A method, apparatus and equipment for predicting knowledge point mastery

Abstract

This application provides a method, apparatus, and device for predicting knowledge point mastery. The method includes: acquiring a knowledge point description sequence, a question sequence for a target object, and an auxiliary information sequence; determining the question embedding feature vector corresponding to the question sequence, the auxiliary embedding feature vector corresponding to the auxiliary information sequence, and the knowledge point embedding feature vector corresponding to the knowledge point description sequence; performing a multi-head attention operation based on the question embedding feature vector and the auxiliary embedding feature vector to obtain an ability feature vector; determining a knowledge space feature vector based on the knowledge point embedding feature vector; performing a multi-head attention operation based on the ability feature vector, the knowledge point embedding feature vector, and the knowledge space feature vector to obtain a knowledge point target feature vector; and predicting the target object's mastery of the target knowledge point for each candidate knowledge point name within the knowledge point graph based on the knowledge point target feature vector and the knowledge point graph. Through the technical solution of this application, the learner's knowledge point mastery can be accurately assessed.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus and device for predicting knowledge mastery. Background Technology

[0002] Knowledge mastery refers to the quantitative assessment indicator of learners' depth of understanding, application ability and transfer level of specific knowledge units (such as concepts, principles, skills, etc.). Its core lies in the observable performance of the degree of knowledge internalization, which is the level of understanding, memorization and application of specific knowledge content in a certain knowledge domain.

[0003] In the field of intelligent education, accurately assessing learners' (such as students) mastery of knowledge points is crucial for achieving personalized learning and precise teaching. However, current technologies cannot accurately assess this mastery. For example, periodic exams can be used, but this method is lagging, cannot reflect mastery in real time, and is time-consuming and labor-intensive, lacking precise evaluation of knowledge point mastery. Summary of the Invention

[0004] This application provides a method for predicting knowledge point mastery, the method comprising:

[0005] The process involves obtaining a knowledge point description sequence, a question sequence for the target object, and an auxiliary information sequence. The knowledge point description sequence includes semantic description information corresponding to each candidate knowledge point name. The auxiliary information sequence includes the knowledge point name, question difficulty, and question response corresponding to each question in the question sequence.

[0006] Determine the question embedding feature vector corresponding to the question sequence, the auxiliary embedding feature vector corresponding to the auxiliary information sequence, and the knowledge point embedding feature vector corresponding to the knowledge point description sequence;

[0007] Multi-head attention operation is performed based on the question embedding feature vector and the auxiliary embedding feature vector to obtain the target object’s ability feature vector for the knowledge point name corresponding to the question sequence;

[0008] The knowledge space feature vector is determined based on the knowledge point embedding feature vector. The knowledge point embedding feature vector includes an initial feature vector for each candidate knowledge point name. The knowledge space feature vector includes multiple fused feature vectors, and each fused feature vector is obtained by fusing multiple initial feature vectors.

[0009] Multi-head attention operation is performed based on the capability feature vector, the knowledge point embedding feature vector and the knowledge space feature vector to obtain the knowledge point target feature vector;

[0010] Based on the target feature vector of the knowledge point and the acquired knowledge point graph, predict the target object's mastery of the target knowledge point for each candidate knowledge point name in the knowledge point graph.

[0011] This application provides a knowledge point mastery prediction device, the device comprising:

[0012] The acquisition module is used to acquire a knowledge point description sequence, a question sequence of the target object, and an auxiliary information sequence; the knowledge point description sequence includes semantic description information corresponding to each candidate knowledge point name; the auxiliary information sequence includes the knowledge point name, question difficulty, and question response corresponding to each question in the question sequence.

[0013] The determining module is used to determine the question embedding feature vector corresponding to the question sequence, the auxiliary embedding feature vector corresponding to the auxiliary information sequence, and the knowledge point embedding feature vector corresponding to the knowledge point description sequence; and to determine the knowledge space feature vector based on the knowledge point embedding feature vector, wherein the knowledge point embedding feature vector includes an initial feature vector for each candidate knowledge point name, and the knowledge space feature vector includes multiple fused feature vectors, and each fused feature vector is obtained by fusing multiple initial feature vectors.

[0014] The processing module is used to perform multi-head attention operations based on the question embedding feature vector and the auxiliary embedding feature vector to obtain the target object's ability feature vector for the knowledge point name corresponding to the question sequence; and to perform multi-head attention operations based on the ability feature vector, the knowledge point embedding feature vector and the knowledge space feature vector to obtain the knowledge point target feature vector.

[0015] The prediction module is used to predict the target object's mastery of the target knowledge point for each candidate knowledge point name in the knowledge point graph, based on the target feature vector of the knowledge point and the acquired knowledge point graph.

[0016] This application provides an electronic device, including: a processor and a machine-readable storage medium, wherein the machine-readable storage medium stores machine-executable instructions that can be executed by the processor; the processor is used to execute the machine-executable instructions to implement the knowledge point mastery prediction method of the above example of this application.

[0017] This application provides a computer program product, which includes a computer program that, when executed by a processor, implements the knowledge point mastery prediction method of the above example of this application.

[0018] This application provides a machine-readable storage medium storing machine-executable instructions that can be executed by a processor; wherein the processor is used to execute the machine-executable instructions to implement the knowledge point mastery prediction method of the above example of this application.

[0019] As can be seen from the above technical solutions, in this embodiment, the ability feature vector can be determined based on the question embedding feature vector and the auxiliary embedding feature vector; the knowledge space feature vector can be determined based on the knowledge point embedding feature vector; the knowledge point target feature vector can be determined based on the ability feature vector, the knowledge point embedding feature vector, and the knowledge space feature vector; and the target knowledge point mastery degree of the target object (i.e., the learner) for each candidate knowledge point name can be predicted based on the knowledge point target feature vector and the knowledge point graph. In this way, the knowledge point mastery degree can be predicted by combining the knowledge point graph, accurately assessing the learner's (e.g., student's) knowledge point mastery degree, reflecting the knowledge point mastery degree in real time, and achieving precise assessment of knowledge point mastery degree. The assessment process is simple and fast. By organizing knowledge points into a graph structure, the hierarchical and dependency relationships between knowledge points are fully utilized, making the assessment results more comprehensive and accurate, achieving precise assessment of the mastery level of each specific knowledge point, and updating the assessment results of knowledge point mastery degree in real time. The assessment results can be interpreted from the perspective of knowledge point association, reflecting both the student's overall learning status and pointing out specific knowledge weaknesses. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating a knowledge point mastery prediction method in one embodiment of this application;

[0021] Figure 2 This is a flowchart illustrating a knowledge point mastery prediction method in one embodiment of this application;

[0022] Figure 3A This is a schematic diagram of the structure of a knowledge point mastery prediction model in one embodiment of this application;

[0023] Figure 3B This is a schematic diagram of the structure of a knowledge point mastery prediction model in one embodiment of this application;

[0024] Figure 3C This is a schematic diagram of the structure of a knowledge point mastery prediction model in one embodiment of this application;

[0025] Figure 4 This is a schematic diagram illustrating the statistical effect of the question difficulty distribution on a sample dataset in one implementation method.

[0026] Figure 5 This is a flowchart illustrating a knowledge point mastery prediction method in one embodiment of this application;

[0027] Figure 6 This is a schematic diagram of the structure of a knowledge point mastery prediction model in one embodiment of this application;

[0028] Figure 7 This is a schematic diagram of the structure of a knowledge point mastery prediction device according to one embodiment of this application;

[0029] Figure 8 This is a hardware structure diagram of an electronic device according to one embodiment of this application. Detailed Implementation

[0030] This application proposes a method for predicting knowledge point mastery, which can be applied to electronic devices. (See [link]). Figure 1 The diagram shown is a flowchart of the method, which may include:

[0031] Step 101: Obtain the knowledge point description sequence, the target object question sequence, and the auxiliary information sequence; the knowledge point description sequence includes the semantic description information corresponding to each candidate knowledge point name; the auxiliary information sequence includes the knowledge point name, question difficulty, and question response corresponding to each question in the question sequence.

[0032] Step 102: Determine the question embedding feature vector corresponding to the question sequence, the auxiliary embedding feature vector corresponding to the auxiliary information sequence, and the knowledge point embedding feature vector corresponding to the knowledge point description sequence.

[0033] Step 103: Perform multi-head attention operation based on the question embedding feature vector and the auxiliary embedding feature vector to obtain the target object's ability feature vector for the knowledge point name corresponding to the question sequence.

[0034] Step 104: Determine the knowledge space feature vector based on the knowledge point embedding feature vector. The knowledge point embedding feature vector includes the initial feature vector of each candidate knowledge point name. The knowledge space feature vector includes multiple fused feature vectors, and each fused feature vector is obtained by fusing multiple initial feature vectors.

[0035] Step 105: Perform multi-head attention operation based on the capability feature vector, the knowledge point embedding feature vector, and the knowledge space feature vector to obtain the knowledge point target feature vector.

[0036] Step 106: Based on the target feature vector of the knowledge point and the acquired knowledge point graph, predict the target object's mastery of the target knowledge point for each candidate knowledge point name in the knowledge point graph.

[0037] For example, obtaining a knowledge point description sequence may include, but is not limited to: obtaining a knowledge point sequence, a question text sequence, and a reference matrix; wherein, the knowledge point sequence includes multiple candidate knowledge point names, the question text sequence includes multiple example question texts, and the reference matrix represents the correspondence between candidate knowledge point names and example question texts; the knowledge point sequence, question text sequence, and reference matrix are input into a large language model, and the semantic description information corresponding to each candidate knowledge point name is determined by the large language model to obtain the knowledge point description sequence; wherein, for each candidate knowledge point name, the example question text corresponding to the candidate knowledge point name is determined by the reference matrix, and semantic analysis is performed on the candidate knowledge point name based on the example question text to obtain the semantic description information corresponding to the candidate knowledge point name.

[0038] For example, performing multi-head attention operations based on question embedding feature vectors and auxiliary embedding feature vectors to obtain the target object's ability feature vector for the knowledge point names corresponding to the question sequence may include, but is not limited to: determining a first Q vector, a first K vector, and a first V vector based on the question embedding feature vector; performing multi-head attention operations based on the first Q vector, the first K vector, and the first V vector to obtain a first intermediate vector; determining a second Q vector, a second K vector, and a second V vector based on the auxiliary embedding feature vector; performing multi-head attention operations based on the second Q vector, the second K vector, and the second V vector to obtain a second intermediate vector; determining a third Q vector and a third K vector based on the first intermediate vector; determining a third V vector based on the second intermediate vector; performing multi-head attention operations based on the third Q vector, the third K vector, and the third V vector to obtain a third intermediate vector, where the third intermediate vector represents the level of mastery at the question level; determining a fourth K vector based on the first intermediate vector; determining a fourth V vector based on the third intermediate vector; determining a fourth Q vector based on the knowledge space feature vector; performing multi-head attention operations based on the fourth Q vector, the fourth K vector, and the fourth V vector to obtain a fourth intermediate vector; and determining the ability feature vector based on the fourth intermediate vector.

