A knowledge tracking method and system for extremely long student answer sequences

Abstract

The present invention discloses a knowledge tracking method and system for extremely long student answer sequences, which belongs to the field of knowledge tracking technology, including: obtaining a student answer sequence, performing embedding calculation on the student answer sequence to obtain a multi-dimensional feature vector; calculating the multi-dimensional feature vector through a Poincare hemisphere space attention mechanism to obtain a similarity attention score; obtaining the Euclidean space attention score of the multi-dimensional feature vector, calculating the similarity attention score and the Euclidean space attention score respectively through an attention mechanism with a linear compensation coefficient to obtain a compensated attention score; obtaining an implicit output based on the compensated attention score and a value vector in a multi-head aggregation attention mechanism; performing prediction based on the implicit output to obtain a student answer prediction; through the above technical solution, the present invention can capture the long-term changes in students' knowledge status and track the students' knowledge learning process more comprehensively and accurately.

Description

Technical Field

[0001] The present invention belongs to the technical field of knowledge tracking, and in particular relates to a knowledge tracking method and system for extremely long student answer sequences. Background Art

[0002] Knowledge tracing involves leveraging students' past learning interaction data to predict their responses to future questions. This relies on accurately modeling the student's knowledge mastery state. With the rapid development of attention mechanisms, numerous attention-based knowledge tracing models have emerged, effectively modeling the student's knowledge mastery state. The Self-Attentive Knowledge Tracing (SAKT) model uses an embedding layer to embed students' past answers into vectors. It then employs a self-attention mechanism to extract the relationship between questions and knowledge points from these answers, thereby modeling the student's knowledge mastery state. The Separated Self-Attention Neural Knowledge Tracing (SAINT) model employs an encoder-decoder architecture, feeding the student's question and answer embedding sequences into an encoder and decoder, respectively. The encoder uses a self-attention mechanism to extract features between the question embedding sequence, while the decoder alternates between a self-attention mechanism and an encoder-decoder attention layer on the student's answer embedding sequence. This input separation facilitates stacking attention layers multiple times to model the student's knowledge mastery state. The knowledge tracking model based on the point-wise attention mechanism (simpleKT) introduces a specific problem difficulty factor with learnable parameters and uses the point-wise attention mechanism to capture the features between students' questions and related knowledge points. It can effectively model students' knowledge mastery status from the perspective of problem difficulty.

[0003] While the three best attention-based knowledge tracking models mentioned above can effectively model students' knowledge mastery, they only target student sequences with a length of 200. When the number of student answers exceeds 200, the student answer sequence is cut into sequences with a maximum length of 200. For example, if the student answer sequence is 1000, the student answer sequence is cut into five sequences of 200. In educational scenarios, this approach prevents the knowledge tracking model from capturing long-term changes in the student's knowledge state when modeling the student's knowledge state, resulting in a more comprehensive and accurate tracking of the student's answers, i.e., the knowledge learning process. Currently, no method has been developed for the case where the student answer sequence exceeds 200 in length in the knowledge tracking task. Summary of the Invention

[0004] In order to solve the above technical problems, the present invention proposes a knowledge tracking method and system for extremely long student answer sequences to solve the problems existing in the above-mentioned prior art.

[0005] To achieve the above objectives, the present invention provides a knowledge tracking method for extremely long student answer sequences, comprising:

[0006] Obtaining a student answer sequence, performing embedding calculation on the student answer sequence to obtain a multi-dimensional feature vector, wherein the multi-dimensional feature vector includes a feature vector of a knowledge point and a feature vector of the question answer;

[0007] The multi-dimensional feature vector is calculated through the Poincare hemispheric spatial attention mechanism to obtain the similarity attention score;

[0008] Obtain the Euclidean space attention score of the multi-dimensional feature vector, calculate the similarity attention score and the Euclidean space attention score respectively through the attention mechanism of the linear compensation coefficient, and obtain the compensated attention score;

[0009] Calculating the compensation attention score and the value vector in the multi-head aggregate attention mechanism through the multi-head aggregate attention mechanism to obtain an implicit output;

[0010] Make predictions based on the implicit output to get the predicted results of the students’ answers.

[0011] Optionally, the student answer sequence includes questions, knowledge points associated with the questions, answers to the questions and time steps of the answers.

