Knowledge tracking method and device based on dual-channel difficulty perception

Abstract

The invention discloses a knowledge tracking method and device based on dual-channel difficulty perception. The method comprises the following steps: firstly, obtaining the statistical difficulty and model evaluation difficulty of a target topic; then, difficulty fusion is carried out on the statistical difficulty and the model evaluation difficulty based on a multi-head attention mechanism; and finally, predicting the accuracy of the target question in the current time step through the knowledge tracking model. According to the method, a multi-dimensional difficulty assessment system is constructed, the model assessment difficulty, the statistical difficulty and the knowledge state of the student are combined, the objectivity and reliability of difficulty assessment are improved, and the learning track and performance can be predicted more accurately. By integrating the dynamic change characteristics of the difficulty perception deviation sequence and the difficulty mastering ratio, the dynamic difficulty adaptability index can accurately describe the adaptability and learning characteristics of students facing different difficulty questions, and the knowledge tracking model can better capture the ability improvement and the adaptability change of the students in the learning process.

Description

Technical Field

[0001] The present invention relates to the technical field of knowledge tracing, and more specifically, to a knowledge tracing method and device based on dual-channel difficulty perception. Background Art

[0002] Knowledge tracing (KT) has become an indispensable part of intelligent tutoring systems, using students' historical response data to evaluate their knowledge mastery and automatically predict their future performance and knowledge states. In the field of education, knowledge tracing can optimize students' learning paths and provide guidance for teachers, enabling teaching strategies and personalized teaching to be improved targeted. By dynamically tracking students' knowledge states, this technology effectively reduces teachers' workload while improving teaching quality. With the rapid development of online education, a large amount of learning behavior data has been recorded, including not only the correctness of students' answers, but also multi-dimensional information such as response time and learning resource utilization patterns. The richness of this data has laid a solid foundation for advancing knowledge tracing technology.

[0003] Currently, the research on knowledge tracing mainly includes three models: Bayesian knowledge tracing, dynamic key-value memory neural network, and deep knowledge tracing. Among them, Bayesian knowledge tracing uses Bayesian networks to model the changes in students' knowledge states during the learning process, but ignores the influence of the order of students' exercises and the correlation between exercises on the prediction results. Deep knowledge tracing is mainly implemented based on RNN and LSTM. Although the prediction accuracy has been improved, currently, deep knowledge tracing only tracks based on students' answer sequences and answer results, ignoring the relationship between exercise difficulty and students' own abilities, and unable to accurately reflect the learning characteristics of different students. Moreover, static difficulty assessment cannot capture the improvement of students' abilities and adaptive changes during the learning process. Summary of the Invention

[0004] In view of this, the present invention provides a knowledge tracing method and device based on dual-channel difficulty perception, which is used to solve the problems that existing knowledge tracing methods ignore the relationship between exercise difficulty and students' own abilities, cannot accurately reflect the learning characteristics of different students, and static difficulty assessment cannot capture the improvement of students' abilities and adaptive changes during the learning process.

[0005] To achieve the above object, the following solutions are proposed:

[0006] A knowledge tracing method based on dual-channel difficulty perception, comprising:

[0007] Obtain the statistical difficulty and model evaluation difficulty of the target question;

[0008] Based on the multi-head attention mechanism, fuse the statistical difficulty and the model evaluation difficulty to obtain a calibrated difficulty score;

[0009] Calculate the difficulty perception deviation sequence of the target question based on the calibrated difficulty score of the target question and the knowledge state of the student at the previous time step;

[0010] Determine the difficulty interval of the target question based on the statistical difficulty, and calculate the difficulty mastery ratio of the student according to the difficulty interval;

[0011] Integrate the dynamic change characteristics of the difficulty perception deviation sequence and the difficulty mastery ratio based on the self-attention mechanism to obtain the dynamic difficulty adaptability index;

[0012] Obtain the input data of the knowledge tracing model based on the dynamic difficulty adaptability index and the knowledge state of the student at the previous time step, and predict the correct rate of the target question at the current time step through the knowledge tracing model.

[0013] Preferably, obtaining the statistical difficulty and the model evaluation difficulty of the target question includes:

[0014] Perform difficulty evaluation on the input question information of the target question through a pre-trained difficulty evaluation model to obtain the model evaluation difficulty of the target question;

[0015] Statistically analyze the historical performance data of the student to obtain the statistical difficulty of the target question.

[0016] Preferably, the process of the difficulty evaluation model for difficulty evaluation includes:

[0017] Obtain the question information of the target question input by the user;

[0018] Perform a structured step-by-step solution to the target question based on the question information to obtain a step-by-step solution plan;

[0019] Perform difficulty evaluation based on the step-by-step solution plan and the question information to obtain the model evaluation difficulty.

[0020] Preferably, the process of obtaining the input data of the knowledge tracing model based on the dynamic difficulty adaptability index and the knowledge state of the student at the previous time step includes:

[0021] Obtain the embedding vector of the dynamic difficulty adaptability index at the current time step;

[0022] Perform element-wise multiplication of the embedding vector and the knowledge state of the student at the previous time step to obtain the input embedding of the knowledge tracing model;

[0023] Perform calculations based on the input embedding, the answer sequence vector, and the knowledge concept vector to obtain the input data.

