Meta-universe-oriented learning guidance method and device, electronic equipment and storage medium

Abstract

The application provides a learning guidance method and device for the metaverse, electronic equipment and storage medium, which deepens the analysis of learning behavior data based on a knowledge enhancement model constructed based on a general large language model to obtain learning semantic states reflecting the cognitive conditions of learners, and combines learning contexts and the learning semantic states to perform joint matching retrieval on first and second layer vector indexes respectively constructed based on concept descriptions and common error examples in a preset teaching knowledge base, and then generates a teaching guidance strategy based on the accurately retrieved target teaching semantic units, thereby comprehensively perceiving the learning situation and dynamic cognitive state of learners and accurately matching multi-dimensional knowledge, and then generating a teaching guidance strategy that is highly targeted and highly consistent with the current cognitive difficulties of learners, thereby effectively improving the accuracy of teaching feedback and the overall effect of learning guidance.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a learning guidance method, apparatus, electronic device, and storage medium for the metaverse. Background Technology

[0002] With the continuous development of intelligent teaching, existing technologies typically employ pre-defined rule mechanisms or static classification models based on a single perspective for instructional interaction. Specifically, existing learning assistance systems mainly collect discrete interaction fragments from learners, classify and discriminate the input data using conventional models, and map them to predetermined task labels or rule bases. When generating instructional feedback, it is usually based on the single question text currently input by the learner, and searches are performed through surface semantics or keyword matching to retrieve pre-set standard answers or fixed error correction prompts from the database to respond to the learner.

[0003] However, the static and fragmented processing method of the aforementioned existing technologies results in only providing single feedback that is corrective after the fact, and it is impossible to provide precise process-oriented teaching guidance strategies for learners. Ultimately, this leads to weak targeting of learning guidance and insufficient rationality of teaching intervention. Summary of the Invention

[0004] This invention provides a learning guidance method, device, electronic device, and storage medium for the metaverse, in order to solve the defects of weak learning guidance and insufficient rationality of teaching intervention in the prior art.

[0005] This invention provides a learning guidance method oriented towards the metaverse, comprising: Obtain learners' learning behavior data; The learning behavior data is input into the knowledge enhancement model to obtain the learning semantic state output by the knowledge enhancement model; the learning semantic state is used to reflect the learner's cognitive status. Using the knowledge enhancement model, the learning context and the learning semantic state in the learning behavior data are used to perform joint matching retrieval corresponding to the first-level vector index and the second-level vector index of the preset teaching knowledge base, respectively, to determine the target teaching semantic unit. Based on the target teaching semantic units, a teaching guidance strategy is generated; The knowledge enhancement model is constructed based on a general large language model; the teaching knowledge base includes multiple teaching semantic units, each of which includes at least a concept description and common error examples; the first-level vector index and the second-level vector index are constructed based on the concept description and the common error examples, respectively.

[0006] According to the present invention, a learning guidance method for a metaverse is provided, wherein the target teaching semantic unit is determined by jointly matching and retrieving the learning context and the learning semantic state in the learning behavior data, respectively, corresponding to the first-level vector index and the second-level vector index of a preset teaching knowledge base, including: Extract the concept description vector of each teaching semantic unit in the first layer vector index, and calculate the first similarity between the learning context and each concept description vector; Extract common error example vectors of each teaching semantic unit in the second-level vector index, and calculate the second similarity between the learning semantic state and each of the common error example vectors; The absolute difference between the learner's current learning difficulty and the corresponding difficulty of each instructional semantic unit is calculated to obtain the difficulty matching degree; The first similarity, the second similarity, and the difficulty matching score are weighted and summed to obtain the retrieval score for each teaching semantic unit; Based on the retrieval score, the target teaching semantic unit is determined from the plurality of teaching semantic units.

[0007] According to a learning guidance method for the metaverse provided by the present invention, the training method of the knowledge enhancement model includes: Construct alignment samples; the input features of the alignment samples include the current learning scenario and the actual learning state, and the output features of the alignment samples include guidance prompts, explanations of principles, and the next action; Calculate the cross-entropy loss between the current output of the general large language model applying the input features and the output features, and calculate the knowledge dependency loss between the current output and the structured teaching knowledge base; Obtain the standard answer output by the large language model without knowledge construction for the input features, calculate the similarity between the current output result and the standard answer, and obtain the answer suppression loss; The model parameters of the general large language model are jointly updated based on the cross-entropy loss, the knowledge dependency loss, and the answer suppression loss to obtain the knowledge enhancement model.

[0008] According to a learning guidance method for the metaverse provided by the present invention, the step of inputting the learning behavior data into a knowledge enhancement model to obtain the learning semantic state output by the knowledge enhancement model includes: The learning intent is obtained by parsing the sequence of operational behaviors in the learning behavior data. Based on the learning intent, the sequence of operational behaviors is mapped to specific teaching knowledge points to obtain concept association results; Based on the learning behavior data, the learner's cognitive motivation is extracted to generate an interpretive cognitive state consisting of cognitive hypotheses, cognitive confidence, and conflicting concepts. The learning semantic state is generated by combining the learning intention, the concept association result, and the interpretive cognitive state.

[0009] According to a learning guidance method for the metaverse provided by the present invention, the step of generating a teaching guidance strategy based on the target teaching semantic unit includes: The learning semantic state, the current teaching objective, and the target teaching semantic unit are used as a common context; The knowledge enhancement model, constrained by teaching roles from multiple teaching perspectives, and based on the common context, performs parallel reasoning to output teaching strategy hypotheses corresponding to each teaching role. By combining the cognitive confidence in the learning semantic state and the current teaching objective, the strategy scores of each teaching strategy hypothesis are negotiated and calculated to obtain the strategy scores of each teaching strategy hypothesis. Based on the strategy score, the teaching guidance strategy is determined from multiple teaching strategy hypotheses.

[0010] According to the learning guidance method for the metaverse provided by the present invention, the teaching strategy is assumed to correspond to different teaching intervention intensities; the multiple teaching positions include at least two of the following: path generation position, diagnostic feedback position, dialogue-inspired position, and developmental assessment position; The path generation stance is used to generate teaching strategy hypotheses that provide the learner with directions of thought or learning path prompts for exploration. The diagnostic feedback stance is used to generate corrective or evaluative teaching strategy hypotheses; The dialogue-inspired stance is used to generate teaching strategy hypotheses for guided questioning in order to obtain the learner's learning ideas and decision-making basis; The developmental assessment stance is used to generate instructional strategy hypotheses that adjust learning strategies and guidance methods.

[0011] This invention also provides a learning guidance device for the metaverse, comprising: Acquire units to obtain learners' learning behavior data; The semantic output unit inputs the learning behavior data into the knowledge enhancement model to obtain the learning semantic state output by the knowledge enhancement model; the learning semantic state is used to reflect the learner's cognitive status. The retrieval unit, through the knowledge enhancement model, utilizes the learning context and the learning semantic state in the learning behavior data to perform joint matching retrieval corresponding to the first-level vector index and the second-level vector index of the preset teaching knowledge base, respectively, to determine the target teaching semantic unit. The guidance unit generates a teaching guidance strategy based on the target teaching semantic unit; The knowledge enhancement model is constructed based on a general large language model; the teaching knowledge base includes multiple teaching semantic units, each of which includes at least a concept description and common error examples; the first-level vector index and the second-level vector index are constructed based on the concept description and the common error examples, respectively.

[0012] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the learning guidance method for the metaverse as described above.

[0013] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the learning guidance method for the metaverse as described above.

[0014] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the learning guidance method for the metaverse as described above.