[0012] Optionally, the process of obtaining the multi-dimensional feature vector includes:

[0013] Encode the knowledge points associated with the question and the answer to the question to obtain the corresponding one-hot encoding, perform weighted calculation on the one-hot encoding to obtain the implicit representation of the knowledge points and the implicit representation of the answer to the question, obtain and calculate the implicit representation of the knowledge points based on the difficulty factor of the question and the embedding of the knowledge points associated with the question to obtain the feature vector of the knowledge points, and calculate the feature vector of the answer to the question based on the implicit representation of the knowledge points and the implicit representation of the answer to the question.

[0014] Optionally, the process of calculating the feature vector of a knowledge point includes:

[0015] In the Poincaré hemisphere spatial attention mechanism, a query vector, a key vector, and a value vector are generated based on a multi-dimensional feature vector, and the query vector and the key vector are mapped from the Euclidean space to the Poincaré hemisphere space. The mapped query vector and the key vector are calculated to obtain a similarity attention score.

[0016] Optionally, the process of calculating the mapped query vector and key vector includes:

[0017] Construct an inequality:

[0018]

[0019] Where Ψ represents the mapping function that maps parameters from Euclidean space to Poincare hemisphere space, Ψ(Q) [：-1] represents the mapping result of the first d-1 dimensions of the query vector Q, Ψ(K) [：-1] represents the mapping result of the first d-1 dimensions of the key vector K, Ψ(Q) d Represents the mapping result of the last dimension of the query vector Q, Ψ(K) d represents the mapping result of the last dimension of the key vector K, and α represents the mapping coefficient from the Euclidean space to the Poincare hemisphere space;

[0020] When the inequality holds:

[0021]

[0022] Where exp(·) represents the exponential function and γ is the scaling coefficient;

[0023] When the inequality does not hold:

[0024]

[0025] Among them, u and v represent different auxiliary parameters.

[0026] Optionally, the process of calculating the similarity attention score and the Euclidean space attention score separately includes:

[0027] In the attention mechanism of linear compensation coefficient, the linear compensation coefficient matrix B and the causal relationship marker matrix C are introduced into the attention score S to obtain the compensated attention score g(S, B, C). The attention score S includes the similarity attention score and the Euclidean space attention score, where:

[0028]

[0029] Among them, Softmax() represents the Softmax function.

[0030] Optionally, the implicit output acquisition process includes:

[0031] The compensation attention scores are used as weights of the value vectors in the multi-head aggregation attention mechanism through the multi-head aggregation attention mechanism, the value vectors in the multi-head aggregation attention mechanism are weightedly aggregated, the weighted aggregation results are spliced, and the splicing results are weighted calculated to obtain implicit output;

[0032] Among them, the compensated attention score includes the compensated similarity attention score and the compensated Euclidean space attention score, wherein the compensated similarity attention score is the score obtained by calculating the similarity attention score through the attention mechanism of the linear compensation coefficient, and the compensated Euclidean space attention score is the score obtained by calculating the Euclidean space attention score through the attention mechanism of the linear compensation coefficient.

[0033] Optionally, the process of making predictions based on implicit outputs includes:

[0034] The implicit output is predicted through a fully connected neural network to obtain the predicted student answers.

[0035] Optionally, the process of obtaining the Euclidean spatial attention score also includes:

[0036] The feature vector of the knowledge point is processed through the dot product attention mechanism to obtain the Euclidean space attention score.

[0037] On the other hand, the present invention also provides a knowledge tracking system for extremely long student answer sequences, which is used to execute the above method.

[0038] Compared with the prior art, the present invention has the following advantages and technical effects:

[0039] 1. By embedding and mapping the student answer sequence in Euclidean space into the Poincare hemisphere space, we can effectively capture the complex hierarchical relationship between student questions and related knowledge points, accurately model the student's knowledge mastery status, and thus provide a more comprehensive and accurate tracking and interpretation of the student's answer process. Based on its effective student answer prediction, we can help students improve their knowledge mastery ability.

[0040] 2. It solves the problem of being unable to handle extremely long student answer sequences in current knowledge tracking tasks, effectively models students' long-term knowledge mastery status, meets the application needs of actual educational scenarios, and can assist teachers in improving teaching quality and realizing personalized teaching.

