A learning content self-adaptive planning method and system based on AIGC and a knowledge graph

Abstract

The application relates to the technical field of intelligent education, and discloses a learning content self-adaptive planning method and system based on AIGC and a knowledge graph, wherein the method comprises the following steps: constructing a knowledge graph represented by a directed acyclic graph; outputting the mastering probability of each knowledge point through a deep knowledge tracking model; determining a target knowledge point based on a graph structure and the mastering probability; calculating a cause-effect influence value of a preceding knowledge point and generating a recommended explanation; determining a content generation mode according to a cognitive load coefficient and a knowledge point out-degree; calling a large language model to generate learning content; and performing forward checking and triggering reverse correction to update the knowledge graph structure after multiple failures of cross-order reference.

Description

Technical Field

[0001] This invention relates to the field of intelligent education technology, specifically to an adaptive planning method and system for learning content based on AIGC and knowledge graphs. Background Technology

[0002] As artificial intelligence technology is increasingly applied in education, intelligent learning systems centered on large language models are gradually being used to automatically generate personalized learning content for learners. Existing systems typically directly call large language models to generate explanations for specific knowledge points and rely on knowledge tracking models to estimate learners' mastery levels to assist in content delivery decisions.

[0003] However, the above approach suffers from the following technical problems. First, when explaining a knowledge point, large language models tend to spontaneously introduce higher-order terms semantically related to that knowledge point, resulting in generated content exceeding the learner's current cognitive level. Existing systems lack systematic and structured verification methods for such out-of-order references. Second, whether the model-generated content fully covers the necessary attributes of the target knowledge point also lacks a quantifiable verification process. Third, when large language models continuously introduce the same higher-order terms in multiple generations of the same knowledge point, existing systems lack a way to transform this signal into a basis for updating the knowledge graph structure. This prevents the knowledge graph from self-correcting structural defects exposed during the generation process, leading to the recurring problem of out-of-order references. These problems collectively restrict the consistency between the content pushed by the intelligent learning system and the learner's cognitive level at the knowledge structure level. Summary of the Invention

[0004] This invention provides a learning content adaptive planning method and system based on AIGC and knowledge graphs, which solves the technical problems in related technologies such as the lack of causal interpretability in learning content recommendations, the risk of generating illusory content, and the inability to adaptively correct the knowledge graph structure.

[0005] In a first aspect, this invention discloses an adaptive planning method for learning content based on AIGC and knowledge graphs, comprising the following steps:

[0006] Obtain the set of knowledge points and prerequisite dependencies in the subject area, and generate a knowledge graph in the form of a directed acyclic graph. Each node carries a set of necessary attributes and a list of prohibited terms.

[0007] Input the learner’s historical question-answering sequence into the deep knowledge tracking model and output the mastery probability of each knowledge point;

[0008] Based on the knowledge graph and the mastery probability of each knowledge point, knowledge points whose mastery probability is lower than the mastery threshold and whose mastery probability of all direct prerequisite knowledge points is not lower than the mastery threshold are selected to form a set of candidate target knowledge points, and the target knowledge points for the current learning are determined.

[0009] For each direct preceding knowledge point of the target knowledge point, calculate its causal impact value on the probability of the target knowledge point being recommended, and identify the preceding knowledge points whose causal impact value exceeds the preset threshold as key bottlenecks.

[0010] The structured causal information of key bottlenecks is filled into a preset explanation template, input into a large language model, and language polishing is performed under the constraint that the causal relationship expression cannot be modified, and recommended explanation text is output.

[0011] Based on the combination of learners’ cognitive load coefficient and the output of target knowledge points, a content generation mode is determined, which includes an atomization generation mode and an aggregation generation mode.

[0012] Based on the determined content generation pattern, prompt words are constructed, and a large language model is called to generate learning content for the target knowledge points. The prompt words all contain the constraint of attaching a list of prohibited terms for the target knowledge points as prohibited items.

[0013] The generated learning content is positively validated from three dimensions: completeness of necessary attribute coverage, accuracy of definition, and compliance of cross-level references. A consistency score is calculated. If the score is not lower than the validation threshold, the learning content is pushed to the learner. Otherwise, specific questions are added as feedback to the prompt words to trigger rewriting.

[0014] The system counts the number of times the same target knowledge point is rejected due to out-of-order reference issues in multiple consecutive independent generation processes, and triggers reverse correction when the preset number of rejections is reached.

[0015] Named entity recognition is performed on the continuously rejected generated content, and the intersection of the results is taken to obtain a candidate set of over-level terms. The candidate terms are matched with nodes in the knowledge graph, and candidate dependency edges are added while maintaining the constraints of the directed acyclic graph. The over-level terms are added to the prohibited term list of the target knowledge point, and the content generation process in step 6 is retried with the updated prohibited term list.

[0016] Furthermore, the deep knowledge tracking model is a recurrent neural network model. The input layer receives the learner's historical question-answering interaction sequence. The input at each time step is an encoding vector of the knowledge point identifier and the answer result, where the knowledge point identifier uses one-hot encoding, and the answer result uses 0 / 1 binary encoding. The output layer uses a sigmoid activation function to map each knowledge point output to... During training, the binary cross-entropy loss function is used, and the Adam optimization algorithm is employed for parameter updates.

[0017] Furthermore, it also includes: obtaining multidimensional behavioral indicators of learners in the current learning session, according to the formula. Calculate the cognitive load factor, where Cognitive load coefficient, This is the normalized value of the average reaction time for each question. The normalized value of the cumulative duration of the current session. This is the normalized value of the current session error rate. , , For the corresponding weight coefficients and satisfying The normalization is a linear mapping using the minimum and maximum values ​​of each indicator in historical session data as the lower and upper bounds.

[0018] Furthermore, the calculation of the causal impact value includes: while keeping the mastery probability of other knowledge points unchanged, temporarily setting the mastery probability of prerequisite knowledge points to a preset high value; re-executing the screening logic under the mastery probability distribution in the counterfactual scenario, performing Monte Carlo sampling on the mastery probability of other prerequisite knowledge points within the estimated uncertainty range, and statistically analyzing the frequency of recommendation of the target knowledge point to obtain the counterfactual recommendation probability. The true recommendation probability is obtained using the same method under the mastery probability distribution in real-world scenarios. According to the formula Calculate the causal impact value, where This represents the causal influence of prerequisite knowledge points on the target knowledge point.

[0019] Furthermore, when the cognitive load coefficient is greater than the first threshold and the out-degree of the target knowledge point is greater than the second threshold, the atomization generation mode is selected; otherwise, the aggregation generation mode is selected. In the atomization generation mode, the prompt words are forced to introduce at most one new term that the learner has not yet encountered in each sentence, and it is forbidden to cite knowledge point terms that are located after the topological level of the target knowledge point and have no direct dependency relationship with the target knowledge point. In the aggregation generation mode, constraints one and two are canceled.

[0020] Furthermore, for each verification dimension, if it passes, the score for that dimension is 1; otherwise, it is 0. The consistency score is the weighted average of the scores of the three dimensions. When the consistency score is less than the verification threshold, the rewriting process is triggered. The maximum number of rewrites is a preset upper limit. If the consistency score is still lower than the verification threshold even after reaching the upper limit, the generation is recorded as a failure and the process enters the reverse correction statistics step.

