CN115545155A

CN115545155A - Multi-level intelligent cognitive tracking method and system, storage medium and terminal

Info

Publication number: CN115545155A
Application number: CN202211150608.3A
Authority: CN
Inventors: 王志锋; 李璐瑶; 左明章; 王继新; 罗恒; 闵秋莎; 董石; 田元; 夏丹; 龙陶陶
Original assignee: Central China Normal University
Current assignee: Central China Normal University
Priority date: 2022-09-21
Filing date: 2022-09-21
Publication date: 2022-12-30

Abstract

The invention belongs to the technical field of individual learning and discloses a multi-level intelligent cognitive tracking method, a multi-level intelligent cognitive tracking system, a storage medium and a terminal, wherein the method comprises the following steps: introducing education target classification in the brucm cognition field, constructing a test question knowledge cognition tensor TKC, collecting learning resources and answering data of a learner, and generating a time sequence answering pair sequence of the learner; a multi-partition attribute cognitive diagnosis method is introduced, and a cognitive level mining model is constructed by combining a deep neural network; and sequencing and coding the cognitive level mining results of the learner to obtain deep characterization features, and constructing a multi-level intelligent cognitive tracking model by combining a self-attention mechanism to further predict the answering performance of the learner on the test question. The invention is beneficial to accurately and finely modeling the whole knowledge structure and the specific hierarchical level of the learner, thereby promoting the personalized learning of the learner and providing a new thought for mining and tracking the cognitive state and the hierarchical level of the learner in the online learning platform.

Description

Multi-level intelligent cognitive tracking method and system, storage medium and terminal

Technical Field

The invention belongs to the technical field of personalized learning, and particularly relates to a multi-level intelligent cognitive tracking method and system, a storage medium and a terminal.

Background

At present, the internet information technology taking intellectualization as a core is developed at a high speed, and the information technology innovation in the education field is standing on the air port developed in the era. In the epidemic situation period of 2020, schools and various network platforms at all levels actively respond to the national 'no-stop-class' call, and the teaching activities are carried out in the modes of live cloud and class cloud, so that the innovation of an 'internet plus' education mode, a learning mode and an education service mode is promoted to a certain extent, and the online transformation of schools and classes is accelerated. Among them, the intelligent evaluation technology represented by learning diagnosis becomes one of the important development contents of education informatization: by constructing the picture of the learner, developing and analyzing the learning behavior, reasonably planning the learning path, efficiently recommending learning resources, finally helping the learner to customize a learning plan, make up for omissions, carry out targeted strengthening exercises and construct a learning and testing closed loop through learning intervention, and really realizing the teaching by factors and personalized learning.

The intelligent cognitive tracking technology is a key technology for portraying the learning behavior of learners. The technology can extract effective information from a response sequence of a learner and a test question knowledge association matrix, mark knowledge examination in the test question and describe the whole test structure, dynamically model and track the cognitive level of the learner according to potential trait differences and test question characteristics of the learner, predict the answer performance of the learner at a certain future moment, mine the cognitive information implied after the learner scores the back, analyze the advantages and the defects of the learner and finally give a targeted and personalized learning suggestion.

At present, intelligent cognitive tracking models proposed by researchers in the field are mainly divided into two categories: static intelligent cognitive tracking based on statistical learning and dynamic intelligent cognitive tracking based on deep learning. Static intelligent cognitive tracking methods based on statistical learning, such as Deterministic input noise And gate (DINA) models, typically model the learner's cognitive level as a binary variable other than 0, i.e., 1, indicating whether the learner is "mastered" or "not mastered" about a certain knowledge, and then continuously update this variable using trial attribute pattern Q matrix And EM algorithm to model the learner's learning process. The structure and the process of the method are clear, the size of the data set has low influence on the performance of the learner, but the method has some defects, for example, the method can only statically diagnose the overall cognitive level of the learner according to the answering condition of the learner within a certain period of time, and cannot track the specific knowledge point at a specific moment. With the development of artificial intelligence and deep learning technology, a dynamic intelligent cognitive tracking method based on deep learning is rapidly developed, which mainly takes a learner response sequence as input, utilizes a recurrent neural network to model the learner response process, uses a hidden state vector to represent the integral cognitive condition of the learner, and continuously updates the hidden state according to the learner response condition. The dynamic intelligent cognitive tracking method based on deep learning is generally high in accuracy and does not depend on excessive prior knowledge. However, due to the inherently less interpretative nature of deep learning techniques, such methods often fail to convincingly interpret the results they produce, thereby limiting their usefulness.

How to effectively utilize professional knowledge in the education field, and when the learning behavior of a learner is represented, the fact that the mastering condition of the learner on each knowledge point is not extremely 'non-0 or 1' is taken into consideration, so that a more accurate and more precise intelligent cognitive tracking model is constructed; how to capture the internal influence among different knowledge points, solve the long-distance dependence problem, reduce the complexity of the model and reduce the computing power required by the model; how to effectively combine the process of the change of the learner cognitive level along with the time with the spatial change process of the mutual influence of the knowledge points becomes the difficult problem of intelligent cognitive tracking.

Through the above analysis, the problems and defects of the prior art are as follows:

(1) The traditional intelligent cognitive tracking method simplifies the mastering condition of a learner on the knowledge points, and extremely uses a variable which is not 0 or 1 to represent the mastering condition of the learner on each knowledge point, which is not consistent with the actual learning process, so that the result granularity of the model is coarse;

(2) The traditional intelligent cognitive tracking method has a single modeling angle, ignores internal influence among knowledge points, cannot 'memorize' information at longer distance, has higher complexity of the model and has higher requirement on computing power;

(3) The traditional intelligent cognitive tracking method usually does not consider the combination of the process of the change of the cognitive level of a learner along with time and the spatial change process of the mutual influence of knowledge points, so that the model has poor interpretability and cannot be well applied to the actual learning process.

Disclosure of Invention

Aiming at the problems in the prior art, the invention provides a multi-level intelligent cognitive tracking method, a multi-level intelligent cognitive tracking system, a storage medium and a terminal.

