Large language model enhanced artificial intelligence knowledge adaptive learning planning system

Abstract

The invention relates to a big language model enhanced knowledge adaptive learning planning system, and belongs to the field of intelligent education. The system comprises a knowledge center module, a learner portrait module, a path planning module and an intelligent learning guiding module, and the knowledge center module extracts entities and relationships from a multi-modal data source by using a large language model to construct a knowledge graph; the learner portrait module collects multi-dimensional learning data of the user and maps the multi-dimensional learning data to corresponding nodes of a knowledge graph, and dynamically deduces a learner portrait through a Bayesian knowledge tracking model; the path planning module generates an initial learning path based on the knowledge graph and the learner portrait, establishes a collaborative filtering analysis model, predicts and optimizes the expected effect of the current learner following the initial learning path in combination with a Bayesian knowledge tracking model, and finally generates a target learning path; and the intelligent learning guiding module generates a standardized knowledge card for each knowledge node on the target learning path through a security retrieval enhancement generation technology.

Description

Technical Field

[0001] This application relates to the field of smart education technology, and in particular to an artificial intelligence knowledge adaptive learning planning system enhanced by a large language model. Background Technology

[0002] In the field of AI education, knowledge systems are rapidly updated and highly interdisciplinary, but existing technologies are still based on static, linear knowledge management paradigms, leading learners to face the dual dilemma of "learning too little to apply" and "learning without knowing how to apply." Current technologies suffer from the following four shortcomings: 1. Lagging knowledge updates; 2. Fragmented knowledge; 3. Rigid and opaque path planning; 4. Inaccurate intelligent assistance. These are detailed below: 1. Knowledge updates are severely lagging, failing to keep pace with the explosive evolution of artificial intelligence (AI) technology. Traditional knowledge systems, primarily built manually, have long update cycles, unable to match the 7.3% monthly growth rate of AI technology, resulting in a generational gap between teaching content and industry practice. The construction and updating of static knowledge graphs require significant expert manpower, are cumbersome processes, and typically involve a 3-6 month delay. This makes static knowledge graphs unsuitable for carrying the high-frequency iteration of technical knowledge in the AI ​​field. By 2023, when breakthrough improvements in diffusion models such as Consistency Distillation had become industry standards, the knowledge graphs and course resources in teaching systems remained at the level of early models such as DDPM, causing learners to acquire knowledge that was already outdated as soon as it was mastered.

[0003] 2. Knowledge is fragmented, lacking an understanding of the essence of technological paradigms and cross-domain connections. In pursuit of completeness, current technologies excessively break down technical concepts, disrupting the overall coherence of knowledge and neglecting the highly interdisciplinary connections within the AI ​​field. Mainstream solutions rely on linear prerequisite relationships, such as "learning convolutional neural networks (CNNs) requires mastering linear algebra," but ignore the inherent connections across knowledge stacks—complex relationships such as the collaborative mechanism between reinforcement learning and generative models, and the deep coupling between hardware-aware training and model compression are not effectively modeled. This flattened representation leads to a lack of domain adaptability in learning path planning, particularly hindering the technical migration of practitioners, such as natural language processing (NLP) engineers transitioning to multimodal system development.

[0004] 3. Rigid and opaque path planning fails to provide personalized and reliable guidance. Existing path recommendation technologies rely on simple rules or static profiles, lacking dynamic adjustments and the ability to leverage collective wisdom. Furthermore, the decision-making process is opaque, resembling a "black box." Path generation algorithms typically depend solely on a single prior relationship graph and the user's basic test scores, ignoring the collective success experiences of similar learners and failing to quantify the applicability of a path to a specific individual. For example, mainstream systems recommend identical learning paths to all NLP engineers but cannot intelligently insert key computer vision nodes for learners transitioning to multimodal development. When suggesting "learning graph neural networks first instead of temporal models," they fail to provide convincing data support, such as predicted success rates or estimated durations, leading to low user trust.

[0005] 4. Inaccurate Intelligent Assistance, Unable to Reliably Support Learning Decisions. Some existing technologies have begun using Integrated Large Language Models (LLMs) for automated question answering and content generation. However, the direct and unconstrained use of LLMs introduces serious risks: LLMs can create knowledge illusions, generating seemingly reasonable but factually incorrect answers. In education, this means it may provide incorrect code examples, misunderstand core definitions of technical concepts (such as confusing "attention mechanisms" with "memory networks"), or cite sources from papers that do not exist at all. Secondly, the knowledge of LLMs is limited to their training data and cannot be guaranteed to be synchronized with the latest technological developments (such as framework version updates and API changes). More seriously, the generated content is unpredictable, and the system cannot ensure that its output strictly conforms to the syllabus or academic standards of a specific course. This uncontrolled generation process is like a "black box," impossible to audit and verify. Once incorrect information is provided, it will directly mislead learners and even lead them to introduce fatal errors in project practice, completely failing to meet the high requirements for accuracy and security in educational applications. Summary of the Invention

[0006] This application provides an artificial intelligence knowledge adaptive learning planning system enhanced by a large language model, which can solve the problems of lagging knowledge updates and rigid content in traditional education systems.