[0021] Furthermore, after the content generation step outputs the generated content text and before the forward verification step is executed, an illusion detection process based on structural perturbation and semantic sensitivity response is also included, including:

[0022] Extract the local subgraph corresponding to the target knowledge point from the knowledge graph. The local subgraph includes the target node, its set of directly preceding knowledge nodes, its set of directly succeeding knowledge nodes, and the dependency edges and attribute annotations between nodes. Generate structured subgraph description text according to a preset text serialization template.

[0023] Multiple sets of preset perturbation operations are performed on the structured subgraph description text. The preset perturbation operations include node replacement perturbation, dependency edge direction perturbation, and attribute deletion perturbation, which are then summarized to form a perturbation subgraph description set.

[0024] The structured subgraph description text and the perturbation subgraph description set are respectively constructed as learning content generation prompts. A large language model is then called to obtain the original generated content and the corresponding generated content for each perturbation. The generated content of each segment is input into a semantic embedding model to obtain semantic vectors, which are then processed according to the formula... Calculate the cosine distance between the original generated content and each perturbed generated content, where For semantic distance, The semantic vector of the original generated content. For the first The semantic vector of the generated content corresponding to the group perturbation;

[0025] Semantic distances are grouped according to perturbation operation type, and the mean and standard deviation are calculated to generate sensitivity response curves;

[0026] The sensitivity response curve is input into the hallucination discrimination module, which outputs hallucination risk markers and location results of suspicious sentence segments. Based on the location results of suspicious sentence segments, standard attribute data of relevant nodes are extracted from the knowledge graph, correction prompts are injected in the form of counterexample comparison, targeted rewriting is performed on suspicious sentence segments, and the verification learning content is output and sent to the positive verification step.

[0027] Furthermore, the hallucination discrimination module outputs the judgment result in the following manner: if the average semantic distance corresponding to the dependency edge direction perturbation type is less than the preset sensitivity threshold, then a hallucination risk label is output; if the average semantic distance corresponding to the dependency edge direction perturbation type is less than the average semantic distance corresponding to the attribute deletion perturbation type, then a hallucination risk label is output; when either of the above two judgment rules is met, the original generated content is segmented according to the sentence boundary, and the similarity change between the semantic vector of each sentence and the semantic vector of the sentence corresponding to each perturbation generated content under the dependency edge direction perturbation type is calculated. Sentences whose absolute value of the similarity change is lower than the sentence-level sensitivity threshold are marked as suspicious segments.

[0028] Furthermore, during the targeted rewriting process, the start and end positions of suspicious sentence segments are marked in the correction prompts, and a constraint instruction is added requiring the large language model to output only the rewritten suspicious sentence segment text. The system then replaces the rewritten sentence segments back to the corresponding positions of the original generated content and splices them together to generate the verification learning content. When calculating the semantic vector, the semantic embedding model segments each text segment according to the sentence boundaries, and uses the mean of the sentence-level semantic vectors as the overall semantic vector representation of the text segment.

[0029] Secondly, this invention discloses a learning content adaptive planning system based on AIGC and knowledge graphs, used to execute the above-mentioned learning content adaptive planning method, including:

[0030] The knowledge graph module is used to obtain the set of knowledge points and prerequisite dependencies in a subject area, generate a knowledge graph in the form of a directed acyclic graph and maintain the structured attributes of each node;

[0031] The mastery assessment module is used to input learners' historical question-answering interaction sequences into the deep knowledge tracking model and output the mastery probability of each knowledge point.

[0032] The target selection module is used to filter the set of candidate target knowledge points based on the knowledge graph and the mastery probability, and to determine the target knowledge points for the current learning session.

[0033] The causal analysis module is used to calculate the causal impact of prerequisite knowledge points on target knowledge points and identify key bottlenecks.

[0034] The explanation generation module is used to fill the explanation template with structured causal information and call the large language model to polish it under the constraint that the causal structure cannot be modified.

[0035] The mode selection module is used to determine the content generation mode based on a combination of cognitive load coefficient and the output degree of target knowledge points.

[0036] The content generation module is used to construct prompt words based on the determined content generation pattern and call the large language model to generate learning content;

[0037] The positive validation module is used to perform validation and calculate consistency scores from three dimensions: completeness of necessary attribute coverage, accuracy of definition, and compliance of out-of-order references.

[0038] The reverse correction module is used to count the number of consecutive out-of-order reference rejections, triggering named entity recognition, node matching, addition of candidate dependency edges, and updating of the list of disabled terms.

[0039] Compared with the prior art, the beneficial technical effects of this invention application are as follows:

[0040] First, by introducing positive verification based on knowledge graph structured attributes into the large language model generation process, the generated content is systematically compared from three dimensions: completeness of necessary attribute coverage, accuracy of definition, and compliance of cross-level references. This allows structural deviations in the generated content to be quantitatively identified and corrected through a rewriting process that carries specific problem information, thereby ensuring that the content pushed to learners is consistent with their current cognitive level at the knowledge structure level.

[0041] Second, the causal relationship of the recommendation explanation is directly calculated from the knowledge graph structure and mastery probability data. The large language model only plays an auxiliary role in language polishing, which makes the causal basis of the recommendation explanation have traceable data sources and reduces the risk of introducing illusory content due to the autonomous reasoning of the large language model.

[0042] Third, by using a reverse correction process, the out-of-order term signals that repeatedly appear during the multiple generation of the large language model are transformed into candidate dependency edges to supplement the knowledge graph, and the list of prohibited terms is updated simultaneously. This allows the knowledge graph to be incrementally improved by utilizing the structural defect signals exposed during the generation process. Subsequent content generation for the same knowledge point will no longer trigger the same out-of-order reference problem under the constraint of the list of prohibited terms, forming a closed-loop correction process in which generation failure signals drive knowledge graph updates, and knowledge graph updates improve the quality of subsequent generation.

[0043] Fourth, the semantic sensitivity response process based on structural perturbation transforms the judgment of the correctness of the dependency relationship of the generated content into a quantifiable structural consistency test. This process relies only on the semantic response of the knowledge graph topology and the large language model, without relying on the coverage of the pre-defined standards. It will not misjudge the correct explanations of semantically equivalent but different expressions. At the same time, it can locate the directional error of the dependency relationship to a specific sentence segment and correct it through targeted rewriting. Attached Figure Description

[0044] Figure 1 This is a flowchart of the adaptive planning method for learning content provided in an embodiment of the present invention;

[0045] Figure 2 This is a schematic diagram comparing the mastery probability and the attainment threshold of each knowledge point provided in the embodiments of the present invention;

[0046] Figure 3 This is a schematic diagram of the normalized values ​​of each behavioral indicator of the cognitive load coefficient provided in an embodiment of the present invention;

[0047] Figure 4 This is a schematic diagram of the difficulty level and out-degree distribution of knowledge graph nodes provided in an embodiment of the present invention;

[0048] Figure 5This is a schematic diagram comparing the semantic distance of each perturbation type with the hallucination discrimination threshold provided in the embodiments of the present invention;

[0049] Figure 6 This is a schematic diagram illustrating the effect of a NumPy array on the recommendation probability of a target knowledge point, as provided in an embodiment of the present invention for calculating causal influence values.

