MOOC recommendation method and device combining large language model and knowledge space reasoning

Abstract

The present invention belongs to the field of MOOC video recommendation, and specifically relates to a MOOC recommendation method and device that combines a large language model and knowledge space reasoning. The method uses the text of a test question to perform data enhancement based on the large language model to obtain enhanced text. The enhanced text and the test question text are converted into vectors and then trained into an embedding model to obtain an optimized embedding training model and the user test question embedding vector. The method obtains knowledge mastery and classification results based on the user's historical learning records, obtains a preference score based on learning monitoring data, constructs a complete knowledge structure based on video collections, knowledge points, and corresponding mapping relationships, obtains a personalized knowledge structure based on the user's learning behavior, calculates the target knowledge state, obtains inner and outer edge video collections, and recommends videos to the user. The method utilizes knowledge enhancement and knowledge space reasoning to obtain more accurate learners' knowledge mastery, video preferences, and knowledge structure characteristics, and recommends MOOC videos that meet user needs.

Description

Technical Field

[0001] The present invention belongs to the field of MOOC video recommendation, and in particular relates to a MOOC recommendation method and device combining a large language model and knowledge space reasoning. Background Art

[0002] In recent years, with the continued advancement of education's digital transformation, online learning, represented by MOOCs, has grown rapidly, with the number of MOOC platform users and video resources exploding. Faced with this massive amount of video resources, learners face information overload and knowledge confusion, making it difficult to quickly and accurately find content that meets their needs.

[0003] In related technologies, MOOC video recommendations primarily employ the following approach: A BERT encoder combined with a multi-head attention mechanism deeply mines the semantic information of MOOC video subtitles. Long-short-term memory neural networks are then used to analyze learner interaction sequences, dynamically constructing learners' knowledge preferences. Finally, video collections are recommended based on these preferences and the similarities between MOOC videos extracted using the attention mechanism. However, this method does not accurately assess learners' knowledge mastery, failing to truly reflect their cognitive state and recommend MOOC videos that match their cognitive level. Alternatively, a deep knowledge tracking model is used to capture learners' cognitive state, a temporal recommendation model is used to capture learners' learning content preferences, and inter-video correlations are integrated to enhance video recommendations. This method mines MOOC video correlations using learners' behavioral data and quantifies the degree of correlation between MOOC video resources using the transition probabilities of learners' viewing behavior sequences. However, this approach does not consider the relationships between MOOC videos from a knowledge structure perspective.

[0004] Regarding the aforementioned related technologies, existing MOOC video recommendation methods based on semantic mining of MOOC video subtitles are limited to the subtitles themselves, failing to accurately assess learners' knowledge mastery, truly reflect their cognitive state, and recommend MOOC videos that match their cognitive level and knowledge structure. Furthermore, existing MOOC recommendation research has not yet explored the use of large language models to enhance text data, nor has it mined and utilized implicit knowledge point relationships within test texts. This has impacted the accuracy of predicting learners' knowledge mastery, and consequently, has limited the effectiveness of MOOC video recommendations. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a MOOC recommendation method and device that combines a large language model and knowledge space reasoning. Based on the knowledge enhancement and knowledge space reasoning of the large language model, a more accurate learner's knowledge point mastery, video preferences and knowledge structure characteristics are obtained, thereby recommending MOOC videos that are more in line with the user's cognitive level and knowledge structure.

[0006] A MOOC recommendation method that combines a large language model with knowledge space reasoning, including:

[0007] Establish a test question knowledge point association diagram based on the training test question database;

[0008] Obtain the test text, perform data augmentation on the test text based on the large language model o1-mini to obtain an enhanced text, wherein the enhanced text includes a parsed text and several similar texts, and perform text embedding processing on the test text, the parsed text, and the several similar texts to obtain a test text vector, a parsed text vector, and an aggregated similar text vector;

[0009] Obtain a graph embedding vector based on the test question knowledge point association graph, and obtain a test question embedding vector based on the graph embedding vector and the question text vector;

[0010] The embedding model is trained using the question text vector, the parsed text vector, the aggregated similar text vector, the graph embedding vector, and the test question embedding vector to obtain an optimized embedding training model;

[0011] Obtain the user's historical learning records, input the test question database into the optimized embedding training model to obtain all test question embedding vectors, and filter all test question embedding vectors based on the historical learning records to obtain the user's test question embedding vector;

[0012] Calculate the knowledge mastery based on historical learning records and the user's test question embedding vector;

[0013] Classify users according to the historical learning records to obtain classification results;

[0014] Obtaining the user's learning monitoring data, calculating the user's concentration score and emotional state score based on the learning monitoring data, calculating the engagement score based on historical learning records, and calculating the preference score based on the concentration score, emotional state score, and engagement score;

[0015] Obtain a video set and the knowledge points corresponding to each video in the video set to form a knowledge point set, obtain a mapping relationship between the video set and the knowledge point set, construct a triple based on the video set, the knowledge point set, and the mapping relationship, generate all knowledge states based on the triple and the disjunctive model, and construct a complete knowledge structure;

[0016] Obtain users' learning behavior and viewing history;

[0017] Obtaining proficiency of different knowledge points based on the viewing history and learning behavior;

[0018] According to the learning behavior, the video set watched by the user, the knowledge point set corresponding to the video set watched by the user, and the corresponding mapping relationship, a plurality of personalized knowledge states are obtained to form a personalized knowledge structure;

[0019] Constructing a knowledge structure matrix based on the individual knowledge structure and the proficiency, calculating the similarity between the knowledge mastery and each column vector in the knowledge structure matrix, obtaining the individual knowledge state with the highest similarity as the target knowledge state, constructing a knowledge state boundary based on the target knowledge state and the complete knowledge structure, and filtering a video set based on the knowledge state boundary to obtain an outer edge video set and an inner edge video set;

[0020] A first recommended video set is obtained based on the classification results and the outer edge video set, and a second recommended video set is obtained based on the preference score and the inner edge video set. The first recommended video set and the second recommended video set are used as recommendation results.

[0021] Optionally, obtaining the test question text, performing data enhancement on the test question text based on a large language model to obtain enhanced text, wherein the enhanced text includes a parsed text and a plurality of similar texts, and performing text embedding processing on the test question text, the parsed text, and the plurality of similar texts to obtain a test question text vector, a parsed text vector, and an aggregated similar text vector includes:

[0022] Based on the question text, generate a parsed text and several similar texts based on the large language model o1-mini;

[0023] Among them, as of November 2024, the o1 series model to which o1-mini belongs is the latest large language model launched by OpenAI;

[0024] The text embedding process includes:

[0025] According to the title text, a word embedding matrix is ​​obtained;

[0026] Get the trainable vector;

[0027] Obtaining an expanded word embedding matrix according to the word embedding matrix and the trainable vector;

[0028] Input the expanded word embedding matrix into the BERT encoding model to obtain the title text vector;

[0029] Performing text embedding processing on the parsed text and a plurality of similar texts to obtain a parsed text vector and a plurality of similar text vectors;

[0030] Aggregate several similar text vectors to obtain an aggregated similar text vector.