[0007] To achieve the above objectives, a first aspect of this application provides a knowledge-adaptive learning planning system enhanced by a large language model. The system includes a knowledge hub module, a learner profiling module, a path planning module, and an intelligent guidance module. The knowledge hub module uses a large language model to extract entities and relationships from multimodal data sources to construct a knowledge graph. The learner profiling module collects multidimensional learning data from users and maps it to corresponding nodes in the knowledge graph, dynamically inferring the learner profile through a Bayesian knowledge tracing model. The path planning module generates an initial learning path based on the knowledge graph and learner profile, establishes a collaborative filtering analysis model, and combines it with the Bayesian knowledge tracing model to predict and optimize the expected effect of the current learner following the initial learning path, ultimately generating a target learning path. The intelligent guidance module generates standardized knowledge cards for each knowledge node on the target learning path using secure retrieval-enhanced generation technology.

[0008] Furthermore, the knowledge graph of the knowledge hub module is a structured knowledge representation, including the relationships between knowledge points. Each node in the knowledge graph represents a knowledge point, and each edge of the knowledge graph represents the relationship between knowledge points. Knowledge graphs are based on graph structures Stored in the form of , where A collection of knowledge points This is a set of relationships between knowledge points, and each edge... express and There exists a type of The relationship.

[0009] Furthermore, the knowledge graph construction process incorporates multiple verification mechanisms, including: A multi-model cross-validation mechanism is set up, and at least two large language models with different architectures are used to generate knowledge entities and relations in parallel. The entities and relations are used to construct the knowledge points and the associations between knowledge points in the knowledge graph. The system labels knowledge associations that are disputed or have low confidence, provides a visual review interface, and establishes an error feedback chain. When users or teachers discover knowledge errors during subsequent use, they can submit correction suggestions through a designated entry point. After receiving the suggestions, the system will restart the verification process to update the knowledge graph.

[0010] Furthermore, the learner profile module supports dynamic feedback correction and establishes a profile deviation feedback channel. Users can raise objections to the mastery status assessed by the system. After receiving the feedback, the system recalculates the mastery probability of the relevant nodes and adjusts the profile.

[0011] Furthermore, the multi-model cross-validation mechanism identifies and corrects potential hallucination content by comparing the responses of different large language models to the same prompt word and combining them with preset knowledge graph factual verification rules. Before the generated content is output, it undergoes a post-processing process, which includes terminology consistency verification and integrity checks. In addition, the system performs a final factual verification to ensure that all generated information can be traced back to the knowledge points in the knowledge graph.

[0012] To achieve the above objectives, a second aspect of this application also provides a planning method for a knowledge adaptive learning planning system enhanced by a large language model. In the learner profiling module, a Bayesian knowledge tracing model is used to dynamically infer the user's understanding of each knowledge point. The hidden mastery state probability is determined; in the path planning module, a personalized preferred path that satisfies knowledge dependencies is generated and optimized through model-based collaborative filtering algorithm and Monte Carlo simulation; in the intelligent learning module, a standardized knowledge card is generated for each knowledge node on the path through secure retrieval enhancement generation technology.

[0013] Furthermore, the learner profiling module uses a Bayesian knowledge tracing model to dynamically infer the user's understanding of each knowledge point. Hidden mastery probability The model includes four core parameters: initial mastery probability. Learning probability Probability of guessing correctly and probability of error ; The update of the mastery probability in the Bayesian knowledge tracing model follows the Bayesian rule:

[0014] in For the observation results, when correct When an error occurs ; After observing one learning attempt, the prior mastery probability for the next time step is:

[0015] in The probability of knowing the state at the current moment. This represents the probability of control at the previous moment. This represents the probability of learning a knowledge point even if one has not mastered it before.

[0016] Furthermore, the path planning module generates an optimal initial path that satisfies knowledge dependencies based on the topology of the knowledge graph and the knowledge points already mastered in the learner profile. The following formula is the specific expression for the optimal initial path:

[0017] in It is the set of all valid paths that satisfy the prerequisites. Factors such as study time and difficulty; The model-based collaborative filtering algorithm uses a matrix factorization model to predict learners. In a specific path Each knowledge node Expected effect score ; The model learns by minimizing the regularized squared error loss function as follows:

[0018]

[0019] in Given a set of known interactions, The global average score and These are biases for learners and knowledge points, used to capture individual and project differences. Regularization coefficients used to control model complexity; After the model is trained, it can predict learners. For path Any unlearned knowledge points Score ; Calculation path Overall predictive mastery ,in This is a set of knowledge points that have already been mastered.

[0020] Furthermore, the path planning module designs Monte Carlo simulations to evaluate the robustness of the path and performs final optimization; For candidate paths ,conduct Each simulation, in which the learner's current state is considered... Based on the historical success rate of knowledge points, a possible learning trajectory is randomly generated, and key indicators such as the estimated total learning time are recorded. and probability of failure This refers to the probability of failure at any critical node along the path; ultimately, a comprehensive path utility function is defined as the final basis for path ranking and optimization:

[0021] in, and It is a normalization factor. Measure the diversity of paths in terms of resource type or learning mode. and These are weighting coefficients, reflecting preferences for different objectives; the system selects the path with the highest utility from the initial path and its variations. Personalized optimal path as the final recommendation .