[0024] Preferably, the process of predicting the correct rate of the target question at the current time step through the knowledge tracing model includes:

[0025] Calculate the knowledge acquisition amount of the student at the current time step according to the input data;

[0026] Based on the knowledge acquisition amount and the knowledge state of the student at the previous time step, calculate to obtain the knowledge state of the student at the current time step;

[0027] Predict the probability of the student's correct answer at the next time step according to the knowledge state at the current time step and the knowledge representation embedding vector at the next time step.

[0028] Preferably, calculate the difficulty mastery ratio of the student through the following formula, and the formula is as follows:

[0029]

[0030] Among them, is the difficulty mastery ratio, σ is the forgetting factor, B k represents the k-th difficulty interval, represents the student's answer result to question i at time step t - 1, represents the number of questions in the k-th difficulty interval at time step t - 1, represents the correct rate in the k-th difficulty interval at time step t - 1.

[0031] Preferably, calculate the knowledge acquisition amount of the student at the current time step according to the following formula, and the formula is as follows:

[0032]

[0033] Among them, z t is the input data, is the initial knowledge acquisition value of the student, is the control output gating, SKG t represents the output knowledge acquisition amount.

[0034] Preferably, calculate the knowledge state of the student at the current time step according to the following formula, and the formula is as follows:

[0035]

[0036] Among them, k t-1 is the knowledge state of the student at the previous time step, k t is the knowledge state of the student at the current time step, is a gating mechanism for controlling the ratio of information passing through k t-1 and SKG t ddai t is the dynamic difficulty adaptability index sequence, is the answer sequence vector.

[0037] A knowledge tracking device based on dual-channel difficulty perception, comprising:

[0038] A difficulty acquisition unit for acquiring the statistical difficulty and model evaluation difficulty of a target question;

[0039] A difficulty calibration unit for fusing the statistical difficulty and model evaluation difficulty based on a multi-head attention mechanism to obtain a calibrated difficulty score;

[0040] A deviation calculation unit for calculating a difficulty perception deviation sequence of the target question based on the calibrated difficulty score of the target question and the knowledge state of the student at the previous time step;

[0041] A mastery ratio calculation unit for determining the difficulty interval of the target question based on the statistical difficulty and calculating the difficulty mastery ratio of the student according to the difficulty interval;

[0042] A dynamic index calculation unit for integrating the dynamic change characteristics of the difficulty perception deviation sequence and the difficulty mastery ratio based on a self-attention mechanism to obtain a dynamic difficulty adaptability index;

[0043] A prediction unit for obtaining the input data of the knowledge tracking model based on the dynamic difficulty adaptability index and the knowledge state of the student at the previous time step, and predicting the correct rate of the target question at the current time step through the knowledge tracking model.

[0044] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0045] (1) The knowledge tracking method based on dual-channel difficulty perception provided by the present invention first acquires the statistical difficulty and model evaluation difficulty of a target question; then, fuses the statistical difficulty and model evaluation difficulty based on a multi-head attention mechanism; finally, predicts the correct rate of the target question at the current time step through a knowledge tracking model. The present invention constructs a multi-dimensional difficulty evaluation system, combines the model evaluation difficulty, statistical difficulty with the knowledge state of the student, not only improves the objectivity and reliability of the difficulty evaluation, but also enables the knowledge tracking model to more accurately predict the learning trajectory and performance.

[0046] (2) The present invention calculates the difficulty perception deviation sequence of the target question based on the calibrated difficulty score of the target question and the knowledge state of the student at the previous time step; determines the difficulty interval of the target question based on the statistical difficulty, and calculates the difficulty mastery ratio of the student according to the difficulty interval; integrates the dynamic change characteristics of the difficulty perception deviation sequence and the difficulty mastery ratio based on the self-attention mechanism to obtain a dynamic difficulty adaptability index. By integrating the dynamic change characteristics of the difficulty perception deviation sequence and the difficulty mastery ratio, the dynamic difficulty adaptability index can accurately describe the adaptability and learning characteristics of the student when facing questions of different difficulties. Therefore, the knowledge tracking model can better capture the ability improvement and adaptability changes of the student during the learning process. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on the provided drawings without creative efforts.

[0048] Figure 1 Schematic diagram of a knowledge tracking method based on dual-channel difficulty perception provided by an embodiment of the present invention;

[0049] Figure 2 Schematic diagram of another knowledge tracking method based on dual-channel difficulty perception provided by an embodiment of the present invention;

[0050] Figure 3 Schematic diagram of a model data processing flow provided by an embodiment of the present invention;

[0051] Figure 4 Schematic diagram of the structure of a knowledge tracking device based on dual-channel difficulty perception provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0052] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0053] First, in combination with Figures 1-3 The knowledge tracking method based on dual-channel difficulty perception provided in this embodiment is introduced. As shown in the figure, the knowledge tracking method includes:

[0054] Step S01: Obtain the statistical difficulty and model evaluation difficulty of the target question.