[0031] Optionally, the using the question text vector, the parsed text vector, the aggregated similar text vector, the graph embedding vector, and the test question embedding vector to train the embedding model to obtain the optimized embedding training model includes:

[0032] According to the correlation between the test questions and the knowledge points, a positive sample set and a negative sample set are obtained;

[0033] Obtaining a first loss according to the question text vector, the parsed text vector, and a first loss function;

[0034] Obtaining a second loss according to the positive sample set, the negative sample set, the aggregated similar text vector, the title text vector, and a second loss function;

[0035] Obtaining a third loss according to the positive sample set, the negative sample set, the graph embedding vector, the aggregated similar text vector, and a third loss function;

[0036] Get a multilayer perceptron;

[0037] Predicting a difficulty value of the test question embedding vector based on the multi-layer perceptron;

[0038] Obtaining a fourth loss based on the predicted difficulty value, the actual difficulty value of the test question, and a fourth loss function;

[0039] summing the first loss, the second loss, the third loss, and the fourth loss to obtain a total loss;

[0040] The model parameters of the embedding model are adjusted according to the total loss to obtain an optimized embedding training model.

[0041] Optionally, the first loss function is expressed as:

[0042] ;

[0043] in, is the title text vector, For test questions The parsed text vector of is the temperature parameter, N is the number of test questions in a batch during training, sim() is the cosine similarity calculation function, To remove the test questions Any test question The temperature parameter is used to adjust the distribution of similarity scores and control the model's ability to distinguish between positive and negative samples.

[0044] The second loss function is expressed as:

[0045] ;

[0046] in, To aggregate similar text vectors, It's a test question The positive sample set, It's a test question A set of negative samples;

[0047] The third loss function is expressed as:

[0048] ;

[0049] ;

[0050] in, and They are respectively the trainable weight matrix and bias coefficient, which are used to map the question node embedding vector to the text embedding space. For test questions The corresponding graph embedding vector;

[0051] The fourth loss function is expressed as:

[0052] ;

[0053] in, is the actual difficulty value of the test question, is the prediction difficulty value.

[0054] Optionally, calculating the knowledge mastery based on historical learning records and the user's test question embedding vector includes:

[0055] According to the historical learning record, the answer record is obtained, and the user's time step is obtained according to the answer record. t The degree of mastery of all knowledge points;

[0056] Performing interactive embedding representation on the answer record, and obtaining the dimension of the interactive embedding representation and the dimension of the test question embedding vector;

[0057] Constructing a static matrix according to the dimension of the test question embedding vector and the number of knowledge points;

[0058] Construct a dynamic matrix based on the dimension of the interaction embedding representation and the number of knowledge points;

[0059] A static weight is calculated based on the user test question embedding vector and the static matrix;

[0060] Get the erase vector and increment vector;

[0061] The dynamic matrix is ​​updated according to the erasure vector, the addition vector and the static weight to obtain an updated dynamic matrix;

[0062] The knowledge mastery degree is obtained according to the updated dynamic matrix.

[0063] Optionally, the obtaining of a plurality of personalized knowledge states based on the learning behavior, the video set watched by the user, the knowledge point set corresponding to the video set watched by the user, and the corresponding mapping relationship to form a personalized knowledge structure includes:

[0064] Calculate the user's proficiency in different knowledge points based on the learning behavior;

[0065] According to the proficiency, the set of videos watched by the user, the set of knowledge points corresponding to the set of videos watched by the user, and the corresponding mapping relationship, a plurality of personalized knowledge states are obtained through a disjunctive model to form a personalized knowledge structure.

[0066] Optionally, constructing a knowledge structure matrix based on the individual knowledge structure and the proficiency, calculating the similarity between the knowledge mastery and each column vector in the knowledge structure matrix, and obtaining the individual knowledge state with the highest similarity as the target knowledge state includes:

[0067] Constructing a knowledge structure matrix based on the individual knowledge structure and the proficiency, wherein rows represent knowledge points, columns represent knowledge states, and elements in the knowledge structure are 0 or proficiency;

[0068] Input the knowledge mastery degree and each column vector in the knowledge structure matrix into a similarity calculation formula to obtain several similarities, and select the individual knowledge state with the highest similarity as the target knowledge state;

[0069] The similarity calculation formula is:

[0070] ;

[0071] in, KM is the knowledge mastery, D is the knowledge structure matrix, is the learner's mastery level vector KM No. elements, is the number of knowledge points, For the The first of the personality knowledge states The proficiency of each knowledge point.

[0072] A MOOC recommendation device that combines a large language model and knowledge space reasoning, comprising:

[0073] Establishing a module for establishing a test question knowledge point association diagram based on a training test question database;

[0074] An enhancement module is used to obtain the test question text, perform data enhancement on the question text based on the large language model o1-mini to obtain enhanced text, wherein the enhanced text includes the parsed text and several similar texts, and perform text embedding processing on the question text, the parsed text, and the several similar texts to obtain a question text vector, a parsed text vector, and an aggregated similar text vector;

[0075] A fusion module is used to obtain a graph embedding vector based on the test question knowledge point association graph, and obtain a test question embedding vector based on the graph embedding vector and the question text vector;

[0076] A training module, configured to train an embedding model using the question text vector, the parsed text vector, the aggregated similar text vector, the graph embedding vector, and the test question embedding vector to obtain an optimized embedding training model;

[0077] The first acquisition module is used to obtain the user's historical learning records, input the test question database into the optimized embedding training model to obtain all test question embedding vectors, and filter all test question embedding vectors according to the historical learning records to obtain the user's test question embedding vector;

[0078] A first calculation module is used to calculate the knowledge mastery based on the historical learning records and the user's test question embedding vector;

[0079] A classification module, configured to classify users according to the historical learning records and obtain classification results;

[0080] A second acquisition module is configured to acquire the user's learning monitoring data, calculate the user's concentration score and emotional state score based on the learning monitoring data, calculate the user's engagement score based on historical learning records, and calculate the user's preference score based on the concentration score, emotional state score, and engagement score;

[0081] A third acquisition module is used to obtain a video set and the knowledge points corresponding to each video in the video set to form a knowledge point set, obtain a mapping relationship between the video set and the knowledge point set, construct a triple based on the video set, the knowledge point set, and the mapping relationship, generate all knowledge states based on the triple and the disjunctive model, and construct a complete knowledge structure;

[0082] The fourth acquisition module is used to obtain the user's learning behavior and viewing history;

[0083] A second calculation module is used to obtain the proficiency of different knowledge points based on the viewing history and learning behavior;

[0084] A third calculation module is configured to obtain a plurality of personalized knowledge states based on the learning behavior, the video set watched by the user, the knowledge point set corresponding to the video set watched by the user, and the corresponding mapping relationship to form a personalized knowledge structure;

[0085] a screening module, configured to construct a knowledge structure matrix based on the individual knowledge structure and the proficiency, calculate the similarity between the knowledge mastery and each column vector in the knowledge structure matrix, obtain the individual knowledge state with the highest similarity as the target knowledge state, construct a knowledge state boundary based on the target knowledge state and the complete knowledge structure, and screen a video set based on the knowledge state boundary to obtain an outer edge video set and an inner edge video set;

[0086] The recommendation module is used to obtain a first recommended video set based on the classification results and the outer edge video set, obtain a second recommended video set based on the preference score and the inner edge video set, and use the first recommended video set and the second recommended video set as recommendation results.

