CN112116092B - Interpretable knowledge level tracking method, system and storage medium - Google Patents

Interpretable knowledge level tracking method, system and storage medium Download PDF

Info

Publication number
CN112116092B
CN112116092B CN202010801341.4A CN202010801341A CN112116092B CN 112116092 B CN112116092 B CN 112116092B CN 202010801341 A CN202010801341 A CN 202010801341A CN 112116092 B CN112116092 B CN 112116092B
Authority
CN
China
Prior art keywords
knowledge
exercise
information
vector
knowledge point
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202010801341.4A
Other languages
Chinese (zh)
Other versions
CN112116092A (en
Inventor
黄昌勤
黄琼浩
李明
***
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhejiang Normal University CJNU
Original Assignee
Zhejiang Normal University CJNU
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Zhejiang Normal University CJNU filed Critical Zhejiang Normal University CJNU
Priority to CN202010801341.4A priority Critical patent/CN112116092B/en
Publication of CN112116092A publication Critical patent/CN112116092A/en
Application granted granted Critical
Publication of CN112116092B publication Critical patent/CN112116092B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/02—Knowledge representation; Symbolic representation
    • G06N5/022—Knowledge engineering; Knowledge acquisition
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • Software Systems (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Biophysics (AREA)
  • Health & Medical Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses an interpretable knowledge level tracking method, an interpretable knowledge level tracking system and a storage medium, wherein the method comprises the following steps: constructing an interpretable knowledge tracking model; performing the following steps on the model: distinguishing embedded knowledge point information, exercise text information and exercise answer record information corresponding to the exercise text information; calculating according to the knowledge point information to obtain a key vector; updating the knowledge level of the knowledge point at the previous moment according to the key vector and the exercise answer record information to obtain the knowledge level at the current moment; calculating to obtain a value vector according to the key vector and the knowledge level of the current moment; and predicting answer information of the exercise text information according to the knowledge point information, the value vector and the exercise answer record information through a multi-dimensional project reaction network block. The method can be effectively applied to online large-scale learning knowledge level state tracking scenes. The method can be applied to the technical field of data mining.