[0041] 3. The multi-head aggregate attention mechanism based on the point product attention mechanism and the Poincaré hemisphere space attention mechanism can dynamically learn feature representations in Euclidean space and Poincaré hemisphere space, and has strong versatility and generalization. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0043] Figure 1Schematic diagram of a method flow in an embodiment of the present invention. DETAILED DESCRIPTION

[0044] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0045] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0046] This paper proposes a knowledge tracking method for extremely long student answer sequences, which mainly includes five steps:

[0047] Step 1: Embed the student answer sequence into a multi-dimensional feature vector, which is a multi-dimensional feature vector representation.

[0048] Step 2: Use the Poincare hemisphere spatial attention mechanism to map the multi-dimensional feature vector of the student's answer sequence from Euclidean space to Poincare hemisphere space, and calculate the complex hierarchical relationship features between the student's questions and the related knowledge points.

[0049] Step 3: Use the attention mechanism with a linear compensation coefficient to enhance the attention mechanism’s ability to capture the features of extremely long student answer sequence lengths.

[0050] Step 4: Use the multi-head aggregate attention mechanism to aggregate the results of the point product attention mechanism and the Poincaré hemisphere space attention mechanism to dynamically learn the feature representations in Euclidean space and Poincaré hemisphere space, thereby enhancing the generalization ability of capturing the features of extremely long student answer sequences.

[0051] Step 5: Use a two-layer fully connected neural network to minimize the binary cross entropy loss function to predict students' answers.

[0052] The technical solution of the present invention uses various types of attention mechanisms to perform feature extraction and task execution in different aspects. The traditional self-attention mechanism is now described:

[0053] The self-attention mechanism is a widely used network model in deep learning models, particularly effective when processing sequential data. It allows each element of an input sequence to be compared with other elements in the sequence to compute a representation of the sequence. The core idea of ​​the self-attention mechanism is to enable the model to focus on information at different positions in the input sequence and generate an output at the current position based on this information.

[0054] The self-attention mechanism works by calculating the correlation between the query, key, and value. Specifically, for each element in the input sequence, the model calculates its similarity with all other elements and assigns weights based on these similarities. Ultimately, the output at each position is the weighted sum of the values ​​at all positions, with the weight determined by the relevance of the query and key. This mechanism enables the model to capture complex dependencies within a sequence, and it performs particularly well when dealing with long-range temporal dependencies.

[0055] Among them, the query vector Q, key vector K and value vector V of the traditional self-attention mechanism are obtained by linear transformation of the input data X. The specific calculation process is as follows: Q = XW Q , K=XW K 、V=XW V , where W Q 、W K 、W V They are the weight matrices used to generate Q, K, and V, respectively. These matrices are all trainable.

[0056] The advantage of its self-attention mechanism is that it can capture long-range dependencies: the self-attention mechanism allows the model to directly refer to other positions in the entire sequence when calculating the representation of a position in the sequence, thereby capturing long-range dependencies. Dynamic weights: The self-attention mechanism generates different weights for each input position. These weights are dynamically calculated based on the current content of the input sequence.

[0057] The following specific description is given for the above technical solution of the present invention:

[0058] Step 1: Embed the student answer sequence into a multi-dimensional feature vector. The multi-dimensional feature vector is a multi-dimensional feature vector representation, specifically including the following:

[0059] Collect students' answer information and form a student answer sequence based on the answer information, where the student answer sequence is {q, c, r, t}, where q is the question; c is the knowledge point associated with the question; r is the answer to the question, 1 indicates a correct answer, and 0 indicates an incorrect answer; t is the time step of the answer, and the question q and the knowledge point c associated with the question are text data.

[0060] First calculate the unique hot encoding of the knowledge point c and the answer r to the question at time step t and

[0061] Then calculate the implicit representation of the knowledge point and the implicit representation of the answer to the question By introducing the problem difficulty factor Embedding of knowledge points associated with the question Calculate the feature vector (embedded representation) of the enhanced knowledge point and the feature vector (embedding representation) of the question answer contained in Among them, the problem difficulty factor Embedding of knowledge points associated with the question All are learnable parameters, which are determined by model training. represents bitwise addition, and ⊙ represents bitwise multiplication.