The invention is realized in such a way, and the multi-level intelligent cognitive tracking method is characterized by comprising the following steps:

step one, introducing education target classification in the brucm Cognition field, constructing a corrected Test-Knowledge Cognition tensor TKC (Test-Knowledge-Cognition), collecting learning resources and answering data of learners, performing feature extraction, and generating a time sequence Test-Knowledge point-brucm level-answering condition answering pair sequence for each learner;

step two, introducing a multi-attribute cognitive diagnosis method based on a time sequence answer pair sequence of the learner and the test question knowledge cognitive tensor TKC integrated with the education target classification in the brucm cognitive field, and constructing a cognitive level mining model by combining a deep neural network to perform multi-level intelligent cognitive mining on the learner;

sequencing and coding learner cognitive level vectors obtained by intelligent cognitive mining to obtain deep characterization features of the learners, introducing a deep self-attention network, and constructing a multi-level intelligent cognitive tracking model; and predicting the answering performance of the learner on the test question according to the tracked learner cognitive level vector and the test question knowledge cognitive tensor TKC.

Further, the first step comprises the following steps:

(1.1) constructing a learning resource set and a learner historical learning data set:

S＝{s ₁ ,s ₂ ,…,s _M }

T＝{t ₁ ,t ₂ ,…,t _N }

K＝{k ₁ ,k ₂ ,…,k _L }

wherein S is a learner set, M is the number of learners, T is a test question set, N is the number of test questions, K is a knowledge point set, and L is the number of knowledge points;

(1.2) starting from a traditional test question knowledge point representation Q matrix, dividing the knowledge point mastery of learners into six cognitive levels, and obtaining a specific and uniform definition of an investigation level facing each knowledge point on a cognitive layer, so as to construct a test question knowledge cognition tensor TKC integrating the education target classification of the Brume cognitive field, wherein the cognitive level range is 0-6, and 0: not mastered; 1: knowing; 2: comprehending; 3: application; 4: analyzing; 5: synthesizing; 6: evaluating;

(1.3) constructing a learner answer matrix R to record the historical answer results of the learner:

wherein r is _mn =0 denotes wrong answer to learner m on test question n, r _mn =1 indicates that the learner m answers correctly on the test question n;

(1.4) processing the learner response record into a time sequence response pair sequence R of learner-test question-knowledge point-Brumm level-response question situation according to the knowledge point corresponding to the learner response question at each moment and the investigation level thereof _s ：

R _s ＝{"user_id":m,

"test_id":n,

"score″:x,

"knowledge_code":l,

"knowledge_level":c}

Wherein M is learner ID, and the value range is that M is more than or equal to 0 and less than or equal to M; n is the test question ID, and the value range is that N is more than or equal to 0 and less than or equal to N; x is the score of the learner m on the test question n, and the value range is 0/1; l is the knowledge point ID of the examination question n investigation, the value range is that L is more than or equal to 0 and less than or equal to L, and when more than one knowledge point of the examination question n investigation is available, L can be expressed as the set of the investigation knowledge point ID; c is the level of the examination question n investigation knowledge point l, the value range is more than or equal to 1 and less than or equal to 6, when l is the set of the investigation knowledge point ID, c can also be expressed as the set of the investigation knowledge point level and respectively corresponds to each dimension in l.

Further, the test question knowledge cognition tensor TKC is as follows:

(t _n ,k _l ,c _i )

wherein, the third dimension i =0 represents the test question n not expecting knowledge point l,1 ≦ i ≦ 6 represents the test question n expecting knowledge points l to c _i And (4) horizontal.

Further, the step two of constructing the cognitive level mining model, and the multi-level intelligent cognitive mining on the learner specifically comprises the following steps:

(2.1) modeling learner factors, including a cognitive level vector F ^s ：

h ^s ＝x ^s ×A

F ^s ＝f _classifier (h ^s )

Wherein x is ^s Is one-hot vector of learner, A is proficiency matrix of all learners, and proficiency vector h of learner can be obtained by multiplying ^s ，f _classfier For the classification function, a learner cognitive level vector F is obtained ^s ；

(2.2) modeling test question factors, wherein the test question factors comprise basic knowledge point relevancy vectors F ^kn And other optional factors F ^other Error parameters and guess parameters of test questions are included:

F ^kn ＝x ^e ×TKC

wherein x is ^e Is a one-hot vector of the test question, TKC is the knowledge cognition tensor of the test question, h ^g Guessing parameters of test questions to represent the probability that the learner answers correctly when the cognitive level of the knowledge points of the test question investigation is not reached, h ^s The error parameters of the test questions represent the probability of wrong answers when the learners reach the cognitive level of the knowledge points examined by the test questions, and B, C is a parameter matrix;

the parameter matrix A, B, C is trained by data learning to obtain: x is to be ^s 、x ^e 、x ^e H obtained by multiplying by a parameter matrix A, B, C respectively ^s 、h ^g 、h ^1-g-s Substituting the project reaction function in the step (2.3) to obtain a prediction result of the cognitive level mining model for the learner to answer, calculating the cross entropy between the prediction result and the actual score of the learner by using the loss function defined in the step (2.4), and training a parameter matrix according to the loss function value and the gradient descent rule;

(2.3) constructing a cognitive level mining model, and predicting a project response function of a learner to answer based on a multi-component attribute cognitive diagnosis method, wherein the project response function comprises the following steps:

training the obtained prediction result by utilizing a deep neural network:

f ₁ ＝φ(W ₁ ×x ^T +b ₁ )

f ₂ ＝φ(W ₂ ×f ₁ +b ₂ )

y＝φ(W ₃ ×f ₂ +b ₃ )

wherein k is _l In order to be a knowledge point ID,

for examination questions at knowledge point k _l The degree of correlation of (a) above,

for learners at knowledge point k _l Upper cognitive level, W ₁ 、b ₁ Are respectively linear regression functions f ₁ Weight coefficient and bias coefficient of (1), W ₂ 、b ₂ Are respectively linear regression functions f ₂ Weight coefficient and bias coefficient of (1), W ₃ 、b ₃ Respectively, weight coefficients and bias coefficients in the linear regression function y.

