1. A meta-cognition ability assessment model based on online learning behaviors and a deep neural network is characterized in that: the meta-cognition ability evaluation model based on the online learning behavior and the deep neural network comprises the following steps in sequence:

1) stage S1 of building a deep neural network model: a bidirectional gating circulation unit integrated with a self-attention mechanism forms a deep neural network model with a double-channel structure;

2) and an S2 stage of acquiring and preprocessing online learning behavior data: an online learning course is selected, basic behavior data of a learner are firstly acquired from an online learning platform, and are processed into a behavior sequence according to behavior occurrence time recorded by the platform, such as 'check-in' → 'test' → 'watching video'. Then, designing a crawler program to obtain interactive text data of the discussion area, carrying out preprocessing operation for removing special characters and stop words on the data, and carrying out manual analysis on the interactive text by using a Broumm cognitive classification theory to obtain the cognitive level of the learner in consideration of the influence of the cognitive level characteristics of the learner contained in the interactive text on a behavior sequence; finally, the data set is divided into training set X according to the proportion of 4:1_(train)And test set X_(test)；

3) Stage S3 of tagging online learning behavior data: the supervised deep neural network requires a labeled data set, and therefore labels the collected online learning behavior data. The method comprises the steps that a meta-cognition ability questionnaire with high credibility is issued to a learner, the learner is divided into two levels of a meta-cognition ability higher level and a meta-cognition ability lower level based on questionnaire scores, and label labeling is carried out on behavior data corresponding to the learner;

4) stage S4 of word vector training: the input of the deep neural network is one or more groups of multi-dimensional features, aiming at the extracted interactive discussion text data and the cognitive level features obtained by analysis, a Word2Vec Word vector tool is mainly adopted to train the interactive discussion text data and the cognitive level features, and therefore corresponding interactive text vectors and cognitive level vectors are obtained; aiming at the behavior sequence data, mainly adopting an Item2Vec algorithm to encode the behavior sequence data so as to obtain a corresponding behavior sequence vector;

5) stage S5 of deep neural network model training: using different model parameters, the training set X_(train)Inputting the data into the deep neural network model constructed in the step 1) for multiple times in small batches, and training the constructed model; wherein, the update gate and the reset gate in the BiGRU use a sigmoid activation function, when the model is output, the used activation function is sigmoid ', the optimizer of the model selects Adam', and the used loss function is a cross entropy loss function;

6) stage S6 of deep neural network model testing: testing data set X obtained in the step 2)_(test)Inputting the data into the trained deep neural network model in the step 5) to obtain two classification results of the learning behavior data of the learner;

7) evaluation of learner meta-cognitive ability stage S7: and (3) taking the binary classification result obtained in the step 6) as the probability p of the level to which the meta-cognition ability belongs, and dividing the level of the meta-cognition ability of the learner into a higher level of the meta-cognition ability and a lower level of the meta-cognition ability by taking a middle value of 0.5 as a boundary.

2. The meta-cognition assessment model based on online learning behavior and deep neural network according to claim 1, characterized in that: in step 2), the cognitive level obtained by the analysis according to the brucm cognitive target classification theory is specifically represented as follows: memory, understanding, application, analysis, evaluation, and creation.

3. The meta-cognition assessment model based on online learning behavior and deep neural network according to claim 1, characterized in that: in the step 4), the behavior sequence vector V and the cognitive level vector C are fused with the fusion vector M to obtain more complex features, the fusion mode is the splicing of the vectors, and the specific calculation is as follows:

M＝contact(V,C)

wherein contact represents the concatenation of the vectors.

4. The meta-cognition assessment model based on online learning behavior and deep neural network according to claim 1, characterized in that: in step 7), the calculation mode of the effect measurement index for the learner meta-cognitive ability evaluation is as follows:

accuracy (session):

recall (Recall):

f1 value (F1-Score):

accuracy (Accuracy):

TP (True Positive, abbreviated as TP) is the number of Positive samples correctly evaluated by the evaluation model;

FP (False Positive, abbreviated as FP) is the number of the misclassified Positive samples;

TN (True Negative, abbreviated TN) represents the number of correctly classified Negative samples;

FN (False Negative, abbreviated as FN) represents the number of misclassified Negative samples;

when the meta cognition ability is layered according to the general theory p, the specific principle is that when the probability value p is greater than 0.5, the higher meta cognition belongs to the higher meta cognition, and when the probability value p is less than or equal to 0.5, the lower meta cognition belongs to the lower meta cognition.

Background

The online learning breaks through the inherent environment of the traditional education wall type and gradually develops into an important form of the current education. In the online learning environment, the learner and the teacher are generally asynchronous online, so that the effective supervision of the learning condition of the learner by the teacher is greatly reduced. Therefore, learners need to have clear knowledge about individuals and environments, and can independently and reasonably plan, monitor and adjust the learning process, thereby realizing effective online learning. Meta-cognition can enable learners to recognize self-cognition level and skill level, and is a key element for the learners to effectively learn online. However, meta-cognition belongs to a psychological process within an individual, which is difficult to judge directly in practice. The meta-cognition ability model is built through the artificial intelligence deep learning technology to accurately evaluate the meta-cognition of the learner, and the student can be helped to become a better learner.

