Knowledge tracking method, medium and equipment

Abstract

The invention discloses a knowledge tracking method, a medium and equipment, and the method comprises the steps: obtaining the situation information of exercises and the absolute difficulty of the exercises, and generating enhanced exercise representation; obtaining the historical time sequence cognitive state of the learner, and obtaining the overall trend of the cognitive state according to the relation between the cognitive state of the learner and the time sequence; acquiring the cognitive state of the learner at the last moment and the current exercise answering result, and dynamically updating to obtain the current cognitive state of the learner according to the latest development area theory based on the enhanced exercise representation and the overall trend of the cognitive state; and based on the current cognitive state of the learner, predicting the probability that the learner correctly answers the target exercise. According to the method, the exercise difficulty and the cognitive time sequence can be dynamically fused through the large language model, knowledge tracking is realized based on the recent development area theory, and the cognitive state evolution and the answer performance of the learner are accurately predicted.

Description

TECHNICAL FIELD

[0001] The present application relates to a knowledge tracing method, medium and device, belonging to the technical field of educational artificial intelligence and cognitive computing. BACKGROUND

[0002] As a core technology of educational artificial intelligence, knowledge tracing (KT) aims to dynamically model the evolution process of learners' cognitive state by analyzing their historical answer sequences, and provide decision basis for personalized teaching. With the development of online education, existing research mainly improves the tracing accuracy through two directions of improving item difficulty modeling and improving large language model (LLM) fusion.

[0003] In the direction of improving item difficulty modeling, early models such as EKT (Exercise-Enhanced KT) use text analysis to measure item semantic difficulty; advanced AKT (Adaptive KT) introduces an item embedding mechanism based on item response theory (IRT) to model item difficulty to obtain more accurate item representation; the latest research DIMKT (Difficulty-Integrated KT) integrates difficulty indicators into the whole process of cognitive state updating. However, these methods only implicitly consider the static difficulty of the item, and do not establish a dynamic interaction mechanism between "cognitive state-difficulty", which leads to the inability to reflect the subjective difficulty perception of learners when answering.

[0004] In the direction of improving large language model (LLM) fusion, the LLMKT scheme uses prompt methods to identify knowledge points in teaching dialogues to generate labeled data to enhance embedded information; the SINKT scheme integrates open world semantics and knowledge graph to enhance knowledge point association modeling. Although progress has been made, existing methods are still limited to the surface text processing capabilities of LLM (such as label generation, semantic expansion), and fail to take advantage of its deep temporal reasoning advantage, making it difficult to capture the long-period dependence relationship of cognitive state.

[0005] At present, the combination of knowledge tracking field and large model is not close enough, and only a huge semantic space is used in the early stage of combination to complete various text information identification, annotation, and concept information expansion. In essence, it is still the processing of text information, and its reasoning ability is not applied to the cognitive state evaluation process. Even if the structured knowledge graph is modeled through the language model, the semantic relationship between the knowledge points of the exercises is analyzed. However, the large language model cannot be combined with the core of knowledge tracking, that is, the analysis of the answering behavior of the learners. In addition, although some knowledge tracking models have taken the difficulty difference of exercises as an important index in the model. However, these models often only consider the influence of the difficulty difference of the exercises on the knowledge state of the learners at the exercise representation layer in actual use, lack of research on dynamic perception of exercise difficulty in the updating process of the cognitive state of the learners, and thus it is difficult to accurately track the dynamic knowledge state change of the learners in the practice process. SUMMARY

[0006] The present application aims to overcome the deficiencies in the prior art and provide a knowledge tracking method, medium and equipment that can dynamically fuse the exercise difficulty and cognitive timing through a large language model, realize knowledge tracking based on the theory of the recent development zone, and accurately predict the cognitive state evolution and answering performance of the learners. To achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0007] In a first aspect, the present application provides a knowledge tracking method, comprising:

[0008] obtaining the situational information of the exercises and the absolute difficulty of the exercises, and generating enhanced exercise representations;

[0009] obtaining the historical timing cognitive state of the learners, and obtaining the overall trend of the cognitive state according to the relationship between the cognitive state and the timing of the learners;

[0010] obtaining the cognitive state of the learners at the last moment and the answering result of the current exercises, and dynamically updating the current cognitive state of the learners based on the enhanced exercise representations and the overall trend of the cognitive state according to the theory of the recent development zone;

[0011] based on the current cognitive state of the learners, the probability of the learners correctly answering the target exercises is predicted.

[0012] In combination with the first aspect, the obtaining of the situational information of the exercises and the absolute difficulty of the exercises, and the generation of the enhanced exercise representations, comprises:

[0013] based on the situational information of the exercises, using a learnable knowledge point weight parameter matrix to represent the situational relationship between the exercises and the knowledge points;

[0014] based on the situational relationship between the exercises and the knowledge points, embedding the knowledge point weight and the exercise absolute difficulty into a multi-layer perception machine to generate exercise embeddings that fuse situational information;

[0015] The exercise embedded with the context information is embedded into a monotone self-attention mechanism of an input forgetting mechanism to obtain an enhanced exercise representation.

[0016] In combination with the first aspect, optionally, the historical time-series cognitive state of the learner is obtained, and a cognitive state overall trend is obtained according to a relationship between the cognitive state of the learner and time.

[0017] Reversible instance normalization is performed on the historical time-series cognitive state of the learner, and a sliding window is used to block in steps of S and a window size of L to obtain a plurality of cognitive state data blocks.

[0018] The pre-trained word embedding in the semantic space of the large language model is linearly mapped to a low-dimensional space to obtain a reduced word embedding.

[0019] According to the multi-head cross-attention mechanism, the reduced word embedding is used to align the cognitive state data blocks across modalities to generate a fused representation after alignment.

[0020] The task prompt prefix of the large language model is sequentially spliced with the fused representation after alignment, and the spliced sequence is input into the large language model for processing to obtain an output hidden state; wherein the task prompt prefix of the large language model includes data description, cognitive state tracking task requirements and cognitive state statistical information.

[0021] The time-series part corresponding to the cognitive state data block is extracted from the output hidden state, and the extracted time-series part is processed to output the cognitive state overall trend.

[0022] In combination with the first aspect, optionally, the cognitive state of the learner at the previous time and the current exercise answer result are obtained, and the current cognitive state of the learner is dynamically updated according to the latest development zone theory based on the enhanced exercise representation and the cognitive state overall trend, including:

[0023] Before answering the exercise, the relative difficulty of the exercise is calculated based on the difference between the enhanced exercise representation and the cognitive state of the learner at the previous time, and the subjective difficulty perception is output by processing the relative difficulty of the exercise through a neural network containing a gated recurrent unit.

