Learning resource recommendation method and system based on knowledge tracking and retrieval enhancement generation

Abstract

The invention provides a learning resource recommendation method based on knowledge tracking and retrieval enhancement generation, and belongs to the technical field of education. The method comprises the steps of obtaining data of a user learning platform and / or learning content uploaded by a user, and constructing a background knowledge base related to a current learning task of the user; acquiring learning interaction behavior data of the user, constructing a knowledge state tracking model, and forming a current knowledge mastering state of the user; wherein the learning interaction behavior data of the user comprises knowledge points, exercises, videos and texts; and outputting personalized learning resource recommendation and question and answer response by utilizing a retrieval enhancement generation model in combination with the background knowledge base and the knowledge mastering state. Therefore, dynamic, personalized and knowledge-accurate learning content recommendation can be realized.

Description

Technical Field

[0001] This invention belongs to the field of educational technology, and more specifically, relates to a method and system for recommending learning resources based on knowledge tracking and retrieval enhancement. Background Technology

[0002] Large Language Models (LLMs) are widely used in education, possessing powerful natural language understanding and generation capabilities. However, their application in education still has the following shortcomings: First, they cannot perceive the learner's knowledge status, resulting in untargeted content recommendations; second, they are prone to generating hallucinations; and third, they cannot access the latest course knowledge, leading to outdated knowledge.

[0003] Traditional educational recommendation systems primarily rely on knowledge tracing models, such as Bayesian Knowledge Tracing (BKT) and Deep Knowledge Tracing (DKT), but they cannot handle complex and open-ended problems. While RAG (Retrieval-Augmented Generation) technology can incorporate external knowledge and improve response accuracy, it lacks integration with the learner's state. In other words, the existing large-scale educational recommendation systems generally suffer from poor knowledge personalization, low accuracy of response content, and a lack of awareness of learning status. Summary of the Invention

[0004] The main purpose of this application is to provide a learning resource recommendation method and system based on knowledge tracing and retrieval enhancement, which can integrate KT (Knowledge Tracing) and RAG mechanisms to achieve dynamic, personalized and knowledge-accurate learning content recommendation.

[0005] To achieve the above objectives, this application proposes a learning resource recommendation method based on knowledge tracing and retrieval enhancement, including: Acquire data from the user's learning platform and / or the learning content uploaded by the user, and build a background knowledge base related to the user's current learning task; Acquire user learning interaction behavior data, construct a knowledge status tracking model, and determine the user's current knowledge mastery status; wherein, the user's learning interaction behavior data includes knowledge points, exercises, videos, and text; By utilizing a retrieval-enhanced generative model, combined with the background knowledge base and the knowledge mastery status, personalized learning resource recommendations and question-and-answer responses are output.

[0006] Furthermore, the acquisition of data from the user's learning platform includes: performing structured extraction and representation of online course resources, and constructing a dynamically updated background knowledge base.

[0007] Furthermore, the construction of the knowledge state tracking model includes: Construct a heterogeneous graph structure that covers learning resources, and use a hierarchical graph neural network to achieve intra-modal and cross-modal information fusion; The learning resources include knowledge points, exercises, videos, and text. The knowledge points are defined through a knowledge graph, and the exercises, videos, and texts are associated with the knowledge points through semantic matching.

[0008] Furthermore, the method of using a hierarchical graph neural network to achieve intra-modal and cross-modal information fusion includes: During intramodal aggregation, a graph attention network is used to extract features from the single-modal subgraphs of the heterogeneous graph structure. In cross-modal alignment, embedding vectors from different modalities are input into a multimodal Transformer model, and semantic alignment is achieved through an interleaved attention mechanism. The final output is a multimodal joint embedding model that comprehensively reflects the semantic associations of the learning resources.

