformation from raw teaching materials to a knowledge network that can be intelligently retrieved and visualized. Finally, the generated digital human image and its expressions and teaching actions are seamlessly integrated with 3D teaching scenes, structured models, and knowledge graphs to form a complete digital course. This significantly shortens the overall process time from material preparation to course delivery and enhances the intelligent and modular teaching generation capabilities.

[0011] Furthermore, the teaching materials include text materials, audio materials, and 3D teaching materials. The process of acquiring the teaching materials involves performing chapter segmentation and semantic analysis based on a large language model to obtain the chapter-level information of the teaching materials, and constructing a multimodal index to generate a course structured model based on the chapter-level information, including:

[0012] Obtain teaching materials, perform chapter segmentation on the text materials based on the large language model, and obtain a list of chapter text segments of the teaching materials;

[0013] Based on the large language model, semantic analysis is performed on the list of chapter text segments segment by segment to obtain the keywords and semantic summaries of each chapter text segment of the teaching materials, and chapter hierarchy information is generated based on the keywords and semantic summaries of the chapter text segments and the list of chapter text segments.

[0014] The chapter-level information is numbered, and a multimodal index entry is constructed based on the numbering between the chapter-level information and text materials, audio materials, and 3D teaching materials;

[0015] Based on the chapter hierarchy information and multimodal index entries, a course structure model is constructed.

[0016] This invention utilizes a large language model to perform fine-grained chapter segmentation and semantic analysis on textual materials, obtaining an accurate list of chapter text segments along with their keywords and summaries, ensuring the logical integrity of the course framework. Subsequently, chapter-level information is numbered, and multimodal index entries for text, audio, and 3D materials are automatically created, achieving fine-grained association between various resources and chapters. Finally, a unified course structure model is constructed using these index entries, providing a consistent and easily expandable data foundation for subsequent action synthesis, knowledge retrieval, and interactive applications, thereby effectively improving the system's efficiency in organizing and managing multi-source teaching materials.

[0017] Furthermore, the step of extracting features from the course structure model based on the large language model to obtain semantic feature vectors, and constructing a course knowledge graph based on the semantic feature vectors and the chapter-level information, includes:

[0018] Based on the large language model, feature extraction is performed on the course structure model to obtain semantic feature vectors;

[0019] Knowledge nodes are created based on the semantic feature vectors and the chapter hierarchy information, and the semantic similarity between the knowledge nodes is calculated.

[0020] Based on the chapter hierarchy information, the chapter parent-child relationship of the knowledge node is obtained, and the association edge between the knowledge nodes is generated based on the semantic similarity and the chapter parent-child relationship, thus completing the construction of the knowledge graph.

[0021] This invention utilizes a large language model to extract semantic feature vectors and performs deep semantic encoding on the text of each chapter. By combining chapter-level information to create knowledge nodes and calculating the semantic similarity between nodes, the inherent connections between concepts can be accurately reflected. Furthermore, based on the parent-child relationship and similarity between chapters, related edges are generated. The resulting knowledge graph not only takes into account the logical structure of the chapters but also reflects the semantic tightness between knowledge points. This facilitates applications such as intelligent question answering, personalized recommendation, and learning path planning, effectively enhancing the completeness and interpretability of knowledge representation.

[0022] Furthermore, the process of generating a digital human image based on 3D facial modeling technology, and generating digital human expression sequences and teaching action sequences based on the digital human image, preset voice data, the course structured model, and preset teaching action library, includes:

[0023] A digital human avatar is generated based on the aforementioned course structure model and 3D facial modeling technology;

[0024] The speech data is subjected to speech framing and prosodic analysis to obtain semantic features;

[0025] Based on the large language model and the course structuring model, teaching intentions are identified, and semantic vectors for each teaching intention are generated.

[0026] The semantic features, semantic vectors, and preset teaching action library are weighted and fused based on the half-cosine curve function to generate a digital human expression sequence and teaching action sequence that are synchronized with the current semantic intent and semantic features.

[0027] This invention generates a high-precision digital human image based on a curriculum structure model and 3D facial modeling technology, ensuring the consistency between the virtual teacher's appearance and the teaching materials. By performing speech framing and prosodic analysis on the audio data to obtain speech features, and combining the teaching intention semantic vector identified by the large language model, a half-cosine weighted fusion algorithm is used to fuse speech, text, and action templates. This results in a sequence of facial expressions and action sequences that not only closely match the rhythm of the explanation but also accurately present the teaching intention. This breaks through the limitations of a fixed action library and achieves a more natural, coherent, and semantically driven interactive digital human demonstration.

[0028] Furthermore, after the knowledge graph construction is completed, the following steps are also included:

[0029] Collect question-and-answer records from users' learning behavior data, and classify the question-and-answer records according to knowledge nodes to obtain question-and-answer data;

[0030] Based on the question-and-answer data, the access frequency, answer accuracy rate and average response time of each knowledge node are statistically analyzed to obtain statistical results;

[0031] Based on the statistical results, the weights of each knowledge node and associated edge in the knowledge graph are dynamically updated using a preset weighted fusion model to dynamically optimize the knowledge graph.

[0032] After the knowledge graph is constructed, this invention continues to collect and classify users' question and answer records, and statistically analyzes the access frequency, answer accuracy, and average response time of each knowledge node to form learning behavior data. Based on these statistical results, a preset weighted fusion model is used to dynamically adjust the weights of knowledge graph nodes and related edges, so that the graph can reflect the user's mastery and weaknesses in real time, thereby achieving adaptive optimization of the knowledge graph and supporting more accurate knowledge push and personalized tutoring.