[0024] During the process of answering the exercise, the knowledge acquisition amount is output by processing the subjective difficulty perception, the current exercise answer result and the absolute difficulty of the exercise through a neural network containing a gated recurrent unit.

[0025] After answering the exercises, the knowledge state indicator is obtained according to the cognitive state of the learner at the last time, subjective difficulty feeling, and the exercise absolute difficulty and the exercise answering result at the current time; the current cognitive state of the learner is obtained by processing the cognitive state of the learner at the last time, the knowledge state indicator and the knowledge acquisition amount through a neural network containing a gated recurrent unit.

[0026] In combination with the first aspect, optionally, the neural network containing a gated recurrent unit processes the exercise relative difficulty, outputs the subjective difficulty feeling, and is expressed by the following formula:

[0027] ,

[0028] wherein, is the subjective difficulty feeling, is the first gate, is a direct output of the exercise relative difficulty after neural network processing;

[0029] The first gate is expressed by the following formula:

[0030] ,

[0031] wherein, is an activation function; is a weight matrix, expressed as , is a dimensional space in a real number field, is an input layer dimension, is an exercise relative difficulty dimension; is a bias term matrix, expressed as , is a dimensional space in a real number field; is the exercise relative difficulty;

[0032] The direct output of the exercise relative difficulty after neural network processing is expressed by the following formula:

[0033] ,

[0034] wherein, is a nonlinear activation function; is a weight matrix, expressed as , is a bias term matrix, expressed as ; is the exercise relative difficulty;

[0035] The exercise relative difficulty is expressed by the following formula:

[0036] ,

[0037] wherein, the learner instantaneous cognitive state, the instantaneous enhanced problem representation.

[0038] In combination with the first aspect, optionally, the knowledge acquisition quantity is output by processing the subjective difficulty perception, the current problem answer result and the problem absolute difficulty through the neural network comprising the gated recurrent unit, and is expressed by the following formula:

[0039] ,

[0040] wherein, the knowledge acquisition quantity, the second gating, the direct output value of knowledge acquisition;

[0041] The second gating is expressed by the following formula:

[0042] ,

[0043] wherein, the activation function; the weight matrix, expressed as , the real number field dimensional space, the subjective difficulty perception dimension, the current problem answer result dimension, the problem absolute difficulty dimension, the input layer dimension; the subjective difficulty perception; the problem absolute difficulty; the concatenation operation; the bias term matrix, expressed as , the real number field dimensional space;

[0044] The direct output value of knowledge acquisition is expressed by the following formula:

[0045] ,

[0046] wherein, the nonlinear activation function; the weight matrix, expressed as It is the subjective feeling of difficulty. is the absolute difficulty of the exercise, For cascade operation, is the bias matrix, expressed as .

[0047] In combination with the first aspect, optionally, the learner's previous cognitive state, knowledge state indicator, and knowledge acquisition amount are processed by a neural network including a gated recurrent unit to obtain the learner's current cognitive state, which is expressed by the following formula:

[0048] ,

[0049] in, For learners Always be aware of your state, For learners Always be aware of your state, is the amount of knowledge acquired, is the knowledge status indicator, which is expressed by the following formula:

[0050] ,

[0051] in, for Activation function; is the weight matrix, expressed as , For the real number field Dimensional space, For learners The overall trend dimension of the cognitive state at each moment, It is the subjective difficulty dimension. The dimension of the answer result for the current exercise. is the absolute difficulty dimension of the exercise, is the input layer dimension; For learners The overall trend of cognitive state at each moment; It is a cascade operation; It is the subjective feeling of difficulty; For learners The answer result at the moment; The absolute difficulty of the exercise; is the bias matrix, expressed as .

[0052] In conjunction with the first aspect, optionally, predicting the probability of the learner correctly answering the target exercise based on the learner's current cognitive state includes:

[0053] Based on the current cognitive state of the learner and the enhanced exercise representation, an ideal probability of the learner correctly answering the target exercise is calculated;

[0054] Based on the guess factor and the error factor corresponding to the target exercise, the ideal probability of the learner correctly answering the target exercise is updated to obtain a final probability of the learner correctly answering the target exercise.

[0055] In a second aspect, the present application provides a computer readable storage medium having stored thereon a computer program / instruction, which, when executed by a processor, implements the steps of the knowledge tracking method of the first aspect.

[0056] In a third aspect, the present application provides a computer device, characterized in that it comprises:

[0057] a memory for storing a computer program / instruction;

[0058] a processor for executing the computer program / instruction to implement the steps of the knowledge tracking method of the first aspect.

[0059] Compared with the prior art, the knowledge tracking method, medium and device provided by the embodiments of the present application have the following beneficial effects:

[0060] The present application obtains the context information of the exercise and the absolute difficulty of the exercise, and generates an enhanced exercise representation. The present application integrates the exercise context information, enriches the difficulty perception of the exercise, and solves the problem of insufficient modeling of exercise semantics and difficulty in traditional models.

[0061] The present application obtains the historical time sequence cognitive state of the learner, and obtains the overall trend of the cognitive state according to the relationship between the cognitive state of the learner and the time sequence. The present application uses a large language model to reprogram the time sequence data into a more natural text prototype of the large language model, and captures the long-time sequence dependence relationship of the cognitive state of the learner in the sequence answer behavior of the learner through the large language model.

[0062] The present application eliminates the distribution offset of the cognitive state sequence through reversible instance normalization and block processing. The present application realizes the alignment of the time sequence data and the text semantics by using dimension reduction word embedding and multi-head cross attention mechanism. The present application combines the task prompt prefix to guide the large language model to understand the cognitive state evolution task. The present application enables the model to capture more long-period dependence of the historical sequence.

[0063] The present application obtains the cognitive state of the learner at the last time and the current exercise answer result, and dynamically updates the current cognitive state of the learner based on the enhanced exercise representation and the overall trend of the cognitive state according to the recent development zone theory. The present application considers the subjective difficulty perception of the learner, and dynamically updates the cognitive state of the learner by using a neural network, thereby realizing the association between the exercise features and the cognitive state of the learner.