[0009] Furthermore, it also includes: By combining multimodal feature increments with historical states, the uncertainty of knowledge states is modeled through latent variables, and formalized as follows: in, Represented as a historical state, It is represented as the multimodal feature increment, where St represents the knowledge state at time point t, and t represents the time point; Introducing a self-supervised learning task to optimize behavioral representations, constructing positive and negative sample pairs, and minimizing contrastive loss: in, To compare loss functions, This is represented as the similarity score for positive sample pairs. Let B represent the similarity score of negative sample pairs, B represent the total number of candidate samples in the batch, and K represent the sample index. By performing masking modeling on the behavior sequence and predicting the masked behavior features through a temporal convolutional network, the knowledge state tracking model's ability to perceive the learning rhythm is enhanced.

[0010] Furthermore, it also includes: Using causal contribution analysis, the Peter-Clark algorithm was used to construct a causal graph to identify the causal effects of key behavioral variables on knowledge state, and their impact was quantified through counterfactual reasoning. A multi-task joint optimization approach is adopted, and a multi-task loss function is used to train a method that balances the knowledge tracking accuracy and interpretability constraints.

[0011] Furthermore, the method of utilizing the retrieval-enhanced generative model, combined with the background knowledge base and the knowledge mastery status, to output personalized learning resource recommendations and question-answering responses includes: For a given recommended learning resource item, a multimodal query vector is constructed, and a hierarchical retrieval is performed in the background knowledge base; wherein, in the precise matching layer, the relevant entity set is located in the knowledge graph through a graph pattern matching algorithm; when retrieval of unstructured knowledge, the most relevant supporting evidence for the query is extracted from the document set based on a cross-modal retrieval mechanism; By employing a hybrid scoring mechanism that balances semantic relevance and domain expertise, a set of document evidence is ultimately obtained. The results of hierarchical retrieval are integrated to form enhanced knowledge units, and then graph neural networks are used for information aggregation and graph embedding computation. Based on the control generation strategy, a template hierarchy structure is used to realize the multi-dimensional and multi-granular generation of recommendation explanations.

[0012] Furthermore, the template hierarchy includes conceptual explanation, procedural explanation, and strategic explanation; wherein, the generation process of the template hierarchy adopts a conditional language model to explain the weight allocation and dynamic control of different knowledge units during generation.

[0013] Furthermore, it also includes introducing evidence to constrain its loss function: in, This can be represented as an alignment loss function, used to constrain the consistency or credibility between recommended content and knowledge evidence. This is represented as the i-th enhanced knowledge unit. This is represented by the semantic tag corresponding to the l-th recommendation. This represents the explanation output generated by the model at position l. Represented as in the given target content and knowledge evidence Under these conditions, the model output The probability of.

[0014] This invention also provides a learning resource recommendation system based on knowledge tracking and retrieval enhancement, comprising: The first module is used to acquire data from the user's learning platform and / or the learning content uploaded by the user, and to build a background knowledge base related to the user's current learning task. The second module is used to acquire users' learning interaction behavior data, construct a knowledge status tracking model, and form the user's current knowledge mastery status; wherein, the user's learning interaction behavior data includes knowledge points, exercises, videos, and text; The third module utilizes a retrieval-enhanced generative model, combined with the background knowledge base and the knowledge mastery status, to output personalized learning resource recommendations and question-and-answer responses.

[0015] Compared with existing technologies, the beneficial effects of this invention are: by constructing a background knowledge base related to the current learning topic and sensing the learner's knowledge mastery and ability status; based on the learning status information and knowledge base content, driving the large language model to generate personalized recommended content, Q&A feedback and learning path suggestions, thereby achieving adaptive optimization of learning recommendations. Attached Figure Description

[0016] Figure 1 A flowchart illustrating the learning resource recommendation method based on knowledge tracking and retrieval enhancement provided in this embodiment of the invention; Figure 2 This is a schematic diagram of a learning resource recommendation system based on knowledge tracking and retrieval enhancement provided in an embodiment of the present invention. Detailed Implementation

[0017] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0018] The following specific examples illustrate the implementation of this disclosure. Those skilled in the art can easily understand other advantages and effects of this disclosure from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. This disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0019] like Figure 1 As shown, this embodiment provides a learning resource recommendation method based on knowledge tracking and retrieval enhancement, including steps S100 to S300: S100. Obtain data from the user's learning platform and / or the learning content uploaded by the user, and construct a background knowledge base related to the user's current learning task.