(2.4) calculating a loss function of the model, the loss function being a cross entropy between the output y and the true label r:

loss _CLMM ＝-∑(r _i logy _i +(1-r _i )log(1-y _i ))

wherein y is a predicted value, and r is a real score; after training, F corresponding to learner ^s The mining result of the learner is obtained, each dimension corresponds to the cognitive level of the learner on the knowledge point, and the cognitive level range is 0-6 (0: not mastered; 1: known; 2: comprehended; 3: applied; 4: analyzed; 5: synthesized; 6: evaluated);

(2.5) updating all weight coefficients (W) in the model according to the loss function value and gradient descent rule obtained in (2.4) ₁ 、W ₂ 、W ₃ ) And bias coefficient (b) ₁ 、b ₂ 、b ₃ )。

Further, the third step includes:

(3.1) constructing a time sequence of the cognitive level of each learner in the brucm education target by using a multi-level cognitive level vector obtained by a cognitive level mining model and combining the answering time information of the learners;

(3.2) coding the cognitive level time sequence of the learner, wherein if the cognitive level time sequence vector of the learner relates to M knowledge points, the corresponding coding length is M-dimension; for a certain knowledge point k, if the learner's cognitive level is a, a is more than or equal to 0 and less than or equal to 6, the value corresponding to the k-th dimension of the code is a + a × M, and thus, the deep characterization feature of each learner's cognitive level vector can be obtained;

(3.3) introducing a deep Self-attention network Transformer, and constructing a multi-level intelligent cognitive tracking model based on the Transformer, wherein the multi-level intelligent cognitive tracking model based on the Transformer comprises an Embedding layer, a Self-attention layer, a Feed forward layer and a Prediction layer;

(3.4) the tracking model performs back propagation based on a loss function formed by a prediction classification and a real label;

(3.5) updating all weight coefficients (W) in the model according to the loss function value and gradient descent rule obtained in (3.4) ⁽¹⁾ 、W ⁽²⁾ ) Bias coefficient (b) ⁽¹⁾ 、b ⁽²⁾ ) And some other parameters (W) ^a 、W ^p 、W ^k 、W ^Q 、W ^K 、W ^V )。

Further, the cognitive level time sequence comprises information including a cognitive level time sequence vector of the learner, a knowledge point ID corresponding to each dimension in the vector and a total length of the time sequence vector;

furthermore, a hyper-parameter max step is set in the cognitive level time sequence, the cognitive level time sequence of all learners is processed into integral multiple of max step, and left filling is carried out by 0 when the length is insufficient, so as to ensure the length to be consistent.

Further, the Embedding layer is an Embedding layer, and an Embedding matrix is established according to the input knowledge point information and the cognitive level information code:

and

both are independent.

Is added with a Mastered embedding matrix after position coding,

is a Knowledge embedding matrix:

M＝x ^a ×W ^a

P＝x ^p ×W ^p

wherein M is an initial Mastered embedding matrix, P is a position coding matrix,

to add a Mastered embedding matrix for position coding, d is the dimension of the embedding layer,

a Knowledge embedding matrix; x is the number of ^a Is a time sequence of learner cognitive level after being coded, x ^p Is the position information of the knowledge point in the sequence, x ^k Is after being codedKnowledge point ID information of (1), W ^a 、W ^p 、W ^k Is a trainable parameter matrix.

Further, the Self-attention layer is a Self-attention layer according to

And

obtaining query (sequence), key (key) and value (value) of input information, and then obtaining output Attention (Q, K, V) from an Attention layer by using a scaled dot product Attention mechanism, wherein the formula is as follows:

wherein, W ^Q 、W ^K 、W ^V Is a trainable parameter matrix and d is a matrix QK ^T Of (c) is calculated.

Further, the Feed forward layer is a forward propagation network, and performs forward propagation on the output of self-attention:

y＝SW ⁽¹⁾ +b ⁽¹⁾

y＝ReLU(y)

y＝yW ⁽²⁾ +b ⁽²⁾

F＝dropout(y)

wherein S is the output of self-attitude layer, W ⁽¹⁾ 、W ⁽²⁾ To be trainedTrained weight parameter matrix, b ⁽¹⁾ 、b ⁽²⁾ For a trainable bias parameter matrix, reLU is an activation function expressed in the form: f (x) = max (0,x), which outputs 0 in case of negative input, then neurons will not be activated, which means that only part of neurons will be activated at the same time, making the network sparse and thus very efficient for computation; the dropout function means that a part of neuron nodes are randomly and temporarily discarded with a certain probability in the deep network training, specifically, dropout acts on each small batch of training data, and due to the mechanism of randomly discarding part of neurons, the dropout function is equivalent to training a neural network with a different structure at each iteration.

Further, the Prediction layer is a full connection layer, the output of self-attribute is propagated forward to obtain a matrix F, and the probability that each knowledge point is mastered by the learner at different levels is output through a Prediction layer and a softmax activation function:

p _i ＝softmax(f _classifier (F))。

further, the tracking model performs back propagation based on a loss function composed of a prediction classification and a real label, where the loss function is:

where K is a class label, and K =7 in the model are level 0: not mastered; level 1: knowing; and (2) level: comprehending; level 3: application; and 4, level 4: analyzing; and (5) level: synthesizing; and (6) level: evaluating; y is the label, the sample class is i, then y _i =1, otherwise y _i =0; output p of the prediction model _i Is the probability that the sample class is i.

Another object of the present invention is to provide a multi-level intelligent cognitive tracking system using the multi-level intelligent cognitive tracking method, wherein the multi-level intelligent cognitive tracking system includes:

the test question knowledge cognition tensor TKC construction module is used for introducing class of education targets in the Brorumm cognition field based on learning resource information of learners, starting from a traditional test question knowledge point representation Q matrix, and dividing mastery of the learners on knowledge points into six capability levels so as to construct a test question knowledge cognition tensor TKC;

the system comprises a response pair sequence generation module, a learning and learning module, a response data extraction module and a response analysis module, wherein the response pair sequence generation module is used for collecting learning resources and response data of learners and extracting characteristics, and generating time sequence question-knowledge point-bloom level-response condition response pair sequences for each learner;

the cognitive level mining module is used for introducing a multi-attribute cognitive diagnosis method based on historical answer records of learners and test knowledge cognitive tensor TKC integrated with educational target classification in the brucm cognitive field, constructing a cognitive level mining model by combining a deep neural network, and performing multi-level intelligent cognitive mining on the learners;

the multi-level intelligent cognitive tracking module is used for sequencing and coding the learner cognitive level vector obtained by intelligent cognitive mining to obtain deep characterization characteristics of the learner, introducing a deep self-attention (Transformer) network, constructing a multi-level intelligent cognitive tracking model, training the model according to the deep characterization characteristics coded by the learner and tracking the cognitive level of the learner at the next moment;

the future performance prediction module is used for predicting the answering performance of the learner on the test question based on the tracked learner cognitive level vector and the test question knowledge cognitive tensor TKC;

the learning ability analysis module is used for analyzing the learning ability of the learner based on the tracked learner cognitive level vector;

and the personalized resource recommendation module is used for performing personalized learner learning resource recommendation based on the tracked learner cognitive level vector and the learning resource information.