[0055] Specifically, as Figure 2 shown, the difficulty evaluation of the present invention does not use a single dimension as the index for difficulty judgment, but combines the evaluation difficulty of the large language model with the evaluation difficulty of mathematical statistics to obtain a calibrated difficulty score in an objective dimension. Among them, through a pre-trained difficulty evaluation model (LLM), based on the question information of the input target question, the model evaluation difficulty of the target question is obtained; by performing mathematical statistics on the historical performance data of students, the statistical difficulty of the target question is obtained.

[0056] (I) Obtaining the model evaluation difficulty

[0057] As Figure 3 shown, in order to extract the implicit difficulty information from the question, a special prompt can be designed to enable the difficulty evaluation model to perform similar question retrieval and difficulty evaluation. Using the chain of thought method, gradually analyze different problem-solving stages, and finally determine the difficulty level of the question.

[0058] First, obtain the question information of the target question input by the user. Adopt a step-by-step solution prompt. This input prompt is used to guide the large language model to complete the problem-solving process, enabling the complexity of the solution to be directly observed. The input question information includes:

[0059] Question text (denoted as gc i ): The original or reconstructed content of the question;

[0060] Question answer (denoted as a i ): The standard answer to the question;

[0061] Brief analysis (denoted as e i ): The brief analysis of the question (if it exists in the dataset, such as the XES3G5M dataset);

[0062] Question type (denoted as pt i ): The type of the question (fill-in-the-blank or multiple-choice);

[0063] Options (only applicable to multiple-choice questions, denoted as opt i ): Include options if it is a multiple-choice question, otherwise empty;

[0064] Knowledge concept (denoted as kc i ): The knowledge points included in the question, including the gradual refinement of knowledge concepts.

[0065] Secondly, based on the problem text, problem answers, brief analysis, problem type, and options in the problem information, the target problem is solved step by step in a structured manner to obtain a step-by-step solution plan. After obtaining the problem information input by the user, the problem information is integrated into a structured step-by-step solution prompt, enabling the difficulty assessment model to generate clear problem-solving steps.

[0066] Finally, the structured problem-solving process not only helps to evaluate the cognitive complexity of the problem but also reflects the depth of knowledge and reasoning difficulty required to solve the problem. Therefore, after obtaining the step-by-step solution plan, the difficulty assessment model applies the difficulty assessment prompt to perform a difficulty assessment based on the step-by-step solution plan and the problem information, obtaining the model-assessed difficulty. The difficulty assessment model can evaluate the difficulty based on the problem text, brief analysis, problem type, knowledge concepts, and step-by-step solution plan. Among them, the step-by-step solution plan can be denoted as sol i , and the model-assessed difficulty is

[0067] Due to the adoption of the Chain of Thought technology, the model can analyze various aspects of the problem through an explicit reasoning process, including concept complexity, the number of problem-solving steps, the degree of knowledge point association, etc., thus obtaining a more accurate and interpretable difficulty assessment result. This evaluation method based on fine-grained analysis not only improves the accuracy of difficulty judgment but also provides specific difficulty attribution basis for educators. Compared with traditional statistics-based evaluation methods, this method can better capture the potential characteristics of problem difficulty, especially in dealing with new problems or situations lacking a large amount of historical data, showing significant advantages.

[0068] (2) Obtaining statistical difficulty

[0069] Statistical difficulty is another important indicator in the difficulty judgment of the present invention. Statistics are based on historical student performance data. The correct rate of all students for a certain problem can be used as the statistical difficulty indicator for that problem. The statistical difficulty of the target problem can be expressed by the following formula:

[0070]

[0071] Among them, is the statistical difficulty of the target problem i, r ij is the answer result of student j for the target problem i (1 for correct, 0 for wrong), N is the total number of students participating in the answer of the target problem i, is the correct rate of the target problem.

[0072] The higher the accuracy rate, the lower the difficulty; the lower the accuracy rate, the higher the difficulty. The range of the statistical difficulty value is between [0, 1], where 1 represents the easiest (all students answer correctly), and 0 represents the most difficult (all students answer wrongly).

[0073] Step S02, based on the multi-head attention mechanism, fuse the statistical difficulty and the model evaluation difficulty.

[0074] Specifically, if only the evaluation difficulty of the large model is used as the question difficulty index, due to the differences between student groups, the question difficulty index is hardly convincing; similarly, if only the mathematical statistics evaluation difficulty is used as the question difficulty index, due to the limited number of students who complete a certain question, the question presents an extremely high or extremely low difficulty value, and objectivity is often lacking. Therefore, the present invention combines the model evaluation difficulty and the statistical difficulty, specifically alleviates the cold start problem, and makes a more refined and objective construction for the implicit difficulty index in the question.