[0022] Furthermore, the specific process of the security retrieval enhancement generation technology includes: S1, from knowledge graphs Searching for target knowledge nodes All authoritative structured information, including its definitions, key concepts, relationships, or prerequisites; S2. Combine the retrieved structured information with prompt word templates containing strict anti-hallucination instructions to construct a double-constrained structured input. The prompt word template guides the large language model to generate accurate, concise, and academically compliant review content, ensuring that the generated content strictly adheres to the factual basis of the knowledge graph while retaining the generative adaptability of the large language model. S3, Invoke the large language model according to The initial content of the knowledge cards is generated. During the generation process, a multi-model cross-validation mechanism is introduced to automatically detect and handle inconsistencies or errors in the generated results.

[0023] This application constructs an architecture of "authoritative knowledge graph + dynamic RAG expansion + whitelist mechanism," which not only ensures the accuracy of the core knowledge but also flexibly integrates cutting-edge technological developments, effectively overcoming the problems of lagging knowledge updates and rigid content in traditional education systems. Based on the fusion decision-making mechanism of "real-time learner profiling + group collaborative filtering + path simulation optimization," the generated paths not only truly fit individual learners but also clearly present the recommendation basis and adjustment logic, greatly enhancing the transparency of the paths and user trust. Through a four-fold protection system of "structured prompt word constraints + RAG content control + multi-model cross-validation + post-processing verification," the powerful generation capabilities of the Large Language Model (LLM) are strictly limited within the safe range defined by the knowledge graph, fundamentally eliminating the illusion problem and enabling its reliable application in educational guidance. The organic synergy of the four modules constructs a complete closed-loop system of "knowledge construction - state tracking - path planning - intelligent guidance - feedback and error correction," realizing a leap from static resource push to dynamic personalized learning support, significantly improving learning efficiency and experience, and laying the foundation for building a next-generation safe, reliable, and highly personalized intelligent education system. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0025] Figure 1 A schematic diagram of an AI knowledge adaptive learning planning system enhanced by a large language model according to this application; Figure 2 Here is a flowchart illustrating the specific process of the security retrieval enhancement generation technology provided in this application; Figure 3 This is a comparison chart showing the performance of this application and two baseline systems on various evaluation metrics, based on the information provided in this application. Figure 4 A contribution analysis diagram for each module provided in this application. Detailed Implementation

[0026] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0027] like Figure 1 As shown, this application discloses an AI-based knowledge adaptive learning planning system enhanced with a large language model, comprising the following four core modules: 1. Knowledge hub module; 2. Learner profiling module; 3. Path planning module; 4. Intelligent guidance module. Details are as follows: 1. Knowledge Hub Module This module forms the authoritative knowledge foundation of the system, employing a multi-layered verification mechanism to ensure the accuracy and reliability of the knowledge. Specifically, it first utilizes a large language model, under rigorously designed structured prompt word constraints, to extract entities and relationships from multimodal data sources such as authoritative textbooks, course outlines, and academic literature. For each knowledge point, in addition to attribute information, an embedding vector is generated using LLM, laying the foundation for subsequent semantic computation and personalized recommendations. Finally, a semi-automated process is employed to construct a domain knowledge graph. This knowledge graph is a structured knowledge representation, containing the relationships between knowledge points. Each node in the knowledge graph represents a knowledge point, and edges represent the relationships between knowledge points. The knowledge graph uses a graph structure. Stored in the form of , where A collection of knowledge points This is a set of relationships between knowledge points, such as prerequisites, inclusion, and dependencies. Each edge... express and There exists a type of The relationship.

[0028] The knowledge graph construction process incorporates multiple verification mechanisms to ensure knowledge accuracy, specifically including: (a) Multi-model cross-validation: At least two large language models with different architectures are used to generate knowledge entities and relations in parallel, and the entities and relations are used to construct the knowledge points and the associations between knowledge points in the knowledge graph; (b) Dynamic error correction feedback: Mark knowledge associations that are controversial or have low confidence, provide a visual review interface, and establish an error feedback link. When users or teachers discover knowledge errors during subsequent use, they can submit correction suggestions through a designated entry point. After receiving the suggestions, the system will restart the verification process to update the knowledge graph.

[0029] Storing knowledge graphs in a graph database and visualizing them can help teachers and students intuitively view the relationships between different knowledge points.

[0030] 2. Learner Profile Module This module is responsible for constructing a dynamically quantified user cognitive model and employs a multi-calibration mechanism to ensure the accuracy and reliability of the profile, precisely reflecting students' knowledge mastery and learning characteristics. It is the core of achieving personalized adaptation. The learner profile construction process specifically includes: The learner profiling module collects multi-dimensional learning data from users, including learning history, grade records, error type identifiers, and interactive behaviors, and maps them to a knowledge graph. On the corresponding node, and extract key features from it.