[0087] A terminal device includes a memory and a processor. The memory stores a computer program that can be run on the processor. When the processor loads and executes the computer program, it adopts a MOOC recommendation method that combines a large language model and knowledge space reasoning.

[0088] A computer-readable storage medium stores a computer program. When the computer program is loaded and executed by a processor, a MOOC recommendation method combining a large language model and knowledge space reasoning is adopted.

[0089] The beneficial effects of the present invention are:

[0090] 1. Using the test question text, perform data enhancement on the test question text based on the large language model to obtain enhanced text, and then perform vector conversion on the enhanced text and the test question text to train the embedding model to obtain an optimized embedding training model, and then obtain the user test question embedding vectors of all test questions in the test question database, obtain the knowledge mastery and classification results based on the user's historical learning records, obtain the preference score based on the learning monitoring data, construct a complete knowledge structure based on the video collection, knowledge points and corresponding mapping relationships, and then combine the user behavior and viewing history to obtain the personalized knowledge structure, and according to the personalized knowledge structure, obtain the personalized knowledge state with the highest similarity as the target knowledge state, and then obtain the outer edge video set and the inner edge video set according to the boundary of the target knowledge state, and obtain the first recommended video set according to the classification result and the outer edge video set, and obtain the second recommended video set according to the preference score and the inner edge video set, and use the first recommended video set and the second recommended video set as the recommendation results. Compared with traditional video recommendations, this application uses data enhancement based on a large language model to enhance the test text, and combines the comparative learning training model to mine the direct, indirect, and high-order knowledge point associations and potential difficulty information in the test text data to obtain a more accurate learner's knowledge point mastery level, thereby recommending MOOC videos that are more in line with their cognitive level to learners.

[0091] 2. Introducing knowledge space reasoning to screen MOOC videos, the system recommends video resources that match learners' knowledge status and structure, while also meeting their personalized needs for consolidating and expanding their learning. This system leverages video features (e.g., the correspondence between different video clips and different knowledge points) and learners' personalized interactive behaviors (e.g., actual video viewing segments, duration, speed, and repeat viewing) and integrates these into the knowledge space reasoning framework to obtain a personalized knowledge structure and state that better reflects the learners' actual circumstances. This system then recommends video resources that meet learners' needs for consolidation and expansion, providing a personalized, intelligent MOOC video recommendation service. BRIEF DESCRIPTION OF THE DRAWINGS

[0092] Figure 1 The figure is a flowchart of a MOOC recommendation method combining a large language model and knowledge space reasoning according to the present invention. DETAILED DESCRIPTION

[0093] A MOOC recommendation method combining large language model and knowledge space reasoning, such as Figure 1 As shown, the present invention includes:

[0094] S1. Establish a test question knowledge point association diagram based on the training test question database.

[0095] Specifically, based on the Q matrix, a question-knowledge point association diagram is created The construction of Is the set of vertices, including all the test questions and knowledge points , It's a picture The undirected edge set of and knowledge points The Q matrix records the association information between test questions and knowledge points, which is given by domain experts. Indicates test questions Examined knowledge points ,and Indicates test questions No knowledge points involved .

[0096] S2. Obtain the test question text, perform data enhancement on the test question text based on a large language model to obtain enhanced text, which includes parsed text and several similar texts, and perform text embedding processing on the test question text, parsed text, and several similar texts to obtain a test question text vector, a parsed text vector, and an aggregated similar text vector.

[0097] Obtain the test text, perform data augmentation on the text based on the large language model o1-mini, and obtain the enhanced text. The enhanced text includes the parsed text and several similar texts. Perform text embedding processing on the text, the parsed text, and several similar texts to obtain the question text vector, the parsed text vector, and the aggregated similar text vector, including:

[0098] According to the title text, a parsed text and several similar texts are generated based on the large language model o1-mini.

[0099] Text embedding processing includes:

[0100] According to the title text, the word embedding matrix is ​​obtained.

[0101] Get the trainable vectors.

[0102] The expanded word embedding matrix is ​​obtained based on the word embedding matrix and the trainable vector.

[0103] Input the expanded word embedding matrix into the BERT encoding model to obtain the title text vector.

[0104] The parsed text and several similar texts are subjected to text embedding processing to obtain a parsed text vector and several similar text vectors.

[0105] Aggregate several similar text vectors to obtain an aggregated similar text vector.

[0106] Specifically, in order to more effectively extract feature information from the test text, we selected the BERT-base-Chinese model to process the test text content and used the P-Tuning method to adjust the BERT model parameters.

[0107] Test Questions The corresponding title text content is ,in Represents text The words, Total words, For text The last word in . Through the word embedding of the BERT model, we get the text The corresponding word embedding matrix ,in, Representative The word embedding vector corresponding to the word, is the last word embedding vector, and and They represent the embedding vectors corresponding to the classification token CLS and sentence separator SEP in the BERT model respectively.

[0108] Then, a set of trainable continuous vectors in the P-Tuning method is used as prefixes to expand the word embedding matrix , and get the final input ,in, It is trainable continuous vectors, PM Indicates the number of trainable continuous vectors.

[0109] Will Input into the pre-trained BERT encoding model and processed by maximum pooling and truncation operations to obtain the embedded representation of the text , as shown below:

[0110] ;

[0111] Among them, BERT( ) represents the pre-trained BERT model, MaxPool( ) represents the maximum pooling operation, Indicates truncating the vector, keeping only the first element.

[0112] Follow the same process to obtain the test questions Embedding vectors corresponding to other text contents: test question analysis text The corresponding embedding vector is , similar test text The corresponding embedding vector is Finally, SN The embedding vectors corresponding to similar test question texts are aggregated to obtain the final embedding representation of similar test question texts. , as shown below:

[0113] .

[0114] S3. Obtain the graph embedding vector based on the test question knowledge point association graph, and obtain the test question embedding vector based on the graph embedding vector and the question text vector.

[0115] Specifically, to ensure that the obtained embedding vector can fully represent the graph To obtain the high-order semantic information of the node, we use the metapath2vec algorithm to obtain the initial embedding vector of the node, and then use the graph convolution based on the graph attention network to update the node embedding to obtain the final embedding representation.

[0116] First, by Perform a random walk on it to generate a node sequence ,in , For the picture The set of all nodes in . Use the skip-gram model to optimize the embedding and obtain the initial embedding vector of the node The optimization goal of the skip-gram model is to maximize the conditional probability, and the probability is calculated as follows:

[0117] ;

[0118] ;

[0119] in, is the embedding representation of the node, is a node In the node sequence In the context, For nodes The context node, is a node Embedded representation The transpose of For any node The embedding representation of Representation node The embedding vector of Represents a compute node and Inner product of embedding vectors.

[0120] Then, we use graph convolution based on graph attention network to aggregate the graph. The high-level semantic information of the node is used to update the embedding representation of the node. The information aggregation process is shown in the following formula:

[0121] ;

[0122] ;

[0123] Among them, W and is a trainable parameter, For nodes After the The embedding vector obtained after the aggregation of the layer graph convolution layer, is a node Neighbor nodes of is the sigmoid activation function.