[0075] As Figure 3 shown, in order to make full use of the complementary advantages of the two methods, the multi-head attention mechanism can be used to fuse the statistical difficulty and the model evaluation difficulty to obtain the calibrated difficulty score. The multi-head attention mechanism can dynamically weigh different aspects of the two difficulty measurements, so as to obtain a more comprehensive and robust difficulty evaluation. The calibrated difficulty score is calculated by the following formula, and the formula is as follows:

[0076]

[0077] where is the calibrated difficulty score.

[0078] After calculating , it can be mapped to the target question i. During the training process, for each time step, there is representing the calibrated difficulty sequence corresponding to the target question at this time step.

[0079] Step S03, calculate the difficulty perception deviation sequence of the target question based on the calibrated difficulty score of the target question and the knowledge state of the students in the previous time step.

[0080] Specifically, as Figure 3 shown, during the learning process of students, the knowledge state can often be regarded as an embodiment of the students' subjective difficulty perception. When a student has a low mastery of a certain concept, their knowledge state in this concept is often low. Therefore, the present invention represents the subjective difficulty of students through the knowledge state. The difficulty perception deviation sequence at the current time step t is calculated by the following formula, and the formula is as follows:

[0081]

[0082] Among them, k t-1 represents the knowledge state of the student at time step t - 1.

[0083] The difficulty perception bias sequence reflects the deviation degree between the student's subjective difficulty perception and the objective difficulty of the question. When the difficulty perception bias sequence is close to zero, that is, dpbs t ≈0, it indicates that the question difficulty matches the student's knowledge level, representing the most ideal learning state; when the difficulty perception bias sequence is significantly positive, that is, dpbs t >>0, it indicates that the objective difficulty of the question is much higher than the student's current knowledge level; when the difficulty perception bias sequence is significantly negative, that is, dpbs t <<0, it indicates that the objective difficulty of the question is much lower than the student's knowledge level.

[0084] Step S04, determine the difficulty interval of the target question based on the statistical difficulty, and calculate the difficulty mastery ratio of the student according to the difficulty interval.

[0085] Specifically, in the process of student personalized modeling, the mastery ratios of students at different difficulty levels show unique patterns among individuals, which is of great significance for personalized modeling. To effectively capture these individual differences, the present invention quantifies the mastery degree of students at different difficulty levels through the Difficulty Mastery Ratio (DMR).

[0086] First, the difficulty range [0, 100] can be equally divided into five intervals, and each interval represents a different difficulty level. Based on the statistical difficulty each target question can be assigned to its corresponding difficulty interval, and the correct rate within each difficulty interval is calculated based on the following formula:

[0087]

[0088] Among them, represents the correct rate at time step t - 1 in the kth difficulty interval, Bk represents the kth difficulty interval, represents the answer result (1 for correct and 0 for wrong) of the student to question i at time step t - 1, represents the number of questions in the kth difficulty interval at time step t - 1.

[0089] As Figure 3 shown, for the student interaction data X i(t-1) , x i(t-1) is the interaction sequence of student i at time t - 1. The correct rate of each difficulty interval will be added to x i(t-1)Among them, by appending, the student interaction sequence changes from a triple to a quadruple. At the same time, considering the natural phenomenon of knowledge decay during the learning process, a forgetting factor is introduced to simulate the progressive forgetting process in students' learning. The difficulty mastery ratio of students is calculated through the following formula, and the formula is as follows:

[0090]

[0091] Among them, is the difficulty mastery ratio, and σ is the forgetting factor.

[0092] Step S05, integrate the dynamic change characteristics of the difficulty perception deviation sequence and the difficulty mastery ratio based on the self-attention mechanism.

[0093] Specifically, in order to more comprehensively evaluate the learning characteristics of students, the difficulty perception deviation sequence and the difficulty mastery ratio are fused to construct a dynamic difficulty adaptability index. The present invention uses TransformerEncoder to utilize the internal multi-head attention mechanism to capture the dynamic characteristics and long-term dependencies in the difficulty perception deviation sequence (DPBS) and the difficulty mastery ratio (DMR). The multi-head attention mechanism enables the model to fully understand the complex interaction patterns between the DPBS and DMR of students at different time points. TransformerEncoder takes the sequence data of DPBS and DMR as input, retains the timing information through position encoding, and at the same time uses the self-attention mechanism to learn the key patterns inside the sequence, thereby effectively integrating the dynamic change characteristics of students' difficulty perception and mastery degree. Specifically, for the sequence data of the difficulty mastery ratio of question i, it is calculated as:

[0094]

[0095] Among them, the statistical difficulty of question i is within the difficulty range (k, k + 20].

[0096] According to the difficulty mastery ratio sequence and the difficulty perception deviation sequence, the dynamic difficulty adaptability index is calculated through the following formula, and the formula is as follows:

[0097]

[0098] The Difficulty Perception Bias Sequence (DPBS) provides a comprehensive difficulty metric that combines objective assessment and subjective perception by integrating calibrated difficulty and students' subjective difficulty; while the Difficulty Mastery Ratio (DMR) provides an important basis for personalized modeling by characterizing students' mastery levels at different difficulty levels. By fusing, the complementary advantages of the two indicators are fully utilized. Among them, DPBS reflects the comprehensiveness of difficulty assessment, and DMR embodies the personalized characteristics of the learning process. By combining, the Dynamic Difficulty Adaptability Index (DDAI) can more accurately describe students' adaptability and learning characteristics when facing questions of different difficulties.