[0050] Figure 7 This is a schematic diagram of the positive verification three-dimensional score provided in an embodiment of the present invention;

[0051] Figure 8 This is a schematic diagram illustrating the comparison of consistency scores between two generated contents provided in an embodiment of the present invention;

[0052] Figure 9 This is a schematic diagram illustrating the satisfaction of the candidate target knowledge point screening conditions provided in the embodiments of the present invention;

[0053] Figure 10 This is a structural block diagram of the adaptive planning system for learning content provided in an embodiment of the present invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0055] As shown in the present invention application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" do not specifically refer to the singular and may also include the plural. Generally speaking, the terms "comprising" and "including" only indicate that explicitly identified steps and elements are included, and "multiple" includes one, two, or more, and these steps and elements do not constitute an exclusive list; the method or apparatus may also include other steps or elements.

[0056] While this application makes various references to certain modules of the system according to embodiments of the present invention, any number of different modules can be used and run on user terminals and / or servers. The modules are merely illustrative, and different aspects of the system and method may use different modules.

[0057] This invention application uses flowcharts to illustrate the operations performed by the system according to embodiments of the invention. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously, as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0058] Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely a part of the embodiments of the present invention, and not all of the embodiments of the present invention. It should be understood that the present invention is not limited to the exemplary embodiments described herein.

[0059] It is worth noting that in this invention application, all data acquisition actions are carried out in compliance with the relevant data protection laws and policies of the country where the location is located, and with the authorization granted by the owner of the relevant device.

[0060] Example 1

[0061] In intelligent learning systems, the learning content pushed to learners needs to meet two requirements simultaneously: First, the knowledge structure of the content should match the learner's current cognitive level, avoiding the introduction of higher-order concepts beyond the learner's existing knowledge reserves; second, the knowledge attributes covered by the content should be complete, without omitting the core components of the target knowledge point. Existing systems typically directly call large language models to generate learning content, which has the following problems: On the one hand, the model tends to spontaneously introduce higher-order terms that are semantically related to a certain knowledge point when explaining it, resulting in generated content that exceeds the learner's current cognitive level; on the other hand, there is a lack of systematic and structured verification methods to ensure that the model-generated content fully covers the necessary attributes of the target knowledge point. In addition, when the large language model continuously introduces the same higher-order term in multiple generations for the same knowledge point, existing systems lack a way to transform this signal into a basis for updating the knowledge structure, causing the knowledge graph to be unable to self-correct as structural defects are exposed during the generation process.

[0062] This implementation provides an adaptive planning method for learning content based on AIGC and knowledge graphs. It addresses the aforementioned technical problems by establishing a bidirectional verification and correction process between the knowledge graph and the content generated by the large language model. This implementation runs on a server-side hardware environment equipped with a graph database, a deep learning inference engine, and a large language model calling interface.

[0063] Firstly, reference Figure 1 According to an embodiment of the present invention, an adaptive planning method for learning content includes the following steps:

[0064] Step 1: Obtain the set of knowledge points and their dependencies in the subject area, and generate a knowledge graph, such as... Figure 4 The figure illustrates the distribution of difficulty levels and out-degree of knowledge graph nodes in an embodiment of the present invention, reflecting the structural characteristics of the knowledge graph, specifically including:

[0065] Obtain the set of knowledge points in a subject area and the prerequisite dependencies between these knowledge points using a directed acyclic graph. Knowledge graphs are generated in the form of [formula], where For a set of nodes, Let be a set of directed edges. Each node in Corresponding to one knowledge point, For node sequence number, Each directed edge in Representing knowledge points It is a knowledge point Direct prerequisites, This represents the target node's index. The topological structure of the directed acyclic graph ensures that there are no circular dependencies between knowledge points, providing a legitimate logical basis for subsequent planning and decision-making.

[0066] It should be noted that each node in the knowledge graph... Includes the following structured attributes: knowledge point name Used for referencing in generated content and interpreted templates; Difficulty level It adopts discrete hierarchical or continuous numerical representation; out-degree Defined as a node Number of directed edges originating from; set of necessary attributes Record a list of the core attributes that must be covered in a complete lesson for this knowledge point; a list of prohibited terms. Initially an empty list, it is dynamically updated during subsequent reverse correction to record high-level terms that must not be cited in the generated content for this knowledge point.

[0067] In this embodiment, the initial data source for the knowledge graph can be one of the following methods or a combination thereof: knowledge points and their dependencies are manually labeled by subject area experts according to curriculum standards and textbook systems; or knowledge point entities and relationships are automatically extracted from textbook texts and teaching outlines using natural language processing algorithms such as named entity recognition and relation extraction, and then stored in the database after expert review. Once the initial knowledge graph is completed, it is persistently stored in the form of a graph database and supports dynamic incremental updates triggered by subsequent reverse corrections.

[0068] Step 2, as follows Figure 2 As shown, the comparison between the mastery probability and the attainment threshold for each knowledge point in this embodiment of the invention reflects the differences in learners' mastery levels on different knowledge points. For example... Figure 3As shown in the illustration, the normalized values ​​of each behavioral indicator of the cognitive load coefficient in this embodiment of the invention reflect the cognitive stress status of the learner in the current conversation. Step 2 inputs the learner's historical answer sequence into the deep knowledge tracing model and outputs the mastery probability of each knowledge point, specifically:

[0069] Obtain the learner's historical question-and-answer interaction sequences, with each record consisting of a knowledge point identifier and a question-and-answer result. Input the historical question-and-answer interaction sequences into a deep knowledge tracing model, and output an estimate of the learner's mastery probability for each knowledge point. , , The closer the value is to 1, the higher the level of mastery. This refers to the node number of the knowledge point. The deep knowledge tracking model is a recurrent neural network model. The input layer receives the learner's historical question-and-answer interaction sequence. The input at each time step in the sequence is an encoded vector of the knowledge point identifier and the answer result. The knowledge point identifier is converted into a numerical vector using one-hot encoding, and the answer result is encoded using 0 / 1 binary encoding. The output layer is the mastery probability vector of each knowledge point, and the sigmoid activation function is used to map the output of each knowledge point to... The training interval uses binary cross-entropy as the loss function and the Adam optimization algorithm for parameter updates to determine the target threshold. As the default configuration, when At that time, it was determined that the learner had basically mastered the knowledge points. .

[0070] In this embodiment of the application, in order to quantify the learner's cognitive stress level in the current session, based on step 2, multidimensional behavioral indicators of the learner in the current learning session are also obtained, and the cognitive load coefficient is calculated. The calculation formula is as follows:

[0071]

[0072] in, This is the normalized value of the average reaction time for each question. The normalized value of the cumulative duration of the current session. This is the normalized value of the current session error rate. , , For the corresponding weight coefficients, satisfying The default value is , , , A higher value indicates a heavier cognitive load. The above normalization uses the minimum and maximum values ​​of each indicator in historical conversation data as the lower and upper bounds of the normalization, linearly mapping the original observations to... interval, making , , Weighted summation can only be performed after the dimensions of the three factors are unified. Cognitive load coefficient The calculations are performed in real time before each planning decision to reflect the learner's immediate cognitive state.

[0073] Step 3: Based on the knowledge graph and the mastery probability of each knowledge point, determine the target knowledge points for the current learning session:

[0074] Based on knowledge graph Probability of mastering each knowledge point Select knowledge points that meet the following two conditions to form a set of candidate target knowledge points. :

[0075] Condition 1: Knowledge Points The probability of mastering the skill is lower than the mastery threshold, i.e. ;

[0076] Condition 2: Knowledge Points In the knowledge graph, the mastery probability of all directly prerequisite knowledge points is no less than the mastery threshold, that is, for all points that meet the requirements... nodes All have ,in For knowledge points The direct prerequisite knowledge nodes. This is the sequence number of the preceding node. For nodes The corresponding probability of mastery.