[0062] Step 2: Use the Poincare hemispheric spatial attention mechanism to map the multi-dimensional feature vector of the student's answer sequence from Euclidean space to Poincare hemispheric space, and calculate the complex hierarchical relationship features between the student's questions and the associated knowledge points, specifically including the following:

[0063] First, the feature vector of the student's answer sequence is used as the input of the attention mechanism. Specifically, the query vector Q=x in the Calais hemispheric spatial attention mechanism t+1 , key vector K={x1,…,x t} and value vector V = {y1, ..., y t}. The subscript t indicates that the parameter corresponds to time step t.

[0064] Map the query vector, key vector, and value vector from the Euclidean space to the Poincare hemisphere space. In the mapping rule, X is used as a vector variable, including the query vector, key vector, and value vector. d Represents the last dimension of vector X, where the last dimension represents the last column in the matrix and the last element in the vector. The dimension corresponds to the column element in the matrix or an element in the vector. [：-1] represents the first d-1 dimensions of vector X, where d represents the total number of dimensions and the first d-1 dimensions represent all data from the 1st dimension to the d-1th dimension. The mapping rule is:

[0065]

[0066] Where α is the mapping coefficient, exp(·) is the exponential function, and Ψ represents the mapping function that maps parameters from Euclidean space to Poincare hemisphere space.

[0067] Then map the query vector Q and key vector K in the attention mechanism to the Poincare hemisphere space:

[0068]

[0069] Finally, the similarity attention score S of the hierarchical perception in the Poincare hemisphere space is calculated hyp , that is, when the inequality:

[0070]

[0071] When it was established,

[0072]

[0073] When the inequality does not hold, then:

[0074]

[0075] Among them, γ is the scale transformation coefficient, and u and v represent different auxiliary parameters.

[0076] Step 3: Use the attention mechanism with a linear compensation coefficient to enhance the attention mechanism's ability to capture the characteristics of extremely long student answer sequences. Specifically, the following steps are included:

[0077] The data processed by the linear compensation coefficient attention mechanism includes the similarity attention score in the Poincare hemisphere space output by the Poincare hemisphere space attention mechanism, and also includes the Euclidean space attention score in the Euclidean space output by the dot product attention mechanism;

[0078] The multi-dimensional feature vector of the student's answer sequence is processed using the dot product attention mechanism to obtain the Euclidean space attention score. In Euclidean space, the attention score of the dot product attention mechanism is the Euclidean space attention score S dot The calculation is as follows:

[0079]

[0080] Where d is the dimension of the key vector, and the superscript T represents the matrix transpose. The query vector, key vector, and value vector of the dot product attention mechanism are the same as those of the Calais hemispheric spatial attention mechanism before mapping, that is, the query vector Q = x t+1 , key vector K={x1,…,x t} and value vector V = {y1, ..., y t}.

[0081] In the attention mechanism of linear compensation coefficient, by introducing the linear compensation coefficient matrix B and the causal relationship label matrix C into the attention score S, the attention score based on the linear compensation coefficient, namely the compensated attention score g(S, B, C), is obtained:

[0082]

[0083] Among them, the attention score S includes the Euclidean space attention score S dot and similarity attention score S hyp , the compensated attention score g(S, B, C) includes the compensated similarity attention score g(S hyp , B, C) and compensated Euclidean spatial attention score g(Sdot , B, C), respectively correspond to the similarity attention score and the Euclidean space attention score, which are the score results after the above calculation through the attention mechanism of the linear compensation coefficient.

[0084] The value of each element of the linear compensation coefficient matrix is:

[0085]

[0086] Among them, i represents the attention head number of the multi-head aggregate attention mechanism, H represents the total number of heads in the multi-head aggregate attention mechanism, and b mn Represents the element in the mth row and nth column of the linear compensation coefficient matrix, where m represents the number of rows and n represents the number of columns.

[0087] Each element of the causal relationship sign matrix is:

[0088]

[0089] Among them, c mn Represents the element in the mth row and nth column of the causal relationship sign matrix, where m represents the number of rows and n represents the number of columns.

