A multi-dimensional student state knowledge tracking method fusing emotion and forgetting mechanism

Abstract

The application discloses a multi-dimensional student state knowledge tracking method fusing emotion and forgetting mechanism, relates to the field of intelligent education, and aims at the problems that the existing method ignores the dynamic influence of emotion on forgetting and lacks a learning feedback mechanism.The application proposes a multi-dimensional state collaborative evolution framework.The method firstly constructs an independent forgetting state which is explicitly regulated by emotion and establishes a knowledge state which is inhibited by the forgetting state;secondly, the knowledge learning gain is innovatively introduced into the emotion updating process, realizing deep coupling of the three-dimensional states of knowledge, emotion and forgetting;finally, the states are fused through cross-dimensional attention, and the basic prediction probability is dynamically calibrated by using subjective difficulty and a prediction uncertainty coefficient.The application effectively simulates a real cognitive psychological process, and significantly improves the accuracy, explainability and individualization level of knowledge state tracking.

Description

Technical Field

[0001] This invention relates to the fields of smart education and knowledge tracking technology, and in particular to a multi-dimensional student status knowledge tracking method that integrates emotion and forgetting mechanisms. Background Technology

[0002] Knowledge tracking is a key technology in the field of smart education. It dynamically models students' mastery of knowledge points by analyzing their historical learning behavior data and predicts their test-taking performance. Accurate knowledge tracking provides crucial information for intelligent tutoring and personalized learning path recommendations, which is significant for achieving individualized instruction, optimizing the allocation of teaching resources, and improving the quality of education. Therefore, accurate modeling of students' knowledge status is an indispensable part of smart education technology research.

[0003] Currently, knowledge tracing modeling is primarily based on deep learning technology. Existing deep knowledge tracing methods use Long Short-Term Memory (LSTM) networks and Recurrent Neural Networks (RNNs) to model students' answer sequences. Subsequently, researchers introduced attention mechanisms, graph neural networks, and other techniques to improve knowledge tracing from multiple dimensions, such as the difficulty of the questions and the relevance of knowledge points. These methods have improved prediction accuracy to some extent, but they mainly focus on the attributes of the questions themselves, lacking in-depth exploration of students' personalized behavioral characteristics, resulting in inaccurate knowledge state modeling and insufficient interpretability of prediction results. Some researchers have found that students' emotional states and forgetting states affect their learning efficiency and knowledge absorption capacity, so they study the forgetting process or emotional information. However, this mainly involves modeling a single dimension and cannot comprehensively capture students' learning states. Therefore, how to effectively integrate students' emotional information and forgetting patterns in knowledge tracing to achieve a more comprehensive and refined model of the student learning process remains a focus of attention. Summary of the Invention

[0004] The purpose of this invention is to propose a multi-dimensional student state knowledge tracking method that integrates emotion and forgetting mechanisms, in order to solve the problems of single-dimensional modeling of student learning state, insufficient characterization of emotional changes and forgetting behavior, and low accuracy and stability of learning outcome prediction in existing technologies, thereby improving the modeling accuracy of students' knowledge mastery state and the reliability of learning outcome prediction.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a multi-dimensional student state knowledge tracking method integrating emotion and forgetting mechanisms, comprising the following steps:

[0006] 1. Construction of learning interaction features

[0007] First, students' continuous answering behavior to exercises is constructed as a learning interaction sequence. From the student interaction sequence, exercise characteristics, answer results, and time intervals are extracted. With review frequency An input feature vector reflecting the forgetting pattern is constructed through nonlinear function mapping. and The student's emotional state information is encoded to generate a continuous representation of emotional observation features. .

[0008] 2. Update of the forgetting state under emotion regulation

[0009] Constructing independently evolving forgetting state vectors This is used to characterize the degree to which students retain existing knowledge. The forgetting state is determined at each time step based on the historical forgetting state. Current forgetting characteristics and emotional state at the previous moment. The update process can be summarized as follows:

[0010]

[0011] in Calculate the basic forgetting decay ; It is the Sigmoid activation function. For element-wise multiplication; , For learnable weight matrix, , It is the bias vector; This is the preset emotional sensitivity coefficient.