[0020] For example, course content (subtitles, text, annotations, etc.) can be extracted from web pages, platforms, or servers using plugins such as Scraper, and then embedded as a vector using the Reader API: in For text fragments, the encoder is a model such as SentenceBERT. For text fragments The encoded vector representation is where j is the index variable and m is the total number of text segments.

[0021] S200. Acquire user learning interaction behavior data, construct a knowledge status tracking model, and form the user's current knowledge mastery status; wherein, the user's learning interaction behavior data includes knowledge points, exercises, videos, and text.

[0022] Specifically, a heterogeneous graph structure is constructed covering four learning resource stages: knowledge points, exercises, videos, and text. Knowledge point nodes are defined with logical relationships such as prerequisites and dependencies through a knowledge graph. Exercise, video, and text nodes are associated with knowledge points through semantic matching (e.g., keyword extraction from exercise stems, and NER alignment between video subtitles and knowledge points). Based on this, a hierarchical graph neural network (HNN) is designed to achieve intra-modal and cross-modal information fusion.

[0023] During intramodal aggregation, a graph attention network is used to extract features from single-modal subgraphs (such as exercise-knowledge point subgraphs) to generate exercise embeddings. Video embedding and text embedding Formalized as: in, This represents the set of knowledge point neighbors of the exercise node. Let Q be the representation vector of the node Q after the l-th layer of neurons is forgotten, and GATconv represents the graph attention convolution operation.

[0024] In cross-modal alignment, the embedding vectors of different modalities are input into a multimodal Transformer, and semantic alignment is achieved through a cross-attention mechanism, which can be formalized as follows: in, The embedding representation vector of the exercise modality. For the embedding representation vector of the video modality, Let H be the embedding representation vector of the text modality, and H be the concatenation of all modality vectors. Let H be the transpose of H. Attn represents the attention similarity score between each modality, and Attn represents the final attention output.

[0025] The final output is a multimodal joint embedding. It can comprehensively reflect the semantic relationships of learning resources.

[0026] To capture the temporal evolution of learning behavior, a dynamic state update mechanism based on neural processes is used. This module increments multimodal features. With historical status Combining, through latent variables The uncertainty of the knowledge state in modeling can be formalized as follows: in, Represented as a historical state, It is represented as the multimodal feature increment, St represents the knowledge state at time point t, and t represents the time point.

[0027] Simultaneously, a self-supervised learning task is introduced to optimize behavioral representations, construct positive and negative sample pairs, and minimize the contrastive loss: in, To compare loss functions, This is represented as the similarity score for positive sample pairs. Let B represent the similarity score of negative sample pairs, B represent the total number of candidate samples in the batch, and K represent the sample index.

[0028] Masking modeling is performed on behavioral sequences (such as dwell time and answering patterns), and the masked behavioral features are predicted through a temporal convolutional network to enhance the model's ability to perceive the learning rhythm.

[0029] To enhance model transparency, a dual-path interpretability module combining causal inference and attention visualization is designed. Causal contribution analysis is used, and the Peter-Clark (PC) algorithm is employed to construct a causal graph, identifying the causal effects of key behavioral variables (such as the number of video replays) on the knowledge state, and quantifying their impact through counterfactual reasoning.

[0030] In terms of training strategy, a multi-task joint optimization approach is adopted, using a multi-task loss function to train a balance between knowledge tracking accuracy and interpretability constraints. in, For the total loss function, The score loss measures the knowledge mastery predicted by the model at time point t. With real labels The differences between them To count the loss of the prediction task, Modeling loss for mask behavior, The KL divergence between two distributions , , These are the weighting coefficients for each loss term.

[0031] To accelerate convergence and improve robustness, a course-based learning strategy is adopted to train the model in stages. In the initial stage, only exercise modal data is used to optimize the prediction of basic knowledge states. In the advanced stage, complex modalities such as video and text are gradually introduced to jointly optimize cross-modal alignment and behavior modeling. In the final stage, the model is fine-tuned on full-modal data to enhance the ability to model long-term knowledge evolution.

[0032] S300. Using a retrieval-enhanced generative model, combined with the background knowledge base and the knowledge mastery status, output personalized learning resource recommendations and question-and-answer responses.

[0033] Specifically, a dual-channel heterogeneous retrieval architecture is adopted to process structured knowledge graphs and unstructured teaching document data separately. For a given recommended learning resource item... Construct a multimodal query vector: in For knowledge component encoder, The encoder is used to represent the user. This is a project feature encoder. Based on this query vector, in the knowledge space... Hierarchical retrieval is performed in the knowledge space. This serves as a background knowledge base.

[0034] At the precise matching layer, graph pattern matching algorithms are used in the knowledge graph. Set of related entities in the middle ,in, To accurately match the set of entities output by the layer, This is the vector representation of the current query segment. For entity e and query Similarity score between them This is the similarity threshold for entity matching.

[0035] The similarity calculation uses graph attention metric. ,in For trainable projection matrices, Here, is the Sigmoid activation function, and e represents an entity node in the knowledge graph. This is the vector representation of the current query segment.

[0036] When retrieving unstructured knowledge, specifically at the semantic extension layer, within the document collection... In this process, based on a cross-modal retrieval mechanism, the most relevant supporting evidence to the query is extracted: in, Let D be the set of unstructured knowledge documents, representing the optimal supporting evidence fragment selected through semantic matching and weighted summation. For query With document fragments The correlation score between the documents is denoted by arg max, which is a maximization operation that selects the highest-scoring item from all document fragments.

[0037] By employing a hybrid scoring mechanism to balance semantic relevance and domain expertise, a document evidence set is ultimately obtained. Then, the knowledge unit construction is enhanced by integrating structured and unstructured search results to form enhanced knowledge units. Then, a graph neural network is used for information aggregation and graph embedding computation: in, For the i-th knowledge enhancement unit, GraphConv is the graph convolution operation. Given the current query vector, GAT is a graph attention network. To enhance the final embedded representation of knowledge units.

[0038] Based on a controlled generation strategy, a template hierarchy is used to achieve multi-dimensional and multi-granular generation of recommendation explanations. The explanation template hierarchy is defined. These correspond to three dimensions: conceptual explanation, procedural explanation, and strategic explanation. The generative process is modeled as a conditional language model. in, To generate the conditional probability of the complete recommendation explanation text, y represents the target output text sequence of the recommendation explanation. For attention modules, knowledge representation The vector is fused with the template control vector t to calculate the context-dependent attention vector, which guides the language generation at each step.

[0039] This includes the weight allocation and dynamic control of different knowledge units during the interpretation generation process: in, Let be the embedding vector of the nth knowledge unit. Let be the attention weight of the nth knowledge unit under the current template t. For template type-specific parameter matrix, knowledge selection is implemented to achieve adaptive interpretation type, where b is the bias term.

[0040] To improve the credibility of the generated results, an evidence-based loss function is introduced: in, This can be represented as an alignment loss function, used to constrain the consistency or credibility between recommended content and knowledge evidence. This is represented as the i-th enhanced knowledge unit. This is represented by the semantic tag corresponding to the l-th recommendation. This represents the explanation output generated by the model at position l. Represented as in the given target content and knowledge evidence Under these conditions, the model output The probability of [something]. This constraint ensures that the interpreted content maintains semantic consistency with the retrieved knowledge.

[0041] A comprehensive assessment of the credibility of the explained content is achieved through a triple verification mechanism, including source traceability verification, logical consistency verification, and adversarial robustness verification.

[0042] Regarding consistency in tracing the source, an explanatory text is established. Mapping matrix with knowledge units Calculate the traceability score of the explanation: The NLI model calculates the statement-level entailment probability and a threshold. Filter low-confidence segments. For the s-th word or sentence, This means selecting the highest-scoring option from all available knowledge units for each sentence.

[0043] In terms of logical consistency verification, a rule set defined by domain experts is introduced. This includes instructional logic rules defined by domain experts, such as the necessity of prior knowledge and the progressive nature of cognitive difficulty, which are verified and explained using Satisfiability Model Theory (SMT). Where C is the set of instructional logic rules defined by domain experts. This represents a rule item within a subset of all teaching rules. This indicates that the SMT solver is used for the logic formula. The satisfiability of the term is verified. If the implication relation holds, the term is true; otherwise, it is false.

[0044] In terms of robustness verification against adversarial attacks, the generated explanatory text... Add perturbation To assess semantic stability and robustness, robustness score for: JSD, or Jensen-Shannon Divergence, is used to measure the symmetric difference between two probability distributions. For the input disturbance term, The model generates a modified distribution to explain y given the original input x. Let represent the probability distribution of the interpretation result u generated by the model under perturbation input.

[0045] The final credibility score is derived by combining the results of the three verifications to arrive at the explanation credibility score: in, , , There are three weighting coefficients. Scoring based on credibility metrics, Scoring based on logical consistency metrics. Scoring is based on robustness indicators.

[0046] During training, a dynamic optimization mechanism is employed, employing a two-stage optimization training strategy. First, a basic generative model is pre-trained on a general teaching corpus, and then fine-grained optimization is performed through reinforcement learning. The reward function is defined as follows: in, , , The BLEU score, calculated using three weighting functions, measures the difference between the generated text y and the reference text. Similarity between them To explain the credibility score of y, Score for diversity.

[0047] like Figure 2 As shown, this embodiment also provides a learning resource recommendation system based on knowledge tracking and retrieval enhancement, including: The first module is used to acquire data from the user's learning platform and / or the learning content uploaded by the user, and to build a background knowledge base related to the user's current learning task. The second module is used to acquire users' learning interaction behavior data, construct a knowledge status tracking model, and form the user's current knowledge mastery status; wherein, the user's learning interaction behavior data includes knowledge points, exercises, videos, and text; The third module utilizes a retrieval-enhanced generative model, combined with the background knowledge base and the knowledge mastery status, to output personalized learning resource recommendations and question-and-answer responses.

[0048] The above description is merely illustrative of the embodiments of the present invention and is not intended to limit the present invention. For those skilled in the art, any modifications, equivalent substitutions, improvements, etc., made without creative effort within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A learning resource recommendation method based on knowledge tracing and retrieval enhancement, characterized in that, include: Acquire data from the user's learning platform and / or the learning content uploaded by the user, and build a background knowledge base related to the user's current learning task; Acquire user learning interaction behavior data, construct a knowledge status tracking model, and determine the user's current knowledge mastery status; wherein, the user's learning interaction behavior data includes knowledge points, exercises, videos, and text; By utilizing a retrieval-enhanced generative model, combined with the background knowledge base and the knowledge mastery status, personalized learning resource recommendations and question-and-answer responses are output.

2. The learning resource recommendation method based on knowledge tracking and retrieval enhancement as described in claim 1, characterized in that, The acquisition of data from the user's learning platform includes: performing structured extraction and representation of online course resources, and constructing a dynamically updated background knowledge base.

3. The learning resource recommendation method based on knowledge tracking and retrieval enhancement as described in claim 1, characterized in that, The construction of the knowledge state tracking model includes: Construct a heterogeneous graph structure that covers learning resources, and use a hierarchical graph neural network to achieve intra-modal and cross-modal information fusion; The learning resources include knowledge points, exercises, videos, and text. The knowledge points are defined through a knowledge graph, and the exercises, videos, and texts are associated with the knowledge points through semantic matching.

4. The learning resource recommendation method based on knowledge tracking and retrieval enhancement as described in claim 3, characterized in that, The method of using hierarchical graph neural networks to achieve intra-modal and cross-modal information fusion includes: During intramodal aggregation, a graph attention network is used to extract features from the single-modal subgraphs of the heterogeneous graph structure. In cross-modal alignment, embedding vectors from different modalities are input into a multimodal Transformer model, and semantic alignment is achieved through an interleaved attention mechanism. The final output is a multimodal joint embedding model that comprehensively reflects the semantic associations of the learning resources.