It is another object of the present invention to provide a computer-readable storage medium storing a computer program which, when executed by a processor, causes the processor to perform the steps of:

introducing education target classification in the brucm cognition field, constructing a test question knowledge cognition tensor TKC, collecting learning resources and answering data of a learner, and generating a time sequence answering pair sequence of the learner; a multi-partition attribute cognitive diagnosis method is introduced, and a cognitive level mining model is constructed by combining a deep neural network; and sequencing and coding the cognitive level mining results of the learner to obtain deep characterization characteristics of the learner, constructing a multi-level intelligent cognitive tracking model by combining a self-attention mechanism, and further predicting the answering performance of the learner on the test questions.

The invention also aims to provide an information data processing terminal, which is used for realizing the multi-level intelligent cognitive tracking system.

In combination with the technical solutions and the technical problems to be solved, please analyze the advantages and positive effects of the technical solutions to be protected in the present invention from the following aspects:

the invention provides a multi-level intelligent cognitive tracking method and a multi-level intelligent cognitive tracking system, which are characterized in that firstly, education target classification in the field of Broume cognition is introduced, a corrected test question knowledge cognition tensor TKC is constructed starting from a traditional test question knowledge point representation Q matrix; learning resources and response data of learners are collected, feature extraction is carried out, and a time sequence test question-knowledge point-bloom level-response condition response pair sequence is generated for each learner. And then, a multi-attribute cognitive diagnosis method is introduced based on a time sequence answer pair sequence of the learner and the test knowledge cognitive tensor TKC integrated with the education target classification in the brucm cognitive field, a cognitive level mining model is constructed by combining a deep neural network, and multi-level intelligent cognitive mining is carried out on the learner. Finally, sequencing and coding learner cognitive level vectors obtained by intelligent cognitive mining to obtain deep characterization characteristics of the learners, and introducing a deep self-attention (Transformer) network to construct a multi-level intelligent cognitive tracking model; and predicting the answering performance of the learner on the test question according to the tracked learner cognitive level vector and the test question knowledge cognitive tensor TKC. The invention is beneficial to accurately and effectively modeling and mining the whole knowledge structure and the specific cognitive level of the learner in a fine-grained manner, thereby promoting the personalized learning of the learner and providing a new idea for tracking and predicting the knowledge structure and the cognitive level of the learner in the online learning platform.

According to the method, the cognitive state of the learner is modeled in a more detailed mode, the cognitive level of the learner at each knowledge point is associated and represented with the test question knowledge cognitive tensor TKC integrated with the education target classification in the brucm cognitive field, the cognitive level of the learner is mined in a fine-grained mode by using the customized cognitive level mining function, and the whole knowledge mastering and the specific cognitive level of the learner can be accurately estimated based on the TKC tensors of different levels.

The method introduces the deep self-attention network transducer and fuses the self-defined knowledge grasping self-attention mechanism into the deep self-attention network transducer, compared with a short-time neural network driven by data, the method can consider the long-time dependency relationship among sequences, not only considers the dominant factor of the actual grasping level of a learner while updating the cognitive level of the learner at each moment, but also integrates the invisible information of mutual influence among knowledge points under the self-attention mechanism, and effectively simulates the change condition of the cognitive level of the learner in the actual learning scene. The invention has higher accuracy of mining and tracking the whole knowledge structure and the specific cognitive level of the learner, is superior to the traditional intelligent cognitive tracking method, and can provide more accurate self-checking information for students and more effective tutoring information for teachers.

The invention effectively excavates the cognitive level of the learner on the knowledge point, considers that the cognitive levels of the learner on different knowledge points can be influenced mutually, and uses a self-attention mechanism to model the association between the knowledge points, thereby improving the accuracy of the model and being closer to the reality.

According to the invention, the cognition level information is defined by adopting the test question knowledge cognition tensor TKC which integrates education target classification in the brucm cognition field, so that the condition that the traditional Q matrix is completely used for rough definition is avoided, the professional knowledge in the education field is fully utilized, and the model effectiveness and the interpretability are improved.

The multi-level intelligent cognitive tracking method provided by the invention can mine the cognitive level of the learner on the knowledge point, and fully considers the long-term dependency relationship among answer sequences, so that the whole knowledge structure and the specific cognitive level of the learner are tracked and predicted, effective mining and tracking information is provided for the learner, the learner is helped to adjust the subsequent learning plan, the gap is checked, and the learning efficiency is improved.

The expected income and commercial value after the technical scheme of the invention is converted are as follows: the invention supports personalized learning by big data and artificial intelligence, realizes the education concept of teaching according to the nature, can be widely applied to the fields of intelligent education systems, intelligent teaching guidance systems, intelligent teaching assistance, self-adaptive learning systems and the like, and has great commercial value.

The technical scheme of the invention fills the technical blank in the industry at home and abroad: according to the invention, by constructing a multi-level intelligent cognitive tracking model, professional knowledge and technology in the education field are fused, and a long-term dependence tracking model with self-defined knowledge mastering self-attention mechanism is designed and fused, so that the cognitive level information with finer granularity can be obtained, and the blank that the cognitive level vector of the learner is constructed only from two angles of '0' and '1' by the technology in the industry and the multi-dimensional, comprehensive and fine-grained mining and tracking of the cognitive level of the learner cannot be carried out is filled, so that the integral knowledge structure and the specific cognitive level of the learner are effectively modeled, and comprehensive and accurate guidance information is provided for the personalized learning of the learner.

The technical scheme of the invention solves the technical problems which are always desired to be solved but are not successfully achieved: when the learner excavates and tracks the cognitive level, people always desire to realize the fine-grained and interpretable excavation mode of the knowledge structure and the cognitive level. The learners' knowledge is not limited to extreme "0" and "1", and the cognitive level is changed with the learning depth or forgetting. The invention aims at the mining of dynamic knowledge structure and cognitive level and solves the technical problem of real-time diagnosis and cognitive level tracking.

The technical scheme of the invention overcomes the technical prejudice that: the traditional intelligent cognitive tracking method simply assumes that the learning state of a learner on a knowledge point is only a mastery state and an unsophisticated state, and the bias greatly hinders the application of intelligent cognitive tracking. The invention starts from professional education field knowledge, constructs a multi-level intelligent cognitive tracking method, breaks through the bias of the traditional technology, and greatly improves the application value of the invention.