[0039] For example, determining the knowledge space feature vector based on the knowledge point embedding feature vector may include, but is not limited to: determining the fifth Q vector, the fifth K vector, and the fifth V vector based on the knowledge point embedding feature vector; performing multi-head attention operation based on the fifth Q vector, the fifth K vector, and the fifth V vector to obtain the fifth intermediate vector; and inputting the fifth intermediate vector into the multilayer perceptron; and performing multilayer perception operation based on the fifth intermediate vector through the multilayer perceptron to obtain the knowledge space feature vector.

[0040] For example, predicting the target object's mastery of the target knowledge point for each candidate knowledge point name in the knowledge point graph based on the target feature vector of the knowledge point and the acquired knowledge point graph may include, but is not limited to: determining a first graph feature and a second graph feature based on the knowledge point graph, wherein the first graph feature includes the initial feature vector of each candidate knowledge point name in the knowledge point graph, and the second graph feature includes the similarity of the initial feature vectors of two candidate knowledge point names in the knowledge point graph; generating a target graph feature based on the target feature vector of the knowledge point and the first graph feature, wherein the target graph feature includes a fused feature vector of each candidate knowledge point name in the knowledge point graph, which is obtained by concatenating the target feature vector of the knowledge point and the initial feature vector; and predicting the target object's mastery of the target knowledge point for each candidate knowledge point name in the knowledge point graph based on the target graph feature, the first graph feature, and the second graph feature.

[0041] For example, predicting the target object's mastery of each candidate knowledge point name in the knowledge point graph based on the target graph features, the first graph features, and the second graph features may include, but is not limited to: inputting the target graph features into a multilayer perceptron, and performing feature mapping operations on the target graph features through the multilayer perceptron to obtain mapped graph features; wherein the feature dimension of the mapped graph features is consistent with the input feature dimension of the graph neural network; inputting the mapped graph features, the first graph features, and the second graph features into the graph neural network, and performing predictions on the graph neural network based on the mapped graph features, the first graph features, and the second graph features to obtain the mastery of the target knowledge point.

[0042] For example, the question sequence also includes the question to be predicted, and the auxiliary information sequence also includes the knowledge point name and question difficulty corresponding to the question to be predicted; the predicted question response corresponding to the question to be predicted can also be determined based on the question embedding feature vector, ability feature vector and knowledge space feature vector; wherein, the predicted question response indicates whether the target object's answer to the question to be predicted is correct or incorrect.

[0043] For example, determining the predicted question response corresponding to the question to be predicted based on the question embedding feature vector, the ability feature vector, and the knowledge space feature vector may include, but is not limited to: determining the sixth Q vector based on the question embedding feature vector, determining the sixth V vector based on the ability feature vector, and determining the sixth K vector based on the knowledge space feature vector; performing a multi-head attention operation on the sixth Q vector, the sixth K vector, and the sixth V vector to obtain the sixth intermediate vector, and inputting the sixth intermediate vector into a multilayer perceptron; performing a multilayer perception operation on the sixth intermediate vector through the multilayer perceptron to obtain the post-operation features, and determining the predicted question response corresponding to the question to be predicted based on the post-operation features.

[0044] As can be seen from the above technical solutions, in this embodiment, the ability feature vector can be determined based on the question embedding feature vector and the auxiliary embedding feature vector; the knowledge space feature vector can be determined based on the knowledge point embedding feature vector; the knowledge point target feature vector can be determined based on the ability feature vector, the knowledge point embedding feature vector, and the knowledge space feature vector; and the target knowledge point mastery degree of the target object (i.e., the learner) for each candidate knowledge point name can be predicted based on the knowledge point target feature vector and the knowledge point graph. In this way, the knowledge point mastery degree can be predicted by combining the knowledge point graph, accurately assessing the learner's (e.g., student's) knowledge point mastery degree, reflecting the knowledge point mastery degree in real time, and achieving precise assessment of knowledge point mastery degree. The assessment process is simple and fast. By organizing knowledge points into a graph structure, the hierarchical and dependency relationships between knowledge points are fully utilized, making the assessment results more comprehensive and accurate, achieving precise assessment of the mastery level of each specific knowledge point, and updating the assessment results of knowledge point mastery degree in real time. The assessment results can be interpreted from the perspective of knowledge point association, reflecting both the student's overall learning status and pointing out specific knowledge weaknesses.

[0045] The technical solutions described above in the embodiments of this application will be explained below in conjunction with specific application scenarios.

[0046] Knowledge mastery refers to a quantitative assessment indicator of a learner's depth of understanding, application ability, and transfer level of a specific knowledge unit (such as concepts, principles, and skills). In the field of intelligent education, accurately assessing learners' (such as students') knowledge mastery is crucial for achieving personalized learning and precision teaching.

[0047] In related technologies, it is impossible to accurately assess the mastery of knowledge points. For example, assessing mastery through regular exams is time-consuming and cannot reflect the mastery in real time; it also lacks precise evaluation. For instance, considering only whether a student's answers are correct ignores the inherent connections between knowledge points, leading to biased assessment results.

[0048] In related technologies, the assessment of knowledge point mastery suffers from the following problems: Ignoring knowledge structure: Treating knowledge points as independent entities fails to fully utilize the hierarchical relationships and interdependencies between them, resulting in incomplete assessment results. Coarse assessment granularity: Only an overall learning level assessment can be provided, unable to refine the mastery of specific knowledge points, hindering targeted teaching interventions. Lack of dynamism: Difficulty in capturing the dynamic changes in students' knowledge status, failing to reflect students' learning progress in a timely manner.

[0049] To address the above findings, this application proposes a method for predicting knowledge point mastery. Based on graph neural networks, it predicts knowledge point mastery, enabling accurate assessment of knowledge point mastery in education and artificial intelligence fields, such as intelligent education evaluation and personalized learning diagnosis. To address the neglect of knowledge structure, a graph neural network is introduced to model the relationships between knowledge points. By organizing knowledge points into a graph structure, the hierarchical and dependency relationships between knowledge points are fully utilized, making the assessment results more comprehensive and accurate. To address the issue of coarse assessment granularity, a fine-grained method for predicting knowledge point mastery based on student answer sequences is proposed. By combining the temporal modeling capabilities of Transformer and the graph structure modeling capabilities of graph neural networks, accurate assessment of the mastery level of each specific knowledge point is achieved. To address the lack of dynamism, a model structure capable of capturing dynamic changes in students' knowledge status is proposed. By analyzing students' answer sequences, the model can update the assessment results of students' knowledge mastery in real time. To address the issue of poor interpretability of assessment results, by integrating the knowledge point graph structure, the model can interpret the assessment results from the perspective of knowledge point relationships, reflecting both the student's overall learning status and identifying specific knowledge weaknesses.

[0050] This application proposes a method for predicting knowledge point mastery, which can be applied to electronic devices such as PCs (Personal Computers), servers, laptops, and smartphones. (See also...) Figure 2 The diagram shown is a flowchart of the method, which may include:

[0051] Step 201: Obtain the knowledge point description sequence, which includes the semantic description information corresponding to each candidate knowledge point name (each knowledge point name is denoted as the candidate knowledge point name).

[0052] For example, a knowledge point sequence can be obtained, which may include multiple candidate knowledge point names. For instance, the knowledge point sequence may include candidate knowledge point name a1, candidate knowledge point name a2, ..., candidate knowledge point name aN, etc. That is, taking N candidate knowledge point names as an example, the N candidate knowledge point names are all known candidate knowledge point names, such as 1000 candidate knowledge point names, etc.

[0053] The knowledge point description sequence can include semantic description information b1 corresponding to candidate knowledge point name a1, semantic description information b2 corresponding to candidate knowledge point name a2, ..., semantic description information bN corresponding to candidate knowledge point name aN, that is, there are a total of N semantic description information. Semantic description information b1 is used to describe the semantic information of candidate knowledge point name a1. Semantic description information b1 is semantic information related to candidate knowledge point name a1, used to explain and describe candidate knowledge point name a1, and is explanatory and descriptive information for candidate knowledge point name a1. For example, if the knowledge point representation of candidate knowledge point name a1 is poor, only providing a knowledge point name label, then candidate knowledge point name a1 can be expanded into a vector representation rich in pedagogical semantics. This vector representation can be the semantic description information b1 corresponding to candidate knowledge point name a1.

[0054] For example, a sequence of question texts can be obtained. The sequence of question texts may include multiple example question texts. Example question texts are text information about example questions. An example question text can be understood as a question. These example question texts may be a subset of example question texts selected from a knowledge base.

[0055] For example, a reference matrix (such as a Q-matrix) can be obtained, which represents the correspondence between candidate knowledge point names and example question texts. For instance, assuming there are N candidate knowledge point names and M example question texts, the reference matrix can include N*M elements. The first element in the first row represents the correspondence between the first candidate knowledge point name and the first example question text; a value of 1 indicates a correspondence exists, and a value of 0 indicates no correspondence exists. The second element in the first row represents the correspondence between the first candidate knowledge point name and the second example question text, and so on. The Mth element in the first row represents the correspondence between the first candidate knowledge point name and the Mth example question text, and the first element in the second row represents the correspondence between the second candidate knowledge point name and the first example question text, and so on. There are no restrictions on this reference matrix.

[0056] For example, a sequence of knowledge points, a sequence of question texts, and a reference matrix can be input into a large language model. The large language model can then determine the semantic description information corresponding to each candidate knowledge point name, resulting in a knowledge point description sequence. This knowledge point description sequence includes the semantic description information corresponding to each candidate knowledge point name.

[0057] For example, for each candidate knowledge point name in the knowledge point sequence, the Large Language Model (LLM) uses the reference matrix to determine the example question text corresponding to the candidate knowledge point name. That is, it uses the reference matrix to find the example question text that corresponds to the candidate knowledge point name from the question text sequence.

[0058] Then, based on the example question text (which may contain multiple example question texts), semantic analysis is performed on the candidate knowledge point name to obtain the semantic description information corresponding to the candidate knowledge point name. For example, by providing the Large Language Model (LLM) with a "name + example question -> explanation" example, the semantic generation capability of the LLM can be used to obtain the semantic description information corresponding to each candidate knowledge point name. There are no restrictions on this process; the LLM can obtain the semantic description information based on the example question text. That is, the semantic description information is semantic information related to the candidate knowledge point name, and the LLM can obtain the semantic description information corresponding to the candidate knowledge point name by analyzing the example question text corresponding to the candidate knowledge point name.