[0124] Finally, after L After the convolution layer, the final embedding vectors of all question nodes are obtained, where the question The corresponding graph embedding vector is .

[0125] The test questions Text embedding vector With graph embedding vector Splicing to get the final test question embedding vector , as shown below:

[0126] ;

[0127] Among them, || represents the vector splicing operation, The dimension is 2× d e .

[0128] S4. Use the question text vector, parsed text vector, aggregated similar text vector, graph embedding vector and test question embedding vector to train the embedding model to obtain an optimized embedding training model.

[0129] Specifically, in order to enable the model to mine the direct knowledge point associations, indirect knowledge point associations, high-order knowledge point associations, and potential difficulty information of the test questions, and learn effective test question embedding representations, we designed the following four tasks to train the parameters of the test question embedding model.

[0130] The embedding model is trained using the question text vector, parsed text vector, aggregated similar text vector, graph embedding vector, and question embedding vector. The optimized embedding training model includes:

[0131] According to the correlation between the test questions and the knowledge points, the positive sample set and the negative sample set are obtained.

[0132] Specifically, based on the need to mine the association of knowledge points in the test text, the positive and negative sample pairs of test questions required for comparative learning and training are set according to the association between test questions and knowledge points. , according to the association relationship of knowledge points, the positive sample set is , and its negative sample set is ,in Indicates test questions The knowledge points examined. Positive sample set Test questions and exam questions in have common knowledge points. On the contrary, the negative sample set Test questions and exam questions in There are no common knowledge points.

[0133] The first loss is obtained according to the title text vector, the parsed text vector and the first loss function.

[0134] Specifically, the task of comparing questions and explanations. By designing comparative learning tasks, the question embedding model is enhanced in its ability to mine the "direct knowledge point associations" in the question text, and the first loss function is used for training.

[0135] A second loss is obtained according to the positive sample set, the negative sample set, the aggregated similar text vector, the title text vector, and the second loss function.

[0136] Specifically, the task of comparing similar test questions is designed to enhance the ability of the test question embedding model to mine the "potential knowledge point association relationship" in the test question text, and the second loss function is used for training.

[0137] The third loss is obtained according to the positive sample set, the negative sample set, the graph embedding vector, the aggregated similar text vector and the third loss function.

[0138] Specifically, the text-image comparison task. By designing comparative learning tasks, the test question embedding model is enhanced in its ability to mine the "high-level knowledge point associations" in the test question text, and the third loss function is used for training.

[0139] Get a multilayer perceptron.

[0140] Calculate the predicted difficulty value of the test question embedding vector based on the multi-layer perceptron.

[0141] The fourth loss is obtained according to the predicted difficulty value, the actual difficulty value of the test question and the fourth loss function.

[0142] Specifically, the test question difficulty prediction task. In order to help the test question embedding model effectively extract the "potential difficulty information" of the test question, based on the test question embedding vector , using a multilayer perceptron to predict the difficulty of the test questions ,in, is a multi-layer perceptron, It is the trainable parameters of the multilayer perceptron, including weights and biases. Based on the actual difficulty value of the test questions , use mean square error as the loss function to calculate the loss value of the test difficulty prediction task, and use the fourth loss function for training.

[0143] The total loss is obtained by summing the first loss, the second loss, the third loss, and the fourth loss.

[0144] Specifically, .

[0145] in, is the total loss, For the first loss, For the second loss, For the third loss, For the fourth loss.

[0146] The model parameters of the embedding model are adjusted according to the total loss to obtain the optimized embedding training model.

[0147] The first loss function is expressed as:

[0148] ;

[0149] in, is the title text vector, For test questions The parsed text vector of is the temperature parameter, N is the number of test questions in a batch during training, sim() is the cosine similarity calculation function, To remove the test questions Any test question The temperature parameter is used to adjust the distribution of similarity scores and control the model's ability to distinguish between positive and negative samples.

[0150] The second loss function is expressed as:

[0151] ;

[0152] in, To aggregate similar text vectors, It's a test question The positive sample set, It's a test question The negative sample set.

[0153] The third loss function is expressed as:

[0154] ;

[0155] ;

[0156] in, and They are respectively the trainable weight matrix and bias coefficient, which are used to map the question node embedding vector to the text embedding space. For test questions The corresponding graph embedding vector.

[0157] The fourth loss function is expressed as:

[0158] ;

[0159] in, is the actual difficulty value of the test question, is the prediction difficulty value.

[0160] S5. Obtain the user's historical learning records, input the test question database into the optimized embedding training model, obtain the embedding vectors of all test questions, filter all the test question embedding vectors according to the historical learning records, and obtain the user's test question embedding vector.

[0161] Specifically, the embedding training model has been trained through the training test question database to obtain an optimized embedding training model. At this time, all data in the entity database are converted into test question embedding vectors, wherein the training test question database is part of the test question database.

[0162] S6. Calculate the knowledge mastery based on historical learning records and user test question embedding vectors.

[0163] The knowledge mastery calculated based on historical learning records and user test embedding vectors includes:

[0164] Get the answer record based on the historical learning record, and get the user's answer record at the time step based on the answer record. t The degree of mastery of all knowledge points.

[0165] Perform interactive embedding representation on the answer records, and obtain the dimensions of the interactive embedding representation and the dimensions of the question embedding vector.

[0166] Specifically, it receives the test question embedding vector passed by the test question embedding pre-training module, and organizes the relevant data into the input form required by the knowledge tracking model based on the historical answer records in the learner's historical learning records.

[0167] Assume that a learner’s answer record is ,in, represents the learner at time step tAnswered the test questions , Indicates that the learner answered the question correctly , Indicates that the learner failed to answer the question correctly According to the learners' answer records , obtain the interaction embedding representation .in, The calculation formula is shown below.

[0168] ;

[0169] Among them, 0 is a dimension with A consistent all-zero vector, || indicates vector concatenation.

[0170] Construct a static matrix based on the dimension of the test question embedding vector and the number of knowledge points.

[0171] A dynamic matrix is ​​constructed according to the dimension of the interaction embedding representation and the number of knowledge points.

[0172] The static weight is calculated based on the user test question embedding vector and the static matrix.

[0173] Get the erase vector and increment vector.

[0174] The dynamic matrix is ​​updated according to the erasure vector, the addition vector and the static weight to obtain an updated dynamic matrix.

[0175] According to the updated dynamic matrix, the knowledge mastery is obtained.

[0176] Specifically, the knowledge tracing model uses Dynamic Key-Value Memory Networks for Knowledge Tracing (DKVMN) to calculate the learner's mastery of knowledge points based on the sorted answer data. DVKMN has a static matrix and a dynamic matrix , where the static matrix The size is , each row represents the vector corresponding to the knowledge point, dynamic matrix The size of , Indicates the number of knowledge points, Represents the test question embedding vector Dimensions, Representing interactive embedding representations dimension.

[0177] First calculate the static weight , the formula is as follows:

[0178] ;

[0179] in, For test questions The embedded vector obtained by pre-training, softmax() is the softmax activation function, Indicates the A knowledge point.

[0180] Next, calculate each time step The erase vector in and increase vector , respectively simulating the learner's forgetting behavior and reinforcement behavior, the formula is as follows:

[0181] ;

[0182] ;

[0183] in, For test questions The corresponding interaction embedding vectors, E and H, are both of size The transformation matrix, is the sigmoid activation function, tanh() is the tanh activation function, and are all trainable bias parameters, is transposed.

