Knowledge tracking method of test question characterization heterogeneous graph based on student emotion

Abstract

The invention relates to the field of education big data mining, graph neural network and student behavior modeling, and provides a knowledge tracking method of a test question characterization heterogeneous graph based on student emotion, the method adopts a heterogeneous graph neural network technology in the deep learning field to characterize multi-dimensional features of test questions, and emotional features generated by students in the learning process are converted into multi-dimensional features of the test questions into multi-dimensional features of the test questions. The method comprises the following steps of: performing modeling on a specific test question, such as concentration degree, confusion degree, boring feeling and contusion feeling, embedding an emotion feature into a specific test question, and modeling a cognitive structure of a learner by adopting a transform mechanism; finally, the test questions with the emotion features of the students and the cognitive structures of the students are fused into a traditional knowledge tracking model, and knowledge tracking and learner performance prediction for different learner groups are achieved. The method can scientifically and comprehensively predict the learning condition of the student, and achieves the purpose of assisting the teacher in precise teaching.

Description

Technical Field

[0001] The present invention relates to the fields of educational big data mining, graph neural networks, and student behavior modeling, and particularly relates to a knowledge tracing method based on a heterogeneous graph of question representations with student emotions. Background Art

[0002] With the continuous development of artificial intelligence and educational big data technologies, precise teaching has received strong technical support. To meet the requirements of precise teaching, it is necessary to analyze the learning trajectories of students from their own perspectives and predict their future learning performances. The core goal of knowledge tracing is to simulate the change of students' knowledge states over time based on their historical answering records. Through the temporal modeling of students' knowledge mastery states, knowledge tracing can accurately track the degree to which students master knowledge points at a specific time point and predict their performances in the next learning interaction based on this. This process can not only reflect the current learning situations of students but also provide important references for educational decision-making and personalized teaching.

[0003] Existing knowledge tracing models can be divided into three categories: (1) probability models; (2) logic models; (3) deep learning-based models.

[0004] The probability models of knowledge tracing assume that the learning process follows a Markov process, and their results are statistically interpretable. A typical one is Bayesian Knowledge Tracing (BKT). BKT uses real-time feedback user interaction modeling and uses a hidden Markov model to model the learner's latent knowledge state as a set of binary variables, where each variable represents whether a certain knowledge skill is understood. However, BKT assumes that once a student masters a certain skill, they will never forget it, which does not conform to the actual learning situation.

[0005] Logic models are a class of models based on logical functions, and their core idea is to predict the probability of correct answers through the learning ability parameters of students and the relevant parameters of questions (such as difficulty and discrimination). Representative logic models include Probabilistic Matrix Factorization (PMF) and Knowledge Mastery Tracing (KPT). PMF is modeled through stage feedback and user interaction data, while KPT is an improvement on the basis of PMF and uses the prior information of users to more accurately track the knowledge proficiency of students.

[0006] Although probability models and logic models have good interpretability for modeling students' knowledge states and it is clear to know the impact of each parameter on the model performance, the learning and cognitive processes of students are affected by many micro factors, and it is difficult for probability models and logic models to fully capture such complex cognitive processes. Therefore, knowledge tracing methods based on deep learning are proposed, and representative ones are the DKT model and GKT. The DKT model first applies recurrent neural networks to the knowledge tracing task and uses the LSTM model to trace the process of students' knowledge proficiency changing dynamically over time. Graph-based Knowledge Tracing (GKT) uses several methods to transform the knowledge structure into a graph, where nodes correspond to skills and edges correspond to their relationships. These graphs are used as the input of the model to predict students' responses. In addition, most current methods focus on time or structural information, consider simple question sequences or interaction sequences, and ignore the importance of students' emotional factors in the learning process. Summary of the Invention

[0007] The purpose of the present invention is to overcome the deficiencies in the above-mentioned prior art and provide a knowledge tracing method based on a heterogeneous graph of test question representation considering students' emotions. By comprehensively using technical methods such as graph neural networks, graph convolution, and Transformer, it systematically deeply explores students' learning behavior patterns, can scientifically and comprehensively predict students' learning situations, and assist teachers in precise teaching.

[0008] The purpose of the present invention is achieved by the following technical measures.

[0009] A knowledge tracing method based on a heterogeneous graph of test question representation considering students' emotions includes the following steps:

[0010] (1) Embedding of students' emotional features. During the learning process, students will generate a series of emotional features, which will affect students' learning. Model these emotional features and embed them into specific test questions.