[0031] This application uses a Bayesian Knowledge Trace (BKT) model to dynamically infer the user's knowledge of each knowledge point. Hidden mastery probability The model includes four core parameters: initial mastery probability. Learning probability Probability of guessing correctly and probability of error Given an observation (correct) Or error The update of probabilities follows Bayes' rule:

[0032] After observing one learning attempt, the prior mastery probability for the next time step is:

[0033] in The probability of knowing the state at the current moment. This represents the probability of control at the previous moment. This represents the probability of learning a knowledge point even if one has not mastered it before.

[0034] To ensure the accuracy of learner profiles, the learner profile module implements a multi-dimensional calibration mechanism. This mechanism integrates multi-dimensional data, covering comprehensive test scores, actual programming assignment performance, and interactive behavior analysis. By utilizing these diverse data sources, it conducts cross-validation to correct potential biases inherent in single data streams. After constructing the learner profile, the module supports dynamic feedback correction, establishing a profile bias feedback channel. Users can raise objections to the system's assessment of mastery status. Upon receiving feedback, the system recalculates the mastery probability of relevant nodes and adjusts the profile, ensuring that the cognitive model aligns with the learner's true state. Ultimately, this results in a learner profile deeply embedded in the knowledge graph structure and continuously evolving. This approach not only mitigates the limitations of relying on a single type of data but also deepens the overall understanding of learners' cognitive and behavioral patterns. Through such meticulous calibration, the module can generate more accurate and reliable learner profiles, which is crucial for achieving personalized learning experiences and targeted educational interventions.

[0035] 3. Path planning module This module is the decision-making center of this application, employing multiple verification and dynamic optimization mechanisms to ensure the effectiveness and reliability of the generated paths. The core innovation of this module lies in "hybrid path planning," a unique combination of collaborative filtering and a Bayesian knowledge tracing model, to achieve truly dynamic personalization.

[0036] The path planning module first generates an optimal initial path that satisfies knowledge dependencies based on the topology of the knowledge graph and the knowledge points already mastered in the learner's profile. This path ensures that students start with the knowledge they have already mastered and gradually learn the knowledge points they haven't yet mastered but which are prerequisites. The path generation uses the graph-based search algorithm A*, with the objective function being to minimize the path length or maximize the knowledge gain. The following is the specific expression for the optimal initial path:

[0037] in It is the set of all valid paths that satisfy the prerequisites. It is influenced by factors such as learning time and difficulty.

[0038] To further personalize learning paths, this module innovatively introduces a model-based collaborative filtering algorithm, employing a matrix factorization model to predict learners. In a specific path Each knowledge node Expected effect score The model learns by minimizing the regularized squared error loss function as follows:

[0039]

[0040] in Given a set of known interactions, The global average score and These are biases for learners and knowledge points, used to capture individual and project differences. Regularization coefficients are used to control model complexity. After model training, it can predict learners. For path Any unlearned knowledge points Score Based on this, the calculation path Overall predictive mastery ,in This is a set of knowledge points that have already been mastered.

[0041] To evaluate the robustness of the path and perform final optimization, this module designs a Monte Carlo simulation. For candidate paths... ,conduct Each simulation is performed based on the learner's current state. Based on the historical success rate of knowledge points, a possible learning trajectory is randomly generated, and key indicators such as the estimated total learning time are recorded. and probability of failure (That is, the probability of failure at any key node on the path). Finally, a comprehensive path utility function is defined, which integrates the above multi-source information and serves as the final basis for path ranking and optimization:

[0042] in, and It is a normalization factor. Measure the diversity of paths in terms of resource type or learning mode. and These are weighting coefficients, reflecting preferences for different objectives. The system selects the path with the highest utility from the initial path and its variations (generated through node insertion, deletion, and replacement). Personalized optimal path as the final recommendation .

[0043] Based on the expected learning outcomes, the initial learning path is optimized and adjusted to generate the final target learning path. The optimization and adjustment of the initial learning path includes at least one of the following operations: 1. If the predicted mastery probability of a certain knowledge node in the path is lower than the failure risk threshold, an auxiliary preparatory knowledge node is inserted before that node; 2. If the predicted total learning time of the entire path exceeds the learner's preset time budget, an equivalent alternative path consisting of more mastered knowledge nodes is recommended to shorten the total time; 3. If the prediction information indicates that similar learners significantly improve their learning outcomes when using a different resource type at a certain knowledge node, the resource type of that node in the path is replaced with the more effective resource type.

[0044] This module also has dynamic feedback and adjustment capabilities, which monitor learners' progress in real time and automatically replan when the deviation from the expected outcome exceeds a threshold. It also provides a manual review interface, allowing teachers or domain experts to review and adjust the generated path to ensure that the path meets the educational and teaching objectives.

[0045] 4. Intelligent Learning Guidance Module This module provides real-time, secure, and intelligent support for the learning process. It employs multiple constraint and verification mechanisms to ensure the accuracy and security of the generated content, addressing the illusion problem that may arise in LLM learning within educational settings. The core of the intelligent learning guidance module is SafeRetrieval Enhanced Generation (SafeRAG) technology, used to generate standardized knowledge cards for each knowledge node along the learning path. The SafeRAG framework directly resolves the contradiction in traditional RAG frameworks within structured education: overly strict retrieval leads to rigid content, while overly lenient retrieval leads to illusion.