[0059] Step 202: Obtain the knowledge point graph. The knowledge point graph is a graph for all candidate knowledge point names (e.g., N candidate knowledge point names). The knowledge point graph can include the N candidate knowledge point names and the connection relationships between the candidate knowledge point names. In this embodiment, there is no restriction on this knowledge point graph.

[0060] For example, a knowledge graph is a structured way of representing knowledge, which can graphically display the relationships between knowledge points. For instance, knowledge points can be concepts, entities, events, etc., while relationships can be inheritance, association, dependency, etc. Through a knowledge graph, the connections between knowledge points can be clearly seen, leading to a better understanding and application of these knowledge points.

[0061] Knowledge graphs have wide applications in many fields. For example, in education, knowledge graphs can help students better understand the relationships between knowledge points, guiding learning and teaching. Teachers can design syllabi and teaching content based on knowledge graphs, and students can use them for self-study and knowledge consolidation. In search engine optimization (SEO), knowledge graphs can help websites provide more accurate and relevant search results. By annotating and associating knowledge points within website content, search engines can better understand the relationships between web pages, improving the quality and accuracy of search results.

[0062] To construct a knowledge graph, a manual construction method can be used. This involves manually creating the knowledge graph, with experts or practitioners in the knowledge domain identifying, classifying, and associating knowledge points to ensure the accuracy and completeness of the relationships between them. Alternatively, a machine learning-based construction method can be used. This method automatically constructs the knowledge graph using machine learning algorithms. By analyzing large corpora and text data, it can automatically identify and extract knowledge points and establish relationships between them, thus accelerating the construction process.

[0063] In one possible implementation, a dual-track strategy can be adopted for the construction of the knowledge point graph, with the two tracks complementing each other to jointly form a comprehensive and accurately related knowledge point graph.

[0064] First, an initial knowledge point graph is constructed based on the knowledge of educational experts. For example, a complete initial knowledge point graph is built based on dataset metadata or guidance from educational experts. The relationships between knowledge points (such as similarity, precedence, etc.) marked by metadata or manual annotation are recorded in the form of an adjacency matrix as a complete initial knowledge point graph. Basic information such as difficulty coefficient and grade attribute, which are statistically derived from strategies or manually annotated, is added to each knowledge point in the initial knowledge point graph to obtain a structured representation of the initial knowledge point graph.

[0065] Then, the initial knowledge point graph is enhanced based on the question data to obtain the enhanced target knowledge point graph, which is then output. The knowledge point graph in subsequent embodiments refers to this target knowledge point graph. For example, question and knowledge point annotation information is collected, knowledge point co-occurrence is statistically analyzed, and then the node2vec algorithm is used to learn the representation of the knowledge points, capturing the structural relationships between them through random walks. Based on the learned knowledge point vector representations, the cosine similarity between knowledge points is calculated, a threshold is set to construct an adjacency matrix, which is used to supplement the initial knowledge point graph constructed from the dataset or human prior knowledge, thus obtaining a complete target knowledge point graph. This target knowledge point graph is then introduced into a graph neural network module, enabling the inference of a student's complete knowledge mastery level based on limited question-answering records.

[0066] After obtaining the target knowledge point graph, the knowledge relationships within it can be enhanced. For example, by adjusting the structure of the target knowledge point graph, such as adding weakly related edges or adjusting edge weights, the model's ability to learn knowledge point relationships can be improved. In practice, this can be achieved using node2vec and rules (such as adding undirected edges / correlation coefficients representing similarity relationships between final-level knowledge points belonging to the same parent knowledge point). This embodiment does not restrict this enhancement process.

[0067] In summary, a knowledge point graph can be obtained. In this embodiment, the process of obtaining this knowledge point graph is not limited. As long as the knowledge point graph can be obtained, subsequent processing can be performed based on the knowledge point graph.

[0068] Step 203: Obtain the question sequence and auxiliary information sequence of the target object. The question sequence may include multiple questions (i.e., the answer sequence of the target object). The auxiliary information sequence may include the knowledge point name, question difficulty and question response, such as correct response or incorrect response, for each question in the question sequence.

[0069] For example, when it is necessary to predict the knowledge mastery level of a user A (such as student A), user A is the target object, and it is necessary to obtain the question sequence of user A. This question sequence can also be called the answer sequence. The question sequence can include multiple question texts and the answer results of each question text. For ease of description, the question text and the answer results of the question text are collectively referred to as a question.

[0070] For example, each question text can be a text-based question, such as a multiple-choice question, fill-in-the-blank question, or word problem. User A can answer each question text, obtaining the answer result for that question text. Based on this, these question texts and the answers for each question text can be combined to obtain a question sequence for the target object. For example, the question sequence may include question c1 (such as a question text and the target object's answer to that question text), question c2, ..., and so on; that is, the question sequence can include multiple questions.

[0071] For example, for each question in the question sequence, we can obtain the corresponding knowledge point name, such as knowledge point name d1 for question c1, knowledge point name d2 for question c2, and so on. This way, we can obtain the knowledge point name for each question in the question sequence. Considering that user A will not answer every question with every candidate knowledge point name, the knowledge point names corresponding to the question sequence are only a subset of the N candidate knowledge point names. For example, if N candidate knowledge point names are 1000, the knowledge point names corresponding to the question sequence are only 50 of them. Each knowledge point name represents the name of a knowledge point in the question text, such as addition, addition operation, addition calculation rules, addition calculation, subtraction, subtraction operation, etc.

[0072] For example, for each question in the question sequence, the difficulty level can be obtained, which represents the difficulty of the question (question text). For instance, it can be divided into three difficulty levels: low, medium, and high, with the question corresponding to any of these levels. Alternatively, it can be divided into four difficulty levels: level one, level two, level three, and level four, with the question corresponding to any of these levels. Or, it can be divided into five difficulty levels: level 0, level 1, level 2, level 3, and level 4, with the question corresponding to any of these levels.

[0073] For example, for each question in the question sequence, the corresponding question response can be obtained, which can be a correct response or an incorrect response. For instance, if user A's response to a question is a correct response, it means that user A's answer to the question is correct; if user A's response to a question is an incorrect response, it means that user A's answer to the question is incorrect.

[0074] In summary, for each question in the question sequence, we can obtain the corresponding knowledge point name, question difficulty, and question response, thus obtaining an auxiliary information sequence. This auxiliary information sequence can include auxiliary information for each question in the question sequence, and the auxiliary information can include the knowledge point name, question difficulty, and question response. For example, assuming there are 50 questions in the question sequence, the auxiliary information sequence will contain 50 pieces of auxiliary information, each including the knowledge point name, question difficulty, and question response.

[0075] In one possible implementation, a student profile can be constructed for each target object (e.g., user A). For example, based on the actual needs of the educational scenario, basic attribute information of the target object can be designed, which may include key features such as grade level, knowledge point mastery, and ability value. Regarding the target object's knowledge point mastery, factors such as knowledge point difficulty, grade adaptability, knowledge point dependencies, and individual differences are comprehensively considered to simulate the distribution of student abilities in a real-world scenario. The subsequent process in this embodiment requires predicting the target object's knowledge point mastery and then constructing the student profile.

[0076] In one possible implementation, to construct a question sequence (i.e., an answer sequence), the answer sequence can be collected based on the target object's historical answer data; that is, all data within the answer sequence is actually collected data. Alternatively, based on the target object's ability score, question difficulty, and knowledge point mastery, the answer sequence for the target object can be generated through a probabilistic model. This process considers the time factor, making the generated data more consistent with the actual learning process; that is, the data within the answer sequence is simulated data. Alternatively, some data within the answer sequence may be actually collected data, while some data may be simulated data.

[0077] For example, after obtaining the answer sequence, it can be expanded. For instance, by appropriately perturbing and transforming the answer sequence, an updated answer sequence can be obtained. For example, perturbation and transformation could refer to operations such as randomly deleting some answer records or adjusting the answer order.

[0078] For example, a dynamic knowledge status update mechanism can be designed to adjust the level of knowledge mastery in real time based on the target audience's answer results. This dynamic knowledge status update mechanism takes into account the interrelationships between knowledge points, reflecting the continuity of the learning process.

[0079] Step 204: Determine the question embedding feature vector corresponding to the question sequence, the auxiliary embedding feature vector corresponding to the auxiliary information sequence, and the knowledge point embedding feature vector corresponding to the knowledge point description sequence.

[0080] For example, a question sequence may include multiple questions, and a question embedding feature vector may include multiple question feature values, with each feature value corresponding one-to-one with a question within the question sequence. For instance, a feature extraction network can be used to extract features from the question sequence to obtain the question embedding feature vector. When extracting features from the question sequence, it is necessary to extract the question feature value corresponding to each question within the sequence. In this way, the question feature values ​​corresponding to all questions can be combined to form the question embedding feature vector.

[0081] For example, the feature extraction network can be a Rasch Embedding module, or other types of embedding modules; there are no restrictions, as long as it can convert the question sequence into question embedding feature vectors. The Rasch Embedding module is an embedding module that maps high-dimensional data (such as text, images, and videos) to a low-dimensional space. The embedding vector is an N-dimensional real-valued vector that represents the input data as points in a continuous numerical space.

[0082] For example, the auxiliary information sequence may include multiple auxiliary information items (such as knowledge point name, question difficulty, and question response), and the auxiliary embedding feature vector may include multiple auxiliary feature values, with each auxiliary feature value corresponding one-to-one with a different auxiliary information item within the auxiliary information sequence. For instance, a feature extraction network can be used to extract features from the auxiliary information sequence to obtain the auxiliary embedding feature vector. When extracting features from the auxiliary information sequence, it is necessary to extract the auxiliary feature value corresponding to each auxiliary information item within the sequence (an auxiliary feature value corresponding to one auxiliary information item can be a single feature value, such as the same feature value for knowledge point name, question difficulty, and question response; or, an auxiliary feature value corresponding to one auxiliary information item can be three feature values, such as different feature values ​​for knowledge point name, question difficulty, and question response). In this way, the auxiliary feature values ​​corresponding to all auxiliary information items can be combined to form the auxiliary embedding feature vector.

[0083] For example, the feature extraction network can be a Rasch Embedding module or other types of embedding networks. There are no restrictions on this, as long as it can convert the auxiliary information sequence into auxiliary embedding feature vectors.