[0011] (2) Embedding of the heterogeneous graph. First, construct the main structure of the heterogeneous graph. The nodes of the heterogeneous graph include emotional nodes, test question nodes, and skill nodes, and these nodes are all obtained through embedding. Then, perform information transfer of heterogeneous nodes and information aggregation of neighbor nodes according to different edge relationships.

[0012] (3) Tracking the cognitive structure processing process. According to the result output in step (2), model the cognitive structure of the test question representation with emotional features and the student's answer interaction sequence, and then predict the student's response according to the student's current cognitive structure and the test question at the next moment.

[0013] In the above technical solution, the embedding of students' emotional features in step (1) is specifically as follows:

[0014] (1-1) Model based on the emotional state of students during the learning process, such as boredom, confusion, concentration, and frustration. Linearly transform multiple variables with emotional characteristics to obtain the potential emotional state of students.

[0015] (1-2) The emotional state of students not only considers their own potential emotions, but also takes into account the content and difficulty of the test questions themselves. Therefore, concatenate the potential emotional state of students and the difficulty of the test questions and then perform a linear transformation to obtain the final emotional characteristics of students, and use the emotional characteristics as emotional nodes and embed them into the heterogeneous graph.

[0016] In the above technical solution, the embedding of the heterogeneous graph in step (2) is specifically as follows:

[0017] (2-1) Construct a heterogeneous graph

[0018] Definition of heterogeneous graph: A heterogeneous graph (Heterogeneous Graph) is a graph structure that contains multiple types of nodes and multiple types of edges, and is used to describe complex relationships and interactions.

[0019] The node types are: Question Node represents each test question task or learning activity during the learning process; Skill Node represents the skills included in the test questions; Emotion Node represents the emotional state (such as boredom, confusion, frustration, etc.) generated by students during the learning process.

[0020] Definition of edge relationship types: Question-Skill Edge represents the relationship between a test question task and a certain skill; Emotion-Question Edge represents the emotional state generated by students when performing a certain test question task, and this edge relationship reflects how the emotional state affects learning behavior; Question-Question Edge represents the order in which students complete a series of test questions.

[0021] (2-2) Transmit heterogeneous node messages

[0022] According to the meta-relationship type, use Transformer to calculate the mutual attention between the three types of nodes, map the question node to a query query vector, map the skill node and the emotion node to the key key vector and the value vector, and calculate the dot product of the query and the key as the attention between nodes; at the same time, create a weight matrix W for each edge type, which can capture different semantic relationships, and finally multiply the attention weights and the value vector passed through the weight matrix W for information transmission between nodes.

[0023] (2-3) Aggregate neighbor node information

[0024] First, calculate the importance of neighbor nodes (skill nodes and sentiment nodes) for the test questions, that is, the attention weights. Generate the transmitted messages from the neighbor nodes. According to the attention weights, aggregate the messages of the neighbor nodes to the test question nodes to generate higher-order test question representations with sentiment and skill features.

[0025] In the above technical solution, the specific process of tracking the cognitive structure processing in step (3) is as follows:

[0026] (3-1) According to the result obtained in step (2), through the heterogeneous graph convolution operation, output the test question representation with sentiment features, and model the cognitive structure of the student using the transformer mechanism.

[0027] (3-2) According to the cognitive structure of the student at each moment t and the test questions at t+1, through a linear layer, predict the student's answering response at the next moment t+1.

[0028] The knowledge tracking method based on the test question representation heterogeneous graph of students' emotions in the present invention uses the heterogeneous graph neural network technology in the deep learning field to represent the multi-dimensional features of test questions, models the emotion features generated by students during the learning process, such as concentration, confusion, boredom, and frustration, and embeds the emotion features into specific test questions. At the same time, the transformer mechanism is used to model the cognitive structure of the learner; finally, the test questions with students' emotion features and the cognitive structure of the students are integrated into the traditional knowledge tracking model to realize knowledge tracking for different learner groups and prediction of learner performance. The present invention can scientifically and comprehensively predict the learning situation of students, achieving the purpose of assisting teachers in precise teaching. Description of the Drawings

[0029] Figure 1 It is a flowchart of the method of the embodiment of the present invention.

[0030] Figure 2 It is a student emotion embedding graph in the embodiment of the present invention.

[0031] Figure 3 It is an example diagram of the structure of the heterogeneous graph.

[0032] Figure 4 It is a computational graph of the heterogeneous information transmission and update process in the embodiment of the present invention.

[0033] Figure 5 It is a cognitive structure processing graph in the embodiment of the present invention. Detailed Embodiments

[0034] To make the objectives, technical solutions, and advantages of the present invention clearer, the following will further describe in detail the embodiments of the present invention with reference to the accompanying drawings.