[0046] Figure 2 A flowchart illustrating the specific techniques for enhancing secure retrieval is provided. The specific process includes: S1, from knowledge graphs Searching for target knowledge nodes All authoritative structured information, including its definitions, key concepts, related relationships, prerequisites, etc.

[0047] S2. Design a multi-layered constraint system that combines the retrieved structured information with prompt word templates containing strict anti-hallucination instructions to construct a double-constrained structured input. This template aims to guide large language models in generating accurate, concise, and academically sound review content, ensuring that the generated content strictly adheres to the factual basis of the knowledge graph while preserving the adaptability of the large language model.

[0048] S3, Invoke the large language model according to The system generates initial content for knowledge cards. During generation, a multi-model cross-validation mechanism is introduced to automatically detect and handle inconsistencies or errors in the generated results, minimizing the illusionary nature of LLM (Learning Power Module) content. This mechanism compares the responses of different LLM models to the same prompt word and combines this with pre-defined factual verification rules for the knowledge graph to identify and correct potential illusory content. Before output, the generated content undergoes a rigorous post-processing process, including terminology consistency checks (ensuring consistency with standard terminology in the knowledge graph) and integrity checks. Furthermore, the system performs a final factual verification to ensure that all generated information is traceable to knowledge points within the knowledge graph.

[0049] The SafeRAG framework ensures that the generated knowledge cards are not only authoritative and accurate, but also clearly and easily understood, effectively solving the illusion problem that may occur in LLM in educational scenarios and achieving a balance between high factual accuracy and generative adaptability.

[0050] Considering the rapid updates in the field of AI knowledge, the intelligent learning module also provides optional online functionality to obtain the latest research updates from a pre-defined whitelist of authoritative sources (such as top conference and journal databases). This updated information is carefully recommended to students in the form of a separate "Latest Research Updates" module, clearly distinguished from the core knowledge cards to avoid confusion and ensure the stability and authority of the core knowledge.

[0051] This module also provides a secure and controllable network function. Responding to user requests, it retrieves the latest dynamic information related to the specific knowledge node only from a preset whitelist of authoritative sources and presents it to the user in an independent format distinct from the standardized knowledge cards, ensuring a balance between the authority of core knowledge and the timeliness of extended information. In the user interface, the latest dynamic information is visually clearly separated from the standardized knowledge cards, and a risk warning label is added to the latest dynamic information stating, "This content is automatically retrieved external dynamics and is for reference only."

[0052] Below, this application will elaborate on the personalized learning path planning and guidance capabilities of the present invention.

[0053] The LLM-enhanced adaptive learning planning system proposed in this application, abbreviated as LKPS (LLM-enhanced Knowledge Pathing System), employs three methods—LKPS, traditional knowledge graph systems (KGS), and collaborative filtering recommendation systems (CFRS)—to provide path planning and guidance support for the "multimodal machine learning" learning topic. The processed learning data is compared and analyzed through learning effectiveness evaluation and visualization methods. By comparing the implementation results of different methods, the significant advantages of this invention in personalized education tasks are verified.

[0054] To verify the performance of this application, comparative and ablation experiments were designed. All experiments were conducted under a simulated "multimodal machine learning" learning theme, which is characterized by strong cross-fertilization of knowledge points and rapid updates, effectively testing the system's adaptability and robustness. The experiments were conducted on a simulated learning platform capable of recording learners' learning behaviors, test scores, and interaction data. We constructed a knowledge graph containing core knowledge points in the field of "multimodal machine learning" and generated corresponding learning materials and test questions based on it. A group of test learners with different learning backgrounds and knowledge levels were recruited to ensure the generalizability of the experimental results.

[0055] This application uses the following four key metrics to quantitatively evaluate the system's performance: 1. Path rationality: This indicator is calculated based on the path generated by the system. Compared with the standard optimal path defined by domain experts The normalized edit distance is used to measure the minimum number of edit operations (insertion, deletion, and replacement of nodes) required to transform one path into another. The calculation formula is as follows:

[0056] in, This indicates the number of nodes in the path. The closer this value is to 1, the more similar the generated path is to the expert path, and the higher its rationality.

[0057] 2. Machine Hallucination Rate: N statements were randomly selected from the learning materials generated by the system for all knowledge points in the test set to form an evaluation sample. Subsequently, three domain experts were hired to independently label each sentence as factually incorrect based on authoritative facts in the knowledge graph. The hallucination rate was calculated using the following formula:

[0058] This application report is based on the average of the annotation results from three experts. The lower the value, the more reliable the content generated by the system.

[0059] 3. Personalization Index: The average mastery probability of nodes in the path. The higher the value, the closer the path is to the learner's current known domain.

[0060] 4. Learning Success Rate: In the simulated learning environment, M simulated learners are configured for each experiment group. A learner is considered "successful" if and only if their mastery of all core knowledge points exceeds the threshold in the final evaluation of path P. This experiment will Set to 0.8. The formula for calculating the learning success rate is:

[0061] To comprehensively evaluate the performance of this application, it is compared with two representative baseline systems: KGS (Knowledge Graph System): a traditional knowledge graph-based learning system that mainly generates fixed learning paths based on the prerequisite relationships between knowledge points, lacking personalization and dynamic adaptability; and CFRS (Collaborative Filtering Recommendation System): a collaborative filtering-based recommendation system that recommends learning paths by analyzing the historical behavior of similar learners, possessing a certain degree of personalization capability, but may ignore the real-time knowledge state of individuals.