Description

Interpretable knowledge level tracking method, system and storage medium
Technical Field
The invention relates to the technical field of data mining, in particular to an interpretable knowledge level tracking method, an interpretable knowledge level tracking system and a storage medium.
Background
Interpretation of terms:
and (3) CTT: classic Test Theory, which regards the student's problem solving ability as a linear fit of learning knowledge level plus noise. It makes a huge contribution to the practice of educational measures.
BKT, Bayesian Knowledge tracking.
DKT, Deep Knowledge training, Deep learning Knowledge level assessment.
DKVMN, Dynamic Key-Value Memory Networks, DKVMN model.
Due to the explosive growth of data of online education resources, on one hand, more autonomous knowledge construction support is provided for online learners, and on the other hand, the learning efficiency and the learning effect of the online learners are seriously influenced by the emergent information lost problems. The knowledge level of the learner is tracked in real time, and effective help can be provided for reasonably optimizing the learning schedule of the learner, scheduling the learning enthusiasm of the students, assisting the teacher in improving the teaching efficiency and the like.
The knowledge level assessment method mainly includes the steps of recording problem solving information of online learners through an online learning system and assessing the problem solving ability of students in real time by using a technical means. In the existing knowledge level assessment model, the method based on the deep learning method achieves better effect than the traditional psychological measurement or machine learning method, but the prediction based on the deep learning model does not have good resolution compared with the traditional method, so that the further application in education is inconvenient. The multidimensional project reflection theory is a project testing method developed from a single-dimensional project theory, the parameter representation and the model prediction have good resolvability, but the parameter estimation of the model and the project marking need the participation of a large number of field experts, so that the method is not suitable for online large-scale learning knowledge level state tracking.
Disclosure of Invention
To solve one of the above technical problems, the present invention aims to: an interpretable knowledge level tracking method, system, and storage medium are provided, which can be effectively applied to an online large-scale learning knowledge level state tracking scenario.
In a first aspect, an embodiment of the present invention provides:
an interpretable knowledge level tracking method, comprising the steps of:
constructing an interpretable knowledge tracking model;
performing the following steps on the interpretable knowledge tracking model:
distinguishing embedded knowledge point information, exercise text information and exercise answer record information corresponding to the exercise text information;
calculating according to the knowledge point information to obtain a key vector;
updating the knowledge level of the knowledge point at the previous moment according to the key vector and the exercise answer record information to obtain the knowledge level at the current moment;
calculating to obtain a value vector according to the key vector and the knowledge level of the current moment;
and predicting answer information of the exercise text information according to the knowledge point information, the value vector and the exercise answer record information through a multi-dimensional project reaction network block.
Further, the knowledge point information includes current knowledge point information and adjacent knowledge point information of the exercise text information, and the embedding knowledge point information includes:
respectively converting the current knowledge point information and the adjacent knowledge point information into feature description vectors;
respectively constructing an embedded dimension reduction matrix of the feature description vector;
carrying out dimension reduction processing on the feature description vector through the embedded dimension reduction matrix;
and embedding the feature description vector subjected to dimension reduction.
Further, the embedding problem text information includes:
acquiring language types of the exercise text information, wherein the language types comprise natural languages and programming languages;
when the language type of the exercise text information is natural language, a first type conversion model is adopted to convert the exercise text information into word vectors;
when the language type of the exercise text information is a programming language, a second type conversion model is adopted to convert the exercise text information into word vectors;
and embedding the word vectors by adopting a three-way GRU (generalized regression Unit) cyclic neural network.
Further, the embedding of the exercise answer record information corresponding to the exercise text information includes:
converting the exercise answer record information into a single-valued vector with a preset dimension;
and carrying out embedding processing on the single-value vector.
Further, the calculating to obtain the key vector according to the knowledge point information includes:
embedding a key memory matrix;
calculating to obtain a current knowledge point key vector according to the current knowledge point information and the key memory matrix;
and calculating to obtain key vectors of the adjacent knowledge points according to the information of the adjacent knowledge points and the key memory matrix.
Further, the updating the knowledge level of the knowledge point at the previous moment according to the key vector and the exercise answer record information to obtain the knowledge level of the current moment includes:
performing first fusion processing on the exercise text information and the exercise answer record information;
acquiring a knowledge graph structure to which the exercise text information belongs;
performing second fusion processing on the current knowledge point information and the adjacent knowledge point information according to the information subjected to the first fusion processing and the knowledge graph structure to which the exercise text information belongs to obtain a fusion feature description vector;
and updating the knowledge level of the knowledge point at the previous moment according to the fusion feature description vector, the current knowledge point key vector and the adjacent knowledge point key vector to obtain the knowledge level at the current moment.
Further, after the step of constructing the interpretable knowledge tracking model, the method further comprises the following steps:
and optimizing an objective function of the interpretable knowledge tracking model by adopting a random gradient descent method.
In a second aspect, an embodiment of the present invention provides:
an interpretable knowledge level tracking system, comprising:
a construction unit for constructing an interpretable knowledge tracking model;
wherein the interpretable knowledge tracking model comprises:
the embedded distinguishing processing module is used for distinguishing embedded knowledge point information, exercise text information and exercise answer record information corresponding to the exercise text information;
the knowledge level interpretable query module is used for calculating to obtain a key vector according to the knowledge point information; calculating to obtain a value vector according to the key vector and the knowledge level of the current moment;
the knowledge level interpretable updating module is used for updating the knowledge level of the knowledge point at the previous moment according to the key vector and the exercise answer record information to obtain the knowledge level at the current moment;
and the interpretable exercise score prediction module is used for predicting the predicted answer information of the exercise text information according to the knowledge point information, the value vector and the exercise answer record information through the multidimensional project response network block.
In a third aspect, an embodiment of the present invention provides:
an interpretable knowledge level tracking system, comprising:
at least one memory for storing a program;
at least one processor configured to load the program to perform the interpretable knowledge level tracking method of the embodiment of the first aspect.
In a fourth aspect, an embodiment of the present invention provides:
a storage medium having stored therein processor-executable instructions for implementing the interpretable knowledge level tracking method of the first aspect embodiment when executed by a processor.