[0035] As Figure 1 shown, an embodiment of the present invention provides a knowledge tracing method for a heterogeneous graph of question representations based on students' emotions, including the following steps:

[0036] (1) Embedding of students' emotional characteristics

[0037] According to the emotional states of students during the learning process, four emotional labels are selected for modeling, namely boredom, concentration, confusion, and frustration. Specifically, as Figure 2 shown:

[0038] Considering that after students complete the questions, their emotional states are affected by two factors, namely emotional latent traits and the complexity of the questions: Emotional latent traits are the personal emotional tendencies of students, such as concentration, confusion, etc.; the complexity of the questions will also affect students' emotions. For example, students may feel confused or anxious when facing questions with higher difficulty.

[0039] First, a multi-layer perceptron is used to linearly combine these four emotional characteristics into students' latent emotional characteristics.

[0040] Secondly, according to the relationship between questions and skills, a learnable parameter matrix is created to represent the difficulty of the questions. The students' latent emotional characteristics and the difficulty of the questions are concatenated and mapped to the same dimension as the question nodes and skill nodes to obtain the embedding of the students' emotional characteristics.

[0041] (2) Embedding of heterogeneous graphs

[0042] (2-1) Construction of heterogeneous graphs

[0043] First, a heterogeneous graph G=(V, E, A V , R E ) is defined, where V represents various types of nodes, E represents the edges connecting different nodes, and A V and R E represent the set of node types and the set of edge relation types.

[0044] The edge relations are defined as three types, namely edge e 1 =(S, Q), from the skill node s to the question node q, and its meta-relation is <s, r 1 , e>, question -- contains -- skill; edge e 2 =(E, Q), indicating that this edge is from the emotion node E to the question node Q, and its meta-relation is <e, r 2 , q>, emotion -- affects -- question; edge e3 =(Q, Q), where the test question node Q to the test question node Q represents the student's answering order, and its meta-relationship is <q, r 3 , q>, test question - order - test question.

[0045] For example, the structure of the heterogeneous graph is as Figure 3 shown. The problem is specifically represented as follows: Given a student's answering sequence, the set of test questions is Q, the set of skills is S, and the combination of student emotions is E. The test questions and skills are embedded through Embedding, and the student node is obtained through the student emotion feature embedding described below. Each test question q contains a specific skill s, and each test question q is affected by a specific emotion e. The nodes and edges of the heterogeneous graph both adopt a dictionary data structure.

[0046] (2 - 2) Transmitting heterogeneous node information

[0047] Adopting the idea of the transformer, through different mappings of the meta-relationship, mutual attention between heterogeneous nodes is obtained, and then information is transmitted through a weight matrix.

[0048] For example, as Figure 4 shown, the figure involves two meta-relations, namely <s, r 1 , q> and <a, r 2 , q>. First, linearly transform the features of the three types of nodes to map them to the same dimension d. The goal of this method is to generate a higher-order test question representation. Therefore, the test question node q to be learned is mapped to a query query vector. The skill node s and the student emotion node a are respectively mapped to a key vector and a value vector.

[0049] Then, create two edge weight matrices w edge1 r1 , w edge2 r2 for message passing to share information between nodes. Using the two weight matrices, calculate the similarity between the query query vector and the key vector. For each test question node q, sum all the attention vectors from its neighbor nodes and perform softmax scaling. The nodes and edges in the heterogeneous graph have different types, and message passing and aggregation need to consider heterogeneity to alleviate the problems brought by the distribution differences of different types of nodes and edges. For the value vectors of skills and emotions, use the same method, and then create two more edge weight matrices w edge3 r1 , w edge4 r2 to capture the influence of edge types on nodes, and transmit the information of skill nodes and emotion nodes to test question nodes, adopting a multi-head message passing mechanism to increase the expressive power of the model.

[0050] (2-3) Aggregate neighbor node information

[0051] The aggregation operation is performed based on all neighbor nodes of the question node q. Information is obtained from the neighbor nodes to update the feature representation of the question node.

[0052] As Figure 4 shown, the attention weights after softmax dot product scaling are multiplied by the value vectors after message passing, and the weighted messages of each neighbor node are aggregated in a concatenated manner. The feature representation of the question node obtained through aggregation is H (l) [q] , which contains the information passed by all neighbor nodes. After the aggregation is completed, the updated representation H (l) [q] of the question node needs to be mapped back to a specific type distribution, introducing a non-linear activation function to enhance the representation ability. Finally, a residual connection is used to add the H (l-1) [q] from the previous layer to the updated representation of the current layer to avoid information loss and alleviate the problem of gradient disappearance.