[0099] Step S06: Based on the dynamic difficulty adaptability index and the student's knowledge state at the previous time step, obtain the input data of the knowledge tracing model, and predict the correct rate of the target question at the current time step through the knowledge tracing model.

[0100] Specifically, after extracting the dynamic difficulty adaptability index sequence, first, calculate the input of the knowledge tracing model; then the knowledge tracing model makes a prediction based on the input data.

[0101] (I) Calculation of the input of the knowledge tracing model

[0102] First, obtain the embedding vector of the dynamic difficulty adaptability index within the current time step.

[0103] Then, perform element-wise multiplication of the embedding vector and the student's knowledge state at the previous time step to obtain the input embedding of the knowledge tracing model. The input embedding is calculated through the following formula:

[0104]

[0105] Among them, is the embedding vector of the dynamic difficulty adaptability index sequence within the current time step, and k t-1 is the student's knowledge state at the previous time step, and * represents element-wise multiplication of the two vectors. is the difficulty mastery ratio sequence of question i at the current time step t, and ddai t is the dynamic difficulty adaptability index sequence, representing the dynamic difficulty adaptability index sequence of questions 1 to n at the current time step t.

[0106] Element-wise multiplication enables the model to capture the interaction relationship between the current knowledge level and the question difficulty by multiplying the dynamic difficulty adaptability index and the knowledge state, thereby more accurately simulating the performance changes of students when facing questions of different difficulties. The model can dynamically adjust the weight of the question difficulty according to the student's knowledge state, better reflecting the personalized characteristics in the learning process.

[0107] Finally, after calculating the input embedding, it is combined with the embedding vector of the student's answer sequence (answer sequence vector) and the embedding vector of the question knowledge concept sequence (knowledge concept vector) to obtain the input data. The input data is calculated by the following formula:

[0108]

[0109] Where is the input answer sequence vector, is the knowledge concept vector, is the concatenation operation.

[0110] Use z t instead of relying solely on x t , to avoid the situation that simply relying on the interaction between difficulty assessment and knowledge state may not fully reflect the actual learning situation of students.

[0111] (2) Knowledge Tracing Model Prediction

[0112] First, calculate the knowledge acquisition of the student at the current time step based on the input data. The specific answering performance of the student contains important real-time feedback information. On the one hand, the student's actual answer reflects their current knowledge mastery; on the other hand, combining the difficulty assessment information can more accurately explain the student's performance. Among them, the knowledge acquisition of the student at the current time step is calculated according to the following formula:

[0113]

[0114] Where z t is the input data, is the initial knowledge acquisition value of the student, is the gate controlling the output, SKG t represents the output knowledge acquisition.

[0115] Then, after obtaining the knowledge acquisition of the student after completing the exercise, calculate based on the knowledge acquisition and the student's knowledge state at the previous time step to obtain the student's knowledge state at the current time step and update the student's knowledge state. The student's knowledge state at the current time step can be calculated through a gating mechanism. By calculating the knowledge pass-through quantity (Knowledge Pass Thought, KPT t ), control the knowledge pass-through quantity at the previous time step, and combine it with the knowledge acquisition at the current time step to obtain the specific knowledge state at the current time step. Therefore, the dynamic difficulty adaptability index at the current time step as well as the knowledge state at the previous time step and the answer at the current time step can be used as inputs, and the output is the student's knowledge state at the current time step.

[0116] Calculate the knowledge state of the student at the current time step according to the following formula, the formula is as follows:

[0117]

[0118] where k t-1 is the knowledge state of the student at the previous time step, and k t is the knowledge state of the student at the current time step, is a gating mechanism used to control the ratio of the information in k t-1 and SKG t ddai t is the dynamic difficulty adaptability index, is the answer sequence vector.

[0119] Finally, after obtaining the knowledge state of the student at the current time step, the probability of the student's correct answer at the next time step can be predicted based on the knowledge state at the current time step and the knowledge representation embedding vector at the next time step. The model input data is composed of the product of the knowledge state of the student at the current time step and the knowledge representation embedding at the next time step, and the probability of the correct answer is output through the sigmoid function. The knowledge representation embedding includes the question, concept, statistical difficulty, and model evaluation difficulty at the next time step. The prediction is carried out through the following formula, the formula is as follows:

[0120]

[0121] where, is the difficulty concatenated embedding vector obtained by embedding calculation of the model evaluation difficulty and the statistical difficulty, is the model evaluation difficulty, is the statistical difficulty, is the knowledge representation embedding, and Embeddlng(x) represents the calculation of the embedding value of element x, is the embedding of the question ID sequence at time step t + 1, is the embedding of the knowledge concept ID at time step t + 1, and y t+1 is the predicted probability value output.