[0012] 3. Knowledge status update inhibited by forgetting

[0013] In the state of knowledge During the update phase, forgetting inhibition mechanisms and emotion regulation mechanisms are introduced to jointly control the process of writing new knowledge.

[0014] First, the basic learning gain is calculated based on the current learning interaction. :

[0015]

[0016] in, Embed vectors for the current exercise. Embedding vectors for knowledge points Embed vectors for the answer results;

[0017] Subsequently, the learning gain was modulated by incorporating emotional and forgetting states to obtain the final knowledge gain. :

[0018]

[0019] in This serves as a gating mechanism for learning efficiency that is regulated by emotions.

[0020] And complete the knowledge state based on the forgetting state. renew:

[0021]

[0022] in, , This is the weight matrix. , It is the bias vector; This is the gain suppression coefficient; It is the hyperbolic tangent activation function.

[0023] 4. Dynamic Evolution of Emotional States Based on Learning Feedback

[0024] Design emotional states The update process not only relies on external observations, but is also regulated by real-time feedback on learning outcomes.

[0025] First, consider historical emotional states. Current characteristics of exercises Current sentiment observation information And the knowledge gains from this learning Calculate candidate emotional states :

[0026]

[0027] Subsequently, the mood update gating coefficient was calculated. :

[0028]

[0029] Ultimately, the emotional state is updated through a gating mechanism:

[0030]

[0031] in, , Weight matrix, , is the bias vector. Where, the gating coefficients are... This mechanism is used to balance the continuity and variability of emotional states, so that learning success can generate positive emotional motivation, while learning setbacks may trigger negative emotional changes, thus forming a closed-loop feedback mechanism between learning effectiveness and emotional changes.

[0032] 5. Multidimensional student state fusion and uncertainty-modulated prediction

[0033] After the update , , Concatenate the sequences along the sequence dimension to construct a hybrid state matrix. :

[0034]

[0035] Will Input a multi-head self-attention network, calculate the mutual attention weight matrix among the three state dimensions using the attention mechanism, and output a comprehensive state representation enhanced by interaction. This indicates that the dynamic importance characteristics of each dimension of the state at the current time step are integrated:

[0036]

[0037] Based on a comprehensive assessment of the student's overall condition, the student's current learning ability is first evaluated, resulting in... :

[0038]

[0039] At the same time, based on students' abilities and the characteristics of the exercises, calculate students' subjective perception of difficulty in the current state. :

[0040]

[0041] in, Weight matrix, This is the bias vector. For smoothing terms, This represents the emotion regulation coefficient; the emotion regulation term is... .

[0042] Furthermore, based on the stability of student states and the ability-difficulty deviation, the prediction uncertainty coefficient is calculated. :

[0043]

[0044] in Indicators representing the stability of student status;

[0045] In the output prediction phase, based on the base prediction probability The prediction results are calibrated using the uncertainty coefficient to obtain the final prediction result:

[0046]

[0047] When uncertainty is high, the prediction results converge toward the neutral value to reduce the prediction risk of the model in situations where students' status fluctuates or information is insufficient.

[0048] in By using three-dimensional states ( , , ), global context The next exercise is embedded and concatenated, and then the fully connected layer is calculated.

[0049] Finally, a multi-task joint loss function is constructed for model training:

[0050]

[0051] in, For the binary cross-entropy loss of the target label after smoothing; The mean squared error loss, weighted by prediction confidence, is used to constrain the emotion state prediction task. This is a balancing factor used to adjust the impact of the two tasks on the total loss. The weights in the equation.

[0052] The beneficial effects of this invention are as follows:

[0053] This invention introduces a forgetting modeling method based on time intervals and review frequency, and creatively combines it with emotion regulation mechanisms to quantify the real learning phenomenon of "emotion influencing forgetting." This mechanism solves the problem that forgetting is only implicitly processed in existing knowledge tracking methods and cannot distinguish between natural decay and emotional interference. It updates knowledge status more closely to the actual human memory patterns, thereby significantly improving the accuracy of predicting long-term learning behavior.