[0090] Step 4: Use a multi-head aggregate attention mechanism to aggregate the results of the dot product attention mechanism and the Poincaré hemisphere space attention mechanism to dynamically learn feature representations in Euclidean space and Poincaré hemisphere space, thereby enhancing the generalization ability of capturing features of very long student answer sequences. Specifically, this includes the following:

[0091] The multi-head aggregate attention mechanism mainly performs weighted aggregation on the compensated score results of the Euclidean space attention score in the Euclidean space and the similarity attention score in the Poincare hemisphere space to obtain the implicit output, as follows:

[0092]

[0093] Among them, V represents the value vector in the multi-head aggregation attention mechanism. The value vectors of the above three attention mechanisms are the same, V = {y1, ..., y t}, Concat is a concatenation operation, h t+1 is the output of the hidden layer in the multi-head aggregate attention mechanism, W h represents the weight matrix of the hidden layer, H hyp Represents the product of the similarity attention score and the value vector, H dot It represents the product of the Euclidean spatial attention score and the value vector. The superscript i represents the attention head number of the dot product attention mechanism in the multi-head aggregation attention mechanism. The superscript j represents the attention head number of the Poincare hemisphere spatial attention mechanism in the multi-head aggregation attention mechanism.

[0094] Step 5: Use a two-layer fully connected neural network to minimize the binary cross entropy loss function to predict student answers. Specifically, it includes the following:

[0095] A two-layer fully connected neural network is used to predict the student's answer. Specifically, the hidden layer output of the multi-head aggregate attention mechanism is used as the input of the fully connected neural network to calculate the binary cross entropy loss. Under the minimum conditions, the student answer prediction situation is the student answer situation predicted by the model for questions and knowledge points for:

[0096]

[0097] Among them, r t+1 represents the student's actual answer, taking values ​​of 0 and 1. It takes a value of 1 when the student answers correctly and a value of 0 when the student answers incorrectly. σ and φ refer to the Sigmoid function and the ReLU function, respectively. All W are learnable parameter matrices in the neural network, where W1 and W2 represent the weight matrices corresponding to the first and second hidden layers of the neural network, and b1 and b2 represent the bias matrices corresponding to the first and second hidden layers of the neural network.

[0098] By using the above technical solution, the present invention has the following advantages:

[0099] 1. By embedding and mapping the student answer sequence in Euclidean space into the Poincare hemisphere space, we can effectively capture the complex hierarchical relationship between student questions and related knowledge points, accurately model the student's knowledge mastery status, and thus provide a more comprehensive and accurate tracking and interpretation of the student's answer process. Based on its effective student answer prediction, we can help students improve their knowledge mastery ability.

[0100] 2. It solves the problem of being unable to handle extremely long student answer sequences in current knowledge tracking tasks, effectively models students' long-term knowledge mastery status, meets the application needs of actual educational scenarios, and can assist teachers in improving teaching quality and realizing personalized teaching.

[0101] 3. The multi-head aggregate attention mechanism based on the point product attention mechanism and the Poincaré hemisphere space attention mechanism can dynamically learn feature representations in Euclidean space and Poincaré hemisphere space, and has strong versatility and generalization.

[0102] In order to better achieve the above technical objectives, the present invention also provides a knowledge tracking system for extremely long student answer sequences to implement the above knowledge tracking method for extremely long student answer sequences, which specifically includes five modules:

[0103] The first module is the student answer sequence embedding module, which is used to obtain the student answer sequence and embed the student answer sequence into a multi-dimensional feature vector.

[0104] The second module is the Poincare hemisphere spatial attention mechanism module, which uses the Poincare hemisphere spatial attention mechanism to map the multi-dimensional feature vector of the student's answer sequence from Euclidean space to Poincare hemisphere space, and calculate the complex hierarchical relationship features between student questions and related knowledge points.

[0105] The third module is an attention mechanism module based on linear compensation coefficients, which is used to use the attention mechanism with linear compensation coefficients. The linear compensation coefficients are used to enhance the attention mechanism's ability to capture the features of extremely long student answer sequence lengths.

[0106] The fourth module is the multi-head aggregate attention mechanism module, which is used to use the multi-head aggregate attention mechanism to dynamically learn the feature representation in Euclidean space and Poincaré hemisphere space by aggregating the results of the point product attention mechanism and the Poincaré hemisphere space attention mechanism, thereby enhancing the generalization ability of capturing the features of extremely long student answer sequences.

[0107] The fifth module is the student answer prediction module, which uses a two-layer fully connected neural network to minimize the binary cross entropy loss function to predict the student answer situation.

[0108] The system technical solution in this embodiment corresponds to the technical solution of the above invention and will not be described in detail here.