[0033] Secondly, the present invention provides a digital course generation system based on a large language model, comprising: a data aggregation layer, an intelligent agent training layer, a digital human interaction layer, and an application service layer;

[0034] The data aggregation layer is used to acquire teaching materials, perform chapter splitting and semantic analysis on the teaching materials based on a large language model, acquire chapter-level information of the teaching materials, and construct a multimodal index based on the chapter-level information to generate a course structured model.

[0035] The intelligent agent training layer is used to extract features from the course structure model based on the large language model, obtain semantic feature vectors, and construct a course knowledge graph based on the semantic feature vectors and the chapter level information.

[0036] The digital human interaction layer is used to generate a digital human image based on 3D facial modeling technology, and to generate digital human expression sequences and teaching action sequences based on the digital human image, preset voice data, the course structure model, and preset teaching action library.

[0037] The application service layer is used to embed the digital human into a 3D teaching scene model based on the digital human image, the digital human expression sequence, and the teaching action sequence, and to synthesize a digital course based on the 3D teaching scene model, the course structure model, and the course knowledge graph.

[0038] Furthermore, the data aggregation layer is used for teaching materials including text materials, audio materials, and 3D teaching materials. The acquisition of teaching materials involves performing chapter segmentation and semantic analysis on the teaching materials based on a large language model to obtain the chapter-level information of the teaching materials, and constructing a multimodal index based on the chapter-level information to generate a course structured model, including:

[0039] Obtain teaching materials, perform chapter segmentation on the text materials based on the large language model, and obtain a list of chapter text segments of the teaching materials;

[0040] Based on the large language model, semantic analysis is performed on the list of chapter text segments segment by segment to obtain the keywords and semantic summaries of each chapter text segment of the teaching materials, and chapter hierarchy information is generated based on the keywords and semantic summaries of the chapter text segments and the list of chapter text segments.

[0041] The chapter-level information is numbered, and a multimodal index entry is constructed based on the numbering between the chapter-level information and text materials, audio materials, and 3D teaching materials;

[0042] Based on the chapter hierarchy information and multimodal index entries, a course structure model is constructed.

[0043] Furthermore, the agent training layer is used to extract features from the course structured model based on the large language model, obtain semantic feature vectors, and construct a course knowledge graph based on the semantic feature vectors and the chapter-level information, including:

[0044] Based on the large language model, feature extraction is performed on the course structure model to obtain semantic feature vectors;

[0045] Knowledge nodes are created based on the semantic feature vectors and the chapter hierarchy information, and the semantic similarity between the knowledge nodes is calculated.

[0046] Based on the chapter hierarchy information, the chapter parent-child relationship of the knowledge node is obtained, and the association edge between the knowledge nodes is generated based on the semantic similarity and the chapter parent-child relationship, thus completing the construction of the knowledge graph.

[0047] Furthermore, the digital human interaction layer is used to generate a digital human image based on 3D facial modeling technology, and to generate digital human expression sequences and teaching action sequences based on the digital human image, preset voice data, the course structure model, and preset teaching action library, including:

[0048] A digital human avatar is generated based on the aforementioned course structure model and 3D facial modeling technology;

[0049] The speech data is subjected to speech framing and prosodic analysis to obtain semantic features;

[0050] Based on the large language model and the course structuring model, teaching intentions are identified, and semantic vectors for each teaching intention are generated.

[0051] The semantic features, semantic vectors, and preset teaching action library are weighted and fused based on the half-cosine curve function to generate a digital human expression sequence and teaching action sequence that are synchronized with the current semantic intent and semantic features.

[0052] Furthermore, the agent training layer is also used for:

[0053] Collect question-and-answer records from users' learning behavior data, and classify the question-and-answer records according to knowledge nodes to obtain question-and-answer data;

[0054] Based on the question-and-answer data, the access frequency, answer accuracy rate and average response time of each knowledge node are statistically analyzed to obtain statistical results;

[0055] Based on the statistical results, the weights of each knowledge node and associated edge in the knowledge graph are dynamically updated using a preset weighted fusion model to dynamically optimize the knowledge graph. Attached Figure Description

[0056] Figure 1 This is a flowchart illustrating a digital course generation method based on a large language model, provided as an embodiment of the present invention. Detailed Implementation

[0057] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0058] The terms "first" and "second," etc., in the specification, claims, and drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such processes, methods, products, or apparatus.

[0059] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments. Example

[0060] See Figure 1 , Figure 1 This is a flowchart illustrating a digital course generation method based on a large language model, provided by an embodiment of the present invention. The embodiment of the present invention provides a digital course generation method based on a large language model, including steps 101 to 104, as detailed below:

[0061] Step 101: Obtain teaching materials, perform chapter segmentation and semantic analysis on the teaching materials based on the large language model, obtain the chapter level information of the teaching materials, and construct a multimodal index based on the chapter level information to generate a course structured model;

[0062] In this embodiment, the teaching materials include text materials, audio materials, and 3D teaching materials. The process of acquiring the teaching materials involves performing chapter segmentation and semantic analysis on the teaching materials based on a large language model to obtain the chapter-level information of the teaching materials, and constructing a multimodal index to generate a course structured model based on the chapter-level information. This includes:

[0063] Obtain teaching materials, perform chapter segmentation on the text materials based on the large language model, and obtain a list of chapter text segments of the teaching materials;

[0064] Based on the large language model, semantic analysis is performed on the list of chapter text segments segment by segment to obtain the keywords and semantic summaries of each chapter text segment of the teaching materials, and chapter hierarchy information is generated based on the keywords and semantic summaries of the chapter text segments and the list of chapter text segments.