[0084] For example, a knowledge point description sequence may include semantic description information corresponding to multiple candidate knowledge point names, and a knowledge point embedding feature vector may include multiple knowledge point feature values, with each knowledge point feature value corresponding one-to-one with a semantic description within the knowledge point description sequence. For instance, a feature extraction network can be used to extract features from the knowledge point description sequence to obtain the knowledge point embedding feature vector. When extracting features from the knowledge point description sequence, it is necessary to extract the knowledge point feature value corresponding to each semantic description within the sequence. This allows all knowledge point feature values ​​to be combined into the knowledge point embedding feature vector. The feature extraction network can be a Rasch Embedding module or other types of embedding networks; there are no restrictions, as long as it can convert the knowledge point description sequence into a knowledge point embedding feature vector.

[0085] In one possible implementation, see Figure 3A The diagram shown illustrates the structure of a knowledge point mastery prediction model. This is merely an example; the network structure of this model is not limited and can be configured according to actual needs. Figure 3A In this context, the knowledge point mastery prediction model can include a first feature extraction network (such as a Rasch Embedding module) and a second feature extraction network (such as a Rasch Embedding module). From... Figure 3A As can be seen, the question sequence can be input into the first feature extraction network (such as the Rasch Embedding module). The first feature extraction network determines the question embedding feature vector corresponding to the question sequence. The question embedding feature vector is a historical sequence, which can be denoted as question embedding q_emb. The auxiliary information sequence can also be input into the first feature extraction network (such as the Rasch Embedding module). The first feature extraction network determines the auxiliary embedding feature vector corresponding to the auxiliary information sequence. The auxiliary embedding feature vector is a historical sequence, which can be denoted as answer embedding s_emb.

[0086] The knowledge point sequence, question text sequence, and reference matrix can be input into a Large Language Model (LLM). The LLM determines the semantic description information (also known as knowledge point explanation information or knowledge point description information) corresponding to each candidate knowledge point name, resulting in a knowledge point description sequence. This knowledge point description sequence includes the semantic description information corresponding to each candidate knowledge point name. Then, the knowledge point description sequence can be input into a second feature extraction network. The second feature extraction network determines the knowledge point embedding feature vector corresponding to the knowledge point description sequence, denoted as knowledge point embedding c_embed. Alternatively, other methods can be used to determine the knowledge point embedding feature vector corresponding to the knowledge point description sequence; there are no restrictions on this.

[0087] Step 205: Perform multi-head attention operation based on the question embedding feature vector and the auxiliary embedding feature vector to obtain the target object's ability feature vector for the knowledge point name corresponding to the question sequence.

[0088] In one possible implementation, see Figure 3A As shown, the knowledge point mastery prediction model can include an MHA (Multi-Head Attention) network. The Q-vector, K-vector, and V-vector of the MHA network can be determined based on the question embedding feature vector and the auxiliary embedding feature vector. Furthermore, the MHA network can process the Q-vector, K-vector, and V-vector to obtain an ability feature vector, which can also be called a dense ability representation z. Since the knowledge point names corresponding to the question sequence are only a subset of N candidate knowledge point names (e.g., 50 out of 1000 candidate knowledge point names), this ability feature vector is the target object's ability feature vector for the 50 knowledge point names corresponding to the question sequence, not the ability feature vector for all N candidate knowledge point names.

[0089] For example, the MHA network is a key component of the Transformer model, enabling it to focus on different parts of the input sequence simultaneously in different representation subspaces. Furthermore, the Transformer model is a neural network architecture that utilizes a self-attention mechanism to process sequential data in parallel.

[0090] For example, the Q-vector, K-vector, and V-vector can refer to the three input representation vectors used in the self-attention mechanism. The Q-vector can represent the query vector, the K-vector can represent the key vector, and the V-vector can represent the numerical vector. In the self-attention mechanism, based on the Q-vector, a weighted distribution is obtained by calculating the similarity between the query vector Q and all key vectors K, which is then used to weight and sum the associated numerical vector V. Simply put, the query vector Q is matched with the key vectors K based on similarity, and the result of the matching is the numerical vector V.

[0091] This embodiment does not impose any restrictions on how the MHA network processes data based on Q-vectors, K-vectors, and V-vectors. It is related to the function of the MHA network. The processing of the MHA network is called multi-head attention operation, that is, the MHA network performs multi-head attention operation based on Q-vectors, K-vectors, and V-vectors. The output features of the MHA network are called capability feature vectors, that is, the output capability feature vectors of the MHA network.

[0092] In one possible implementation, see Figure 3B The diagram shown is a structural schematic of the knowledge point mastery prediction model. The MHA network of the knowledge point mastery prediction model can include four MHA subnetworks, denoted as MHA subnetwork 1, MHA subnetwork 2, MHA subnetwork 3 and MHA subnetwork 4. The number of MHA subnetworks can be more or less, and there is no limit to the number of MHA subnetworks.

[0093] See Figure 3B As shown, the first Q-vector, first K-vector, and first V-vector of MHA sub-network 1 can be determined based on the question embedding feature vector. For example, the question embedding feature vector can be used as the first Q-vector, the first K-vector, and the first V-vector of MHA sub-network 1. A first intermediate vector can be obtained by performing a multi-head attention operation based on the first Q-vector, first K-vector, and first V-vector. For instance, MHA sub-network 1 can perform a multi-head attention operation based on the first Q-vector, first K-vector, and first V-vector, and the output feature of MHA sub-network 1 is called the first intermediate vector.

[0094] See Figure 3BAs shown, the second Q-vector, second K-vector, and second V-vector of the MHA sub-network 2 can be determined based on this auxiliary embedding feature vector. For example, this auxiliary embedding feature vector can be used as the second Q-vector, the second K-vector, and the second V-vector of the MHA sub-network 2. A second intermediate vector can be obtained by performing a multi-head attention operation based on the second Q-vector, second K-vector, and second V-vector. For example, the output feature of the MHA sub-network 2 can be called the second intermediate vector by performing a multi-head attention operation based on the second Q-vector, second K-vector, and second V-vector.

[0095] See Figure 3B As shown, the third Q-vector and third K-vector of MHA subnetwork 3 can be determined based on the first intermediate vector, and the third V-vector of MHA subnetwork 3 can be determined based on the second intermediate vector. For example, the first intermediate vector can be used as the third Q-vector of MHA subnetwork 3, the first intermediate vector can be used as the third K-vector of MHA subnetwork 3, and the second intermediate vector can be used as the third V-vector of MHA subnetwork 3. The third intermediate vector can be obtained by performing a multi-head attention operation based on the third Q-vector, the third K-vector, and the third V-vector. For example, MHA subnetwork 3 can perform a multi-head attention operation based on the third Q-vector, the third K-vector, and the third V-vector, and the output feature of MHA subnetwork 3 is called the third intermediate vector. For example, the third intermediate vector is used to represent the problem-level mastery h; that is, the third intermediate vector can also be called the problem-level mastery vector, reflecting the target object's mastery of the problem level.

[0096] See Figure 3BAs shown, the fourth K vector of MHA subnetwork 4 can be determined based on the first intermediate vector, the fourth V vector of MHA subnetwork 4 can be determined based on the third intermediate vector, and the fourth Q vector of MHA subnetwork 4 can be determined based on the knowledge space feature vector (know_params, the method of obtaining the knowledge space feature vector is described in subsequent embodiments). For example, the first intermediate vector can be used as the fourth K vector of MHA subnetwork 4, the third intermediate vector can be used as the fourth V vector of MHA subnetwork 4, and the knowledge space feature vector can be used as the fourth Q vector of MHA subnetwork 4. The fourth intermediate vector can be obtained by performing a multi-head attention operation based on the fourth Q vector, the fourth K vector, and the fourth V vector. For example, MHA subnetwork 4 can perform a multi-head attention operation based on the fourth Q vector, the fourth K vector, and the fourth V vector, and the output feature of MHA subnetwork 4 is called the fourth intermediate vector. After obtaining the fourth intermediate vector, the ability feature vector (i.e., dense ability representation z) can be determined based on the fourth intermediate vector. For example, the fourth intermediate vector can be used as the ability feature vector, or the ability feature vector can be obtained by operating on the fourth intermediate vector, that is, the ability feature vector of the target object for the knowledge point name corresponding to the question sequence.

[0097] Step 206: Determine the knowledge space feature vector based on the knowledge point embedding feature vector. The knowledge point embedding feature vector includes the initial feature vector of each candidate knowledge point name. The knowledge space feature vector includes multiple fused feature vectors, and each fused feature vector is obtained by fusing multiple initial feature vectors.

[0098] For example, the knowledge point embedding feature vector (c_embed) can include multiple knowledge point feature values ​​corresponding to multiple candidate knowledge point names. The knowledge point feature values ​​can be called the initial feature vector. That is, the knowledge point embedding feature vector includes N initial feature vectors corresponding to N candidate knowledge point names.

[0099] When determining the knowledge space feature vector (know_params) based on the knowledge point embedding feature vector, multiple initial feature vectors can be fused to obtain a fused feature vector. The number of fused feature vectors can also be multiple, but it is less than the number of initial feature vectors. That is, the fusion operation is used to reduce the dimensionality of the knowledge point embedding feature vector. Thus, the knowledge space feature vector can include multiple fused feature vectors. For example, four initial feature vectors from the knowledge point embedding feature vector can be fused to obtain one fused feature vector. In this way, N initial feature vectors can be fused to obtain N / 4 fused feature vectors, and the knowledge space feature vector can include N / 4 fused feature vectors. Of course, this is just an example of the fusion operation, and there are no limitations on this operation.

[0100] In one possible implementation, see Figure 3A and Figure 3B As shown, the knowledge point mastery prediction model can include an MHA network (different from the MHA network in step 205, it is a different MHA network) and an MLP network. The fifth Q-vector, fifth K-vector, and fifth V-vector of the MHA network can be determined based on the knowledge point embedding feature vector. For example, the knowledge point embedding feature vector can be used as the fifth Q-vector, the fifth K-vector, and the fifth V-vector of the MHA network. A fifth intermediate vector can be obtained by performing a multi-head attention operation based on the fifth Q-vector, fifth K-vector, and fifth V-vector. For example, the MHA network can perform a multi-head attention operation based on the fifth Q-vector, fifth K-vector, and fifth V-vector, and the output feature of the MHA network is called the fifth intermediate vector. After obtaining the fifth intermediate vector, it can be input into the MLP network (Multi-Layer Perceptron).