The invention has the beneficial effects that: the invention divides and embeds knowledge point information, exercise text information and exercise answer record information corresponding to the exercise text information in the interpretable knowledge tracking model after constructing the interpretable knowledge tracking model, calculates key vectors of the knowledge point information, updates the knowledge level at the previous moment of the knowledge point according to the key vectors and the exercise answer record information to obtain the knowledge level at the current moment, calculates value vectors according to the key vectors and the knowledge level at the current moment, predicts the predicted answer information of the exercise text information according to the knowledge point information, the value vectors and the exercise answer record information through a multi-dimensional project reaction network block, and leads the interpretable knowledge tracking model to have no need of participation of a large number of domain experts by utilizing the good parameter representation mode, the analytically predicted output and the universal approximation characteristic of the multi-dimensional project reaction network block, the method can be effectively applied to online large-scale learning knowledge level state tracking scenes.
Drawings
FIG. 1 is a flow diagram of data processing for interpreting a knowledge tracking model in accordance with one embodiment of the present invention;
FIG. 2 is a schematic diagram of a structure of an interpretable knowledge tracking model according to an embodiment of the present invention.
Detailed Description
The invention is described in further detail below with reference to the figures and the specific embodiments. The step numbers in the following embodiments are provided only for convenience of illustration, the order between the steps is not limited at all, and the execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art.
In the following description, reference is made to "some embodiments" which describe a subset of all possible embodiments, but it is understood that "some embodiments" may be the same subset or different subsets of all possible embodiments, and may be combined with each other without conflict.
Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of the present application only and is not intended to be limiting of the application.
The embodiment of the invention provides an interpretable knowledge level tracking method, which can be applied to a server, and the server can interact with a plurality of learning terminals.
Before the explanation of the specific embodiments, the following symbols are defined, and the specific contents are shown in table 1:
TABLE 1
Figure BDA0002627500720000041
Figure BDA0002627500720000051
Figure BDA0002627500720000061
Figure BDA0002627500720000071
Next, the following explanation is made:
assume that in a learning education system, a course has S {.., S {.i,. } the students study, and the course has Q practice problems. The knowledge points of the course can be regarded as a graph G ═ V, E with a structure, and V represents a knowledge point set
Figure BDA0002627500720000072
Certain dependency relationship exists between the knowledge points
Figure BDA0002627500720000073
And these knowledge points can be further abstracted to N1Knowledge concept
Figure BDA0002627500720000074
For a specific student s'iHis problem record can be represented as a sequence set seq (s'i)＝{(c1,q1,r1),...,(ct,qt,rt),...,(cT,qT,rT) Is of s'i∈S、
Figure BDA0002627500720000075
qtE.g. Q, at the same time (c)t,qt,rt) Indicating congestionSet of related knowledge points ctProblem qtGiven answer r by the student at time step tt. Exercise question qtRepresented by a set of sequences whose elements are each word from the text content of the problem, can be written as
Figure BDA0002627500720000076
In the present embodiment, answers to the problem rtIs a Boolean value, which is used to indicate whether the corresponding problem question is right or wrong.
Further, in the present embodiment, it is assumed that each student has a knowledge level varying with time for each knowledge point in each step t, and the rule of updating this knowledge level with time is as follows:
when a student answers a problem, the associated set of knowledge points ctEach knowledge point of the set of knowledge points
Figure BDA0002627500720000078
And a set of neighboring knowledge points NtThe knowledge level is updated.
Thus, each student problem score prediction method can be described as:
given the exercise material and records within a learning time step 1 to T, the knowledge level of each knowledge point of the learner can be updated by constructing a knowledge level assessment model, using the exercise records to predict the exercise q to be done nextT+1Answer r of (1)T+1。
In view of the above explanation, the implementation process of the present embodiment includes the following steps:
s1, constructing an interpretable knowledge tracking model; the interpretable knowledge tracking model is the knowledge level assessment model, and the specific structure of the interpretable knowledge tracking model is shown in fig. 2. The interpretable knowledge tracking model of the present embodiment differs from existing neural network models in that: the distinguishing embedding is carried out in the model input stage to provide a basis for subsequent model interpretability.
Specifically, the steps shown in fig. 1 are performed on the interpretable knowledge tracking model of step S1:
s21, distinguishing embedded knowledge point information, exercise text information and exercise answer record information corresponding to the exercise text information; this step is carried out with N1Concept of individual knowledge
Figure BDA0002627500720000077
And different embedding is carried out on the different knowledge points for classification basis.
In some embodiments, the knowledge point information includes current knowledge point information and neighboring knowledge point information of the problem text information.
Wherein the embedding knowledge point information includes:
respectively converting the current knowledge point information and the adjacent knowledge point information into feature description vectors;
the method comprises the following specific steps:
set c of current knowledge point informationtConversion to N1Individual knowledge concept feature description vector
Figure BDA0002627500720000081
Wherein
Figure BDA0002627500720000082
The elements of the vector
Figure BDA0002627500720000083
If it is coincident with
Figure BDA0002627500720000084
And is
Figure BDA0002627500720000085
Otherwise
Figure BDA0002627500720000086
Set N of adjacent knowledge point informationtConversion to N1Description vector of each knowledge characteristic
Figure BDA0002627500720000087
Wherein the content of the first and second substances,
Figure BDA0002627500720000088
the elements of the vector
Figure BDA0002627500720000089
If it is coincident with
Figure BDA00026275007200000810
And is
Figure BDA00026275007200000811
Otherwise
Figure BDA00026275007200000812
Since the above-constructed knowledge point feature description vectors are too sparse and high dimensional, dimension reduction processing is required. In this embodiment, the embedded dimension reduction matrix of the feature description vector is respectively constructed; and carrying out dimension reduction processing on the feature description vector by embedding a dimension reduction matrix.
Specifically, the process of this step is as follows:
aiming at the feature description vector of the knowledge point contained in the current knowledge point, an embedded dimension reduction matrix is constructed
Figure BDA00026275007200000813
Handle
Figure BDA00026275007200000814
Reducing vitamin
Figure BDA00026275007200000815
The process is shown in equation 1, equation 2, and equation 3:
Figure BDA00026275007200000816
Figure BDA00026275007200000817
Figure BDA00026275007200000818
wherein the content of the first and second substances,
Figure BDA00026275007200000819
if it is not
Figure BDA00026275007200000820
Element(s)
Figure BDA00026275007200000821
Otherwise
Figure BDA00026275007200000822
Aiming at the feature description vectors of adjacent knowledge points, an embedded dimension reduction matrix is constructed
Figure BDA00026275007200000823
Handle
Figure BDA00026275007200000824
Reducing vitamin
Figure BDA00026275007200000825
The process is shown in equation 4 and equation 5:
Figure BDA00026275007200000826
Figure BDA00026275007200000827
and finally, embedding the feature description vector subjected to dimension reduction.
The embedding of the exercise answer record information corresponding to the exercise text information comprises:
converting the exercise answer record information into a single-valued vector with a preset dimension; and then embedding the single-value vector. For example, assuming that the current time step for a student to answer is t, the answer is recorded as (c)t,qt,rt) Then, the exercise answer record is converted into dimension d2The transformation process of the single-valued vector of (2) can be as shown in equation 6:
Figure BDA0002627500720000091
the embedded problem text information comprises:
acquiring language types of the exercise text information, wherein the language types comprise natural languages and programming languages;
when the language type of the exercise text information is natural language, a first type conversion model is adopted to convert the exercise text information into word vectors; the first class transformation model may be word2 vec.