[0065] The chapter-level information is numbered, and a multimodal index entry is constructed based on the numbering between the chapter-level information and text materials, audio materials, and 3D teaching materials;

[0066] Based on the chapter hierarchy information and multimodal index entries, a course structure model is constructed.

[0067] In this embodiment, the system first acquires course materials uploaded by teachers, including text materials (such as PPT slides and lecture notes), audio materials (audio files), and 3D teaching materials (such as models or animation assets). During the acquisition of teaching materials, these multimodal resources are stored in object storage, and the metadata of each record is registered in the resource management module. Next, a pre-tuned chapter segmentation engine based on a large language model is invoked to perform boundary detection and semantic segmentation on each segment of the text materials, outputting a "chapter text segment list" (each chapter segment contains a set of consecutive text paragraphs and their source identifiers). Subsequently, a semantic analyzer based on the same large language model is used to sequentially perform keyword extraction and summary generation on each chapter text segment, obtaining a core terminology set and a brief topic description for each chapter segment. Based on the "chapter text segment list," "keyword set," and "semantic summary," chapter hierarchy information (including chapter number, parent-child relationship, and paragraph range) is constructed. Based on this, the system assigns chapter-level information numbers and associates them with corresponding text paragraphs, audio timestamps, and 3D teaching resource identifiers to generate a multimodal index entry for each chapter. Each entry records mapping relationships such as "chapter number → list of text paragraph IDs," "audio timestamp range," and "3D material ID." Finally, the platform aggregates all multimodal index entries in chapter order to generate a "course structured model" that can be used by downstream modules.

[0068] In this embodiment, the course structure model retains the chapter tree structure and semantic overview, and also realizes the fine association between text, voice and 3D materials, serving as a unified data foundation for subsequent functional modules such as knowledge graph construction, digital human generation and intelligent question answering.

[0069] This invention utilizes a large language model to perform fine-grained chapter segmentation and semantic analysis on textual materials, obtaining an accurate list of chapter text segments along with their keywords and summaries, ensuring the logical integrity of the course framework. Subsequently, chapter-level information is numbered, and multimodal index entries for text, audio, and 3D materials are automatically created, achieving fine-grained association between various resources and chapters. Finally, a unified course structure model is constructed using these index entries, providing a consistent and easily expandable data foundation for subsequent action synthesis, knowledge retrieval, and interactive applications, thereby effectively improving the system's efficiency in organizing and managing multi-source teaching materials.

[0070] Step 102: Based on the large language model, extract features from the course structure model to obtain semantic feature vectors, and construct a course knowledge graph based on the semantic feature vectors and the chapter level information;

[0071] In this embodiment, the step of extracting features from the course structure model based on the large language model to obtain semantic feature vectors, and constructing a course knowledge graph based on the semantic feature vectors and the chapter-level information, includes:

[0072] Based on the large language model, feature extraction is performed on the course structure model to obtain semantic feature vectors;

[0073] Knowledge nodes are created based on the semantic feature vectors and the chapter hierarchy information, and the semantic similarity between the knowledge nodes is calculated.

[0074] Based on the chapter hierarchy information, the chapter parent-child relationship of the knowledge node is obtained, and the association edge between the knowledge nodes is generated based on the semantic similarity and the chapter parent-child relationship, thus completing the construction of the knowledge graph.

[0075] In this embodiment, a pre-trained and fine-tuned large language model is loaded to perform deep semantic encoding on each chapter and its multimodal index entries in the generated course structured model, thereby extracting its semantic feature vectors. Specifically, the large language model uses a self-attention mechanism to perform contextual fusion on the chapter text, keywords, abstracts, and related 3D material descriptions, outputting fixed-dimensional vector representations to fully characterize the semantic connotation of each chapter. Subsequently, the system maps the semantic feature vectors of each chapter and their corresponding chapter-level information to knowledge nodes. These knowledge nodes include conceptual entities at the "chapter" level and can be further refined to the "knowledge point" level.

[0076] In this embodiment, after generating knowledge points, the semantic association strength between any two nodes is measured by calculating their vector cosine similarity, and node pairs exceeding a preset threshold are marked as "semantically related". Finally, the system combines the parent-child relationships defined in the chapter hierarchy information to establish a "parent-child" structural edge for each knowledge node; and based on the aforementioned semantic similarity judgment results, adds "cross-chapter" or "same-level" association edges between nodes. Thus, a course knowledge graph that retains the course organization structure and incorporates deep semantic connections is constructed in the graph database, providing a solid data foundation for subsequent intelligent question answering, learning path recommendation, and knowledge visualization.

[0077] In this embodiment, semantic feature vectors are extracted using a large language model to perform deep semantic encoding on the text of each chapter; knowledge nodes are created by combining chapter-level information and the semantic similarity between nodes is calculated, which can accurately reflect the inherent relationship between concepts; then, related edges are generated based on the parent-child relationship of chapters and similarity. The knowledge graph finally constructed not only takes into account the logical structure of chapters, but also reflects the semantic tightness between knowledge points, which is helpful for applications such as intelligent question answering, personalized recommendation and learning path planning, and effectively enhances the completeness and interpretability of knowledge representation.