Meta-cognition is not easy to observe and obtain, and the current research aiming at a Meta-cognition capability model is a newer research field, is basically in the theoretical research stage of the model, and is lack of an effective model construction technology for showing the Meta-cognition capability model. Although the existing researchers analyze and research the meta-cognition strategy of the learner from the aspect of learning behavior analysis, the existing researchers mainly excavate the behavior based on the traditional machine learning technology, and cannot meet the requirement of accurate and intelligent evaluation on the meta-cognition ability of the learner.

Disclosure of Invention

Aiming at the defects in the prior art, the invention aims to provide a meta-cognition ability evaluation model based on online learning behaviors and a deep neural network.

In order to achieve the above object, the meta-cognition ability assessment model based on online learning behavior and deep neural network provided by the invention specifically comprises the following steps:

1) stage S1 of building a deep neural network model: a bidirectional gating circulation unit integrated with a self-attention mechanism forms a deep neural network model with a double-channel structure;

2) and an S2 stage of acquiring and preprocessing online learning behavior data: an online learning course is selected, basic behavior data of a learner are firstly acquired from an online learning platform, and are processed into a behavior sequence according to behavior occurrence time recorded by the platform, such as 'check-in' → 'test' → 'watching video'. Then, designing a crawler program to acquire interactive text data of a discussion area, and performing preprocessing operation for removing special characters and stop words on the data, and in addition, considering that the cognitive level characteristics of learners contained in the interactive text can influence a behavior sequence, manually analyzing the interactive text by using a Broume cognitive classification theory to acquire the cognitive level of the learners, such as memory, understanding, application, analysis, evaluation and creation; finally, the data set is divided into training set X according to the proportion of 4:1_(train)And test set X_(test)；

3) Stage S3 of tagging online learning behavior data: the supervised deep neural network requires a labeled data set, and therefore labels the collected online learning behavior data. The method comprises the steps that a meta-cognition ability questionnaire with high credibility is issued to a learner, the learner is divided into two levels of a meta-cognition ability higher level and a meta-cognition ability lower level based on questionnaire scores, and label labeling is carried out on behavior data corresponding to the learner;

4) stage S4 of word vector training: the input of the deep neural network is one or more groups of multi-dimensional features, aiming at the extracted interactive discussion text data and the cognitive level features obtained by analysis, a Word2Vec Word vector tool is mainly adopted to train the interactive discussion text data and the cognitive level features, and therefore corresponding interactive text vectors and cognitive level vectors are obtained; aiming at the behavior sequence data, mainly adopting an Item2Vec algorithm to encode the behavior sequence data so as to obtain a corresponding behavior sequence vector, and fusing the behavior sequence vector with a cognitive level vector to obtain a fusion vector with more complex characteristics;

5) stage S5 of deep neural network model training: using different model parameters, the training set X_(train)Inputting the data into the deep neural network model constructed in the step 1) for multiple times in small batches, and training the constructed model; wherein, the update gate and the reset gate in the BiGRU use a sigmoid activation function, when the model is output, the used activation function is sigmoid ', the optimizer of the model selects Adam', and the used loss function is a cross entropy loss function;

6) stage S6 of deep neural network model testing: testing data set X obtained in the step 2)_(test)Inputting the data into the trained deep neural network model in the step 5) to obtain two classification results of the learning behavior data of the learner;

7) evaluation of learner meta-cognitive ability stage S7: and (3) taking the binary classification result obtained in the step 6) as the probability p of the level to which the meta-cognition ability belongs, and dividing the level of the meta-cognition ability of the learner into a higher level of the meta-cognition ability and a lower level of the meta-cognition ability by taking a middle value of 0.5 as a boundary.

The cognitive level obtained by the analysis according to the brucm cognitive target classification theory in the step 2) is specifically represented as follows: memory, understanding, application, analysis, evaluation, and creation;

in the step 4), the behavior sequence vector V, the cognitive level vector C and the fusion vector M which is subjected to fusion to obtain more complex characteristics are fused, the fusion mode is the splicing of the vectors, and the specific calculation is as follows:

M＝contact(V，C)

wherein contact represents the concatenation of vectors;

the effect measurement indexes for the learner meta-cognitive ability evaluation in the step 7) are as follows:

TABLE 1 Meta-cognitive ability evaluation model Effect metrics

TP (True Positive, abbreviated as TP) is the number of Positive samples correctly evaluated by the evaluation model;

FP (False Positive, abbreviated as FP) is the number of the misclassified Positive samples;

TN (True Negative, abbreviated TN) represents the number of correctly classified Negative samples;

FN (False Negative for short) represents the number of misclassified Negative samples.

When the meta-cognition ability is layered according to the general theory P, the specific principle is that when the probability value P is larger than 0.5, the meta-cognition is higher, and when the probability value P is less than or equal to 0.5, the meta-cognition is lower.