[0101] After obtaining the fifth intermediate vector, the MLP network can perform multilayer perceptron operations based on this fifth intermediate vector to obtain the knowledge space feature vector (know_params). For example, the MLP network can process based on this fifth intermediate vector, and the output feature of the MLP network is called the knowledge space feature vector. This embodiment does not impose restrictions on how the MLP network processes based on the fifth intermediate vector; it depends on the function of the MLP network. For example, an MLP network is an artificial neural network that can consist of an input layer, one or more hidden layers, and an output layer, used to handle complex nonlinear relationships. The MLP network is a fully connected layer used to concatenate (fuse) the feature vectors in the fifth intermediate vector (such as the N initial feature vectors corresponding to the N candidate knowledge point names), and can map the fused feature vectors to the input space of the subsequent MHA network through the fully connected layer.

[0102] Step 207: Perform multi-head attention operation based on the capability feature vector, the knowledge point embedding feature vector, and the knowledge space feature vector to obtain the knowledge point target feature vector.

[0103] For example, the seventh Q vector, seventh K vector, and seventh V vector of the MHA network can be determined based on the capability feature vector, the knowledge point embedding feature vector, and the knowledge space feature vector. The MHA network can perform multi-head attention operation based on the seventh Q vector, seventh K vector, and seventh V vector to obtain the knowledge point target feature vector. That is, the output feature of the MHA network is used as the knowledge point target feature vector.

[0104] In one possible implementation, see Figure 3A and Figure 3B As shown, the knowledge point mastery prediction model may include an MHA network (different from the MHA networks in steps 205 and 206, it is a different MHA network). The seventh V vector of the MHA network can be determined based on the ability feature vector; for example, the ability feature vector can be used as the seventh V vector of the MHA network. The seventh K vector of the MHA network can be determined based on the knowledge space feature vector; for example, the knowledge space feature vector can be used as the seventh K vector of the MHA network. The seventh Q vector of the MHA network can be determined based on the knowledge point embedding feature vector; for example, the knowledge point embedding feature vector can be used as the seventh Q vector of the MHA network. A multi-head attention operation can be performed based on the seventh Q vector, seventh K vector, and seventh V vector to obtain the knowledge point target feature vector. For example, the MHA network can perform a multi-head attention operation based on the seventh Q vector, seventh K vector, and seventh V vector, and the output feature of the MHA network is called the knowledge point target feature vector.

[0105] For example, after inputting the ability feature vector, the knowledge point embedding feature vector, and the knowledge space feature vector into the MHA network, the MHA network calculates the correlation between the knowledge point embedding feature vector and the ability feature vector (i.e., the student's ability vector). In other words, the knowledge point target feature vector reflects the correlation between the knowledge point embedding feature vector and the ability feature vector. For instance, the knowledge point embedding feature vector can be used as the Q-vector (query) of the MHA network, the knowledge space feature vector as the K-vector (key) of the MHA network, and the ability feature vector as the V-vector (value) of the MHA network. Through attention calculation, the MHA network can find the most relevant student ability dimension for each knowledge point, obtaining the degree of correlation between each knowledge point and the student's ability vector, thus obtaining a more accurate representation of the knowledge point mastery level.

[0106] For example, after obtaining the target feature vector of a knowledge point, the MHA network can input the target feature vector into the next network layer (such as an MLP network). Alternatively, the MHA network can normalize the target feature vector of the knowledge point and input the normalized target feature vector into the next network layer. Alternatively, the MHA network can perform residual connection processing on the target feature vector of the knowledge point and input the residual connection-processed target feature vector into the next network layer. Alternatively, the MHA network can perform both normalization and residual connection processing on the target feature vector of the knowledge point and input the residual connection-processed target feature vector into the next network layer.

[0107] Step 208: Generate target graph features based on the knowledge point target feature vector and the first graph features. The target graph features may include the fusion feature vector of each candidate knowledge point name in the knowledge point graph. The fusion feature vector is obtained by concatenating the knowledge point target feature vector and the initial feature vector.

[0108] For example, a knowledge point graph is a graph representing all candidate knowledge point names (e.g., N candidate knowledge point names). This graph can include the N candidate knowledge point names and the connections between them, reflecting these connections. Based on this, a first graph feature can be determined. This first graph feature includes the initial feature vector of each candidate knowledge point name within the knowledge point graph, such as the initial feature vectors of the N candidate knowledge point names. For instance, for each candidate knowledge point name within the knowledge point graph, feature extraction can be performed to obtain its initial feature vector. The initial feature vectors of all candidate knowledge point names can then constitute the first graph feature.

[0109] Furthermore, a second graph feature can be determined based on the knowledge point graph. This second graph feature includes the similarity between the initial feature vectors of two candidate knowledge point names within the knowledge point graph. For example, this second graph feature can be an N*N dimensional feature matrix. The elements in the first row represent the similarity between the first candidate knowledge point name and the initial feature vector of each candidate knowledge point name. For instance, if the first element is 1 (indicating a similarity of 1), the second element represents the similarity between the initial feature vectors of the first and second candidate knowledge point names, and so on. The elements in the second row represent the similarity between the second candidate knowledge point name and the initial feature vector of each candidate knowledge point name, and so on.

[0110] For example, for two candidate knowledge point names that are connected (the knowledge point graph includes N candidate knowledge point names and the connections between them; the graph can be used to determine which candidate knowledge point names are connected and which are not), the similarity between their initial feature vectors can be calculated. This similarity can be a value between 0 and 1. For two candidate knowledge point names that are not connected, the similarity between their initial feature vectors can be 0.

[0111] For example, after obtaining the target feature vector of the knowledge points and the first graph features, the target feature vector of the knowledge points can be concatenated with each initial feature vector within the first graph features to obtain the fused feature vector corresponding to that initial feature vector. This results in a total of N fused feature vectors for candidate knowledge point names. For instance, concatenating the target feature vector of the knowledge point with the initial feature vector of candidate knowledge point name a1 (e.g., the target feature vector precedes or follows the initial feature vector) yields the fused feature vector of candidate knowledge point name a1. Concatenating the target feature vector of the knowledge point with the initial feature vector of candidate knowledge point name a2 yields the fused feature vector of candidate knowledge point name a2, and so on, resulting in N fused feature vectors. Thus, the target graph features can include fused feature vectors of N candidate knowledge point names.

[0112] Step 209: Based on the target graph features, the first graph features, and the second graph features, predict the target object's mastery of the target knowledge points for each candidate knowledge point name in the knowledge point graph.

[0113] For example, see Figure 3A and Figure 3B As shown, the knowledge mastery prediction model includes MLP networks and GCN networks. See [link / reference]. Figure 3C The diagram shown is a structural illustration of a knowledge point mastery prediction model; only a portion of the model's structure is illustrated here. From... Figure 3C As can be seen, the concept map is input into the MLP network and the GCN network. For example, based on the concept map, first and second map features are determined. The first map feature is input into the MLP network, and the first and second map features are input into the GCN network. Furthermore, the MHA network can input the target feature vector of the knowledge points into the MLP network. Thus, in the MLP network, the target map feature can be generated based on the target feature vector of the knowledge points and the first map feature; this process can be seen in step 208. Alternatively, the target map feature can be generated based on the target feature vector of the knowledge points and the first map feature during the processing before the MLP network, and then this target map feature is input into the MLP network.

[0114] For example, after obtaining the target graph features, a feature mapping operation is performed on the target graph features using an MLP network (i.e., a multilayer perceptron) to obtain the mapped graph features. The feature dimension of the mapped graph features is consistent with the input feature dimension of the graph neural network (i.e., a GCN network). In other words, after concatenating and fusing the output features of the MHA network (such as the target feature vector of the knowledge point) with the first graph features (i.e., the initial feature vector of the candidate knowledge point name) to obtain the enhanced knowledge point representation (i.e., the target graph features), the fused target graph features can be mapped to the input space of the graph neural network using the MLP network. That is, the target graph features are mapped to the required input dimension of the graph neural network using the MLP network. For example, if the input feature dimension of the graph neural network is A*B, then the target graph features are mapped to A*B features using the MLP network, and this A*B feature is denoted as the mapped graph features.

[0115] For example, after obtaining the mapped graph features, the mapped graph features can be input into the graph neural network, along with the first and second graph features. In this way, the graph neural network makes predictions based on the mapped graph features, the first graph features, and the second graph features to obtain the mastery level of the target knowledge point. This embodiment does not limit this prediction process; the graph neural network only needs to be able to obtain the mastery level of the target knowledge point based on the mapped graph features, the first graph features, and the second graph features.

[0116] For example, the mapped graph features are determined based on the target feature vector of knowledge points, and the target feature vector of knowledge points is determined based on the ability feature vector. Therefore, the mapped graph features can reflect the ability feature vector of the target object to the knowledge point names corresponding to the question sequence, that is, the target object's ability to some candidate knowledge point names. Since the first graph feature reflects the initial feature vector of each candidate knowledge point name in the knowledge point graph, and the second graph feature reflects the similarity of the initial feature vectors of two candidate knowledge point names in the knowledge point graph, by combining the mapped graph features, the first graph feature, and the second graph feature, we can obtain the target object's mastery of the target knowledge points for each candidate knowledge point name in the knowledge point graph based on the target object's ability to some candidate knowledge point names. For example, based on the target object's ability to 50 candidate knowledge point names, we can obtain the target object's mastery of the target knowledge points for 1000 candidate knowledge point names, and thus obtain the target object's mastery of the target knowledge points for all candidate knowledge point names.

[0117] For example, this embodiment does not impose any restrictions on how the graph neural network makes predictions based on the mapped graph features, the first graph features, and the second graph features. It depends on the capabilities of the graph neural network. The graph neural network only needs to be able to output the target object's mastery of the target knowledge points for all candidate knowledge point names.

[0118] For example, a graph neural network can also be called a graph convolutional neural network. A graph neural network can use a GCN (Graph Convolutional Network) network. A GCN network is a neural network used to process graph structure data. By passing information between graph nodes and their neighbors, it learns the embedded representation of nodes. A GCN network can be used to process the relationships between knowledge points.

[0119] For example, a single-layer GCN network can be used for prediction based on the mapped graph features, the first graph features, and the second graph features. Alternatively, a multi-layer GCN network can be used, also based on the mapped graph features, the first graph features, and the second graph features. A two-layer GCN network will be used as an example. In practical applications, the number of layers in the GCN network can be adjusted according to the complexity of the knowledge point graph and the sparsity of the dataset to reasonably utilize the prior knowledge of the knowledge point graph to assist in predicting knowledge point mastery.