[0078] In this embodiment, after the knowledge graph construction is completed, the following steps are also included:

[0079] Collect question-and-answer records from users' learning behavior data, and classify the question-and-answer records according to knowledge nodes to obtain question-and-answer data;

[0080] Based on the question-and-answer data, the access frequency, answer accuracy rate and average response time of each knowledge node are statistically analyzed to obtain statistical results;

[0081] Based on the statistical results, the weights of each knowledge node and associated edge in the knowledge graph are dynamically updated using a preset weighted fusion model to dynamically optimize the knowledge graph.

[0082] In this embodiment, after the knowledge graph is constructed, the acquisition module collects the question-and-answer records generated by students during the question-and-answer interaction in real time. Based on the predefined knowledge node index in the knowledge graph, each question-and-answer detail (including the question content, system answer, whether the answer is correct, and response time, etc.) is mapped to the corresponding node. Subsequently, the statistical analysis engine summarizes and calculates indicators such as access frequency, accuracy rate, and average response time of the mapped question-and-answer data according to the node dimension in batch processing or streaming mode. Finally, the system calls the pre-designed weighted fusion model to merge the above statistical results with the original graph weights according to the preset weight strategy, dynamically adjusts the weight value of each knowledge node and its associated edges in the graph database, realizes real-time feedback and adaptive optimization of the knowledge graph on students' learning behavior, and thus continuously improves the accuracy of intelligent agent question answering and the effect of personalized recommendation.

[0083] After the knowledge graph is constructed, this invention continues to collect and classify users' question and answer records, and statistically analyzes the access frequency, answer accuracy, and average response time of each knowledge node to form learning behavior data. Based on these statistical results, a preset weighted fusion model is used to dynamically adjust the weights of knowledge graph nodes and related edges, so that the graph can reflect the user's mastery and weaknesses in real time, thereby achieving adaptive optimization of the knowledge graph and supporting more accurate knowledge push and personalized tutoring.

[0084] Step 103: Generate a digital human image based on 3D facial modeling technology, and generate a digital human expression sequence and a teaching action sequence based on the digital human image, preset voice data, the course structure model, and preset teaching action library;

[0085] In this embodiment, the step of generating a digital human image based on 3D facial modeling technology, and generating a digital human expression sequence and a teaching action sequence based on the digital human image, preset voice data, the course structure model, and a preset teaching action library, includes:

[0086] A digital human avatar is generated based on the aforementioned course structure model and 3D facial modeling technology;

[0087] The speech data is subjected to speech framing and prosodic analysis to obtain semantic features;

[0088] Based on the large language model and the course structuring model, teaching intentions are identified, and semantic vectors for each teaching intention are generated.

[0089] The semantic features, semantic vectors, and preset teaching action library are weighted and fused based on the half-cosine curve function to generate a digital human expression sequence and teaching action sequence that are synchronized with the current semantic intent and semantic features.

[0090] In this embodiment, firstly, based on the chapter text, keywords, and associated 3D teaching material information in the aforementioned course structured model, 3D facial modeling technology (such as the 3DMM-based face reconstruction algorithm) is invoked to reconstruct a digital human facial geometric mesh with an accuracy of 0.1mm, and a complete digital human image is generated by combining expression-driven parameters. Next, the voice data provided by the teacher is subjected to voice framing and prosodic analysis to extract voice features such as phoneme boundaries, stress positions, and speech rate changes to reflect the rhythm and emotional tone of the explanation. Then, the system combines the chapter text in the course structured model with the large language model, identifies the teaching intentions contained in each chapter or key sentence through a self-attention mechanism, and outputs the corresponding semantic vectors as content guidelines for action generation. Finally, the system uses a half-cosine curve function to weight and fuse the above-mentioned voice features, teaching intention semantic vectors, and facial expression and gesture action fragments in a pre-collected and labeled teaching action library to generate a digital human expression sequence and teaching action sequence that are strictly aligned with the voice prosody and accurately convey the teaching intentions. This achieves a natural, coherent, and semantically driven dynamic demonstration of the digital human in different teaching scenarios.

[0091] In this embodiment, a high-precision digital human image is generated based on a course structured model and 3D facial modeling technology, ensuring the consistency between the virtual teacher's appearance and the teaching materials. By performing speech framing and prosodic analysis on the speech data to obtain speech features, and combining the teaching intention semantic vector identified by the large language model, a half-cosine weighted fusion algorithm is used to fuse speech, text, and action templates. This makes the generated expression sequence and action sequence highly consistent with the rhythm of the explanation and accurately present the teaching intention, breaking through the limitations of a fixed action library and realizing a more natural, coherent, and semantically driven interactive digital human demonstration.

[0092] Step 104: Based on the digital human image, the digital human expression sequence, and the teaching action sequence, embed the digital human into a 3D teaching scene model, and synthesize a digital course based on the 3D teaching scene model, the course structure model, and the course knowledge graph.