When the language type of the exercise text information is a programming language, a second type conversion model is adopted to convert the exercise text information into word vectors; the second class conversion model may be code2 vec. In the step, the statement block is taken as a unit, and the exercise text information is converted into word vectors.
After the word vector is obtained, a three-way (tri-directional) GRU recurrent neural network is adopted to carry out embedding processing on the word vector. Specifically, a three-way (tri-directional) GRU recurrent neural network is adopted to carry out embedding processing on the exercise text information. The front bidirectional (bidirectional) GRU recurrent neural network is mainly used for learning the time-space dependence of text contents, and the rear unidirectional (uni-directional) GRU recurrent neural network is mainly used for coding learned information to obtain global embedded representation of the text contents of exercise contents.
Specifically, the process of learning text spatio-temporal dependence information based on the bidirectional GRU recurrent neural network is as follows:
the working flow of the cyclic unit of the bidirectional GRU cyclic neural network for learning the spatiotemporal information of the text content of the exercise is described in the following formula 7:
Figure BDA0002627500720000092
wherein the content of the first and second substances,
Figure BDA0002627500720000093
input weight matrix
Figure BDA0002627500720000094
Cyclic network weights
Figure BDA0002627500720000095
And bias weight
Figure BDA0002627500720000096
Are all parameters that need to be trained in the network block.
In the embodiment, because the bidirectional cyclic neural network is adopted, the word vector hidden state in the forward direction is recorded as
Figure BDA0002627500720000097
The word vector hidden state in the backward direction is expressed as
Figure BDA0002627500720000098
After the processing of formula 7, the hidden states of all word vectors in the forward and backward directions can be obtained, and in order to construct the embedded representation of the global problem text content, the obtained hidden state vectors need to be fused, and the embodiment adopts average fusion processing, that is, the average fusion processing is adopted, that is, the hidden state vectors are subjected to forward and backward hidden states
Figure BDA0002627500720000101
Since the content in the problem text contains different vocabularies, such as natural language, programming language, etc., in different semantic spaces, it needs to perform weighting and equivalent processing, and this embodiment uses the attention mechanism and current knowledge embedding pair
Figure BDA0002627500720000102
The processing is carried out, and the operation formula is shown as formula 8:
Figure BDA0002627500720000103
wherein alpha ismIs composed of
Figure BDA0002627500720000104
And
Figure BDA0002627500720000105
the value of attention in between.
The problem text global embedding encoder based on the GRU recurrent neural network comprises the following specific processes:
after a fused word vector hidden state sequence is obtained
Figure BDA0002627500720000106
The unidirectional GRU recurrent neural network is adopted as a global problem text content embedded representation encoder for the sequence, and the processing procedure is shown as formula 9:
Figure BDA0002627500720000107
wherein GRU (. cndot.,) represents a recurrent neural network, the parameters of which are respectively an input sequence and a parameter list to be trained, the internal structure of which is the same as that of formula 7, and a weight matrix is simultaneously input
Figure BDA0002627500720000108
Cyclic network weights
Figure BDA0002627500720000109
And bias weight
Figure BDA00026275007200001010
Are all parameters that need to be trained in the network block.
Since the above-constructed model requires a memory matrix to store and track the hidden states of the learner's knowledge level, existing knowledge hidden states lack interpretable parametric representations. Therefore, the embodiment proposes a knowledge level parameter memory matrix with definite meaning, which includes a key memory matrix for providing knowledge point component query and a value memory matrix for recording hidden state of knowledge level, and also provides a method for matching knowledge level interpretable query and knowledge level interpretable update, and the process is as follows:
s22, calculating according to the knowledge point information to obtain a key vector; the key vectors include a current knowledge point key vector and an adjacent knowledge point key vector.
In some embodiments, this step is specifically as follows:
embedding a key memory matrix; the key memory matrix has the performance of storing long-term stable knowledge.
Calculating to obtain a current knowledge point key vector according to the current knowledge point information and the key memory matrix;
and calculating to obtain key vectors of the adjacent knowledge points according to the information of the adjacent knowledge points and the key memory matrix.
In particular, one dimension d may be taken3×N1The matrix is a key memory matrix that has the capability of storing long-term stable knowledge, denoted as K. The knowledge point component information of each problem can be inquired from the key memory matrix through specific operation, the inquired vector is called key vector, and its meaning is the influence weight of the related knowledge point of the problem to the knowledge level, including the current knowledge point key vector betatAnd neighboring knowledge point key vector gammat. Assuming that the current time step is t, the answer is recorded as (c)t,qt,rt) About problem qtKnowledge point set ctAnd its neighboring knowledge points NtThe key vectors of (a) can be obtained by the operation of equation 10:
Figure BDA0002627500720000111
wherein the content of the first and second substances,
Figure BDA0002627500720000112
representing a Hadamard multiplication.
S23, updating the knowledge level of the knowledge point at the previous moment according to the key vector and the exercise answer record information to obtain the knowledge level at the current moment;
in some embodiments, this step may be implemented by:
performing first fusion processing on the exercise text information and the exercise answer record information; the first fusion process refers to the information aggregation stage.
Acquiring a knowledge graph structure to which the exercise text information belongs;
performing second fusion processing on the current knowledge point information and the adjacent knowledge point information according to the information subjected to the first fusion processing and the knowledge graph structure to which the exercise text information belongs to obtain a fusion feature description vector; the second fusion process refers to a value update phase.
And updating the knowledge level of the knowledge point at the previous moment according to the fusion feature description vector, the current knowledge point key vector and the adjacent knowledge point key vector to obtain the knowledge level at the current moment.
Specifically, the knowledge level updating of the step mainly plays a role in tracking and updating the knowledge level of the learner. Assuming the current time step is t, the question answer record information is (c)t,qt,rt) As shown in FIG. 2, the knowledge level update network block of this step will update the value-memory matrix from the previous time step vt-1Updating to the current knowledge level state vtAnd a time sequence memory function is provided for subsequent knowledge level tracking and exercise score prediction. Updating the content of the network block by the knowledge level mainly comprises two processes: an information aggregation phase and a value update phase.
The specific process of the information aggregation stage is as follows:
assuming the current time step is t, the question answer record information is (c)t,qt,rt) First, the embedded representation x of the text information content of the problemtRelative answer r to the exercise text informationt ECarrying out polymerization by using the formula 11:
Figure BDA0002627500720000121
then, the knowledge point information c corresponding to the exercisetAnd its adjacent knowledge points NtBased on information aggregation phase acquisition
Figure BDA0002627500720000122
Further fusing the knowledge point diagram structure related to the course to obtain a fused feature vector
Figure BDA0002627500720000123
The fusion can be realized by formula 12, formula 13 and formula 14:
Figure BDA0002627500720000124
Figure BDA0002627500720000125
Figure BDA0002627500720000126
wherein the content of the first and second substances,
Figure BDA0002627500720000127
fneighboris a fully-connected neural network function that is used to decide how current knowledge point set information is passed to neighboring knowledge point sets.
Obtaining a fused feature vector
Figure BDA0002627500720000128