[0077] Figure 9 The illustration shows the satisfaction of the candidate target knowledge point screening conditions in the embodiments of the present invention, demonstrating the combined screening effect of mastery probability threshold and prerequisite knowledge point attainment conditions.

[0078] In the candidate target knowledge point set If the set of candidate target knowledge points is not empty, Determine the learning objectives and knowledge points for the current session. , This refers to the node number of the target knowledge point. If the set of candidate target knowledge points... If the value is empty, it indicates that the learner has already met all the knowledge points they can currently access, and the learner is prompted to proceed to the next learning stage.

[0079] It should be noted that the above set of candidate target knowledge points Prioritizing knowledge points refers to ranking the candidate target knowledge points in a set. In the case of non-empty, consider the difficulty level of the knowledge points. With out The candidate knowledge points are ranked, and those with moderate difficulty and a relatively wide range of application are prioritized as the learning objectives for this session. .

[0080] Step 4: Calculate the causal impact of prerequisite knowledge points on the target knowledge point:

[0081] For target knowledge points Let its set of direct prerequisite knowledge points in the knowledge graph be . .for Each prerequisite knowledge point Perform the following sub-steps to calculate its causal impact value on the probability of the target knowledge point being recommended. :

[0082] Step 401, Construct a counterfactual scenario: While keeping the mastery probability of all other knowledge points unchanged, temporarily change the prior knowledge points... Set the probability of mastery to a high value Default value .

[0083] Step 402, Re-run the target knowledge point selection logic: Under the above counterfactual knowledge probability distribution, re-execute the target knowledge point selection logic in step 3 to calculate the target knowledge points. The probability of being selected into the candidate target knowledge point set and ultimately recommended in a counterfactual scenario. .

[0084] Figure 6 This illustration shows the calculation of causal influence value in an embodiment of the present invention, displaying the impact of changes in the probability of mastering prior knowledge points on the recommendation probability of target knowledge points in the form of a NumPy array.

[0085] It should be noted that the selection logic in step 3 is based on the deterministic judgment of conditions one and two, and under a given probability distribution of mastery, the target knowledge points... Whether it is selected into the candidate target knowledge point set The result is a deterministic 0 or 1. Through the Except The probability of mastering other prerequisite knowledge points is sampled using Monte Carlo methods within its estimated uncertainty range. The screening logic is then repeatedly executed and statistics are compiled. The frequency of recommendations is obtained; similarly, The same Monte Carlo sampling method was also used to statistically analyze the probability distribution under the real-world scenario. The frequency of recommendations was obtained.

[0086] Step 403, calculate the causal impact value:

[0087]

[0088] in, Prerequisite knowledge points For target knowledge points The causal impact value, This represents the probability of a target knowledge point being recommended under the mastery probability distribution in the current real-world scenario. To include prerequisite knowledge points The probability of mastery is temporarily set to The probability that the target knowledge point will be recommended.

[0089] Step 404, Key Bottleneck Determination: If Then the prerequisite knowledge points Determined to affect the target knowledge point The key bottlenecks that are recommended are included in the scope of recommendation explanation generation.

[0090] It should be noted that the causal impact value This indicates that the prerequisite knowledge points are assumed. In a counterfactual situation that has been fully grasped, the target knowledge point The increase in the probability of being recommended relative to the real-world situation. Causal influence value. The larger the value, the more significant the impact of the mastery of the prerequisite knowledge point on whether the target knowledge point can be recommended.

[0091] After identifying the key bottlenecks in step 4, structured causal information is filled into the template and polished by a large language model to generate recommended explanations.

[0092] For all prerequisite knowledge points identified as key bottlenecks, structured information will be used. Fill in the preset explanation template to generate a structured explanation draft containing real causal evidence. Prerequisite knowledge points The name, For target knowledge points The name, This represents the corresponding causal impact value. The structured explanation draft is input into the large language model for natural language polishing. The prompts explicitly include constraints: the large language model is prohibited from independently reasoning, adding, modifying, or deleting any causal expressions during the polishing process; the large language model can only optimize the language expression while preserving the original causal structure. The polished recommended explanation text is then output.

[0093] Step 5: Based on the combination of cognitive load coefficient and the output degree of target knowledge points, determine the content generation mode:

[0094] Based on the learner's current cognitive load coefficient Degree from target knowledge point The combination of conditions determines the content generation mode, where For target knowledge points Corresponding out-degree:

[0095] like and If so, then the atomization generation mode is selected;

[0096] If the above conditions are not met, i.e. or If so, then select the aggregation generation mode.

[0097] It should be noted that when learners have a heavy cognitive load and the target knowledge point has a high degree of probability, the knowledge point itself involves many conceptual branches and has high content complexity. In this case, using high information density aggregated content push can easily aggravate cognitive overload. Therefore, the atomized generation mode is used to reduce the cognitive density of a single content push. Conversely, when learners have a low cognitive load or the target knowledge point has few conceptual branches, the aggregated generation mode is used to improve the learning value and information density of the content.

[0098] Step 6: Construct prompt words based on the generation pattern, and call the large language model to generate learning content for the target knowledge points:

[0099] Based on a defined content generation pattern, prompt words are constructed, and a large language model is invoked to generate target knowledge points. The learning content.

[0100] In atomized generation mode, the prompt words include the following mandatory constraints:

[0101] Constraint 1: Each sentence may introduce at most one new term that the learner has not yet encountered. The scope of learned knowledge points is determined by the probability of mastery in the knowledge graph. Definition of a set of knowledge points;

[0102] Constraint 2: It is prohibited to use terms corresponding to knowledge points in the knowledge graph that are located after the topological level of the target knowledge point and have no direct dependency relationship with the target knowledge point in the generated content;

[0103] Constraint 3: Target knowledge points List of prohibited terms All terms in the prompt are appended to the prompt words, explicitly requiring that the large language model not use the aforementioned terms in the generated content. For target knowledge points The corresponding list of prohibited terms.

[0104] In the aggregation generation mode, prompts allow generated content with high information density, such as comprehensive comparisons, knowledge summaries, and multi-concept association analyses. The prompts still include Constraint 3, but Constraints 1 and 2 are removed. Compliance of out-of-order citations is now checked by subsequent positive validation. The large language model outputs the generated content text, which then enters the subsequent validation process.

[0105] Figure 7 This diagram illustrates the scores for each dimension after performing forward validation according to an embodiment of the present invention. Figure 8 The comparison of consistency scores between the two generated content is shown, demonstrating that positive verification can effectively distinguish between high-quality and low-quality generated content.

[0106] Step 7: Perform a positive check on the generated content and output a consistency score:

[0107] The content generated by the large language model is combined with the target knowledge points in the knowledge graph. The structured attributes are compared, and validation is performed from the following three dimensions:

[0108] Step 701, Necessary Attribute Coverage Integrity Verification: Check whether the generated content completely covers the knowledge graph. The necessary attribute set All core attributes specified in the standard are assessed through keyword matching or semantic similarity calculation to determine whether the generated content adequately describes each necessary attribute. If any necessary attribute is not covered, it is recorded as an attribute missing issue. For target knowledge points The corresponding set of necessary attributes.