[0184] The final matrix To update, the formula is as follows:

[0185] .

[0186] In the final value matrix In , each row vector corresponds to the mastery status of a knowledge point. The average of each row vector is taken, and the average result is regarded as the learner's mastery of the corresponding knowledge point. Indicates the learner's mastery of all knowledge points, where Indicates learners' understanding of knowledge points degree of mastery.

[0187] S7. Classify users according to historical learning records to obtain classification results.

[0188] Specifically, user classification is based on their steady-state learning characteristics. Steady-state learning characteristics refer to the stable, personalized traits exhibited by learners during the learning process, reflecting their relatively fixed cognitive and behavioral patterns. Therefore, based on learning style theory and personality test theory, this paper uses learner data from the past two years to conduct an in-depth analysis and assessment of learners' steady-state learning characteristics.

[0189] First, based on the Feld-Silverman model, the present invention uses learning behavior data to model eight characteristics of learners, including dynamic, active, visual, linguistic, sensory, intuitive, sequential, and comprehensive. These characteristics are quantified through learning behavior data such as the number of times learners participate in discussions, the frequency of watching videos, the type of videos watched, and the length of learning time. Then, the Myers-Briggs Type Indicator (MBTI) is applied to collect learners' scores on the four personality dimensions of attention direction, cognitive style, decision-making style, and lifestyle through questionnaires. Then, the K-means clustering algorithm is used to divide learners into C categories, each learner is represented by a twelve-dimensional vector, where The value of each dimension represents the learner's Finally, the silhouette coefficient (SC) and Calinski-Harabasz (CH) index are calculated to evaluate the clustering effect, and the cluster with the best clustering effect is selected. C value.

[0190] S8. Obtain the user's learning monitoring data, calculate the user's concentration score and emotional state score based on the learning monitoring data, calculate the engagement score based on historical learning records, and calculate the preference score based on the concentration score, emotional state score, and engagement score.

[0191] Specifically, after each learner finishes watching a video, the module evaluates the learner's preference for the video from three perspectives: emotional state, concentration, and participation, based on the learner's viewing status in the "learner monitoring video data" and the "learner's historical learning record" data.

[0192] a) Emotional state. We obtain facial image data of learners watching the video from the “learner monitoring video data” and use a convolutional neural network (CNN) to process the facial images, identify the emotions of each image, and obtain the emotional vector of the learner during the viewing process. Assume that the emotional vector of a learner is ,in, Indicates the number of facial image data, The emotion expressed by the facial image, Express negativity, Express neutrality, Indicates positive. The sum of all element values ​​in the vector is averaged to obtain the learner's emotional state score. .

[0193] b) Concentration. Determine the learner's concentration based on the learner's head posture and facial expression when watching the "learner monitoring video data". Use the 6DRepNet algorithm to calculate the pitch angle and yaw angle of the head posture, determine the learner's head state (normal state, slight deflection, significant deflection), and calculate the concentration score based on the head state by weighted summation of the frequency of occurrence of the three head states. Calculate the concentration score based on facial expression based on the emotion vector obtained in a). The average of the concentration score based on head state and the concentration score based on facial expression is taken as the learner's final concentration score .

[0194] c) Engagement. Receive the learner's personalized interactive behavior when watching the video from the "learner's historical learning record", such as viewing time, number of repeated viewings, whether to double speed, whether to skip, etc. Use the above features to calculate the learner's engagement score , as shown below:

[0195] ;

[0196] ;

[0197] in, Indicates the actual viewing time of learners. Indicates the total length of the video. Score_repeat, Score_ speed, Score_skip The scores for repeated viewing, double-speed viewing, and skipping are respectively determined based on the learner's actual behavior and the engagement calculation rules. The engagement calculation rules include rules for repeated viewing, double-speed viewing, and skipping. Tables 1 through 3 show the engagement calculation rules, with Table 1 covering repeated viewing, Table 2 covering double-speed viewing, and Table 3 covering skipping.

[0198] Table 1 Repeat viewing behavior rules

[0199]

[0200] Table 2 Rules for viewing at double speed

[0201]

[0202] Table 3 Skip-viewing behavior rules

[0203]

[0204] The above three scores 、 、 The learner's preference for the video is obtained by weighted summation. The learner's preference for all videos he has watched constitutes the learner's dynamic learning characteristics and is passed to the "MOOC video recommendation module".

[0205] S9. Obtain the video set and the knowledge points corresponding to each video in the video set to form a knowledge point set, obtain the mapping relationship between the video set and the knowledge point set, construct triples based on the video set, the knowledge point set and the mapping relationship, generate all knowledge states based on the triples and the disjunction model, and construct a complete knowledge structure.

[0206] Specifically, knowledge space reasoning can help analyze cognitive dependencies between knowledge points, facilitating learning in a more scientific and systematic manner. This module applies knowledge space theory to construct learners' knowledge structures within a disjunctive model framework by combining video features (e.g., how different video clips correspond to different knowledge points) with learner interaction data.

[0207] Assume that the video collection is , knowledge point set , the triple (VS, K, r) is called the video-knowledge point temporal mapping, where It's a video The total duration (unit: minutes), is a temporal mapping from VS to K, where T is a set of time intervals. This mapping satisfies the following conditions: , ,in and Knowledge points In the video The timestamps of the start and end of the occurrence of . Under the disjunctive model, for any ,Depend on P The generated knowledge state is Traverse all subsets of the knowledge point set K and obtain all knowledge states generated by the knowledge point mapping (VS, K, r) under the disjunctive model to form a complete knowledge structure .

[0208] S10. Obtain the user's learning behavior and viewing history.

[0209] S11. Obtain the proficiency of different knowledge points based on viewing history and learning behavior.

[0210] S12. According to the learning behavior, the video set watched by the user, the knowledge point set corresponding to the video set watched by the user, and the corresponding mapping relationship, several personalized knowledge states are obtained to form a personalized knowledge structure.

[0211] Based on the learning behavior, the video set watched by the user, the knowledge point set corresponding to the video set watched by the user, and the corresponding mapping relationship, several personalized knowledge states are obtained to form the personalized knowledge structure including:

[0212] Based on learning behavior, the user's proficiency in different knowledge points is calculated.

[0213] According to the proficiency, the set of videos watched by the user, the set of knowledge points corresponding to the set of videos watched by the user, and the corresponding mapping relationship, a number of personalized knowledge states are obtained through the disjunctive model to form a personalized knowledge structure.

[0214] Specifically, the learner's viewing behavior (such as double-speed viewing, repeated viewing) will affect the learner's proficiency in the relevant knowledge points required to watch the video, and thus affect the subsequent determination of the learner's knowledge status. We define the proficiency of the knowledge points corresponding to the video based on the learner's behavior during the video viewing process. ,in Indicates that learners watch the video and understand the knowledge points The proficiency is , whose value range is [0, 1]. If the learner watches the video vs Contains knowledge points The proficiency The calculation is as follows:

[0215] ;

[0216] in, Indicates baseline proficiency, Indicates the current speed. Indicates the learner's baseline speed (i.e. the speed at which learners usually watch videos). , Indicates the current number of repeated views. represents the number of baseline repeat views for the learner, 、 Are weight coefficients, satisfying .