[0120] The first layer of the GCN network can transform and message-pass the mapped graph features, first graph features, and second graph features. Through the message-passing mechanism of the graph structure, the prediction result of each knowledge point can take into account the information of related knowledge points. The first layer of the GCN network can transform features and perform message passing between knowledge points. Non-linear activation functions and dropout operations can also be added to the first layer of the GCN network to enhance the model's expressive power. The second layer of the GCN network is used to map features to the category space of knowledge point mastery, that is, to obtain the target object's mastery of the candidate knowledge point name.

[0121] For example, in the above processing, the input dimension can also be configured, which represents the dimension of the student's ability vector. The node feature dimension of the knowledge point can be configured, which is used to represent the feature information of the knowledge point. The number of input channels of the graph convolutional network can be configured, which is used to control the dimension of feature transformation. There are no restrictions on the parameters of the knowledge point mastery prediction model.

[0122] In one possible implementation, after obtaining the target audience's mastery level of the candidate knowledge point names, data validation can be performed. Data validation can include rationality validation and distribution validation. Rationality validation refers to verifying the rationality and authenticity of the generated data (i.e., the mastery level of the target knowledge point) through evaluation by educational experts and statistical analysis. Distribution validation refers to ensuring that the generated data (i.e., the mastery level of the target knowledge point) meets the expected distribution requirements in all dimensions (such as difficulty distribution, knowledge point coverage, etc.). See also Figure 4 As shown, this is a schematic diagram illustrating the statistical effect of the question difficulty distribution on the sample dataset. Figure 4The distribution is validated using visual statistical methods such as bar charts and probability density curves. Figure 4 neutron Figure 1 (i.e., the lower subgraph) represents a histogram of the difficulty of all questions. Figure 2 (i.e., the upper side subgraph) represents the difficulty PDF (probability density function) of all problems.

[0123] This completes the knowledge point mastery prediction, which can predict the target object's mastery of the target knowledge point for each candidate knowledge point name within the knowledge point graph. In the above process, a knowledge point mastery prediction model is used to predict the target knowledge point mastery. The training process for this model can include:

[0124] Collect sample data, which includes a question sequence and an auxiliary information sequence for the sample object. The question sequence may include multiple questions (i.e., the answer sequence of the sample object), and the auxiliary information sequence may include the knowledge point name, question difficulty, and question response for each question in the question sequence.

[0125] Based on the sample data, steps 204-209 are performed to obtain the sample object's mastery level of each candidate knowledge point name within the knowledge point graph. Furthermore, when acquiring the sample data, the sample object's mastery level of each candidate knowledge point name within the knowledge point graph can also be labeled.

[0126] Based on this, the loss value can be calculated based on the knowledge point mastery of the samples and the knowledge point mastery of the labels. The network parameters of the knowledge point mastery prediction model can be adjusted based on the loss value. The goal of the adjustment is to make the loss value smaller and smaller, that is, to make the knowledge point mastery of the samples and the knowledge point mastery of the labels closer and closer. After multiple iterations, the knowledge point mastery prediction model is converged. The converged knowledge point mastery prediction model can then be output, thus completing the training process of the knowledge point mastery prediction model.

[0127] When calculating the loss value based on the knowledge mastery of sample knowledge points and the knowledge mastery of label knowledge points, a semi-sparse loss function can be used, or other loss functions can be used; there are no restrictions on which loss function to use. For example, to avoid class imbalance that could lead to biased model training, a semi-sparse loss function can be used to calculate the loss value. See the following formula for an example of a loss function: In the loss function described above, It can represent the loss value. This can represent the cross-entropy loss value. This represents the level of mastery of the knowledge points related to the tags. This represents the sample's mastery of knowledge points, i.e., the predicted value of the mastery of knowledge points.

[0128] This application proposes a method for predicting knowledge point mastery. Besides predicting the mastery of target knowledge points, it can also predict the response to the question being predicted, i.e., predict whether the answer is correct or incorrect. See [link to relevant documentation]. Figure 5 The diagram shown is a flowchart of the method, which may include:

[0129] Step 501: Obtain the knowledge point description sequence, which includes the semantic description information corresponding to each candidate knowledge point name (each knowledge point name is denoted as the candidate knowledge point name).

[0130] Step 502: Obtain the knowledge point graph. The knowledge point graph is a graph of all candidate knowledge point names. The knowledge point graph includes N candidate knowledge point names and the connection relationships between the candidate knowledge point names.

[0131] Step 503: Obtain the question sequence and auxiliary information sequence of the target object. The question sequence may include multiple completed questions (i.e., the answers to questions that the target object has completed) and questions to be predicted (i.e., the answers to questions that the target object has not completed, and it is necessary to predict whether the answer to this question may be correct or incorrect). The auxiliary information sequence may include the knowledge point name, question difficulty and question response, such as correct response or incorrect response, for each completed question in the question sequence. The auxiliary information sequence may also include the knowledge point name and question difficulty for the questions to be predicted, but does not include the question response for the questions to be predicted.

[0132] For example, compared to step 203, the question sequence is supplemented with a question to be predicted, and the auxiliary information sequence is supplemented with the knowledge point name and question difficulty corresponding to the question to be predicted. In this way, the prediction question response corresponding to the question to be predicted can be predicted through the knowledge point mastery prediction model.

[0133] Step 504: Determine the question embedding feature vector corresponding to the question sequence, the auxiliary embedding feature vector corresponding to the auxiliary information sequence, and the knowledge point embedding feature vector corresponding to the knowledge point description sequence.

[0134] Step 505: Perform multi-head attention operation based on the question embedding feature vector and the auxiliary embedding feature vector to obtain the target object's ability feature vector for the knowledge point name corresponding to the question sequence.

[0135] Step 506: Determine the feature vector of the knowledge space based on the embedded feature vector of the knowledge point.

[0136] Step 507: Perform multi-head attention operation based on the capability feature vector, the knowledge point embedding feature vector, and the knowledge space feature vector to obtain the knowledge point target feature vector.

[0137] Step 508: Generate target graph features based on the knowledge point target feature vector and the first graph features. The target graph features may include the fusion feature vector of each candidate knowledge point name in the knowledge point graph. The fusion feature vector is obtained by concatenating the knowledge point target feature vector and the initial feature vector.

[0138] Step 509: Based on the target graph features, the first graph features, and the second graph features, predict the target object's mastery of the target knowledge points for each candidate knowledge point name in the knowledge point graph.

[0139] For example, steps 501-509 are similar to steps 201-209, and will not be described again here.

[0140] Step 510: Based on the question embedding feature vector, ability feature vector, and knowledge space feature vector, determine the predicted question response corresponding to the question to be predicted; wherein, the predicted question response can indicate whether the target object's answer to the question to be predicted is correct or incorrect.

[0141] For example, see Figure 6 The image shown is a schematic diagram of the knowledge point mastery prediction model. Figure 3A and Figure 3B In comparison, the knowledge mastery prediction model adds an MHA network and an MLP network.

[0142] After obtaining the question embedding feature vector, ability feature vector, and knowledge space feature vector, the sixth Q-vector, sixth V-vector, and sixth K-vector of the MHA network are determined based on these vectors. For example, the sixth Q-vector is determined based on the question embedding feature vector, the sixth V-vector based on the ability feature vector, and the sixth K-vector based on the knowledge space feature vector. Alternatively, the question embedding feature vector can be used as the sixth Q-vector of the MHA network, the ability feature vector as the sixth V-vector, and the knowledge space feature vector as the sixth K-vector. Then, a multi-head attention operation is performed on the sixth Q-vector, sixth K-vector, and sixth V-vector to obtain the sixth intermediate vector. For example, the MHA network can perform a multi-head attention operation on the sixth Q-vector, sixth K-vector, and sixth V-vector, and the output feature of the MHA network is called the sixth intermediate vector. After obtaining the sixth intermediate vector, it can be input into the MLP network.

[0143] After obtaining the sixth intermediate vector, the MLP network can perform multilayer perceptron operations based on this sixth intermediate vector to obtain post-operated features. For example, the MLP network can process based on this sixth intermediate vector, and the output features of the MLP network are called post-operated features. This application does not limit how the MLP network processes based on this sixth intermediate vector.

[0144] After obtaining the post-operation feature, the predicted question response can be determined based on this feature. For example, the post-operation feature can be the predicted question response. If the post-operation feature (i.e., the predicted question response) has a first value (e.g., 1), it indicates that the answer is correct; if the post-operation feature (i.e., the predicted question response) has a second value (e.g., 0), it indicates that the answer is incorrect. Alternatively, the post-operation feature can be input into other network layers, which can then make predictions based on it to obtain the predicted question response. For example, if the predicted question response has a first value, it indicates that the answer is correct; if the predicted question response has a second value, it indicates that the answer is incorrect.

[0145] Thus, the prediction of knowledge point mastery and the prediction of the question to be predicted are completed. It can predict the target audience's mastery of the target knowledge point for each candidate knowledge point name within the knowledge point graph, and predict the target audience's answer to the question to be predicted. In the above process, a knowledge point mastery prediction model is used to predict the mastery of the target knowledge point. The training process of the knowledge point mastery prediction model can include:

[0146] Sample data is collected, including a sequence of questions and a sequence of auxiliary information. The question sequence can include multiple questions (e.g., 10 questions), and the auxiliary information sequence can include the knowledge point name, difficulty level, and response for each question. For example, with 10 questions, if the first question is considered completed, the second question is considered to be predicted, the response for the first question is a normal value, and the responses for the remaining questions are set to zero. If the first and second questions are considered completed, the third question is considered to be predicted, the responses for the first and second questions are normal values, and the responses for the remaining questions are set to zero, and so on, thus obtaining multiple sets of sample data.

[0147] Based on the sample data, steps 504-510 are executed to obtain the sample object's mastery of each candidate knowledge point name in the knowledge point graph and the sample object's answer to the question to be predicted (i.e., the sample question response). When obtaining the sample data, the sample object's mastery of each candidate knowledge point name in the knowledge point graph can be labeled, and the sample object's labeled question response to the question to be predicted can be labeled (when the third question is the question to be predicted, the question response of the third question is the labeled question response).

[0148] Based on this, a first loss value can be calculated based on the mastery of sample knowledge points and the mastery of labeled knowledge points. A second loss value can be calculated based on the responses to sample questions and labeled questions. A target loss value can then be calculated based on the first and second loss values, such as by weighting the first and second loss values. The network parameters of the knowledge point mastery prediction model are then adjusted based on the target loss value. The goal of this adjustment is to minimize the target loss value, meaning that the mastery of sample knowledge points becomes increasingly closer to the mastery of labeled knowledge points, and the responses to sample questions become increasingly closer to the responses to labeled questions. After multiple iterations, this process continues until the knowledge point mastery prediction model converges. The converged knowledge point mastery prediction model is then output, completing the training process of the knowledge point mastery prediction model.