Then, proceed withKnowledge level update, this embodiment combines the current knowledge point key vector β of the query with the unidirectional GRU recurrent neural network appropriate for the problem scenariotAnd neighboring knowledge point key vector gammatThe state of knowledge level of the trainee is updated, and the updating operation process is shown as formula 15:
Figure BDA0002627500720000129
wherein, Vt jIs a value memory matrix VtThe vector of the j-th column of the image,
Figure BDA0002627500720000131
and
Figure BDA0002627500720000132
belonging to the weight parameters to be trained in the network block.
S24, calculating to obtain a value vector according to the key vector and the knowledge level of the current moment; in the step, a value vector is obtained by calculation according to the key vector of the current knowledge point information and the knowledge level of the previous moment. Which is to track the level of knowledge of the student. In particular, with a d4×N1The memory matrix of (a) represents the hidden state of the trainee's knowledge level and is recorded as a value memory matrix V. Assuming that the current time step is t, the answer is recorded as (c)t,qt) When the learner predicts the exercise result, it needs to search out the knowledge level vector associated with the exercise knowledge point from the value memory matrix V, called value vector, and record it as value vector
Figure BDA0002627500720000133
The query process is shown in equation 16:
Figure BDA0002627500720000134
wherein, softmax (·, ·) represents a vector-wise part connected with softmax neural network, the first parameter is network input, and the second parameter is a weight parameter which needs training for the network.
After the embedding, querying and updating of the knowledge points are completed, step S25 is executed to perform answer prediction.
And S25, predicting answer information of the exercise text information according to the knowledge point information, the value vector and the exercise answer record information through the multi-dimensional project reaction network block.
Specifically, in this step, after the knowledge level state of the student is obtained, the answer condition of the student to the problem needs to be predicted. The method combines the good analytic performance of the multidimensional project reaction theory and the universal approaching property of the neural network capable of being programmed in a micro mode, and is suitable for large-scale online student exercise solvability performance prediction.
Specifically, the parameters of the multidimensional project reaction theory model are estimated by using a micro-programming method, and the model is shown as formula 17:
Figure BDA0002627500720000135
wherein the content of the first and second substances,
Figure BDA0002627500720000136
showing the mastery of the knowledge concept by the trainee at time step t,
Figure BDA0002627500720000137
a value representing the influence of the mastery conditions on student answer conditions, dnIndicating the difficulty of the problem, cnGuess coefficients representing answers, cθThe presetter is used for controlling the student ability value range.
The multi-dimensional project reaction theory model parameters are solved by combining the multi-dimensional project reaction theory and a micro-programming method.
First, in estimating the ability of a student, the parameter θ is mainly solvedtAnd alphatWherein the multidimensional project response network block is also called multidimensional project response theoryTheory model, multidimensional project reaction network Block with knowledge level of student Vt jAnd problem text content xtKey vector beta of current problem knowledge pointtAs an input, a micro-programmable method is used to obtain the desired parameter estimation, and the process can be described as the following formula 18:
Figure BDA0002627500720000141
wherein, WθAnd
Figure BDA0002627500720000142
bθand
Figure BDA0002627500720000143
when evaluating the difficulty parameter of the exercise, the key vector beta of the knowledge point is usedtAnd embedding x of problem text contenttImplementing the parameter c for the input of a network block using a microprogrammable methodn、dnThe specific process of solving (2) is shown in equation 19:
Figure BDA0002627500720000144
wherein, WcAnd
Figure BDA0002627500720000145
bcand
Figure BDA0002627500720000146
therefore, the achievement prediction output module is constructed completely,
Figure BDA0002627500720000147
for the result prediction output of the exercises, the above statements show that the model has good parameter representation and resolvable result prediction output, also has an efficient parameter solving method, and can be applied to online big exercisesAnd tracking and evaluating the knowledge level of the scale students, and realizing interpretable prediction output of the student exercise scores.
Before the above model is applied specifically, in some embodiments, the objective function of the interpretable knowledge tracking model is optimized by a stochastic gradient descent method.
Specifically, the model of the present embodiment is an end-to-end, fully differentiable neural network model, and therefore, the parameters that need to be trained include: exercise text embedded
Figure BDA0002627500720000148
Part of the key-value pair query process
Figure BDA0002627500720000149
Knowledge level updating network blocks
Figure BDA00026275007200001410
Multidimensional item reflects { W of a network block*,b*}. During the t-th time step, the model of this embodiment answers question qtGiving predictive answers
Figure BDA00026275007200001411
The actual answer made by the student is recorded as a case rt. The aim of the model training is to adjust the above parameters based on the observed sequence s'i＝{(c1,q1,r1),(c2,q2,r2),...,(cT,qT,rT) To minimize the objective function equation 20:
Figure BDA00026275007200001412
since the model is fully differentiable, the objective function is optimized by the stochastic gradient descent algorithm, so that the training of the model is completed.
In summary, the embodiment of the invention utilizes the good parameter representation mode, the analytically predictable output and the universal approximation characteristic of the multidimensional project response network block, so that the constructed interpretable knowledge tracking model can be effectively applied to the online large-scale learning knowledge level state tracking scene without the participation of a large number of field experts.
An embodiment of the present invention provides an interpretable knowledge level tracking system, including:
a construction unit for constructing an interpretable knowledge tracking model;
wherein, as shown in fig. 2, the interpretable knowledge tracking model comprises:
the embedded distinguishing processing module is used for distinguishing embedded knowledge point information, exercise text information and exercise answer record information corresponding to the exercise text information;
the knowledge level interpretable query module is used for calculating to obtain a key vector according to the knowledge point information; calculating to obtain a value vector according to the key vector and the knowledge level of the current moment;
the knowledge level interpretable updating module is used for updating the knowledge level of the knowledge point at the previous moment according to the key vector and the exercise answer record information to obtain the knowledge level at the current moment;
and the interpretable exercise score prediction module is used for predicting the predicted answer information of the exercise text information according to the knowledge point information, the value vector and the exercise answer record information through the multidimensional project response network block.
The content of the embodiment of the method of the invention is all applicable to the embodiment of the system, the function of the embodiment of the system is the same as the embodiment of the method, and the beneficial effect achieved by the embodiment of the system is the same as the beneficial effect achieved by the method.
An embodiment of the present invention provides an interpretable knowledge level tracking system, including:
at least one memory for storing a program;
at least one processor configured to load the program to perform the interpretable knowledge level tracking method of the embodiment of the first aspect.
The content of the embodiment of the method of the invention is all applicable to the embodiment of the system, the function of the embodiment of the system is the same as the embodiment of the method, and the beneficial effect achieved by the embodiment of the system is the same as the beneficial effect achieved by the method.
Furthermore, a storage medium is provided, in which processor-executable instructions are stored, which when executed by a processor are used to implement the interpretable knowledge level tracking method of an embodiment of the first aspect.
While the preferred embodiments of the present invention have been illustrated and described, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims.