[0015] The learning guidance method, device, electronic device, and storage medium provided by this invention for the metaverse utilize a knowledge enhancement model built on a general large language model to deeply analyze learning behavior data to obtain learning semantic states that reflect learners' cognitive status. Combining the learning context with these semantic states, the method performs joint matching retrieval of first-layer and second-layer vector indices in a preset teaching knowledge base, constructed based on concept descriptions and common error examples. Based on the precisely retrieved target teaching semantic units, it generates teaching guidance strategies, achieving comprehensive perception of learners' learning situations and dynamic cognitive states, and multi-dimensional accurate knowledge matching. This results in the generation of highly targeted teaching guidance strategies that closely match learners' current cognitive difficulties, effectively improving the accuracy of teaching feedback and the overall effectiveness of learning guidance. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating the learning guidance method for the metaverse provided by this invention. Figure 2 This is a flowchart illustrating the two-layer joint retrieval method provided by the present invention; Figure 3 This is a flowchart illustrating the training method of the knowledge enhancement model provided by the present invention; Figure 4 This is a flowchart illustrating the output method of the teaching guidance strategy provided by the present invention; Figure 5 This is a schematic diagram of the learning guidance system provided by the present invention; Figure 6 A schematic diagram of the structure of the learning guidance device for the metaverse provided by the present invention; Figure 7 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0019] It should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of embodiments of the present invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0020] It should be noted that all actions to obtain information or data in this application are carried out in compliance with the relevant data protection laws and policies of the country where the application is located, and with the authorization of the owner of the relevant device.

[0021] To address the aforementioned issues, this invention provides a learning guidance method oriented towards the metaverse, which enables intelligent learning enhancement by comprehensively perceiving the learning context, dynamically coordinating decision-making, and effectively supporting the learning process, thereby improving the accuracy and immersion of learning guidance. Figure 1 This is a flowchart illustrating the learning guidance method for the metaverse provided by this invention, as shown below. Figure 1 As shown, the method includes: Step 110: Obtain learner learning behavior data.

[0022] Here, "learner" refers to an individual participating in learning interactions within a learning environment such as a metaverse, which could be a student of any age. Additionally, "learning behavior data" refers to a collection of data reflecting learners' various interactions and operations within the learning context.

[0023] Specifically, in real-world scenarios, acquiring learner behavior data can be achieved by combining learner operational behavior data, voice interaction data, and learning trajectory information collected in a 3D scene. For example, operational behaviors can include information such as viewpoint switching, object selection, and object manipulation; voice interaction can include voice or text input; and learning trajectories can include information such as task completion status and dwell time.

[0024] Step 120: Input the learning behavior data into the knowledge enhancement model to obtain the learning semantic state output by the knowledge enhancement model.

[0025] The learning semantic state is used to reflect the learner's cognitive status; the knowledge enhancement model is based on the construction of a general large language model.

[0026] Here, the knowledge enhancement model refers to an intelligent model with reliable professional knowledge and stable reasoning ability, used for semantic interpretation and decision reasoning in the learning process. This knowledge enhancement model is built upon a general-purpose large language model, meaning its underlying foundation uses a general-purpose large language model with powerful generalization and emergence capabilities. It has undergone specific structured knowledge construction and instruction alignment enhancement training tailored to the teaching domain, transforming it from a general-purpose model that simply answers questions directly into a specialized model capable of guiding instruction based on the learning state.

[0027] In addition, the learning semantic state here refers to the vectorized representation generated by the model after semantic parsing the learning behavior sequence, which reflects the learner's cognitive process and level of understanding. For example, it can be a vectorized representation reflecting the learner's cognitive assumptions, uncertainties and conflicting concepts.

[0028] Specifically, learning behavior data can be input into a knowledge enhancement model, which then analyzes the intentions behind the behavior to output a learning semantic state. This allows for a deeper perception and characterization of the learner's cognitive state. For example, the learning semantic state can reflect whether the learner is engaging in exploratory attempts or experiencing cognitive conflict.

[0029] Step 130: Using the knowledge enhancement model, the learning context and the learning semantic state in the learning behavior data are used to perform joint matching retrieval corresponding to the first-level vector index and the second-level vector index of the preset teaching knowledge base, respectively, to determine the target teaching semantic unit.

[0030] The teaching knowledge base includes multiple teaching semantic units, and each teaching semantic unit includes at least a concept description and common error examples; the first-level vector index and the second-level vector index are constructed based on the concept description and the common error examples, respectively.

[0031] Here, the learning context refers to the background environment information related to the current learning task extracted from learning behavior data. The pre-built teaching knowledge base refers to a pre-constructed structured data set containing a large amount of professional knowledge in the teaching domain. It should be noted that the pre-built teaching knowledge base includes multiple teaching semantic units, which refer to multiple knowledge blocks within the pre-built teaching knowledge base divided into the smallest teachable cognitive granularity to avoid semantic fragmentation caused by traditional retrieval. Each teaching semantic unit includes at least a concept description and common error examples. The concept description refers to the basic definition or descriptive text used to explain specific knowledge content to learners; common error examples refer to typical error cases used to provide learners with targeted error correction prompts. In addition, it may also include difficulty levels or teaching strategies.

[0032] In one embodiment, any instructional semantic unit in the instructional knowledge base can be represented as ;in, This represents the nth instructional semantic unit; These are conceptual descriptions used to explain knowledge content; This is the reasoning process used to explain the principles of knowledge to learners; Examples of common errors are provided to offer learners targeted tips. This represents the difficulty level, used to control the depth of instruction. This is a teaching strategy used to indicate a better approach. It should be noted that the construction... At that time, except It can be empty, the rest cannot be empty; except The values ​​are floating-point numbers between 0 and 100; the rest are text. It should be noted that the original teaching materials used to build the teaching knowledge base can be collected from various sources, including textbook texts, transcribed texts of instructional videos, expert explanation strategies, and examples of common learner errors.

[0033] In addition, the first-level vector index and the second-level vector index constitute a two-level retrieval structure of the knowledge base. The first-level vector index is constructed based on the concept description and is used to represent the basic knowledge semantics of the unit. The second-level vector index is constructed based on the common error examples and is used to represent the state semantics related to specific cognitive errors.

[0034] Therefore, the target instructional semantic unit here refers to the instructional semantic unit that best matches the current learning context and learning semantic state after matching and retrieval.

[0035] Specifically, learning context can be extracted from learning behavior data using a knowledge enhancement model, and the structured vectors of the learning context can be used to perform preliminary retrieval and matching against the first-level vector index of the pre-set teaching knowledge base. Simultaneously, the output learning semantic state can be used to match against the second-level vector index of the pre-set teaching knowledge base. Thus, by combining the context's need for basic knowledge with the learning state's need for error correction, a joint matching retrieval is performed, ultimately calculating the retrieval score for each teaching semantic unit, thereby determining the target teaching semantic unit most suitable for the current context.

[0036] It should be noted that, unlike the traditional retrieval method that relies solely on question text for single semantic matching, this embodiment of the invention introduces a joint matching retrieval mechanism with a two-layer vector index. This mechanism combines the learning context with the learning semantic state that reflects the cognitive situation, enabling the accurate retrieval of teaching semantic units that not only meet the current task background but also highly match the learner's immediate cognitive difficulties. This ensures the relevance and targeted nature of subsequent guidance strategies from the underlying data source.

[0037] Step 140: Generate a teaching guidance strategy based on the target teaching semantic unit.

[0038] Here, instructional guidance strategies refer to intervention plans or guidance actions designed to help learners overcome current cognitive obstacles and achieve learning goals.

[0039] Specifically, after identifying the target teaching semantic unit, the teaching guidance content or error correction prompts contained in that unit can be directly obtained and transformed into specific teaching guidance strategies.

[0040] The method provided in this invention uses a knowledge enhancement model built on a general large language model to deeply analyze learning behavior data to obtain a learning semantic state that reflects the learner's cognitive situation. It then combines the learning context with this learning semantic state to perform joint matching retrieval of the first and second layer vector indices in a preset teaching knowledge base, constructed based on concept descriptions and common error examples. Based on the precisely retrieved target teaching semantic units, it generates teaching guidance strategies, achieving a comprehensive perception of the learner's learning context and dynamic cognitive state, and multi-dimensional accurate knowledge matching. This results in the generation of highly targeted teaching guidance strategies that closely match the learner's current cognitive difficulties, effectively improving the accuracy of teaching feedback and the overall effectiveness of learning guidance.

[0041] Based on any of the above embodiments Figure 2 This is a flowchart illustrating the two-layer joint retrieval method provided by the present invention, as follows: Figure 2 As shown, the method includes: Step 210: Extract the concept description vector of each teaching semantic unit in the first layer vector index, and calculate the first similarity between the learning context and each concept description vector.