[0149] As can be seen from the above technical solutions, in this embodiment, combining knowledge point graph prediction of knowledge point mastery can accurately assess learners' (e.g., students') knowledge point mastery, reflect it in real time, and achieve precise assessment of knowledge point mastery. The assessment process is simple and fast. By organizing knowledge points into a graph structure, the hierarchical and dependency relationships between knowledge points are fully utilized, making the assessment results more comprehensive and accurate. This enables precise assessment of the mastery level of each specific knowledge point and real-time updates of the assessment results. The assessment results can be interpreted from the perspective of knowledge point association, reflecting both the student's overall learning status and specific knowledge weaknesses. In the process of predicting knowledge point mastery based on graph neural networks, data construction and processing are crucial. Due to the temporal, hierarchical, and relational characteristics of educational data, multiple aspects are considered, including individual student differences, the relationships between knowledge points, and the dynamic changes in the answering process. By modeling the relationships between knowledge points through graph neural networks, the prediction results not only consider individual knowledge points but also utilize information from related knowledge points. For example, in mathematics, the prediction of mastery of "solving two linear equations in two variables" considers the mastery of related knowledge points such as "one linear equation in one variable," making the prediction results more accurate and reasonable. Through multi-head attention mechanisms, the correlation between each knowledge point and each dimension of the student's ability vector can be clearly displayed. This interpretable prediction result helps teachers better understand students' learning status and provides a basis for teaching decisions. Graph neural networks have powerful generalization capabilities; even for combinations of knowledge points not present in the training data, they can make reasonable predictions based on the relationships between knowledge points. They can adapt to different subject knowledge systems and have good transfer capabilities. The model structure is simple, with a small number of parameters, low resource consumption for training and inference, and low requirements for the scale of training data, achieving good results on smaller datasets. For intelligent question recommendation scenarios, the most suitable practice questions can be recommended to students based on the predicted mastery of knowledge points, achieving personalized learning. For learning diagnostic report scenarios, detailed learning diagnostic reports can be generated for teachers and parents, intuitively showing the student's mastery of each knowledge point. For learning path planning scenarios, the optimal learning path can be planned for students based on the dependencies between knowledge points and the student's mastery level.

[0150] Based on the same concept as the above method, this application proposes a knowledge point mastery prediction device, see [link to relevant documentation]. Figure 7 The diagram shown is a structural schematic of the device, which may include:

[0151] The acquisition module 71 is used to acquire a knowledge point description sequence, a question sequence of the target object, and an auxiliary information sequence; the knowledge point description sequence includes semantic description information corresponding to each candidate knowledge point name; the auxiliary information sequence includes the knowledge point name, question difficulty, and question response corresponding to each question in the question sequence.

[0152] The determining module 72 is used to determine the question embedding feature vector corresponding to the question sequence, the auxiliary embedding feature vector corresponding to the auxiliary information sequence, and the knowledge point embedding feature vector corresponding to the knowledge point description sequence; and to determine the knowledge space feature vector based on the knowledge point embedding feature vector, wherein the knowledge point embedding feature vector includes an initial feature vector for each candidate knowledge point name, and the knowledge space feature vector includes multiple fused feature vectors, each of which is obtained by fusing multiple initial feature vectors.

[0153] Processing module 73 is used to perform multi-head attention operation based on the question embedding feature vector and the auxiliary embedding feature vector to obtain the target object's ability feature vector for the knowledge point name corresponding to the question sequence; and to perform multi-head attention operation based on the ability feature vector, the knowledge point embedding feature vector and the knowledge space feature vector to obtain the knowledge point target feature vector.

[0154] The prediction module 74 is used to predict the target object's mastery of the target knowledge point for each candidate knowledge point name in the knowledge point graph based on the target feature vector of the knowledge point and the acquired knowledge point graph.

[0155] For example, when the acquisition module 71 acquires the knowledge point description sequence, it is specifically used for:

[0156] Obtain a knowledge point sequence, a question text sequence, and a reference matrix; wherein, the knowledge point sequence includes multiple candidate knowledge point names, the question text sequence includes multiple example question texts, and the reference matrix represents the correspondence between candidate knowledge point names and example question texts;

[0157] The knowledge point sequence, the question text sequence, and the reference matrix are input into a large language model. The large language model determines the semantic description information corresponding to each candidate knowledge point name, thus obtaining the knowledge point description sequence. Specifically, for each candidate knowledge point name, the reference matrix determines the example question text corresponding to that candidate knowledge point name. Based on the example question text, semantic analysis is performed on the candidate knowledge point name to obtain the semantic description information corresponding to that candidate knowledge point name.

[0158] For example, when the processing module 73 performs a multi-head attention operation based on the question embedding feature vector and the auxiliary embedding feature vector to obtain the ability feature vector of the target object for the knowledge point name corresponding to the question sequence, it specifically performs the following steps: determining a first Q vector, a first K vector, and a first V vector based on the question embedding feature vector; performing a multi-head attention operation based on the first Q vector, the first K vector, and the first V vector to obtain a first intermediate vector; determining a second Q vector, a second K vector, and a second V vector based on the auxiliary embedding feature vector; performing a multi-head attention operation based on the second Q vector, the second K vector, and the second V vector to obtain a second intermediate vector; determining a third Q vector and a third K vector based on the first intermediate vector; determining a third V vector based on the second intermediate vector; performing a multi-head attention operation based on the third Q vector, the third K vector, and the third V vector to obtain a third intermediate vector, where the third intermediate vector represents the level of mastery at the question level; determining a fourth K vector based on the first intermediate vector; determining a fourth V vector based on the third intermediate vector; determining a fourth Q vector based on the knowledge space feature vector; performing a multi-head attention operation based on the fourth Q vector, the fourth K vector, and the fourth V vector to obtain a fourth intermediate vector; and determining the ability feature vector based on the fourth intermediate vector.

[0159] For example, when determining the knowledge space feature vector based on the knowledge point embedding feature vector, the determining module 72 is specifically used to: determine the fifth Q vector, the fifth K vector, and the fifth V vector based on the knowledge point embedding feature vector; perform multi-head attention operation based on the fifth Q vector, the fifth K vector, and the fifth V vector to obtain the fifth intermediate vector; input the fifth intermediate vector to the multilayer perceptron; and obtain the knowledge space feature vector by performing multilayer perception operation based on the fifth intermediate vector through the multilayer perceptron.

[0160] For example, when the prediction module 74 predicts the target object's mastery of the target knowledge point for each candidate knowledge point name in the knowledge point graph based on the target feature vector of the knowledge point and the acquired knowledge point graph, it specifically performs the following steps: determining a first graph feature and a second graph feature based on the knowledge point graph, wherein the first graph feature includes the initial feature vector of each candidate knowledge point name in the knowledge point graph, and the second graph feature includes the similarity of the initial feature vectors of two candidate knowledge point names in the knowledge point graph; generating a target graph feature based on the target feature vector of the knowledge point and the first graph feature, wherein the target graph feature includes a fused feature vector of each candidate knowledge point name in the knowledge point graph, the fused feature vector being obtained by concatenating the target feature vector of the knowledge point and the initial feature vector; and predicting the target object's mastery of the target knowledge point for each candidate knowledge point name in the knowledge point graph based on the target graph feature, the first graph feature, and the second graph feature.

[0161] For example, when the prediction module 74 predicts the target object's mastery of each candidate knowledge point name in the knowledge point graph based on the target graph features, the first graph features, and the second graph features, it specifically performs the following: inputting the target graph features into a multilayer perceptron, and performing a feature mapping operation based on the target graph features through the multilayer perceptron to obtain mapped graph features; wherein the feature dimension of the mapped graph features is consistent with the input feature dimension of the graph neural network; inputting the mapped graph features, the first graph features, and the second graph features into the graph neural network, and performing a prediction based on the mapped graph features, the first graph features, and the second graph features through the graph neural network to obtain the target knowledge point mastery.

[0162] For example, the question sequence further includes questions to be predicted, and the auxiliary information sequence further includes the knowledge point name and question difficulty corresponding to the questions to be predicted; the prediction module 74 is further configured to:

[0163] Based on the question embedding feature vector, the ability feature vector, and the knowledge space feature vector, the predicted question response corresponding to the question to be predicted is determined; wherein, the predicted question response indicates whether the target object's answer to the question to be predicted is correct or incorrect.

[0164] When the prediction module 74 determines the prediction question response corresponding to the question to be predicted based on the question embedding feature vector, the ability feature vector, and the knowledge space feature vector, it is specifically used for:

[0165] The sixth Q vector is determined based on the question embedding feature vector, the sixth V vector is determined based on the ability feature vector, and the sixth K vector is determined based on the knowledge space feature vector.

[0166] A multi-head attention operation is performed based on the sixth Q vector, the sixth K vector, and the sixth V vector to obtain the sixth intermediate vector, and the sixth intermediate vector is input to the multilayer perceptron.

[0167] The multilayer perceptron performs multilayer perceptron operation based on the sixth intermediate vector to obtain post-operation features, and determines the prediction question response corresponding to the question to be predicted based on the post-operation features.

[0168] Based on the same concept as the above method, this application proposes an electronic device, see [link to previous application]. Figure 8As shown, it includes: a processor 81 and a machine-readable storage medium 82, wherein the machine-readable storage medium 82 stores machine-executable instructions that can be executed by the processor 81; the processor 81 is used to execute the machine-executable instructions to implement the knowledge point mastery prediction method disclosed in the above example of this application.

[0169] Based on the same application concept as the above method, this application embodiment also provides a machine-readable storage medium storing a plurality of computer instructions, which, when executed by a processor, can implement the knowledge point mastery prediction method disclosed in the above examples of this application.

[0170] The aforementioned machine-readable storage medium can be any electronic, magnetic, optical, or other physical storage device that can contain or store information, such as executable instructions, data, etc. For example, machine-readable storage media can be: RAM (Random Access Memory), volatile memory, non-volatile memory, flash memory, storage drives (such as hard disk drives), solid-state drives, any type of storage disk (such as optical discs, DVDs, etc.), or similar storage media, or combinations thereof.

[0171] Based on the same concept as the methods described above, this application also provides a computer program product, which may include a computer program. When executed by a processor, the computer program implements the knowledge point mastery prediction method disclosed in the examples above.