[0109] Step 702, Definition Accuracy Verification: Check whether the definitions of the target knowledge points and related concepts in the generated content are consistent with the pre-stored standard definitions in the knowledge graph. If there are incorrect definitions that contradict the standard definitions, they are recorded as definition errors.

[0110] Step 703, Over-level reference compliance verification: Check whether the generated content references higher-level concepts that are beyond the learner's current cognitive level, that is, terms corresponding to higher-level knowledge points in the knowledge graph that are not yet within the scope of the learner's mastered knowledge points. If so, it is recorded as an over-level reference problem.

[0111] Based on the verification results of the above three dimensions, the consistency score is calculated. For each dimension, if the dimension passes validation, the score for that dimension is 1; otherwise, it is 0; consistency score. This is the weighted average of the scores across the three dimensions, with each dimension having equal weight by default. ,in , , The scores correspond to three dimensions: completeness of necessary attribute coverage, accuracy of definition, and compliance with out-of-order references. If the verification passes, the generated content will be pushed to the learner; otherwise... If the problem is not resolved, a rewrite process is triggered. The specific issues discovered during this validation are added as structured feedback to the prompt words, and the process returns to step 7, requiring the large language model to make targeted corrections. The maximum number of rewrites is three. If the consistency score is high after three consecutive rewrites... If the result is still below the threshold, record the generation failure and proceed to step 9 to determine whether to trigger reverse correction.

[0112] It should be noted that the consistency score threshold This corresponds to the scenario where at least two of the three dimensions pass the validation. Under the default equal-weight configuration, validation will pass only if the sum of the scores of the three dimensions is not less than 2.4, meaning that at most one dimension is allowed to have a validation problem; if two or more dimensions fail the validation simultaneously, the rewrite process will be triggered.

[0113] Step 8: Count the number of out-of-order reference rejections and determine if reverse correction is triggered.

[0114] Statistics on the same target knowledge points The number of times the forward check is rejected due to out-of-order reference issues in three consecutive independent generation attempts. If all three attempts are rejected due to out-of-order reference issues, the reverse correction process is triggered, and step 9 is executed; otherwise, the reverse correction is not triggered, and the process ends.

[0115] It should be noted that if the large language model continuously introduces the same high-order term in multiple independent generation processes, it indicates that there may be a strong semantic relationship between the term and the target knowledge point in the objective knowledge system. However, the current knowledge graph has not yet recorded this dependency relationship, which is a deficiency in the knowledge graph structure and needs to be supplemented through reverse correction.

[0116] Step 9: Perform reverse correction to update the knowledge graph and the list of prohibited terms.

[0117] After triggering the reverse correction, the following sub-steps are executed:

[0118] Step 901, Over-level term extraction: Perform named entity recognition on the generated content that is rejected three times consecutively, extract the set of terms marked as over-level references by positive verification in each generated content, and take the intersection of the three extraction results to obtain a high-confidence candidate set of over-level terms. The input to the named entity recognition algorithm is the rejected generated content text, and the output is a set of term entities.

[0119] Step 902, Knowledge Graph Node Matching: For the candidate set of advanced terms Each cross-level term in In knowledge graphs Node set Node matching is performed in the process, where Candidate set of terms for exceeding the level Elements in the set are matched using either exact name matching or semantic similarity matching based on word vectors, with the semantic similarity threshold set to 0.85 by default; if in the node set Find with corresponding nodes , If the matched node number is found, proceed to step 903; if no matching node is found, proceed to step 903. Recorded as a candidate for new knowledge points awaiting manual review.

[0120] Step 903, Add candidate dependency edges: For nodes that have been matched with the knowledge graph For terms that are beyond the target knowledge point, check if they exist in the current knowledge graph. To the node The directed edges, i.e., the checks Check if the condition is met; if the directed edge does not exist, and the knowledge graph still satisfies the directed acyclic graph constraint after adding the edge, then automatically add a candidate dependency edge. The edge is then mapped to the knowledge graph and marked as a candidate edge, awaiting manual confirmation by domain experts or verification by a certain amount of subsequent data before being upgraded to a formal edge.

[0121] Step 904, Disable term list update: Regardless of whether node matching was successful in step 902, skipped terms will be removed. Add target knowledge points List of prohibited terms This prevents the recurrence of out-of-order citation issues from the perspective of prompt words.

[0122] After completing the above reverse correction, the content generation process in step 6 is retried with the updated list of prohibited terms.

[0123] It should be noted that this implementation method introduces positive verification based on the structured attributes of knowledge graphs into the process of generating learning content from a large language model. It systematically compares the generated content from three dimensions: completeness of necessary attribute coverage, accuracy of definition, and compliance of cross-level references. This allows structural deviations in the generated content to be quantitatively identified and corrected through a rewriting process, thereby ensuring that the content pushed to learners is consistent with their current cognitive level at the knowledge structure level.

[0124] To address the risk of illusion inherent in large language models during causal explanation generation, this implementation method directly calculates the causal impact value from the knowledge graph structure and mastery probability data. The large language model only plays an auxiliary role in language polishing, and the causal basis is provided by the structured computation results, so that the causal relationship of the recommendation explanation has a traceable data source.

[0125] To address the phenomenon of large language models continuously introducing the same high-order terms in multiple generation processes, this implementation method transforms this signal into a structural update basis for the knowledge graph through a reverse correction process. Candidate dependency edges are added to the knowledge graph, and the list of prohibited terms is updated simultaneously. This allows the knowledge graph to incrementally improve itself by utilizing the structural defect signals exposed during the generation process. Subsequent content generation for the same knowledge point will no longer trigger the same type of out-of-order reference problem under the constraint of the list of prohibited terms. This forms a closed-loop correction process in which generation failure signals drive knowledge graph updates, and knowledge graph updates improve the quality of subsequent generation.

[0126] Secondly, refer to Figure 10 This invention also discloses a learning content adaptive planning system based on AIGC and knowledge graphs, used to execute any of the above-mentioned learning content adaptive planning methods, including: a knowledge graph module, used to acquire a set of knowledge points and pre-dependencies in a subject area, generate a knowledge graph in the form of a directed acyclic graph and maintain the structured attributes of each node; a mastery assessment module, used to input the learner's historical answer interaction sequence into a deep knowledge tracking model and output the mastery probability of each knowledge point; a target selection module, used to select a set of candidate target knowledge points based on the knowledge graph and mastery probability and determine the target knowledge point for the current learning session; and a causal analysis module, used to calculate the causal influence value of pre-existing knowledge points on the target knowledge point and determine the key... The system comprises the following modules: a bottleneck module; an explanation generation module, which fills the explanation template with structured causal information and uses a large language model to refine it under the constraint of prohibiting modification of the causal structure; a mode selection module, which determines the content generation mode based on a combination of cognitive load coefficient and the output degree of the target knowledge point; a content generation module, which constructs prompt words based on the determined content generation mode and uses a large language model to generate learning content; a forward verification module, which performs verification and calculates consistency scores from three dimensions: completeness of necessary attribute coverage, accuracy of definition, and compliance of cross-level references; and a reverse correction module, which counts the number of consecutive cross-level reference rejections and triggers named entity recognition, node matching, addition of candidate dependency edges, and updating of the list of disabled terms.