[0122] Next, the embodiments of the present invention verify the performance of the knowledge tracking model of the present application through a comparative experiment, and the process is as follows:

[0123] 1. Dataset selection

[0124] Two public datasets are selected to train and evaluate the model:

[0125] XES3G5M is a large-scale dataset that contains a large number of questions and auxiliary information of related Knowledge Components (KCs). The XES3G5M dataset was collected from a real online mathematics learning platform and contains 7,652 questions, 865 KCs, and 5,549,635 learning interactions from 18,066 students.

[0126] The Eedi (NIPS34) dataset is sourced from the UK Eedi online education platform and collects middle school students' mathematics learning data. What makes this dataset special is that it contains a large number of math problems with pictures, and each question is in the form of a multiple-choice question. The dataset contains more than 15 million answer records, involving 120,000 students, 27,613 questions, and 388 different course topics.

[0127] 2. Comparison Models

[0128] For the multi-label sentiment classification task, a total of 16 superior-performing baselines were compared under each dataset. These methods were divided into 7 categories, among which the currently best-performing memory network-based methods were mainly compared. Compare the models that are superior in performance under each dataset. To verify the effectiveness of DDKT, 8 different baselines were used. All models were trained and evaluated with default parameters and trained on the A40 cluster. Details of all baselines are shown below:

[0129] DKT is the first model to introduce deep learning into the field of knowledge tracing and is implemented based on recurrent neural networks (RNN / LSTM). This model evaluates students' knowledge states by modeling the learning process of students and predicts students' mastery of questions and corresponding knowledge concepts.

[0130] SAKT is the first model to directly apply Transformer to the knowledge tracing task and is implemented based on the self-attention mechanism. This model proposes a self-attention model to capture the long-term dependencies between students' learning records, effectively improving the accuracy of knowledge state tracing.

[0131] DTransformer is a two-layer framework model based on Transformer, which ensures the stability of the model through contrastive learning. This model proposes a novel architecture to trace students' learning activity patterns and improves the generalization ability of the model through contrastive learning.

[0132] DKVMN is a knowledge tracing model based on memory networks, which is implemented using a static key matrix and a dynamic value matrix. This model defines that the key matrix stores latent knowledge concepts, the value matrix stores the student knowledge state, and updates the student knowledge state through read and write operations.

[0133] DIMKT is a knowledge tracing model that takes into account the item difficulty effect and is implemented based on a deep learning architecture. This model innovatively integrates item difficulty into the learning process, establishes the relationship between the student knowledge state and the item difficulty level, and significantly improves the prediction performance.

[0134] ReKT is a simple yet powerful knowledge tracing model implemented based on the FRU (Forget-Retrieve-Update) architecture. This model models the student knowledge state from multiple perspectives, including problem, concept, and domain knowledge states, and achieves excellent performance while maintaining simplicity.

[0135] extraKT is a knowledge tracing model that extends the context window of the attention model through length extrapolation. This model effectively solves the limitations of the attention mechanism in processing long sequences and can maintain stable performance even when the context window size changes.

[0136] simpleKT is a simple but powerful knowledge tracing baseline model implemented based on the ordinary dot product attention function. Inspired by the Rasch model in psychometrics, this model captures the individual differences between problems with the same knowledge concept by explicitly modeling problem-specific variations, and achieves excellent prediction performance while maintaining the simplicity of the model.

[0137] stableKT is a model improved based on simpleKT, which enhances the length generalization ability by applying a linear bias to the attention scores. This model uses a multi-head aggregation module to capture individual differences and complex hierarchical relationships, significantly improving the prediction performance and length generalization ability.

[0138] 3. Experimental Settings

[0139] First, deploy GLM-4 as a local large language model (LLM) to extract item difficulty. The difficulty coefficient uses a percentage system, ranging from 0 to 100, and a higher value indicates greater item difficulty. During the data loading phase, the statistical difficulty of the items is calculated for computing the difficulty-aware bias sequence. Following the Pykt benchmark format, each dataset is divided into 5 folds, where each record represents a student's learning sequence. Use fold0 as the test set and folds 1 to 4 as the training set.

[0140] In terms of model optimization, the Adam optimizer and MSE Loss are used as the loss function. The learning rate is set to 0.0001, the dropout rate is 0.2, the MLP embedding dimension is 256, and the forgetting factor used to calculate the difficulty mastery ratio is set to 0.8. In the multi-head attention mechanism, 4 attention heads are set. All experiments are conducted on the same A40 cluster to minimize the impact of hardware differences.

[0141] In terms of model evaluation, AUC and accuracy, which are widely used in knowledge tracing tasks, are adopted as evaluation metrics. Among them, AUC measures the ability of the model to distinguish correct and incorrect responses, while accuracy directly reflects the prediction accuracy. Finally, early stop is set to 10 epochs to reduce the training time.