[0053] (3) Cognitive structure processing process

[0054] After obtaining the question representation with sentiment features, the cognitive structure of the student is modeled. As Figure 5 shown: Using the transformer mechanism, the self-attention mechanism is used for the student's question sequence and interaction sequence respectively. Here, the student's question sequence is already the question representation with sentiment features after graph convolution. Then, the attention mechanism is used to model the student's cognitive structure. After the above process, the cognitive structure sequence {H 1 ,H 2,…… H T} of the student at each moment is obtained. Combining the hidden state H T at each moment with the question representation vector q t+1 at the next moment T+1, the prediction result r t+1 at the T+1 moment can be obtained.

[0055] In the prediction layer, the current cognitive structure h t of the student and the question q t+1 at the next moment are concatenated and linearly transformed. Then, the relu function is used to increase the non-linear features of the model. An appropriate dropout operation is performed in the second layer, and the sigmoid activation function is used. After the calculations of these two layers, the final prediction result r t+1 is obtained, which represents the probability that the learner answers the question q t+1 successfully at the T+1 moment.

[0056] The content not described in detail in this specification belongs to the prior art well-known to those skilled in the art.

[0057] The above are only the preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. A knowledge tracking method based on a heterogeneous graph of test question representation based on student emotions, characterized by The method comprises the following steps: (1) Embedding students’ emotional characteristics: Students will generate a series of emotional characteristics during the learning process, which will affect their learning. Emotional characteristics are modeled and embedded into specific test questions. (2) Embedding of heterogeneous graphs. First, the main structure of the heterogeneous graph is constructed. The nodes of the heterogeneous graph include emotion nodes, question nodes, and skill nodes. These nodes are obtained through embedding. Then, information transmission of heterogeneous nodes and information aggregation of neighboring nodes are performed based on different edge relationships. (3) Tracking the cognitive structure processing process: Based on the output of step (2), the test question representation with emotional characteristics and the student's answer interaction sequence are used to model the student's cognitive structure, and then the student's answer response is predicted based on the student's current cognitive structure and the test question at the next moment.

2. The knowledge tracking method based on the heterogeneous graph of test question representation based on student emotions according to claim 1 is characterized in that The embedding of student emotion features in step (1) specifically includes: (1-1) Modeling is done based on the students’ emotional states during the learning process. The emotional states include concentration, confusion, boredom, and frustration. Multiple variables with emotional characteristics are linearly transformed to obtain the students’ potential emotional states. (1-2) The students’ potential emotional state and the difficulty of the test questions are concatenated and then linearly transformed to obtain the students’ final emotional characteristics. The emotional characteristics are then used as emotional nodes and embedded into the heterogeneous graph.

3. The knowledge tracking method based on the heterogeneous graph of test question representation based on student emotions according to claim 1 is characterized in that The embedding of the heterogeneous graph in step (2) specifically includes: (2-1) Constructing a heterogeneous graph The node types are: question node, skill node, and emotion node. Question node represents each question task or learning activity in the learning process. Skill node represents the skills contained in the question. Emotion node represents the emotional state of students in the learning process. Definition of edge relationship types: The test-question relationship edge represents the relationship between a test task and a skill; the emotion-question relationship edge represents the emotional state of students when performing a test task. This edge relationship reflects how the emotional state affects learning behavior; the test-question relationship edge represents the order in which students complete a series of test questions; (2-2) Transmitting heterogeneous node information According to the meta-relation type, Transformer is used to calculate the mutual attention between the three nodes, map the test question node to a query vector, map the skill node and the emotion node to the key vector and the value vector, and calculate the dot product of the query and the key as the attention between the nodes; at the same time, a weight matrix W is created for each edge type to capture different semantic relationships, and finally the attention weights passed by the weight matrix W are multiplied by the value vector to transfer information between nodes; (2-3) Aggregate neighbor node information First, the importance of neighbor nodes, namely skill nodes and emotion nodes, to the test questions, namely the attention weight, is calculated, and the transmitted messages are generated from the neighbor nodes. According to the attention weight, the messages of the neighbor nodes are aggregated to the test question nodes to generate a higher-order test question representation with emotion and skill characteristics.

4. The knowledge tracking method based on the heterogeneous graph of test question representation based on student emotions according to claim 1 is characterized in that The cognitive structure processing tracking in step (3) specifically includes: (3-1) Based on the results obtained in step (2), the test question representation with emotional features is output through heterogeneous graph convolution operation, and the student's cognitive structure is modeled using the transformer mechanism; (3-2) Based on the students’ cognitive structure at each time point t and the test questions at t+1, a linear layer is used to predict the students’ responses at the next time point t+1.

Images