[0217] Different learning behaviors (such as actual viewing segments, duration, etc.) will also lead to changes in the learner's video-knowledge point temporal mapping relationship, thereby forming different knowledge structures. u The knowledge structure of CS uAt the same time, build CS u The following constraints need to be met during the process:

[0218] ;

[0219] in, Representing learners Watch the video vs Knowledge points The actual duration of Represents a video vs Knowledge points The total duration of .

[0220] For example, , complete video-knowledge point temporal mapping

[0221]

[0222] Complete knowledge structure for .

[0223] For learners , whose viewing behavior is watching part of the video , video part , full video And a 1-minute video , the specific video-knowledge point temporal mapping relationship is

[0224] , Its corresponding knowledge structure For learners , whose viewing behavior is to watch the entire video ,video Parts and videos The specific video-knowledge point temporal mapping relationship is

[0225] , , , its corresponding knowledge structure Obviously, the knowledge structure 、 and It's different.

[0226] For learners Any knowledge state , knowledge status KS Corresponding knowledge proficiency Defined as:

[0227] ;

[0228] ;

[0229] in, Representing knowledge status KS Knowledge points The proficiency is . Knowledge state KS All knowledge points included Knowledge points corresponding to the video The mean proficiency, N Knowledge state KS Contains knowledge points The number of videos.

[0230] For example, based on the learner's viewing speed and repeated viewing behavior, the proficiency of the knowledge points corresponding to each video is calculated. The knowledge proficiency of the videos watched is , , , then for the knowledge structure One of the knowledge states in , and its corresponding knowledge proficiency . And learners The knowledge proficiency of the videos watched is , , , for knowledge structure One of the knowledge states in , and its corresponding knowledge proficiency Obviously, each learner has different viewing behaviors, and the knowledge proficiency corresponding to the knowledge structure is also different.

[0231] S13. Construct a knowledge structure matrix based on individual knowledge structure and proficiency, calculate the similarity between knowledge mastery and each column vector in the knowledge structure matrix, obtain the individual knowledge state with the highest similarity as the target knowledge state, construct the knowledge state boundary based on the target knowledge state and the complete knowledge structure, and filter the video set based on the knowledge state boundary to obtain the outer edge video set and the inner edge video set.

[0232] According to the individual knowledge structure and proficiency, a knowledge structure matrix is ​​constructed. The similarity between the knowledge mastery and each column vector in the knowledge structure matrix is ​​calculated. The individual knowledge state with the highest similarity is obtained as the target knowledge state, including:

[0233] A knowledge structure matrix is ​​constructed based on individual knowledge structure and proficiency, where rows represent knowledge points, columns represent knowledge states, and elements in the knowledge structure are 0 or proficiency.

[0234] Specifically, according to the learner (user) knowledge structure And its knowledge point set The mapping relationship and the proficiency of the knowledge points corresponding to each video , construct the knowledge structure matrix D, where the rows represent knowledge points and the columns represent knowledge states. The elements in the matrix are usually 0 or the proficiency obtained in step 1 .For example , Indicates the The knowledge state does not include knowledge points, on the contrary, Indicates the The knowledge state contains Knowledge points and proficiency is By calculating the learner's knowledge mastery level obtained by the "knowledge mastery level assessment module" KM The cosine similarity with each column vector in the knowledge structure matrix D is used to obtain the knowledge state that is most similar to the current learner's knowledge mastery level, and it is used as the learner's initial target knowledge state.

[0235] The knowledge mastery degree and each column vector in the knowledge structure matrix are input into the similarity calculation formula to obtain several similarities, and the individual knowledge state with the highest similarity is selected as the target knowledge state.

[0236] The similarity calculation formula is:

[0237] ;

[0238] in, KM is the knowledge mastery, D is the knowledge structure matrix, is the learner's mastery level vector KM No. elements, is the number of knowledge points, For the The first of the personality knowledge states The proficiency of each knowledge point.

[0239] For learners Knowledge structure , any knowledge state , the inner edge of KS , KS outer edge The inner edge of the knowledge state represents the videos that learners have recently watched. The content of these videos may be easily forgotten, so learners need to review and consolidate this content. The outer edge of the knowledge state represents the videos that need to be studied and expanded next. These videos can provide guidance for teachers' teaching plans or learners' self-study paths.

[0240] S14. Obtain a first recommended video set based on the classification results and the outer edge video set, obtain a second recommended video set based on the preference score and the inner edge video set, and use the first recommended video set and the second recommended video set as recommendation results.

[0241] Specifically, step 1: selection of videos to be expanded based on steady-state learning features

[0242] Based on the classification results from the steady-state learning feature evaluation, we calculate the preferences of all learners in different categories for video teaching styles. For example, learners in a certain category prefer videos with strong interactive teaching. We then filter the outer edge candidate videos based on learner preferences to obtain a set of recommended videos for the next category.

[0243] Step 2: Selection of videos to be consolidated based on dynamic learning features

[0244] Receive dynamic learning features and obtain the learner's preference for previously viewed videos. Based on the resulting inner edge candidate video set, for videos in the inner edge candidate video set that the learner has already watched, if the learner's preference score for the video is greater than or equal to the learner's average preference score for previously viewed videos, then the video is retained in the set of videos to be consolidated. Otherwise, the video is removed from the candidate video set. Simultaneously, videos in the inner edge candidate video set that the learner has not watched are retained in the set of videos to be consolidated.

[0245] Step 3: Comprehensive recommendation

[0246] The video collection for the "to-expand" category obtained in step 1 and the video collection for the "to-consolidate" category obtained in step 2 are combined as the final recommendation result and recommended to the learner. The resulting recommended video collection can help learners consolidate weak knowledge points and expand their learning of new knowledge.

[0247] A MOOC recommendation device that combines a large language model and knowledge space reasoning, comprising:

[0248] A module is established to establish a test question knowledge point association diagram based on a training test question database.

[0249] The enhancement module is used to obtain the question text of the test question, perform data enhancement on the question text based on the large language model, and obtain the enhanced text. The enhanced text includes the parsed text and several similar texts. The question text, the parsed text and several similar texts are subjected to text embedding processing to obtain the question text vector, the parsed text vector and the aggregated similar text vector.

[0250] The fusion module is used to obtain the graph embedding vector based on the test question knowledge point association graph, and obtain the test question embedding vector based on the graph embedding vector and the question text vector.

[0251] The training module is used to train the embedding model using the question text vector, the parsed text vector, the aggregated similar text vector, the graph embedding vector, and the test question embedding vector to obtain an optimized embedding training model.

[0252] The first acquisition module is used to obtain the user's historical learning records, input the test question database into the optimized embedding training model, obtain the embedding vectors of all test questions, filter all test question embedding vectors according to the historical learning records, and obtain the user's test question embedding vector.

[0253] The first calculation module is used to calculate the knowledge mastery based on historical learning records and user test question embedding vectors.

[0254] The classification module is used to classify users according to historical learning records and obtain classification results.