Drawings

Fig. 1 is a flowchart of a multi-level intelligent cognitive tracking method according to an embodiment of the present invention;

fig. 2 is a tensor diagram of a multi-level intelligent cognitive tracking method provided by an embodiment of the present invention;

FIG. 3 is an exemplary diagram of the test knowledge cognitive TKC tensor provided by the embodiment of the invention;

FIG. 4 is an exemplary radar chart of learner cognitive level provided by an embodiment of the present invention;

fig. 5 is a schematic diagram of a multi-level intelligent cognitive tracking method provided in an embodiment of the present invention, (a) is a schematic diagram of a method, and (b) is a schematic diagram of a cognitive level mining model;

fig. 6 is a structural block diagram of a multi-level intelligent cognitive tracking system according to an embodiment of the present invention;

in the figure: 1. the test knowledge cognition tensor TKC construction module is used for constructing a test knowledge cognition tensor TKC; 2. a response pair sequence generation module; 3. a cognitive level mining module; 4. a multi-level intelligent cognitive tracking module; 5. a future performance prediction module; 6. a learning ability analysis module; 7. and a personalized resource recommendation module.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail with reference to the following embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

This section is an explanatory embodiment expanding on the claims so as to fully understand how the present invention is embodied by those skilled in the art.

As shown in fig. 1, the multi-level intelligent cognitive tracking method provided in the embodiment of the present invention includes the following steps:

s101, introducing Braun cognition field education target classification, starting from a traditional test question knowledge point representation Q matrix, and constructing a corrected test question knowledge cognition tensor TKC; collecting learning resources and response data of learners and carrying out feature extraction to generate a time sequence test question-knowledge point-bloom level-response condition response pair sequence for each learner;

s102, introducing a multi-attribute cognitive diagnosis method based on a time sequence answer pair sequence of the learner and the test question knowledge cognitive tensor TKC integrated with the education target classification in the brucm cognitive field, and constructing a cognitive level mining model by combining a deep neural network to perform multi-level intelligent cognitive mining on the learner;

s103, sequencing and coding learner cognitive level vectors obtained by intelligent cognitive mining to obtain deep characterization characteristics of the learners, introducing a deep self-attention (Transformer) network, and constructing a multi-level intelligent cognitive tracking model; and predicting the answering performance of the learner on the test question according to the tracked learner cognitive level vector and the test question knowledge cognitive tensor TKC.

Some symbols appearing in the embodiments of the present invention are shown in table 1.

TABLE 1 symbolic description

Further, the S101 includes:

S＝{s ₁ ,s ₂ ,…,s _M }

T＝{t ₁ ,t ₂ ,…,t _N }

K＝{k ₁ ,k ₂ ,…,k _L }

(1.1.1) data information collected from the "C language program and design" class of the book of the department of artificial intelligence education, 2021, university of learning and university in china is shown in table 2;

TABLE 2 data set information

Number of students	51
		Knowledge points	17
Number of questions	124
		Highest level of knowledge points	3
Total number of replies recorded	6165
		Maximum number of student responses	124
The shortest recorded number of student's answers	10

(1.1.2) preprocessing a data set, deleting response records with missing values, and sequencing the response records according to a time sequence;

(1.2) starting from a traditional test question knowledge point representation Q matrix, dividing mastery of learners on knowledge points into six cognitive levels, and obtaining specific and uniform definitions of investigation levels of each knowledge point on a cognitive layer, so as to construct a test question knowledge cognition tensor TKC integrating education target classification in the Brum cognitive field, wherein the cognitive level range is 0-6, and 0: not mastered; 1: knowing; 2: comprehending; 3: application; 4: analyzing; 5: synthesizing; 6: evaluating;

wherein r is _mn =0 denotes wrong answer to learner m on test question n, r _mn =1 indicates that learner m answered correctly on test question n;

(1.4) processing the learner response record into a time sequence response pair sequence R of learner-test-knowledge point-bloom level-response condition according to the knowledge point corresponding to the learner response question at each moment and the investigation level thereof _s ：

R _s ＝{"user_id":m,

"test_id":n,

"score″:x,

"knowledge_code":l,

"knowledge_level":c}

In the C language data set, the example of the sequence of the learner's time-series answer pairs can be expressed as follows:

R _s ＝{"user_id":23,

"test_id":11,

"score":1,

"knowledge_code":2,5,

"knowledge_level":1,3}

as shown in fig. 2 and 3, the cognition tensor TKC for the test question knowledge is:

(t _n ,k _l ,c _i )

wherein, the third dimension i =0 represents the test question n non-investigation knowledge point l,1 ≦ i ≦ 6 represents the test question n investigation knowledge points l to c _i And (4) horizontal. According to the classification of the educational target in the cognitive field provided by Broume, the educational experts label the knowledge points and levels of each examination question investigation, and construct different examination question knowledge capability tensors TKC. In the C language data set, the highest level of the examination question investigation knowledge points is the layer 3 application.

Further, the step S102 of constructing a cognitive level mining model, and performing multi-level intelligent cognitive mining on the learner specifically includes the following steps:

(2.1) modeling learner factors including a cognitive level vector F ^s ：

h ^s ＝x ^s ×A

F ^s ＝f _classifier (h ^s )

Wherein x is ^s Is one-hot vector of learner, A is proficiency matrix of all learners, and proficiency vector h of learner can be obtained by multiplying ^s ，f _classfier For the classification function, a learner cognitive level vector F is obtained ^s An example radar map of learner cognitive level is shown in FIG. 4;

F ^kn ＝x ^e ×TKC

the parameter matrix A, B, C is trained by data learning to obtain: x is to be ^s 、x ^e 、x ^e H obtained by multiplying the parameter matrixes A, B, C respectively ^s 、h ^g 、h ^1-g-s Substituting the project reaction function in the step (2.3) to obtain a prediction result of the cognitive level mining model responding to the learner, calculating the cross entropy between the prediction result and the actual score of the learner by using the loss function defined in the step (2.4), and training a parameter matrix according to the loss function value and the gradient descent rule;

training the obtained prediction result by utilizing a deep neural network:

f ₁ ＝φ(W ₁ ×x ^T +b ₁ )

f ₂ ＝φ(W ₂ ×f ₁ +b ₂ )

y＝φ(W ₃ ×f ₂ +b ₃ )

wherein k is _l Is known toThe identification point ID is a point of interest,

for learners at knowledge point k _l Upper cognitive level, W ₁ 、b ₁ Are respectively linear regression functions f ₁ Weight coefficient and bias coefficient of (1), W ₂ 、b ₂ Are respectively a linear regression function f ₂ Weight coefficient and bias coefficient of (1), W ₃ 、b ₃ Respectively, weight coefficients and bias coefficients in the linear regression function y.