Claims (6)

1. An interpretable knowledge level tracking method, comprising the steps of:
constructing an interpretable knowledge tracking model;
performing the following steps on the interpretable knowledge tracking model:
distinguishing embedded knowledge point information, embedded problem text information and problem answer record information embedded corresponding to the problem text information;
calculating to obtain a key vector according to the embedded knowledge point information;
updating the knowledge level of the knowledge point at the previous moment according to the key vector and the exercise answer record information embedded into the exercise text information corresponding to the exercise text information to obtain the knowledge level at the current moment;
determining a value vector from a value memory matrix according to the key vector and the knowledge level of the current moment, wherein the value memory matrix is used for recording a knowledge level hidden state;
predicting answer information of the exercise text information according to the embedded knowledge point information, the value vector and the exercise answer record information embedded corresponding to the exercise text information through a multi-dimensional project reaction theory and a neural network;
the knowledge point information comprises current knowledge point information and adjacent knowledge point information of the exercise text information; the embedded knowledge point information comprises:
respectively converting the current knowledge point information and the adjacent knowledge point information into feature description vectors;
respectively constructing an embedded dimension reduction matrix of the feature description vector;
carrying out dimension reduction processing on the feature description vector through the embedded dimension reduction matrix;
embedding the feature description vector subjected to dimension reduction;
the embedded problem text information comprises:
acquiring language categories of exercise text information, wherein the language categories comprise natural languages and programming languages;
when the language type of the exercise text information is natural language, a first type conversion model is adopted to convert the exercise text information into word vectors, wherein the first type conversion model is word2 vec;
when the language type of the exercise text information is a programming language, a second type conversion model is adopted to convert the exercise text information into word vectors, wherein the second type conversion model is code2 vec;
embedding the word vectors by adopting a three-way GRU (generalized regression Unit) cyclic neural network;
the embedding of the exercise answer record information corresponding to the exercise text information comprises:
converting the exercise answer record information into a single-valued vector with a preset dimension;
embedding the single-value vector;
the calculating to obtain the key vector according to the embedded knowledge point information comprises:
embedding a key memory matrix for providing knowledge point component queries;
calculating to obtain a current knowledge point key vector according to the current knowledge point information and the key memory matrix, wherein the key vector comprises a vector inquired from the key memory matrix through calculation of knowledge point component information of each exercise, and the vector represents influence weight of an exercise related knowledge point on a knowledge level;
calculating to obtain key vectors of the adjacent knowledge points according to the information of the adjacent knowledge points and the key memory matrix;
the determining a value vector from a value memory matrix includes:
when the learner exercise achievement is predicted, knowledge level vectors relevant to exercise knowledge points are inquired from the value memory matrix, and the relevant knowledge level vectors are used as value vectors.
2. The method of claim 1, wherein said updating the knowledge level at the previous moment of the knowledge point according to the key vector and the question answer record information embedded in the question answer record information corresponding to the question text information to obtain the knowledge level at the current moment comprises:
performing first fusion processing on the embedded exercise text information and the exercise answer record information embedded corresponding to the exercise text information;
acquiring a knowledge graph structure to which the embedded exercise text information belongs;
according to the information after the first fusion processing and the knowledge graph structure to which the embedded problem text information belongs, performing second fusion processing on the current knowledge point information and the adjacent knowledge point information to obtain a fusion feature description vector;
and updating the knowledge level of the knowledge point at the previous moment according to the fusion feature description vector, the current knowledge point key vector and the adjacent knowledge point key vector to obtain the knowledge level at the current moment.
3. The interpretable knowledge level tracking method of claim 1, further comprising, after the step of constructing an interpretable knowledge tracking model, the steps of:
and optimizing an objective function of the interpretable knowledge tracking model by adopting a random gradient descent method.
4. An interpretable knowledge level tracking system, comprising:
a construction unit for constructing an interpretable knowledge tracking model;
wherein the interpretable knowledge tracking model comprises:
the embedded distinguishing processing module is used for distinguishing embedded knowledge point information, embedded exercise text information and exercise answer record information embedded corresponding to the exercise text information;
the knowledge level interpretable updating module is used for calculating to obtain a key vector according to the embedded knowledge point information; updating the knowledge level of the knowledge point at the previous moment according to the key vector and the exercise answer record information embedded into the exercise text information corresponding to the exercise text information to obtain the knowledge level at the current moment;
the knowledge level interpretable query module is used for determining a value vector from a value memory matrix according to the key vector and the knowledge level of the current moment, and the value memory matrix is used for recording the hidden state of the knowledge level;
the interpretable exercise score prediction module is used for predicting the predicted answer information of the exercise text information according to the embedded knowledge point information, the value vector and the exercise answer record information embedded corresponding to the exercise text information through a multi-dimensional project reaction theory and a neural network;
the knowledge point information comprises current knowledge point information and adjacent knowledge point information of the exercise text information; the embedded knowledge point information comprises:
respectively converting the current knowledge point information and the adjacent knowledge point information into feature description vectors;
respectively constructing an embedded dimension reduction matrix of the feature description vector;
carrying out dimension reduction processing on the feature description vector through the embedded dimension reduction matrix;
embedding the feature description vector subjected to dimension reduction;
the embedded problem text information comprises:
acquiring language categories of exercise text information, wherein the language categories comprise natural languages and programming languages;
when the language type of the exercise text information is natural language, a first type conversion model is adopted to convert the exercise text information into word vectors, wherein the first type conversion model is word2 vec;
when the language type of the exercise text information is a programming language, a second type conversion model is adopted to convert the exercise text information into word vectors, wherein the second type conversion model is code2 vec;
embedding the word vectors by adopting a three-way GRU (generalized regression Unit) cyclic neural network;
the embedding of the exercise answer record information corresponding to the exercise text information comprises:
converting the exercise answer record information into a single-valued vector with a preset dimension;
embedding the single-value vector;
the calculating to obtain the key vector according to the embedded knowledge point information comprises:
embedding a key memory matrix for providing knowledge point component queries;
calculating to obtain a current knowledge point key vector according to the current knowledge point information and the key memory matrix, wherein the key vector comprises a vector inquired from the key memory matrix through calculation of knowledge point component information of each exercise, and the vector represents influence weight of an exercise related knowledge point on a knowledge level;
calculating to obtain key vectors of the adjacent knowledge points according to the information of the adjacent knowledge points and the key memory matrix;
the determining a value vector from a value memory matrix includes:
when the learner exercise achievement is predicted, knowledge level vectors relevant to exercise knowledge points are inquired from the value memory matrix, and the relevant knowledge level vectors are used as value vectors.
5. An interpretable knowledge level tracking system, comprising:
at least one memory for storing a program;
at least one processor configured to load the program to perform the interpretable knowledge level tracking method of any one of claims 1-3.
6. A storage medium having stored therein processor-executable instructions for implementing the interpretable knowledge level tracking method of any one of claims 1-3 when executed by a processor.
CN202010801341.4A 2020-08-11 2020-08-11 Interpretable knowledge level tracking method, system and storage medium Active CN112116092B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010801341.4A CN112116092B (en) 2020-08-11 2020-08-11 Interpretable knowledge level tracking method, system and storage medium