[0062] Table 1 and Figure 3 This demonstrates the performance of this application and two limit systems across various evaluation metrics. (See Table 1 and...) Figure 3It can be seen that this application significantly outperforms KGS and CFRS in all evaluation metrics. Regarding path rationality, this application has the highest path rationality (0.93), indicating that by combining the topological structure of the knowledge graph with real-time learner profiles, it can generate paths that better conform to cognitive patterns and the actual situation of learners. Although KGS is also based on a knowledge graph, its path rationality is slightly inferior due to the lack of dynamic perception of individual learning states. CFRS, relying primarily on group behavior, may not be able to fully capture the strict dependencies between knowledge points, thus having the lowest path rationality. Regarding the machine illusion incidence rate, this application has the lowest machine illusion incidence rate (0.04), thanks to its strict multi-layer constraint cue word engineering, multi-model arbitration, and factual verification in its SafeRAG mechanism. KGS, not involving LLM generation, has a relatively low illusion incidence rate (0.08), but may still suffer from information bias due to untimely knowledge updates. CFRS, potentially directly referencing or aggregating LLM-generated content without rigorous verification, has the highest illusion incidence rate (0.15). In terms of personalization index, this application performs best (0.88), which fully verifies the effectiveness of its hybrid path planning strategy (combining BKT and CF). Through Bayesian knowledge tracing to finely model individual knowledge mastery and supplemented by the collective intelligence of collaborative filtering, this application can provide learners with a highly customized learning experience. KGS has the lowest personalization index due to its fixed path. Although CFRS has some personalization, it lacks a deep understanding of the individual's real-time knowledge state, so its personalization level is not as high as this application. In terms of learning success rate, this application has the highest learning success rate (0.86), indicating that it can more effectively help learners master target knowledge. This is closely related to its advantages in path rationality, low illusion rate, and high personalization. The learning success rates of KGS and CFRS are relatively low, reflecting their shortcomings in adapting to learners' needs and providing reliable learning content.

[0063] Table 1. Comparison of performance data of this application and two baseline systems on various evaluation indicators.

[0064] In summary, the comparative experimental results fully demonstrate the significant advantages of this application in providing personalized, reliable, and efficient AI knowledge adaptive learning planning.

[0065] To gain a deeper understanding of the contribution of each core module of this application to the overall system performance, we conducted a series of ablation experiments. We constructed several variants of this application and observed their impact on various metrics by removing or simplifying specific modules.

[0066] LKPS-Full: A complete LKPS system containing all core modules, serving as a benchmark for performance comparison.

[0067] LKPS-CF: The collaborative filtering component has been removed. Path planning now relies solely on the knowledge graph topology and real-time learner profiles, no longer referencing the success patterns of similar learners to evaluate the contribution of collective intelligence to personalized recommendations.

[0068] LKPS-MC: The Monte Carlo simulation validation step has been removed. Path optimization is now based solely on the initial prediction model, and risk simulations and robustness assessments based on historical data are no longer performed to validate the simulation process's role in identifying potential learning bottlenecks and improving path success rates.

[0069] LKPS-DA: Removed dynamic adjustment capability. The learning path is fixed after initial generation and cannot be optimized based on the learner's real-time progress. It is used to evaluate the impact of the system's dynamic adaptability on long-term learning outcomes.

[0070] LKPS-RAG: The SafeRAG mechanism is removed. The intelligent learning module directly uses LLM to generate content without retrieval enhancement and strict constraints. It is used to verify the necessity of SafeRAG in controlling hallucinations.

[0071] LKPS-Base: The basic model, retaining only the core knowledge graph path generation functionality. It removes all advanced components such as SafeRAG, collaborative filtering, Monte Carlo simulation, and dynamic tuning, representing the lower limit of the system's performance.

[0072] Table 2 and Figure 4 The performance of this application and its variants on various evaluation metrics is demonstrated. The results of the ablation experiments reveal the key roles of each core module of this application: Table 2. Contribution Analysis of Each Module

[0073] Contribution of Collaborative Filtering: Comparing this application (Personalization Score=0.88, Learning Success Rate=0.85) and LKPS-CF (Personalization Score=0.62, Learning Success Rate=0.72), it can be found that removing the collaborative filtering module leads to a significant decrease in the personalization index and learning success rate. This indicates that collaborative filtering plays a key role in introducing collective intelligence, enhancing path personalization, and improving learning outcomes.

[0074] Contribution of Monte Carlo simulation validation: Comparing this application (PR=0.92, LSR=0.85) and LKPS-MC (PR=0.87, LSR=0.68), it was found that removing the Monte Carlo simulation validation step significantly reduced the rationality of the path and the learning success rate. This highlights the importance of multi-model validation in ensuring the quality of knowledge graph construction and content generation, and indirectly affects the effectiveness of the path.