[0054] This invention constructs a multidimensional state unit that co-evolves knowledge, emotion, and forgetting, and combines it with a cross-dimensional attention interaction mechanism to model the deep dependencies between different state dimensions. This design solves the problem of traditional models that only track the degree of knowledge mastery and ignore the interaction between psychological state and memory strength, avoids information loss caused by simple feature splicing, and improves the precision of students' comprehensive cognitive state modeling.

[0055] This invention establishes a prediction uncertainty quantification and adaptive calibration mechanism that combines learning state stability with subjective difficulty bias. This strategy addresses the problem of deep learning models making overconfident predictions during cold starts or when student states fluctuate abnormally, effectively reducing prediction bias in high-risk scenarios and improving robustness and reliability. Attached Figure Description

[0056] Figure 1 This is a flowchart of the method of the present invention.

[0057] Figure 2 This is an overall architecture diagram of the method described in this invention.

[0058] Figure 3 This is a schematic diagram of a multidimensional state co-evolution mechanism.

[0059] Figure 4 This is a flowchart for multi-task prediction and dynamic calibration. Detailed Implementation

[0060] The following describes a multi-dimensional student state knowledge tracking method that integrates emotion and forgetting mechanisms, proposed in this invention, with specific embodiments. This embodiment uses a specific implementation process as an example to describe the model structure, state update calculation, and prediction process, but does not constitute a limitation on the scope of protection of this invention.

[0061] This invention proposes a multi-dimensional student state knowledge tracking method that integrates emotion and forgetting mechanisms. This embodiment is implemented based on a deep learning framework, and the overall system architecture is as follows: Figure 1 As shown, it mainly includes student learning sequence construction, three-dimensional state coupling update, cross-dimensional state interaction, and multi-task prediction and calibration.

[0062] 1. Learning to construct interactive sequences

[0063] Obtain students' continuous responses to exercises from the assist2012 dataset and construct them into a learning interaction sequence. For any student, the learning interaction at time step t is represented as: ( , , , )

[0064] in, Indicates the current exercise identifier. This indicates the corresponding knowledge point identifier. This indicates the student's answer. Indicates the timestamp of the response.

[0065] Based on the learned interaction sequence, time interval features are extracted. Review frequency characteristics The forgotten feature vectors generated by nonlinear mapping are denoted as follows: and The calculation is as follows:

[0066]

[0067]

[0068] in, , and , These are learnable linear layer parameters;

[0069] Students' emotional state Discretize the encoding (d_m = 6 buckets) and map it to a sentiment observation vector.

[0070] In this embodiment, the learning sequence length T is set to 100. When the sequence length is less than T, zero padding is used; when it exceeds T, sliding window segmentation is performed.

[0071] 2. Three-dimensional state initialization

[0072] Initialize three types of state vectors for each student: knowledge state Emotional state and the state of forgetting .in, , , All are represented as d-dimensional vectors; in this embodiment, d is 64. At the initial time t=0, each state vector is initialized to a zero vector or a random vector.

[0073] 3. Forgotten State Update Calculation

[0074] At time step t, based on the forgotten state of the previous time step... and the current forgotten feature vector , First, calculate the basic forgetting decay:

[0075]

[0076] Subsequently, the emotional state of the previous moment is introduced. The controlled emotion regulation factor is multiplicatively adjusted for the basal forgetting decay to obtain the forgetting state at the current moment:

[0077]

[0078] in, It is the Sigmoid activation function. For element-wise multiplication; , For learnable weight matrix, It is the bias vector; This is the preset emotional sensitivity coefficient.

[0079] 4. Knowledge Status Update Calculation

[0080] First, we combine the basic interaction features: exercise embedding vectors. Knowledge point embedding vector , Answer result embedding vector Calculate the basic learning gain:

[0081]

[0082] Calculating emotion-regulated learning efficiency gating :

[0083]

[0084] When writing new knowledge into the state, explicitly introduce the current forget state. As a suppression term, the final learning gain is obtained:

[0085]

[0086] in, This represents the gain suppression coefficient. The knowledge state at the current moment is updated by combining the historical knowledge retention:

[0087]

[0088] in, , This is the weight matrix. , It is the bias vector; It is the hyperbolic tangent activation function;

[0089] 5. Emotional state update calculation

[0090] The emotional state of the previous moment Current Exercise Embedded External sentiment observation and the knowledge learning gain splicing and calculating candidate emotional states :

[0091]

[0092] And calculate the mood update gating

[0093]

[0094] Ultimately, emotional state Update as follows:

[0095]

[0096] in, , Weight matrix, , This is the bias vector.