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010801341.4A CN112116092B (en) 2020-08-11 2020-08-11 Interpretable knowledge level tracking method, system and storage medium

Publications (2)

Publication Number Publication Date
CN112116092A CN112116092A (en) 2020-12-22
CN112116092B true CN112116092B (en) 2022-03-25

Family

ID=73803985

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010801341.4A Active CN112116092B (en) 2020-08-11 2020-08-11 Interpretable knowledge level tracking method, system and storage medium

Country Status (1)

Country Link
CN (1) CN112116092B (en)

Families Citing this family (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112766513B (en) * 2021-01-29 2024-05-07 华中师范大学 Knowledge tracking method and system for memory collaboration
CN112818100B (en) * 2021-01-29 2023-12-22 华中师范大学 Knowledge tracking method and system for integrating question difficulty
CN113239699B (en) * 2021-04-15 2023-11-07 北京师范大学 Depth knowledge tracking method and system integrating multiple features
CN113724110A (en) * 2021-08-27 2021-11-30 中国海洋大学 Interpretable depth knowledge tracking method and system and application thereof
CN114155124B (en) * 2022-02-07 2022-07-12 山东建筑大学 Test question resource recommendation method and system
CN114911975B (en) * 2022-05-05 2024-04-05 金华航大北斗应用技术有限公司 Knowledge tracking method based on graph attention network
CN116306773B (en) * 2022-09-13 2024-03-26 西湖大学 Interpretable neural network high-dimensional data dimension reduction analysis system
CN116578694B (en) * 2023-04-21 2024-01-09 华南师范大学 Disentangled knowledge tracking method, system, device and medium