[0255] The second acquisition module is used to obtain the user's learning monitoring data, calculate the user's concentration score and emotional state score based on the learning monitoring data, calculate the participation score based on historical learning records, and calculate the preference score based on the concentration score, emotional state score and participation score.

[0256] The third acquisition module is used to obtain the video collection and the knowledge points corresponding to each video in the video collection to form a knowledge point set, obtain the mapping relationship between the video collection and the knowledge point set, construct triples based on the video collection, knowledge point set and the mapping relationship, generate all knowledge states based on the triples and disjunction model, and construct a complete knowledge structure.

[0257] The fourth acquisition module is used to obtain the user's learning behavior and viewing history.

[0258] The second calculation module is used to obtain the proficiency of different knowledge points based on viewing history and learning behavior.

[0259] The third calculation module is used to obtain a number of personalized knowledge states based on learning behavior, a set of videos watched by the user, a set of knowledge points corresponding to the set of videos watched by the user, and corresponding mapping relationships to form a personalized knowledge structure.

[0260] The screening module is used to construct a knowledge structure matrix based on individual knowledge structure and proficiency, calculate the similarity between knowledge mastery and each column vector in the knowledge structure matrix, obtain the individual knowledge state with the highest similarity as the target knowledge state, construct the knowledge state boundary based on the target knowledge state and the complete knowledge structure, and screen the video collection based on the knowledge state boundary to obtain the outer edge video collection and the inner edge video collection.

[0261] The recommendation module is used to obtain a first recommended video set based on the classification results and the outer edge video set, obtain a second recommended video set based on the preference score and the inner edge video set, and use the first recommended video set and the second recommended video set as recommendation results.

[0262] An embodiment of the present application also discloses a terminal device, including a memory and a processor. The memory stores a computer program that can be run on the processor. When the processor loads and executes the computer program, it adopts a MOOC recommendation method that combines a large language model and knowledge space reasoning.

[0263] Among them, the terminal device can be a computer device such as a desktop computer, a laptop computer or a cloud server, and the terminal device includes but is not limited to a processor and a memory. For example, the terminal device can also include input and output devices, network access devices and buses, etc.

[0264] Among them, the processor can adopt a central processing unit (CPU). Of course, according to actual usage, other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. can also be adopted. The general-purpose processor can adopt a microprocessor or any conventional processor, etc., and this application does not impose any restrictions on this.

[0265] Among them, the memory can be an internal storage unit of the terminal device, such as the hard disk or memory of the terminal device, or it can be an external storage device of the terminal device, such as a plug-in hard disk, smart memory card (SMC), secure digital card (SD) or flash memory card (FC) equipped on the terminal device, etc., and the memory can also be a combination of the internal storage unit and the external storage device of the terminal device. The memory is used to store computer programs and other programs and data required by the terminal device. The memory can also be used to temporarily store data that has been output or is to be output. This application does not impose any restrictions on this.

[0266] Among them, through this terminal device, a MOOC recommendation method combining a large language model and knowledge space reasoning in the above embodiment is stored in the memory of the terminal device, and is loaded and executed on the processor of the terminal device for easy use.

[0267] An embodiment of the present application also discloses a computer-readable storage medium, and the computer-readable storage medium stores a computer program. When the computer program is executed by a processor, a MOOC recommendation method combining a large language model and knowledge space reasoning in the above embodiment is adopted.

[0268] Among them, the computer program can be stored in a computer-readable medium, the computer program includes computer program code, the computer program code can be in the form of source code, object code, executable file or certain middleware, etc. The computer-readable medium includes any entity or device that can carry computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that computer-readable medium includes but is not limited to the above-mentioned components.

[0269] Among them, through this computer-readable storage medium, a MOOC recommendation method combining a large language model and knowledge space reasoning in the above embodiment is stored in a computer-readable storage medium, and is loaded and executed on a processor to facilitate the storage and application of the above method.

[0270] Those skilled in the art should understand that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of protection of the present application is limited to these examples. In line with the present application, the technical features in the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and there are many other variations of different aspects of one or more embodiments of the present application as above, which are not provided in detail for the sake of simplicity.

Claims

1. A MOOC recommendation method combining a large language model and knowledge space reasoning, characterized by: include: Establish a test question knowledge point association diagram based on the training test question database; Obtain the test text, perform data augmentation on the test text based on the large language model o1-mini to obtain an enhanced text, wherein the enhanced text includes a parsed text and several similar texts, and perform text embedding processing on the test text, the parsed text, and the several similar texts to obtain a test text vector, a parsed text vector, and an aggregated similar text vector; Obtain a graph embedding vector based on the test question knowledge point association graph, and obtain a test question embedding vector based on the graph embedding vector and the question text vector; The embedding model is trained using the question text vector, the parsed text vector, the aggregated similar text vector, the graph embedding vector, and the test question embedding vector to obtain an optimized embedding training model; Obtain the user's historical learning records, input the test question database into the optimized embedding training model to obtain all test question embedding vectors, and filter all test question embedding vectors based on the historical learning records to obtain the user's test question embedding vector; Calculate the knowledge mastery based on historical learning records and the user's test question embedding vector; Classify users according to the historical learning records to obtain classification results; Obtaining the user's learning monitoring data, calculating the user's concentration score and emotional state score based on the learning monitoring data, calculating the engagement score based on historical learning records, and calculating the preference score based on the concentration score, emotional state score, and engagement score; Obtain a video set and the knowledge points corresponding to each video in the video set to form a knowledge point set, obtain a mapping relationship between the video set and the knowledge point set, construct a triple based on the video set, the knowledge point set, and the mapping relationship, generate all knowledge states based on the triple and the disjunctive model, and construct a complete knowledge structure; Obtain users' learning behavior and viewing history; Obtaining proficiency of different knowledge points based on the viewing history and learning behavior; According to the learning behavior, the video set watched by the user, the knowledge point set corresponding to the video set watched by the user, and the corresponding mapping relationship, a plurality of personalized knowledge states are obtained to form a personalized knowledge structure; A knowledge structure matrix is ​​constructed based on the individual knowledge structure and the proficiency, and the similarity between the knowledge mastery and each column vector in the knowledge structure matrix is ​​calculated to obtain the individual knowledge state with the highest similarity as the target knowledge state. A knowledge state boundary is constructed based on the target knowledge state and the complete knowledge structure. A video collection is filtered based on the knowledge state boundary to obtain an outer edge video collection and an inner edge video collection, where the inner edge video collection is the video that the user has recently watched, and the outer edge video collection is the video that the user needs to learn and expand next. A first recommended video set is obtained based on the classification results and the outer edge video set, and a second recommended video set is obtained based on the preference score and the inner edge video set. The first recommended video set and the second recommended video set are used as recommendation results.

2. The method for recommending MOOCs that combines a large language model with knowledge space reasoning as claimed in claim 1, wherein: The method of obtaining the test question text, performing data enhancement on the test question text based on the large language model o1-mini to obtain an enhanced text, wherein the enhanced text includes a parsed text and several similar texts, and performing text embedding processing on the test question text, the parsed text and the several similar texts to obtain a test question text vector, a parsed text vector and an aggregated similar text vector includes: Based on the question text, generate a parsed text and several similar texts based on the large language model o1-mini; The text embedding process includes: According to the title text, a word embedding matrix is ​​obtained; Get the trainable vector; Obtaining an expanded word embedding matrix according to the word embedding matrix and the trainable vector; Input the expanded word embedding matrix into the BERT encoding model to obtain the title text vector; Performing text embedding processing on the parsed text and a plurality of similar texts to obtain a parsed text vector and a plurality of similar text vectors; Aggregate several similar text vectors to obtain an aggregated similar text vector.