[0097] 6. Cross-dimensional state interaction and prediction

[0098] After the update , , Stacking them according to the state dimension forms a three-dimensional state matrix:

[0099]

[0100] in, This represents a stacking operation in the state dimension.

[0101] Based on the three-dimensional state matrix The next step is to use the problem embedding vector as the query vector. and will Mapped to key vectors respectively Sum value vector The cross-dimensional state interaction results are calculated using a multi-head attention mechanism:

[0102]

[0103] The cross-dimensional attention output This represents the comprehensive response of different state dimensions to the next exercise in the current learning context. Based on this, a linear mapping and nonlinear transformation are applied to the attention output to obtain the student's comprehensive state representation at time step t:

[0104]

[0105] in, and These are the learnable mapping weight matrix and the bias term, respectively. This represents the activation function.

[0106] 7. Multi-task prediction and uncertainty calibration

[0107] Based on comprehensive state representation The student's ability assessment value is obtained by mapping. With state stability At the same time, the characteristics of the exercises are mapped to objective difficulty. Introducing emotion regulation items Calculate subjective difficulty :

[0108]

[0109] in, For smoothing terms, =0.5. This indicates that the stronger the ability, the lower the subjective difficulty; the stronger the negative emotions, the higher the subjective difficulty.

[0110] And further calculate the prediction uncertainty coefficient :

[0111]

[0112] Finally, the basic prediction probability Perform calibration:

[0113]

[0114] Output the final probability of answering the question. And predict the emotional state in the next moment in parallel:

[0115]

[0116] When uncertainty is extremely high ( The prediction results tend to be 0.5 (random guess), thus reducing the model's risk of hallucination.

[0117] During model training or online prediction, the student learning interaction sequence is processed step by step according to the above steps to dynamically update the student's knowledge state, emotional state, and forgetting state, and outputs predicted information on the student's future answer results and emotional changes. This embodiment verifies that the present invention can accurately track students' knowledge mastery, while dynamically sensing their emotional changes and forgetting risks, providing a reliable basis for personalized teaching.

Claims

1. A multi-dimensional student state knowledge tracking method integrating emotion and forgetting mechanisms, characterized in that, Includes the following steps: S1: Construct a learning interaction sequence by analyzing students' continuous responses to exercises, including exercise information, knowledge points, answer results and interaction time, and extract time intervals, review frequency and emotional features based on the sequence to construct a multi-dimensional input vector; S2: Construct an independently evolving forgetting state vector, the update of which is explicitly regulated by the emotional state of the previous moment; and establish a knowledge state update mechanism that is inhibited by the forgetting state. S3: Calculate the knowledge learning gain based on the learning interaction feedback at the current moment, and input the knowledge learning gain into the emotional state update process to achieve coupled update of the three-dimensional states of knowledge, emotion, and forgetting; S4: Apply cross-dimensional attention weighting to the updated 3D state to obtain a comprehensive state representation enhanced by interaction; S5: Perform multi-task prediction based on the comprehensive state representation, calculate subjective difficulty and prediction uncertainty coefficient, use the uncertainty coefficient to calibrate the basic prediction probability to output the final answer probability, and predict the emotion at the next moment.

2. The multi-dimensional student state knowledge tracking method integrating emotion and forgetting mechanisms as described in claim 1, characterized in that, The three-dimensional state coupling update process described in S2 and S3 includes the following sub-steps: Step A: Calculate the basal forgetting decay and introduce the emotional sensitivity coefficient. Use the emotional state of the previous moment to construct a moderating factor to multiplicatively correct the basal forgetting decay and obtain the independent forgetting state at the current moment. Step B: Calculate the basic learning gain and learning efficiency gating, introduce a gain suppression coefficient, use the current forgetting state to suppress the basic learning gain, and update the current knowledge state in combination with the amount of historical knowledge retained. Step C: Construct a multi-source emotional stimulus input containing the knowledge learning gain calculated in Step B, calculate the emotion update gating, and use the gating to update the current emotional state.