[0093] In this embodiment, a pre-built 3D teaching scene model is first loaded into a 3D rendering engine (such as Unity or Unreal Engine). This model includes various visual elements such as classroom scenery, experimental devices, or example objects. Then, the digital human image generated by the 3D facial modeling and motion synthesis module, along with the corresponding expression sequence and teaching action sequence, are sequentially jittered into the scene. The real-time demonstration in the virtual environment is driven by skeletal skinning and facial driving parameters. Next, the camera path and lens switching strategy within the scene are automatically configured according to the chapter order in the course structure model and the association logic of each knowledge node in the knowledge graph to ensure the coherence and emphasis of the digital human's explanation. At each knowledge node or chapter node, the corresponding 3D teaching model (such as neural network structure, mechanical principle diagram, etc.) can be automatically triggered for local interactive display, and the demonstration duration and lens focus are dynamically adjusted through the node weight information in the graph. Finally, the rendering engine synthesizes all chapter fragments into a high-definition video with subtitles and interactive annotations, and outputs a complete digital course file. This achieves a deep integration of digital humans, 3D scenes, and knowledge models, while also ensuring the structured presentation and visual interactive effects of the teaching content.

[0094] This invention also provides a digital course generation device based on a large language model, comprising: a data aggregation layer, an intelligent agent training layer, a digital human interaction layer, and an application service layer;

[0095] The data aggregation layer is used to acquire teaching materials, perform chapter splitting and semantic analysis on the teaching materials based on a large language model, acquire chapter-level information of the teaching materials, and construct a multimodal index based on the chapter-level information to generate a course structured model.

[0096] The intelligent agent training layer is used to extract features from the course structure model based on the large language model, obtain semantic feature vectors, and construct a course knowledge graph based on the semantic feature vectors and the chapter level information.

[0097] The digital human interaction layer is used to generate a digital human image based on 3D facial modeling technology, and to generate digital human expression sequences and teaching action sequences based on the digital human image, preset voice data, the course structure model, and preset teaching action library.

[0098] The application service layer is used to embed the digital human into a 3D teaching scene model based on the digital human image, the digital human expression sequence, and the teaching action sequence, and to synthesize a digital course based on the 3D teaching scene model, the course structure model, and the course knowledge graph.

[0099] In this embodiment, the data aggregation layer is used for teaching materials including text materials, audio materials, and 3D teaching materials. The acquisition of teaching materials involves performing chapter segmentation and semantic analysis on the teaching materials based on a large language model to obtain the chapter-level information of the teaching materials, and constructing a multimodal index based on the chapter-level information to generate a course structured model, including:

[0100] Obtain teaching materials, perform chapter segmentation on the text materials based on the large language model, and obtain a list of chapter text segments of the teaching materials;

[0101] Based on the large language model, semantic analysis is performed on the list of chapter text segments segment by segment to obtain the keywords and semantic summaries of each chapter text segment of the teaching materials, and chapter hierarchy information is generated based on the keywords and semantic summaries of the chapter text segments and the list of chapter text segments.

[0102] The chapter-level information is numbered, and a multimodal index entry is constructed based on the numbering between the chapter-level information and text materials, audio materials, and 3D teaching materials;

[0103] Based on the chapter hierarchy information and multimodal index entries, a course structure model is constructed.

[0104] In this embodiment, the agent training layer is used to extract features from the course structured model based on the large language model, obtain semantic feature vectors, and construct a course knowledge graph based on the semantic feature vectors and the chapter-level information, including:

[0105] Based on the large language model, feature extraction is performed on the course structure model to obtain semantic feature vectors;

[0106] Knowledge nodes are created based on the semantic feature vectors and the chapter hierarchy information, and the semantic similarity between the knowledge nodes is calculated.

[0107] Based on the chapter hierarchy information, the chapter parent-child relationship of the knowledge node is obtained, and the association edge between the knowledge nodes is generated based on the semantic similarity and the chapter parent-child relationship, thus completing the construction of the knowledge graph.

[0108] In this embodiment, the digital human interaction layer is used to generate a digital human image based on 3D facial modeling technology, and to generate digital human expression sequences and teaching action sequences based on the digital human image, preset voice data, the course structure model, and a preset teaching action library, including:

[0109] A digital human avatar is generated based on the aforementioned course structure model and 3D facial modeling technology;

[0110] The speech data is subjected to speech framing and prosodic analysis to obtain semantic features;

[0111] Based on the large language model and the course structuring model, teaching intentions are identified, and semantic vectors for each teaching intention are generated.

[0112] The semantic features, semantic vectors, and preset teaching action library are weighted and fused based on the half-cosine curve function to generate a digital human expression sequence and teaching action sequence that are synchronized with the current semantic intent and semantic features.

[0113] In this embodiment, the agent training layer is further used for:

[0114] Collect question-and-answer records from users' learning behavior data, and classify the question-and-answer records according to knowledge nodes to obtain question-and-answer data;

[0115] Based on the question-and-answer data, the access frequency, answer accuracy rate and average response time of each knowledge node are statistically analyzed to obtain statistical results;

[0116] Based on the statistical results, the weights of each knowledge node and associated edge in the knowledge graph are dynamically updated using a preset weighted fusion model to dynamically optimize the knowledge graph.