[0142] 4. Monolingual Datasets

[0143] The performance comparison results between the knowledge tracing model (DDKT) of the present invention and the baseline model on three datasets are shown in Table 1. The results show that the DDKT model of the present invention is superior to other baselines in all aspects, especially in capturing complex learning patterns and dealing with diverse educational scenarios. In large-scale datasets such as XES3G5M, the DDKT model of the present invention shows superior performance compared to other baseline models, with the accuracy improvement ranging from 0.5% to 2%. In smaller-scale datasets such as Eedi, the DDKT model of the present invention also achieves the best performance, with the AUC improvement ranging from 1% to 6%. In addition, the DDKT model proposed by the present invention performs well on sparse datasets.

[0144] Table 1

[0145]

[0146] 5. Ablation Study

[0147] In the ablation experiment, the effects of three main components of DDKT on the model performance are studied: the difficulty mastery ratio, the difficulty perception bias sequence, and the Transformer layer used for feature combination.

[0148] DDKTw / o DMR: This variant removes the calculation component of the personalized mastery ratio used to track the learning status of students;

[0149] DDKT w / o DPBS: This variant removes the calculation component of the difficulty perception bias sequence that combines objective and subjective difficulty metrics;

[0150] DDKT w / o Transformer: This variant replaces the Transformer architecture with ordinary Multihead Attention to combine the input embeddings.

[0151] As shown in Table 2, the changes in the AUC and accuracy of the model after removing or replacing these modules were tested. On small datasets such as Eedi, the results showed that removing each module led to a decrease in AUC and accuracy. Therefore, each module is essential and there is a close collaborative relationship among them. Regarding DPBS, when only using statistical difficulty estimation for difficulty assessment, the following effects occur: after removing the model's assessment of difficulty, some questions perform worse due to the cold start problem, which directly results from the statistical evaluation difficulty value becoming too large or too small due to data sparsity; at the same time, due to the lack of the model's assessment of difficulty, the process of establishing the calibrated difficulty score is also missing. The model should adjust the weights based on the model's assessment of difficulty to calibrate the statistical evaluation difficulty, and the calibrated difficulty score is a more accurate difficulty calibration value that directly affects the model's performance.

[0152] Table 2

[0153]

[0154] For DMR, removing the DMR calculation component results in the absence of student personalized modeling, which directly causes the model to be unable to determine the student's difficulty mastery level during the learning process. For example, when a student consistently performs well on high-difficulty algebra problems but often encounters difficulties in basic arithmetic, traditional models may overestimate or underestimate their overall mathematical ability. DMR can capture this subtle performance pattern and provide a more accurate portrait of the student's knowledge state for the model. The TransformerEncoder contains Multi-headAttention and its feed-forward layer, and a dropout layer is added to prevent overfitting and avoid getting stuck in local optima. Therefore, using Transformer can not only perform the work of Multi-head Attention but also prevent the model from overfitting and effectively capture the long-range dependencies in the student's learning sequence through its self-attention mechanism.

[0155] Next, the knowledge tracing device based on dual-channel difficulty perception provided by the embodiments of the present invention will be described. The knowledge tracing device based on dual-channel difficulty perception described below can be correspondingly referred to the knowledge tracing method based on dual-channel difficulty perception described above.

[0156] First, in combination with Figure 4 , an introduction to the knowledge tracing device based on dual-channel difficulty perception will be given. As Figure 4 shown, the knowledge tracing device based on dual-channel difficulty perception may include:

[0157] A difficulty acquisition unit 100 is configured to acquire the statistical difficulty and the model evaluation difficulty of a target question;

[0158] A difficulty calibration unit 200 is configured to perform difficulty fusion on the statistical difficulty and the model evaluation difficulty based on a multi-head attention mechanism to obtain a calibrated difficulty score;

[0159] A deviation calculation unit 300 is configured to calculate a difficulty perception deviation sequence of the target question based on the calibrated difficulty score of the target question and the knowledge state of the student at the previous time step;

[0160] A mastery ratio calculation unit 400 is configured to determine a difficulty interval of the target question based on the statistical difficulty and calculate the difficulty mastery ratio of the student according to the difficulty interval;

[0161] A dynamic index calculation unit 500 is configured to integrate the dynamic change characteristics of the difficulty perception deviation sequence and the difficulty mastery ratio based on a self-attention mechanism to obtain a dynamic difficulty adaptability index;

[0162] A prediction unit 600 is configured to obtain input data of a knowledge tracing model based on the dynamic difficulty adaptability index and the knowledge state of the student at the previous time step, and predict the correct rate of the target question at the current time step through the knowledge tracing model.

[0163] An embodiment of the present invention further provides a storage medium, which can store a program suitable for being executed by a processor, and the program is used to implement each processing flow in the foregoing knowledge tracing solution based on dual-channel difficulty perception.

[0164] Finally, it should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover a non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "including a..." does not exclude the existence of additional identical elements in the process, method, article or device including the element.

[0165] Each embodiment in this specification is described in a progressive manner, and the key points of each embodiment are the differences from other embodiments. The same and similar parts among the embodiments can be referred to each other.

[0166] The foregoing description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A knowledge tracking method based on dual-channel difficulty perception, characterized in that: include: Obtain the statistical difficulty and model evaluation difficulty of the target question; Based on the multi-head attention mechanism, the statistical difficulty and model evaluation difficulty are integrated to obtain the calibration difficulty score; Calculate the difficulty perception deviation sequence of the target question based on the calibrated difficulty score of the target question and the student's knowledge status at the previous time step; Determine the difficulty range of the target questions based on statistical difficulty, and calculate the students' difficulty mastery ratio based on the difficulty range; Based on the self-attention mechanism, the dynamic change characteristics of the difficulty perception bias sequence and the difficulty mastery ratio are integrated to obtain the dynamic difficulty adaptability index; The input data of the knowledge tracking model is obtained based on the dynamic difficulty adaptability index and the knowledge status of the students in the previous time step, and the correct rate of the target question in the current time step is predicted by the knowledge tracking model.

2. The knowledge tracking method based on dual-channel difficulty perception according to claim 1 is characterized in that: The statistical difficulty and model evaluation difficulty of obtaining the target question include: The difficulty assessment model completed in advance performs difficulty assessment based on the input target question information to obtain the model assessment difficulty of the target question; The historical performance data of students are counted to obtain the statistical difficulty of the target questions.

3. The knowledge tracking method based on dual-channel difficulty perception according to claim 2 is characterized in that: The difficulty assessment model performs a difficulty assessment process, including: Get the topic information of the target topic entered by the user; Based on the question information, the target question is solved step by step in a structured manner to obtain a step-by-step solution; The difficulty is assessed based on the step-by-step solution and question information to obtain the model assessment difficulty.

4. The knowledge tracking method based on dual-channel difficulty perception according to claim 1 is characterized in that: The process of obtaining input data of the knowledge tracking model based on the dynamic difficulty adaptability index and the knowledge state of the student at the previous time step includes: Get the embedding vector of the dynamic difficulty adaptability indicator at the current time step; Perform element-wise multiplication of the embedding vector and the student’s knowledge state at the previous time step to obtain the input embedding of the knowledge tracking model; The input data is obtained by calculation based on the input embedding, answer sequence vector and knowledge concept vector.

5. The knowledge tracking method based on dual-channel difficulty perception according to claim 4 is characterized in that: The process of predicting the accuracy of the target question at the current time step by using the knowledge tracking model includes: Calculate the student's knowledge acquisition at the current time step based on the input data; Calculate based on the knowledge acquisition amount and the knowledge state of the student in the previous time step to obtain the knowledge state of the student in the current time step; The probability of the student's correct answer in the next time step is predicted based on the knowledge state of the current time step and the knowledge representation embedding vector of the next time step.

6. The knowledge tracking method based on dual-channel difficulty perception according to claim 5 is characterized in that: The students' difficulty mastery ratio is calculated by the following formula: in, is the difficulty mastering ratio, σ is the forgetting factor, B k represents the kth difficulty interval, represents the student's answer to question i at time step t-1, represents the number of questions in the kth difficulty interval at time step t-1, Represents the accuracy rate in the kth difficulty interval at time step t-1.

7. The knowledge tracking method based on dual-channel difficulty perception according to claim 6 is characterized in that: The amount of knowledge acquired by the student at the current time step is calculated according to the following formula: Among them, z t For input data, Acquire value for students’ initial knowledge, For control Output gating, SKG t Represents the amount of knowledge acquired in the output.

8. The knowledge tracking method based on dual-channel difficulty perception according to claim 7 is characterized in that: The knowledge state of the student at the current time step is calculated according to the following formula: Among them, k t-1 is the student’s knowledge state at the previous time step, k t is the student’s knowledge state at the current time step, is a gating mechanism used to control k t-1 and SKG t Information passing ratio in , ddai t is a dynamic difficulty adaptability indicator sequence, is the answer sequence vector.

9. A knowledge tracking device based on dual-channel difficulty perception, characterized in that: include: A difficulty acquisition unit, used to obtain the statistical difficulty and model evaluation difficulty of the target question; The difficulty calibration unit is used to integrate the statistical difficulty and model evaluation difficulty based on the multi-head attention mechanism to obtain a calibration difficulty score; a deviation calculation unit, for calculating a difficulty perception deviation sequence of a target question based on a calibrated difficulty score of the target question and a knowledge state of the student at a previous time step; A mastery ratio calculation unit, used to determine the difficulty range of the target question based on the statistical difficulty, and calculate the student's difficulty mastery ratio according to the difficulty range; A dynamic index calculation unit is used to integrate the dynamic change characteristics of the difficulty perception bias sequence and the difficulty mastery ratio based on the self-attention mechanism to obtain a dynamic difficulty adaptability index; The prediction unit is used to obtain input data of the knowledge tracking model based on the dynamic difficulty adaptability index and the knowledge state of the student at the previous time step, and predict the accuracy of the target question at the current time step through the knowledge tracking model.

Images