The meta-cognition ability evaluation model based on the online learning behavior and the deep neural network provided by the invention has the following beneficial effects: the online learning behavior and deep neural network model can automatically, accurately and intelligently evaluate the meta-cognitive ability of online learners. In addition, the Item2Vec algorithm is adopted in the model to carry out vectorization representation on the behavior sequence data, the inherent semantic information of the behavior data can be reserved, and the meta-cognition feature extraction from various online learning behavior data is considered, so that more accurate meta-cognition capability assessment can be realized. The method can realize more accurate and visual representation of the current meta-cognition ability state of the online learner, and the meta-cognition of the learner is displayed by learning behavior data, so that a reference is provided for teachers to better realize layered teaching.

Drawings

FIG. 1 is a flow chart of a meta-cognitive ability assessment model for explaining the contents of the invention based on online learning behaviors and a deep neural network;

FIG. 2 is a schematic diagram of a deep neural network model constructed in the present invention.

Detailed Description

The invention is described in further detail below with reference to the following drawings and specific embodiments.

As shown in fig. 2, the meta-cognitive ability assessment model based on online learning behavior and deep neural network provided by the present invention is mainly formed by a two-channel structure formed by a bidirectional gated cyclic unit (BiGRU) fused with Self-Attention mechanism (Self-Attention); the method can evaluate the meta-cognition ability of the learner according to the online learning behavior of the learner, and mainly divides the learner and the meta-cognition ability into a person with higher meta-cognition ability and a person with lower meta-cognition ability. The system specifically comprises an input layer, a vector layer, a feature extraction layer, a feature fusion layer and an output layer, and is specifically implemented as follows:

1) an input layer: a batch of online learners are selected, basic learning behaviors such as 'login', 'check-in' and the like are extracted one by one from the MOOC online learning platform according to the ID of each learner, and the basic learning behaviors are arranged into behavior sequence data according to the time sequence of the occurrence of the behaviors. Meanwhile, a crawler program is compiled by means of a Python + Selenium frame to crawl interactive text discussion data issued by a learner in a course discussion area, and the text data are analyzed according to a Broumu cognitive classification theory to obtain the cognitive level of the learner contained in the text data. And finally, inputting the behavior sequence, the cognitive level and the interactive text into the model together.

2) Vector layer: the input data is converted from unstructured text data into a spatial vector representation which can be recognized by a computer. Aiming at learning behavior sequence data, adopting Item2Vec algorithm to carry out each behavior i_iConverted into corresponding d-dimensional behavior vector v_i∈R^dFor each piece of behavior sequence data I_i＝{i₁→i₂→…→i_nThe corresponding behavior vectors are combined into a behavior sequence vector V_i∈R^n×d. For cognitive level, One-hot coding is firstly carried out, then the cognitive level is mapped to a high-dimensional vector space to form a lookup table, and the One-hot coding of each cognitive level corresponds to a d-dimensional vector in the lookup table to represent C_i∈R^dThe method considers the behavior sequence characteristics of the learner on the basis of the current cognitive level of the learner, and fuses the learning behavior sequence and the current cognitive level to obtain fusionResultant eigenvector M_i∈R^(n+1)×d. Aiming at interactive text data, a Word2Vec model is adopted to convert sentence text consisting of n words into d-dimensional Word vector representation W_i∈R^n×d。

3) A feature extraction layer: mainly comprises a Self-extension mechanism and a BiGRU network. The Self-Attention mechanism can better focus on important learning behavior sequences or important words related to meta cognition in the behavior sequence data. Two vectors M obtained for vector layer output_i∈R^(n+1)×dAnd W_i∈R^n×dFurther inputting the information into a Self-Attention mechanism to respectively obtain corresponding Attention weight values, and then weighting the Attention weight values to the vector to obtain a vector N endowed with Attention characteristics_iAnd S_i. The bidirectional GRU network can better extract the time sequence in the behavior sequence and the context semantic information in the interactive text. Vector N to be output through attention mechanism layer_iAnd S_iFurther inputting the feature vector into a BiGRU network to obtain a feature vector U_iAnd G_i∈R^k。

4) A characteristic fusion layer: the method mainly extracts the meta-cognition feature vector U from the two-channel Att-BiGRU model_iAnd G_i∈R^kFusing, outputting fused characteristics Z belonging to R through a full connection layer^k。

5) An output layer: and finally outputting the probability p of the category to which the meta-cognitive competence hierarchy of the learner belongs through a softmax function according to the output of the feature fusion layer. When the probability value p is greater than 0.5, the higher meta cognition is the one, and when p is less than or equal to 0.5, the lower meta cognition is the one.

The probability p is calculated as follows:

p＝softmax(W_cZ+b_c)

and finally, calculating a loss function value by using the real label, and further performing parameter optimization by using a random gradient descent algorithm.

The loss function is calculated as follows:

wherein the content of the first and second substances,the probability that the meta-cognitive ability of learner d belongs to category j. The training goal of the model is to minimize the cross-entropy loss function.