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110297817A (en) * 2019-06-25 2019-10-01 哈尔滨工业大学 A method of the structure of knowledge is constructed based on personalized Bayes's knowledge tracing model
CN110378818A (en) * 2019-07-22 2019-10-25 广西大学 Personalized exercise recommended method, system and medium based on difficulty
CN110704732A (en) * 2019-09-19 2020-01-17 广州大学 Cognitive diagnosis-based time-sequence problem recommendation method

Family Cites Families (16)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9704102B2 (en) * 2013-03-15 2017-07-11 William Marsh Rice University Sparse factor analysis for analysis of user content preferences
CN107025219B (en) * 2017-04-19 2019-07-26 厦门大学 A kind of word insertion representation method based on internal Semantic hierarchy
CN107067849B (en) * 2017-04-21 2020-10-30 杭州博世数据网络有限公司 Parent end system based on cloud teaching platform
CN107545146A (en) * 2017-09-08 2018-01-05 北京辅仁淑凡软件科技有限公司 A kind of Psychological Evaluation method that can adaptively set a question
CN107729987A (en) * 2017-09-19 2018-02-23 东华大学 The automatic describing method of night vision image based on depth convolution loop neutral net
CN107729497B (en) * 2017-10-20 2020-08-14 同济大学 Word embedding deep learning method based on knowledge graph
CN108920705B (en) * 2018-07-19 2020-11-03 邓北平 Knowledge point identification coding method and device
CN109598995B (en) * 2019-01-08 2020-11-17 上海健坤教育科技有限公司 Intelligent teaching system based on Bayesian knowledge tracking model
CN110377707B (en) * 2019-07-25 2024-04-02 中国科学技术大学 Cognitive diagnosis method based on depth item reaction theory
CN110428010B (en) * 2019-08-05 2022-09-06 中国科学技术大学 Knowledge tracking method
CN110991645B (en) * 2019-11-18 2024-03-29 广东宜学通教育科技有限公司 Self-adaptive learning method, system and storage medium based on knowledge model
CN110941723A (en) * 2019-11-18 2020-03-31 广东宜学通教育科技有限公司 Method, system and storage medium for constructing knowledge graph
CN110930274B (en) * 2019-12-02 2023-04-28 中山大学 Practice effect evaluation and learning path recommendation system and method based on cognitive diagnosis
CN111159419B (en) * 2019-12-09 2021-05-25 浙江师范大学 Knowledge tracking data processing method, system and storage medium based on graph convolution
CN111475645B (en) * 2020-03-17 2024-04-30 平安国际智慧城市科技股份有限公司 Knowledge point labeling method, knowledge point labeling device and computer readable storage medium
CN111461442B (en) * 2020-04-07 2023-08-29 中国科学技术大学 Knowledge tracking method and system based on federal learning

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110297817A (en) * 2019-06-25 2019-10-01 哈尔滨工业大学 A method of the structure of knowledge is constructed based on personalized Bayes's knowledge tracing model
CN110378818A (en) * 2019-07-22 2019-10-25 广西大学 Personalized exercise recommended method, system and medium based on difficulty
CN110704732A (en) * 2019-09-19 2020-01-17 广州大学 Cognitive diagnosis-based time-sequence problem recommendation method

Also Published As

Publication number Publication date
CN112116092A (en) 2020-12-22

Similar Documents

Publication Publication Date Title
CN112116092B (en) Interpretable knowledge level tracking method, system and storage medium
Chen et al. Prerequisite-driven deep knowledge tracing
CN110264091B (en) Student Cognitive Diagnosis Method
CN110428010B (en) Knowledge tracking method
Wang et al. Deep hierarchical knowledge tracing
CN110941723A (en) Method, system and storage medium for constructing knowledge graph
CN111695779A (en) Knowledge tracking method, knowledge tracking device and storage medium
CN115545160B (en) Knowledge tracking method and system for multi-learning behavior collaboration
Lu et al. CMKT: Concept map driven knowledge tracing
CN113591988B (en) Knowledge cognitive structure analysis method, system, computer equipment, medium and terminal
CN114021722A (en) Attention knowledge tracking method integrating cognitive portrayal
CN115545155A (en) Multi-level intelligent cognitive tracking method and system, storage medium and terminal
Duan Automatic generation and evolution of personalized curriculum based on genetic algorithm
CN116136870A (en) Intelligent social conversation method and conversation system based on enhanced entity representation
CN115238036A (en) Cognitive diagnosis method and device based on graph attention network and text information
CN115310520A (en) Multi-feature-fused depth knowledge tracking method and exercise recommendation method
CN114971066A (en) Knowledge tracking method and system integrating forgetting factor and learning ability
CN114676903A (en) Online prediction method and system based on time perception and cognitive diagnosis
Swamy et al. Deep knowledge tracing for free-form student code progression
CN114707775B (en) Knowledge tracking model training method, tracking method, device, equipment and medium
Wang et al. Textbook Enhanced Student Learning Outcome Prediction
Song et al. A deep knowledge tracking model integrating difficulty factors
CN117743699B (en) Problem recommendation method and system based on DKT and Topson sampling algorithm
CN115952838B (en) Self-adaptive learning recommendation system-based generation method and system
CN114155124B (en) Test question resource recommendation method and system

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant