Course knowledge point learning path recommendation system based on multi-modal data

Abstract

The invention discloses a curriculum knowledge point learning path recommendation system based on multi-modal data. The curriculum knowledge point learning path recommendation system comprises a basic processing layer and an information alignment reasoning layer. And the basic processing layer is used for extracting knowledge themes and knowledge points from the original multi-modal course materials and integrating mastering condition probabilities among knowledge pairs. The information alignment reasoning layer is used for processing mastering condition probabilities among the knowledge themes, the knowledge points and the integrated knowledge pairs, performing structured representation on the knowledge points in a triple form, aligning the knowledge points in the related knowledge themes, and finally predicting a pre-repair relationship among the knowledge points through a judgment model to form a learning path so as to obtain a learning result; and completing the recommendation of the course knowledge point learning path. According to the method, traditional modes of manually designing rule normal forms and manually marking structured teaching resources are reduced, and knowledge points and relationships are automatically extracted from unstructured or semi-structured teaching documents by integrating advanced natural language processing and machine learning technologies.

Description

Technical Field

[0001] This application relates to the fields of data processing and educational knowledge engineering technology, and in particular to a course knowledge point learning path recommendation system based on multimodal data. Background Technology

[0002] In existing knowledge graph construction and instructional analysis systems, the capture of knowledge points and the construction of relationships typically rely on highly structured or semi-structured teaching resources. The mainstream technical approach involves manually or semi-automatically extracting entities and relationships from text using manually designed rules or limited natural language processing tools. Ultimately, this knowledge point data is invoked in the form of service interfaces for downstream tasks such as question-answering systems and recommendation systems.

[0003] However, knowledge graphs constructed based on the above methods usually have the following drawbacks.

[0004] Low automation and high labor costs: Existing methods heavily rely on manual rule design and annotation for specific domains and format textbooks. Once the domain of the text materials changes, the original processing methods become ineffective, making it difficult to achieve batch processing across disciplines and documents. This strong dependence results in poor system generalization ability and low automation levels.

[0005] The knowledge graph suffers from a flat structure, lacking hierarchy and connections: traditional methods generate knowledge point sets that are mostly independent triples, lacking effective mechanisms to identify and organize the hierarchical structure between knowledge points (such as part-whole relationships). These deep semantic relationships are crucial for clearly conveying the structure of the knowledge system to users, assisting large language models in knowledge reasoning and argumentation, and accurately identifying users' knowledge blind spots. Without these relationships, the value of a knowledge graph is greatly diminished.

[0006] Insufficient utilization of multi-source heterogeneous data: Most existing solutions only process teaching materials and fail to effectively integrate and utilize multimodal teaching data (such as knowledge point mastery distribution data reflecting students' learning status, knowledge point course requirement documents providing knowledge point topics, and assisting in generating question lists for specific interest-related knowledge points, etc.). However, this multi-source heterogeneous data contains rich information and is of great value for in-depth exploration of the supporting relationships between knowledge points. Summary of the Invention

[0007] To address the above problems, this invention proposes a course knowledge point learning path recommendation system based on multimodal data.

[0008] The multimodal data-based knowledge point prerequisite relationship judgment system is structurally divided into two parts: a basic processing layer and an information alignment and reasoning layer. The basic processing layer is primarily responsible for extracting knowledge topics, knowledge points, and the conditional probabilities of mastery between integrated knowledge pairs from the original multimodal course materials. Subsequently, the aforementioned knowledge topics, knowledge points, and the conditional probabilities of mastery between integrated knowledge pairs are processed in the information alignment layer. This process structures the knowledge points into triples and aligns them with relevant knowledge topics. Finally, the judgment model predicts the prerequisite relationships between knowledge points, forming learning paths and recommending learning paths for the course knowledge points.

[0009] The multimodal course materials input to the basic processing layer consist of four parts: teaching material documents, question lists, course requirement documents, and knowledge point mastery distribution data. Different processing methods are required for different materials.

[0010] The course requirement document, question list, and knowledge point mastery distribution data do not require complex processing: The course requirement document refers to syllabus-like documents containing a large number of subject knowledge modules and topics. This system directly extracts the key topics from the course requirement document as the objects for subsequent triplet knowledge point alignment; the question list contains questions on the importance of the textbook, helping the large language model grasp the key points of the textbook, allowing the system to ignore a large number of invalid sentences in the textbook, reducing the burden on generation and inference tasks, and can be presented in plain text format; the knowledge point mastery distribution shows each student's mastery of a certain knowledge point, denoted as 1 if a student masters a knowledge point, and 0 otherwise. The knowledge point mastery distribution is key data for the subsequent knowledge point prerequisite relation determination task, providing key feature elements for the vector representation of knowledge points.

[0011] The processing of teaching materials involves four modules: material segmentation, knowledge point extraction, triple extraction, and verification and merging. Finally, valid and reasonable knowledge triples are sent to the information reasoning alignment layer for topic classification. The above operations require pre-configuration of the large language model and embedding model, necessitating model download or the use of mainstream API services.

[0012] Material Segmentation Module: To facilitate better extraction and summarization, the original teaching materials need to be segmented. First, a large language model is used to summarize each chapter, ensuring the consistency of the chapter's main idea is not disrupted by the segmentation process. Then, each independent chapter is segmented at the sentence level.

[0013] The knowledge point extraction module filters and discards a large number of irrelevant sentences based on the chapter summary and question list, extracting and storing only relevant knowledge statements from the text blocks. This reduces unnecessary workload for subsequent triple extraction. A large language model is used to atomically decompose the questions in the question list, resulting in a set of atomic question lists. An embedding model is then used to embed and fuse these atomic questions, obtaining relevant key vectors by summing and embedding each question from the chapter summary and atomic question lists. A similarity threshold is then set. Each sentence is multiplied by the key vector, and then compared with a similarity threshold. Compare and retain those with similarity greater than a threshold. The results are used to achieve filtering.

[0014] The triple extraction module transforms the sentences selected by the knowledge point extraction module into triple structures for storage. This process uses a large language model to perform entity and relation extraction from the text.

[0015] The verification module performs a validity check. During the validity check phase, the triplet structure data is extracted using regularization. If it does not conform to the regularization rules, the entity or relation is discarded. Then, each field is checked, including format checks, NOT NULL checks, validity checks, and validation checks. If a verification error occurs, all information related to that triplet is ignored.

[0016] The information reasoning alignment layer includes a knowledge point topic classification module and a prerequisite relationship judgment module. It processes the mastery conditional probabilities between knowledge topics, knowledge points, and integrated knowledge pairs, structurally represents knowledge points as triples, aligns them with relevant knowledge topics, and finally predicts the prerequisite relationships between knowledge points through a judgment model, forming learning paths and recommending learning paths for course knowledge points.

[0017] The knowledge point topic classification module is used to summarize the knowledge points from the triples output by the verification module into a hierarchical structure according to the course requirements. Based on the course requirements, k topics are extracted. A BERT encoder is used to semantically embed n knowledge points and k topic terms from the teaching materials document to obtain semantic vector representations of the topic term knowledge points. Based on the topic semantic vector sequence, a sample sequence is generated using data augmentation, and the two sequences are integrated. The minimum Euclidean distance between the elements of the integrated sequence is used to obtain a distance sequence, and the overall distance sequence elements are scaled using the scale of the element with the maximum Euclidean distance. The encoder classification model is trained by integrating the sequence and scaling factors to ensure that the classification mapping relationship holds.

[0018] The specific process is as follows: Deploying a pre-trained BERT model enables the conversion of natural language into vector encoding, transforming the topics or knowledge points represented by the triples output by the verification module into token symbols. <cls>Prefix symbols and <sep>Sequence representation of suffix symbols; using the sequence 's' to represent keywords or knowledge points, and 'X' to represent the vector sequence after 's' is encoded by the BERT encoder: The keywords and knowledge points are used as inputs to the BERT encoder, and the first vector after encoding is taken as follows: , as keywords or knowledge points Vector representation. A set of topic vectors is obtained by traversing and integrating all keywords and knowledge points. With knowledge point vector set .

[0019] Topic Vectors Add random noise to obtain m data-enhanced samples. , the original sample With enhanced samples The combination yields a new sequence of vectors with index k. This is called the k-th sample group. Traverse all indices to obtain a sequence of K new vector groups. Then, integrate these K new vector group sequences according to index i to obtain a sequence, such as... , and record as Furthermore, in the above sequence Find the corresponding distance sequence based on , of which elements for elements in the sequence With elements The minimum Euclidean distance is the minimum distance between elements that do not have a subset of data in the current dataset. Finally, the distance is scaled for each element of N. The above , , is the negation symbol. The remainder symbol is used. Represents a sequence The largest element in the system.

[0020] Furthermore, the above The sequence serves as input to an encoding classification model (consisting of a transformer encoder layer and a classification task head), whose parameters need to be trained. Vector The encoded vector is obtained after passing through the transformer layer. The classification head then processes the data to obtain the final classification result. The classification head parameters and encoder layer parameters are trained independently: the classification head parameters are updated using the cross-entropy loss function; the encoder layer parameters are updated using the loss function defined by the following formula, where... For index functions, when the index The value is 1 when the index is 1. The time value is 0.

[0021] Through topic vector sets Complete the training of the encoding classification model; convert the knowledge point vectors... As input to the trained encoding classification model, the topic classification result is obtained through prediction.

[0022] Since topic classification tasks lack readily available training samples, this module trains model parameters by constructing samples, significantly reducing the cost of manually compiling samples. Furthermore, recognizing the difficulty in classifying subject knowledge points due to the similarity and difficulty in distinguishing their semantic vectors, this module utilizes parameters... The topic semantic vector is dynamically learned and adjusted to better learn and classify similar topics.

[0023] Priority relationship judgment module: Use elements Let represent the i-th element of the subsequence of knowledge point A and knowledge point B respectively. Combining knowledge points A and B results in a sequence. Input the query sequence into the BERT encoder and extract its corresponding query sequence from the last transformer layer of the encoder. key sequence , means as follows: Taking knowledge point pair A and B as an example, extract the subsequence corresponding to the index. , and , And the summation and mean are used to represent knowledge point A. and Vectors; similarly, we can obtain knowledge point B. and .use These represent the sequence A in its corresponding The starting and ending indices of the sequence; using These represent the B sequence in its corresponding The sequence contains the starting and ending indices. Furthermore, as described in the knowledge point topic classification module, knowledge points A and B are used as separate inputs to obtain their encoded representations. The semantically relevant features are obtained by performing click operations on the query vectors and key vectors of the two knowledge points A and B mentioned above. and ; using the classification codes of A and B to obtain ,in This is a normalized click operation. Semantic relevance is also considered. , and classification code value , It has semantic classification features and , , , The adjustment coefficient is set to a custom value. The semantic classification features of the remaining pairwise knowledge points are handled in the same way as described above.

[0024] The distribution data of knowledge points were statistically analyzed and summarized into the following data table, with symbols A, B, C, and D representing the corresponding knowledge points, as shown in Table 1.

[0025] Table 1

[0026] The prior knowledge relationship determination is based on pairwise judgment to achieve a global prior knowledge relationship determination. P(A) represents the probability of mastering knowledge point A, P( P(B|A) represents the probability of not mastering knowledge point A; P(B|A) represents the probability of mastering knowledge point B given that knowledge point A is mastered; the above probabilities are all derived from the distribution data of mastery, and the rest are similar.

[0027] Using the statistical table and semantic classification features mentioned above, we construct pairwise feature vectors. The following section uses knowledge point A and knowledge point B as references to construct feature vectors. , This represents the feature vector where knowledge point A is a prerequisite for knowledge point B. (Select) The elements in each row and column constitute the aforementioned pre-prepared eigenvector. This feature vector can be considered as consisting of 5 parts and a total of 17 elements: the first part, which respectively grasps... Under the condition of mastering Conditional probability Part Two, mastering each As a condition not known Conditional probability Part Three, not mastered Under the condition of mastering Conditional probability Part Four, not mastered Even under these conditions, it was not mastered. Conditional probability Part 5, Semantic Classification Feature Elements The first four parts each contain four elements, ordered according to the priority of prerequisite knowledge points, which corresponds to the order in which the conditional probability values ​​are described. Integrating these five parts yields the final representation of the prerequisite feature vector. And iterate through all pairs of knowledge points; there are n knowledge points that can yield n results. (n-1) prior eigenvectors.

[0028] The prior relation prediction module consists of a multilayer perceptron (MLP) and a fractional mapping (S) part. The MLP maps prior feature vectors to scalars, and the mapping relationship is as follows: To ensure the antisymmetry of the output and achieve the three-class classification task, a score mapping is designed. . With ordered vector pairs For example, there is a mapping On the other hand, there are ordered vector pairs. , mapping .pass The input order guarantees This classifies two opposing cases into opposite results. The sg function consists of the hyperbolic tangent function tanh(z) and a custom threshold. The common definition is expressed as follows: .

[0029] Beneficial effects of this invention: Compared with traditional technical solutions, the method and system for determining the prior relationship of knowledge points based on multimodal data proposed in this invention have the following significant advantages and beneficial effects: This invention significantly reduces the traditional methods of manually designing rule paradigms and manually annotating structured teaching resources. By integrating advanced natural language processing and machine learning technologies, it achieves automated extraction of knowledge points and relationships from unstructured or semi-structured teaching documents. The system possesses powerful cross-document and cross-disciplinary batch processing capabilities, effectively solving the core pain points of low efficiency and poor scalability in knowledge point acquisition.

[0030] Building upon automated extraction, the system introduces an innovative hierarchical induction mechanism. Through a knowledge point topic classification module, it automatically categorizes scattered and disorganized knowledge points into a unified topic framework based on course requirements, forming a hierarchical knowledge structure with hierarchical relationships. Furthermore, the prerequisite relationship judgment module collaboratively processes multi-source heterogeneous teaching data, thereby more accurately and reliably inferring the implicit prerequisite relationships and dependencies between knowledge points, ultimately leading to recommended learning paths for each knowledge point. Attached Figure Description

[0031] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent from the following detailed description taken in conjunction with the accompanying drawings. Several embodiments of the invention are illustrated in the drawings by way of example and not limitation, wherein: Figure 1 This diagram illustrates the overall architecture and processing flow of an embodiment of the present invention. Figure 2 This invention illustrates a flowchart of the knowledge point topic classification module according to an embodiment of the invention. Figure 3 A flowchart of the prior relationship determination module according to an embodiment of the present invention is shown; Figure 4 The experimental results loss curves of an embodiment of the present invention are shown. Detailed Implementation

[0032] Embodiments of the present invention will now be described with reference to several exemplary embodiments. It should be understood that these embodiments are provided merely to enable those skilled in the art to better understand and implement the present invention, and are not intended to limit the scope of the invention in any way. Rather, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the invention to those skilled in the art. The present invention will be further explained and described below in conjunction with the accompanying drawings: The technical solutions of the embodiments of this application are described in detail below with reference to the accompanying drawings.

[0033] Figure 1 This invention illustrates a course knowledge point learning path recommendation system based on multimodal data, according to an embodiment of the present invention. See Figure 1 This method uses many modular units, which can be summarized as a basic processing layer and an information alignment processing layer. The basic processing layer mainly focuses on extracting relevant information from the document to be processed; the information alignment processing layer mainly focuses on information classification and prediction and the integration of different heterogeneous information sources.

[0034] The basic processing layer will complete the extraction of three types of information: extracting relevant keywords from the course requirements document, extracting triplet knowledge points by combining the question list and teaching material documents, and integrating the conditional probabilities of all pairwise knowledge points from the knowledge point mastery distribution data.

[0035] The multimodal course materials input to the basic processing layer consist of four parts: teaching material documents, question lists, course requirement documents, and knowledge point mastery distribution data. Different processing methods are required for different materials.

[0036] The information alignment processing layer primarily focuses on information prediction and integration. The knowledge point topic classification module aggregates knowledge points in triplet form into corresponding knowledge topics. Two knowledge topics constitute a classification judgment task, and by constructing all classification tasks, all corresponding prerequisite relationship predictions are obtained, and knowledge point learning path recommendations are built.

[0037] The information reasoning alignment layer includes a knowledge point topic classification module and a prerequisite relationship judgment module. It processes the mastery conditional probabilities between knowledge topics, knowledge points, and integrated knowledge pairs, structurally represents knowledge points as triples, aligns them with relevant knowledge topics, and finally predicts the prerequisite relationships between knowledge points through a judgment model, forming learning paths and recommending learning paths for course knowledge points.

[0038] Figure 2 The following flowchart illustrates a knowledge point topic classification module according to an embodiment of the present invention, which includes the following processes: Semantic vector representation of topics and knowledge points: Topics and knowledge points in natural language form are used as input to a pre-trained BERT dataset. A specific topic or knowledge point undergoes semantic encoding by BERT to obtain its vector sequence X. The first position of the vector sequence X, i.e. , as a semantic expression of the aforementioned themes or semantic expression of knowledge points By applying all the topics and knowledge points to the above process, we obtain a set of topic word vectors. With knowledge point vector set .

[0039] Data augmentation and sequence integration :by Using random noise as the anchor point, a set of m data-augmented samples is generated. , anchor point With data augmentation samples Combining to obtain Iterate through all indices i to obtain k new vector sequences. Then, integrate these k sequences according to their index i to obtain a new sequence, denoted as . .

[0040] Distance sequence and scaling factor wi: sequential Using an element as a basis, find the minimum Euclidean distance between that element and all other elements not in the same group to obtain a distance sequence. Distance sequence Divide all elements of the sequence by the largest value in the sequence to obtain the corresponding scaling factor. .

[0041] Model parameter training: through Sequence and scaling factor The encoding classification model is trained, and the model weights are obtained through backpropagation using the loss function. The model training is then complete.

[0042] Topic Classification: Knowledge Point Vector Set As input to the above model, a classification prediction for each piece of knowledge is obtained.

[0043] Figure 3 The following flowchart illustrates a priori relationship determination module according to an embodiment of the present invention, which includes the following processes: Feature element acquisition: This includes the following two aspects: obtaining the query and key vector representation and classification encoding of knowledge points; and summarizing the distribution features of knowledge point mastery.

[0044] Knowledge point query, key vector representation, and classification encoding acquisition: All knowledge points are used individually as input to the knowledge point topic classification module to obtain their corresponding codes. Furthermore, every two knowledge points form a judgment group. Each judgment group combines the sequence of knowledge points in order to form its semantic sequence representation. The corresponding query and key vector sequences are then extracted using the last layer of the BERT encoder's transformer. The query and key vector representations for each knowledge point are obtained by averaging the corresponding query and key vector groups according to their sequence indices. Each query and key vector group is then subjected to dot-multiplication normalization, and the corresponding encoding is also subjected to dot-multiplication normalization to finally obtain all semantically relevant features. Coefficients are set. Obtain semantic classification features .

[0045] Summary of Knowledge Point Mastery Distribution Characteristics: Statistical analysis of the mastery distribution data yields the conditional probability values ​​of each pair of knowledge points. All conditional probability values ​​related to knowledge points within the discrimination group are then arranged according to a pattern to obtain a knowledge point conditional probability sequence.

[0046] Feature vector representation of knowledge point pairs: The semantic classification features mentioned above... Add it to the end of the knowledge point conditional probability sequence to obtain the prior feature vector.

[0047] First, train the parameters of the relationship prediction model: prepare a labeled dataset for model training and learning, and obtain the corresponding model weights.

[0048] Prerequisite relationship prediction and classification: All judgment groups are used as input, and the prerequisite relationships between knowledge points are judged in pairs to obtain the prerequisite relationships of all knowledge points.

[0049] experiment: To verify the effectiveness of the proposed method, it was compared with two baselines on the LecureBank dataset (1202 data points in total). The evaluation metrics included Accuracy (ACC), F1-score (F1) and AUC.

[0050] Commonly used Christ learning classifiers were selected as the traditional baseline, including: NB (Naive Bayes), SVM (Support Vector Machine), and RF (Random Forest Ensemble).

[0051] Graph representation and graph neural network baselines were selected: MHAVGAE: A graph representation learning model based on Variational Graph Autoencoder (VGAE) (including multi-head attention mechanism / variants) that learns latent representations of nodes / graphs and performs downstream classification.

[0052] HGAPNet: A baseline for hierarchical graph attention networks that aggregates neighborhood information through attention mechanisms to model more complex structural dependencies.