3. The method for recommending MOOCs that combines a large language model with knowledge space reasoning as claimed in claim 1, wherein: The training of the embedding model using the question text vector, the parsed text vector, the aggregated similar text vector, the graph embedding vector, and the test question embedding vector to obtain an optimized embedding training model includes: According to the correlation between the test questions and the knowledge points, a positive sample set and a negative sample set are obtained; Obtaining a first loss according to the question text vector, the parsed text vector, and a first loss function; Obtaining a second loss according to the positive sample set, the negative sample set, the aggregated similar text vector, the title text vector, and a second loss function; Obtaining a third loss according to the positive sample set, the negative sample set, the graph embedding vector, the aggregated similar text vector, and a third loss function; Get a multilayer perceptron; Predicting a difficulty value of the test question embedding vector based on the multi-layer perceptron; Obtaining a fourth loss based on the predicted difficulty value, the actual difficulty value of the test question, and a fourth loss function; summing the first loss, the second loss, the third loss, and the fourth loss to obtain a total loss; The model parameters of the embedding model are adjusted according to the total loss to obtain an optimized embedding training model.

4. The method for recommending MOOCs based on a large language model and knowledge space reasoning as claimed in claim 1, wherein: The calculation of the knowledge mastery based on the historical learning records and the user's test question embedding vector includes: According to the historical learning record, the answer record is obtained, and the user's time step is obtained according to the answer record. t The degree of mastery of all knowledge points; Performing interactive embedding representation on the answer record, and obtaining the dimension of the interactive embedding representation and the dimension of the test question embedding vector; Constructing a static matrix according to the dimension of the test question embedding vector and the number of knowledge points; Construct a dynamic matrix based on the dimension of the interaction embedding representation and the number of knowledge points; A static weight is calculated based on the user test question embedding vector and the static matrix; Get the erase vector and increment vector; The dynamic matrix is ​​updated according to the erasure vector, the addition vector and the static weight to obtain an updated dynamic matrix; The knowledge mastery degree is obtained according to the updated dynamic matrix.

5. The method for recommending MOOCs that combines a large language model with knowledge space reasoning as claimed in claim 1, wherein: According to the learning behavior, the video set watched by the user, the knowledge point set corresponding to the video set watched by the user, and the corresponding mapping relationship, a plurality of personalized knowledge states are obtained to form a personalized knowledge structure including: Calculate the user's proficiency in different knowledge points based on the learning behavior; According to the proficiency, the set of videos watched by the user, the set of knowledge points corresponding to the set of videos watched by the user, and the corresponding mapping relationship, a plurality of personalized knowledge states are obtained through a disjunctive model to form a personalized knowledge structure.

6. The method for recommending MOOCs combining a large language model and knowledge space reasoning as claimed in claim 1, wherein: The step of constructing a knowledge structure matrix based on the individual knowledge structure and the proficiency, calculating the similarity between the knowledge mastery and each column vector in the knowledge structure matrix, and obtaining the individual knowledge state with the highest similarity as the target knowledge state includes: Constructing a knowledge structure matrix based on the individual knowledge structure and the proficiency, wherein rows represent knowledge points, columns represent knowledge states, and elements in the knowledge structure are 0 or proficiency; Input the knowledge mastery degree and each column vector in the knowledge structure matrix into a similarity calculation formula to obtain several similarities, and select the individual knowledge state with the highest similarity as the target knowledge state; The similarity calculation formula is: ； in, KM is the knowledge mastery, D is the knowledge structure matrix, is the learner's mastery level vector KM No. elements, m is the number of knowledge points, For the The first of the personality knowledge states The proficiency of each knowledge point.

7. A MOOC recommendation device combining a large language model and knowledge space reasoning, characterized by: include: Establishing a module for establishing a test question knowledge point association diagram based on a training test question database; An enhancement module is used to obtain the test question text, perform data enhancement on the question text based on the large language model o1-mini to obtain enhanced text, wherein the enhanced text includes the parsed text and several similar texts, and perform text embedding processing on the question text, the parsed text, and the several similar texts to obtain a question text vector, a parsed text vector, and an aggregated similar text vector; A fusion module is used to obtain a graph embedding vector based on the test question knowledge point association graph, and obtain a test question embedding vector based on the graph embedding vector and the question text vector; A training module, configured to train an embedding model using the question text vector, the parsed text vector, the aggregated similar text vector, the graph embedding vector, and the test question embedding vector to obtain an optimized embedding training model; The first acquisition module is used to obtain the user's historical learning records, input the test question database into the optimized embedding training model to obtain all test question embedding vectors, and filter all test question embedding vectors according to the historical learning records to obtain the user's test question embedding vector; A first calculation module is used to calculate the knowledge mastery based on the historical learning records and the user's test question embedding vector; A classification module, configured to classify users according to the historical learning records and obtain classification results; A second acquisition module is configured to acquire the user's learning monitoring data, calculate the user's concentration score and emotional state score based on the learning monitoring data, calculate the user's engagement score based on historical learning records, and calculate the user's preference score based on the concentration score, emotional state score, and engagement score; A third acquisition module is used to obtain a video set and the knowledge points corresponding to each video in the video set to form a knowledge point set, obtain a mapping relationship between the video set and the knowledge point set, construct a triple based on the video set, the knowledge point set, and the mapping relationship, generate all knowledge states based on the triple and the disjunctive model, and construct a complete knowledge structure; The fourth acquisition module is used to obtain the user's learning behavior and viewing history; A second calculation module is used to obtain the proficiency of different knowledge points based on the viewing history and learning behavior; A third calculation module is configured to obtain a plurality of personalized knowledge states based on the learning behavior, the video set watched by the user, the knowledge point set corresponding to the video set watched by the user, and the corresponding mapping relationship to form a personalized knowledge structure; A screening module is used to construct a knowledge structure matrix based on the individual knowledge structure and the proficiency, calculate the similarity between the knowledge mastery and each column vector in the knowledge structure matrix, obtain the individual knowledge state with the highest similarity as the target knowledge state, construct a knowledge state boundary based on the target knowledge state and the complete knowledge structure, and screen the video collection based on the knowledge state boundary to obtain an outer edge video collection and an inner edge video collection, wherein the inner edge video collection is the video that the user has recently watched, and the outer edge video collection is the video that the user needs to learn and expand next; The recommendation module is used to obtain a first recommended video set based on the classification results and the outer edge video set, obtain a second recommended video set based on the preference score and the inner edge video set, and use the first recommended video set and the second recommended video set as recommendation results.

8. A terminal device comprising a memory and a processor, characterized in that: The memory stores a computer program that can be run on the processor. When the processor loads and executes the computer program, the method according to any one of claims 1 to 6 is adopted.

9. A computer-readable storage medium storing a computer program, wherein: When the computer program is loaded and executed by a processor, the method according to any one of claims 1 to 6 is adopted.

Images