[0042] Here, the concept description vector of each teaching semantic unit in the first-level vector index refers to the structured high-dimensional vector extracted from the concept description content contained in each teaching semantic unit (TSU) in the pre-set teaching knowledge base through a specific text encoding model. Here, the first similarity refers to a quantitative indicator used to measure the degree of matching between the background of the current learning task and the basic knowledge in the teaching semantic unit.

[0043] Specifically, this can be achieved by extracting the concept description vectors of each teaching semantic unit from the first-level vector index of the pre-defined teaching knowledge base, while simultaneously obtaining the structured vector of the current learning context. Then, by calculating the spatial distance or angle between the structured vector of the learning context and the concept description vector of each teaching semantic unit, the first similarity between the learning context and each concept description vector can be calculated. For example, the cosine similarity calculation method can be used to calculate the cosine value between the learning context vector and the concept description vector as the first similarity.

[0044] In yet another embodiment, the vectorized representation of any instructional semantic unit in the instructional knowledge base is as follows: ; In the formula, The vectorized representation of the i-th instructional semantic unit in the instructional knowledge base; This represents the concept description vector of the i-th instructional semantic unit; The vector representing the reasoning process of the i-th instructional semantic unit; This represents the vector of common error examples for the i-th instructional semantic unit; The instructional strategy vector representing the i-th instructional semantic unit. ； The learnable weights are represented by each vector. This is an Embedding model.

[0045] It should be noted that by calculating the first similarity between the learning context and the concept description vector, retrieval can be performed based on the similarity of teaching functions, accurately matching the knowledge that is suitable for prompting in the current learning scenario, rather than simply relying on literal question retrieval, thus improving the basic relevance of recalled knowledge.

[0046] Step 220: Extract common error example vectors of each teaching semantic unit in the second-level vector index, and calculate the second similarity between the learning semantic state and each of the common error example vectors.

[0047] Here, the common error example vectors for each instructional semantic unit in the second-level vector index refer to structured vectors generated by the coding model based on targeted error examples or cognitive conflict records collected in the instructional semantic unit.

[0048] Specifically, common error example vectors for each instructional semantic unit are extracted from the second-level vector index. Next, using the learning semantic state vector representing the learner's current cognitive state output by the knowledge enhancement model, a second similarity is calculated between the learning semantic state vector and each common error example vector. Similarly, this second similarity can be calculated using a cosine similarity algorithm to determine the degree of fit between the learner's current cognitive state and the common error patterns accumulated in the database.

[0049] It should be noted that by calculating the second similarity between the learned semantic state and the vector of common error examples, the learner's internal cognitive state is directly used as the retrieval basis. This enables the knowledge enhancement model to achieve accurate error correction and guidance based on the learner's dynamic learning state, effectively making up for the shortcomings of traditional methods that cannot intervene in cognitive state.

[0050] Step 230: Calculate the absolute difference between the learner's current learning difficulty and the corresponding difficulty of each teaching semantic unit to obtain the difficulty matching degree.

[0051] Here, "current learning difficulty" refers to the appropriate level of challenge assessed based on the learner's learning progress and historical performance. "Difficulty corresponding to each instructional semantic unit" refers to the difficulty level parameter used to control the depth of instruction, labeled for each instructional semantic unit during the construction of the knowledge base.

[0052] Furthermore, the absolute difference here refers to the result obtained by subtracting the two difficulty values ​​mentioned above and taking the absolute value. Here, difficulty matching refers to an indicator used to evaluate whether the difficulty of the extracted teaching unit matches the learner's current cognitive level.

[0053] Specifically, the system can obtain the learner's current learning difficulty recorded by the system and extract the corresponding pre-set difficulty from each instructional semantic unit. Then, the absolute difference between the current learning difficulty and the corresponding difficulty of each instructional semantic unit is calculated. Furthermore, this absolute difference can be mapped by introducing a specific activation function, such as using a sigmoid function to calculate the absolute difference, directly constructing a difficulty matching function, and finally obtaining the difficulty matching degree to reflect the difficulty fit.

[0054] It should be noted that a difficulty control dimension was introduced during the search process to avoid pushing overly difficult or overly simple teaching content to learners, thereby ensuring the tiered and reasonable nature of the teaching guidance.

[0055] Step 240: The first similarity, the second similarity, and the difficulty matching score are weighted and summed to obtain the retrieval score of each teaching semantic unit.

[0056] Here, the retrieval score refers to the final score that comprehensively evaluates the degree of matching between each instructional semantic unit and the overall current learning context.

[0057] Specifically, the calculated first similarity, second similarity, and difficulty matching score can be obtained separately. A first weight for the first similarity, a second weight for the second similarity, and a third weight for the difficulty matching score are pre-defined. Then, the first similarity is multiplied by the first weight, the second similarity by the second weight, and the difficulty matching score by the third weight. The products of these three factors are then summed to obtain the comprehensive retrieval score for each teaching semantic unit.

[0058] Here, the comprehensive retrieval score between learning behavior data and any teaching unit can be calculated using the following formula, as shown below: ; In the formula, This represents the comprehensive retrieval score between learning behavior data and teaching unit i; They represent Second similarity and difficulty matching The corresponding weighting coefficients; where, Indicates the learning context, Represents the concept description vector of teaching unit i; Represents the learning semantic state; a vector of common error examples for teaching semantic unit i. Match() is a difficulty matching function that can directly construct the learning difficulty of the learner at the current time t. The difficulty level corresponding to this teaching semantic unit The activation function result of the absolute value difference, such as .

[0059] Step 250: Based on the retrieval score, determine the target teaching semantic unit from the plurality of teaching semantic units.

[0060] Specifically, after calculating the retrieval score of each teaching semantic unit, the teaching semantic units are sorted in descending order of retrieval score. Then, based on the maximum value calculation function, a preset number of teaching semantic units with the highest retrieval scores are returned, and these are identified as target teaching semantic units for use in generating specific teaching guidance strategies. Here, the identification of target teaching semantic units can be achieved through the following formula, as shown below: ; In the formula, This represents the top K teaching semantic units with the highest retrieval scores; This indicates that multiple instructional semantic units are sorted in descending order of search scores. That is, the top K instructional semantic units by search score can be used as target instructional semantic units.

[0061] The method provided in this invention calculates a first similarity by matching the concept description with the learning context, calculates a second similarity by matching incorrect examples with the learning semantic state, and calculates a difficulty matching degree by combining the learning difficulty difference. Finally, a retrieval score is obtained by weighted summation to determine the target teaching unit. This achieves multi-dimensional and three-dimensional retrieval based on context, cognitive state, and teaching difficulty, effectively overcoming the limitations of traditional retrieval methods that rely solely on question text. This ensures that the recalled teaching knowledge naturally meets the characteristics of being explanable, suggestible, and error-correctable, significantly improving the accuracy and personalization of teaching guidance strategies.

[0062] Based on any of the above embodiments Figure 3 This is a flowchart illustrating the training method for the knowledge enhancement model provided by this invention, as shown below. Figure 3 As shown, the method includes: Step 310: Construct alignment samples. The input features of the alignment samples include the current learning scenario and the actual learning state, and the output features of the alignment samples include guidance prompts, explanations of principles, and the next action.

[0063] Here, aligned samples refer to training data pairs used for instruction-supervised fine-tuning of the base language model. Here, the current learning context in the input features refers to the specific task context faced by the learner in the interactive environment. Here, the guidance prompts, explanations of principles, and next actions in the output features refer to the expected feedback content structure provided by the model from the teacher's perspective.

[0064] Specifically, data can be extracted from collected expert teaching records and historical interaction logs to construct aligned samples of teaching instructions. The input features are formatted as a combination of the current learning context and a learning state vector reflecting the learner's cognitive state, denoted as input x. Correspondingly, the output features are formatted as a sequence containing prompts, explanations of principles, and the next learning action, denoted as output.

[0065] It should be noted that by constructing aligned samples containing guiding prompts, the conventional large language model directly generates the final answer based on the input. Instead, by standardizing the format of output features, the knowledge enhancement model is forced to learn a dynamic guiding logic based on heuristics and feedback.

[0066] Step 320: Calculate the cross-entropy loss between the current output of the general large language model applying the input features and the output features, and calculate the knowledge dependency loss between the current output and the structured teaching knowledge base.

[0067] Here, cross-entropy loss refers to the basic classification loss that measures the difference between the model's actual output sequence and the labeled output sequence. Knowledge dependency loss refers to a quantitative penalty term used to assess the degree of dependence and fidelity of the model's output content on externally retrieved professional teaching knowledge.

[0068] Specifically, an open-source large-scale model, such as GLM5 or Qwen3.5-9B, can be selected as the base model, and LoRA technology can be used for instruction-supervised fine-tuning. The input feature x is input into the general large-scale language model to obtain the current output result y.

[0069] First, calculate the current output result y and the output features used as labels. Cross-entropy loss between The specific calculation formula can be expressed as: ; Subsequently, to ensure that the output of the knowledge enhancement model is based on evidence, the K most similar instructional semantic units retrieved from the structured instructional knowledge base based on the input features are extracted. And calculate the knowledge dependency loss between the current output and the encoded features of these K teaching semantic units. Here, the specific formula for calculating the knowledge dependency loss is: ; In the formula, This indicates knowledge dependency loss.

[0070] It should be noted that by calculating the cross-entropy loss and knowledge dependency loss separately, not only is the fluency and accuracy of the language generated by the model guaranteed, but the knowledge illusion that is easy to occur during the fine-tuning process is also effectively suppressed, and the knowledge-enhancing model is encouraged to rely heavily on the reliable teaching content provided by the external structured teaching knowledge base.

[0071] Step 330: Obtain the standard answer output by the large language model without knowledge construction for the input features, and calculate the similarity between the current output result and the standard answer to obtain the answer suppression loss.

[0072] The standard answer here refers to the final solution or direct answer given by a standard, un-taught, large language model for the same input features, which can be denoted as... Additionally, the answer suppression loss here refers to a loss function specifically designed to penalize the model for its tendency to directly output the final answer.

[0073] Specifically, the same input features can be fed into a conventional large language model that has not undergone knowledge construction to obtain the standard answer it generates. Then, the current output y of the fine-tuned model is compared with the standard answer. The cosine similarity between the two parameters is used to calculate the answer suppression loss, which is then combined with hyperparameters. The specific calculation formula is as follows: ; In the formula, This indicates that the answer suppresses loss; This is represented as a pre-defined hyperparameter used to control the maximum tolerable answer similarity.

[0074] It should be noted that by introducing a standard answer and calculating the answer suppression loss, the similarity between the current model output and the directly given standard answer is constrained. This prompts the knowledge enhancement model to find a balance during training, ensuring that the output conforms to the correct knowledge direction while being strictly restricted from directly giving the final result. In this way, the model can truly learn to use guided methods to help students think independently.

[0075] Step 340: Based on the cross-entropy loss, the knowledge dependency loss, and the answer suppression loss, jointly update the model parameters of the general large language model to obtain the knowledge enhancement model.

[0076] Specifically, the calculated cross-entropy loss, knowledge dependency loss, and answer inhibition loss are weighted and summed to construct the final guided constraint total loss function, expressed as follows: ; In the formula, This represents the total loss due to guiding constraints; and These represent the weight coefficients corresponding to the answer suppression loss and the knowledge dependence loss, respectively.

[0077] Subsequently, the gradient is calculated based on the total loss L using the backpropagation algorithm, and the parameters of the low-rank matrix attached to the general language model via LoRA are jointly updated. After iterative training and convergence, the model parameters are fixed, ultimately resulting in a knowledge-enhanced model that not only possesses general language capabilities but also deeply understands teaching guidance strategies.

[0078] The method provided in this invention optimizes the parameters of a large language model by combining cross-entropy loss, knowledge dependency loss, and innovative answer suppression loss. While effectively suppressing the model from directly giving the final answer and eliminating knowledge illusion, it forces the knowledge enhancement model to generate heuristic guidance prompts and explanations of principles based on the retrieved structured teaching knowledge. This achieves a key leap from a general question-answering model to a professional knowledge enhancement model with exclusive teaching guidance capabilities.

[0079] Based on any of the above embodiments, step 120 includes: The learning intent is obtained by parsing the sequence of operational behaviors in the learning behavior data. Based on the learning intent, the sequence of operational behaviors is mapped to specific teaching knowledge points to obtain concept association results; Based on the learning behavior data, the learner's cognitive motivation is extracted to generate an interpretive cognitive state consisting of cognitive hypotheses, cognitive confidence, and conflicting concepts. The learning semantic state is generated by combining the learning intention, the concept association result, and the interpretive cognitive state.

[0080] Specifically, firstly, the intent is parsed from the operational behavior sequences in the learning behavior data to obtain the learning intent. Here, the operational behavior sequence refers to the continuous action records extracted from the learning behavior data and statistically encoded according to a sliding window over time. Here, the learning intent refers to the intent classification label assigned by the augmentation model after high-level abstraction of the operational behavior sequence, such as exploratory attempt, verification of local rules, adherence to a false hypothesis, or strategy-free trial and error.

[0081] In practical implementation, the sequence of operational behaviors within the corresponding sliding time window is extracted from the collected learning behavior data, and formally represented as follows: The sequence of actions is then input into a pre-trained knowledge enhancement model. In the first layer, behavioral intent parsing is performed. The knowledge enhancement model outputs the learning intent at the current moment through pattern recognition of continuous behavioral trajectories. The calculation process can be represented as follows: ; In the formula, This represents the learning intent corresponding to the sequence of operational behaviors at time t; This represents the sequence of operational behaviors in the learning behavior data.

[0082] Then, combining the learning intention, the sequence of operational behaviors is mapped to specific teaching knowledge points to obtain the concept association result. Here, the concept association result refers to the structured teaching semantic unit directly corresponding to or involved in the learner's current behavior. In specific implementation, after obtaining the learning intention, the specific operational data at the current moment is combined with the learning intention, and the knowledge enhancement model is used again to call the preset teaching knowledge base for a second layer of concept association mapping. By analyzing the inherent relationship between the action object and the teaching syllabus through the knowledge enhancement model, the corresponding concept association result is output. The calculation process can be represented as follows: ; In the formula, This represents the conceptual association result corresponding to the sequence of operational behaviors at time t; This represents the sequence of operational behaviors in the learning behavior data. It should be noted that by mapping these sequences of operational behaviors to specific teaching knowledge points, precise anchoring between interactive behaviors and the teaching knowledge base is achieved. This ensures that subsequent diagnosis and intervention have clear knowledge direction and prevents guidance from going astray.

[0083] Furthermore, based on learning behavior data, learners' cognitive motivations are extracted to generate an interpretive cognitive state composed of cognitive hypotheses, cognitive confidence, and conflicting concepts. Here, cognitive motivation refers to the internal thought logic driving the learner's current behavior. Cognitive hypotheses refer to the learner's currently actively constructed, potentially biased, understanding logic. Cognitive confidence, or uncertainty, refers to the learner's degree of certainty about their current understanding. Conflicting concepts refer to contradictions or errors in the learner's current cognition. The interpretive cognitive state is a comprehensive interpretive characteristic encompassing all three.

[0084] In practice, the input operation sequence B and the concept association results obtained from the mapping are combined. and learning intentions In the knowledge enhancement model, the third layer of cognitive state analysis is performed. The learner's cognitive motivation is inferred through the knowledge enhancement model, and an interpretive cognitive state is output. The calculation process can be represented as follows: ; In the formula, This represents the interpretative cognitive state at time t. Understandably, in this process, the knowledge enhancement model delves into the contradictions in the learner's mind, using these contradictions as the optimal entry point for instruction.

[0085] Finally, by integrating the learning intention, concept association results, and interpretive cognitive state, a learning semantic state is generated. In practice, the features obtained from the previous layers are combined and encapsulated to construct a unified, structured learning semantic state. Specifically, the concept association results... Explanatory cognitive state and cognitive confidence in the assessed explanatory cognitive states. Combined, they form the semantic state of learning. Therefore, learning semantic states not only includes summarizing past behaviors, but also includes a quantitative confidence level of the current cognitive level.

[0086] The method provided in this invention constructs a three-layer sequential parsing architecture through intent parsing, concept mapping, and deep cognitive extraction. This generates a structured learning semantic state rich in cognitive assumptions and conflicting concepts, deeply analyzing the real cognitive motivations and knowledge blind spots behind learners' interactive behaviors. This enables the system to understand students' thought processes like a real teacher, thus laying a solid and interpretable semantic foundation for subsequently developing highly personalized heuristic intervention strategies that directly address cognitive issues.

[0087] Based on any of the above embodiments Figure 4 This is a flowchart illustrating the output method of the teaching guidance strategy provided by the present invention, as shown below. Figure 4 As shown, the method includes: Step 410: The learning semantic state, the current teaching objective, and the target teaching semantic unit are used as a common context.

[0088] Here, the current learning objective refers to the expected knowledge mastery standard that learners are expected to achieve at the current stage, determined by the course progress and learning phase, and is dynamically updated as the learning progresses. Here, the shared context refers to the data structure that provides unified background knowledge for multi-agent collaborative reasoning, specifically manifested as a shared instructional semantic workspace for multiple agents.

[0089] Specifically, the learning semantic state obtained through parsing and the target teaching semantic unit determined by matching can be extracted, while the current teaching objective recorded in real time can be obtained. These three are integrated into a unified data structure to construct a shared teaching semantic workspace, that is, to construct a common context for subsequent multi-perspective decision-making and reasoning.

[0090] Step 420: Using the knowledge enhancement model, and constrained by teaching roles from multiple teaching perspectives, parallel reasoning based on the common context is used to output teaching strategy hypotheses corresponding to each teaching role.

[0091] Here, the constraint of multiple teaching positions refers to injecting identity instructions or constraint prompts representing different teaching perspectives into a unified underlying model, which can be denoted as... This refers to the teaching role constraint of the i-th teaching position. Furthermore, the teaching strategy hypothesis refers to the set of preliminary intervention plans and teaching explanations generated by the knowledge enhancement model under specific teaching role constraints, representing its unilateral position within the current shared context.

[0092] In practical implementation, multiple virtual agent roles with different functions can be configured on a unified knowledge enhancement model foundation. The knowledge enhancement model can then use a pre-constructed common context as input, combined with the specific constraints of each role, to perform parallel reasoning computations and output the teaching strategy hypotheses corresponding to the teaching role. Each agent... The character modeling reasoning process can be formally represented as: ; After each teaching role performs independent calculations, the teaching strategy hypotheses generated by each agent based on its own functional position are collected, forming a multi-perspective set of decision trajectories. Therefore, the common context at this point can be further represented as... ,in, This represents the common context at time t. This represents the teaching objective at time t. This represents the different teaching strategy hypotheses of different agents regarding the semantic state of learning, based on their own teaching perspectives. ,in, These represent the teaching strategy hypotheses of the path-generating agent, the diagnostic feedback agent, the dialogue-inspired agent, and the developmental evaluation agent regarding the learning semantic state, respectively.

[0093] It should be noted that, under a unified base model, collaborative teaching decisions from multiple perspectives are achieved through role constraint prompts and shared workspaces. This eliminates the need to deploy and call multiple independent language models, thus significantly reducing computational resource consumption while ensuring the diversity and comprehensiveness of teaching decisions.

[0094] Step 430: Combining the cognitive confidence in the learning semantic state and the current teaching objective, negotiate and calculate the strategy scores of each teaching strategy hypothesis to obtain the strategy scores of each teaching strategy hypothesis.

[0095] Here, cognitive confidence refers to a quantitative assessment of whether a learner truly understands the current knowledge point, directly reflecting the stability of their cognition. Negotiation calculation refers to the process of introducing a specific arbitration mechanism to determine and quantify the initiability of interventions that may conflict between different roles. Strategy scoring refers to the probability or score assigned to each candidate strategy hypothesis after comprehensive evaluation, reflecting the rationality of its current implementation.

[0096] In practical implementation, a decision-making mechanism responsible for conflict negotiation, such as the Accessor agent role, is activated. This involves receiving a set of potentially conflicting teaching strategy hypotheses output by the preceding teaching roles, and extracting the cognitive confidence level from the learner's semantic state and the current teaching objective. This is then combined with the learner's cognitive understanding confidence level. With current teaching tasks and objectives Under these conditions, we hypothesize the teaching strategies for each intervention level in the decision space. Each solution is scored independently, and the strategy score is calculated.

[0097] It should be noted that by combining confidence level and objective for negotiation calculation, the intervenability of teaching is determined for conflicting assumptions generated by different teaching positions. This effectively solves the problem of over-intervention or under-intervention caused by direct reliance on a single output in traditional intelligent teaching systems, and ensures a high degree of matching between feedback intensity and cognitive state.

[0098] Step 440: Based on the strategy score, determine the teaching guidance strategy from multiple teaching strategy hypotheses.

[0099] In practice, after completing the negotiation and evaluation, the final strategy score for each candidate teaching behavior type within the decision space is obtained. By finding the maximum score, the intervention behavior type with the highest current score is selected. The execution mechanism can be represented as solving for the action that maximizes the probability distribution. Here, based on the strategy score, the teaching guidance strategy is determined from multiple teaching strategy hypotheses, which can be achieved through the following formula, as shown below: ; In the formula, Indicate the hypotheses of each teaching strategy Strategy score; This represents the cognitive confidence level in the learned semantic state; Indicate teaching objectives; Represents the decision-making space. This represents the decision type corresponding to the highest strategy score; where, As can be seen from the formula, the decision content corresponding to the maximum score will be used as the final teaching guidance strategy to be implemented at the current moment.

[0100] In one embodiment, an Accessor agent can be used to determine the teaching guidance strategy from multiple teaching strategy hypotheses based on strategy scoring. It is understood that the Accessor agent does not simply select candidate strategies, but rather acts as a mechanism to control the level of teaching intervention. The teaching strategy hypotheses generated by different teaching roles essentially correspond to different levels of teaching intervention, and may simultaneously satisfy local rationality but conflict with each other; for example, diagnostic correction and exploratory approaches cannot be simultaneously valid.

[0101] Understandably, by scoring the hypotheses of each teaching strategy, a stable and rational group decision-making path can be formed in a single model reasoning process, avoiding the strategy oscillations caused by multiple rounds of interaction and direct output of the model without intervention, and improving the coherence of teaching intervention.

[0102] The method provided in this invention relies on a unified knowledge enhancement model to construct a shared teaching semantic workspace. By assigning constraints to different roles, it generates multi-dimensional strategy hypotheses through parallel reasoning. It innovatively uses cognitive confidence and teaching objectives to negotiate and score each hypothesis to establish the final strategy. This effectively avoids the perspective limitations of single-model decision-making and the waste of computing power from stacking multiple models. By constructing a negotiation evaluation system, it resolves intervention conflicts under multi-objective teaching positions and generates an optimal teaching guidance strategy that has the characteristics of group rationality and is adapted to learners' cognitive fluctuations in real time. This greatly improves the stability and rationality of intelligent assisted teaching.

[0103] Based on any of the above embodiments, the teaching strategies are assumed to correspond to different levels of teaching intervention; the multiple teaching positions include at least two of the following: path generation position, diagnostic feedback position, dialogue-inspired position, and developmental assessment position.

[0104] Here, the intensity of instructional intervention refers to the level of intervention of instructional actions on learners' autonomous learning process, ranging from level 0 (no intervention) to level 5 (strong intervention), encompassing different levels of action types in the decision-making space, such as non-intervention, questioning, guidance, strategy prompting, and evaluation. Multiple instructional stances refer to a collection of analytical perspectives simulating different teacher roles, specifically including path generation stances, diagnostic feedback stances, dialogue-inspired stances, and developmental assessment stances. In practical applications, at least two of these should be configured to form a multi-perspective synergy.

[0105] In practical implementation, when achieving multi-agent collaboration, at least two virtual agents corresponding to different teaching perspectives can be configured based on the underlying knowledge enhancement model. Each agent is given corresponding functional role constraints, enabling it to perform parallel reasoning according to its own pedagogical perspective when facing the same shared teaching semantic workspace. The teaching strategy hypotheses output by these roles essentially correspond to different levels of teaching intervention, thus forming a set of candidate strategy hypotheses covering different depths of intervention, providing rich material for subsequent conflict negotiation and interventionability determination.

[0106] It should be noted that by setting up multiple teaching positions corresponding to different levels of teaching intervention, the limitation of traditional single-model systems that can only output a single type of feedback is broken. A multi-tiered pool of teaching intervention candidates with diverse intervention granularities is constructed, thus providing the necessary foundation for achieving truly individualized instruction and adaptive feedback.

[0107] The path generation stance is used to generate teaching strategy hypotheses that provide the learner with directions of thought or learning path prompts for exploration. The diagnostic feedback stance is used to generate corrective or evaluative teaching strategy hypotheses; The dialogue-inspired stance is used to generate teaching strategy hypotheses for guided questioning in order to obtain the learner's learning ideas and decision-making basis; The developmental assessment stance is used to generate instructional strategy hypotheses that adjust learning strategies and guidance methods.

[0108] Specifically, regarding the path generation stance, it acts as a guide agent, its core task being to provide directional exploration suggestions when learners are stuck in cognitive impasses or at a loss. In practice, when the underlying knowledge enhancement model receives the constraints of the path generation stance, it analyzes, based on the current cognitive state and learning objectives, what the learner should do next from the current state. Using this reasoning logic, it generates teaching strategy hypotheses that provide learners with directions for exploration or hints for learning paths. For example, it might generate prompts such as guiding the observation of specific local features of a 3D model, or provide alternative strategy branches for solving the current task.

[0109] Understandably, by generating stances through path configuration, the approach avoids directly providing answers when learners encounter difficulties. Instead, by offering scaffolded path guidance, it effectively protects learners' desire to explore and cultivates their ability to explore independently and solve problems.

[0110] Regarding the diagnostic feedback stance, it is similar to playing the role of a diagnostic agent (Critic), focusing on rigorously identifying errors, loopholes, or imperfections in the current learning behavior and cognitive state.

[0111] In practical implementation, when using a diagnostic feedback stance for parallel computation, the knowledge enhancement model focuses on what errors exist in the current interaction data or how it can be improved. Based on this analysis, instructional strategy hypotheses containing corrective or evaluative feedback are directly generated. For example, when it detects that a learner is operating based on a flawed physics assumption, the stance's output will explicitly point out the logical fallacy and provide targeted corrective feedback.

[0112] By configuring a diagnostic feedback stance, learners' cognitive biases and operational errors can be captured and corrected in a timely and sensitive manner, preventing erroneous concepts from becoming entrenched in learners' minds, and fundamentally improving the accuracy of knowledge acquisition and learning efficiency.

[0113] The dialogic heuristic stance is used to develop instructional strategies for generating guided questions, in order to understand learners' learning approaches and decision-making processes. The dialogic heuristic stance corresponds to the role of the Interviewer (Interviewer Agent, a heuristic intelligent agent), aiming to deepen the learning process through Socratic interactive methods such as question-and-answer sessions.

[0114] In practical implementation, when reasoning based on a dialogue-heuristic stance, the core optimization goal of the knowledge enhancement model is to design questions that encourage learners to express their true thoughts. Based on this, guided questions and dialogue interaction content are generated, proactively posing questions to learners. Through this interactive process, learners are prompted to answer, thereby obtaining their learning thought processes and decision-making basis, thus strengthening the entire reflective learning process.

[0115] It should be noted that by configuring dialogue-inspired stances, learners are encouraged to express their implicit thought processes explicitly. This not only helps the system collect deeper cognitive state data more accurately, but more importantly, it promotes learners' self-reflection and the development of metacognitive abilities.

[0116] The developmental assessment stance is used to adjust learning strategies and guidance methods. It corresponds to the role of the assessment agent (Mentor), focusing on macro-level competency development and long-term learning planning.

[0117] In practical implementation, from a developmental assessment perspective, the knowledge enhancement model no longer focuses solely on individual operations or localized errors. Instead, it comprehensively assesses which abilities should be emphasized in the next stage of learning, based on the learner's current progress and the overall course objectives. Based on this long-term perspective, learning strategies and guidance methods are adjusted at a macro level, resulting in hypotheses for adjusting these strategies and guidance methods.

[0118] It should be noted that by configuring a developmental assessment stance, it ensures that every micro-level teaching intervention does not deviate from the macro-level long-term competency development goals, achieving an organic unity between local interactive feedback and overall teaching progress, and preventing the fragmentation of the teaching process.

[0119] The method provided in this invention, based on the setting of educational roles including path generation, diagnostic feedback, dialogue inspiration, and developmental assessment, generates in parallel multiple teaching strategy hypotheses corresponding to different teaching intervention intensities. It achieves a comprehensive simulation of classic teacher support roles in real teaching scenarios, and covers the full-scale teaching intervention needs from micro-level knowledge correction and heuristic questioning to macro-level strategy guidance through a multi-dimensional perspective. It constructs a strategy candidate space with strong educational theory support, and significantly improves the richness and professional level of intervention methods in intelligent learning systems.

[0120] Based on any of the above embodiments Figure 5 This is a schematic diagram of the learning guidance system provided by the present invention, as shown below. Figure 5 As shown, the system mainly consists of a learner interaction module, a knowledge construction and enhancement model, and a core processing engine that includes a teaching semantic understanding module, a multi-agent collaboration module, a guidance and feedback generation module, and a feedback update and adjustment module. Through the collaborative work of these modules, the system as a whole constructs a complete closed loop from data perception, semantic understanding, multi-role decision-making to dynamic feedback. First, the learner interaction module, serving as the system's data input, is responsible for collecting real-time interactive event streams of learners within the metaverse scenario. Each moment's behavior record is encoded into a structured vector, containing operational behaviors (such as perspective switching, object selection and manipulation, path selection, etc.), voice interactions (such as voice or text input), and learning trajectory information (such as task completion status, dwell time, number of repeated operations, etc.). The system statistically analyzes and encodes these behavioral sequences within a sliding time window, outputting abstract learner time-series behavioral records as the foundational data for subsequent state perception and understanding.

[0121] Figure 5The independent knowledge construction and enhancement model on the left is the core foundation of the entire system's intelligent reasoning, providing support for semantic understanding, collaborative decision-making, and feedback generation in the core processing engine on the right. It is worth noting that, to ensure high-quality support across different teaching domains, the system rigorously evaluates and selects the foundation models (knowledge enhancement models) trained using different general-purpose large language models during the offline phase. Unlike traditional natural language processing tasks that rely solely on bilingual evaluation understudy (BLEU) based on n-gram accuracy and recall-oriented understudy for gisting evaluation (ROUGE), a dedicated teaching evaluation metric, PedagogyScore, can also be used. This metric comprehensively evaluates the quality of the prompts provided by the model, whether it demonstrates clear reasoning, and whether it can identify and correct student errors. By assigning different weight parameters, it ensures that the selected knowledge enhancement model not only has clear language expression and complete knowledge points but also possesses excellent teaching capabilities. Here, during the training phase of the knowledge enhancement model, the evaluation function used to evaluate the optimal knowledge enhancement model can be expressed as: ; In the formula, This represents the evaluation score of the knowledge enhancement model; These represent the coefficients of each evaluation indicator. It's understandable that weighting parameters are set. Ensure the model is clearly expressed. Ensure the knowledge points output by the model are complete. To ensure effective model instruction, different weight parameters can be set for different types of tasks. As a teaching evaluation indicator, it can be calculated using the following formula, as shown below: ; In the formula, This indicates the quality of the prompts given by the evaluation model. Assess whether a clear line of reasoning is demonstrated. The evaluation model assesses its ability to identify and correct student errors. It should be noted that the evaluation of the knowledge enhancement model is based on teaching ability indicators, not just language indicators, ensuring that high-quality knowledge construction and enhancement models applicable to instructional guidance are available across different domains.