[0053] LCPRE: Structural Representation Learning Baseline (for capturing representations of local / global structural features), serving as a contrast to another class of graph representation methods.

[0054] DGCPL: Jointly models the structural dependencies of nodes / relationships on graph-structured data, learns the latent representations of samples through graph convolution / message passing mechanisms, and performs downstream classification predictions based on these representations. The loss result is as follows: Figure 4 As shown in Table 1 below: Table 1 < / sep> < / cls>

Claims

1. A course knowledge point learning path recommendation system based on multimodal data, characterized in that, It consists of two parts: a basic processing layer and an information alignment and inference layer. The basic processing layer is used to extract knowledge topics, knowledge points, and mastery conditional probabilities between integrated knowledge pairs from the original multimodal course materials. The information alignment reasoning layer is used to process the mastery conditional probabilities between knowledge topics, knowledge points, and integrated knowledge pairs. It represents knowledge points in a structured form as triples and aligns them with relevant knowledge topics. Finally, it predicts the prerequisite relationships between knowledge points through a judgment model, forms a learning path, and completes the recommendation of learning paths for course knowledge points.

2. The course knowledge point learning path recommendation system based on multimodal data according to claim 1, characterized in that, The input multimodal course materials for the basic processing layer consist of four parts: teaching materials documents, a list of questions, course requirements documents, and knowledge point mastery distribution data.

3. The course knowledge point learning path recommendation system based on multimodal data according to claim 2, characterized in that, In the basic processing layer, the course requirement document, question list, and knowledge point mastery distribution are processed as follows: The course requirements document refers to the syllabus-like document material, which contains subject knowledge and knowledge topics. The key topics in the course requirements document are extracted and used as the objects for subsequent triplet knowledge point alignment; the question list contains questions on the importance of the textbook, which help the large language model to grasp the key points of the textbook, in plain text form. The distribution of knowledge point mastery represents each student's mastery of relevant knowledge points; mastering a knowledge point is recorded as 1, and not mastering it is recorded as 0.

4. The course knowledge point learning path recommendation system based on multimodal data according to claim 2, characterized in that, The basic processing layer includes a material segmentation module, a knowledge point extraction module, a triple extraction module, and a verification and merging module for processing teaching material documents. The material segmentation module is used to segment the original teaching material text into blocks. First, a large language model is used to summarize each chapter to obtain a chapter summary. Then, each independent chapter is segmented into blocks at the sentence level. Knowledge Extraction Module: Used to filter and discard irrelevant sentences based on chapter summaries and question lists, and extract and store relevant knowledge statements from text blocks; The large language model is used to atomically decompose the questions in the question list to obtain a set of atomic question lists. An embedding model is then used for embedding fusion, combining the chapter summary and each question in the atomic question list to obtain relevant key vectors. A similarity threshold is then set. Each sentence is multiplied by the key vector, and then compared with a similarity threshold. Compare and retain those with similarity greater than a threshold. The results are used to achieve filtering; The triple extraction module converts the sentences selected by the knowledge point extraction module into triple structures and saves them. This process uses a large language model to extract entities and relations from the text. The verification module is used to perform legality checks. During the legality check phase, the triple structure data is extracted using regularization. If it does not conform to the regularization rules, the entity or relation is discarded first. Then, each field is checked, including format check, non-empty check, legality check, and validity check. If a verification check error occurs, all information related to the triple is ignored.

5. The course knowledge point learning path recommendation system based on multimodal data according to claim 2, characterized in that, The information reasoning alignment layer includes a knowledge point topic classification module and a prerequisite relationship judgment module.