(2.4) calculating a loss function of the model, the loss function being the cross entropy between the output y and the true tag r:

loss _CLMM ＝-∑(r _i logy _i +(1-r _i )log(1-y _i ))

wherein y is a predicted value, and r is a real score; after training, F corresponding to learner ^s The mining result of the learner is obtained, each dimension corresponds to the cognitive level of the learner on the knowledge point, the cognitive level range is 0-6, and the cognitive level range applied to the C language data set is 0-3 (0: not mastered; 1: known; 2: comprehended; 3: applied);

The experimental parameter settings of the embodiment of the present invention are shown in table 3.

The experimental development platform used the pytorch framework. Dividing answer data of all learners according to weeks, and respectively mining cognitive levels to obtain cognitive level vectors of the learners at different time nodes. In this step, AUC and ACCURACY were used as evaluation indexes, and the experimental results are shown in table 4.

Table 3 experimental parameters set-one

Parameter(s)	Value of
		batch_size	32
epoch	200/500/800
		dropout	0.5
learning rate	0.001
		linear_num	1024/512/256/128

TABLE 4 Experimental results of step one

Time	AUC	ACC
			week2	0.7650	0.8000
week3	0.7533	0.8900
			week4	0.8663	0.8000
week5	0.6756	0.6600
			week6	0.8181	0.7450
week9	0.9077	0.8500
			week11	0.7478	0.7200
week14	0.8555	0.7867
			week15	0.8022	0.7667
week16	0.7337	0.7000

Further, the step S103 includes:

(3.1) constructing a time sequence of the cognitive level of each learner in the brucm education target by combining a multi-level cognitive level vector obtained by a cognitive level mining model and response time information of the learners;

(3.2) coding the cognitive level time sequence of the learner, wherein if the cognitive level time sequence vector of the learner relates to M knowledge points, the corresponding coding length is M-dimension; for a certain knowledge point k, if the learner's cognitive level is a (a is more than or equal to 0 and less than or equal to 6), the value corresponding to the k-th dimension of the code is a + a × M, so that the deep characterization feature of each learner's cognitive level vector can be obtained;

(3.4) the tracking model performs back propagation based on a loss function composed of a prediction classification and a real label;

furthermore, a hyper-parameter max step is set in the cognitive level time sequence, all the cognitive level time sequence of learners are processed into integral multiple of max step, and left filling is carried out by 0 when the length is not enough, so as to ensure the length to be consistent.

Furthermore, the Embedding layer is an embedded layer and is used for inputtingThe knowledge point information and the cognition level information are encoded to establish an Embedding matrix:

and

both are independent.

Is added with a Mastered embedding matrix after position coding,

is a Knowledge embedding matrix:

M＝x ^a ×W ^a

P＝x ^p ×W ^p

a Knowledge embedding matrix; x is the number of ^a Is a time sequence of learner cognitive level after being coded, x ^p Is the position information of the knowledge point in the sequence, x ^k Is the encoded knowledge point ID information, W ^a 、W ^p 、W ^k Is a trainable parameter matrix.

Further, the Self-attention layer is a Self-attention layer according to

And

obtaining query, key and value of input information, and then obtaining output Attention (Q, K, V) from an Attention layer by adopting a scaled dot product Attention mechanism, wherein the formula is as follows:

y＝SW ⁽¹⁾ +b ⁽¹⁾

y＝ReLU(y)

y＝yW ⁽²⁾ +b ⁽²⁾

F＝dropout(y)

wherein S is the output of self-attitude layer, W ⁽¹⁾ 、W ⁽²⁾ For trainable weight parameter matrices, b ⁽¹⁾ 、b ⁽²⁾ For a trainable bias parameter matrix, reLU is an activation function expressed in the form: f (x) = max (0,x) which outputs 0 if the input is negative thenNeurons will not be activated, which means that only part of the neurons will be activated at the same time, making the network very sparse and thus very efficient for the calculation; the dropout function means that a part of neuron nodes are randomly and temporarily discarded with a certain probability in the deep network training, specifically, dropout acts on each small batch of training data, and due to the mechanism of randomly discarding part of neurons, the dropout function is equivalent to training a neural network with a different structure at each iteration.

p _i ＝softmax(f _classifier (F))。

where K is a category label, K =7 in the model, and K =4 in the model applied to the C language dataset are respectively level 0: not mastered; level 1: knowing; and (2) level: comprehending; and (3) level: application; y is the label, the sample class is i, then y _i =1, otherwise y _i =0; output p of the prediction model _i Is the probability that the sample class is i.

The second experimental parameter set of the embodiment of the present invention is shown in table 5.

Table 5 experimental parameter set two

Parameter(s)	Value of
		batch_size	32
epoch	200/500/800
		dropout	0.2
learning rate	0.01
		max step	10
hidden_num	128

A schematic diagram of the multi-level intelligent cognitive tracking method provided by the embodiment of the invention is shown in fig. 5.

As shown in fig. 6, the multi-level intelligent cognitive tracking system according to the embodiment of the present invention includes:

the test question knowledge cognition tensor TKC construction module 1 is used for introducing class of education targets in the brucm cognition field based on learning resource information of learners, starting from a traditional test question knowledge point representation Q matrix, and dividing mastery of the learners on knowledge points into six capability levels so as to construct a test question knowledge cognition tensor TKC;

the answer pair sequence generation module 2 is used for collecting learning resources and answer data of learners and extracting characteristics, and generating time sequence question-knowledge point-bloom level-answer condition answer pair sequences for each learner;

the cognition level mining module 3 is used for introducing a multi-attribute cognition diagnosis method based on the historical answer records of learners and the test question knowledge cognition tensor TKC integrated with the education target classification in the brucm cognition field, constructing a cognition level mining model by combining a deep neural network, and performing multi-level intelligent cognition mining on the learners;

the multi-level intelligent cognitive tracking module 4 is used for sequencing and coding the learner cognitive level vectors obtained by intelligent cognitive mining to obtain deep characterization characteristics of the learners, introducing a deep self-attention (Transformer) network, constructing a multi-level intelligent cognitive tracking model, training the model according to the deep characterization characteristics coded by the learners and tracking the cognitive level of the learners at the next moment;

the future performance prediction module 5 is used for predicting the response performance of the learner on the test question based on the tracked learner cognitive level vector and the test question knowledge cognitive tensor TKC;

the learning ability analysis module 6 is used for analyzing the learning ability of the learner based on the tracked learner cognitive level vector;

and the personalized resource recommendation module 7 is used for performing personalized learner learning resource recommendation based on the tracked learner cognitive level vector and the learning resource information.