[0127] Example 2

[0128] In the adaptive planning method for learning content described in Example 1, forward verification compares the generated content with the pre-stored standard definition entries in the knowledge graph. However, the number of pre-stored standard definition entries in the knowledge graph is limited, making it difficult to exhaustively list all semantically equivalent expressions of the same knowledge point in different teaching contexts. This leads to explanations that are semantically correct but whose expression does not completely match the pre-stored definitions being misjudged as incorrect definitions. Furthermore, the illusion generated by the large language model when generating learning content often manifests as implicit alteration of the dependencies between prior concepts in the knowledge graph, such as changing the knowledge point... It is a knowledge point The preceding dependency relationship is implicitly replaced with the opposite dependency direction in the generated content. This type of structural illusion cannot be identified by the definition accuracy check described in step 802 of Example 1.

[0129] Based on Example 1, this implementation method addresses the two types of problems mentioned above in the accuracy verification step of step 7 by introducing an illusion detection process based on structural perturbation and semantic sensitivity response. This transforms the judgment of the correctness of the generated content dependency relationship from the coverage of the pre-stored dependency definition into a quantifiable structural consistency test.

[0130] According to an embodiment of the adaptive planning method for learning content of the present invention, after steps 1 to 7 of Embodiment 1, the method further includes the following steps:

[0131] Step 2-1: Extract local subgraphs of the target knowledge points from the knowledge graph and generate structured subgraph description text:

[0132] After generating content text from the large language model output, it is derived from the knowledge graph. Extract target knowledge points Corresponding local subgraph Local subgraph It contains the following elements: target node ,and The set of directly related prior knowledge nodes ,and The set of directly related successor knowledge nodes This includes the set of dependency edges between the aforementioned nodes and the structured attribute annotations of each node. Based on local subgraphs... The set of nodes, the set of dependent edges, and the attribute annotations are used to generate a structured subgraph description text according to a preset text serialization template. Structured subgraphs describe text Describe the names, attributes, and directions of dependencies between nodes in the local subgraph in a complete natural language format.

[0133] It should be noted that local subgraphs The scope is limited to the direct predecessor and direct successor nodes of the target node, rather than a full graph traversal. The purpose is to control the information density of the structured description text within a range suitable for a single processing of a large language model, while focusing on the topological neighborhood information most directly related to the judgment of the correctness of dependency relationships.

[0134] Step 2-2: Perform multiple sets of preset perturbation operations on the structured subgraph description text to generate a set of perturbed subgraph descriptions.

[0135] Text describing structured subgraphs Execute multiple sets of preset perturbation operations, each set of perturbation operations generates a perturbation subgraph description text, and the results are summarized to form a perturbation subgraph description set. ,in The text describing the perturbation subgraph corresponding to the first group of perturbations. The text describing the perturbation subgraph corresponding to the second group of perturbations. For the first The text describing the perturbation subgraph corresponding to the group perturbation. This represents the total number of perturbation operations. The default perturbation operations include the following three types:

[0136] Type 1 perturbation, node replacement perturbation: in the structured subgraph description text Random selection Some of the preceding knowledge nodes in the knowledge graph are replaced with the names of other knowledge points in the knowledge graph that do not have a direct dependency on the target knowledge point, and node replacement perturbation description text is generated.

[0137] Perturbation type two, edge-direction perturbation: in the structured subgraph description text Select target node For the dependency edges between the node and its direct predecessor, the dependency direction description is adjusted to the opposite direction to generate a dependency edge direction perturbation description text. It should be noted that the dependency edge direction perturbation here is only used to test semantic sensitivity. It is not required that the perturbed subgraph meets the directed acyclic graph constraint. This perturbation description text is only used as input prompt words and does not update the knowledge graph structure.

[0138] Type 3 perturbation, attribute deletion perturbation: in the structured subgraph description text Remove non-critical attribute annotations from each node, retain the node name and dependency description, and generate attribute deletion perturbation description text.

[0139] It should be noted that the node replacement perturbation is used to test whether the content generated by the large language model is sensitive to the specific identity of the prior knowledge points; the dependency edge direction perturbation is used to test whether the content generated by the large language model is sensitive to changes in the direction of the dependency relationship. This type of perturbation directly corresponds to the typical manifestation of implicit structural illusion; the attribute deletion perturbation is used to test whether the content generated by the large language model depends on attribute information rather than simply the dependency relationship structure, serving as a control baseline.

[0140] Steps 2-3 involve calling the large language model based on the original subgraph description text and the perturbed subgraph description set, respectively, to obtain the set of explanation content:

[0141] Describe text using structured subgraphs Each perturbation subgraph description text in the perturbation subgraph description set is used to construct a cue word for learning content. The task instruction part of each cue word is consistent, requiring the large language model to generate target knowledge points based on the provided subgraph description text. The explanation content, specifically the subgraph description text in the prompts, is replaced with the corresponding original or perturbed versions. A large language model is then used to process the above. The prompt words retrieve the original generated content. and the generated content corresponding to each disturbance This will be compiled into a collection of explanatory content. ,in The generated content corresponds to the text describing the original subgraph. This is the generated content corresponding to the first group of perturbations. For the first The generated content corresponding to the group perturbation. The disturbance group number and .in The generated content text output in step 6 of Example 1 is the same object.

[0142] Steps 2-4 involve inputting the set of content to be explained into the semantic embedding model to calculate the semantic distance between the original generated content and each perturbed generated content:

[0143] Each text segment in the content set is input into the semantic embedding model to obtain the semantic vector representation of each segment. The semantic embedding model is a pre-trained language model. The input layer receives the word sequence of natural language text, and the output layer outputs a dense semantic vector of fixed dimensions. Training adopts a contrastive learning approach, with the similarity loss of semantically similar text pairs as the optimization objective. Let... The semantic vector is Disturbance generated content The semantic vector is Both are output by the same semantic embedding model and have the same dimensionality. The cosine distance between the original generated content semantic vector and the generated content semantic vector corresponding to each perturbation is calculated as a semantic distance metric, using the following formula:

[0144]

[0145] in, For the original generated content and the first Group perturbations correspond to the semantic distance between generated content. The larger the value, the more significant the difference between the two pieces of content in the semantic space.

[0146] In this embodiment of the application, in order to improve the stability of semantic distance calculation for short text content, based on steps 2-4, each text segment in the content set is further segmented according to sentence boundaries, and the average semantic vector at the sentence level is used as the overall semantic vector representation of the text segment, replacing the method of directly embedding and encoding the entire text, thereby reducing the impact of text length differences on the semantic vector calculation results.

[0147] Figure 5 The diagram illustrates the comparison between the semantic distance of each perturbation type and the hallucination discrimination threshold in an embodiment of the present invention. It can be seen that the average semantic distance of the perturbation type in the dependent edge direction is significantly lower than that of the attribute deletion perturbation type, triggering the hallucination risk labeling, specifically through the following steps.

[0148] Steps 2-5: Group and statistically analyze semantic distances according to perturbation operation type to generate sensitivity response curves:

[0149] Semantic distances for each group are determined according to the perturbation operation type defined in step 2-2. Grouping. For each type of disturbance. Calculate the statistical distribution of semantic distances corresponding to all perturbation groups under this type. The statistics include the mean. with standard deviation ,in This is a disturbance operation type. Disturbance type The mean of the semantic distance. Disturbance type The standard deviation of the semantic distance. Plot the perturbation operation type on the horizontal axis and the statistical mean of the corresponding semantic distance. Using the vertical axis as the ordinate, generate the sensitivity response curve. .

[0150] It should be noted that the sensitivity response curve This reflects the following pattern: if the large language model correctly understands the dependency structure in the knowledge graph, then when the dependencies in the subgraph describing the text are perturbed, the generated content should exhibit significant semantic changes, i.e., the type of perturbation in the dependency edge direction corresponds to... It should be significantly higher than the attribute deletion perturbation type corresponding to ,in The mean semantic distance corresponding to the perturbation type of the dependent edge direction. Remove the semantic distance mean corresponding to the perturbation type for the attribute; conversely, if Low or with If they are similar, it indicates that the content generated by the large language model is not sensitive to the direction of dependency relationships, and there is a risk of implicit structural illusion.

[0151] Steps 2-6: Determine the risk of hallucination based on the sensitivity response curve and locate suspicious segments:

[0152] Sensitivity response curve Input the hallucination detection module to obtain the hallucination risk assessment result. The hallucination detection module consists of three sub-units: Decision Rule 1, Decision Rule 2, and sentence-level sensitivity localization. The judgment result of the preceding sub-unit determines whether the subsequent sub-unit is triggered. The input of the hallucination detection module is the sensitivity response curve. The output includes hallucination risk markers and location results for suspicious sentence segments. .

[0153] Rule 1, Insensitive Decision: If the semantic distance mean corresponding to the perturbation type of the dependent edge direction is... Below the preset sensitivity threshold (default value) If the content generated by the large language model is not sensitive to perturbations in the direction of dependency, then an illusion risk label will be output.

[0154] Judgment Rule 2, Directional Anomaly Judgment: If If the sensitive direction is opposite to the expected direction, then an illusion risk marker will be output.

[0155] If either of the above two judgment rules is met, the sentence-level sensitivity positioning subunit is triggered: the original generated content is processed. Segment by sentence boundaries, calculate the similarity change between the semantic vector of each sentence and the semantic vector of the sentence corresponding to each perturbation generated under the perturbation type of dependency edge direction, and select sentences with similarity changes below the sentence-level sensitivity threshold. (default value) Sentences marked as suspicious segments are output in the original generated content. Location positioning results If neither of the above two judgment rules is met, then the original generated content is judged to be invalid. There is no risk of structural illusion due to dependency relationships, so proceed directly to step 7 of Implementation Example 1 to perform forward verification.

[0156] It should be noted that in the sentence-level sensitivity localization subunit, the original generated content The semantic vector of each sentence and the perturbation type of the dependency edge direction. Group perturbation generated content The cosine similarity between the semantic vectors of sentences corresponding to the middle position is compared to the similarity between the sentence and the original generated content. The cosine similarity between the semantic vectors of corresponding sentences is the difference, which represents the magnitude of the similarity change. When the absolute value of this difference is lower than the sentence-level sensitivity threshold... When the condition is met, it indicates that the semantics of the sentence have hardly changed after the direction of the dependency relationship is disturbed, and it is judged as a suspicious sentence segment.

[0157] Steps 2-7: Perform targeted rewriting based on the location results of suspicious sentence segments, and output the verified learning content:

[0158] Obtain the location results of suspicious sentence segments The knowledge graph nodes involved in the process, from the knowledge graph Extract the standard attribute data of the above nodes, including the node name. Required attribute set And the direction information of the dependent edges between nodes. The above standard attribute data is injected into the correction prompt words in the form of a counterexample comparison: the correction prompt words also provide structured subgraph description text. The correct dependency representation in the text, as well as the incorrect dependency representation after perturbation of the dependency edge direction, are required to be compared with the correct representation to locate suspicious sentence segments in the large language model. Suspicious segments marked in the code will be rewritten in a targeted manner, and the original generated content must not be modified. The text does not identify any segments as suspicious. A large language model is invoked to process and correct the prompts, outputting the rewritten verification learning content. The learning content will be verified. Step 7 of Example 1 is sent to perform a positive verification, and the subsequent process is the same as steps 8 to 9 of Example 1.

[0159] In this embodiment of the application, to avoid the large language model making collateral modifications to non-suspicious sentence segments during the targeted rewriting process, based on steps 2-7, the start and end positions of suspicious sentence segments are explicitly marked in the correction prompts, and constraint instructions are added to the correction prompts, requiring the large language model to output only the rewritten suspicious sentence segment text, and the system to replace the rewritten sentence segments back to the original generated content. The corresponding positions are spliced ​​to generate verification learning content. Instead of having the entire text rewritten and output by a large language model.

[0160] It should be noted that, addressing the limitation of the accuracy verification in step 802 of Embodiment 1, which relies on the coverage of pre-stored standard definitions, this implementation introduces a semantic sensitivity response process based on structural perturbation. This transforms the judgment of the correctness of the generated content dependency relationships into a quantifiable structural consistency test. Since the calculation of the sensitivity response curve depends only on the topological structure of the knowledge graph and the semantic response of the large language model, and not on the coverage of pre-stored standard definitions, it avoids misjudgment of semantically equivalent but differently expressed correct content, thus overcoming the misjudgment problem caused by the limited number of pre-stored definition entries.

[0161] To address the problem that implicit structural hallucinations are difficult to identify through forward definition comparison, this implementation method relies on the semantic distance mean of edge direction perturbation types. The consistency determination with the expected direction transforms the implicit attribute of whether the large language model correctly understands the direction of dependency into an observable sensitivity directional feature. This allows dependency directional errors that could not be identified by keyword matching or semantic similarity comparison to be located to specific sentence segments. These errors are then corrected through a targeted rewriting process that injects standard attribute data in the form of counterexamples, thereby ensuring the accuracy of the content ultimately included in the learning path planning in terms of the dependency logic of the knowledge graph.

[0162] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A learning content adaptive planning method based on AIGC and knowledge graph, characterized in that, Includes the following steps: Step 1: Obtain the set of knowledge points and prerequisite dependencies in the subject area, and generate a knowledge graph in the form of a directed acyclic graph. Each node carries a set of necessary attributes and a list of prohibited terms. Step 2: Input the learner's historical question-answering sequence into the deep knowledge tracking model and output the mastery probability of each knowledge point; Step 3: Based on the knowledge graph and the mastery probability of each knowledge point, select knowledge points whose mastery probability is lower than the mastery threshold and whose mastery probability of all direct prerequisite knowledge points is not lower than the mastery threshold to form a set of candidate target knowledge points and determine the target knowledge points for this learning session. Step 4: For each direct predecessor knowledge point of the target knowledge point, calculate its causal impact value on the probability of the target knowledge point being recommended, and identify predecessor knowledge points whose causal impact value exceeds the preset threshold as key bottlenecks. Step 5: Based on the combination of learners' cognitive load coefficient and the output of target knowledge points, determine the content generation mode, which includes atomization generation mode and aggregation generation mode. Step 6: Construct prompt words based on the determined content generation pattern, and call the large language model to generate learning content for the target knowledge points. The prompt words all contain the constraint of attaching a list of prohibited terms for the target knowledge points as prohibited items. Step 7: Perform positive validation on the generated learning content from three dimensions: completeness of necessary attribute coverage, accuracy of definition, and compliance of cross-level references. Calculate the consistency score. If the score is not lower than the validation threshold, the learning content will be pushed to the learner. Otherwise, specific issues will be added as feedback to the prompt words to trigger rewriting.

2. The adaptive planning method for learning content according to claim 1, characterized in that, Step 2 further includes: obtaining multidimensional behavioral indicators of the learner in the current learning session, according to the formula. Calculate the cognitive load factor, where Cognitive load coefficient, This is the normalized value of the average reaction time for each question. The normalized value of the cumulative duration of the current session. This is the normalized value of the current session error rate. , , For the corresponding weight coefficients and satisfying The normalization is a linear mapping using the minimum and maximum values ​​of each indicator in historical session data as the lower and upper bounds.

3. The adaptive planning method for learning content according to claim 1, characterized in that, The calculation of the causal influence value in step 4 includes: while keeping the mastery probability of other knowledge points unchanged, temporarily setting the mastery probability of the prerequisite knowledge points to a preset high value; re-executing the screening logic of step 3 under the mastery probability distribution in the counterfactual scenario, performing Monte Carlo sampling on the mastery probability of other prerequisite knowledge points within the estimated uncertainty range, and statistically analyzing the frequency ratio of the target knowledge points being recommended to obtain the counterfactual recommendation probability. The true recommendation probability is obtained using the same method under the mastery probability distribution in real-world scenarios. According to the formula Calculate the causal impact value, where This represents the causal influence of prerequisite knowledge points on the target knowledge point.

4. The adaptive planning method for learning content according to claim 1, characterized in that, In step 5, when the cognitive load coefficient is greater than the first threshold and the out-degree of the target knowledge point is greater than the second threshold, the atomization generation mode is selected; otherwise, the aggregation generation mode is selected. In the atomized generation mode, the prompt words are subject to a mandatory constraint that each sentence can introduce at most one new term that the learner has not yet encountered, and it is prohibited to cite knowledge point terms that are located after the topological level of the target knowledge point and have no direct dependency on the target knowledge point; the constraint is removed in the aggregated generation mode.

5. The adaptive planning method for learning content according to claim 1, characterized in that, In step 7, if each verification dimension passes, the score for that dimension is 1; otherwise, it is 0. The consistency score is the weighted average of the scores of the three dimensions. When the consistency score is less than the verification threshold, the rewriting process is triggered. The maximum number of rewrites is a preset upper limit. If the consistency score is still lower than the verification threshold even after reaching the upper limit, the generation is recorded as a failure.

6. The adaptive planning method for learning content according to claim 5, characterized in that, The system counts the number of times the same target knowledge point is rejected due to out-of-order reference issues in multiple independent generation processes. When the number of rejections reaches a preset number, a reverse correction is triggered. Named entity recognition is performed on the continuously rejected generated content, and the intersection of the results is taken to obtain a candidate set of out-of-order terms. The candidate terms are then matched with nodes in the knowledge graph. Candidate dependency edges are added while maintaining the constraints of the directed acyclic graph. The out-of-order terms are then added to the prohibited term list of the target knowledge point. The updated prohibited term list is used to re-trigger the content generation step.

7. The adaptive planning method for learning content according to claim 1, characterized in that, After the generated content text is output in step 6 and before the positive verification is performed in step 7, an illusion detection process based on structural perturbation and semantic sensitivity response is also included, including: Extract the local subgraph corresponding to the target knowledge point from the knowledge graph. The local subgraph includes the target node, its set of directly preceding knowledge nodes, its set of directly succeeding knowledge nodes, and the dependency edges and attribute annotations between nodes. Generate structured subgraph description text according to a preset text serialization template. Multiple sets of preset perturbation operations are performed on the structured subgraph description text. The preset perturbation operations include node replacement perturbation, dependency edge direction perturbation, and attribute deletion perturbation, which are then summarized to form a perturbation subgraph description set. The structured subgraph description text and the perturbation subgraph description set are respectively constructed as learning content to generate prompt words. The large language model is called to obtain the original generated content and the generated content corresponding to each perturbation. The generated content of each segment is input into the semantic embedding model to obtain semantic vectors, according to the formula. Calculate the cosine distance between the original generated content and each perturbed generated content, where For semantic distance, The semantic vector of the original generated content. For the first The semantic vector of the generated content corresponding to the group perturbation; Semantic distances are grouped according to perturbation operation type, and the mean and standard deviation are calculated to generate sensitivity response curves; The sensitivity response curve is input into the hallucination discrimination module, which outputs hallucination risk markers and location results of suspicious segments. Based on the location results of suspicious sentence segments, standard attribute data of relevant nodes are extracted from the knowledge graph, correction prompts are injected in the form of counterexample comparison, targeted rewriting is performed on the suspicious sentence segments, and the verification learning content is output and sent to step 7.

8. The adaptive planning method for learning content according to claim 7, characterized in that, The hallucination discrimination module outputs the judgment result in the following manner: if the average semantic distance corresponding to the perturbation type of the dependent edge direction is less than the preset sensitivity threshold, then a hallucination risk label is output; if the average semantic distance corresponding to the perturbation type of the dependent edge direction is less than the average semantic distance corresponding to the perturbation type of the attribute deletion, then a hallucination risk label is output. When either of the above two judgment rules is met, the original generated content is segmented according to the sentence boundary, and the similarity change between the semantic vector of each sentence and the semantic vector of the corresponding sentence of each perturbation generated content under the perturbation type of the dependency edge direction is calculated. Sentences whose absolute value of the similarity change is lower than the sentence-level sensitivity threshold are marked as suspicious segments.

9. The adaptive planning method for learning content according to claim 8, characterized in that, During the targeted rewriting process, the start and end positions of suspicious sentence segments are marked in the correction prompts, and a constraint instruction is added requiring the large language model to output only the rewritten suspicious sentence segments. The system then replaces the rewritten sentence segments with the corresponding positions of the original generated content and splices them together to generate the verification learning content. When calculating the semantic vector, the semantic embedding model segments each text segment according to the sentence boundaries and uses the mean of the sentence-level semantic vectors as the overall semantic vector representation of the text segment.

10. A learning content adaptive planning system based on AIGC and knowledge graph, used to execute the learning content adaptive planning method according to any one of claims 1 to 9, characterized in that, include: The knowledge graph module is used to obtain the set of knowledge points and prerequisite dependencies in a subject area, generate a knowledge graph in the form of a directed acyclic graph and maintain the structured attributes of each node; The mastery assessment module is used to input learners' historical question-answering interaction sequences into the deep knowledge tracking model and output the mastery probability of each knowledge point. The target selection module is used to filter the set of candidate target knowledge points based on the knowledge graph and the mastery probability, and to determine the target knowledge points for the current learning session. The causal analysis module is used to calculate the causal impact of prerequisite knowledge points on target knowledge points and identify key bottlenecks. The explanation generation module is used to fill the explanation template with structured causal information and call the large language model to polish it under the constraint that the causal structure cannot be modified. The mode selection module is used to determine the content generation mode based on a combination of cognitive load coefficient and the output degree of target knowledge points. The content generation module is used to construct prompt words based on the determined content generation pattern and call the large language model to generate learning content; The positive validation module is used to perform validation and calculate consistency scores from three dimensions: completeness of necessary attribute coverage, accuracy of definition, and compliance of out-of-order references. The reverse correction module is used to count the number of consecutive out-of-order reference rejections, triggering named entity recognition, node matching, addition of candidate dependency edges, and updating of the list of disabled terms.

Images