[0075] Contribution of dynamic updates: Comparing this application (Path Rationality = 0.92, Learning Success Rate = 0.85) and LKPS-DA (Path Rationality = 0.89, Learning Success Rate = 0.65), removing the dynamic update mechanism leads to a decrease in path rationality and learning success rate. This indicates that dynamic updates of the knowledge graph and learner profiles are crucial for the system to adapt to the rapid development of the AI ​​field and real-time changes in learner status.

[0076] Contribution of the SafeRAG mechanism: Comparing this application (Hallucination Rate=0.03, Learning Success Rate=0.85) and LKPS-RAG (Hallucination Rate=0.25, Learning Success Rate=0.67), it is clear that the SafeRAG mechanism makes a significant contribution to reducing the occurrence rate of machine illusions and improving the learning success rate. When SafeRAG is removed, the occurrence rate of illusions increases sharply and the learning success rate decreases significantly. This directly proves the irreplaceable role of SafeRAG in ensuring the reliability of intelligent learning content.

[0077] The role of LKPS-Base: As the simplest version of this application, LKPS-Base has the lowest performance across all metrics, which further emphasizes the necessity of integrating all high-level modules in LKPS to achieve a high-performance adaptive learning system.

[0078] This application first constructs a "multimodal machine learning" knowledge graph to ensure the integrity and authority of the knowledge structure. Subsequently, learning paths and guidance support are generated for the same group of test learners using LKPS, KGS, and CFRS methods, respectively. Learning effectiveness data (such as knowledge point mastery rate, learning time, user satisfaction, etc.) under each method are recorded, and then comparative analysis is conducted. Figure 3 and Figure 4The study presents the knowledge mastery status change curves of learners under three different methods. The analysis results clearly show that the LKPS method proposed in this application demonstrates superior learning effectiveness. Compared to the KGS and CFRS methods, LKPS can more efficiently improve learners' mastery of core knowledge, its learning path is more aligned with individual ability changes, and the proportion of learners encountering learning bottlenecks is significantly reduced, exhibiting stronger personalized adaptability.

[0079] In particular, the learning curve of this application exhibits a stable and rapid upward trend, indicating that it can effectively identify learner status and dynamically optimize subsequent paths. This performance, when dealing with the highly interdisciplinary and rapidly iterating topic of "multimodal machine learning," can ensure learning effectiveness through real-time profile updates, collective intelligence, and a secure learning guidance mechanism. In contrast, the learning effects of KGS and CFRS methods show obvious stagnation periods and setbacks, with unnatural transitions between different knowledge modules, leading to reduced learning efficiency and a degraded learning experience.

[0080] This experiment demonstrates the effectiveness and advantages of the system in personalized education, particularly in rapidly developing technology education fields such as artificial intelligence. Precise learning status tracking and dynamic path adjustment can significantly improve learning efficiency and depth of understanding. The experimental results verify the superiority of the technical solution in practical applications, providing a solid theoretical and empirical foundation for the application of this invention in the field of smart education.

[0081] This application also provides a computer program product, including computer instructions, which, when executed by a processor, implement each step of the aforementioned large language model-enhanced artificial intelligence knowledge adaptive learning planning system and method, and can achieve the same beneficial effects as the aforementioned large language model-enhanced artificial intelligence knowledge adaptive learning planning system and method. To avoid repetition, these will not be described again here.

[0082] In the embodiments provided in this application, it should be understood that the disclosed methods and apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between devices or units, and may be electrical, mechanical, or other forms.

[0083] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can be physically included separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0084] The integrated units implemented as software functional units described above can be stored in a computer-readable storage medium. These software functional units, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute some steps of the transmission and reception methods described in the various embodiments of this application. 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.

[0085] The above description is the preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principles described in this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A knowledge-adaptive learning planning system enhanced by a large language model, characterized in that, The system includes a knowledge hub module, a learner profiling module, a path planning module, and an intelligent learning guidance module. The knowledge hub module uses a large language model to extract entities and relationships from multimodal data sources to construct a knowledge graph. The learner profiling module collects multidimensional learning data from users and maps it to corresponding nodes in the knowledge graph, and dynamically infers learner profiles through a Bayesian knowledge tracing model. The path planning module generates an initial learning path based on a knowledge graph and learner profiles, establishes a collaborative filtering analysis model and combines it with a Bayesian knowledge tracing model to predict and optimize the expected effect of the current learner following the initial learning path, and finally generates the target learning path. The intelligent learning guidance module generates standardized knowledge cards for each knowledge node on the target learning path through secure retrieval and enhanced generation technology.

2. The knowledge adaptive learning planning system for large language model enhancement according to claim 1, characterized in that, The knowledge graph of the knowledge hub module is a structured knowledge representation, including the relationships between knowledge points. Each node in the knowledge graph represents a knowledge point, and each edge of the knowledge graph represents the relationship between knowledge points. Knowledge graphs are based on graph structures Stored in the form of , where A collection of knowledge points This is a set of relationships between knowledge points, and each edge... express and There exists a type of The relationship.