In order to prove the creativity and the technical value of the technical scheme of the invention, the part is the application example of the technical scheme of the claims on specific products or related technologies.

The multi-level intelligent cognitive tracking method provided by the application embodiment of the invention is applied to a computer device, the computer device comprises a memory and a processor, the memory stores a computer program, and the computer program causes the processor to execute the steps of the multi-level intelligent cognitive tracking method when being executed by the processor.

It should be noted that the embodiments of the present invention can be realized by hardware, software, or a combination of software and hardware. The hardware portion may be implemented using dedicated logic; the software portions may be stored in a memory and executed by a suitable instruction execution system, such as a microprocessor or specially designed hardware. Those skilled in the art will appreciate that the apparatus and methods described above may be implemented using computer executable instructions and/or embodied in processor control code, such code being provided on a carrier medium such as a disk, CD-or DVD-ROM, programmable memory such as read only memory (firmware), or a data carrier such as an optical or electronic signal carrier, for example. The apparatus and its modules of the present invention may be implemented by hardware circuits such as very large scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of hardware circuits and software, e.g., firmware.

The embodiment of the invention achieves some positive effects in the process of research and development or use, and has great advantages compared with the prior art, and the following contents are described by combining data, diagrams and the like in the test process.

In addition, the method is compared with the experimental result of the traditional intelligent cognitive tracking method, and the AUC and ACCURACY are used as evaluation indexes, and the comparison result is shown in Table 6.

TABLE 6 results of the experiment

Method	AUC	ACC
			Traditional intelligent cognitive tracking method	61.82％	57.36％
Method for producing a composite material	67.91％	60.79％

According to experimental results, compared with the traditional intelligent cognitive tracking method, the AUC and the ACC of the multi-level intelligent cognitive tracking method are respectively improved by 6.09% and 3.43%, and the method provided by the invention is proved to be capable of better mining the mastery degree of learners on different knowledge points, tracking the cognitive level of learners at different moments, having higher accuracy on the answer performance prediction of learners and being more effective than the traditional intelligent cognitive tracking method.

The experimental result shows that the test question knowledge cognition tensor TKC and the cognition level mining model which are constructed by the multi-level intelligent cognition tracking method and are integrated with the education target classification in the brucm cognition field are effective to fine-grained mining of the cognition level of the learner, and the whole knowledge mastering and the specific level of the learner can be accurately estimated based on TKC tensors of different levels; in addition, a deep self-attention network transducer is introduced and a self-defined knowledge grasping self-attention mechanism is fused into the deep self-attention network transducer, a prediction model capable of considering long-term dependence among sequences is designed, and modeling of the cognitive level of a student is beneficial to modeling of the cognitive level change process of a learner in two senses of time and space; the learning level of the learner is updated at each moment, meanwhile, not only are explicit factors of actual answers of the learner considered, but also invisible information of mutual influence among knowledge points under the self-attention mechanism is provided, the change situation of the knowledge state of the learner in an actual learning scene is effectively simulated, and the whole knowledge structure and the specific learning level of the learner are more accurately mined and tracked.

The above description is only for the purpose of illustrating the present invention and the appended claims are not to be construed as limiting the scope of the invention, which is intended to cover all modifications, equivalents and improvements that are within the spirit and scope of the invention as defined by the appended claims.

Claims

1. A multi-level intelligent cognitive tracking method is characterized by comprising the following steps:

constructing a test question knowledge cognition tensor TKC, collecting data, extracting characteristics and generating an answer pair sequence; a multi-partition attribute cognitive diagnosis method is introduced, and a cognitive level mining model is constructed by combining a deep neural network; and (3) constructing a multi-level intelligent cognitive tracking model by combining a self-attention mechanism, and predicting the answering performance of the learner on the test question.

2. The multi-tier intelligent cognitive tracking method of claim 1, wherein the multi-tier intelligent cognitive tracking method specifically comprises the steps of:

step one, introducing education target classification in the bloom cognition field, constructing a corrected test question knowledge cognition tensor TKC, collecting learning resources and answer data of learners, performing feature extraction, and generating a time sequence test question-knowledge point-bloom level-answer condition answer pair sequence for each learner;

3. The multi-level intelligent cognitive tracking method of claim 2, wherein step one comprises:

S＝{s ₁ ,s ₂ ,…,s _M }

T＝{t ₁ ,t ₂ ,…,t _N }

K＝{k ₁ ,k ₂ ,…,k _L }

Wherein M is learner ID, and the value range is that M is more than or equal to 0 and less than or equal to M; n is the test question ID, and the value range is that N is more than or equal to 0 and less than or equal to N; x is the score of the learner m on the test question n, and the value range is 0/1; l is the knowledge point ID inspected by the test question n, the value range is that L is more than or equal to 0 and less than or equal to L, and L can be expressed as the set of the inspected knowledge point ID when more than one knowledge point is inspected by the test question n; c is the level of the examination question n investigation knowledge point l, the value range is more than or equal to 1 and less than or equal to 6, when l is the set of the investigation knowledge point ID, c can also be expressed as the set of the investigation knowledge point level and respectively corresponds to each dimension in l;

the test question knowledge cognition tensor TKC is as follows:

(t _n ,k _l ,c _i )

4. The multi-level intelligent cognitive tracking method of claim 2, wherein the second step of constructing a cognitive level mining model, and performing multi-level intelligent cognitive mining on the learner specifically comprises the following steps:

(2.1) modeling the learner factor, which is the cognitive level vector F ^s ：

h ^s ＝x ^s ×A

F ^s ＝f _classifier (h ^s )

Wherein x is ^s Is one-hot vector of learner, A is proficiency parameter matrix of all learners, and proficiency vector h of learner can be obtained by multiplying ^s ，f _classfier For the classification function, a learner cognitive level vector F is obtained ^s ；

F ^kn ＝x ^e ×TKC

training the obtained prediction result by utilizing a deep neural network:

f ₁ ＝φ(W ₁ ×x ^T +b ₁ )

f ₂ ＝φ(W ₂ ×f ₁ +b ₂ )

y＝φ(W ₃ ×f ₂ +b ₃ )

wherein k is _l In order to be a knowledge point ID,

for learners at knowledge point k _l Upper cognitive level, W ₁ 、b ₁ Are respectively linear regression functions f ₁ Weight coefficient and bias system in (1)Number, W ₂ 、b ₂ Are respectively linear regression functions f ₂ Weight coefficient and bias coefficient of (1), W ₃ 、b ₃ Respectively are a weight coefficient and a bias coefficient in the linear regression function y;