[0172] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, embodiments of this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0173] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for predicting the mastery of knowledge points, characterized in that, The method includes: The process involves obtaining a knowledge point description sequence, a question sequence for the target object, and an auxiliary information sequence. The knowledge point description sequence includes semantic description information corresponding to each candidate knowledge point name. The auxiliary information sequence includes the knowledge point name, question difficulty, and question response corresponding to each question in the question sequence. Determine the question embedding feature vector corresponding to the question sequence, the auxiliary embedding feature vector corresponding to the auxiliary information sequence, and the knowledge point embedding feature vector corresponding to the knowledge point description sequence; Multi-head attention operation is performed based on the question embedding feature vector and the auxiliary embedding feature vector to obtain the target object’s ability feature vector for the knowledge point name corresponding to the question sequence; The knowledge space feature vector is determined based on the knowledge point embedding feature vector. The knowledge point embedding feature vector includes an initial feature vector for each candidate knowledge point name. The knowledge space feature vector includes multiple fused feature vectors, and each fused feature vector is obtained by fusing multiple initial feature vectors. Multi-head attention operation is performed based on the capability feature vector, the knowledge point embedding feature vector and the knowledge space feature vector to obtain the knowledge point target feature vector; Based on the target feature vector of the knowledge point and the acquired knowledge point graph, predict the target object's mastery of the target knowledge point for each candidate knowledge point name in the knowledge point graph.

2. The method according to claim 1, characterized in that, Obtain the sequence of knowledge point descriptions, including: Obtain a knowledge point sequence, a question text sequence, and a reference matrix; wherein, the knowledge point sequence includes multiple candidate knowledge point names, the question text sequence includes multiple example question texts, and the reference matrix represents the correspondence between candidate knowledge point names and example question texts; The knowledge point sequence, the question text sequence, and the reference matrix are input into a large language model. The large language model determines the semantic description information corresponding to each candidate knowledge point name, thus obtaining the knowledge point description sequence. Specifically, for each candidate knowledge point name, the reference matrix determines the example question text corresponding to that candidate knowledge point name. Based on the example question text, semantic analysis is performed on the candidate knowledge point name to obtain the semantic description information corresponding to that candidate knowledge point name.

3. The method according to claim 1, characterized in that, The multi-head attention operation based on the question embedding feature vector and the auxiliary embedding feature vector to obtain the target object's ability feature vector for the knowledge point names corresponding to the question sequence includes: The first Q vector, the first K vector, and the first V vector are determined based on the question embedding feature vector. The first intermediate vector is obtained by performing a multi-head attention operation based on the first Q vector, the first K vector, and the first V vector. The second Q vector, the second K vector, and the second V vector are determined based on the auxiliary embedded feature vector. A multi-head attention operation is then performed based on the second Q vector, the second K vector, and the second V vector to obtain the second intermediate vector. The third Q vector and the third K vector are determined based on the first intermediate vector, the third V vector is determined based on the second intermediate vector, and a multi-head attention operation is performed based on the third Q vector, the third K vector and the third V vector to obtain the third intermediate vector, which is used to represent the degree of mastery at the problem level. The fourth K vector is determined based on the first intermediate vector, the fourth V vector is determined based on the third intermediate vector, the fourth Q vector is determined based on the knowledge space feature vector, and the fourth intermediate vector is obtained by performing a multi-head attention operation based on the fourth Q vector, the fourth K vector, and the fourth V vector. The capability feature vector is determined based on the fourth intermediate vector.

4. The method according to claim 1, characterized in that, The step of determining the knowledge space feature vector based on the knowledge point embedding feature vector includes: Based on the knowledge point embedding feature vector, a fifth Q vector, a fifth K vector, and a fifth V vector are determined. A multi-head attention operation is performed on the fifth Q vector, the fifth K vector, and the fifth V vector to obtain a fifth intermediate vector, which is then input into a multilayer perceptron. The multilayer perceptron performs a multilayer perception operation based on the fifth intermediate vector to obtain the knowledge space feature vector.

5. The method according to claim 1, characterized in that, The step of predicting the target object's mastery of the target knowledge point for each candidate knowledge point name in the knowledge point graph based on the target feature vector of the knowledge point and the acquired knowledge point graph includes: Based on the knowledge point graph, a first graph feature and a second graph feature are determined. The first graph feature includes the initial feature vector of each candidate knowledge point name in the knowledge point graph, and the second graph feature includes the similarity of the initial feature vectors of two candidate knowledge point names in the knowledge point graph. Based on the knowledge point target feature vector and the first graph feature, a target graph feature is generated. The target graph feature includes a fusion feature vector of the name of each candidate knowledge point in the knowledge point graph. The fusion feature vector is obtained by concatenating the knowledge point target feature vector and the initial feature vector. Based on the target graph features, the first graph features, and the second graph features, predict the target object's mastery of the target knowledge points for each candidate knowledge point name in the knowledge point graph.

6. The method according to claim 5, characterized in that, The step of predicting the target object's mastery of target knowledge points for each candidate knowledge point name within the knowledge point graph, based on the target graph features, the first graph features, and the second graph features, includes: The target graph features are input into a multilayer perceptron, and the multilayer perceptron performs feature mapping based on the target graph features to obtain mapped graph features; wherein the feature dimension of the mapped graph features is consistent with the input feature dimension of the graph neural network; The mapped graph features, the first graph features, and the second graph features are input into the graph neural network. The graph neural network then makes predictions based on the mapped graph features, the first graph features, and the second graph features to obtain the mastery level of the target knowledge point.

7. The method according to any one of claims 1-6, characterized in that, The question sequence also includes questions to be predicted, and the auxiliary information sequence also includes the knowledge point name and question difficulty corresponding to the questions to be predicted; the method further includes: Based on the question embedding feature vector, the ability feature vector, and the knowledge space feature vector, the predicted question response corresponding to the question to be predicted is determined; wherein, the predicted question response indicates whether the target object's answer to the question to be predicted is correct or incorrect.

8. The method according to claim 7, characterized in that, The step of determining the predicted question response corresponding to the question to be predicted based on the question embedding feature vector, the ability feature vector, and the knowledge space feature vector includes: The sixth Q vector is determined based on the question embedding feature vector, the sixth V vector is determined based on the ability feature vector, and the sixth K vector is determined based on the knowledge space feature vector. A multi-head attention operation is performed based on the sixth Q vector, the sixth K vector, and the sixth V vector to obtain the sixth intermediate vector, and the sixth intermediate vector is input to the multilayer perceptron. The multilayer perceptron performs multilayer perceptron operation based on the sixth intermediate vector to obtain post-operation features, and determines the prediction question response corresponding to the question to be predicted based on the post-operation features.

9. A knowledge point mastery prediction device, characterized in that, The device includes: The acquisition module is used to acquire a knowledge point description sequence, a question sequence of the target object, and an auxiliary information sequence; the knowledge point description sequence includes semantic description information corresponding to each candidate knowledge point name; the auxiliary information sequence includes the knowledge point name, question difficulty, and question response corresponding to each question in the question sequence. The determining module is used to determine the question embedding feature vector corresponding to the question sequence, the auxiliary embedding feature vector corresponding to the auxiliary information sequence, and the knowledge point embedding feature vector corresponding to the knowledge point description sequence; and to determine the knowledge space feature vector based on the knowledge point embedding feature vector, wherein the knowledge point embedding feature vector includes an initial feature vector for each candidate knowledge point name, and the knowledge space feature vector includes multiple fused feature vectors, and each fused feature vector is obtained by fusing multiple initial feature vectors. The processing module is used to perform multi-head attention operations based on the question embedding feature vector and the auxiliary embedding feature vector to obtain the target object's ability feature vector for the knowledge point name corresponding to the question sequence; and to perform multi-head attention operations based on the ability feature vector, the knowledge point embedding feature vector and the knowledge space feature vector to obtain the knowledge point target feature vector. The prediction module is used to predict the target object's mastery of the target knowledge point for each candidate knowledge point name in the knowledge point graph, based on the target feature vector of the knowledge point and the acquired knowledge point graph.

10. The apparatus according to claim 9, characterized in that, The processing module performs a multi-head attention operation based on the question embedding feature vector and the auxiliary embedding feature vector to obtain the target object's ability feature vector for the knowledge point name corresponding to the question sequence. Specifically, this is used for: The first Q vector, the first K vector, and the first V vector are determined based on the question embedding feature vector. The first intermediate vector is obtained by performing a multi-head attention operation based on the first Q vector, the first K vector, and the first V vector. The second Q vector, the second K vector, and the second V vector are determined based on the auxiliary embedded feature vector. A multi-head attention operation is then performed based on the second Q vector, the second K vector, and the second V vector to obtain the second intermediate vector. The third Q vector and the third K vector are determined based on the first intermediate vector, the third V vector is determined based on the second intermediate vector, and a multi-head attention operation is performed based on the third Q vector, the third K vector and the third V vector to obtain the third intermediate vector, which is used to represent the degree of mastery at the problem level. The fourth K vector is determined based on the first intermediate vector, the fourth V vector is determined based on the third intermediate vector, the fourth Q vector is determined based on the knowledge space feature vector, and the fourth intermediate vector is obtained by performing a multi-head attention operation based on the fourth Q vector, the fourth K vector, and the fourth V vector. The capability feature vector is determined based on the fourth intermediate vector.

11. The apparatus according to claim 9, characterized in that, The prediction module, based on the target feature vector of the knowledge point and the acquired knowledge point graph, specifically predicts the target object's mastery of the target knowledge point for each candidate knowledge point name in the knowledge point graph when it is used for: Based on the knowledge point graph, a first graph feature and a second graph feature are determined. The first graph feature includes the initial feature vector of each candidate knowledge point name in the knowledge point graph, and the second graph feature includes the similarity of the initial feature vectors of two candidate knowledge point names in the knowledge point graph. Based on the knowledge point target feature vector and the first graph feature, a target graph feature is generated. The target graph feature includes a fusion feature vector of the name of each candidate knowledge point in the knowledge point graph. The fusion feature vector is obtained by concatenating the knowledge point target feature vector and the initial feature vector. Based on the target graph features, the first graph features, and the second graph features, predict the target object's mastery of the target knowledge points for each candidate knowledge point name in the knowledge point graph.

12. The apparatus according to any one of claims 9-11, characterized in that, The question sequence also includes questions to be predicted, and the auxiliary information sequence includes the knowledge point names and question difficulty corresponding to the questions to be predicted; the prediction module is further used for: Based on the question embedding feature vector, the ability feature vector, and the knowledge space feature vector, the predicted question response corresponding to the question to be predicted is determined; wherein, the predicted question response indicates whether the target object's answer to the question to be predicted is correct or incorrect.

13. An electronic device, characterized in that, include: A processor and a machine-readable storage medium, the machine-readable storage medium storing machine-executable instructions that can be executed by the processor; The processor is configured to execute machine-executable instructions to implement the method of any one of claims 1-8.

Images