[0122] Within the core processing engine, the instructional semantic understanding module receives time-series behavioral data from the learner interaction module and invokes the underlying knowledge construction and enhancement model to perform in-depth modeling. This module maps the learner's raw actions into cognitive semantic representations that can be used for instructional reasoning, outputting a structured learning semantic state that includes cognitive motivations, hypotheses, and conflicts, thereby providing accurate semantic input for subsequent intervention decisions.

[0123] Subsequently, the multi-agent collaboration module, based on a unified knowledge construction and enhancement model, receives the learned semantic state and performs parallel analysis and negotiation. This module simulates the collaborative behavior of different teaching support roles in real-world teaching processes, such as the aforementioned guiding, diagnosing, inspiring, and evaluating roles, as well as the Accessor role specifically responsible for conflict negotiation. Within a shared teaching semantic workspace, it generates an optimal teaching guidance strategy that includes action types and strategy assumptions.

[0124] Next, the guidance and feedback generation module generates multimodal guidance information and feedback content that can be directly applied to the metaverse scene based on the determined teaching guidance strategy. Specifically, this module maps the teaching guidance strategy into multimodal guidance signals. For example, this can be achieved through the following formula, as shown below: ; In the formula, The multimodal guidance signal corresponding to the teaching guidance strategy at time t; Indicates the teaching guidance strategy at time t; This represents the semantic state of learning. Among them, instructional guidance strategies... , The type of instructional behavior selected by the Accessor agent in the decision space. Hypotheses about the teaching strategies given for the corresponding agents.

[0125] Understandably, the system employs differentiated feedback strategies for different types of decision-making actions: if the strategy is non-intervention, the system merely maintains environmental records without generating explicit prompts. To maintain learners' space for independent exploration; if intervention is needed, specific guidance methods are provided, including language guidance (such as voice prompts, multi-turn dialogues), scene guidance (such as highlighting, path arrows, gaze guidance, environmental change prompts, etc.), and operation feedback (such as operation correction, constraint prompts, etc.), which are fed back to learners in real time through the scene interface of the metaverse system.

[0126] Finally, the feedback update and adjustment module dynamically updates the learning status and learning objectives based on the learner's response to the guidance and feedback. New behavioral sequences generated by the learner are re-collected by the learner interaction module, prompting the instructional semantic understanding module to recalculate the current state and affecting the entire instructional semantic workspace. Simultaneously, this module also dynamically updates the current learning objectives based on learning progress using the underlying model. The steps for updating the learning objectives can be represented by the following formula, as shown below: ; In the formula, This represents the teaching objective at time t+1; Let represent the teaching objective at time t. This represents the learned semantic state at time t; This represents a knowledge enhancement model. Understandably, this mechanism allows teaching strategies to gradually converge as learners' cognitive states change, achieving continuous optimization of individualized guidance and ultimately forming an efficient and stable perception-decision-feedback-update closed-loop adjustment mechanism within the system.

[0127] Based on any of the above embodiments Figure 6 This is a schematic diagram of the learning guidance device for the metaverse provided by the present invention, as shown below. Figure 6 As shown, the device includes: Obtain Unit 610 to acquire learner learning behavior data; The semantic output unit 620 inputs the learning behavior data into the knowledge enhancement model to obtain the learning semantic state output by the knowledge enhancement model; the learning semantic state is used to reflect the learner's cognitive status. The retrieval unit 630, through the knowledge enhancement model, uses the learning context and the learning semantic state in the learning behavior data to perform joint matching retrieval corresponding to the first-level vector index and the second-level vector index of the preset teaching knowledge base, respectively, to determine the target teaching semantic unit. The guidance unit 640 generates a teaching guidance strategy based on the target teaching semantic unit; The knowledge enhancement model is constructed based on a general large language model; the teaching knowledge base includes multiple teaching semantic units, each of which includes at least a concept description and common error examples; the first-level vector index and the second-level vector index are constructed based on the concept description and the common error examples, respectively.

[0128] The device provided in this invention uses a knowledge enhancement model built on a general large language model to deeply analyze learning behavior data to obtain a learning semantic state that reflects the learner's cognitive situation. It then combines the learning context with this learning semantic state to perform joint matching retrieval of the first and second layer vector indices in a preset teaching knowledge base, constructed based on concept descriptions and common error examples. Based on the precisely retrieved target teaching semantic units, it generates teaching guidance strategies, achieving a comprehensive perception of the learner's learning context and dynamic cognitive state, and multi-dimensional accurate knowledge matching. This results in the generation of highly targeted teaching guidance strategies that closely match the learner's current cognitive difficulties, effectively improving the accuracy of teaching feedback and the overall effectiveness of learning guidance.

[0129] Based on any of the above embodiments, the retrieval unit is specifically used for: Extract the concept description vector of each teaching semantic unit in the first layer vector index, and calculate the first similarity between the learning context and each concept description vector; Extract common error example vectors of each teaching semantic unit in the second-level vector index, and calculate the second similarity between the learning semantic state and each of the common error example vectors; The absolute difference between the learner's current learning difficulty and the corresponding difficulty of each instructional semantic unit is calculated to obtain the difficulty matching degree; The first similarity, the second similarity, and the difficulty matching score are weighted and summed to obtain the retrieval score for each teaching semantic unit; Based on the retrieval score, the target teaching semantic unit is determined from the plurality of teaching semantic units.

[0130] Based on any of the above embodiments, the device further includes a training unit, which is specifically used for: Construct alignment samples; the input features of the alignment samples include the current learning scenario and the actual learning state, and the output features of the alignment samples include guidance prompts, explanations of principles, and the next action; Calculate the cross-entropy loss between the current output of the general large language model applying the input features and the output features, and calculate the knowledge dependency loss between the current output and the structured teaching knowledge base; Obtain the standard answer output by the large language model without knowledge construction for the input features, calculate the similarity between the current output result and the standard answer, and obtain the answer suppression loss; The model parameters of the general large language model are jointly updated based on the cross-entropy loss, the knowledge dependency loss, and the answer suppression loss to obtain the knowledge enhancement model.

[0131] Based on any of the above embodiments, the semantic output unit is specifically used for: The learning intent is obtained by parsing the sequence of operational behaviors in the learning behavior data. Based on the learning intent, the sequence of operational behaviors is mapped to specific teaching knowledge points to obtain concept association results; Based on the learning behavior data, the learner's cognitive motivation is extracted to generate an interpretive cognitive state consisting of cognitive hypotheses, cognitive confidence, and conflicting concepts. The learning semantic state is generated by combining the learning intention, the concept association result, and the interpretive cognitive state.

[0132] Based on any of the above embodiments, the guiding unit is specifically used for: The learning semantic state, the current teaching objective, and the target teaching semantic unit are used as a common context; The knowledge enhancement model, constrained by teaching roles from multiple teaching perspectives, and based on the common context, performs parallel reasoning to output teaching strategy hypotheses corresponding to each teaching role. By combining the cognitive confidence in the learning semantic state and the current teaching objective, the strategy scores of each teaching strategy hypothesis are negotiated and calculated to obtain the strategy scores of each teaching strategy hypothesis. Based on the strategy score, the teaching guidance strategy is determined from multiple teaching strategy hypotheses.

[0133] Based on any of the above embodiments, the teaching strategies are assumed to correspond to different levels of teaching intervention; the multiple teaching positions include at least two of the following: path generation position, diagnostic feedback position, dialogue-inspired position, and developmental assessment position. The path generation stance is used to generate teaching strategy hypotheses that provide the learner with directions of thought or learning path prompts for exploration. The diagnostic feedback stance is used to generate corrective or evaluative teaching strategy hypotheses; The dialogue-inspired stance is used to generate teaching strategy hypotheses for guided questioning in order to obtain the learner's learning ideas and decision-making basis; The developmental assessment stance is used to generate instructional strategy hypotheses that adjust learning strategies and guidance methods.