[0109] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A knowledge tracking method for extremely long student answer sequences, characterized by: include: Obtaining a student answer sequence, performing embedding calculation on the student answer sequence to obtain a multi-dimensional feature vector, wherein the multi-dimensional feature vector includes a feature vector of a knowledge point and a feature vector of the question answer; The multi-dimensional feature vector is calculated through the Poincare hemispheric spatial attention mechanism to obtain the similarity attention score; Obtain the Euclidean space attention score of the multi-dimensional feature vector, calculate the similarity attention score and the Euclidean space attention score respectively through the attention mechanism of the linear compensation coefficient, and obtain the compensated attention score; Calculating the compensation attention score and the value vector in the multi-head aggregate attention mechanism through the multi-head aggregate attention mechanism to obtain an implicit output; Make predictions based on implicit outputs to get predictions of students’ answers; The process of calculating the similarity attention score and the Euclidean space attention score respectively includes: In the attention mechanism of linear compensation coefficient, the linear compensation coefficient matrix B and the causal relationship marker matrix C are introduced into the attention score S to obtain the compensated attention score g(S, B, C). The attention score S includes the similarity attention score and the Euclidean space attention score, where: Among them, Softmax() represents the Softmax function; The process of obtaining implicit output includes: The compensation attention scores are used as weights of the value vectors in the multi-head aggregation attention mechanism through the multi-head aggregation attention mechanism, the value vectors in the multi-head aggregation attention mechanism are weightedly aggregated, the weighted aggregation results are spliced, and the splicing results are weighted calculated to obtain implicit output; Among them, the compensated attention score includes the compensated similarity attention score and the compensated Euclidean space attention score, wherein the compensated similarity attention score is the score obtained by calculating the similarity attention score through the attention mechanism of the linear compensation coefficient, and the compensated Euclidean space attention score is the score obtained by calculating the Euclidean space attention score through the attention mechanism of the linear compensation coefficient; The process of making predictions based on implicit outputs involves: The implicit output is predicted through a fully connected neural network to obtain the predicted student answers.

2. The method according to claim 1, characterized in that The student answer sequence includes questions, knowledge points associated with the questions, answers to the questions and time steps of the answers.

3. The method according to claim 2, characterized in that The process of obtaining the multi-dimensional feature vector includes: Encode the knowledge points associated with the question and the answer to the question to obtain the corresponding one-hot encoding, perform weighted calculations on the one-hot encodings to obtain the implicit representation of the knowledge points and the implicit representation of the answer to the question, obtain and calculate the implicit representation of the knowledge points based on the question difficulty factor and the embedding of the knowledge points associated with the question to obtain the feature vector of the knowledge points, perform bitwise addition calculations on the implicit representation of the knowledge points and the implicit representation of the answer to the question to obtain the feature vector of the answer to the question.

4. The method according to claim 1, wherein The process of calculating the feature vector of a knowledge point includes: In the Poincaré hemisphere spatial attention mechanism, a query vector, a key vector, and a value vector are generated based on a multi-dimensional feature vector, and the query vector and the key vector are mapped from the Euclidean space to the Poincaré hemisphere space. The mapped query vector and the key vector are calculated to obtain a similarity attention score.

5. The method according to claim 4, characterized in that The process of calculating the mapped query vector and key vector includes: Construct an inequality: Where Ψ represents the mapping function that maps parameters from Euclidean space to Poincare hemisphere space, Ψ(Q) [：-1] represents the mapping result of the first d-1 dimensions of the query vector Q, Ψ(K) [：-1] represents the mapping result of the first d-1 dimensions of the key vector K, Ψ(Q) d Represents the mapping result of the last dimension of the query vector Q, Ψ(K) d represents the mapping result of the last dimension of the key vector K, α represents the mapping coefficient of the Euclidean space to the Poincare hemisphere space, and d represents the total number of dimensions; When the inequality holds: Where exp(·) represents the exponential function and γ is the scaling coefficient; When the inequality does not hold: Among them, u and v represent different auxiliary parameters.

6. The method according to claim 1, characterized in that The process of obtaining the Euclidean spatial attention score also includes: The feature vector of the knowledge point is processed through the dot product attention mechanism to obtain the Euclidean space attention score.

7. A knowledge tracking system for extremely long student answer sequences, characterized by: Used to perform the method according to any one of claims 1 to 6.

Images