3. The knowledge adaptive learning planning system for large language model enhancement according to claim 2, characterized in that, The knowledge graph is constructed using a multi-factor verification mechanism, including: A multi-model cross-validation mechanism is set up, and at least two large language models with different architectures are used to generate knowledge entities and relations in parallel. The entities and relations are used to construct the knowledge points and the associations between knowledge points in the knowledge graph. The system labels knowledge associations that are disputed or have low confidence, provides a visual review interface, and establishes an error feedback chain. When users or teachers discover knowledge errors during subsequent use, they can submit correction suggestions through a designated entry point. After receiving the suggestions, the system will restart the verification process to update the knowledge graph.

4. The knowledge adaptive learning planning system for large language model enhancement according to claim 3, characterized in that, The learner profiling module supports dynamic feedback correction and establishes a profile deviation feedback channel. Users can raise objections to the mastery status assessed by the system. After receiving the feedback, the system recalculates the mastery probability of relevant nodes and adjusts the profile.

5. The knowledge adaptive learning planning system for large language model enhancement according to any one of claims 1-4, characterized in that, The multi-model cross-validation mechanism identifies and corrects potential hallucination content by comparing the responses of different large language models to the same prompt word and combining them with preset knowledge graph factual verification rules. Before the generated content is output, it undergoes a post-processing process, which includes terminology consistency verification and integrity checks. In addition, the system performs a final factual verification to ensure that all generated information can be traced back to the knowledge points in the knowledge graph.

6. A planning method for a knowledge adaptive learning planning system enhanced by a large language model, characterized in that, In the learner profiling module, a Bayesian knowledge tracing model is used to dynamically infer the user's understanding of each knowledge point. The hidden mastery state probability is determined; in the path planning module, a personalized preferred path that satisfies knowledge dependencies is generated and optimized through model-based collaborative filtering algorithm and Monte Carlo simulation; in the intelligent learning module, a standardized knowledge card is generated for each knowledge node on the path through secure retrieval enhancement generation technology.

7. The planning method for the knowledge adaptive learning planning system with large language model enhancement according to claim 6, characterized in that, The learner profiling module uses a Bayesian knowledge tracing model to dynamically infer the user's understanding of each knowledge point. Hidden mastery probability The model includes four core parameters: initial mastery probability. Learning probability Probability of guessing correctly and probability of error ; The update of the mastery probability in the Bayesian knowledge tracing model follows the Bayesian rule: in For the observation results, when correct When an error occurs ; After observing one learning attempt, the prior mastery probability for the next time step is: in The probability of knowing the state at the current moment. This represents the probability of control at the previous moment. This represents the probability of learning a knowledge point even if one has not mastered it before.

8. The planning method for the knowledge adaptive learning planning system with large language model enhancement according to claim 7, characterized in that, The path planning module generates an optimal initial path that satisfies knowledge dependencies based on the topology of the knowledge graph and the knowledge points already mastered in the learner profile. The following formula is the specific expression for the optimal initial path: in It is the set of all valid paths that satisfy the prerequisites. Factors such as study time and difficulty; The model-based collaborative filtering algorithm uses a matrix factorization model to predict learners. In a specific path Each knowledge node Expected effect score ; The model learns by minimizing the regularized squared error loss function as follows: in Given a set of known interactions, The global average score and These are biases for learners and knowledge points, used to capture individual and project differences. Regularization coefficients used to control model complexity; After the model is trained, it can predict learners. For path Any unlearned knowledge points Score ; Calculation path Overall predictive mastery ,in This is a set of knowledge points that have already been mastered.

9. The planning method for the knowledge adaptive learning planning system with large language model enhancement according to claim 8, characterized in that, The path planning module designs Monte Carlo simulations to evaluate the robustness of the path and performs final optimization; For candidate paths ,conduct Each simulation, in which the learner's current state is considered... Based on the historical success rate of knowledge points, a possible learning trajectory is randomly generated, and key indicators such as the estimated total learning time are recorded. and probability of failure This refers to the probability of failure at any critical node along the path; ultimately, a comprehensive path utility function is defined as the final basis for path ranking and optimization: in, and It is a normalization factor. Measure the diversity of paths in terms of resource type or learning mode. and These are weighting coefficients, reflecting preferences for different objectives; the system selects the path with the highest utility from the initial path and its variations. Personalized optimal path as the final recommendation .

10. The planning method for a knowledge adaptive learning planning system with large language model enhancement according to any one of claims 6-9, characterized in that, The specific process of the security retrieval enhancement generation technology includes: S1, from knowledge graphs Searching for target knowledge nodes All authoritative structured information, including its definitions, key concepts, relationships, or prerequisites; S2. Combine the retrieved structured information with prompt word templates containing strict anti-hallucination instructions to construct a double-constrained structured input. The prompt word template guides the large language model to generate accurate, concise, and academically compliant review content, ensuring that the generated content strictly adheres to the factual basis of the knowledge graph while retaining the generative adaptability of the large language model. S3, Invoke the large language model according to The initial content of the knowledge cards is generated. During the generation process, a multi-model cross-validation mechanism is introduced to automatically detect and handle inconsistencies or errors in the generated results.

Images