3. The multi-dimensional student state knowledge tracking method integrating emotion and forgetting mechanisms according to claim 2, characterized in that, The calculation steps for the independent forgetting state described in step A are as follows: Based on the previous moment's forgotten state Current forgotten feature vector , Calculate the basic forgetting decay : ； Then, the emotional state of the previous moment is introduced. The control adjustment factor is used to calculate the forgetting state at the current time. : ； in, This is a state of forgetting from the previous moment. It is the Sigmoid activation function. For element-wise multiplication; , For learnable weight matrix, It is the bias vector; This is the preset emotional sensitivity coefficient.

4. The multi-dimensional student state knowledge tracking method integrating emotion and forgetting mechanisms according to claim 2, characterized in that, The calculation steps for the knowledge state described in step B are as follows: First, embed the current problem into the vector. Knowledge point embedding vector Embedded vectors of answer results The data is spliced ​​together, and the knowledge learning gain is calculated through fully connected layers. : ； Calculating emotion-regulated learning efficiency gating : ； Introducing a state of forgetting The inhibition term for new knowledge acquisition is combined with learning efficiency gating to calculate the final learning gain. : ； Update the current knowledge status based on the amount of historical knowledge retained. : ； in, , This is the weight matrix. , It is the bias vector; This is the gain suppression coefficient; It is the hyperbolic tangent activation function.

5. The multi-dimensional student state knowledge tracking method integrating emotion and forgetting mechanisms according to claim 2, characterized in that, The calculation steps for the emotional state described in step C are as follows: The emotional state of the previous moment Current Exercise Embedded External sentiment observation and the knowledge learning gain splicing and calculating candidate emotional states With Emotion Update Gating : ； ； Finally, update your current emotional state. Time step update: ； in, , Weight matrix, , This is the bias vector.

6. The multi-dimensional student state knowledge tracking method integrating emotion and forgetting mechanisms according to claim 1, characterized in that, The multi-task prediction and calibration process described in S5 includes the following sub-steps: Step D: Based on student status Mapping yields student ability assessment values With state stability At the same time, the characteristics of the exercises are mapped to objective difficulty. ; Step E: Calculate subjective difficulty Taking into account the stability of the state and the deviation between subjective difficulty and student ability assessment values, the uncertainty coefficient of the prediction is quantified. ; Step F: Evaluate the basic prediction probability based on the prediction uncertainty coefficient Perform linear interpolation calibration and output the final answer probability. .

7. The multi-dimensional student state knowledge tracking method integrating emotion and forgetting mechanisms according to claim 6, characterized in that, The calculation formulas for each indicator in steps E and F are as follows: Introducing emotion regulation items Subjective difficulty calculation formula: ； Prediction uncertainty coefficient Calculation formula: ； In parallel, the three-dimensional states , , Global context The next exercise is embedded and concatenated, and the basic prediction probability is calculated through a fully connected layer. Then output the final probability of answering the question. : ； ； in, For smoothing terms, For adjustment coefficients, , In response to the weights and biases of the prediction layer, To take the absolute value.

8. The multi-dimensional student state knowledge tracking method integrating emotion and forgetting mechanisms according to claim 1, characterized in that, The method also constructs a multi-task joint loss function for model training: ； in, For the binary cross-entropy loss of the target label after smoothing; The mean squared error loss, weighted by prediction confidence, is used to constrain the emotion state prediction task. This is a balancing factor used to adjust the impact of the two tasks on the total loss. The weights in the equation.

9. A method for tracking multi-dimensional student state knowledge by integrating emotion and forgetting mechanisms as described in any one of claims 1-8, characterized in that, include: Feature construction module: Processes student exercise answer interaction logs and generates multidimensional embedded representations of forgetting and emotional features; State Evolution Module: Introduces emotion regulation and forgetting inhibition mechanisms to jointly update knowledge, emotions, and forgetting states; Cross-dimensional interaction module: Based on attention mechanism, it captures the deep dependencies between knowledge, emotion, and forgetting state; Prediction Feedback Module: Used to calculate subjective difficulty and uncertainty, and output the probability of answering questions and the prediction results of emotional state.

Images