5. The learning resource recommendation method based on knowledge tracking and retrieval enhancement as described in claim 3, characterized in that, Also includes: By combining multimodal feature increments with historical states, the uncertainty of knowledge states is modeled through latent variables, and formalized as follows: in, Represented as a historical state, It is represented as the multimodal feature increment, where St represents the knowledge state at time point t, and t represents the time point; Introducing a self-supervised learning task to optimize behavioral representations, constructing positive and negative sample pairs, and minimizing contrastive loss: in, To compare loss functions, This is represented as the similarity score for positive sample pairs. Let B represent the similarity score of negative sample pairs, B represent the total number of candidate samples in the batch, and K represent the sample index. By performing masking modeling on the behavior sequence and predicting the masked behavior features through a temporal convolutional network, the knowledge state tracking model's ability to perceive the learning rhythm is enhanced.

6. The learning resource recommendation method based on knowledge tracking and retrieval enhancement as described in claim 3, characterized in that, Also includes: Using causal contribution analysis, the Peter-Clark algorithm was used to construct a causal graph to identify the causal effects of key behavioral variables on knowledge state, and their impact was quantified through counterfactual reasoning. A multi-task joint optimization approach is adopted, and a multi-task loss function is used to train a method that balances the knowledge tracking accuracy and interpretability constraints.

7. The learning resource recommendation method based on knowledge tracing and retrieval enhancement as described in claim 1, characterized in that, The method of using a retrieval-enhanced generative model, combined with the background knowledge base and the knowledge mastery status, to output personalized learning resource recommendations and question-answering responses includes: For a given recommended learning resource item, a multimodal query vector is constructed, and a hierarchical retrieval is performed in the background knowledge base; wherein, in the precise matching layer, the relevant entity set is located in the knowledge graph through a graph pattern matching algorithm; when retrieval of unstructured knowledge, the most relevant supporting evidence for the query is extracted from the document set based on a cross-modal retrieval mechanism; By employing a hybrid scoring mechanism that balances semantic relevance and domain expertise, a set of document evidence is ultimately obtained. The results of hierarchical retrieval are integrated to form enhanced knowledge units, and then graph neural networks are used for information aggregation and graph embedding computation. Based on the control generation strategy, a template hierarchy structure is used to realize the multi-dimensional and multi-granular generation of recommendation explanations.

8. The learning resource recommendation method based on knowledge tracing and retrieval enhancement as described in claim 7, characterized in that, The template hierarchy includes conceptual explanation, procedural explanation, and strategic explanation; wherein, the generation process of the template hierarchy adopts a conditional language model to explain the weight allocation and dynamic control of different knowledge units during generation.

9. The learning resource recommendation method based on knowledge tracing and retrieval enhancement as described in claim 7, characterized in that, This also includes introducing evidence to constrain its loss function: in, This can be represented as an alignment loss function, used to constrain the consistency or credibility between recommended content and knowledge evidence. This is represented as the i-th enhanced knowledge unit. This is represented by the semantic tag corresponding to the l-th recommendation. This represents the explanation output generated by the model at position l. Represented as in the given target content and knowledge evidence Under the given conditions, the model output The probability of.

10. A learning resource recommendation system based on knowledge tracing and retrieval enhancement, characterized in that, include: The first module is used to acquire data from the user's learning platform and / or the learning content uploaded by the user, and to build a background knowledge base related to the user's current learning task. The second module is used to acquire users' learning interaction behavior data, construct a knowledge status tracking model, and form the user's current knowledge mastery status; wherein, the user's learning interaction behavior data includes knowledge points, exercises, videos, and text; The third module utilizes a retrieval-enhanced generative model, combined with the background knowledge base and the knowledge mastery status, to output personalized learning resource recommendations and question-and-answer responses.

Images