[0064] The present application is based on a three-stage dynamic updating mechanism of the recent development zone theory, and the relative difficulty of the exercises is processed by a gating loop unit before the exercises are answered, so as to obtain the subjective difficulty perception of the learners and accurately quantify the individual differences of the learners; the subjective difficulty perception of the learners, the answering results and the absolute difficulty, and the knowledge acquisition amount are combined during the process of answering the exercises, so as to dynamically evaluate the adaptability of the recent development zone; and the knowledge state indicator is used to calibrate the cognitive state after the exercises are answered, so that the state updating conforms to the laws of educational psychology;

[0065] The present application predicts the probability of the learners to correctly answer the target exercises based on the current cognitive state of the learners; the ideal prediction probability is corrected by introducing the guess factor and the error factor, so as to accurately evaluate the cognitive state of the learners, and make the prediction results conform to the cognition of psychometrics, and the cognitive state is interpretable;

[0066] The present application can dynamically fuse the exercise difficulty and the cognitive time sequence by means of a large language model, realize knowledge tracking based on the recent development zone theory, and accurately predict the evolution of the cognitive state of the learners and the answering performance. BRIEF DESCRIPTION OF DRAWINGS

[0067] Figure 1 is a flowchart of a knowledge tracking method in embodiment 1 of the present application;

[0068] Figure 2 is a flowchart of obtaining the overall trend of the cognitive state in the knowledge tracking method in embodiment 1 of the present application;

[0069] Figure 3 is a flowchart of updating the current cognitive state of the learners in the knowledge tracking method in embodiment 1 of the present application. DETAILED DESCRIPTION

[0070] The present application will be further described below in combination with the drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present application, and cannot be used to limit the protection scope of the present application.

[0071] Embodiment 1:

[0072] As shown in the figure, the present embodiment provides a knowledge tracking method, which comprises: Figure 1 obtaining the context information of the exercises and the absolute difficulty of the exercises, and generating enhanced exercise representations;

[0073] obtaining the historical time sequence cognitive state of the learners, and obtaining the overall trend of the cognitive state according to the relationship between the cognitive state of the learners and the time sequence;

[0074]

[0075] ​Acquire the learner's previous cognitive state and the current problem solving result, based on the enhanced problem representation and the overall trend of cognitive state, dynamically update the learner's current cognitive state according to the zone of proximal development theory;

[0076] Based on the current cognitive state of the learner, predict the probability of the learner correctly solving the target problem.

[0077] The specific implementation steps are as follows:

[0078] Step 1: Acquire the context information of the problem and the absolute difficulty of the problem, and generate the enhanced problem representation.

[0079] Step 1.1: Based on the context information of the problem, use the learnable knowledge point weight parameter matrix to represent the context relationship between the problem and the knowledge point.

[0080] The learnable knowledge point weight parameter matrix is represented by the following formula:

[0081] , formula (1),

[0082] In formula (1), is the learnable knowledge point weight parameter matrix, and the symbol represents element-by-element multiplication, is the initialized knowledge point weight parameter, and the symbol represents element-by-element multiplication. If a problem corresponds to multiple knowledge points, it is necessary to ensure that the sum of the knowledge point weight parameters of this problem is 1, and the weight parameters of the corresponding knowledge points are evenly distributed.

[0083] In formula (1), is the matrix of one-to-one correspondence between the problem and the knowledge point, which is represented by the following formula:

[0084] , formula (2),

[0085] In formula (2), the number of rows of the matrix is the number of problems , the number of columns of the matrix is the number of knowledge points , if the problem and the knowledge point have a relationship, it is 1, and if they do not have a relationship, it is 0.

[0086] During learning, the learnable knowledge point weight parameter matrix is constantly updated, and the context relationship between the problem and the knowledge point is updated. After learning, in order to maintain the normalization constraint, the softmax normalization is realized:

[0087] , formula (3),

[0088] In formula (3), The weight parameter of the learnable knowledge point obtained after learning is processed, and the vector elements of the knowledge concept examined by each question after processing are 1.

[0089] The absolute difficulty of the question is composed of the problem difficulty and the knowledge point difficulty, and is represented by the correct answer rate. The calculation formula is as follows:

[0090] , formula (4),

[0091] In formula (4), , and are the problem difficulty, the knowledge point difficulty and the absolute difficulty of the question respectively, is the number of learners who answered correctly on the question, is the number of learners who answered correctly on the knowledge point, is the total number of learners who answered the question or the knowledge point, is the predetermined question level, is the predetermined knowledge point level, and the default setting , is a concatenation symbol.

[0092] The embodiment uses the learnable knowledge point weight parameter matrix to represent the context relationship between the question and the knowledge point, specifically the context space relationship, and realizes the construction of the complex "knowledge point-question" question context by integrating the absolute difficulty of the question.

[0093] Step 1.2: Based on the context relationship between the question and the knowledge point, the knowledge point weight embedding and the absolute difficulty of the question embedding are input into the multilayer perception machine to generate a question embedding that integrates context information.

[0094] It is expressed by the following formula:

[0095] , formula (5),

[0096] In formula (5), is the question embedding that integrates context information obtained after the multilayer perception machine is integrated; is the knowledge point weight embedding; is the absolute difficulty of the question embedding; is a concatenation operator; is a weight matrix, represented as , is a dimensional space in the real number field, is the number of knowledge points, is the dimension of the absolute difficulty of the question embedding, is the input layer dimension; is a bias term matrix, represented as , is a real number field dimensional space.

[0097] Step 1.3: Embedding the problem with contextual information into the monotonic self-attention mechanism of the input forgetting mechanism to obtain enhanced problem representation.

[0098] It is expressed by the following formula:

[0099] , formula (6),

[0100] In formula (6), is the time step corresponding to the query vector, is the time step corresponding to the key vector, is the time step corresponding to the value vector, is the problem embedding with contextual information, is a learnable weight matrix;

[0101] In formula (6), is calculated using the function is normalized after calculating the correlation between time step and time step , is the time step corresponding to the key vector, is the input layer dimension, i.e. the single sample dimension of the input sample sequence , is the similarity score of the forgetting mechanism between time step and time step , is the sum of the similarity of the forgetting mechanism of all positions in the sequence;

[0102] In formula (6), is a learnable decay rate parameter; is a time distance measure between time step and time step , time step corresponds to the last problem related to the same knowledge point as the current problem; is the dimension of ;

[0103] In formula (6), is the single-head attention output obtained after attention calculation at time step ;

[0104] In formula (6), is a weight matrix, represented as , is a real number domain dimensional space, is the dimension of the multi-head attention output after splicing, is the input layer dimension; is a splicing operation; is a bias term matrix, denoted as , is a real number domain dimensional space.

[0105] The embodiment fuses problem context information, enriches the difficulty perception of the problem, and solves the problem of insufficient modeling of problem semantics and difficulty in traditional models.

[0106] Step 2: Obtain the historical time sequence cognitive state of the learner, and obtain the overall trend of the cognitive state according to the relationship between the cognitive state of the learner and the time sequence;

[0107] According to the constructivist learning theory, it can be known that the cognitive state of the learner is a continuous development process, and the acquisition of new knowledge is an expansion and deepening on the basis of the original knowledge, rather than isolated or sudden. The cognitive state of the learner in a period of time is continuous, and increases and decreases within a certain range. However, the existing knowledge tracking method cannot effectively fit the overall trend of the cognitive state of the learner.

[0108] The embodiment considers the long sequence information capturing ability of the large language model, captures the time sequence information of the cognitive state of the learner in the sequence answer record of the learner according to the cognitive state of the learner in a period of time, masters the overall trend of the cognitive state of the learner in the time progression process, and more comprehensively evaluates the ability and potential of the learner.

[0109] Step 2.1: As shown in Figure 2 , reversible instance normalization is performed on the historical time sequence cognitive state of the learner, and a sliding window is used to block with a step size S and a window size L, to obtain a plurality of cognitive state data blocks.

[0110] In the embodiment, the reversible instance normalization is implemented to eliminate the sequence distribution offset problem by using a learnable affine transformation parameter.

[0111] The cognitive state data block is denoted as , each cognitive state data block is linearly mapped to a dimension , that is, .

[0112] In the embodiment, is the number of cognitive state data blocks, .

[0113] Step 2.2: As shown in Figure 2As shown in Figure 1, the pre-trained word embeddings in the semantic space of the large language model are linearly mapped to the low-dimensional space to obtain the reduced-dimensional word embeddings.

[0114] In this embodiment, in order to avoid creating a dense reprogramming space, a linear mapping method is used to embed the pre-trained words into (where V is the vocabulary size and D is the hidden layer dimension of the large language model) is mapped to a smaller word embedding space ,in , effectively alleviating the dimensionality expansion problem of traditional word embedding reprogramming strategies.

[0115] Step 2.3: If Figure 2 As shown in the figure, based on the multi-head cross-attention mechanism, the dimensionality-reduced word embedding is used to align the cognitive state data blocks across modalities and generate an aligned fusion representation.

[0116] Step 2.3.1: In the multi-head cross attention mechanism, each attention head The query matrix is ​​obtained according to the text mode in the reduced word embedding, and the key matrix and value matrix are obtained according to the temporal mode in the reduced word embedding, which can be expressed as follows:

[0117] , formula (7),

[0118] In formula (7), For the The query matrix of the attention heads, For the cognitive state data blocks, is the query matrix coefficient, expressed as , For the real number field dimensional space, is the cognitive state data block dimension, is the dimension of the multi-head cross attention mechanism, expressed as , For floor operation, is the number of heads in the multi-head cross attention mechanism;

[0119] In formula (7), No. The key matrix of the attention heads, is the word embedding space after dimensionality reduction, expressed as , For the real number field dimensional space, is the vocabulary size after dimensionality reduction, is the hidden dimension of the large language model, is the bond matrix coefficient, expressed as , is the multi-head cross-attention mechanism dimension;

[0120] in formula (7), is the key matrix of the i-th attention head, is the value matrix coefficient, denoted as . .

[0121] Step 2.3.2: Compile the output of each attention head , which is represented by the following formula:

[0122] , formula (8),

[0123] in formula (8), is the compilation result, is the attention function, is the normalized exponential function, is the multi-head cross-attention mechanism dimension.

[0124] Step 2.3.3: Aggregate the compilation results of the i-th attention head to obtain: , wherein, is the aggregation result, is a real number field dimensional space.

[0125] Step 2.3.4: Linearly transform the aggregation result to obtain a cross-modal fusion representation output that matches the hidden layer dimension of the large language model, i.e., the aligned fusion representation.

[0126] Step 2.4: As shown in Figure 2 , concatenate the task prompt prefix of the large language model with the aligned fusion representation in sequence, and input the concatenated sequence into the large language model for processing to obtain the output hidden state.

[0127] The task prompt prefix of the large language model includes data description, cognitive state tracking task requirements, and cognitive state statistical information.

[0128] In this embodiment, the data description includes dataset size, number of learners, number of interactions, etc.

[0129] In this embodiment, the cognitive state tracking task requirement is the specific goal of cognitive state tracking.

[0130] In this embodiment, the cognitive state statistical information includes maximum / minimum / average cognitive state, change trend, etc.

[0131] ​The task prompt prefix of the large language model is concatenated with the aligned fusion representation in sequence, which is represented by the following formula:

[0132] , formula (9),

[0133] In formula (9), is a discrete symbol sequence obtained after processing, is a word segmenter, is data description, is cognitive state tracking task requirement, is cognitive state statistical information, is text concatenation.

[0134] In formula (9), is the concatenation result of the task prompt prefix and the fusion representation sequence, is an embedding function for converting a discrete sequence into a continuous vector, is a real number field, dimensional space, is the sequence length dimension of, is the hidden layer dimension of the large language model.

[0135] The embedding of the text prompt (the concatenation result of the task prompt prefix and the fusion representation sequence) is processed by the large model and the processed time series feature is concatenated in the sequence dimension to form the input embedding of the complete large language model, and then input into the large language model to obtain the last layer hidden state as the output .

[0136] Step 2.5: Extract the time series part corresponding to the cognitive state data block from the output hidden state, process the extracted time series part, and output the overall trend of the cognitive state.

[0137] From the output , discard the prefix part, keep the dimension, project the output cognitive state data block to the cognitive state data of T time steps through flattening and linear mapping, and obtain the final cognitive state output . It is represented by the following formula:

[0138] , formula (10),

[0139] In formula (10), is the processing of the large language model, is the concatenation operation, is a real number field, dimensional space. ​​

[0140] In formula (10), is the spatial variation matrix, expressed as , For the real number field Dimensional space; is the dimension flattening function, To extract the last Position to last The tensor slicing operation of the sub-tensor at each position, For the real number field dimensional space, T is the time step dimension, It is the dimension of cognitive status data block.

[0141] Specifically, this embodiment eliminates the distribution offset of cognitive state sequences through reversible instance normalization and block processing; adopts dimensionality reduction word embedding and multi-head cross-attention mechanism to achieve alignment of time series data with text semantics; combines task prompt prefixes to guide the large language model to understand cognitive state evolution tasks; and enables the model to capture more long-term dependencies of historical sequences.

[0142] This embodiment uses a large language model to reprogram time series data into a more natural text prototype of the large language model, and uses the large language model to capture the long-term dependencies of the learner's cognitive state in the learner's sequential answering behavior.

[0143] Step 3: Obtain the learner's previous cognitive state and the current exercise answer result. Based on the enhanced exercise representation and the overall trend of cognitive state, the learner's current cognitive state is dynamically updated according to the theory of the zone of proximal development.

[0144] like Figure 3 As shown, based on the theory of the zone of proximal development, the learner's cognitive state is dynamically updated, and the impact of the two dimensions of relative difficulty and absolute difficulty on the learner's cognitive state evaluation is considered. A dynamic knowledge tracking mechanism of "cognitive pre-assessment-development zone adaptation-state progressive calibration" is constructed before, during, and after answering exercises.

[0145] Step 3.1: Before answering an exercise, the relative difficulty of the exercise is calculated based on the difference between the enhanced exercise representation and the learner's cognitive state at the previous moment. The relative difficulty of the exercise is processed by a neural network containing a gated recurrent unit to output a subjective difficulty feeling.

[0146] Step 3.1.1: Calculate the relative difficulty of the exercise based on the difference between the enhanced exercise representation and the learner's previous cognitive state.

[0147] The learner's subjective feeling of difficulty of the exercise, that is, the relative difficulty of the exercise, is used to indicate the learner's current cognitive level and can be used to judge whether the learner can answer the exercise independently.

[0148] The item relative difficulty in the embodiment is the difference between the learner's cognitive state at the previous time and the enhanced item representation at the current time when answering the item. It is expressed by the following formula:

[0149] , formula (11),

[0150] In formula (11), is the learner's cognitive state at the previous time, is the enhanced item representation at the current time.

[0151] Step 3.1.2: Process the item relative difficulty through a neural network containing a gated recurrent unit, and output the subjective difficulty perception.

[0152] It is expressed by the following formula:

[0153] , formula (12),

[0154] In formula (12), is the subjective difficulty perception, is the first gate, is the direct output of the item relative difficulty after neural network processing.

[0155] In formula (12), the first gate is expressed by the following formula:

[0156] , formula (13),

[0157] In formula (13), is the activation function; is the weight matrix, expressed as , is a dimensional space in the real number field, is the input layer dimension, is the item relative difficulty dimension; is the bias term matrix, expressed as , is a dimensional space in the real number field; is the item relative difficulty.

[0158] In formula (12), the direct output of the item relative difficulty after neural network processing is expressed by the following formula:

[0159] , formula (14),

[0160] In formula (14), is a nonlinear activation function; is a weight matrix, denoted as , is a bias term matrix, denoted as ; is the relative difficulty of the exercise.

[0161] Step 3.2: During the exercise answering process, the subjective difficulty perception, the current exercise answering result, and the exercise absolute difficulty are processed by a neural network containing a gated recurrent unit, and the knowledge acquisition amount is output.

[0162] In assessing whether the learner has reached the "zone of proximal development", the knowledge acquisition representation is used, focusing on whether the learner can successfully solve exercises beyond their current knowledge level, especially the learner's performance when facing challenging questions.

[0163] Specifically, when the learner faces exercises higher than their current knowledge level and successfully solves them, the learner will achieve deep mastery of the knowledge point and gain significant knowledge improvement. Conversely, if the exercise is too simple for the learner, even if the answer is correct, the learner's mastery of the knowledge point and overall cognitive state improvement will be relatively limited.

[0164] In summary, the learner's knowledge acquisition is closely related to their existing development level and actual answer results. Therefore, the learner's existing development level is combined with the actual answer results and the difficulty of the exercise itself, i.e., the absolute difficulty, to comprehensively evaluate their knowledge acquisition.

[0165] which is expressed by the following formula:

[0166] , formula (15),

[0167] In formula (15), is the knowledge acquisition amount, is the second gate, is the direct output value of knowledge acquisition.

[0168] In formula (15), the second gate is expressed by the following formula:

[0169] , formula (16),

[0170] In formula (16), is an activation function; is a weight matrix, denoted as , is a dimensional space in the real number field, is the subjective difficulty perception dimension, a dimension of the result of the current exercise, a dimension of the absolute difficulty of the exercise, a dimension of the input layer; a dimension of the subjective difficulty perception; a dimension of the absolute difficulty of the exercise; a dimension of the concatenation operation; a dimension of the bias term matrix, denoted as , a dimension of the real field dimensional space.

[0171] In equation (15), the direct output value of knowledge acquisition is represented by the following equation:

[0172] , equation (17),

[0173] In equation (17), is a nonlinear activation function; is a weight matrix, denoted as a dimension of the subjective difficulty perception, a dimension of the absolute difficulty of the exercise, a dimension of the concatenation operation, a dimension of the bias term matrix, denoted as .

[0174] Step 3.3: After answering the exercise, the knowledge state indicator is obtained according to the learner's cognitive state at the last time, the subjective difficulty perception, the result of the current exercise and the absolute difficulty of the exercise; the current cognitive state of the learner is obtained by processing the learner's cognitive state at the last time, the knowledge state indicator and the knowledge acquisition amount through a neural network containing a gated recurrent unit.

[0175] After answering the exercise, the cognitive state of the learner needs to be updated moderately and objectively.

[0176] In the update, the cognitive state improvement of the learner during the practice process is comprehensively considered with the past cognitive state. The basis for evaluation includes the characteristics of the learner's cognitive state under the time characteristics, the given answer and the difficulty of the exercise. If the learner answers correctly in the exercise, it is considered that he or she has reached the "zone of proximal development", and the cognitive state acquisition achieved in the answering process should be focused on. On the contrary, if the learner answers incorrectly, it indicates that he or she cannot basically obtain new knowledge on this exercise, and the improvement amplitude of his or her cognitive state is small. In this case, more attention should be paid to the learner's previous cognitive state.

[0177] In the evaluation, the knowledge state indicator is considered.

[0178] Step 3.3.1: According to the learner's previous moment cognitive state, subjective difficulty feeling, current exercise answer result and exercise absolute difficulty, the knowledge state indicator is obtained.

[0179] is expressed by the following formula:

[0180] , formula (18),

[0181] In formula (18), is the activation function; is the weight matrix, expressed as , is a dimensional space in the real field, is the overall trend dimension of the learner's previous moment cognitive state, is the subjective difficulty feeling dimension, is the current exercise answer result dimension, is the exercise absolute difficulty dimension, is the input layer dimension; is the overall trend of the learner's previous moment cognitive state; is the cascade operation; is the subjective difficulty feeling; is the answer result of the learner at the previous moment; is the exercise absolute difficulty; is the bias matrix, expressed as .

[0182] Step 3.3.2: The learner's current cognitive state is obtained by processing the learner's previous moment cognitive state, knowledge state indicator and knowledge acquisition quantity through a neural network containing a gated recurrent unit.

[0183] is expressed by the following formula:

[0184] , formula (19),

[0185] In formula (19), is the learner's previous moment cognitive state, is the learner's previous moment cognitive state, is the knowledge acquisition quantity, is the knowledge state indicator.

[0186] This embodiment, based on the three-stage dynamic update mechanism of the zone of proximal development theory, processes the relative difficulty of exercises before answering them through a gated recurrent unit to determine the learner's subjective perception of difficulty, accurately quantifying individual differences among learners. During the exercise-answering process, the learner's subjective perception of difficulty, answer results, absolute difficulty, and knowledge acquisition are combined to dynamically assess the adaptability of the zone of proximal development. After answering the exercises, the knowledge state indicator is used to calibrate the cognitive state, ensuring that the state update conforms to the laws of educational psychology. Taking into account the learner's subjective perception of difficulty, a neural network is used to dynamically update the learner's cognitive state, establishing a connection between exercise characteristics and the learner's cognitive state.

[0187] Step 4: Based on the learner's current cognitive state, predict the probability that the learner will correctly answer the target exercise.

[0188] Step 4.1: Based on the learner's current cognitive state and the enhanced problem representation, calculate the ideal probability that the learner will correctly answer the target problem.

[0189] Leveraging learners The current cognitive state at the moment and The enhanced exercise representation of the target exercise at each moment simulates the learner's answering process and predicts the ideal probability of the learner correctly answering the target exercise.

[0190] It is expressed by the following formula:

[0191] , formula (20),

[0192] In formula (20), for The ideal probability that a learner will correctly answer the target exercise at any given moment, For learners Always be aware of your state, for Enhanced problem representation for moment-target problems.

[0193] Step 4.2: Based on the guess factor and error factor corresponding to the target exercise, update the ideal probability that the learner correctly answers the target exercise to obtain the final probability that the learner correctly answers the target exercise.

[0194] It is expressed by the following formula:

[0195] , formula (21),

[0196] In formula (21), for The final probability that the learner correctly answers the target exercise at any given moment, for The guess factor for the time target exercises, for a failure factor of the moment target exercise.

[0197] In formula (21), a guess factor of the moment target exercise and a failure factor of the moment target exercise , is expressed by the following formula:

[0198] In formula (22),

[0199] In formula (22), is a regularization operation, is a ReLU activation function, , is a weight matrix, denoted as ; , is a bias item matrix, denoted as .

[0200] The embodiment introduces a guess factor and a failure factor to correct the ideal prediction probability, realizes accurate evaluation of the cognitive state of the learner, and makes the prediction result consistent with the psychology of measurement. The cognitive state has interpretability.

[0201] To sum up, the embodiment can dynamically fuse the exercise difficulty and the cognitive time sequence through a large language model, realize knowledge tracking based on the recent development zone theory, and accurately predict the evolution of the cognitive state and the performance of the learner.

[0202] Embodiment 2:

[0203] The embodiment selects three real online education data sets to evaluate the knowledge tracking method provided in embodiment 1.

[0204] The three real online education data sets selected in the embodiment are ASSISTments2017, Ednet and ASSISTments2012. These three data sets all contain exercises, knowledge points, answer results, submission times and answer durations and other key attributes, and the sizes of the data sets are different, so that the performance of the model can be accurately evaluated from different dimensions. In the embodiment, the three real online education data sets are processed, and the part of the answer records with data missing is removed to ensure the integrity and accuracy of the data.

[0205] Among them, the ASSISTments2017 data set comes from the data mining competition in 2017, containing 942816 interaction records, involving 686 students and 102 exercises.

[0206] Among them, the Ednet dataset is the largest publicly available interactive educational system dataset known to date, collected by the artificial guided system Santa, containing a total of 131317236 interaction information.

[0207] Among them, the ASSISTments2012 dataset is derived from the free online tutoring platform ASSISTments in the 2012-2013 academic year, in which the learning records related to knowledge point missing have been recorded in detail. This dataset contains 6123270 interaction records, involving 46674 students and 179999 exercises.

[0208] In order to accurately reflect the influence of the answer sequence on the cognitive state of the learner, the learning records of the learners are arranged in order according to the answer sequence. In order to ensure the efficiency of the knowledge tracking method (model) processing, the length of the input sequence is uniformly set to a fixed value of 200, and the answer records are divided into multiple fixed-length input sequences. For the part that is less than 200 after cutting, zero vectors are used for padding. At the same time, in order to improve the accuracy and reliability of the evaluation, the data with less than 30 answer records is excluded to fully ensure the representativeness of the sample.

[0209] In the training phase of the knowledge tracking method (model), 20% of the dataset is selected as the test set. Before training, the standard K-fold cross-validation process (K=5) is followed to preprocess all the datasets, and in each cross-validation, the training set and the validation set are divided according to the ratio of 5:1. In addition, the learning rate is set to 0.002, and the learning rate is reduced by 50% every 5 epochs to achieve the optimal solution. All parameters are randomly initialized in a uniform distribution. The dimension parameter is uniformly set to the number of knowledge points. In order to start the training process, all parameters are randomly initialized to a uniform distribution. All hyperparameters are learned on the training set, and the model that performs best on the validation set is used to evaluate the test set.

[0210] Training and evaluation environment: Linux operating system, CPU model Intel(R) Xeon(R) Gold 6248R CPU@ 3.00GHz * 96, memory 62.5GB, display chip model Tesla V100.

[0211] The control methods selected in this embodiment include: DKT, DKVMN, SAKT, AKT, DIMKT, StableKT and UKT.

[0212] DKT (Deep Knowledge Tracing, Deep Knowledge Tracing) is the earliest KT model based on deep learning, which uses RNN / LSTMs to evaluate the knowledge state of students.

[0213] DKVMN (Dynamic Key-Value Memory Networks) is a KT model based on memory networks, which defines a matrix key that stores latent KCs and a matrix value that stores the student's knowledge state; using read-write operations, the student's knowledge state is updated over time.

[0214] SAKT (Self-Attentive Knowledge Tracing) is a self-attention model that directly applies a transformer to the KT task, capturing long-term dependencies between student learning records.

[0215] AKT (Attentive Knowledge Tracing) is an attention-based knowledge tracing model with context awareness, which defines a knowledge retriever that obtains the student's dynamic knowledge state through an attention mechanism; using item difficulty, it improves item embedding through IRT, and designs two self-attention encoders to learn the context representation of items and answers.

[0216] DIMKT (Difficulty-Integrated Knowledge Tracing) is a KT model that explicitly integrates difficulty levels into problem representations and establishes a relationship between learners' knowledge states and item difficulty levels during practice.

[0217] StableKT (Stable Knowledge Tracing) is a KT model that uses a multi-head aggregation module to capture individual differences and learn from short sequences while maintaining consistent performance in long sequences and capturing hierarchical relationships between problems and related knowledge points.

[0218] UKT (Uncertainty-aware Knowledge Tracing) is an uncertainty-aware knowledge tracing model that first uses a random distribution embedding to represent uncertainty in student interactions and uses a self-attention mechanism to capture the transition of state distribution in learners' learning behavior.

[0219] Among them, DKT and DKVMN models are knowledge tracing methods that do not consider item difficulty, AKT and DIMKT are knowledge tracing methods that consider item difficulty. SAKT is an item-centered knowledge tracing model, and StableKT and UKT models are emerging models in recent years.

[0220] The knowledge tracing method provided in Embodiment 1 (A Large Language Model Enhanced Framework for Difficulty-Aware and Temporal-Adaptive Knowledge Tracing, referred to as LLM-DTKT model) is compared with the above seven typical models, from the necessity of considering the difficulty of the exercises and the superiority of dynamically considering the difficulty of the exercises, and compared with the emerging models, to show the effectiveness of the knowledge tracing method provided in Embodiment 1.

[0221] Specifically, the performance of the learner on the exercise at the future time step t is predicted, and the cognitive state of the learner is predicted by learning the exercise answer records of the learner at time steps 1 to t-1 in the data set. In order to evaluate the reliability, three evaluation indexes are used: area under the curve (AUC), accuracy (ACC) and root mean square error (RMSE). This embodiment compares the prediction performance of the LLM-DTKT model and other baseline models, and the comparison results are shown in Table 1. The bold part is the result of the method provided in Embodiment 1, and the underlined part is the optimal result.

[0222] Table 1 Comparison results

[0223]

[0224] Firstly, according to Table 1, the results of the knowledge tracing method provided in Embodiment 1 (LLM-DTKT model) on the ASSISTments2017 dataset and the Ednet dataset are better than those of all models, and on the ASSIST2012 dataset, there is only a 0.15% gap with the optimal result. The excellent performance of the knowledge tracing method provided in Embodiment 1 (LLM-DTKT model) shows that it is necessary and valuable to consider the complex exercise situation information representation, capture the long sequence information of the learner's answer, and dynamically perceive the difficulty of the exercise.

[0225] Secondly, the result of the knowledge tracing method provided in Embodiment 1 (LLM-DTKT model) on the ASSISTments2012 dataset has a 0.15% gap with the optimal result, which is due to the fact that in this dataset, one problem corresponds to one knowledge point, and the exercise situation embedding cannot effectively learn the complex situation relationship between the exercise and the knowledge point. In the ASSISTments2017 and Ednet datasets, the exercise and the knowledge point are in a one-to-many relationship, and the knowledge tracing method provided in Embodiment 1 (LLM-DTKT model) can more effectively learn the exercise situation.

[0226] In summary, the embodiment 1 can dynamically fuse the exercise difficulty and cognitive timing through a large language model, realize knowledge tracking based on the recent development zone theory, and accurately predict the cognitive state evolution and answer performance of the learners.

[0227] Embodiment 3:

[0228] The embodiment provides a computer readable storage medium, and computer programs / instructions are stored on the computer readable storage medium.

[0229] Embodiment 4:

[0230] The embodiment provides a computer device, and the computer device comprises:

[0231] a memory configured to store computer programs / instructions;

[0232] a processor configured to execute the computer programs / instructions to implement the steps of the knowledge tracking method in the embodiment 1.

[0233] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage, etc.) containing computer-usable program code.

[0234] The present application is described with reference to flowcharts and / or block diagrams of the method, device (system), and computer program product according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to produce a machine, so that the instructions executed by the computer or other programmable data processing devices generate a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 an apparatus for performing the functions specified in the flowcharts and / or block diagrams. Figure 1 an apparatus for performing the functions specified in the flowcharts and / or block diagrams.

[0235] These computer program instructions can also be stored in a computer readable storage medium that can guide the computer or other programmable data processing devices to work in a specific manner, so that the instructions stored in the computer readable storage medium produce a product including instruction apparatus, which implements the functions specified in the flowcharts and / or block diagrams. Figure 1 an apparatus for performing the functions specified in the flowcharts and / or block diagrams.Figure 1 the function specified in the one or more blocks.

[0236] These computer program instructions can also be loaded into computer or other programmable data processing devices, so that a series of operation steps are performed on the computer or other programmable data processing devices to generate computer-implemented processes, so that the instructions executed on the computer or other programmable data processing devices provide processes for implementing the flow Figure 1 the flow or flows and / or blocks Figure 1 the steps of the function specified in the one or more blocks.

[0237] The embodiments of the present application are described above with reference to the accompanying drawings, but the present application is not limited to the specific embodiments described above, and the specific embodiments described above are merely illustrative, but not restrictive, and those of ordinary skill in the art can make many forms under the inspiration of the present application without departing from the purpose of the present application and the scope protected by the claims, which are all within the protection of the present application.

Claims

1. A knowledge tracking method, characterized in that: include: Obtain contextual information about the exercises and the absolute difficulty of the exercises to generate enhanced exercise representations; Obtain the learner's historical temporal cognitive state, and obtain the overall trend of the cognitive state based on the relationship between the learner's cognitive state and the temporal sequence; Obtain the learner's previous cognitive state and current exercise answer results, and based on the enhanced exercise representation and overall trend of cognitive state, dynamically update the learner's current cognitive state according to the theory of the zone of proximal development; Based on the learner's current cognitive state, the probability of the learner correctly answering the target exercise is predicted.