6. The course knowledge point learning path recommendation system based on multimodal data according to claim 5, characterized in that, The knowledge point topic classification module is used to summarize the knowledge points from the triples output by the verification module according to the course requirements, forming a hierarchical structure. The specific process is as follows: Deploy the pre-trained BERT model, and use tokens to represent the topics or knowledge points in the triples output by the verification module. <cls>Prefix symbols and <sep> Sequence representation of suffix symbols; using 's' sequences to represent keywords or knowledge points;< / sep> < / cls> The topic and knowledge points are used as inputs to the BERT encoder, and the first vector after encoding is taken as follows: , as keywords or knowledge points Vector representation; traversing and integrating all topics and knowledge points to obtain a set of topic vectors. With knowledge point vector set ; Topic Vectors Add random noise to obtain m data-enhanced samples. , the original sample With enhanced samples The combination yields a new sequence of vectors with index k. Iterate through all indices to obtain k new vector sequences, and then integrate the k new vector sequences in order of index i to obtain the sequence. ; Additionally, in the sequence Find the corresponding distance sequence based on , of which elements for elements in the sequence With elements The minimum Euclidean distance, and for Scaling each element ; in the above , , is the negation symbol. The remainder symbol is used. Represents a sequence The largest element in; The above Sequences are used as input to the encoding classification model, and vectors are used as input. The encoded vector is obtained after passing through the transformer layer. The classification head is used to obtain the final classification result; the classification head parameters and encoder layer parameters are trained independently: the classification head parameters are updated backward using the cross-entropy loss function; the encoder layer parameters are updated backward using the loss function defined by the following formula, where... For index functions, when the index The value is 1 when the index is 1. The time value is 0: Through topic vector sets Complete the training of the coding classification model; knowledge point vectors As input to the trained encoding classification model, the topic classification result is obtained through prediction.

7. The course knowledge point learning path recommendation system based on multimodal data according to claim 6, characterized in that, The prior relationship judgment module uses elements Let these represent the i-th elements of the subsequences of knowledge point A and knowledge point B, respectively. Combining knowledge point A and knowledge point B results in a sequence. Input the query sequence into the BERT encoder and extract its corresponding query sequence from the last transformer layer of the encoder. key sequence : For knowledge point pairs A and B, extract the subsequence corresponding to the index. , and , And the summation and mean are used to represent knowledge point A. and Vectors; similarly, this leads to knowledge point B. and ;use These represent the sequence A in its corresponding The starting and ending indices of the sequence; using These represent the B sequence in its corresponding The sequence includes the starting and ending points of the index; furthermore, as described in the knowledge point topic classification module, knowledge points A and B are used as separate inputs to obtain their encoded representations. The semantically relevant features are obtained by performing click operations on the query vectors and key vectors of the two knowledge points A and B mentioned above. and ; using the classification codes of A and B to obtain ,in To normalize click operations; to integrate semantic relevance. , and classification code value , It has semantic classification features and , , , Custom adjustment coefficient; Based on semantic classification features and Construct feature vectors and predict the prior relation prediction output.

8. The course knowledge point learning path recommendation system based on multimodal data according to claim 7, characterized in that, The semantic classification feature and The feature vector is constructed, and the prior relation prediction is as follows: The prior knowledge relationship determination is based on pairwise judgment to achieve a global prior knowledge relationship determination; P(A) represents the probability of mastering knowledge point A, P( P(B|A) represents the probability of not mastering knowledge point A; P(B|A) represents the probability of mastering knowledge point B given that knowledge point A is mastered; the above probabilities are all derived from the distribution data of mastery, and the rest are similar; Construct feature vectors using knowledge point A and knowledge point B. , This represents the feature vector of knowledge point A as a prerequisite for knowledge point B; select The elements in each row and column constitute the aforementioned pre-prepared eigenvector. ; The first part of this feature vector, respectively, grasps... Under the condition of mastering Conditional probability Part Two, mastering each As a condition not known Conditional probability Part Three, not mastered Under the condition of mastering Conditional probability Part Four, not mastered Even under these conditions, it was not mastered. Conditional probability ; Part 5, Semantic Classification Feature Elements The first four parts each contain four elements, ordered according to the priority of prerequisite knowledge points, which corresponds to the order in which the conditional probability values ​​are described. Integrating these five parts, the final result is the expression of the prerequisite feature vector. And iterate through all pairs of knowledge points; there are n knowledge points that can yield n results. (n-1) prior eigenvectors; The prior relation prediction module includes a multilayer perceptron (MLP) and a fractional mapping (S) part; the MLP maps prior feature vectors to scalars. Design score mapping ;in The function consists of the hyperbolic tangent function tanh(z) and a custom threshold. The common definition is expressed as follows: .

Images