[0134] Figure 7 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 7As shown, the electronic device may include: a processor 710, a communication interface 720, a memory 730, and a communication bus 740, wherein the processor 710, the communication interface 720, and the memory 730 communicate with each other through the communication bus 740. The processor 710 can call logical instructions in the memory 730 to execute a learning guidance method oriented towards the metaverse. This method includes: acquiring learner learning behavior data; inputting the learning behavior data into a knowledge enhancement model to obtain a learning semantic state output by the knowledge enhancement model; the learning semantic state reflecting the learner's cognitive state; using the knowledge enhancement model, and utilizing the learning context in the learning behavior data and the learning semantic state, performing joint matching retrieval corresponding to the first-level vector index and the second-level vector index of a preset teaching knowledge base to determine the target teaching semantic unit; generating a teaching guidance strategy based on the target teaching semantic unit; the knowledge enhancement model being constructed based on a general large language model; the teaching knowledge base including multiple teaching semantic units, each teaching semantic unit including at least a concept description and common error examples; the first-level vector index and the second-level vector index being constructed based on the concept description and the common error examples, respectively.

[0135] Furthermore, the logical instructions in the aforementioned memory 730 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0136] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the learning guidance method for the metaverse provided by the above methods. The method includes: acquiring learner learning behavior data; inputting the learning behavior data into a knowledge enhancement model to obtain a learning semantic state output by the knowledge enhancement model; the learning semantic state is used to reflect the learner's cognitive situation; using the knowledge enhancement model, and utilizing the learning context in the learning behavior data and the learning semantic state, performing joint matching retrieval corresponding to the first-level vector index and the second-level vector index of a preset teaching knowledge base to determine the target teaching semantic unit; generating a teaching guidance strategy based on the target teaching semantic unit; the knowledge enhancement model is constructed based on a general large language model; the teaching knowledge base includes multiple teaching semantic units, each of which includes at least a concept description and common error examples; the first-level vector index and the second-level vector index are constructed based on the concept description and the common error examples, respectively.

[0137] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the learning guidance method for the metaverse provided by the above methods. The method includes: acquiring learner learning behavior data; inputting the learning behavior data into a knowledge enhancement model to obtain a learning semantic state output by the knowledge enhancement model; the learning semantic state being used to reflect the learner's cognitive status; using the knowledge enhancement model, and utilizing the learning context in the learning behavior data and the learning semantic state, performing joint matching retrieval corresponding to the first-level vector index and the second-level vector index of a preset teaching knowledge base, respectively, to determine the target teaching semantic unit; generating a teaching guidance strategy based on the target teaching semantic unit; the knowledge enhancement model being constructed based on a general large language model; the teaching knowledge base including multiple teaching semantic units, each teaching semantic unit including at least a concept description and common error examples; the first-level vector index and the second-level vector index being constructed based on the concept description and the common error examples, respectively.

[0138] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0139] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0140] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A learning guidance method oriented towards the metaverse, characterized in that, include: Obtain learners' learning behavior data; The learning behavior data is input into the knowledge enhancement model to obtain the learning semantic state output by the knowledge enhancement model; The learning semantic state is used to reflect the learner's cognitive status; Using the knowledge enhancement model, the learning context and the learning semantic state in the learning behavior data are used to perform joint matching retrieval corresponding to the first-level vector index and the second-level vector index of the preset teaching knowledge base, respectively, to determine the target teaching semantic unit. Based on the target teaching semantic units, a teaching guidance strategy is generated; The knowledge enhancement model is constructed based on a general large language model; the teaching knowledge base includes multiple teaching semantic units, each of which includes at least a concept description and common error examples; the first-level vector index and the second-level vector index are constructed based on the concept description and the common error examples, respectively.

2. The learning guidance method for the metaverse as described in claim 1, characterized in that, The step of using the learning context and the learning semantic state in the learning behavior data to perform joint matching retrieval corresponding to the first-level vector index and the second-level vector index of the preset teaching knowledge base, respectively, to determine the target teaching semantic unit includes: Extract the concept description vector of each teaching semantic unit in the first layer vector index, and calculate the first similarity between the learning context and each concept description vector; Extract common error example vectors of each teaching semantic unit in the second-level vector index, and calculate the second similarity between the learning semantic state and each of the common error example vectors; The absolute difference between the learner's current learning difficulty and the corresponding difficulty of each instructional semantic unit is calculated to obtain the difficulty matching degree; The first similarity, the second similarity, and the difficulty matching score are weighted and summed to obtain the retrieval score for each teaching semantic unit; Based on the retrieval score, the target teaching semantic unit is determined from the plurality of teaching semantic units.

3. The learning guidance method for the metaverse as described in claim 1, characterized in that, The training methods for the knowledge enhancement model include: Construct alignment samples; the input features of the alignment samples include the current learning scenario and the actual learning state, and the output features of the alignment samples include guidance prompts, explanations of principles, and the next action; Calculate the cross-entropy loss between the current output of the general large language model applying the input features and the output features, and calculate the knowledge dependency loss between the current output and the structured teaching knowledge base; Obtain the standard answer output by the large language model without knowledge construction for the input features, calculate the similarity between the current output result and the standard answer, and obtain the answer suppression loss; The model parameters of the general large language model are jointly updated based on the cross-entropy loss, the knowledge dependency loss, and the answer suppression loss to obtain the knowledge enhancement model.

4. The learning guidance method for the metaverse according to any one of claims 1 to 3, characterized in that, The step of inputting the learning behavior data into the knowledge enhancement model to obtain the learning semantic state output by the knowledge enhancement model includes: The learning intent is obtained by parsing the sequence of operational behaviors in the learning behavior data. Based on the learning intent, the sequence of operational behaviors is mapped to specific teaching knowledge points to obtain concept association results; Based on the learning behavior data, the learner's cognitive motivation is extracted to generate an interpretive cognitive state consisting of cognitive hypotheses, cognitive confidence, and conflicting concepts. The learning semantic state is generated by combining the learning intention, the concept association result, and the interpretive cognitive state.

5. The learning guidance method for the metaverse according to any one of claims 1 to 3, characterized in that, The generation of teaching guidance strategies based on the target teaching semantic units includes: The learning semantic state, the current teaching objective, and the target teaching semantic unit are used as a common context; The knowledge enhancement model, constrained by teaching roles from multiple teaching perspectives, and based on the common context, performs parallel reasoning to output teaching strategy hypotheses corresponding to each teaching role. By combining the cognitive confidence in the learning semantic state and the current teaching objective, the strategy scores of each teaching strategy hypothesis are negotiated and calculated to obtain the strategy scores of each teaching strategy hypothesis. Based on the strategy score, the teaching guidance strategy is determined from multiple teaching strategy hypotheses.

6. The learning guidance method for the metaverse as described in claim 5, characterized in that, The teaching strategies are assumed to correspond to different levels of teaching intervention; the multiple teaching positions include at least two of the following: path generation position, diagnostic feedback position, dialogue-inspired position, and developmental assessment position. The path generation stance is used to generate teaching strategy hypotheses that provide the learner with directions of thought or learning path prompts for exploration. The diagnostic feedback stance is used to generate corrective or evaluative teaching strategy hypotheses; The dialogue-inspired stance is used to generate teaching strategy hypotheses for guided questioning in order to obtain the learner's learning ideas and decision-making basis; The developmental assessment stance is used to generate instructional strategy hypotheses that adjust learning strategies and guidance methods.

7. A learning guidance device for the metaverse, characterized in that, include: Acquire units to obtain learners' learning behavior data; The semantic output unit inputs the learning behavior data into the knowledge enhancement model to obtain the learning semantic state output by the knowledge enhancement model; the learning semantic state is used to reflect the learner's cognitive status. The retrieval unit, through the knowledge enhancement model, utilizes the learning context and the learning semantic state in the learning behavior data to perform joint matching retrieval corresponding to the first-level vector index and the second-level vector index of the preset teaching knowledge base, respectively, to determine the target teaching semantic unit. The guidance unit generates a teaching guidance strategy based on the target teaching semantic unit; The knowledge enhancement model is constructed based on a general large language model; the teaching knowledge base includes multiple teaching semantic units, each of which includes at least a concept description and common error examples; the first-level vector index and the second-level vector index are constructed based on the concept description and the common error examples, respectively.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the learning guidance method for the metaverse as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the learning guidance method for the metaverse as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the learning guidance method for the metaverse as described in any one of claims 1 to 6.

Images