[0117] In this embodiment, the digital course generation system adopts a four-layer B / S architecture. The data aggregation layer is mainly responsible for the acquisition and preprocessing of course materials. Teachers upload text materials (PPT, lecture notes), audio materials (audio files), and 3D teaching materials to the cloud or local server through a browser or local client. The data cleaning module in the aggregation layer calls the fine-tuned large language model to perform chapter splitting and semantic analysis on the text materials, outputting a "chapter text segment list," keywords, and summaries. It also combines audio frame timestamps and 3D material identifiers to construct a multimodal index, ultimately generating a course structured model. The intelligent agent training layer extracts features from the structured model based on the same large language model, obtains the semantic feature vectors of each chapter, and combines chapter hierarchical information to create knowledge nodes, semantic similarity between computing nodes, and parent-child relationships in the graph database to dynamically construct and optimize the course knowledge graph. The digital human interaction layer uses the course structured model and 3D facial modeling technology to generate a high-precision digital human image, combining speech prosody features and LLM. The identified teaching intent vector is used to generate a sequence of facial expressions and teaching actions synchronized with semantics and speech through a half-cosine weighted fusion algorithm. This digital human is then embedded into a 3D teaching scene model and co-rendered with a knowledge graph and structured model to output digital course videos with subtitles and interactive annotations. The application service layer seamlessly integrates service modules such as the intelligent agent question-and-answer platform, teacher lesson preparation toolbox, and student learning path recommendation engine with the above generation process, enabling teachers to easily create, manage, and publish digital courses, and allowing students to initiate questions and answers, view learning progress, and receive personalized learning path recommendations in real time during the learning process.

[0118] In this embodiment, the system collaboratively completes digital course generation according to a four-layer architecture: First, in the data aggregation layer, teachers upload "Machine Learning" PPT courseware through a browser. The platform automatically calls a large language model to perform chapter splitting and semantic analysis on the PPT text materials, extracting the titles, keywords, and summaries of each chapter, and constructing a multimodal index and course structure model by combining the corresponding 3D teaching materials and voice scripts; Next, in the agent training layer, the system calls LLM based on the structure model to perform in-depth semantic feature extraction, generating semantic feature vectors for each chapter, constructing a mind map-style knowledge graph, and generating two query subsets, "teaching cases" and "ideological and political cases," in the knowledge base based on a hybrid question-and-answer mode; Subsequently, in the digital human interaction... In the interlayer, the platform reconstructs a digital human image based on 3D facial modeling technology. It then combines this image with 3D models such as the structured course model and the neural network structure represented by the knowledge graph. Through a semi-cosine weighted fusion algorithm, it integrates the semantic vector of teaching intent with the prosodic features of speech, triggering dynamic actions such as "gestures" to generate facial expressions and body sequences. This is then rendered and synthesized in real time into a subtitled explanatory video. Finally, in the application service layer, when a student asks a question, the agent simultaneously retrieves the dedicated course knowledge base and the pre-training data of the large model to output the answer. If the teacher marks it as incorrect, the feedback is injected into the agent's training layer through a recall-correction-retraining process, enabling the knowledge base to be dynamically updated iteratively. This closed-loop improvement of answer accuracy and the continuous optimization and teaching adaptability of the course agent are ensured.

[0119] In this embodiment, the digital course generation system is also a course intelligent agent teaching assistant system. Student users can initiate various types of intelligent question-and-answer services such as "AI-guided learning," "AI-assisted learning," "AI-assisted training," or "AI-expanded learning" at any time through the front-end learning interface. The interactive prompts or intelligent agents are created and maintained by teachers in the background. The system associates courseware, teaching videos, generated teaching cases with ideological and political cases, prompts, and intelligent agent resources according to the multi-layered chapter structure of the course. Students can quickly search and call the corresponding content in any chapter dimension. The question-and-answer module supports three modes: dedicated data question-and-answer (limited to the course knowledge base), mixed data question-and-answer (course knowledge base and large model pre-training data), and dedicated question-and-answer based on chapter intelligent agents to meet learning needs of different depths and breadths. In addition, the system provides learning plan formulation and resource recommendation functions, automatically generating personalized learning paths based on students' knowledge mastery. Students can also conduct online assessments of their learning effects at any time to grasp their learning progress and weaknesses.

[0120] In this embodiment, the platform's "Intelligent Agent Square" allows administrators or teachers to publish intelligent agents related to learning, courses, and even daily life. It supports accessing third-party intelligent agents via web links or SDK parameters, and AI can automatically generate introductions and configure access permissions. Teachers can also create and manage prompt word templates for each chapter, similarly supporting AI-generated introductions and status settings. They can also centrally initiate prompt-word or intelligent agent-guided Q&A sessions in the AI ​​workbench. The system has a built-in courseware document training function: after students or teachers upload untrained files and select training, they can build a dedicated course knowledge base in the large model. If the training results are unsatisfactory, they can recall and modify the files for retraining at any time. Simultaneously, the platform can automatically generate mind map-style course graphs based on chapter hierarchy and related data, and provide learning record analysis and graph display, learning path recommendations, and problem tracking services. Teachers and administrators can also query all Q&A conversation records of students within a class for teaching effectiveness evaluation and personalized tutoring. The online learning module supports resume playback and download of materials. The learning interface presents the course structure intuitively in the form of mind maps, helping students to organize the knowledge. In addition, the system integrates intelligent assisted programming functions, providing students with one-stop programming support such as code generation, parsing, optimization, error checking and formatting, which greatly improves learning efficiency and experience.