(2.4) calculating a loss function of the model, wherein the loss function is a cross entropy between the output predicted value y and the real label r:

loss _CLMM ＝-∑(r _i logy _i +(1-r _i )log(1-y _i ))

after training, F corresponding to learner ^s The mining result of the learner is obtained, and each dimension corresponds to the cognitive level of the learner on the knowledge point;

(2.5) updating all weight coefficients W in the model according to the loss function value obtained in (2.4) and the gradient descent rule ₁ 、W ₂ 、W ₃ And bias coefficient b ₁ 、b ₂ 、b ₃ ；

The parameter matrix A, B, C is obtained by training through data learning, and the process is as follows:

firstly, x is ^s 、x ^e 、x ^e H obtained by multiplying the parameter matrixes A, B, C respectively ^s 、h ^g 、h ^1-g-s Substituting the project reaction function to obtain a prediction result of the cognitive level mining model for the learner;

secondly, calculating the cross entropy between the prediction result and the learner real score by using the loss function, and training the parameter matrix according to the loss function value and the gradient descent rule.

5. The multi-level intelligent cognitive tracking method of claim 2, wherein the third step comprises:

(3.1) constructing a brucm education target cognition level time sequence of each learner by combining a multi-level cognition level vector obtained by a cognition level mining model and answering time information of the learner, wherein the cognition level time sequence comprises the cognition level time sequence vector of the learner, a knowledge point ID corresponding to each dimension in the vector and the total length of the time sequence vector;

(3.2) coding the cognitive level time sequence of the learner, wherein if the cognitive level time sequence vector of the learner relates to M knowledge points, the corresponding coding length is M-dimension; for the knowledge point k, if the learner's cognitive level is a, a is more than or equal to 0 and less than or equal to 6, the value corresponding to the k-th dimension of the code is a + a × M, and thus the deep characterization feature of each learner's cognitive level vector is obtained;

(3.5) updating the weight coefficient W in the model according to the loss function value obtained in (3.4) and the gradient descent rule ⁽¹⁾ 、W ⁽²⁾ Bias coefficient b ⁽¹⁾ 、b ⁽²⁾ And a parameter W ^a 、W ^p 、W ^k 、W ^Q 、W ^K 、W ^V 。

6. The multi-level intelligent cognitive tracking method as claimed in claim 5, wherein the cognitive level time sequence is set with a hyper-parameter max step, all the learner's cognitive level time sequence is processed into integer times of max step, and left side filling is performed with ' 0 ' when the length is insufficient, so as to ensure the length is consistent;

the Embedding layer is an Embedding layer, and an Embedding matrix is established according to the input knowledge point information and the cognitive level information code:

and

the two exist independently;

the above-mentioned

Is a Mastered embedding matrix added with position codes, the position codes are

Is a Knowledge embedding matrix:

M＝x ^a ×W ^a

P＝x ^p ×W ^p

wherein M is an initial Mastered embedding matrix, P is a position coding matrix, and d is the dimension of an embedded layer; x is the number of ^a Is a time sequence of learner cognitive level after being coded, x ^p Is the position information of the knowledge point in the sequence, x ^k Is the encoded knowledge point ID information, W ^a 、W ^p 、W ^k Is a trainable parameter matrix.

7. The multi-level intelligent cognitive tracking method of claim 5, wherein said Self-attention layer is a Self-attentive layer, according to said Self-attentive layer

And said

Obtaining the sequence query, key and value of the input information, and expressing as:

the scaled dot product Attention mechanism is used to derive the output Attention (Q, K, V) from the Attention layer expressed as:

8. The multi-level intelligent cognitive tracking method of claim 5, wherein the Feed forward layer is a forward propagation network, forward propagation is performed on the output of self-attention, a dropout function is used for discarding a part of neuron nodes randomly, and then iterative training is performed to obtain a matrix F, wherein the process is as follows:

y＝SW ⁽¹⁾ +b ⁽¹⁾

y＝ReLU(y)

y＝yW ⁽²⁾ +b ⁽²⁾

F＝dropout(y)

wherein S is the output of self-attitude layer, W ⁽¹⁾ 、W ⁽²⁾ For trainable weight parameter matrices, b ⁽¹⁾ 、b ⁽²⁾ For a trainable bias parameter matrix, reLU is an activation function expressed in the form: f (x) = max (0,x);

the Prediction layer is a full connection layer, the matrix F outputs the probability that each knowledge point is mastered by the learner at different levels through a Prediction layer and a softmax activation function:

p _i ＝softmax(f _classifier (F))；

the tracking model performs back propagation based on a loss function formed by prediction classification and real labels, wherein the loss function is as follows:

where K is a category label, K =7, including level 0: not mastered; level 1: knowing; and (2) level: comprehending; and (3) level: application; and 4, level 4: analyzing; and (5) level: synthesizing; and (6) level: evaluating; the sample class of label y is i, then y _i =1, otherwise y _i ＝0。

9. A multi-tier intelligent cognitive tracking system implementing the multi-tier intelligent cognitive tracking method of any one of claims 1-8, the multi-tier intelligent cognitive tracking system comprising:

the test question knowledge cognition tensor TKC construction module is used for introducing class of education targets in the brucm cognition field based on learning resource information of learners, starting from a traditional test question knowledge point representation Q matrix, and dividing mastery of the learners on knowledge points into six capability levels so as to construct a test question knowledge cognition tensor TKC;

the cognitive level mining module is used for introducing a multi-attribute cognitive diagnosis method based on historical answer records of learners and test knowledge cognitive tensors TKC integrated with education target classification in the brucm cognitive field, constructing a cognitive level mining model by combining a deep neural network, and performing multi-level intelligent cognitive mining on the learners;

the multi-level intelligent cognition tracking module is used for sequencing and coding the learner cognition level vector obtained by intelligent cognition mining to obtain deep characterization features of the learner, introducing a deep self-attention network, constructing a multi-level intelligent cognition tracking model, and training the model and tracking the cognition level of the learner at the next moment according to the coded deep characterization features of the learner;

10. A computer-readable storage medium, storing a computer program which, when executed by a processor, causes the processor to perform the steps of the multi-level intelligent cognitive tracking method of any one of claims 1-8.