2. The knowledge tracking method according to claim 1, characterized in that: The step of obtaining contextual information of the exercise and the absolute difficulty of the exercise to generate an enhanced exercise representation includes: Based on the contextual information of the exercises, a learnable knowledge point weight parameter matrix is ​​used to represent the contextual relationship between exercises and knowledge points. Based on the contextual relationship between exercises and knowledge points, the knowledge point weight embedding and the exercise absolute difficulty embedding are input into the multi-layer perceptron to generate exercise embedding that incorporates contextual information. The exercises that incorporate contextual information are embedded in the monotonic self-attention mechanism of the input forgetting mechanism to obtain enhanced exercise representation.

3. The knowledge tracking method according to claim 1, characterized in that: The acquiring of the learner's historical time-series cognitive state and obtaining the overall trend of the cognitive state based on the relationship between the learner's cognitive state and the time sequence include: Implement reversible instance normalization on the learner's historical temporal cognitive state, and use a sliding window with a step size S and a window size L to divide it into blocks to obtain multiple cognitive state data blocks; Linearly map the pre-trained word embeddings in the semantic space of the large language model to a low-dimensional space to obtain reduced-dimensional word embeddings; Based on the multi-head cross-attention mechanism, the dimensionality-reduced word embedding is used to align the cognitive state data blocks across modalities and generate an aligned fusion representation. Sequentially concatenating the task prompt prefix of the large language model with the aligned fusion representation, and inputting the concatenated sequence into the large language model for processing to obtain an output hidden state; wherein the task prompt prefix of the large language model includes data description, cognitive state tracking task requirements, and cognitive state statistical information; A time series portion corresponding to the cognitive state data block is extracted from the output hidden state, the extracted time series portion is processed, and an overall trend of the cognitive state is output.

4. The knowledge tracking method according to claim 1, characterized in that: The acquisition of the learner's previous cognitive state and current exercise answer results, based on the enhanced exercise representation and overall trend of cognitive state, dynamically updates the learner's current cognitive state according to the theory of the zone of proximal development, including: Before answering an exercise, the relative difficulty of the exercise is calculated based on the difference between the enhanced exercise representation and the learner's cognitive state at the previous moment, and the relative difficulty of the exercise is processed through a neural network containing a gated recurrent unit to output a subjective difficulty feeling; During the exercise, a neural network including a gated recurrent unit processes the subjective difficulty perception, the current exercise answer result, and the absolute difficulty of the exercise, and outputs the amount of knowledge acquired; After answering the exercises, the knowledge state indicator is obtained based on the learner's previous cognitive state, subjective difficulty feeling, current exercise answer result and absolute difficulty of the exercise; the learner's previous cognitive state, knowledge state indicator and knowledge acquisition amount are processed through a neural network containing a gated recurrent unit , and obtain the learner’s current cognitive state.

5. The knowledge tracking method according to claim 4, characterized in that: The relative difficulty of the exercises is processed by a neural network including a gated recurrent unit, and the subjective difficulty feeling is output, which is expressed by the following formula: ， in, It is the subjective feeling of difficulty. For the first gate, It is the direct output of the relative difficulty of the exercises after being processed by the neural network; First gate , expressed by the following formula: ， in, for Activation function; is the weight matrix, expressed as , For the real number field dimensional space, is the input layer dimension, is the relative difficulty dimension of the exercises; is the bias matrix, expressed as , For the real number field Dimensional space; The relative difficulty of the exercises; Direct output of the relative difficulty of the exercises after neural network processing , expressed by the following formula: ， in, is a nonlinear activation function; is the weight matrix, expressed as , is the bias matrix, expressed as ; The relative difficulty of the exercises; Relative difficulty of exercises , expressed by the following formula: ， in, For learners Always be aware of your state, for Constantly enhanced problem representation.

6. The knowledge tracking method according to claim 4, characterized in that: The subjective difficulty perception, the current exercise answer result and the absolute difficulty of the exercise are processed by a neural network including a gated recurrent unit, and the knowledge acquisition amount is outputted, which is expressed by the following formula: ， in, is the amount of knowledge acquired, For the second gate, It is the direct output value of knowledge acquisition; Second gate , expressed by the following formula: ， in, for Activation function; is the weight matrix, expressed as , For the real number field dimensional space, It is the subjective difficulty dimension. The dimension of the answer result for the current exercise. is the absolute difficulty dimension of the exercise, is the input layer dimension; It is the subjective feeling of difficulty; The absolute difficulty of the exercise; It is a cascade operation; is the bias matrix, expressed as , For the real number field Dimensional space; Direct output value of knowledge acquisition , expressed by the following formula: ， in, is a nonlinear activation function; is the weight matrix, expressed as It is the subjective feeling of difficulty. is the absolute difficulty of the exercise, For cascade operation, is the bias matrix, expressed as .

7. The knowledge tracking method according to claim 4, characterized in that: The learner's previous cognitive state, knowledge state indicator, and knowledge acquisition amount are processed by a neural network including a gated recurrent unit to obtain the learner's current cognitive state, which is expressed by the following formula: ， in, For learners Always be aware of your state, For learners Always be aware of your state, is the amount of knowledge acquired, is the knowledge status indicator, which is expressed by the following formula: ， in, for Activation function; is the weight matrix, expressed as , For the real number field dimensional space, For learners The overall trend dimension of the cognitive state at each moment, It is the subjective difficulty dimension. The dimension of the answer result for the current exercise. is the absolute difficulty dimension of the exercise, is the input layer dimension; For learners The overall trend of cognitive state at each moment; It is a cascade operation; It is the subjective feeling of difficulty; For learners The answer result at the moment; The absolute difficulty of the exercise; is the bias matrix, expressed as .

8. The knowledge tracking method according to claim 1, characterized in that: The method of predicting the probability of the learner correctly answering the target exercise based on the learner's current cognitive state includes: Based on the learner's current cognitive state and the enhanced problem representation, the ideal probability of the learner correctly answering the target problem is calculated; Based on the guess factor and error factor corresponding to the target exercise, the ideal probability of the learner correctly answering the target exercise is updated to obtain the final probability of the learner correctly answering the target exercise.

9. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instruction is executed by a processor, the steps of the knowledge tracking method described in any one of claims 1 to 8 are implemented.

10. A computer device, characterized in that: include: Memory, used to store computer programs / instructions; A processor is configured to execute the computer program / instructions to implement the steps of the knowledge tracing method according to any one of claims 1 to 8.

Images