[0121] In this embodiment, teachers can also quickly create and manage courses through this system: First, teachers enter the course name and core keywords in the "Course Management" interface. The system calls a large language model to automatically generate a course introduction and supports setting multi-level access permissions for students, teachers, or administrators. Subsequently, teachers can quickly build a course chapter system by manually adding chapters hierarchically, importing XMind files, generating chapter structures with AI with one click, or automatically extracting chapters directly from uploaded PPTs / documents. In the chapter management panel, teachers can drag and drop to adjust hierarchical relationships, upgrade or demote nodes, show / hide chapters, edit titles, or batch delete useless nodes. Under each chapter, teachers can upload or associate teaching materials in various formats such as documents, videos, and compressed files to build a resource library, and generate teaching cases and ideological and political cases using natural language input. The generated content can be freely output or use preset templates. The resource management function supports uploading, editing, moving, downloading, and pinning of multi-level materials, realizing courseware reuse and unified management. Teachers can centrally configure multiple intelligent agents in the AI ​​workbench, including assigning access permissions by chapter or importing third-party intelligent agents, and setting their display and calling strategies in the "Intelligent Agent Square". The platform also allows multiple teachers to co-write courses, and invite students to join in batches through invitation codes, student IDs or lists, and set personalized access and public access strategies for different classes. In addition, the system includes a one-click teaching assistant toolbox, which allows teachers to directly access teaching cases, ideological and political cases, intelligent question generation, intelligent assessment, video generation and prompt word templates, etc., to greatly improve lesson preparation efficiency and teaching quality.

[0122] In this embodiment, an intelligent course learning and tutoring system for students is constructed, integrating multi-level course content management, intelligent agent question answering, personalized learning support, and AI teaching assistant tools. The system allows teachers to create course structures through an AI workbench, manually adding chapters, importing mind maps, automatically extracting chapters from PPTs, or calling AI to generate course chapters with a single click, flexibly building a multi-level course system. Each chapter can be associated with teaching resources such as courseware, videos, teaching cases, and ideological and political cases, while also supporting the generation of case content from natural language and outputting it according to templates. During the learning process, students can conduct AI question answering based on courseware content, supporting three modes: dedicated data question answering, mixed data question answering, and chapter-based intelligent agent question answering, comprehensively improving the depth of course comprehension and interactive experience. The system also supports students in creating learning plans, self-assessing, generating learning graphs, analyzing learning records, recommending learning paths, and querying personal and class learning question answering records. Teachers can create and manage prompts and intelligent agents, configure third-party access methods, control their display and access permissions, and utilize courseware training functions to build a dedicated course knowledge base. The platform also provides an "Intelligent Agent Plaza" where students can access intelligent agents related to their lives and studies, enabling 24 / 7 learning support. Furthermore, the system integrates one-stop intelligent assisted programming functions, covering tasks such as code generation, parsing, optimization, error checking, and formatting, empowering students to improve their interdisciplinary abilities. This solution effectively achieves a deep integration of course content and intelligent interaction, providing systematic support for personalized learning and intelligent teaching.

[0123] In this embodiment of the invention, a terminal device is also provided, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the above-described digital course generation method based on a large language model.

[0124] In this embodiment of the invention, a computer-readable storage medium is also provided, which includes a stored computer program, wherein the computer program controls the device where the computer-readable storage medium is located to execute the above-described digital course generation method based on a large language model when it is running.

[0125] For example, a computer program can be divided into one or more modules, one or more of which are stored in memory and executed by a processor to perform the present invention. The one or more modules can be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in a terminal device.

[0126] The terminal device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor, memory, and display. Those skilled in the art will understand that the above components are merely examples of terminal devices and do not constitute a limitation on the terminal device. It may include more or fewer components, or combinations of certain components, or different components. For example, the terminal device may also include input / output devices, network access devices, buses, etc.

[0127] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal device, connecting all parts of the terminal device through various interfaces and lines.

[0128] Memory can be used to store computer programs and / or modules. The processor implements various functions of the terminal device by running or executing the computer programs and / or modules stored in the memory, and by accessing data stored in the memory. Memory can mainly include a program storage area and a data storage area. The program storage area can store the operating system, application programs required for at least one function (such as sound playback, text conversion, etc.), etc.; the data storage area can store data created based on the use of the mobile phone (such as audio data, text message data, etc.). In addition, memory can include high-speed random access memory, and can also include non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart media cards (SMC), secure digital cards (SD cards), flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0129] In this invention, modules for generating digital courses based on large language models, if implemented as software functional units and sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. Computer-readable media can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. Those skilled in the art can understand and implement this invention without any inventive effort.

[0130] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for generating digital courses based on a large language model, characterized in that, include: Acquire teaching materials, including text materials, audio materials, and 3D teaching materials; perform chapter segmentation processing on the text materials based on the large language model to obtain a list of chapter text segments of the teaching materials; Based on the large language model, semantic analysis is performed on the list of chapter text segments segment by segment to obtain the keywords and semantic summaries of each chapter text segment of the teaching materials, and chapter hierarchy information is generated based on the keywords and semantic summaries of the chapter text segments and the list of chapter text segments. The chapter-level information is numbered, and after being numbered, it is associated with the corresponding text data paragraphs, audio data timestamps, and 3D teaching data resource identifiers to generate a multimodal index entry for each chapter. All multimodal index entries are compiled in chapter order to construct a course structure model; Load a pre-trained and fine-tuned large language model, perform deep semantic encoding on each chapter and its multimodal index entries in the generated course structured model, and extract semantic feature vectors. Knowledge nodes are created based on the semantic feature vectors and the chapter hierarchy information, and the semantic similarity between the knowledge nodes is calculated. Based on the chapter hierarchy information, the chapter parent-child relationship of the knowledge node is obtained, and based on the semantic similarity and the chapter parent-child relationship, the association edge between the knowledge nodes is generated to complete the construction of the knowledge graph; A digital human avatar is generated based on the aforementioned course structure model and 3D facial modeling technology; The speech data is subjected to speech framing and prosodic analysis to obtain semantic features; Based on the large language model and the course structuring model, teaching intentions are identified, and semantic vectors for each teaching intention are generated. The semantic features, semantic vectors and preset teaching action library are weighted and fused based on the semi-cosine curve function to generate a digital human expression sequence and teaching action sequence that are synchronized with the current teaching intention and semantic features. The 3D teaching scene model is loaded into the 3D rendering engine; the generated digital human image, corresponding expression sequence, and teaching action sequence are jittered into the scene in sequence. The camera path and lens switching strategy in the scene are automatically configured according to the chapter order in the course structure model and the association logic of each knowledge node in the knowledge graph; at each knowledge node or chapter node, the corresponding 3D teaching scene model is automatically triggered to perform local interactive display, and the demonstration duration and lens focus are dynamically adjusted through the node weight information in the graph; the rendering engine synthesizes all chapter fragments into a high-definition video with subtitles and interactive annotations, and outputs a complete digital course file.

2. The digital course generation method based on a large language model as described in claim 1, characterized in that, After the knowledge graph construction is completed, the following is also included: Collect question-and-answer records from users' learning behavior data, and classify the question-and-answer records according to knowledge nodes to obtain question-and-answer data; Based on the question-and-answer data, the access frequency, answer accuracy rate and average response time of each knowledge node are statistically analyzed to obtain statistical results; Based on the statistical results, the weights of each knowledge node and associated edge in the knowledge graph are dynamically updated using a preset weighted fusion model to dynamically optimize the knowledge graph.

3. A digital course generation system based on a large language model, used to implement the digital course generation method according to any one of claims 1-2, characterized in that, include: Data aggregation layer, intelligent agent training layer, digital human interaction layer, and application service layer; The data aggregation layer is used to acquire teaching materials, perform chapter splitting and semantic analysis on the teaching materials based on a large language model, acquire chapter-level information of the teaching materials, and construct a multimodal index based on the chapter-level information to generate a course structured model. The intelligent agent training layer is used to extract features from the course structure model based on the large language model, obtain semantic feature vectors, and construct a course knowledge graph based on the semantic feature vectors and the chapter level information. The digital human interaction layer is used to generate a digital human image based on 3D facial modeling technology, and to generate digital human expression sequences and teaching action sequences based on the digital human image, preset voice data, the course structure model, and preset teaching action library. The application service layer is used to embed the digital human into a 3D teaching scene model based on the digital human image, the digital human expression sequence, and the teaching action sequence, and to synthesize a digital course based on the 3D teaching scene model, the course structure model, and the course knowledge graph.

4. The digital course generation system based on a large language model as described in claim 3, characterized in that, The teaching materials include text materials, audio materials, and 3D teaching materials. The data aggregation layer is used to acquire the teaching materials, perform chapter segmentation and semantic analysis on the teaching materials based on a large language model, obtain the chapter-level information of the teaching materials, and construct a multimodal index based on the chapter-level information to generate a course structured model, including: Obtain teaching materials, perform chapter segmentation on the text materials based on the large language model, and obtain a list of chapter text segments of the teaching materials; Based on the large language model, semantic analysis is performed on the list of chapter text segments segment by segment to obtain the keywords and semantic summaries of each chapter text segment of the teaching materials, and chapter hierarchy information is generated based on the keywords and semantic summaries of the chapter text segments and the list of chapter text segments. The chapter-level information is numbered, and a multimodal index entry is constructed based on the numbering between the chapter-level information and text materials, audio materials, and 3D teaching materials; Based on the chapter hierarchy information and multimodal index entries, a course structure model is constructed.

5. A digital course generation system based on a large language model as described in claim 4, characterized in that, The intelligent agent training layer is used to extract features from the course structured model based on the large language model, obtain semantic feature vectors, and construct a course knowledge graph based on the semantic feature vectors and the chapter-level information, including: Based on the large language model, feature extraction is performed on the course structure model to obtain semantic feature vectors; Knowledge nodes are created based on the semantic feature vectors and the chapter hierarchy information, and the semantic similarity between the knowledge nodes is calculated. Based on the chapter hierarchy information, the chapter parent-child relationship of the knowledge node is obtained, and the association edge between the knowledge nodes is generated based on the semantic similarity and the chapter parent-child relationship, thus completing the construction of the knowledge graph.

6. The digital course generation system based on a large language model as described in claim 5, characterized in that, The digital human interaction layer is used to generate a digital human image based on 3D facial modeling technology, and to generate digital human expression sequences and teaching action sequences based on the digital human image, preset voice data, the course structure model, and a preset teaching action library, including: A digital human avatar is generated based on the aforementioned course structure model and 3D facial modeling technology; The speech data is subjected to speech framing and prosodic analysis to obtain semantic features; Based on the large language model and the course structuring model, teaching intentions are identified, and semantic vectors for each teaching intention are generated. The semantic features, semantic vectors, and preset teaching action library are weighted and fused based on the half-cosine curve function to generate a digital human expression sequence and teaching action sequence that are synchronized with the current teaching intention and semantic features.

7. A digital course generation system based on a large language model as described in claim 5, characterized in that, The agent training layer is also used for: Collect question-and-answer records from users' learning behavior data, and classify the question-and-answer records according to knowledge nodes to obtain question-and-answer data; Based on the question-and-answer data, the access frequency, answer accuracy rate and average response time of each knowledge node are statistically analyzed to obtain statistical results; Based on the statistical results, the weights of each knowledge node and associated edge in the knowledge graph are dynamically updated using a preset weighted fusion model to dynamically optimize the knowledge graph.

Images